Numer. Math.

41,373-398 (1983) Numerische

Mathematik
9 Springer-Verlag1983

A Semi-Implicit Mid-Point Rule for Stiff Systems

of Ordinary Differential Equations
G. Bader and P. Deuflhard
Universit~it Heidelberg,Institut ftir Angewandte Mathematik, Irn Neuenheimer Feld 293,
D-6900 Heidelberg 1 (Fed. Rep.)

Summary. The paper introduces a new semi-implicit extrapolation method

especially designed for the numerical solution of stiff systems of ordinary
differential equations. The existence of a quadratic asymptotic expansion in
terms of the stepsize is shown. Moreover, the new discretization is analyzed
in the light of well-known stability models. The efficiency of the new in-
tegrator is clearly demonstrated by solving a series of challenging test prob-
lems including real life examples.
Subject Classifications: AMS(MOS): 65L05; CR: 5.17.

O. Introduction

For many years, the numerical solution of stiff ordinary differential equations
(ODEs) has been an active field of research motivated by challenging real life
applications. The presently most effective algorithm for this type of problems
seems to be the BDF method of Gear [10] in the advanced version due to
Hindmarsh [13]. As this method is based on an implicit discretization, it re-
quires the iterative solution of a nonlinear algebraic system per each integration
step. In view of this feature, recent attention has focussed on so-called semi-
implicit methods, which require the solution of just one linear system per each
step (or stage). Examples of this latter kind are the Rosenbrock-Wanner meth-
ods (see e.g. Kaps and Rentrop [15]) or the new extrapolation method to be
introduced in the present paper.
The latter method is based on a new type of discretization, which is called
the semi-implicit mid-point rule, as it has been derived from the well-known
explicit mid-point rule (see Gragg [11, 12], and Bulirsch and Stoer [2]). In
Sect. 1, this semi-implicit mid-point discretization is shown to permit a qua-
dratic asymptotic expansion in terms of the stepsize h. This expansion and cer-
tain additional stability considerations (see Sect. 2) back the expectation that
the new method should be useful for solving stiff problems. Details of the al-
gorithmic realization are described in Sect. 3. Finally, in Sect. 4, numerical
374 G. Bader and P, Deuflhard

comparisons of the new extrapolation code with the codes due to Gear and
Hindmarsh [10, 13] and Kaps and Rentrop [15] are given.

1. Quadratic Asymptotic Expansions

1.I. Preliminary Considerations
Consider the initial value problem for a system of n first-order differential
equations (autonomous for simplicity of the subsequent representation)

Y'=f(Y) (1.1)
y(0)=y0.
Upon applying the explicit mid-point rule as realized in the algorithm due to
Gragg [12], Bulirsch and Stoer [2 3 one obtains:

a) rto: = Yo
r/p =Yo +hf(yo)
(1.2)
b) k = 1..... [: qk+ P=t/k- i +2hf(r/k),
c) S e: = ~1[~le- t +2q~+r/e+ t].

The discretization error for both tl(t; h) and S(t; h) is known to have an asymp-
totic h2-expansion of the form (with t'. =Eh fixed):
N
t/(t; h ) - y ( t ) = ~. [ u j ( t ) + ( - 1)e"vj(t)], hEJ+EN+ ~(t; h) h =N+2
~=1 for he[0, HI (1.3)
where N depends on the order of differentiability o f f The functions uj, vj are
known to satisfy the linear systems (dropping inhomogeneous terms):

u)= fy(y(t)) uj + . . .
(1.4)
v] = -fy(y(t)) Vj + ....

As an illustrative trivial example, let

y ( t ) = e - t + e - ~ ~ 1 7 6 1 7 6-', for t "sufficiently large",

which roughly yields

uj(t)~e -t , vj(t)~e +1~176176j = 1,2,...,N. (1.5)

Because of this behaviour, the functions vj are called the "weakly unstable"
components. The final replacement of re by Se just eliminates v 1 in the asymp-
totic expansion.
A Semi-Implicit Mid-Point Rule for Stiff Systems 375

The above illustrating example clearly demonstrates why the GBS-algo-

rithm does not work for stiff ODEs. In order to treat this class of problems,
one might use the transformation (compare Lawson [16]):
y(t) ='.e A' c(t) (1.6)

in terms of the matrix exponential of some (n, n)-matrix A to be adapted. Thus

system (1.1) is transformed into the system
c ' = e - A t [ f (eat c)-- A eAt c]. (1.1')

Now, discretize this equation by means of the explicit mid-point rule and use
(1.6) for back-substitution. Then, with the notation

f (y):= f (y) - A y,
one obtains the following discretization:

a) F/o:= Yo
Fll: = eAh(y o + hf(yo)),
b) k = l , . . . , • : FIk+l:=eah(eAhTlk_l+2hf(Flk)), (1.2')
C)
Se:=z[e
- 1
rle_l+2Fle+e-AhT]e+l].
A h --

Now, following the lines of the construction of (1.2'), one may easily verify the
existence of the asymptotic expansion
N
F/(t; h ) - y ( t ) = ~ [ff~(t)+(-1)e~j(t)].hZJ+ff~u+~(t; h).h :u+z, (1.3')
j=l

where uj, vj satisfy

~) =f,(y(t)) ~j +...
(1.4')
=fr(y(t)) gj + 2(A - f , ) gj + ....

A natural choice of A seems to be

A: =f,(Yo)' (1.7)
With this choice, the above illustrating example yields
ff~(t) ~ e-', gj(t) ~ e-', (1.5')

i.e. the "weak instability" seems to be removed - at least in the framework of

the asymptotic analysis for h--*0. This observation certainly motivates a more
general interest in discretization schemes similar to (1.2').
As (1.2') only contains J] this discretization may naturally be associated
with the ODE system
y ' - A y =f(y). (1.8)
In fact, the two-step scheme (1.2'.b) can be generated by

y' - Ay--* 1 [e- ah q(t + h) - e ah rl(t- h)]. (1.9)

Z/'/
376 G. B a d e r a n d P. D e u f l h a r d

For actual realization, however, one will replace the matrix exponential by
some Pad~ approximation, say
enh~E(Ah). (1.10)

Thus, in order to preserve the symmetry h~--~-h, (1.9) will be replaced by

y' - A y ~ h [E( - A h) q(t + h) - E(A h) ~l(t- h)], (1.11)

where consistency requires that

E(0)=I, d~hh=o=A. (1.12)

This approach has already been presented in [5].

In what follows, the simple choice

E(Ah)'.=I+hA (1.13)

will be made. This choice will turn out to produce the outstanding candidate
among the discretization schemes of the type (1.11) - see the stability con-
siderations of Sect. 2.1. In fact, even the exact matrix exponential will be ruled
out for stability reasons! Thus, in contrast to any approaches as the one of
Lawson [16], the discretization proposed here should not be regarded as an
"exponential fitting" method.

1.2. Semi-implicit Scheme

With the notation
A: =fr(Yo), f(y):=f(y)-Ay
the preceding considerations lead to the following discretization scheme:

a) r/o" =Yo
r/,: = (I - h A)- 1 [Yo + hf(yo)],
(1.14)
b) k = l ..... d: rlk+l:=(I-hA)-l[(I+hA)rlk_l+2hf(qk)],
c) - 1
S2m:=~(~/2m+ 1 +q2m_ 0 for ( = 2 m .

From the derivation, (1.14.b) is called the semi-implicit mid-point rule. This
two-step scheme is started by the semi-implicit Euler step (1.14.a), which may
be derived in a similar way as the mid-point step (1.14.b). The final step
(1.14.c), which is due to I-1], performs some kind of smoothing that will be
analyzed in Sect. 2.2 below. Note that this final step is not just obtained by
transforming Gragg's final smoothing step (1.2.c) via (1.2'.c) with (1.10).
Theorem 1. Let y denote the unique solution of the IVP (1.1) and r/(t; h)-~le the
solution of the discretization scheme (1.14.a,b) for t = f h fixed. Let for some
inner product ( . , . ) the following relation hold:
(x, A x ) <__~( x , x ) - ~ llxll 2. (1.15)
A Semi-lmplicit Mid-Point Rule for Stiff Systems 377

Assume that fG C 2N+ 2 and that

{If ( u ) - f (w)[t < L [{u-wll. (1.16)

Then there exists some H > 0 subject to

H#<l-e (1.17)

for some e in the range 0 < e < 1 such that, for all hG[0,H], the following asymp-
totic expansion holds:
N

r/(t; h ) - y ( t ) = ~. [ui(t)+(-1)ev~(t)] h2J+EN+:(t; h)h 2N+2 (1.18)

j=:
where u~, vj satisfy
t -- 1
u j - (A + fr) u~ + ~ [aj(t) + bj(t) + cj(t) + d~(t)] (1.19)
v~.= (A - fy) vi + 89[aj(t) + bj(t) - ci(t ) - dj(t)]

with ai, bj, e~, dj defined in the proof (see (1.25) and (1.30)), and EN+ t is uni-
formly bounded for all h~EO, HI.
In particular, with the additional notation

for s=O,
the following estimates hold:

a) for ~<0:
lIEN+ a(t; h)ll <[le2Lt+ ctt q~(2Lt), (1.20)

b) for # > 0 , I ~ H < l - e , 0 < e < l , ~,:=L+/~

e

lIEN+ l(t; h)ll <fl- e 2Lt +ctt ~o(2Lt),

g e

where the definition of ct, fl is given by the proof

Remark 1. F o r any one-step method of order p applied to the well-known
scalar test equation of Dahlquist [3] one will obtain the discretization error

]R(~.h)- eahl <=h p+ 1. %+ l,

where R is some rational function. Let 2 h = c 1 = c o n s t , then

IR(2h)- e ah] = c2 = const,

which, in turn, leads to
%+ : ~ 12Ip+ 1 for h ~ 0 .
378 G. Bader and P. Deuflhard

Now, recall that extrapolation methods are just a special case of one-step
methods (cf. [26]). As a consequence, without further restrictions on y and f
(and their derivatives), the uniform bound on EN+ 1 must contain at least poly-
nomials (of degree (2N+2)) in [IA[I. Recall that with explicit discretizations
e llallt appears.

Remark 2. Recall that (1.16) implies that

IlfyH- IIfy(y)-AII = Ilfy(y)-fy(yo)ll < L. (1.16')

Thus, if f is linear, then L = 0 holds. If f is nonlinear, then there exists some
neighbourhood of Yo within which L is sufficiently small - even if the Lipschitz
constant of f is extremely large! From (1.20), the basic stepsize H is - due to
stability requirements - seen to be roughly restricted by
1
H<Hmax~f~, if ~<0.

If H ~ H . . . . then accuracy requirements are said to determine the basic step-

size, i.e. only higher order derivatives rather than the first order effect "stiffness"
regulate the size of the basic integration step (see also [26]).
Proof (I) The first part of the proof follows the lines of a proof given by Stetter
[22, 23] for symmetric two-step methods. For ease of notation, let

/~."=2h, g(y):=Ay, Xk: =q2k' Z k : = ~ Z k + 1"

With this notation, the scheme (1.14.a,b) is equivalent to the following recur-
sion
a) Xk+l:--=Xk+h[g(Xk+2"~Xk)+f(Zk)],
(1.21)

with

xo:=yo, zo:=xo+~[g(zo)+f(xo)].
This semi-implicit recursion is well-defined, if

I - hA nonsingular,

which is guaranteed by the assumption (1.17).

Next, the additional transformation

w~, = zk - ~ [g(zk) + f(xk)]

leads to the symmetric representation

zk =wk + w~+ 1 /7 [f(x~+ 1)-f(x~)]. (1.22)

2 4
A Semi-Implicit Mid-Point Rule for Stiff Systems 379

Thus, (1.21) is equivalent to the one-step scheme

a) Xk + 1 -- Xk 1
ft =~ [g(Xk+ 1) + g(Xk)] +f(Zk),
(1.21')
Wk+ 1 -- Wk 1
b) =~ EY(~+ ~)+ f(x~)] + g(z~)

with z k expressed by (1.22) and

XO=Wo=y O.

For kT~0, his difference system is seen to be consistent with the differential
system
a) x'-Ax=f(w), x(0)=y o,
(1.23)
b) w'-Aw=f(x), w(0)= y o .
Thus x ( t ) - w ( t ) = y ( t ) follows. Moreover (1.21') is invariant under the common
interchanges
Xk+--')'Xk+ 1~ Wk ~4'Wk+ 17 h~--)" - ~ .

These properties imply the existence of a quadratic asymptotic expansion for

both x(t; h) and w(t; h) - see Stetter [22, 23]. Finally, by virtue of (1.22), z(t; h)
is also seen to have an asymptotic hZ-expansion.
(II) The second part of the proof aims at the verification of the results (1.19)
and (1.20). For this purpose, one may follow the lines of the type of proof
given by Gragg Ell] in his dissertation. On the basis of (I), introduce
N
a) x(t; h)=:y(t)+ ~ ~i(t)h2~ + X(t; h)h 2N+2,
i= 1 (1.24)
N
b) z(t; h) = :y(t) + ~ (,(t) h 2i + Z(t; h) h 2 ~ +2,
i=l

where one formally identifies

x(t2k ; h)=rl(t2k; h), z(t2k+ 1 ; h)=q(tzk +1 ; h).

Insertion of these expansions into (1.21) and collection of equal powers in h
leads to the ODEs (for # > 1).

a) ~ - A ~e =J~(Y) ~e + ae(t) + be(t),

(1.25)
b) ~ - A ~e= fy(Y) r + ce(t) + de(t)
with the initial values

c) G(o)=o,
G(O)= _

Z~
•(2k--
d--k
1)tA~I
~vl|
380 G. Bader and P. Deuflhard

and the definitions

d) ~o(t)- (o(t)- y(t)

c+(t),= - W -~ '" )

,=2 ~ f<~ A =:e=,be(t)h2'+B~ 0 .... ,~.; Z(t; h) hZN+ 2; h) h 2N+2,

N 1 N
~. f<O(y) A ~i =: ~ dc(t ) h2e + D~ +, (~0 ..... ~. ; X(t; h) h 2N+2 ; h) h 2N+2
i=2 9 g=l

where
N
A(:= ~ (k(t) h2k+z(t; h) h 2N+2,
k=l
N
A~:= ~, ~k(t) h2k-Fx(t; h)h 2N+2,
k=l

bl(t)=-dl(t)=-O.

The results (1.25) have essentially already been found by Mirbeth [-17]. In or-
der to verify the boundedness of the remainder terms X, Z, one may con-
veniently introduce the further notation:
1 1
S f<N+I)(Y +6"AOA~N+I d6
( N + 1)! ~=o
----:B~+ 1((o .... , (n ; Z(t; h) h 2N+ 2; h) h 2N+2,
1 1
fm+a)(y+f'A~)A~N+l d6
( N + 1)! ~=o
=..O~+l(~O .... ,~,; X(t; h)h2N+2; h)h 2N+2,
N 1 (~2N+ 1 - 2 i
R(t;h)'.= ~ ~ (2N+l-2i)!r
i=0 6=0
N 1 ~2N+1-21
P(t; h):=,~=o'= ~=_o
~ (2N + 1 - 2 i ) I (12N+2-2i)(tq-r~h) dr~'
,_ 0 DN+I,=O
Note that, from the definitions of at(t ), ce(t), one has
~i,~iEC 2N+ 3-2i,

so that both R and P may be uniformly bounded. With these preparations, the
following system of nonlinear difference equations may be shown to hold:
A Semi-ImplicitMid-Point Rule for StiffSystems 381

a) ( I - h A ) [X(t +h; h)+ R(t +h; - h ) ] - ( I +hA) [ X ( t - h ; h ) + R ( t - h ; h)]

=2h[BN+ i('; Z(t; h) hZN+2; h)+ fy(y) Z(t; h)],
(1.26)
b) (I - h A) [Z(t + h; h) + P(t + h; h)] - (1 + h A) [ Z ( t - h; h) + P ( t - h; h)]
= 2h[ON+ 1('; X (t; h) hZN+2; h)+ fr(y ) X(t; h)].
As starting values, one obtains

c) x(0; h)=0
N+ 1 r(2k) [O~ (1.26)
Z(h; h ) = - e ( h ; - h ) + ( I - h A ) - i ~ ~N+l-k~ v,
k= 0 (2k) !
For the convenience of the subsequent estimates, one may introduce the func-
tions
q~(t; h)'.=(I + h A) R ( t - h ; h ) - ( I - h A) R(t + h; -h) + 2h BN+ l(. ; h),
~(t; h):=(I+hA)P(~-h; h ) - ( I - h A ) P ( t + h ; - h ) + 2 h DN+I('; h).

These functions are odd with respect to h so that

4,(t; 0)= 9'(t; 0)=0.

Moreover, they are Cl-functions with respect to h. In order to derive the es-
timates (1.20), one may adopt a technique similar to the one used by Dahlquist
and Lindberg [-4-]: let I1"LIdenote the norm induced by the inner product ( . , . )
defined in (1.15). For the time being, let /~<0, which readily implies that (for
the subordinate matrix norm) for all h > 0

II(I-hA)- 1 (I + h a ) l I < 1
1 (1.27)
<1.
I](I-hA)-lll <1 - h # =

As a consequence, the following estimates may be shown

ll(I-ha) -1 ~(t; h)[I < e l h e : = m a x ( e l , e 2 ) < 0%

[[(I-hZ) -1 ~(t; h)lq<e2h
HZ(h; h)[[ </~< oo.

Note that, as a consequence of (1.25), the above bounds e,/3 depend on bounds
for polynomials in IIAI[ via the dependence upon ~i, ~i - compare Remark 1
above. The additional implicit dependence upon the terms X, Z will be dis-
cussed at the end of the proof.
As a next step, introduce the upper bounds

[[X(t2k; h)l[ ~E2k

k=0, 1, ..., m.
IlZ(t2k+ 1; h)ll <E2k+ 1
With these preparations, the difference equations (1.26) may be shown to lead
to the following majorizing difference equation
382 G. Bader and P. Deuflhard

Eo:=0, El:= fl
(1.28)
Ek+ 1: = E k - 1 -t- 2hL E k -t- hot, k = l ..... :.

From this, by means of repeated induction, one may come to the upper bound

E k </~(1 + 2hL)k + ct (1 +2hL) k - 1

2L (1.29)

which, by means of standard techniques, yields the uniform upper bound

e 2 Lt _ 1
liE(t; h)N < [3e2Lt + ct ~
2L

This verifies (1.20.a). Next, assume that # > 0 and #H < 1 - ~ for some ee]0, 1[.
In this situation, (1.27) must be replaced by

I1(I- h A ) -I (I +hA)ll < 1 + h #

=l-h#
(1.27')
1 1 1
H(I-hA)-~[]< < <
=l-h# =I-Hp=e"

As a consequence, the above estimates are replaced according to

l-h#' l-h#' l-h#'

These modifications lead to the modified majorizing difference equation

E 0, =0, El" = 1 - h # ' (1.28')

l+h# 2hL hc~
Ek+ 1 -- 1 - h # Ek- 1 + l_---~ Ek + 1 ~ h ~ ~

In turn, (1.28') then leads to the associated estimate

[ 2h(L+#)~ k
/1-+2h(L+#)\k ~1-~ i ~ ] 1
Ek <
=l-h# ~ i---~-~ ) + e 2(/,+#) (1.29')

The associated uniform bound is

Ote 2Lt- 1
HE(t; h)[l =<fl-e2s -
~ 2L

with L , = L + # . This verifies (1.20.b). Upon formally re-inserting the above

bounds into the estimates defining e and //, one obtains an implicit condition
on a maximum stepsize H so that all the above estimates hold uniformly for
A Semi-ImplicitMid-Point Rule for StiffSystems 383

h~[0,H]. Any restrictions on H, however, are merely caused by higher order

derivatives o f f rather than by the first-order effect "stiffness". To see this prop-
erty, let f be linear. Then f=BN+I=DN+I=O and L = 0 . In this case, the
ODEs (1.25.a) and (1.25.b) and the difference equations (1.26.a) and (1.26.b) are
decoupled. As a consequence, for /~< 0 no stability restrictions arise from the
above estimates. In particular, the bounds (1.20) simplify to

liE(t; h)ll<fl+at for # < 0 , L = 0 .

A detailed examination shows that fl, ~ contain polynomials in IIA[I of the

same degree as predicted in Remark 1 above. A similarly detailed examination
for the nonlinear case would be technically rather complicated: it shows that a,
fl contain polynomials in [IA]I as in the linear case and additional bounds for
higher derivatives of f and y.
Finally, in order to verify (1.18) and (1.19), just introduce the new variables

uj(t).'= 89 + ~j(t)), vj(t) := 89 -- ~j(t)). (1.30)

This completes the proof. 9

Theorem2. Let $2,,=S(t2~ ;h) be defined by the above final step (1.14.c). Then,
under the assumptions of Theorem 1, S has an asymptotic h2-expansion of the
form
N
S(t; h)=y(t)+ ~ ~(t) h21+/~N+ l(t; h) h 2N+2, (1.31)
i=1

with fire+1 uniformly bounded. In comparison with (1.18) and with the notation
~j(t):=Uj(t)--Vj(t),
the following results hold."
(o(t)=--y(t)
~(t)
a) (~(t)= k=o (2k)! { = 1 .... , N + I , (1.32)

b) (I + h A)/~N + 1(0; h) = (I - h A)- ' (N +1 (0).

Moreover, if A = 0 (explicit mid-point rule), then

a) (e(0) = 0 f = 1, ..., X + 1, (1.33)
b) /~u+ 1(0; h)=0.

Proof. Notation as in the proof of Theorem 1. By definition one obtains

2S(t; h)=z(t +h; h) +z(t-h; h)

N
= ~ [(i(t+h)+(i(t-h)] h21+[Z(t+h; h)+Z(t-h; h)] h 2u+2
i=0
384 G. Bader and P. Deuflhard

The standard Taylor expansion of the (i then leads to

S(t; h)= ~ h2` ~ ~(e2~(t)~h zu+2/~u+l(t; h)

e=o k=O (2k)!
with

/~s+l(t; h),=89 +h; h)+ P(t +h;-h)+ Z(t-h; h)+ P(t-h; h)]. (1.34)
Since Z and P are uniformly bounded, the same is true for /~. Moreover, the
representation (1.32.a) is verified. Next, in order to verify (1.32.b), apply (1.26.b)
for t = 0 (recalling that t/(0; h)=yo) , which reads

(I+hA) [ Z ( - h ; h)+P(-h; h)] = ( I - h A ) [Z(h; h)+P(h; -h)].

Upon inserting this identity into (1.34) for t = 0 and using (1.26.c) with (1.32.a),
the result (1.32.b) arises. Moreover, if A = 0 , then (1.25.c) immediately yields
(1.33a), which, by means of (1.32.b) then confirms (1.33.b). 9
Remark. Note that the above theorem including the corollary could be modi-
fied in such a way that the existence of a quadratic asymptotic expansion is
also verified for non-autonomous right-hand side f(t, y). If Ilfttl is "large", how-
ever, the addition of the trivial ODE t'= 1 is recommended.
Finally, two features of the above theory seem to be worth mentioning:
(1) The discretization (1.14) removes the "weak instability" just as the dis-
cretization with the exact matrix exponential - compare (1.4') and (1.19).
(2) The theory also covers the case when A is just some "essential part" of
the matrix fy(Yo) that is responsible for the stiffness of the ODE system. Recall
that for some real life applications a natural splitting into stiff and non-stiff
components is possible - compare Hofer [14]. In view of this property, auto-
matic partitioning is presently under investigation.

2. Stability Considerations

In this section, the discretization (1.14) is studied by means of certain simple

intuitive model problems such as the inhomogeneous model given by Prothero
and Robinson [18] or the well-known homogeneous A-stability model suggest-
ed by Dahlquist [3].

2.1. Inhomogeneous Stability Model

In [18], the following type of model problem has been treated

y' = 2(y - g(t)) + g'(t), t _>_0. (2.1)

Under the assumption
Re(2) < 0
A Semi-ImplicitMid-Point Rule for StiffSystems 385

the above ODE has the steady state solution

y(t)-g(t).

Therefore, a desirable property of any discretization would be that the steady

state solution of the ODE is also a solution of the associated difference equa-
tion - at least, up to a polynomial of a certain degree. This idea will now be
applied to the following type of mid-point rule

E( - A h) rl(t + h) - E ( A h) q ( t - h) = 2 h f (t, rl(t)), (2.2)

which has been derived in (1.11). Note that (2.2) is just the non-autonomous
version of (1.11). One may identify the quantities

A:=2, f(t):=g'(t)-2g(t)

and use the standard notation

z:=2h.

Moreover, already the restriction to the linear special case

g(t):=g o + a t (2.3)

will lead to a strict determination of only one Pad6 approximation. For the
above specification, (2.2) is equivalent to

(go + a t) I-E( - z) - E(z) + 2z] + a h [E( - z) + E ( z ) - 2] = 0.

Hence, any steady state solution (2.3) is a solution of the above difference
equation, if the following two conditions hold:

E(z) - E( - z) = 2z
(2.4)
E(z) + E ( - z) = 2.

These equations readily yield

E(z) := 1 + z, (2.5)

the specification that has been used throughout Sect. 1.2 above - see (1.13).
One may also observe that the exact exponential ez is ruled out by the equa-
tions (2.4)! In view of the above outstanding property for the simple Pad6-
approximation (2.5), the concentration of the present study to this special case
seems to be justified.

2.2. Homogeneous Stability M o d e l

As the discretization (1.14) is linear in f, the stability analysis of linear homo-
geneous systems of the kind
y'= A y (2.6)
386 G. Bader and P. Deuflhard

reduces to the analysis of the well-known complex scalar equation (I-3])

y' : 2y, y(0) = Yo (2.7)

2OE, Re(A)<0.

For the above special ODE and with the notation

z:=fyh=2h, Re(z)<0,

the discretization (1.14) leads to (for m > 1):

1 ( l + z ] "`-I
q2m-1-- 1--Z \ l - - z ! YO
=(I+z~ m
~]2m \~ZI Yo (2.8)

1 (1 + z~ "-1
S2m = ( 1 - z ) 2 \ l - - z f Yo"

One may observe that for Re(z)<0

I~/el< lYol for all f = 1, 2, (2.9)

IS2,,I < lYol. . . . '
which means that the discretization is A-stable. Moreover, for Re(z)~ - ~ , the
following asymptotic behaviour is seen:

q 2 , , ~ ( - 1)"yo
(-1)"
q2,,- 1~ Yo~ 0 (2.10)
Z
g2 m__a.(-- 1)m-1
Z2 Y0~0"

This means that, for f odd, the r/e are strongly absolutely stable. In particular,
the final step due to [1] performs some additional smoothing.
The above stability considerations have been made under the assumption
that
A =fr(Y) = constant.
In the more general nonlinear case one has

fr(Y) ~ constant.

In this situation, the reduction of an O D E system to a single scalar equation is

certainly possible, if A and fy(y) can be commonly diagonalized. Under this
assumption, one may reasonably introduce

Zo:=Ah, Re(z0)<0.
A Semi-Implicit Mid-Point Rule for Stiff Systems 387

U p o n treating z o and z as independent variables, one might expect to include

the effect of rounding and discretization errors as well as variations of the
Jacobian into the analysis - thus modelling in some respect nonlinear ODEs.
For
Z 0 :~= Z,

the analysis of the total discretization (including extrapolation) has recently

been done by Hairer et al. [26]. A lot of insight, however, can already be
gained from the analysis of the semi-implicit mid-point step as given in [-1].
The associated characteristic equation is

(1 -Zo)~2-2(Z-Zo)Z-(l+zo)=O. (2.11)

Let zj(Z, Zo), j= 1, 2 denote the characteristic roots of (2.11). Then the associated
stability region is defined by
a(zo) := {z~ffg[lzi(z, Zo)[ < 1, j = 1, 2}.
For the b o u n d a r y 0G one may introduce the p a r a m e t e r representation
"c(~)=e ie.
This representation leads to

~G(zo)={zetl?.lZ-Zo= - Z o cos q~+ i sin~o, q ~ [ 0 , 27z]}, (2.12)

which means that the stability regions are elliptic. In Fig. 1, such regions are
shown for various z o. F o r z o approaching the imaginary axis, the elliptic do-
mains shrink d o w n to the degenerate case of just the interval [0, 2Zo], if z o lies
on the imaginary axis. This behaviour nicely reflects the fact that the concept
of absolute stability actually ends on the imaginary axis.

[m (z~

Fig. 1. Stability regions G(zo)for zo approaching the imaginary axis

388 G. Bader and P. Deuflhard

3. Algorithmic Realization

The actual implementation of an extrapolation method based on the above

discretization (1.14) requires some further consideration. The details to be de-
scribed in the present section are realized in the standard routine METAN 1.

3.1. Discretization
The realization of the (n, n)-matrix

A.. =fy(Yo), (3.1)

which is required only once per basic integration step, is done either by evalua-
tion of the analytic Jacobian (if such a representation is easily available - see
e.g. [7]) or by finite difference approximation (to be counted for, at least, n
additional evaluations of f). The replacement of the "exact" Jacobian by an
approximation is theoretically backed by the analysis of Sects. 1.2 and 2.2.
Compact scheme. For an efficient implementation of (1.14) one introduces the
auxiliary variables
Ak: = q k + 1 --?]k
and uses the relation
( 1 - h A ) - 1 (I + h A ) = 2 ( I - h A ) - 1 - I .

Thus one obtains the following compact scheme:

a) qo' =Yo
A o' = (I - h A ) - 1 hf(yo) '

b) k = l ..... # - 1 : qk:=tlk_lq-Ak_l
Ak' = AR- 1 + 2 ( 1 - h A ) - 1 [ h f (qk)_ Ak - 1], (3.2)
C) ~]g,:~-/~_ 1 -t- AE- 1
zle,= (I - h A ) - a [-hf 07e)-Ar 1]

This realization avoids the time-consuming evaluation of f ( y ) = f ( y ) - A y

(which might anyway have caused extinction of leading digits). Moreover, since

A k = O(h),

a rounding error stabilization for h ~ 0 is thus obtained.

Solution of the linear systems. The above matrix ( I - h A ) is decomposed in
(3.2.a) using Gaussian elimination with column pivoting and possible row per-
mutations. As a consequence, the linear systems in (3.2.b, c)just require for-
ward and backward substitutions. (For large sparse systems the linear equation
solvers should be modified, of course.) A subtle question is the construction of
A Semi-Implicit Mid-Point Rule for Stiff Systems 389

a monitor for the detection of a singular or "nearly singular" matrix. In order

to have a more satisfactory theoretical basis for any such monitor, some scal-
ing offy is realized in M E T A N 1, which results in a scaling of the arising linear
system, since
D - l(I - hfy) D = I - h ( D - ~fy D) (3.3)

for any nonsingular (diagonal) weighting matrix. As a consequence, the scaled

variables
D-1A

are actually computed in (3.2). On this scaled basis, a matrix is regarded as

being "nearly singular" either, if an exactly zero pivot element occurs in the
Gaussian decomposition or, if the differences A are too large in the sense that

[]D-1A{] >~r, ~>>1, (3.4)

where D is the above weighting matrix. Note that this latter test is formally
invariant under componentwise re-gauging of the variable y. If one of the two
tests is activated, then the integration step is repeated with a reduced stepsize.
Remark. The authors do not favour any monitoring devices that involve any
condition number of the matrix ( I - h A ) . In fact, typical linear systems arising
from extremely stiff ODE systems usually exhibit extremely ill-conditioned
matrices occurring in nevertheless well-conditioned linear systems. As an exam-
ple, a linear system of the type

may occur after the decomposition. Here cond(A) ,,~-1, while the solution x 1
= x 2 = l is well-defined for all e~0. The peculiarity of such systems is that in
the singular limiting case (here e ~ 0 ) the right-hand sides remain in the column
space of the matrix. For a treatment of the rounding error propagation in such
cases see Wilkinson [25].

3.2. Extrapolation
The existence of the quadratic asymptotic expansion (1.31) permits one to ap-
ply Richardson extrapolation. For this purpose, the problem is repeatedly dis-
cretized over the same basic step, say [0, H], with successively decreasing step-
sizes H
hi , = - - , i = 1, 2..... (3.6)
nl

The integers n i are usually arranged in a sequence

~={nl,nz . . . . },

which characterizes the extrapolation method under consideration.

390 G. Bader and P. Deuflhard

In view of (1.18), all integers E=n~ in the discretization (1.14)

S t e p s i z e sequence.
should be chosen either even or odd. In order to obtain actual smoothing by
the final step (1.14.c), the ni should be chosen to be even. Moreover, recall from
(2.10) that the approximations S,i for e z behave like
,, __.(_ 1)m,- 1
g2 Z2 for n i = 2 m i

for R e ( z ) ~ - oo. In this asymptotic case, one may reasonably require that the
sign of the above approximation agrees with the sign of e z along the negative
real axis, i.e. that
S2mi>0,
which readily leads to the representation

n i := 2(2j/+ 1), j~N. (3.7)

In addition, one may recall the Toeplitz condition for extrapolation (compare
Gragg [12]) in the form
5
ni-<a<l, a:=~. (3.8)
Hi+ 1

The special value of a has been selected on an empirical basis. These con-
siderations lead to the sequence

9.= {2, 6, 10, 14, 22, 34, 50.... }

which has been implemented in METAN 1.

Type o f e x t r a p o l a t i o n . For the classical GBS algorithm, rational extrapolation
has been recommended [2]. In recent investigations [6], however, p o l y n o m i a l
extrapolation has appeared to be preferable in both the GBS integrator and
the stiff integrator presented herein. Therefore, rational extrapolation is only
left as a possible internal option in METAN 1.
In [61, a new order and stepsize control for extrapolation
S t e p s i z e control.
methods has been suggested, which is realized in the present version of ME-
TAN 1 - for details see that paper. The only question to be discussed here is
the question of which order should be used to describe the dependence of the
discretization error upon the basic stepsize H. For the subsequent analysis, let
e~ denote some norm of the discretization error associated with column k of
the extrapolation tableau. Then the following approximation is usually applied
in the stepsize control (k = I ..... kmax):

ek(H)-- Yk IIEk + I ( H ; h)It H2 k (3.9)

with/~ as defined in Theorem 2.
If A - - 0 (explicit mid-point rule), then

/~k+ 1(0; h)=0 (cf. (1.33.b)),

A Semi-Implicit Mid-Point Rule for Stiff Systems 391

which implies that

/~k+ I(H; h) = (9(H) (3.10.a)
and, as a consequence
ek(H) = (9(H2k+ 1). (3.10.b)

The analysis for A4:0 is less simple, since then (1.32.b) replaces (1.33.b). As
(I + hA) may well be singular, a more detailed analysis is necessary. On the basis
and with the notation of (1.34) together with (1.26), one may derive the follow-
ing difference equation (dropping the indices (k+ 1)):

(I-hA)(E(t+h; h)-(I +hA) ff~(t-h; h) (3.11)

= h [D(t + h; h) +fy(y(t + h)) X(t + h; h) + D(t - h; h) +fy(y(t - h)) X ( t - h; h)].
The specification of (3.11) to the argument t=h then leads to

/~(2h; h ) = ( I - hA) -2 CN+ 1(0) + h ( I - h A ) -1 [D(2h; h)+fy(y(2h)) X(2h; h)], (3.12)

where (1.32.b) and (1.26.c) have been used. Obviously, the second right-hand
term is (9(h), whereas the first right-hand term is just (9(1). If the second term
dominates, then
1l/~(2h; h)ll = (9(h) (3.13)
for "sufficiently large" h. This starting point together with the recursion (3.11)
then implies (3.10.a) for sufficiently large h. On the other hand, if the first term
dominates, then
[b/~(2h; h)ll =(9(1). (3.13')
This starting point together with (3.11) then implies

a) tl/~(H; h)t I =(9(1), (3.10')

b) ek(H) = (9(H2 k).
Hence, a comparison of the two right-hand terms in (3.12) seems to be interest-
ing. For/~ N 0, a straightforward estimation (in terms of the quantities used in
the proofs of Sect. 1.2) yields

II(I-hA) -a D(2h; h)ll <=4,

I[(I-hA)-lfyX(2h; h)ll < L [IX(2h; h)ll,
tlX(2h; h)ll <h(~+2L/?).
These estimates may be combined to give an estimate for the second right-
hand term in (3.12):
IIh(I-ha) -1 [D(2h; h)+ fy(y(2h)) X(2h; h)]ll <h[~ +hL(c~ + ZLfl)].
This result may help to interpret the situation described by (3.12). I f f is linear,
then D = f = 0 by definition, which implies that ~ = L = 0. Hence, for linear f the
second term vanishes so that then (3.10') applies. However, for highly nonlinear
392 G. Bader and P. Deuflhard

f, the second right-hand term in (3.12) may be large compared with the first
term, which represents the "smoothing" performed by the final step. This
smoothing will be the more significant, the stiffer the problem is. Hence, for
nonlinear and extremely stiff ODE systems, the representation (3.10) based on
(3.13) will be appropriate. This interpretation nicely agrees with numerical ex-
periments that have been made both under the assumption (3.10) and under
the assumption (3.10'). Roughly speaking, the even orders as in (3.10') have
appeared to be slightly preferable in linear problems, whereas the odd orders
as in (3.10) have appeared to be significantly preferable in more difficult non-
linear problems. A further analysis shows that assuming (3.10) for the stepsize
control in a situation where (3.10') would be more appropriate, just leads to a
more cautious stepsize strategy - thus improving the robustness of the algo-
rithm. Hence, the odd order assumption (3.10) has been implemented in the
stepsize control of METAN 1.

4. Numerical Comparisons

In the present section, the F O R T R A N subroutine METAN1 (based on the

semi-implicit mid-point rule) is compared with the implicit multistep code
G E A R (due to Gear [10] and Hindmarsh [13]) and the Rosenbrock-Wanner
code G R K F 4 (due to Kaps and Rentrop [15] in the form published by Stoer
and Bulirsch [24]). All numerical experiments have been run on the
IBM 370/168 of the University of Heidelberg in F O R T R A N double precision
(with about 16 decimal digits). The numerical examples to be presented are
documented in terms of the quantities
TOL - prescribed relative precision
TIME - total computing time on the IBM 370/168
NFEV - number of evaluations of the right-hand side f of the ODE sys-
tems
NDEC - number of Gaussian decompositions.
In the standard form of the above routines, numerical difference approxi-
mations of the Jacobian are used, if not stated differently.

4.1. Sample set S T I F F D E T E S T (due to Enright et al. [8])

In contrast to [8], but in agreement with [15], all 25 test problems of this
standard set were started with a common initial stepsize H = 1.D-3. The usual
problem class summary is given in Table 1 for the new extrapolation code
METAN 1. Note that the linear test problems (classes A, B) clearly dominate
the computing time. In order to permit a not too biased comparison with
GEAR, Table 2 gives the comparative results both including and excluding test
problem B 5 (see also [15]).
A Semi-lmplicit Mid-Point Rule for Stiff Systems 393

Table 1. Problem class summary for METAN 1

Class TOL NFEV NDEC TIME Is]

A 1.D-2 485 62 0.16

1.D-4 1014 117 0.38
1.D-6 1934 174 0.69
B 1.D-2 880 117 "0.23
1.D-4 2450 259 0.70
1.D-6 4844 412 1.33
C 1.D-2 572 88 0.13
1.D-4 1212 156 0.28
I.D-6 2511 251 0.59
D 1.D-2 489 79 0.09
1.D-4 998 128 0.18
1.D-6 2064 223 0.38
E 1.D-2 846 131 0.18
1.D-4 1501 193 0.32
1.D-6 2861 323 0.65

Table 2. Comparative results over STIFF DETEST including/exclud-

ing test problem B 5

Code TOL NFEV NDEC TIME

GEAR 1.D-2 6486/3357 440/321 2.58/0.86

1.D-4 9099/5783 556/426 3.89/2.02
1.D-6 9729/8491 638/583 3.67/3.10
GRKF4 1.D-2 5781/5313 814/762 1.01/0.92
1.D-4 12564/11277 1871/1730 2.40/2.17
1.D-6 43275/39360 6481/6046 9.98/9.24
METAN 1 I.D-2 3272/3062 477/450 0.79/0.73
1.D-4 7175/6337 853/771 1.86/1.59
1.D-6 14214/12499 1383/1247 3.64/3.14

Comments: GEAR fails in D3 (no final result) and B5 (wrong final

result) with TOL = 1.D-2

From Table 2, the following observations can be made:

(1) Comparison G E A R - METAN 1:
Both codes are essentially equivalent in TIME, if the example B 5 is excluded.
The bad performance of G E A R in B5 (and also in class B: linear problems
with eigenvalues of the Jacobian approaching the imaginary axis) is theoreti-
cally well understood from the "stiff stability" property of the BDF-method
(see Gear ['10]). Moreover, for T O L = 1.D-6, the multistep method requires
significantly less function evaluations, but has a considerably larger amount of
overhead time - a situation, which is similar to the one known from the non-
stiff case. Finally, the code METAN 1 seems to have a slightly more robust
stepsize control.
394 G. Bader and P. Deuflhard

(2) Comparison G R K F 4 - METAN 1

Both of the semi-implicit methods seem to have no trouble in class B. ME-
TAN 1 requires significantly less function evaluations and Gaussian decom-
positions. For T O L = I . D - 2 and T O L = I . D - 4 , however, this effect is nearly
compensated by the extremely low overhead of the G R K F 4 code. For
T O L < 1.D-4, the authors of G R K F 4 anyway do not recommend their code
(see [15]).
Summarizing, the new code METAN 1 compares favourably to the other
two codes that have been selected for the comparison. On the other hand,
serious doubts about the usefulness of the above test set have recently been
formulated by Shampine [20]. Therefore, one should not exaggerate a favour-
able performance of any stiff integrator in a comparison over this particular
test set. Rather, more challenging test problems will be included here.

4.2. Kidney Model Problem (due to Babugka, see Scott and Watts [19])
Originally, this ODE system arises in the context of a two-point boundary
value problem, where
Y'I=a~(Y3--Yl)

Y'z = - a(Y3 - Yl)

Y's = 1 [b_c(y 3 _ys)_ay3(y 3 -YO] (4.1)
Y4
YX=a(Y3 -Yt)
C
Ys = -~(Ys-Y3)
with
a=100, b=0.9, c=1000, d=10,
yt(O) =y2(O)=y3(O)= 1, y4(O)= - 10
For the integration over [0, 1], the following choices of the missing initial val-
ue have been made (see [19]):

a) Ys(0) =0.990268835,
b) ys(0) =0.99 , (4.2)
c) ys(0)=0.9
A peculiarity of the problem is that it is non-stiff for the first value, but rather
stiff for the second and the third value. Moreover, the solution of the initial
value problem is extremely sensitive to small changes of the initial values ys(0)
- see Fig. 2. Comparative results for both analytic Jacobian and numerical
difference approximation of the Jacobian are listed in Table 3. Due to the
sensitivity of the problem, the final accuracies obtained by the three codes were
not comparable for TOL > 1.D-6.
A Semi-Implicit Mid-Point Rule for Stiff Systems 395

y, ltl_-- /
/
2- /"
104. Ys(O)=0.9/.

3- /
2- /
103" /"
, /
3-
2- f"
102.=_ /" Ys(O)=O99 / j
-: / I
~- / /
/
2- / I
lO1-~ / I
,7 / //
3-1 / ' / y~(0)=0990268635
2-] .J lJ
100. ~
0 0.2 0 Z, 0.6 0.8 1.0 t
Fig. 2. Graph yl(t) of the IVP (4.1) for 3 different initial values ys(0)

Table 3. Comparison of computing times in the kidney

model problem (TOL= I.D-6) with analytical/numerical
Jacobian for three initial values ys(0)
(4.2.a) (4.2.b) (4.2.c)

GEAR 0.05/0.06 0.09/0.10 0.18/0.18

GRKF4 0.06/0.08 0.17/0.40 0.23/7.52
METAN 1 0.13/0.13 0.35/0.36 0.59/0.71

In this challenging example, the code G E A R is clearly best. For analytical

Jacobian, the code G R K F 4 is nearly comparable with GEAR, whereas the
code M E T A N 1 exhibits a rather poor performance. For numerical difference
approximation of the Jacobian G E A R and M E T A N 1 show essentially the pre-
vious behaviour. The performance of G R K F 4, however, drops out drastically
- thus possibly indicating that the underlying discretization might be rather
sensitive to the quality of the Jacobian approximation. The bad performance of
the code M E T A N 1 may be understood on the basis of the theory of Sect. 1.2:
in the present example, the Jacobian varies rather fast within the integration
interval [0, 1] - so L in (1.21) is rather large, which, in turn, leads to a stability
restriction for the basic stepsize H. Such a stability restriction should lead to
an order restriction (see [6]), which is actually observed in the course of the
integration.
Remark. A similar behaviour occurs when new variables for Yx-Y3 and Y3-Y5
are introduced9
396 G. Bader and P. Deuflhard

4.3. Glycolytic Pathway Problem (due to Garfinkel and Hess [9])

In [7], a software package called L A R K I N has been presented that is es-
pecially adapted to the numerical solution of large ODE systems arising from
chemical reaction kinetics. For comparison purposes, the above standard in-
tegrators GEAR, G R K F 4 , and M E T A N 1 have been modified to sparse ver-
sions GEARS, G R K F 4 S , and METAS1, which commonly use the analytical
Jacobian, the sparse solver due to Sherman [21], and an appropriate scaling
device. For details see [7].
The present example with n = 6 5 ODEs arises from biochemistry. In order
to give an idea about the complexity of the problem, the sparse pattern of the
(65,65)-Jacobian is presented in Fig. 3. The results obtained by solving the
problem for three different tolerances are arranged in Table 4+ As it turns out,

| OLU
~1 H~S
9 O*~ " .
4 L?F
5 H|P
MKU .":::.
?
I
AOP
GL~ e , ~ 1 4 9 1 4 9 ,~149 9
9 .. : 9
lit!
10 IS~ o 9
It ISG
12 FRF
1) PPIt
14 pFr 9 I 9149149
15 rPP 9 1 4 9 1 4 9
1~ PFI
IT kl, O 9 ... 9 9
11 AZ,F
19 ALA
;10 CA[* "::,.,',,.,
31 DIIA
~l ?In
l] TID
24 ? J ~ i i t l i ii
315 tllt~
2~ /rOD
2? ItOll D
28 DP|
2'I Fie 9 e~ 9 a o 9
30 ROll
31 OG~
l~ Dim D 9 i
l) I~i Q 6 9 9149
34
35 I
36
17 I~P

151 t~n
40 |gt(* 9 9 19
41 ILIIP
4) I'SP
41 Ill!
44 l~rlc O I
4s I'I~0
441 Pyp
4T PT?
48
49
SO
P~it
LDO
LOL
"!!!..."
1t N
54 Oll
SS DII
S6 %.|
S? OZY
SO XS! e.
S9 asp I ,.
60 ~rp
It Ol11 9 o9
il COle
45 ImtJl m
14 PPP
45 A4P

Fig. 3. Sparse pattern of the Jacobian in the glycolytic pathway problem

A Semi-Implicit Mid-Point Rule for Stiff Systems 397

Table 4. Computing time required to solve the

glycolytic pathway problem
1.D-2 1.D-4 1.D-6

GEARS 2.87 6.42 11.98

GRKF 4S 9.63 28.47 > 90.00
METAS 1 2.36 4.94 10.83

M E T A S 1 is slightly faster t h a n G E A R S in this e x a m p l e - with a s i m i l a r speed-

u p factor as in the s a m p l e set S T I F F D E T E S T . Surprisingly, however, the
code G R K F 4 S d r o p s o u t of the c o m p a r i s o n . T h i s b e h a v i o u r has a p p e a r e d to
b e typical for a wide class of large scale c h e m i c a l kinetics p r o b l e m s t h a t h a v e
b e e n solved m e a n w h i l e - for a selection of further e x a m p l e s see [7].

Conclusion

I n this paper, a n e w efficient e x t r a p o l a t i o n m e t h o d b a s e d o n a s e m i - i m p l i c i t

d i s c r e t i z a t i o n has b e e n presented, T h e existence of a n a s y m p t o t i c h Z - e x p a n s i o n
is theoretically p r o v e d even for extremely stiff O D E systems (with large L i p -
schitz c o n s t a n t of f ) . T h e n u m e r i c a l results s h o w t h a t ( a p a r t f r o m the perfor-
m a n c e in the k i d n e y m o d e l p r o b l e m of Sect. 3.2) the n e w stiff i n t e g r a t o r c o m -
pares f a v o u r a b l y to existing i n t e g r a t o r s - even in large scale stiff O D E systems.

Acknowledgements. The authors wish to thank G. Dahlquist and H.J. Stetter for extremely
helpful comments on a former draft of this paper. Moreover, they are indebted to their colleague
E. Hairer for his careful reading of the manuscript.

References

1. Bader, G.: Numerische Behandlung steifer Differentialgleichungen mit einer modifizierten Mit-
telpunktsregel. Technische Universit~it Mtinchen, Institut ftir Mathematik: Diploma thesis, 1977
2. Bulirsch, R., Stoer, J.: Numerical Treatment of Ordinary Differential Equations by Extrapola-
tion Methods. Numer. Math. 8, 1-t3 (t966)
3. Dahlquist, G.: A Special Stability Problem for Linear Multistep Methods. BIT 3, 27-43 (1963)
4, Dahlquist, G., Lindberg, B.: On Some Implicit One-Step Methods for Stiff Differential Equa-
tions. Royal Institute of Technology, Stockholm: Tech. Rep. TRITA-NA-7302, 1973
5. Deuflhard, P.: A Study of Discretization Schemes due to Hersch with Application to Stiff
Differential Systems. Lecture given at the Oberwolfach Meeting on "Numerical Treatment of
Differential Equations", July 4-July 10, 1976 (Unpublished)
6. Deuflhard, P.: Order and Stepsize Control in Extrapolation Methods. Numer. Math. 41, 399-
422 (1983)
7. Deuflhard, P., Bader, G., Nowak, U.: LARKIN - A Software Package for the Numerical
Simulation of LARge Systems arising in Chemical Reaction KINetics. Springer Series
chem. Phys. 18, 38-55 (1981)
8. Enright, W.H., Hull, T.E., Lindberg, B.: Comparing Numerical Methods for Stiff Systems of
ODEs. BIT 15, 10-48 (1975)
9. Garfinkel, D., Hess, B.: Metabolic Control Mechanisms VII. A Detailed Computer Model of
the Glycolytic Pathway in Ascites Cells. J. Bio. Chem. 239, 971-983 (1964)
398 G. Bader and P. Deuflhard

10. Gear, C.W.: Numerical Initial Value Problems in Ordinary Differential Equations. New York:
Prentice Hall, 1971
11. Gragg, W.B.: Repeated Extrapolation to the Limit in the Numerical Solution of Ordinary
Differential Equations. University of California, Los Angeles: Thesis, 1963
12. Gragg, W.B.: On Extrapolation Algorithms for Ordinary Initial Value Problems. SIAM J. Nu-
mer. Anal. 2, 384-404 (1965)
13. Hindmarsh, A.C.: GEAR-Ordinary Differential Equation System Solver. Lawrence Livermore
Laboratory: Tech. Rep. UCID-30001, Rev. 3 (Dec. 1974)
14. Hofer, E.: A Partially Implicit Method for Large Stiff Systems of ODEs with only Few Equa-
tions Introducing Small Time-Constants. SIAM J. Numer. Anal. 13, 645-663 (1976)
15. Kaps, P., Rentrop, P.: Generalized Runge-Kutta Methods of Order Four with Step Size Con-
trol for Stiff Ordinary Differential Equations. Numer. Math. 33, 55-68 (1979)
16. Lawson, J.D.: Generalized Runge-Kutta Processes for Stable Systems with Large Lipschitz
Constants. SIAM J. Numer. Anal. 4, 372-380 (1967)
17. Mirbeth, M.: Eine modifizierte Mittelpunktsregel zur numerischen Integration steifer Systeme
von Differentialgleichungen. Technische Universit~it Miinchen, Institut f'tir Mathematik: Di-
ploma thesis, 1975
18. Prothero, A., Robinson, A.: On the Stability and Accuracy of One-Step Methods for Solving
Stiff Systems of Ordinary Differential Equations. Math. Comput. 28, 145-162 (1974)
19. Scott, M.R., Watts, H.A.: A Systematized Collection of Codes for Solwng Two-Point Bound-
ary-Value Problems. SANDIA Lab., Albuquerque: Tech. Rep. SAND 75-0539, 1975
20. Shampine, L.F.: Evaluation of a Test Set for Stiff ODE Solvers. ACM TOMS 7, 409-420 (1981)
21. Sherman, A.H.: NSPIV-A FORTRAN Subroutine for Gaussian Elimination with Partial
Pivoting. University of Texas at Austin: Tech. Rep. TR-65, CNA-118, 1977
22. Stetter, H.J.: Symmetric Two-Step Algorithms for Ordinary Differential Equations. Comput. 5,
267-280 (1970)
23. Stetter, H.J.: Analysis of Discretization Methods for Ordinary Differential Equations. Springer
Tracts in Natural Philosophy. Berlin, Heidelberg, New York: Springer, 1973
24. Stoer, J., Bulirsch, R.: Einfiihrung in die Numerische Mathematik II. Berlin, Heidelberg, New
York: Springer, 1973
25. Wilkinson, J.H.: The Algebraic Eigenvalue Problem. Clarendon, Oxford, 1965
26. Hairer, E., Bader, G., Lubich, C.: On the Stability of Semi-Implicit Methods for Ordinary
Differential Equations. BIT 22, 211-232 (1982)

Received March 30, 1981/January 18, 1983