ima j appl math 1968 oliver 399 409
TRANSCRIPT
-
7/23/2019 Ima j Appl Math 1968 Oliver 399 409
1/11
-
7/23/2019 Ima j Appl Math 1968 Oliver 399 409
2/11
4 0 0 J. OLIVER
2 .
Error Propagation
in
Initial Value Problem s
Consider first
the
system
y&)
=
Pr.i(x)
yi
(x) + ...+P
r
Jx)yJLx) +
RXx)
( r = l , . . . , n )
(la)
or, in the usual notation,
y'(x) = P(x)y(x)+R(x), (lb)
where we wish to evaluate some or all ofthecomponents
Y
r
(x)
ofaparticular solution
Y(x) over some range
[a,b],
Y(a) being specified (the theory may be extended to more
general initial conditions).
Suppose initiallyweemploy a linear one-step forward integration method to produce
a computed solution Y(x,) over a set of discrete pointsx,where
a=
* o < x i ,(*) {s=
1,2,...,/ ),
(4)
for some real Ai,...,/^ with A
t
^ ^2>. .. >
A*,
where the components (p,,i(x),...,
t
^(x)
of
,(jc)
are slowly varying functions of x.(In the special case of a constant matrix P
having distinct eigenvalues
Xi,...,
A, the,{x) are its eigenvectors, and thus independent
ofx.)In order to substitute fore
Jt
, in (2b) we denote byW
tir
(x)the cofactor of4>
lif
(x)
in
W{x) =
.. .
nt
i(x)
(5)
atUniv
ersityofSouthCarolina-ColumbiaonNovember24,2010
imamat.oxfordjournals.org
Downloadedfrom
http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/ -
7/23/2019 Ima j Appl Math 1968 Oliver 399 409
3/11
INSTABILITY
OF
DIFFERENTIAL SYSTEMS
401
which leads
to
'
n
f
B
th Ix ~\ W (x
^
Yl x }
=
Q
E | E
W
(
X
J
Y ^ ) ^
X
J
)> X
exp{A,(x,-Xj)}
written
so as to
display bo th
the
local character
of
the errors ,
and the
manner
in
which
they are propagated.
The behaviour
of
the functions
which may in theory assume any values; we have written the result in this form,
however, since
in
many practical cases these quantities will
not
greatly exceed unity
in
magnitude. Similarly
the
initial effect
of the
propagated relative error
is
dependent
upon Yj(Xj)/YXxj),
and
upon
the
local truncation
and
rounding error. This latter
quantity depends, amongst other factors, uponthew ord-lengthof themachine: thus,
in the case of a binary machine having a mantissa in floating point mode of t bits,
we expect the quantity
Uxj)
=
\eJLxj)/YjLxj)\ (8)
to
be
either zero
or
else
of
order
2~ at
least.
So
long
as Y(*
y
) ~
Y(xj),
any
non-zero
tiXj) shouldbe at leastof this order, resultingin a certain minimum levelof relative
error, even
if the
local*truncation error
is
made very small.
The
effect
of a
relative
error
in one
component
Y,(xj) of the
required solution will
be
magnified
in
another
component
Y
r
(xj)in the
rat io
|Y&C j)IY
r
(xj)\ but
this will only
be
serious
if the
ratio
is larg e;wecommentonthis further inSection6.
Of greater concern, however, is the manner in which these local errors are pro-
pagated,andthisislargely de terminedby thebehaviourof
as
x
t
increases.It isimmed iately obvious that if, for example, some
Y
p
(x)
issuch th at
\exp(A,x)/Y
p
(x)\ increases rapidly withxfor one or moreA,,thenthe computed value
Y
p
(x)
will eventually
be
swamped
by
multiples
of
these complementary solutions.
N ow it is possible to derive a variety of elegant theoretical definitions of inherent
stability, based
on the
criterion that
the
propagated errors should
not
grow,
in
some
sense,asx increases. In practice, however, what matters is whether the propagated
relative errors grow
so
rapidly
as to
prevent at tainme nt
of the
required accuracy over
the particular range of interest in a given computational environment, and this is
necessarily
a
relative decision.
In a
similar manner,
we
shall speak
of one
solution
dominat ing another as meaning that their ratio increases appreciably in magnitude
as
x
increases, with ou t defining this precisely.
Thus the method will be ineffective for any particular solution
Y
p
(x)
which is
strongly dominatedby one or moreof exp(A,x),but it isimportant to realize that it
may neverthelessbeeffective for those
Y
r
{x)
whichare not dominated by any of the
complementary solutions.The disastrous error propagation in Y
p
(x) will eventually
cause
Y
p
(x)to
behave like
a
m ultiple
of its
dom inant complementary solution, rather
atUniversityofSouthCarolina-Columbi
aonNovember24,2010
imamat.oxfordjournals.org
Downloadedfrom
http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/ -
7/23/2019 Ima j Appl Math 1968 Oliver 399 409
4/11
4
J. OLIVER
than like
Y
p
(x),
thus altering the order of the rounding errorst]
p
{xj),but this will not
of itself induce instability in the other solutions.
We have thus far analysed error propaga tion in terms of
the
fundamental solutions
of
the
original differential system, whereas a practical one-step method approximates
it by a system of linear difference equations having a particular solution Y*(x,) and
complementary solutions vW*(x,). One desirable feature of any method is, however,
that the local trunca tion errors in each of the required solutions
Y
r
(x )
should be small,
and achieving this will usually ensure that any dominant complementary solution of
the differential equations has a corresponding dom inant solution in the approximating
system. Consequently any inherent instability in the original problem almost always
leads to similar error propagation in any discrete system which closely approximates it.
The same is true both of linear multi-step methods, with the added complication of
the spurious complementary solutions which the replacement system of difference
equations possesses, and also of global methods of solution such as Chebyshev series
(although their theory is not yet well developed).
It may of course be possible to represent the differential system by a discrete one
whose particular solutions closely approximate the ones required, but in which the
dominant complementary solutions are replaced by ones with a milder form of
behaviour. An example of this, involving Chebyshev collocation applied to a single
fourth order equation, is cited by Oliver (1968a). However, this remedy will not
usually be feasible, and so an alternative method of solving an inherently unstable
initial value problem must be found: we return to this point jn Section 5.
3 .
Error Propagation in Boundary Value Problems
The stability of the two-point boundary value problem may be analysed similarly.
For simplicity, we take the following type of conditions:
y
r
(a)
=
Y
r
(a \
0
= \,...,n-m)
yp(r)(b) = YpO)(b),(r =
n-m+l),...,n)
where
1
-x )} exp
{^(x -b)}
,.Z
+l
W(a,b)
with the stability characteristics largely dependent upon the terms
We conclude, therefore, that inherent instability in Y
p
(x)with respect to the boundary
conditions at x = b will result if and only if exp (A,^) is sufficiently dominated by
Y
p
(x).
Finally, to investigate the effect of errors introduced during the computations,
we suppose that a linear one-step formula, of the form
y(xd = A(xOy(*,-i)+*(*
-
7/23/2019 Ima j Appl Math 1968 Oliver 399 409
6/11
4 0 4 J. OLIVER
Under the above assumption regarding the complementary solutions, we may write
**/)= t cJ'Xxd+S'Mxj)0 =0,1,...,;)
1 = 1
t j \ .,N ) (17)
where the constants are determined by (16) and by equating the representations (17)
at their common pointXj.This gives
0
0
Ija) 0
0 t > -
m +
0 >( )
0
.)(*)
dt
An,
0
.(18)
Using similar algebraic manipulation to tha t given earlier, substitution for
c
r
and
d
r
in (17) shows that each component
i,(
x
j)
contributes an am ount e^ x , ) to (*,) where
(19)
{
Xj
)
V(xJ)
the determinants F having been norm alized as previously. Examination of the
terms which govern propagation, namely
indicates that perturbations will propagate unstably in Y
p
(x ) if and only if Y
p
(x ) is
dominated byexp
( *+i*)
o r
itself dominatesexp
{X^pc),
agreeing with the conclusions
already reached in regard to changes in the boundary conditions. The perturbation
{xj) obviously encompasses local truncation error, and rounding errors can also be
regarded inthislightsee, for example, an analogous treatment of recurrence relations
by Oliver (1968b)though it is more difficult to assess their probable order of
magnitude than in the initial value case.
The comments m ade in Section 2 regarding the relationship between the comple-
mentary solutions of the differential system and those of the discrete approximating
system apply equally well here.
atUniversityofSouthCarolina-Columbi
aonNovember24,2010
imamat.oxfordjournals.org
Downloadedfrom
http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/ -
7/23/2019 Ima j Appl Math 1968 Oliver 399 409
7/11
INSTABILITY OF DIFFERENTIAL SYSTEMS 4 0 5
4. Choice of Boundary Conditions
It
is obvious from the foregoing analysis that for a differential system having
exponential-type complementary solutions and particular solutions of a somewhat
similar nature, there is an appropriate choice of boundary conditions which lead to
a well-posed problem for each Y
r
(x). More precisely, if YXx) is dominated by
exp(A,x)fors= 1,..., m but not for s=m
+
1,...,n, then boundary conditions of the
form (9) are most effective. In particular, forward integration of an initial value
problem should be employed when Y
r
(x ) is not dominated by any of the comple-
mentary solutions, and of course integration in the reverse direction if it does not
dominate any of them.
Although a particular boundary value problem may be inherently stable with
respect to the required solutions, it suffers from the major disadvantage that jury
methods of solution entail considerable computation and storage. Shooting methods
will normally be impracticable, due to inherent instability, although Conte (1966)
suggests that this instability may
be
avoided
by
orthogonalizing
the
computed solutions
as the forward integration proceeds. We stress, however, that the boundary value
problem itself m ust be well-posed if
this
approach is to succeed.
If the available boundary conditions specify an inherently unstable system for one
or more of the required solutions, and no further conditions are available, then no
improvement
is
possibleindeed, the effect of the instability cannot
even be
recognized
in one computed solution. Although this may be true of a mathematical system,
however, in physical problems some additional information is usually available,
perhaps in the form of vaguely-stated conditions of boundedness. For example,
Midgeley (1966) cites the case of the linearized equation for vertical wave propagation
in a vertically inhomogeneous gas, for which the relative behaviour of the various
complementary solutions is known from physical considerations. This mathematically
redundant information may indicate that the problem as specified is ill-posed, and
tha t boundary conditions of a different nature should, if possible,
be
obtained and used.
Furthermore, if an exact well-posed problem cannot be formulated, a knowledge of
the general relative behaviour of the various solutions can itself lead to a stable
system.
5. Replacement of
n
Hi-posed Initial Value Problem
To illustrate this, we return to the initial value system in which Y(a) is specified,
and suppose tha t
a
particular solutionY
p
(x )
is
required which
is
dominated
by exp
(A,*),
(s
= l,...,m ), to such an extent that forward integration methods are ineffective.
Assume also that no other conditions are available, except the knowledge that Y
p
(x )
is dominated by exp(b,
and consider the solution
z(x)
specified by the
boundary conditions
zj(a)= YJta) (s = l , . . . ,n-m )
Note that the detailed choice of boundary conditions is unimportant; what matters
is tha t onlyn mof
the
exact ones atx = ashould be used, and zero values of Y^x)
atUniversityofSouthCarolina-Columbia
onNovember24,2010
imamat.oxfordjournals.org
Downloadedfrom
http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/ -
7/23/2019 Ima j Appl Math 1968 Oliver 399 409
8/11
4 0 6
J.
OLIVER
substituted nearx = x
M
for the remainder. z(x) existsjf
# 0, (21)
and this we assume to be true. Following the usual procedure of introducing
normalized determ inants in Cramer's rule, the relative difference betweenz(x)and
is found to be of the form
(22)
and thus the more exp(k^x)dominates Y
p
(x), the closer will z
p
(x,) approximate to
YJxd over[a, b].
A similar analysis to tha t performed in Section 3 shows that this differential system
for z
p
{x^ is well-conditioned to truncation and rounding errors, so that this is an
effective method for evaluating an approximation to the required solution. The
parameter x
M
is taken sufficiently large so that the required accuracy is obtained,
though as always in such algorithms the optimum choice ofx
u
is a non-trivial decision
for an analogous problem in recurrence relations see Olver (1964, 1967).
Midgeley (1966) has suggested a method using step-by-step methods and quadrature
to evaluate each of the complementary solutions in turn in order of decreasing dom in-
ance. This avoids the necessity for solving a boundary value problem, although he
effectively solves one for each complementary solution, using an appropriate number
of arbitrary initial conditions with zero values at the opposite end of the range. The
method is riot, however, immediately applicable to non-homogeneous equations.
Finally we note that even when the nature of the complementary solutions is not
apparent from physical considerations, their behaviour can be determined by a process
such as Midgeley's, or equivalently by solving directly a series of boundary value
problems for the homogeneous form of the differential equations using m arbitrary
conditions at some point to the left of x = a,and
nm
zero conditions at x
u
> b,
for m = n,n1,...,0. Provided the complementary solutions are of exponential type,
this will indicate their behaviour (including those with nearly equal dominance), and
if the nature of the required solution is known then a well-conditioned boundary
value problem can often be constructed.
6. Numerical Examples
We take first an example based on that of Conte (1966):
y'^x) = -
1-8368>'
1
(X)-6-2976>'
2
(X)-1008}'
4
(X)
+
10-08(1
+ x
2
/2) ,
y'
2
(x)=
-6-2976>>
1
(x)
+ l-8368.y
2
(x)-10-08.K3(x)-49-875 exp(20x), (23)
y'
3
(x )
= -10-08 j>
2
(x)+3-4832y
3
(x)-ll-9424>>
4
(x)-
37-40625
exp (20x)
+11
-9424(1
+
x
2
/2),
y't(x) = -1008j '
1
(x)- l l -9424y3(x)-3-4832^(x) +
x +3-4832(1+x
2
/2).
atUniversityofSouthCarolina-Columbia
onNovember24,2010
imamat.oxfordjournals.org
Downloadedfrom
http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/ -
7/23/2019 Ima j Appl Math 1968 Oliver 399 409
9/11
INSTABILITY
OF DIFFERENTIAL SYSTEMS 4 0 7
This is, conveniently, the special case of a constant matrix P, and the complement-
ary solutions vary exactly like the exponentials
exp
(20*),
exp(*) , ex p (- *) and
exp] (- 2 0 * ). Using the appropriate initial values to specify the particular solution
Yi(x)= 0-76506 exp (20*)-2
exp
(*)/3-2exp( - * ) / 3 ,
Y
2
(x) = -
2-65283
exp (20*) -
exp
(x)/2+8 exp ( - x)/9,
F
3
(*) = - 0-64575 exp
(20*)
+
3 exp
(*)/8 +
2
exp ( - *)/3,
(
'
F
4
(*) = l+*2/2+sinh(*) ,
we note tha t Y
4
(x)is strongly dominated by exp (20*), whereas the othe r solutions are
not. Consequently we expect that forward integration is inherently stable for all
solutions except
YA(X).
To illustrate this without introducing the complication of parasitic solutions, we
approximate the differential system at equally spaced points *
;
by
- i )} . (25)
Taking an interval h of 2~
7
, and using (25) recursively over [0,2] on KDF9 (which
has a 40-bit mantissa) with the initial values appropriate to the differential system
produces a reasonab ly accurate approximation to Y\(x),as Table
1
shows, and simil-
arly for Yz{x)and Yi(x),whereas the computed values of
Y*(x)
are rapidly swamped
by the propagated error. Employing a smaller interval would of course improve the
accuracy of the first three computed solutions, but would not remove the inherent
instability for YA{X).
TABLE 1
Forward integration
for
first differential system
*
0-25
0-5
0-75
1
1-25
1-5
1-75
2
Computed
1 1 3 x 1 0 2
1-71x104
2-55x106
3-80x108
5-67x1010
8-46x1012
1-26x1013
1-89x1017
Tito
Relative er ror
6-89x10-3
1-24x10-2
1-79x10-2
2-36x10-2
2-93x 10-2
3-51 x 10-2
4 0 9 x 1 0 - 2
4-68x 10-2
Y4.x)
Computed
+ 6 -9 5 x 1 0 -1
- 2 1 0 x 1 0 2
- 5 0 1 x 1 0 4
- 1 0 2 x l 0
7
- 1 - 9 4 x 1 0 9
-3-50x1011
- 6 1 4 x 1 0 1 3
- 1 0 5 x 1 0 1 6
Exact
1-28
1-65
2 1 0
2-68
3-38
4-25
5-32
6-63
Since Y-t(*) is dominated by only one of the complementary solutions, it would
seem at first sight that a boundary value problem in which three conditions were
specified at * = 0, and a further one for some * > 0, would be well-conditioned. This
is indeed the case, yet computed values of Y*(x)again exhibit catastrophic error
accum ulation. This is not due to growth of local errors, however, but rathe r to relative
errors in Yi(*,),
Y
2
(xj)
and
Y
3
(xj)
producing greatly magnified errors in Y^Xj),
particularly for large *,. Referring to (19), the relevant term is Y&Cj)IY
r
(xj), with
equivalent ones in (12) and (13). Consequently the problem is well-conditioned in one
atUniversityofSouthCarolina-ColumbiaonNovember24,2010
imamat.oxfordjournals.org
Downloadedfrom
http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/ -
7/23/2019 Ima j Appl Math 1968 Oliver 399 409
10/11
4 8
J. OLIVER
sense, yet not in another, and there seems little remedy in such a situation (except
perhapstoproduce a fourth order differential equation for
YA(X)),
though itisthought,
fortunately, to be of infrequent occurrence in practice.
A more common problem is that typified by the same system (23), but with the
terms inexp(20*) omitted from the equations for
y
2
(x)
and
y
3
(x).\With
appropriate
initial conditions, the solution is
Yi(pc) = - 2 exp (*)/3 -
2
exp (- x)/3,
Y
2
(x) = - exp(x)/2+8 exp (- x)/9,
Y
3
(x) = 3exp(x)/8+2exp(-x)/3,
Y
A
{x ) =
(26)
and
the
initial value prob lem
is
ill-conditioned with respect
to all the
particular
solutions
Y
r
(x),
asillustratedbythe resultsofTable 2, obtained w hentheapproxima-
ting system (25)
was
solved recursively over [0,2] w ith
an
interval
of
2~
7
.
TABLE 2
Second differential system: relative errorsincomputedsolutions
X
0
0-25
0-5
0-75
1 0
1-25
1-5
1-75
2 0
Forward
Y i x )
0
3-12x10-7
1-31x10-6
2-02x10-3
2-24x10-3
2-74x10-1
3-29x101
3-92x103
4-63x103
integration
Y*x)
0
1-02x10-6
1-91x10-
218x10-5
2-29x10-3
2-72x10-1
3-24x101
3-88x103
4-67x105
Boundary
Yi(x)
0
4-98X10-6
1-88x10-3
3-88x10-3
6-20x10-5
8-63x10-5
111x10-4
1-34x10-4
108x10-4
value method
Y4.x)
-4 -7 3 X10-11
1-63x10-3
2-79x10-5
3-76x10-5
4-69
x 10-5
5-68x10-3
6-75x10-3
7-90x10-3
4-31x
10-s
If,
however,
one of
the initial conditions
a tx =0 is
discarded
in
favour
of
o ne
for
some positivex,and the approximating system solvedby a boundary value method,
then
as
expected the re
is no
evidence
of
inherent instability. Fu rtherm ore, we may use
the methodofSection5, withm = 1inthis case,togood effect. Forexample, solving
the system (25) subject
to
*,(0)= YJ[0) (s = 1,2,3)
24(2-5)= 0,
{27)
with
a
fairly large interval
of
2~
3
produces
the
approximations
to
Yi{x)
and
YA(X)
shown in Table 2. More accurate approximations can be obtained with a smaller
interval,or byapplication ofdeferred correction proce dures .
7. Conclusion
W e have investigated the inhe rent stability of systems of linear differential equ ation s
having complementary solutionsofexponential type,and shown under a numberof
reasonable assumptions that
to
any particular solution
of
exponential type there will,
atUniversityofSouthCarolina-ColumbiaonNovember24,2010
imamat.oxfordjournals.org
Downloadedfrom
http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/http://imamat.oxfordjournals.org/ -
7/23/2019 Ima j Appl Math 1968 Oliver 399 409
11/11