CHAPTER 12

A Moving Grid Method for One-Dimensional PDEs

Based on the Method of Lines*
J. G. Verwert
J. G. Blomt
R. M. Furzelandt
P.A. Zegelingt

Abstract. Even in one space dimension the numerical solution of time-dependent partial
differential equations is often complicated due to large local gradients in the solution that evolve in
time. The sharp moving gradients limit the efficient application of easy-to-use method of lines
schemes that work on a fixed space grid. In such a situation the use of an adaptive or moving grid
can often improve the efficiency and accuracy of the numerical computation. The method described
in this paper integrates in a moving reference frame. The grid movement is based on the principle of
spatial equidistribution of nodes and is regularized by employing a grid-smoothing technique in
space and time. The spatial grid-smoothing ensures that the ratio of adjacent grid intervals is res-
tricted, thus controlling clustering and grid expansion. The temporal grid-smoothing serves to obtain
a smooth progression of the grid for evolving time. The spatial discretization is based on standard
central differencing since we aim at a large problem class. For the numerical integration in time we
use a sophisticated BDF code. In many cases this stiff solver can be used in a similar easy way as
on a fixed grid. In other, more difficult cases, some parameter tuning may be required to optimally
govern the grid movement. The performance of the method is numerically illustrated.

1. Introduction. We consider systems of partial differential equations (PDEs) in one space

dimension,
u1 = f (u,x,t), (I.la)
with the initial and boundary conditions

* For the CWI/Shell project 'Adaptive Grids', P.A. Zegeling has received support from the
'Netherlands Foundation for the Technical Sciences' (STW), future Technical Science Branch of the
Netherlands Organization for the Advancement of Research (NWO) (contract no. CWI55.092). J.G.
Verwer has received financial support from the US Army Research Office in London which made it
possible for him to participate in the Workshop at RPI (ref. no. R&D 6168-MA-06).
tcentre for Mathematics and Computer Science (CWI), P.O. Box 4079, 1009 AB Amsterdam,
The Netherlands.
*Koninklijke/Shell-Laboratorium, Amsterdam (Shell Research B.V.), P.O. Box 3003, 1003 AA
Amsterdam, The Netherlands.

160
 MOVING-GRID METHOD 161

u(x, 0) = u 0 (x), XL < x < xR and b(u,x,t) = 0, x == xL, XR, t > 0. (l.lb)
Here f and b are spatial differential operators and it is tacitly assumed that the problems under con-
sideration are well-posed and that they possess a unique solution. The differential operator f is sup-
posed to be of at most 2-nd order. In particular, we are concerned with problems with disparate
space and time scales giving rise to solutions with large space-time gradients. However, we do not
consider genuinely discontinuous shock solutions as those arising in first order hyperbolic problems.
Problems with disparate space and time scales occur in many applications from the engineering sci-
ences and often an adaptive or moving grid can improve the efficiency and accuracy of a numerical
computation.
The method described here is based on the method of lines (MOL) which is a well-known
approach for numerically solving PDE problems such as (1.1). In the MOL approach the discretiza-
tion of the PDE is carried out in two stages. In the first stage the space variables are discretized on
a selected space mesh, normally chosen a priori for the entire calculation, so as to convert the PDE
problem into a system of, usually stiff, ordinary differential equations (ODEs) with time as indepen-
dent variable. The second stage then deals with the numerical integration in time of this stiff ODE
system to generate the desired numerical solution. With this MOL approach in mind, several sophis-
ticated PDE packages have been developed in recent years, notably for one-space-dimensional prob-
lems (see e.g. [2, 3, 8, 10, 11, 14, 15]). These MOL packages greatly benefit from the very successful
developments of automatic stiff ODE solvers. In particular, the implicit Gear-type BDF solvers
play a prominent role here. Gear-type solvers have proved to be efficient, robust and reliable, in that
they work for a broad class of problems and usually solve the stiff ODE system under consideration
in an accurate and efficient way. The experiences with MOL packages have revealed clearly that this
is also true of semi-discrete PDE problems on fixed space grids. However, for solutions possessing
large space-time gradients, like travelling wave fronts or emerging boundary and interior layers, a
grid held fixed for the entire calculation can be computationally inefficient, since this grid will
almost certainly have to contain a very large number of nodes. In such cases, a moving grid pro-
cedure that attempts to adjust automatically both the space and the time-stepsizes is likely to be
more successful in efficiently resolving critical regions of high spatial and temporal activity.
The method described in this paper is of Lagrangian type and, at the semi-discrete level,
automatically moves continuous-time grid lines to regions of high spatial activity. The grid move-
ment underlies the principle of spatial equidistribution of nodes and employs regularization tech-
niques borrowed from Dorfi and Drury [4]. The spatial discretization is based on standard central
differencing since we aim at a large problem class. For the numerical integration in time we use a
sophisticated BDF code [2, 3, 11 ]. From the users point of view it is of interest to note that this stiff
solver can be used in a similar easy way as in the conventional (non-moving) approach. Some
parameter tuning is required to govern the regularization of the grid movement as well as to optim-
ise the efficiency. Needless to say, tuning is an important issue since the need for tuning is in
conflict with robustness and ease of use. The numerical study of [7), where a comparison is
presented between our current method, the adaptive moving-grid method of Petzold [12], and the
moving-finite-element method (MFE) of Miller, shows that in this respect the current method com-
pares favourably with the MFE method.
In Section 2 we introduce the semi-discretization in a moving reference frame, completely in
line with the common MOL approach. In Section 3 we give the moving-grid equation that deter-
mines the continuous-time grid trajectories implicitly in terms of the semi-discrete solution on this
grid. Section 4 is devoted to a discussion of the two grid-smoothing procedures that are used to reg-
ularize the grid movement. In Section 5 we discuss the complete semi-discrete system and its numer-
ical integration. Section 6 presents results of numerical experiments with three different example
problems and the final Section 7 is devoted to a brief conclusion.

2. The semi-discrete PDE. Virtually all of the space mesh adapting techniques for time-
dependent problems attempt to move the nodes in such a way that, in regions of high spatial
activity, there is enough spatial resolution. In other words, the construction of these methods is
aimed at minimizing the number of space nodes relative to a certain level of spatial accuracy. On
the other hand, in most time-dependent applications large spatial gradients are accompanied by
large temporal gradients, the standard example being provided by the simple running wave form
u(x,t) == w(x -et). It is thus natural not only to minimize the computational effort put into the
162 Verwer et al.

spatial discretization, but also to attempt to minimize the computational effort put into the time
integration. Note that on a non-moving mesh a steep wave form such as u(x,t) = w(x -et) will
require standard time-stepping techniques, including the sophisticated Gear methods, to use small
time-steps. The reason for this is that as the moving front passes a grid point, the solution at this
grid point will change very rapidly and so small time steps are then necessary to retain accuracy.
The above observation naturally leads one to consider the Lagrangian discretization approach where
the grid is moved continuously along with the solution with the aim of reducing these rapid transi-
tions. Note. however, that it is not always possible to reduce them simultaneously in space and time
(see [7, 16) for a more comprehensive discussion).
We start our derivation at the semi-discrete level. Thus, completely in line with the common
MOL approach, consider smooth, continuous-time trajectories
XL = Xo < ... < X;(t) < x.+1(t) < ... < XN+l = XR for t;;;.. 0, (2.1)
which are, as yet, unknown. Introduce, along x(t) = X,(t), the total derivative
u' = x'ux + Ut = x; Ux + f (u,X;(t),t), 1 ,.;; i ,.;; N, (2.2)
and spatially discretize, for each fixed t, the space operators aI ax and f so as to obtain the semi-
discrete system
u; = x; [(U;+J - u,_i)/(X;+1 - x,_i)J + F;, t > o. I,.;; i,.;; N. (2.3)
As usual, U;(t) represents the semi-diserete approximation to the exact PDE solution u at the point
(x,t) = (X;(t),t) and F, is the finite difference replacement for f (u,x,t) at this point. Note that the
standard, central difference approximation for ux is used. It is supposed that F; is also based on
standard, 3-point, central differencing. Further it is of interest to observe that at this stage of
development the only errors introduced are the space discretization errors. With the associated grid
functions
X = (Xi. · · ·, XNf, U = [Uf, · · ·, Uk]r, F =[Ff, · · ·, Fkf,
D; = (U;+1 - U;-1)/(X;+1 - X;-i), D = [Df, · · ·, Dkf,
we reformulate (2.3) in the more compact form
U' = X' D + F,
0 t > 0, U(O) given, (2.4)
which represents the semi-discrete system to be numerically integrated in time. The notation X'oD
means that Xi is to be multiplied with all components of the vector D1•
In the discussion to follow, we neglect the treatment of boundary conditions, since these are
dealt with in the usual way. We also wish to emphasize that for convection-diffusion problems with
steep gradient or near-shock behaviour, the use of central differencing of first order terms is not
ideal and one would probably consider stable upwind or flux-corrected approximations, since other-
wise any deviation from an ideal Lagrangian grid movement, assuming this exists, readily results in
unphysical oscillatory solutions. It is emphasized that the actual generation of the moving grid is the
central issue here and that other spatial discretizations can be easily implemented.

3. The moving-grid equation.

3.1. Spatial equidistribution. We shall construct an equation that defines the time-dependent
grid X implicitly in terms of the continuous-time solution U. This grid equation underlies the fami-
liar notion of spatial equidistribution. Introduce the point concentration values
n; = (AX;)- 1, !:l.X; = X;+ 1 - X1, 0..;; i..;; N, (3.l)
and the spatial equidistribution equation
n;-1/M;-1 = n;IM;, 1..;; i..;; N, (3.2)
where M; ~ a > 0 represents a monitor value that reflects spatial variation over the i-th subinterval
[X;, X;+d· Typically, M; is a semi-discrete replacement of a solution functional m(u) containing
one or more spatial derivatives. For example, the 1-st derivative functional (in scalar form; the
change for systems is obvious)
 m{11) =(a + (u,) 2
central

> 0 sen.es to en;,1.ue that

"~'""'""''"" <i 1~
leads to the wdl-k11ow11 monitor
unifonn intervals. Ail re~ults '""'"w•~·•
the monitor or its modifo.:ation
tion curvat1.m:I could be used.

3.2. The grid-smoothing prol:t>d111res.

U However, as well known. for ,.,.,,.,.,,,-,,1
equidistribution needs to
this we now introduce two g,nd-~.rm:iothirig
ing a smooth grid, and the other for :·W1)u1mo
grid-smoothing procedures amounts
modifications and delay a more nm-.r.•hM•<i,:••
The grid-smoothing is
numerically 'anti-diffu:;ed' counterpart~
iiu .:::: II() - 11.iic+ l Xn i -- nri).
ii, = n, - K(K-:- I )(n, +1 - 2n, +- n, _1 ), K > 0. I ~ 1 ~ ,'I/ - I.
nN = l'IN -- K(K + !)(nv ·- l - llN ),

which results in the now

n; 1/ kl, _ 1 = ri, M,.
The first and last equation in involve the 'zero concentration gradient' boundary conditions

where n .. and nN + 1 correspond to the artificial X 1 and ~ 2 , respectively. In [7], and

also in {4], the similar conditions n 0 = n 1 , nN ... 1 =
nN have been used. However, these imply that
the first and last monitor values, and Af,v, respectively, are removed from the moving grid equa-
tion (in (3.6) the index i then runs from 2 to N -· l). This is not appropriate in cases where the
boundary monitor values are much larger than the interior ones, like, e.g., in Problem I of Section 6
during the generation of the steep flame front at the right boundary. The present boundary condi-
tions overcome this deficiency.
The introduction of the 'anti-diffused' point concentrations is equivalent to a certain smooth-
ing procedure for the monitor function (see Section 4), thus ensuring that the adjacent point concen-
trations are restricted such that

(3.7)
This condition implies that the grid we compute is locally bounded and, most importantly, provides
a natural way to control clustering and grid expansion. While the monitor function detenhines the
relative shape of X, the value of " and .V determine the level of clustering. Further. for a given N
and a given monitor function distribution, the choice of K detennines the minimum and maximum
interval lengths. In actual application, a value of K of about I or 2 is recommended so that modestly
graded space grids are obtained. In all our experiments we have used the (rather conservative)
default value K =
2. Recall that the grading of the space grid plays an important role in controlling
space discretization errors (see, for example, [6J).
When combined with the spatial grid-smoothing, the temporal grid-smoothing is effected by
replacing the system of algebraic equations (3.6) by the following system of differential equations
(ii,_, + Tiii-dlM,-1 =<ii,+ Tni)1M,, 1 > o, 1.;,:; j.;;; N. (3.8)
The introduction of the derivatives of the point concentrations serves to prevent the grid movement
I 64 Verwer et al.

from adjusting solely to new monitor values. Instead, the use of (3.8) forces the grid to adjust over
a time interval of length -r from old to new monitor values, i.e. the parameter -r acts as a delay factor
(see Section 4). The aim here is to avoid temporal oscillations and hence to obtain a smoother pro-
gression of X(t). These oscillations can arise in grids generated via spatial equidistribution tech-
niques, because when applied to solutions with extremely large gradients, the numerical monitor
values are very sensitive to small perturbations in the grid and vice versa. With oscillatory trajec-
tories it is certain that near steep fronts one or more components in the ODE system rapidly vary
for evolving time. This is detrimental for the numerical time stepping and also causes difficulty in
the Newton solution of the sets of nonlinear algebraic equations that arise in the implicit time
integration with the stiff solver.
In contrast to the choice of K, the choice of a good value for T is less simple. Increasing T too
much results in a grid that lags too far behind any moving spatial transition. In fact, for sufficiently
large values of T a non-moving grid results. Fortunately, our numerical experience (see Section 6)
indicates that in many situations temporal grid-smoothing is actually redundant. We owe this to the
spatial grid-smoothing which also helps to prevent the grid from oscillating. However, in situations
where smoothing in time is advisable, it makes sense to attempt to choose -r close to the anticipated
temporal step size value such that, over one or a few time levels, the influence of past monitor
values is felt. The discussion of the next section is aimed at providing more insight in this matter.

4. Discussion of the smoothing procedures.

4.1. Preliminaries. Equations (3.8) are based on the relation

t > 0, 0.;;;; i.;;;; N, (4. la)
where c = c(t) is the proportionality constant involved. This proportionality constant is solution
dependent and in fact also depends on the parameters -r and ic. This dependence is suppressed in
our notation and we shall use c(t) as a generic notation for, possibly, different constants of propor-
tionality. Using p. =
K(K+ 1), we first rewrite in (3.5) as n;
no= - p.111 + (l+p.)no,
n; = - + (1 + 2p.)n; -
p.ll; + 1 µ.n; - i, l .;;;; i .;;;; N-1, (4.lb)
nN = - fJJIN-1 + (l+µ.)nN.
For initial conditions we suppose a given concentration distribution n;(O), 0 .,;;;;; i .,;;;;; N, that has been
subjected already to the spatial grid-smoothing procedure, i.e. the initial grid satisfies (3.6) at t = 0.
For the actual practice this is a natural assumption because the space smoothing is also applied at
later times. Violation of this assumption makes it likely that already within the first time-step the
grid is forced to undergo a large change. However, in principle, an initial grid not satisfying (3.6)
can be used.
We have N + 1 equations for the N + 1 unknowns n;, 0 .;;;; i ..;;; N, if we consider the propor-
tionality constant c(t) and the monitor values M;(t) as being given. In fact, for the analysis
presented in the remainder of this section it is convenient to uniquely represent the N + l concentra-
tions n;(t) for t ;a.. 0 in terms of the initial concentrations n;(O) and the values c(t), M;(t) as
described below. First, solving (4.la) yields the nonlinear Volterra integral equation system
I

ii;(t) = e-tlT[n;(O) + [-r- 1es 1•c(s)M;(s)ds], t ;a.. 0, O.,;;;;; i.,;;;;; N, (4.2)

where n;(O) is determined by n;(O) through (4.lb). We have a system of nonlinear Volterra integral
equations because the monitor function values M; depend on all concentrations in a nonlinear way.
Second, the matrix M associated to the system of linear equations (4.lb), i.e.
(4.3)
is symmetric, positive definite. Hence, M is non-singular and the point concentrations n; are
uniquely expressed into by n;
(4.4)
 llilOViNG-GRID METHOD l 65

from system !4JJ

form of the :-o!u1ion c'f I.his svstem.

LEM1v1A 4. . Lei ~·

'I L 1 ··1' I'

; c l1

where

with, for k -== l and 2.

an = -11:1 D, all - I +K,oiD.

D = \ +· "-,v.
PROOF The characteristic of with (4.lb) has the
roots 11 and i· 1• so that the associated nomc1ge1ne(l1Js
..;. i ~ N,
where C ! , C: are con.stants. A recursion is
easily checked to be
N- I
11 1.pa, =(I +2K)- 1 2; vl 1 -11n1 , 0 ~ i ~IV,
J I

which is (4.51 with the firs: and last 1enn omitted. Hence, the general solution of(4.lb) reads
v i
n, = C 1 w·' ·1" C 2 1" +(I +2.x;) 1 v" .1lii 1, L
J ,, I

where the two constants C • C 2 serve to matdi the boundary conditions, i.e. the first and last equa-
tion of An elementary calculation kads 10 (4.5). The introduction of the auxiliary quantities
~~ only serves to express the solution in this specific fom1. 0

At first sight expression (4.5) is a bit complicated by the incorporation of the boundary condi-
tions. Neglecting these leads to the more transparent expression
N-1
n; = n+21er' 2:r:'-J!ii1 (4.6)
j =l ,

given in 14]. The rele\ant point in all this is the appearance of the 'smoothing kernel' 11: 1 - 1 i. Note
that 0 < I' < I.
Next the equidistribution equation (4.la) is taken into account, i.e. we now simply substitute
ii1 =--= cAf1 into (4.5) to obtain
s
n, = c( l + 21e)- l 2; ~·ii : lv!1, 0,;;;; i ,;;;; N. (4.7)
} :.::0
- -
where M1 = Af1 for I ,;;;; j :;;;;; N - l and Mo and l'>1N are defined in exactly the same way as Vo
-
and V.v in (4.5). Like\i\ise, (4.6) then reads
lV - I
n, == c( I +211:) 1 2; pi' --,1, (4.T)
J ~I
l 66 Ven,ver et al.

COROLLARY 4. L the anti-diffused concentrations ii, proportional to lH, is equivalent to tak-

ing the concentrations n, proportional to the smoothed monitor value

A;= -.1i l't.-f.}' n

'--'

REMARK 4. I. A trivial consequence of the proportionality of n, to the positive 'monitor' values A 1 , is

that all concentrations n, remain positive which means that the spatial grid-smoothing cannot lead
to node crossing. Of course, this is also a direct consequence of the grid ratio condition (3.7).
Further it is of interest to note that all values ii, are positive too, which can be concluded from the
two following observations. first, all iii are either positive or negative, as they are proportional to
Second, if all ii, < 0, then all n, must be negative which is a contradiction. 0

The motivation behind the spatial grid-smoothing lies in the desirable grid condition (3.7)
which serves to control clustering and grid expansion:

THEOREM 4. I. The spatial grid-smoothing restricts the concentrations n, such that (3.7) is satisfied.
PROOF Consider (4.7). From the inequalities i -11 .:;;; Ii I
+I and 0 < 1• < ! we directly
deduce
,,. - N -
n,tn,+1 = v-1r::~:J1l·!•-;i+l 2:M1"11-;-llJ.,;;; P-1.
j =O J =O

because all terms in the numerator are smaller than or equal to the corresponding terms in the
denominator. In a similar simple way the left-hand side inequality of (3.7) is proved. 0

In the proof, the size of the 'monitor values' M; plays no role whatsoever. only the fact that
they are positive is used. As a matter of fact, for any randomly chosen set of positive values M,.
condition (3.7) is satisfied. This is an attractive feature \vith respect to robustness, but also makes it
difficult to precisely quantify the effect of the space smoothing on the original equidistributing grid.
An additional complicating factor. in this respect, is the effect of the 'zero concentration gradient'
boundary conditions. although ha"ing
(t ~ 0)
is a natural restriction and certainly advantageous with respect to spatial accuracy near the boun-
dary. Further, while neglecting the boundaries, the averaged expression (4.T) looks very natural.
Our practical experience is that the spatial grid-smoothing procedure leads to a point distribution
where the monitor function will determine the relative shape of the distribution and the value of "
and N the level of clustering. We refer to Dorfi and Drury [4] for a numerical illustration.
h is of interest to observe that, for a given N, the choice of " determines the minimum and
maximum interval lengths. In actual application, the minimum should be related to the expected
small scale features in the solution to be computed. Suppose that in a transition from small to large
space gradients and back, a solution requires a local refinement in a grid with a factor of 10"'. Let
N1tx: be the number of points in this transition region. Then, if the point concentration variation is
bounded by I + I / K, it follows from
(l+ l / K)0.5N... = HY",
that N1oc is at least
N1oc = 2mln(l0)/ln(l + l Ix:):::::; 4.6m /ln(I +I/ K). (4.8)
For example, for m =
3 and K = 1, 2, 3, we have, respectively. N10, ~ 20. 34 and 48. Note that
the factor of 0.5 above accounts for the fact that a local grid refinement is supposed to be followed
by a local grid expansion. Using the 'rule of thumb' (4.8), one can make a quick (but somewhat
crude) estimate of the number of points needed for a particular problem by summing the minimum
number required to solve each small scale feature [4].
 MOVING-GRID METHOD 167

REMARK 4.2. The range of summation in (4.7) may be changed without violating the grid ratio con-
~tion. For example, if only the direct neighbouring monitor values are used, n; becomes propor-
tional to
i+I
A;= ~ p\i-J\Mj = PM;-1 + M; + PM;+i. I .;;;; i.;;;; N-1,
j=i-1
while condition (3.7) remains valid. This suggests, for example, to realize the grid smoothing directly
via the rule
(X; - X;-1)A;-1 = (Xi+1 - X;)A;, 2.;;;; i.;;;; N-l. (4.9)
We have not tested this alternative. Note that this technique preserves the 3-point coupling in X, but
a drawback is that M; becomes coupled to M;- 2 , M;- 1 and M;+I· Another obvious alternative
which comes to mind is to perform the smoothing on the AX; values rather than on the point con-
centrations. The AX; values are then replaced by
G; = AX; - K(1e+ lXAX;+J - 2AX; + AX;-1), IC> 0,
so as to obtain the grid equation system
G;-1M;- 1 = G;M;. (4.10)
This smoothing procedure also leads to a grid X satisfying condition (3.7) and to slightly simpler
equations (certainly so after the temporal grid-smoothing). As yet we don't know whether this par-
ticular choice of smoothing is better or worse than that based on the point concentrations. D

4.3. Temporal grid-smoothing. In terms of equidistribution, temporal grid-smoothing means

that Tiif + ii; is taken proportional to the monitor values M;, as can be seen in equation (4.la). The
introduction of the derivative of the point concentration implies that the grid movement is no longer
dictated by solution values at the current time level t, but also depends on past solution values. By
preventing the grid movement from adjusting solely to new monitor values at time t, we hope to
introduce a smoothing effect so as to avoid oscillatory trajectories X;(t), t ;;.. 0.
Let us examine the solution for n;(t) in the following form (cf. (4.2)), where At represents a
typical stepsize that is taken in a numerical time integration:
I

n;(t) = e-b.t 11 n;(t-At) + j T- 1 e<•- 1>11 c(s)M;(s)ds, t ;;;;.. !::.t, 0 .;;;; i ...:; N. (4.11)
t-Al

We see that n;(t) is determined by the sum of e-Ai 11 ii;(t -t::.t) and a weighted average of values
c(s)M;(s) over the interval [t -t::.t,t]. The weighting is determined by the size of 'T and is exponen-
tially decaying for backward time values. One can see that 'T acts as a delay factor for the grid
movement and that the influence of past solution values is exponentially decaying.
For T-+0, n;(t)-+c(t)M;(t) whereas n;(t)-+n;(t -At) as T-+oo. It follows that for sufficiently
large values of T a non-moving grid results. This means that increasing 'T too much will result in a
grid that lags too far behind any moving steep spatial transition. On the other hand, too small
values for 'T render no smoothing effect. In situations where temporal grid-smoothing is advisable, it
makes sense to choose 'T close to the anticipated At-values, so that over one or a few time levels the
influence of past monitor values is felt. This suggests allowing 'T vary with !::.t. Note that so far we
have assumed that r is constant over the whole range of integration.
For an alternative interpretation of the smoothing in time procedure, it is illustrative to exam-
ine the implicit Euler discretization (l-st order BDF formula) of the equation
-rAXi (AX;)- 2 + (AX;)- 1 = cM;, t > 0, 0 .;;;; i .;;;; N, (4.12)
which arises from (4.la) by putting" = 0 and by substituting
dn;ldt = -AXfl(AX;)2.
Spatial grid-smoothing is omitted here to simplify the presentation. Observe that, apart from the
spatial smoothing, it is just this semi-discrete equation which is numerically integrated in time after
elimination of the constant of proportionality (see Section 5). Let y =
TI !::.t. Then the implicit
Euler replacement of (4.12) is given by
168 Vel'Vv'er et at

k ;;i. I. 0 <i.;;; N. (4.13)

where A.X,,1< is the approximation to A.X. at time I = tk, tk = tk _ 1 + At and t 0 = 0. This fully
discrete relation shows that, instead of taking (A.X1.k )- 1 proportional to M,,k, with numerical tem-
poral grid-smoothing we take the entire grid point expression at the left-hand side of (4.13) propor-
tional to M,,k. This term contains only grid values. The contribution from the previous time-level
should introduce the desired smoothing effect. For the special choice T = flt, the simple equidistri-
bution relation
(4.14)
results. Observe that for the higher order BDF formulas, similar equidistribution relations are found,
the only difference being that then AX;,k - I is replaced by a linear combination of such differences
over more previous time-levels.
Finally, the following result shows that smoothing in time does not interfere with the grid-
ratio condition (3.7):

LEMMA 4.2. The combined space-time grid-smoothing restricts the concentrations n; such that (3.7)
is satisfied.
PROOF For condition (3.7) to hold, the actual.size of the values Mj is irrelevant, according to the
proof of Theorem 4.1. It is sufficient that all M1 > 0. It thus suffices to prove that the solutions ii;
of the di!ferential equations (4.la), as given in (4.2), remain positive for all t ;;;;. 0, since this implies
that all M1 > 0 (see Lemma 4.1 ). First we recall that ii;(O) > 0, as shown in Remark 4. l. Now sup-
pose that at a certain time t' the constant of proportionality c(t) becomes negative (if c(t) > 0 for
all t, the proof is complete). Then, since M; > 0, a right neighbourhood of t = t' exists where all
n1(t) will decrease. Because all entries of the matrix M- 1 arising in equation (4.3) are positive (see
again Lemma 4.1 or observe that M is a Stieltjes matrix), all point concentrations n;(t) will also
decrease in this right neighbourhood. This is impossible since the interval [xL, xRl is fixed. Hence
we have a contradiction for the assumption that c(t) can be negative and the proof is complete. 0

REMARK 4.3. The temporal grid-smoothing discussed here is closely related to that suggested in
[l, 9]. The main difference lies in the fact that in [l, 9] the derivative of X; is introduced directly
into an equidistribution equation based on nodal values X;, whereas here the equation for the con-
centration values n; is modified. This leads to a different system of grid equations when written in
terms of X; and 0 x:.
5. The complete semi-discrete system.

5.1. The moving-grid equation in tenns of nodal values. Inserting

n; = (A.X;)- 1, ni = -AX;t(AX;)2 (5.l)
into (3.8) leads to the moving-grid equation system that is actually used. Its i-th equation reads,
2.;;; i,,;;;; N-1,

-'T[ M;-1(!::..X;-2 r]x:-2+ (5.2)

+T[ M;(flX,_if /J: + 1+2µ.

M;-1(AX;-d
2 +
M;-1(AX;-2
r]x:-1 +
[ µ. + I +2µ + l + 2µ + µ ] x~ +
- 7' M;(aX; -1 ) 2 M;(AX; 'l M; - I (AXl-1 )2 M; -1 (AX; 'l I

+ [ µ. + I +2µ + µ ] Xi + l +
T M;(~+if M;(!::i.X;f M;-1(!::i.X;f

- T [ M;(t::..i;+d ]x:+i =
 MOVING-GRID METHOD 169

= [- ___!__
&X,+1
+ .l±l£.
&x, - ___!_l/M
u, 1
- ' '
l- -L +
&x, ~
u,_, - __/L_]IM
ax1_ 2 ·- I
•

The 1-st and N-th equation slightly differ due to the boundary condition:s and are easily found. Note
that, away from the boundary, the nodal points X,+ 2 , X,+ 1• X,, X,_ 1, X, -i are coupled with the
nodal point velocities x; +l• Xi +i. x;. x; i. x; _2 and the monitor values M, __ 1, M,.
For future reference. system (5.2). together with the 1-st and N-th equation, is represented in
the form of the nonlinear ODE system
TB(X,U)X' = g(X.U) (5.3)
where B is the N x N penta-diagonal matrix associated to the left-hand side part of (5.2). In order
that we have a genuine ODE system, it is required that B(X,U) is non-singular for any X, U. If no
time smoothing is carried out. i.e. T = O. we are left with the algebraic system
g(X,U) = 0, (5.4)
which represents the equidistribution relation combined with spatial grid-smoothing.

REMARK 5.1. An alternative and somewhat simpler moving-grid equation system that has essentially
the same smoothing properties as (5.2) is obtained by putting µ. = 0 in its left-hand side. This
renders B tri-diagonal and symmetric positive definite. In terms of point concentrations, the result-
ing system reads Tn: + n,
= cM, (cf. (4.la)), which shows that the temporal grid-smoothing is car-
ried out on the concentration values n, rather than on n,. D

5.2. The complete semi-discrete system and its numerical integration. Systems (2.4) and (5.3)
together form the complete semi-discrete system that is numerically integrated in time,
TBX' = g. t > 0, X(O) given, (5.5a)
U' - X' D = F,
0 t > 0, U(O) given. (5.5b)
In case of Dirichlet boundary conditions, the total number of equations and unknowns is
(NPDE + l)•N. where NPDE is the number of components of the original PDE problem (1.1). For
other types of boundary conditions, the number of equations and unknowns slightly differs. The
supposed non-singularity of the matrix B trivially implies that for T > 0 we have a genuine ODE
system; for T = 0 we have a DAE system of index one. The large matrix that multiplies the deriva-
tives X', U' in (5.5) has a rather simple. lower block-triangular structure. We cannot exploit this
advantage since the system is numerically integrated with an implicit method. The Newton iteration
matrix involved contains the partial derivative matrices of g and F with respect to X and U, or
approximations thereof, and hence the lower block-triangular structure is lost. It is therefore compu·
tationally more attractive to change the order of unknowns so as to obtain a band-matrix. When
using the order · · · , U; _ 1, X, _ 1 • U;. X1, U; + 1• X1 + 1, • • • , the band-width for the Newton matrix
becomes 4•(NPDE + 1) + I. This is based on the fact that we work with standard 3-point central
differences for the spatial operators, that X is 5-point coupled, and that the monitor Mi is given by
(3.4).
For the numerical integration of the above semi-discrete system, one can use, in principle, any
stiff method designed to solve linearly implicit systems of the present type. The results of the next
section have been obtained with the BDF code DASSL (version of 830315) [ll). A similar code is
the LSODI-based BDF code of the SPRINT package (2, 3]. We have experimented with both these
codes (see also [7]) and since they are very much alike, the choice between the two should be of
minor influence to the performances observed. This indeed turns out to be true in the case of suc-
cessful runs. However, in some cases we have experienced a rather different performance. With both
codes and for different problems runs were interrupted due to fatal Newton errors, especially so
when using extremely fine grids. This could be due to the fact that in our experiments the local error
and Newton convergence test has been applied to Xi and not to &X1• Also, with moving grid
methods a poor prediction of Xi can be generated in the preparation of the actual BDF step, thus
causing convergence problems for the Newton solver. These aspects need further attention (e.g. in a
study along the lines of Petzold and Lotstedt (13)).
From the user's point of view it is of interest to note that DASSL, and likewise the stiff solver
I 70 Verwer et al.

of SPRINT, are used in the same way as in the conventional, non-moving MOL approach. Apart
from providing a subroutine for the semi-discrete system (numerical differencing for Jacobians was
used) and specifying the initial values and required output times, one must define only the local
absolute and relative error tolerances, atol and rtol, the desired local error norm, and an optional
initial time-step value At 0 • Throughout we have used atol = rtol : = TOL and the standard
weighted Euclidean norm; TOL and At 0 will be specified with the three example problems in the
next section.
The method parameters for the grid are N, the number of moving points, the grid-smoothing
parameters "and r, and the constant a of the monitor (cf. (3.4))

M, = l
a+ NPDE- 1
NPDE
::Z
J =I
(U,+ 1,1 - U;_/'fl(X;+1 - X;Y
]y,
(5.6)

The choice a = I yields the common arc-length monitor; this we have used throughout, unless
noted otherwise. For " the default val1.1e 2 was selected, while r was simply put equal to zero. Addi-
tional tests have shown that for the three example problems below the temporal grid-smoothing is
redundant, which is of course a favourable situation. We wish to emphasize, however, that for other
problems a positive value for -r may lead to a better performance. As observed previously, this
aspect deserves more attention.

6. Numerical examples. We present numerical results for three different example problems. In
the plots the solid or dashed lines represent accurate reference solutions (obtained from [16]) while
the marks represent the generated PDE approximations. Integration information, which serves to
show the time-stepping efficiency of the process, is presented in terms of STEPS = total number
of successful time steps, JACS = total number of Jacobian evaluations, and BS = total number
of back solves. The two latter quantities determine, to a great extent, th~ CPU time needed to com-
plete the integration over the specified time interval.

6.1. Problem I: The Dwyer-Sanders flame propagation model. This model, first proposed as a
test example in [5], simulates several basic features of flame propagation. It has two components, a
mass density 11 and a temperature v. The PDE system is given by
311/3t = 32u/3x 2 - uf(v), 0 < x < l, 0 < t .i;;; .006,
3v!3t = 32 vl3x 2 + uf(v), 0 < x < I, 0< t .io;; .006,
where f (v) = 3.52• IO"exp(-4/v). The initial functions are u(x, 0) = l, v(x, 0) = 0.2 (0 or;;; x .io;; I)
and the boundary conditions are given by
3u/3x (0,t) = avtax (0,t) = 0,
3u!3x (l,t) = 0 and v(l,t) = 0.2 + t 10.0002 (t .io;; 0.0002), v(l,t) = 1.2 (I ;;;;. 0.0002).
The given function for v at the right boundary represents a heat source that generates a steep flame
front. When v reaches its maximum, this front starts to propagate from right to left at a relatively
high speed. The speed of propagation of the front is almost constant. At the final time t = 0.006,
the front has come close to the left boundary.
The initial grid X(O) was taken uniform with N = 40. A uniform start grid provides a difficult
test since the method rapidly must refine near x = 1 in order to accurately simulate the fast genera-
tion of the front. The remaining parameters to be specified are At 0 = 10- 6 and TO L = 10- 4 • In
passing we note that the error control mechanism of DASSL may reduce the specified initial stepsize
At 0 • In the present experiment At 0 was reduced to .1276• 10- 6 •
Fig. 6.1 shows plots of the grid and the computed temperature front for a range of output
times. The costs of the run amount to STEPS = 148, BS = 410, JACS = 52. Inspection of the
plots justifies the conclusion that the grid movement and the accuracy of the approximation are very
satisfactory over the entire time interval (also for the density which is not shown here). The small
lump for early times is genuine and is contaminated with only very little overshoot (not visible
here). For later times the numerical front is slightly too fast. These small errors are spatial, i.e., they
remain if many more time steps are spent and disappear if more space points are used. For exam-
ple, for N = 80 and TOL = 10- 4 , which costs STEPS = 164, BS = 492, JACS = 66, the
 MOVING-GRID METHOD 171

-
'o a.
.& ID ..
~

Cl
.n ..
~

Cl
.;. ...0

... ,.;
Cl -.
. .. co
x
~
Cl

>
Cl
C'll •
0

N
..;·

Cl
0
0.6 0.8 1.0 o.o 0.2 0.1 0.6 0.8 1.0
x x
FIGURE 6.1. Problem I (N =
40). Grid and temperature front at times
t = .15*10- 3 , .3*10- 3 , .6*10- 3 (.6*10- 3> .6*10- 2 •

approximations are exact up to plotting accuracy. Admittedly, 80 moving points for this problem is
quite a lot. It turns out that a relatively large number of points are wasted in the front, especially
for N = 80, while there are not too many near the foot and the top. We owe this to the arc-length
monitor. A comparison with results shown in [16], where a second derivative monitor is used that
deemphasizes the front and places more points where the curvature is largest, suggests that imple-
mentation of a second derivative monitor in the current algorithm would improve the spatial accu-
racy.

6.2. Problem II: A 'hot spot' problem from combustion theory. This problem is described in
Adjerid & Flaherty [ l] as a model of a single-step reaction with diffusion and reads
au1at = a2 u/ax 2 + D(I +a - u)exp(-o/u), 0 < x < l, t > 0,
au1ax (0,t) = 0, u(l,t) = 1, t > 0,
u(x,O)=l, Oo;;;;xo;;;;I,
where D = Re8 I (ao) and R, o, a are constant numbers. The solution represents a temperature of a
reactant in a chemical system. For small times the temperature gradually increases from unity with
a 'hot spot' forming at x = 0. At a finite time, ignition occurs, causing the temperature at x 0 to =
increase very rapidly to l + a. A flame front then forms and propagates towards x I at high =
speed. The degree of difficulty of the problem is very much determined by the value of 8. Following
[I, 7, 16], we have selected the problem parameters a = l, o = 20, R = 5. The problem reaches a
steady state once the flame propagates to x =
1. For the current choice of parameters, the steady
state is reached slightly before time t = 0.29, which we take as the end point. We use times
t = 0.26, 0.27, 0.28, 0.29 for output. It is noted that for t =
0.26 the reference solution is not
sufficiently accurate near x = 0, but it is very accurate for the remaining output times [16].
For the numerical process, two solution phases should be distinguished, viz. the formation of
the 'hot spot' with the flame front (the ignition phase) and the prop?.gation of this front to the right
end point x = 1 (the propagation phase). Accurate handling of the formation of the 'hot spot' and
the ignition is of importance. The ignition proceeds very rapidly, causing a widely different time
scale, so that variable steps in time are a necessity. A difficulty is that the code must detect the start
of the ignition very accurately at the right time, so that the step size can be rapidly reduced to a
level small enough to simulate this ignition in a sufficiently accurate way. Small errors at this time
point result in significantly larger global errors later on. Some trial and error tests have revealed that
the BDF code needs at least a time tolerance value TOL of 10- 5 , while using an initial step size of
10-s [7]. These are the values we have used. The small tolerance does not cause any problems with
 172 Verwer et al.

I0 .
0

! ..
~

•
=:::
0 ;:.
..
-i

.... x
!
0 ..
:::> '":

! ..
"!

! o.o 0.2 0.1 0.1 0.1 1.0

- o.o
"!
0.2 0.1 0.1 o.e 1.0
x x
FIGURE 6.2. Problem II (N = 40). Grid and flame front at times t = .26, .27, .28, .29.
the high-order integrators.
Figure 6.2 shows a plot of the computed grid and the flame front on this grid for the four
specified output times, using 40 moving nodes. The costs of this experiment amount to
STEPS = 136, BS = 382, JACS =35. The 'hot spot' nature is clearly visible from the grid. The
numerical flame appears to be too slow, but is almost in the right position for t = .27 and .28 (the
plot at t= .29 is the steady state solution). As for the previous problem, it is the spatial error that
dominates and decreasing TOL gives no further improvement. Changing N to 80 yields a very accu-
rate solution (up to plotting accuracy), while there is no great increase in the number of time steps,
viz. STEPS =159, BS =
423, JACS =
37. Inspection of the solution shows that, similar as for
Problem I, there are quite a few points in the flame front, but not very many at the top. Also here a
curvature monitor would improve the spatial accuracy, see [16] for comparison. Finally we refer to
[7] where results for a range of values T > 0 are shown.

6.3. Problem ID: Waves travelling in opposite directions. Our third example problem is a two-
component, semi-linear hyperbolic system, the solution of which is given by two waves travelling in
opposite directions (copied from [10], see also (7, 16]). The system is
oulot =- oulox - IOOuv,
ovlot = + ovlox - IOOuv,
fort> O'and -0.5 < x < 0.5, and the solution is subjected to homogeneous Dirichlet boundary
conditions and to the initial condition
u(x, 0) = 0.5(1 + cos(IQ.zrx)) for x E [-0.3, -0.l] and u(x, 0) = 0 otherwise,
v(x, 0) = 0.5(1 + cos(IQ.zrx)) for x E [ +0.1, +0.3] and v(x, 0) = 0 otherwise.
Note that these are functions with a mere C 1 continuity, which represent wave pulses located· at
x = -0.2 and x =0.2, respectively. Initially, while the pulses are separated, the nonlinear term
lOOuv vanishes, so that for t > 0 these waves start to move with speed 1 and without change of
shape, u to the right and v to the left. At t = 0.1 they collide at x = 0 and the nonlinear term
becomes nonzero, resulting in a nonlinear interaction leading to changes in the shapes and speeds of
the waves. Specifically, the crests of the waves collide a little beyond t = 0.25 and they have
separated again at approximately t =
0.3, so that from this time on the solution behaviour is again
dictated by the linear advection terms. At the nonlinear interaction, the pulses lose their symmetry
and experience a decrease in amplitude.
DASSL has been applied with N = 40, TOL =
10- 3 and M0 = 10- 5 . For convenience, we
have again used a uniform start grid. However, unlike the two previous problems, this uniform grid
does not satisfy the constraint (5.4) which it should if 'T =
0. To circumvent this start up difficulty,
 MOVING-GRID METHOD 173

"'d

.
d

...d

""' N
d

a
0
-ll.5 -0.3 -0.! 0.1 0.3 0.5
x
ci
t~ -
ci

..
0
t
l
I I
I
-1-
..
ci
I I
I
t I

"'...; I
I
I
..
I ..,
ci
t-:.
x t I t-:.
><
..
I
I
0 :::> .. I => ...
.,; Q I
I d
I
I
I
+I
fic; i
I

t'
N N
Q d

0 a
d d
-o.s -0.3 -0.1 0.1 11.:!I o.s -0.S -0.3 -0.1 0.1 0.3 o.s
x x
ci ci

.
•
~
d

...d
...
'"
...d
§ t-:. t-:.
~ >< ><
=> .. => ...
~ 0 d

!:::
iQ N
d
N
d
~
0 Cl
d d
-o.s -0.l -0.1 0.1 0.3 o.s
~
0.1
x x

FIGURE 6.3. Problem UI (N = 40). Grid and solution at times t = O.l,0.25,0.3,0.5.

 J 74 Vervver et al

pu! " :.mall \ ' so that we are in an ODE situation and an~· grid can be used to
DASSL then lowers our of ~t 0 to . 3* 10 · 0 and completes the
integration l l ! successful steps {46 up to t 327 back-solves and 78 Jacobian evalua-
tion~. The vahie r is of c,~urse small. so that, verv soon after the start, we are
dose 10 !he r :cc si!uati,m. ll is =
that if ,. 0 and. we start on a grid satisfying
or dioose r in case of a uniform start grid, the number of required steps will
he smaller ! see abo
Fig. 6.3 shows the grid and the numerical approximations at the specified output times. We
see !hat the movem.:nt mimics the interaction an<l point out that the \isible inaccuracies
are due to a somewhat choice for TOL and the number of points. These inaccuracies will
vanish if more points are used and again we remark that a curvature monitor would probably lead
to significantly more accuracy (see (16]). In the present experiment we have replaced the (regulari-
zation) constant a = I of the arc-length monitor by 0.1. The reason is that when the waves have
separated they are no longer very steep. with the result that the value 1.0 is somewhat too large for
obtaining sufficient refinement in the vicinity of the two waves, at least for N = 40. With this
number of points it is also necessary that, after the pulses separate. the grid refines properly in the
vicinity of the waves, else spurious oscillations become visible. Recall that after the separation we
are just solving the first order hyperbolic model problem using standard central differences. This
experimenl shows that it is de~irable that the regularization constant of the arc-length monitor func-
twn be made solution-dependent, in some way or another. On the other hand, the results published
in l Hi] indicate that with a curvature monitor this is less important.

7. Conclusions. This work has been carried out in connection with a joint CWI/Shell project
on 'Adaptive Grids'. One of the aims of this project is to develop a reliable, robust and efficient ID
moving-grid method, based on the method of lines, which can be used in almost the same easy way
as existing MOL packages that integrate on a non-moving grid. The demand of ease of use requires
that. as far as possible, the user should be relieved from fine tuning the grid movement. The results
obtained so far justify the conclusion that the technique discussed in this paper goes a long way
towards fulfilling the above requirements.
An important feature is the grid-smoothing capability involving the two method parameters "
and r. The meaning of K is very dear and for general use K can be taken equal to, say. l or 2. At the
present stage of development, the actual choice to be made for r is less dear. Fortunately, our
numerical experience indicates that in many cases it is possible to simply put T =
0 or to select T
really small. so that the grid movement is almost exclusively dictated by the spatial equidistribution
at the forward time level. The numerical results also suggest very clearly to implement a curvature
monitor as in [l6].
Finally we should mention that. in a few instances, the stiff solvers interrupted the integration
due to a Newton convergence test failure, especially so when using extremely fine grids. This could
be due to the fact that. in the experiments reported, the local error and New1on convergence test
was applied to X, and not to .3.X,. Also poor prediction of the velocities may have caused difficulties
for the Newton solver. These aspects need further attention (e.g. in a study along the lines of Pet-
zold and Lotstedt [13]).

REFERENCES
L S. Adjerid and J.E. Flaherty. A Moving Finite Element Method with Error Estimation and
Refinement for One-Dimensional Time Dependent Partial Differential Equations, SIAM J. Numer.
Anal., 23 (1986), pp. 778-796.
2. M. Berzins and R.M. Furzeland, A User's Manual for SPRINT - A Versatile Software Package
for Solving Systems of Algebraic, Ordinary and Partial Differential Equations: Part I - Algebraic and
Ordinary Differential Equations, Report TNER.85.058, Thornton Research Centre, Shell Research
Ltd.. U.K., 1985.
3. M. Berzins and R.M. Furzeland, A User's Manual for SPRINT - A Versatile Software Package
for Solving Systems of Algebraic, Ordinary and Partial Differential Equations: Part 2 - Solving Partial
~ Differential Equations, Report No. 202, Department of Computer Studies, The University of Leeds,
1986.
4. E.A. Dorfi and L. O'C. Drury, Simple Adaptive Grids for 1-D initial Value Problems, J. Comput.
 MOVING-GRID METHOD 175

Phys., 69 (1987), pp. 175-195.

5. H.A. Dwyer and B.R. Sanders, Numerical Modeling of Unsteady Flame Propagation, Report
SAND77-8275, Sandia National Laboratories, Livermore, USA, 1978.
6. R.M. Furzeland, The Construction of Adaptive Space Meshes for the Discretisation of Ordinary and
Partial Differential Equations, Report TNER.85.022, Thornton Research Centre, Shell Research Ltd.,
U.K., 1985.
7. R.M. Furzeland, J.G. Verwer and P.A. Zegeling, A Numerical Study of Three Moving Grid
Methods for One-Dimensional Partial Differential Equations which are based on the Method of Lines,
Report NM-R8806, Centre for Mathematics and Computer Science (CWI), Amsterdam, 1988. (Sub-
mitted to J. Comput. Physics.)
8. A.C. Hindmarsh, ODE Solvers for Use with the Method of Lines, in Advances in Computer
Methods for Partial Differential Equations-IV, R. Vichnevetsky and R.S. Stepleman, ed., IMACS,
1981, pp. 312-316.
9. J.M. Hyman and M.J. Naughton, Static Rezone Methods for Tensor-product grids, in Proc.
SIAM-AMS Conference on Large Scale Computation in Fluid mechanics, SIAM, Philadelphia, PA,
1984.
10. N.K. Madsen, MOLAG: A Method of Lines Adaptive Grid Interface for Nonlinear Partial
Differential Equations, in PDE Software: Modules, Interfaces and Systems, B. Engquist and T.
Smedsaas, ed., North-Holland, 1984.
11. L.R. Petzold, A description of DASSL: A Differential! Algebraic System Solver, in /MACS Trans.
on Scientific Computation, R.S. Stepleman, ed., 1983.
12. L.R. Petzold, Observations on an Adaptive Moving Grid Method for One-Dimensional Systems of
Partial Differential Equations, Applied Numerical Mathematics, 3 (1987), pp. 347-360.
13. L.R. Petzold and P. LOtstedt, Numerical Solution of Nonlinear Differential Equations with Alge-
braic Constraints Il: Practical Implications, SIAM J. Sci. Stat. Comput., 7 (1986), pp. 720-733.
14. R.F. Sincovec and N.K. Madsen, Software for Nonlinear Partial Differential Equations, ACM
Trans. Math. Software, l (1975), pp. 232-260.
15. R.F. Sincovec and N.K. Madsen, Algorithm 494, PDEONE, Solutions of Systems of Partial
Differential Equations, ACM Trans. Math. Software, l (1975), pp. 261-263.
16. J.G. Verwer, J.G. Blom and J.M. Sanz-Serna, An Adaptive Moving Grid Method for One-
Dimensional Systems of Partial Differential Equations, Report NM-R8804, Centre for Mathematics
and Computer Science (CWI), Amsterdam, 1988. (To appear in J. Comput. Phys. in 1989.)