BIT 23 (1983).

1I8-131

NUMERICAL SMALL PARAMETER METHOD

FOR STIFF ODE:S
HAN TIAN-MIN
Computing Center. Academia Sinica, P.O. Box 2719, Peking, People's Republic of China

Abstract.
A numerical method for stiff ODE:s is proposed in this paper. The Jacobian matrix is not needed.
The results of 30 examples are presented.

1. Introduction.

At present, the effective numerical methods for stiff ODE:s are implicit and
usually Newton-Raphson iteration is used to solve the non-linear equations
which result. It is impossible to avoid the matrix inversions and matrix-vector
products required by the method. The number of operations is O(n3) and O(n2)
respectively, where n is the dimension of the equations.
For most implicit methods a matrix inversion is required only when the step"
size changes or the iteration is not converging, but at least one matrix-vector
product is required at each step. This difficulty would be present even if an
LV decomposition and back-substitutions were used. If n is large, for instance
n > 30, the integration becomes very expensive for all these methods. Using
the technique of sparse matrices is one way to overcome this difficulty, but
it will complicate the programme greatly. This paper is a continuation of the
author's work [4]. We propose a numerical method for stiff equations, without
use of the Jacobian matrix. A great deal of numerical experiment have been
done. The numerical results show that our method is efficient, especially when
n is very large.

2. Derivation of the method.

For the initial value problem

(1) y' = f(t, y), y(O) = Yo

we have, using the backward Euler method

(2) hY~+l = Yn+l-Yn'

Received February I, 1982. Revised May 11, 1982.

 NUMERICAL SMALL PARAMETER METHOD 000 119

From the differential equation we get

(3)

where e > 0 is a small parameter.

Adding (2) and (3) we get

(4)

Setting xn+ 1 = hy~+ l' P = h/(h+ e) (p < 1) we can re-write (4) as

(5)

Equation (2) can be written

(6) Yn+1 =yn+xn+1 °

Introducing the parameter IX, 0 < IX < 1, and setting Xn = hy~, we can re-
write (6) as

(7)

and combining (5) and (7), we get the integration scheme:

Xn+ 1 = p[ef(tn+ l' Yn+ 1) + Yn+ 1- Yn]

(8)
{ Yn+1 = Yn+(1-IX)Xn+IXXn+1 o

In practice we adopt the iterative form as follows:

(i) Y~~1 = Yn+Xn

(9) (ii) X~L//) = p[ef(tn+ l' y~L~1)+ y~L~1- Yn]
{
(iii) y~L++/) = Yn+(l-lX)xn+IXX~L//)

3. Truncation error of the method.

In order to get the local truncation error, we rewrite (8) as follows

(10) Yn+1-Yn = (1-IX)Xn+IXXn+1

= (1-IX)p[ej" + Yn - Yn-1] + IXp[efn+ 1+ Yn+ 1- Yn].
120 HAN TlAN-MIN

Expanding Yn+ 1 and Yn-l in Taylor series, we get

(11) , _j,
Y" - n-2
11-2O'+p
I-p
"
hYn+
(a2-61).2''h "Yn + ...
(notice the relation h(1-p) = hli/(h+li) = lip); the local truncation error is es-
sentially

I I-2O'+p 21/
(12)
-2 I-p hYn'

From (11), if a = t, we get

I I - 20' + p h2 1/ _ I P h2 1/ _ I 1 h3 1/
2 1-P Yn - 2 1-P Yn - 2 ~ Yn

which means that the method is second order with the local truncation error

I 1 3 1/ 1 3 1/1
(13) - 2 ~h Yn +-rrh Y + ....

From (12), if a =1= t, the method is first order.

4. Stability of the method.

In order to study the stability of the method we substitute the model equation
Y' = AY into (10), and get the characteristic equation

(1-O'p-O'pliA)Jl.2_(1 +p-2O'p-O'pGA.+pliA)Jl.+p(1-O') = o.
The stability region in the GA.-planeis enclosed by the following curve

A _ (1- O'p)ei21J- (1 + P - 2O'p)eilJ+ (1- o')p

(14)
e - O'pei21J+(I_O')peilJ .

The outline of stability regions, for a = 0.9, 0.7, 0.4 when p = I (h -+ 00),
is given in figure I.
If a = 0.1, 0.2, 0.3, 0.4, the stability regions are in the left half plane and
inside the closed curve. For a = 0.6, 0.7, 0.8 the situation is different, and the
region may spread over the left and right half planes and outside the curve.
The stability region becomes large when a changes its value from small to large.
If we keep a = constant, the region becomes large when p goes from large to
 NUMERICAL SMALL PARAMETER METHOD ... 121

0.6
0<=0·7 0.5
P='
o.i
0.3

Fig. I.

~2 -I
Fig. 2.

Stability regions in different cases

as explained in the text.

Fig. 3.

small, for instance, for (X = 0.9, if p = 1 a small part of the stability region
must be removed from the left half plane and if p = 0.8 the stability region
includes the entire left half plane.
In figure 2 we give the BA.-planestability region for (X = 0.5 and p = I, 0.5, 0.1;
because of symmetry we only give the upper half part.
When B = I, we may draw the A.-planestability region; we then consider h
as a parameter. For B = I, h = I, (X = 0.9,0.7,0.5; 0.3, 0.1 the A.-planestability
region is given in figure 3.
When (X = 1, our method is equivalent to the Backward Euler Method.
Making (X = I in (11), we can see the agreement between the two local truncation
errors of our method and Euler's. On the other hand, we can change the
BA.-planestability region into the region on the hA.-plane.In fact, in (14) we set
BA. = B(X+ iy) and obtain
. 1-p _. B • B •
BX+IBY = -- (l-e .11) = - (1-cos 0)+1- sm 0
p h h

i.e: hx = I-cos 0, h = sin 0.

The case (X = 0 is equivalent to the explicit method

see [9]. In fact, setting (X = 0, {3= 1 +P in (10) and noticing the relationship
Bp = h(1- p), we can prove this point. The hA.-planestability region is inside the
following closed curve:
122 HAN TIAN-MIN

{
hX ( 2h) P
= 1+--;- (cosO-I) = 2_p(cosO-l)
hy = sinO

which is an ellipse with the minor half axis I; when e --+ 0 (P --+ 2) the major
axis of this ellipse extends to infinity along the negative real axis.

5. Convergence condition of iteration and rate of convergence.

For general non-linear equations

Y' = F(t, Y)

if Ak = rxk + iPk are eigenvalues of the Jacobian matrix, where rxk < 0, then for
o< h < 00 the convergence condition of the iterative scheme (9) is

(l ~ k ~ n).

For any e > 0, the sufficient condition of convergence for the iterative scheme
(9) is h max IAkl< I. (Here the condition rxk < 0 is not necessary.) Without loss
of generality, we prove it for the model equation

Y' = Ay·

In fact, from hlAI < I, we have hIAI+h/e < I +h/e but JhA+h/el ~ hIAI+h/e
so we have IhA+h/el < I +h/e. Hence lehA+hl < e+h or h/(e+h)11 +eAI < I
i.e. pll +eAI < 1. This is what we need for convergence in scheme (9) for the
model equation Y' = Ay.
In order to study the rate of convergence it is enough to consider the model
equationy' = Ay. Let y~a~1 express the solution of equation (8) (heref(tn+ l' Yn+ 1)
= AYn+ 1)' then

(a) _ (l-rxP)Yn+(l-rx)Xn
Yn+ 1 - l-rxp(eA+ I)
and from scheme (9) we have

(L+l)_
Yn+l . - Yn +(1 -rx )Xn + rxp[1 (L)
eAYn+l (L)
+Yn+l-Y J

L
= [rxp(eA+ I)]<L+l)y~OJ1 + L [rxp(eA+ 1)J[(I-rxP)Yn+(I-rx)xnJ
i=O
 NUMERICAL SMALL PARAMETER METHOD .•. 123
1- [ocp(eA+ l)](L+ 1)
= [OCp(eA+1)](L+l)y~OJ1 + ------[(l-OCp)Yn+(1-OC)Xn]
l-ocp(eA+ 1)
= [ocp(eA+ 1)](L+ l)y~OJ1 + y~L - [OCp(eA+1)](L+ l)y~~L

and finally we get

Y(L+1)_y(~)
n+l n+l
= [OCp(eA+l)](L+ 1)(y(O) _y(~) )
n+l n+l'

This is the expression ~e wanted to find.

In some cases the rate of convergence for the iteration can be accelerated by
Aitken extrapolation (cf. [1]).

6. The considerations in practical computation.

(1) Choice of oc.

According to the stability analysis, we see that the stability region of the
method becomes large when oc changes from zero to 1. (In Fig. 1 we give the
situation for p = 1; for p +- 1 we have similar results). At the same time the
rate of convergence becomes slower. On the other hand, the local truncation
error reaches its minimum when oc = -t. Considering these three factors, for the
general computation program we suggest taking oc = !.

(2) Choice of e.
In most stiff systems, at the beginning of the computation, some components
decrease (or increase) sharply. During this period accuracy considerations
require that the step-size h be very small, so we take e = 1. After the transient
stage the parameter e should be reduced, but this is done by the program
automatically; for details, see [5]. For some kinds of systems, taking e ~
2/(max IAI), it may be kept constant in the entire process of computation. Among
30 examples we have worked on, 27 of them were done by setting e = 1 at the
beginning.

(3) Control of error.

We use truncation error expression (12) to control the accuracy. In our·
program we use (Xn+ 1 -xn)/h2 to express y~' approximately. Expression (12) is
written as follows:

(15)

If oc = 1- and e = 1, according to (13) we ought to use

(16) _ ~h3y"
~ n
+ ..1..h
12
3y'"
n
124 HANTIAN-MIN

to control the local truncation error, but this will add some complexity to the
program, so we do not treat this case separately, but always use (IS) to control
the error. This means that we only take the first term in (16), but this seems
to work well in practice.

7. Numerical experiments.

The numerical experiments have been done on 30 examples, 25 of them taken

from Enright et al. [2]. The numerical results are given in the appendix. These
25 problems reflect various aspects of stiff ODE :s. Four problems have been
taken from three textbooks [3, 6, 7]. The remaining one has been constructed
artificially by myself. It comes from a partial differential equation and we use
it to show the capability of our method in the situation when N is very large.

EXAMPLE
I. On the unit square we consider heat conduction equation

The initial and boundary conditions are as follows:

u(O, x, y) = I - sin fJx sin fJy

u(t, 0, y) = l_e-a./J2t sin fJy

u(t, I,y) = l-e-a./J2t(sinfJ+sinfJy)

u(t, x, 0) = l-e-a.fJ2tsinfJx

u(t, x, I) = l-e-a.fJ\sin fJx+sin fJ).

The theoretical solution is

u(t, x, y) = l-e-a.fJ2t(sin fJx+sin fJy).

We take (X = I, fJ = 5, Llx = Lly = 1/20, and treat it by the method of lines.

A system of ordinary differential equations

V' = AV+b

is formed in this way with the dimension 361.

EXAMPLE
2.
Yl(O) = 1 I [3], p. 210.
Y2(0) = O.
 NUMERICALSMALLPARAMETERMETHOD... 125

Eigenvalues: - I, - 1000.

Theoretical solution:

EXAMPLE3. Y' = A Y [6], p. 229.

A = [-~~ -~~
-20
20
J
40 -40 -40

Eigenvalues: - 2, - 40 ± 40;.
Theoretical solution:
Yl = !e-21+te-401(cos 40t+sin 40t)
Y2 = !e- 21 - te- 401( cos 40t + sin 40t)
1 Y3 = -e-40I(cos 4Ot-sin 40t)

EXAMPLE4. Y' = UZ-UBUY [3], p. 219.

1
with U = -
2
[I J 1
1
1
-1
1

1
1
-1
1
1

Y(O) ~ [~1]
PI = 1000, fJ2 = 800, P3 = -10, P4 = 0.001.

Eigenvalues: - 1002 -+ - 1000, - 802 -+ - 800,

8 -+ -10, -2.001 -+ -0.001

Theoretical solution:
126 HAN TIAN-MIN

EXAMPLE 5. [7J, p. 268.

During the course of computation L~Yi must be kept constant, e.g. = 1. In

table 1 we give the results of these 5 problems. For each of them we take
Q( = 0.5, E1 = O. In examples 1,2, 3 single precision is used and in examples 4

and 5 we use double precision. If the theoretical solution is known we give the
error of the last step; for example 5 we give the numerical solution of each Yi
and the value of L~Yi - 1.

Parameter list of table 1.

eo the initial value of parameter e
ef the final value of parameter e
TF we expect integration over the range (0, TF)
RTF the real range is (0, RTF) in our computation
HO the initial stepsize
HF the final stepsize
NS number of steps over the range (0, RTF)
NF number of derivative evaluations over the range (0, RTF)
NE number of derivative evaluations by the explicit Euler method over the
the range (0, TF)
Error the maximum absolute value of difference between theoretical solution
and numerical solution for each component in the last step.
N, TOL, ETA have the same meaning as before.
Table 1.
Example N eo ef TF RTF TOL ETA

I 361 I 0.65xlO-J 1.0 1.005 10-2 IO-J

2 2 I 0.14 X 10-2 20.0 20.02 10-4 10-5
3 3 I 0.33 X 10-1 20.0 20.08 10-4 10-5
4 4 I 0.59 X 10-1 1000 1044 10-6 10-7
5 3 0.2xIO-J 0.2 X IO-J 1000 1000 10-6 10-7

Example HO HF NS NF NE Error

I 0.0001 0.0288 160 160 1600 0.023

2 0.0001 0.2432 662 1301 10000 0.68xlO-J
3 0.005 0.125 202 644 800 0.54 xlO-5
4 10-9 4.965 1442 3931 0.5 X 106 0.42 x IO-J
5 0.001 0.1232 22014 22059 0.5 X 107 *)

*) Note: At the end of the integration range YI = 0.330570, Y2 = 0.1958 x 10- 5, YJ = 0.669428,
I-L~ Y. = 2 xIO- ll
.
 NUMERICAL SMALL PARAMETER METHOD ... 127

8. Discussion and further work.

In the spirit of B. Lindberg [8J, using the strategy:

(1) equal maximal global errors along the trajectories
(2) choosing optimal stepsize sequences according to a criterion on the error lstep,
we can get a better understanding of the new method.
Let WE' KE, WE' K£ express the total amount of work and the amount of work
of each step for the Backward Euler Method and the new one, according to
[8J, p. 878 (7.12),

WE = KE {m~x[1y'I(t)1 DY"(S)I-1!Z ds J}. f:1y'I(t)11!Zdt· (2GE)-1

l¥. = K£ {m~x[1y'I(t)1
t1y"(S)I-Z!3dS J}l!Z. f:1y'I(t)11!3 dt· (sGE)-t

K£ = O(N).

Here GE stands for the maximum global error in the interval and N for the
dimension of the equation.
If the system is very stiff we must take s very small and the new method has
disadvantages. But for moderately stiff problems we take s = 10-4, GE = 10-4,
then l/GE = (sGE)-t and

Hence we come to the conclusion that the new method is very favourable for
large moderately stiff problems.
The numerical results indicate that the method is comparatively inefficient
for problems Az, A3, D 4 (see appendix). But in another experiment, using the
information of the eigenvalues, we have:

80 = 8£ TOL NS NF Y(RTF)

Az 0.001 10-4 1926 2143 YI = 0.9984-1 Y9 = 0.8998

A3 0.0002 10-4 2308 2572 YI = -0.1111-1 Y4 = 0.1112
D4 0.0005 10-4 324 333 YI = 0.5515 Y3 = -0.1806-5

Comparing the results: 25760/2143 ~ 12, 14061/2572 ~ 5, 19029/333 ~ 57.

Why is that so? It may be an interesting problem for further investigation.
128 HAN TIAN-MIN

Appendix

In this appendix we give the numerical results of 25 problems which come

from Enright et aI. [2], p. 28-33. We use a star "*,, to indicate problems cal-
culated with double precision. The construction of Table 2 differs somewhat
from Table 1; we have omitted the two terms NE and Error. The reason is that
the eigenvalues of some problems are variable and the theoretical solution is
unknown. However, we can estimate NE from the eigenvalues, variable or
constant, as given in [2]. The correctness of the results can be verified from
the theoretical solution, if it is known, or by comparing the results with dif-
ferent parameters for same problem.
[Note 1] The parameter El, i.e. the lower bound of e, is useless except in C4,
Cs, D1. If we do not have any information about the system, we
can make it equal to a small positive number, say El = 10-9 or
even El = O. In problem C4, if we set El = 0, then for TOL = 10-4
we have ef = 0.00368 which is much less than 0.02 that we expect
(because the maximum eigenvalue of this system is 100). If e is too
small during the computation it will destroy the accuracy of the
results. For this problem, in the case we know the maximum eigen-
value, we put El = 0.02, and the actual result is ef = 0.01463 which
is a little smaller than 0.02. This seems contradictory with the meaning
of El (the lower bound of e), but we code our program this way:
if e < El we do not reduce e further. For TOL = 10-6, if El = 0,
e may reach 0.5352 x 10-4 and the stepsize becomes very very small
and we interrupt the computation. When El = 0.02 we have ef =
0.01734 and get a satisfactory numerical solution.
[Note 2] In problem Cs for TOL = 10-4 if El = 0, then ef = 0.004877 and
ifEl = 0.02, then ef = 0.01951.
For TOL = 10-6 if El = 0.02 then ef = 0.01734. In all these 25
problems, the parameter El is important only for C4 and Cs in the case
TOL = 10-6. From this we draw the conclusion: For problems with
strong nonlinear coupling from smooth to transient components, the
parameter El is important and for others we can set El = O.
[Note 3] For high precision (say TOL = 10-6) we want to choose El through
problem D1• We run this problem with lower precision (TOL = 10-4)
getting ef = 0.03704. In the case TOL = 10-6 we setEl = 0.03704
getting ef = 0.02926.
[Note 4] D6 is only one of these 25 problems for which integration was not
completed over the region (0, TF). In the region (0, 0.4853) we have
spent one minute of computer time. Further computation with the
current stepsize will be very expensive, so we decided to interrupt.
Here TOL = 10-6•
 NUMERICAL SMALL PARAMETER METHOD ... 129

[Note 5] Instead of taking Xf = 1 in [3], we take Xf = 100, i.e. put TF = 100

in problem E2• The numerical results show that the solution of this
problem is a periodical vibration with the amplitude ±2 (approxi-
mately) and there are nine wave crests in the region (0, 100).

Finally, we want to point out that in all these 30 problems overflow occurred
in only three of them, namely example 5 of table I and D6, E5, when Go = 1.
Using information of the eigenvalues we chose Go = 2/m~ IAI instead.

Table 2. Results of experiments on 25 problems from Enright et al. [2l

N Ef TF RTF HO HF NS NF Y(RTF)

A. 4 0.03292 20.0 21.37 0.01 1.642 295 781 Iyil < 0.8 x 10-4
4 0.03292 20.0 20.04 0.01 0.1170 930 2816 ly;I < 0.3 X 10-4

A2 9 0.001219 120 120.0 0.0005 0.009027 11221 25760 YI = 0.09983, Y9 = 0.8998

9 0.3806-3 120 120.0 0.0005 26.73 14797 20217 YI = 0.09945, Y9 = 0.8995

AJ 4 0.2572-3 20.0 20.0 IO-s 0.2731-2 7555 14061 YI = -0.1313-2, Y4 = 0.1321

4 0.1070-3 20.0 20.0 IO-s 0.3513-2 11833 21756 YI = -0.1283-2, Y4 = 0.1285

A4 10 0.1608-4 1.001 IO-s 0.133r2 1623 2279 YI = 0.3255, lylOl < 10-12
10 0.3568-4 1.0 IO-s 0.4161-3 5238 9898 YI = 0.3653, lylOl < 10-16

B. 4 0.02469 20.0 20.02 0.007 0.01956 1839 6056 YI = -0.5043-3, Y4 = 0.0

4 0.02195 20.0 20.00 0.007 0.01528 6846 20132 YI = -0.2065-s, Y4 = 0.0

B2 6 0.3333 20.0 20.12 0.01 0.2883 l3"5 453 IYd < 10-" Y6 = 0.1323
6 1 20.0 20.02 0.01 0.09123 593 1744 IYd < 10-8, Y6 = 0.1350

B3 6 0.2222 20.0 20.18 0.01 0.2883 157 513 IYd < 10-9, Y6 = 0.1311
6 1 20.0 20.09 0.01 0.1026 617 1859 IYd < 10-8, Y6 = 0.1340

B4 6 0.2053 20.0 20.07 0.01 0.08779 240 1026 lyll < 10-4, Y6 = 0.1319
6 0.6667 20.0 20.00 0.01 0.08659 808 2642 lyll < 10-8, Y6 = 0.1352

Bs 6 0.02601 20.0 20.00 0.01 0.8341-2 2492 10603 IYd < 10-3, Y6 = 0.1352
6 0.02601 20.0 20.01 0.01 0.8455-2 3078 11886 lyll < IO-s, Y6 = 0.1351

C1 4 0.08779 20.0 20.78 0.01 1.169 159 576 YI = 0.5345-3, Y4 = 0.20000-1

4 0.1975 20.0 20.16 0.01 3.329 747 2334 YI = 0.4006-3, Y4 = 0.20000-1

C2 4 0.01951 20.0 20.85 0.01 1.110 206 660 YI = 2.000, Y4 = 0.4003-1

4 0.02195 20.0 20.02 0.01 0.1284 811 2196 YI = 2.000, Y4 = 0.4003-1

C3 4 0.02195 20.0 20.07 0.01 0.1461 252 792 YI = 2.000, Y4 = 0.4333

4 0.01734 20.0 25.10 0.01 6.58 868 3109 YI = 2.000, Y4 = 0.4333

C4 4 note I 20.0 20.10 0.01 0.2574 396 1438 YI = 2.000, Y4 = 0.41993

4 20.0 20.02 0.01 0.1475 2034 7934 YI = 2.000, Y4 = 0.41983
 130 HAN TIAN-MIN

N ef TF RTF HO HF NS NF Y(RTF)

Cs 4 note 2 20.0 20.01 0.01 0.9669 406 1716 y, = 2.000, Y4 = 0.37115

4 20.0 20.08 om 0.2777 2444 12599 y, = 2.000, Y4 = 0.37125

D, 2 0.03704 400 400.1 0.017 0.1104 1328 1767 Y, = 21.48, Y2 = 26.12

2 note 3 400 400,0 0.017 0.02247 6945 19053 Y, = 20.59, Y2 = 24.85

D2 3 0.5144-3 40.0 40.05 0.00001 0.05040 2045 3706 Y, = 0.6992, Y3 = 28.93

3 0.2855-3 40.0 40.00 0.00001 0.8464-2 12246 20983 Y, = 0.7058, Y3 = 27.28

D3 4 0.7316-2 20.0 20.03 0.000025 0.06568 350 631 Y, = 0.6396, Y4 = 0.3169

4 0.5781-2 20.0 20.04 0.000025 0.8099 750 2355 Y, = 0.6396, Y4 = 0.3169

D4 3 0.6510-3 50.0 50.0 0.00029 0.5976-2 9652 19029 Y, = 0.5956, Y3 = 0.4920- 5

3 0.5787-3 50.0 50.0 0.00029 0.2412-2 11391 18618 Y, = 0.5946, Y3 = 0.1993 - 5

Ds 2 0.1736-2 100 100 0.0001 0.04157 206 301 Y, = -0.9988, Y2 = 0.9927

2 0.1829-2 100 100 0.0001 0.998T2 1984 3507 Y, = -0.9906, Y2 = 0.9822

D*6 3 0.5-7 0.4853 10-9 0.3371-5 20650 20683 Y, = 0.8524, Y2 = 0.1476

note 4

E*, 4 0.9259-2 1.009 0.0068 0.03873 27 126 Y, = 0.9997-8, Y4 = 0.2596-5

4 0.9259-2 1.000 0.0068 0.118 61 316 Y, = 0.9999-8, Y4 = 0.2246-6

E*2 2 0.1975 100 100.1 0.001 0.1501 3791 12794 note 5

2 1.0 100 100.0 0.001 0.04174 10750 37042

E3 3 0.02469 500 500.2 0.01 1.483 1331 2171 Y, = -0.158T2, Y3 = 26.43

3 0.00578 500 500.1' 0.01 0.1567 9919 16985 Y, = 0.4226-2, Y3 = 25.59

E4 4 0.2710-3 1000 1000 0.00005 0.1305 24288 29228 Y, = 19.96, Y4 = 0.03573

4 0.7136-4 1000 1303 0.00005 650 7897 18505 Y, = 19.85, Y4 = 0.1514

E*5 4 10-5 1000 1005 0.00005 7.981 30 5 Y, = 0.1760-2, ly41 < 10-8
4 10-5 1000 1001 0.00005 2.306 199 247 Y, = 0.1755-2, IY41 < 10-8

Note: The results in the first line of each problem correspond to TOL = 10-4, ETA = 10-5, the second'
line correspond to TOL = 10-6, ETA = 10-7• Except in D6 and Es we set eo = 1.

Acknowledgement

The author is most grateful to the referee for very helpful suggestions and useful
comments.

REFERENCES

1. G. Dahlquist, A. Bjorck, Numerical Methods, Prentice-Hall, Englewood Cliffs, N.J. (1974).

2. W. H. Enright, T. E. Hull and B. Lindberg, Comparing numerical methods for stiff systems of
O.D.E:s, BIT 15 (1975), 10-48.
3. C. W. Gear, Numerical Initial Value Problems in Ordinary Differential Equations, Prentice-Hall,
Englewood Cliffs, N.J. (1971).
 NUMERICAL SMALL PARAMETER METHOD ... 131

4. Han Tian-Min, A numerical methodfor solving the initial value problems of stiff ordinary differential
equations, Scientia Sinica Vol. XIX No.2 (1976), 180-198.
5. Han Tian-Min, A FORTRAN program for stiff O.D.E:s by a small parameter method, Technical
Report of The First National Applied Software Congress, Guilin, China (1981).
6. J. D. Lambert, Computational Methods in Ordinary Differential Equations, John Wiley and Sons,
London (1973).
7. L. Lapidus and J. H. Seinfeld, Numerical Solution of Ordinary Differential Equations, Academic
Press, New York (1971).
8. B. Lindberg, Characterization of optimal stepsize sequences for methods for stiff differential
equations, SIAM J. Numer. Anal. Vol. 14, No.5 (1977), 859-887.
9. Z. Zlatev and P. G. Thomsen, A one-parameter family of explicit multistep methods with long
stability regions, Institute for numerical analysis, Technical University of Denmark Report No.
NI· 77-06 (1977).