Discrete-continuous model

conversion
L. S. Shieh and H. Wang
Department of Electrical Engineering, University of Houston, Houston,
Texas 77004, USA

IX. E. Yates
Guidance and Control Directorate, US Army Missile Command, Redstone Arsenal,
Alabama 35809, USA
(Received March 1980)

This paper presents methods for model conversions of continuous-time

state-space equations and discrete-time state-space equations. An improved
geometric-series method is presented for converting continuous-time models
to equivalent discrete-time models. Also, a direct truncation method, a
matrix continued fraction method and a geometric-series method are
presented for converting discrete models to equivalent continuous models.
As a result, many well-developed theorems and methods in either con-
tinuous or discrete domains can be effectively applied to a suitable model
in either domain.

Introduction Continuous-to-discrete conversion

Many industrial processes and social systems are formu- Consider a linear, time-invariant continuous state equation:
lated via discrete models. Chemical processes and
economic systems are typical examples. The recursive J&(t) =&X,(0 + &U,(r) (la)
features of these discrete models and the recent availability X,(O) = (Y (lb)
of high performance low cost microprocessesors have
whereXOERn,UOERm,A,ERnxn,andB,~Rnxm.
enabled a new consideration of control and analysis of
The exact solution of equation (1) is:
these systems. On the other hand, many control systems
are formulated via continuous models for which many
theories and practical methods have been developed.
Many large practical control systems consist of both
continuous-time and discrete-time subsystems. For effec-
x,(t)
s
= at(t) x,,(o) + @.c(t- A>&Jo(X) dh
0
Pa)

tive analysis and synthesis of these composite systems where :

it is often necessary to convert a discrete subsystem to an
equivalent continuous model and vice versa. a,(t) = eact = [eAcrlk = [@d(T)lk= Qd(kT) (2b)
Harris1 has recently studied the problem of converting Q&T) 4 eAcT (2c)
a discrete state equation to an equivalent continuous model.
In his method a logarithmic matrix function is converted t=kT for k=O,1,2,... (2d)
to a logarithmic scalar function via a model transformation a),(t) is a continuous state transition matrix and Qd(kT) is
such that both approximate and exact continuous models a discrete state transition matrix, and T is the sampling
can be formulated from the discrete models. On the other period. U,,(t) may be a known analytical function or a set
hand, Shieh et al.’ have converted continuous-time state of discrete data whose function is unknown. To simplify
equations to equivalent discrete-time models via a block- the convolution integral in equation (2a), U,,(t) can be
pulse function method394 and a geometric-series method.’ approximated by using either of the following piecewise-
In this paper, the geometric-series method' will be improved constant functions:
and extended for accurately converting continuous state
equations to equivalent discrete models. Also, a direct lJd(t) 4 U,.(kT) = U,(kT) = G,(t)
truncation method, a matrix continued fraction method, for kTGt<(k+l)T;
and a geometric-series method are proposed for converting
discrete state equations to equivalent continuous equations. k=0,1,2,... (3a)

0307-904X/80/060449-07/$02.00
0 1980 IPC Business Press Appl. Math. Modelling, 1980, Vol 4, December 449
Discrete-continuous model conversion: L. S. Shieh et al.

or Cd3 = [Z, - ;A,T+_iZ(A,T)'-

_i&(A,T)3]-'
U&t) 4 U&T) = 1 [U&T) + U&T + T)] = UC@) x [I,+ :A,T+~(A,T)~+~~,(A,T)~]
=Q;~P~
forkT<t<(k+l)T; (8~)
k=0,1,2,... (3b) cd4=[~n-:~c~+
&(A,T)~-&(A~T)~
where U,.(kT) is a rectangular wave which can be con- + i&ii&T)41-’
sidered as obtained by applying a sampler and zero-order -
X [I, + ~A,T+~(AcT)2+ &4(A,T)3
hold device to Uo(t), while (I,(kT) is another rectangular
wave formed by sampling and polygonal holding the Uo(t). + i&&W)“1 = Q2f’4 WI
The approximate state due to Ud(t) is Xd(t). The new state
The corresponding input matrices Hdj for j = 1, 2,3,4 are:
equation becomes:
Hdl= T[I,- $A,T]-'B, (94
ri,(t) =&%(t) + MJ&) (4a)
HLi2= T[I,- :A,T+ &(Ac~)2]-lB, (9b)
X,(O) = X,(O) (4b)
The solution of equation (4) evaluated at t = kT is: Ham= T[I,- :A,T+ $(A,T)~-&(A,T)~]-1

k-l
x [I,+ -71i(~W)~l& (9c)
X&kT)= @(T)X&O)+ c @&kT-jT- T)HJI,O'T) Hd4= T[~,-~A,T+&(A,T)~-&(A,T)~
j=O
&&T)21Bc
+i&(A,T)"l-'[I,+ (94
fork=1,2,... (Sa)
where : Gdj in equation (8) are approximate models of the exponen-
tial matrix function &c* obtained by taking a finite
T number of dominant terms and approximating the infinite
~~ = eAc@B,da = [@&T)-In]A;'Bc summation of the other terms of the Taylor-series matrix
(5b)
s expansion. For example, G,, can be expressed as:
0
;A,T] for T< 2/1IA,11
Gd, =[I,- :A,T]-'[I,+
I,is an n x II identity matrix. Dropping the Tin Xd(kT)
and U&T) in equation (S), the corresponding discrete (lOa)
model becomes :
tiIn+A,T+;(A,T)2+J~3&(A,TY (lob)
Xd(k + l)= GdXd(k)+HdUd(k) (6a)
while the exact expression is:
Xd(O)=Xo(O) (6b)
where : Gd = eAc* = I,ta,T+;(A,r)'

Gd6 @,(T)=eACT (6~)

(1 la>
Hd =[Gd-Zn]A;'Bc (64
For a large scale system, on-line evaluation of the exact
transition matrix eAcT is a difficult task. An approximate =(,+A,.T+;(A,T)'+ f &A,Ty (lib)
j=3 j!
transition matrix is often used for on-line operations of a
control system. Using the geometric-series method, Shieh The first three dominant terms in both equations (lob)
et a1.2 have shown that Gd can be expressed using approxi- and (1 lb) are identical and the other terms differ in their
minants of eAcr as: weighting factors 1/2j-l in equation (lob) and l/j! in
Gd = eAc* = [e-:&*]-’ [ef&T] equation (1 lb). It has been shown that if Gdl and Hdl of
equations (8a) and (9a) and U,(t)(=U,(kT)) inequation
%QT'Pii?Gdj for j=l,2,... (7a) (3b) are used in equation (6) to solve the discrete-time
system of equation (l), the solution so obtained is identical
where :
to that obtained by using the Walsh function approach5
and the block-pulse function approach.4 Also, it has been
Qj= [I,- ; AT] +:$I(i2;:;ji)(i!f)
[In (A,T)‘]
Pb) shown that equation (6) may be used to obtain the numeri-
cal integration of a stiff state equation.4
From equation (7) we observe that [I, - (1/2j)A,T]-'
pi = [L + $LT] [L +;;; (;ol,;!) CM)‘] (7~) is a geometric series that converges if T < 2j/ll A,II and a
better approximation can be obtained if T< 2i/llA,II.
Thus, we can improve the approximants of equation (7a)
T<2j/llAJ, j= 1,2,... (74 as follows.
where 1)AJ is the matrix norm of the matrix A,. Rewriting equation (7a) and modifying the sampling
Some approximate discrete-time system matrices Gdj for period T with T, 2 Tfm where m is a positive integer
j=1,2,3,4are: yields :
G,, = [In - $A,T]-'[I,
+ &T]= Q;'P1 @a) Gd= [e- k&7’- [e;@-]
UW
Gd2 = [In- kA,T+ &(A,T)*]-' = {[,-tAcT”~-‘[~~AcTm]}m
(12b)
x [I,+~A,T+~(Ac~Z]= Q;'P2 = { ,,-t&T]-1 ,e’“cT]>”
(8b) (12c)

450 Appl. Math. Modelling, 1980, Vol 4, December

 Discretecontinuous model conversion: L. S. Shieh et al.

2 Gg (124 Ifm=4,wehave:
* lAPi>“’ i Ggi
z {Qf (12e) 4

Ggl=

(12f)
where :

T<
WlA,I
(1&J) =I,t
T,,, = T/m

A, = AC/m
ACT t gAc~)2 t __ ’

1
(AcT)3
~j = In - 1 A,T 22(!9

i-1
[ 2(j)(m)
t- ’ &T>4 + 2& (AcT)’+ . . (1SC)
[
(-l)‘(i - i)
x I*+C W)‘] (1W 23tS> 45

i=l (2’)(j)(i!)(m)’

1
An exact expansion for eAcT is:
1
pi = I, + - AJ
L 2(j)(m) Gd = e*cT = I, + ACT t &A,T)~ + - 1 (AcT)3

L
j-l
(i - 9
2y1.q
x m+C (ACT)’ (12i)
i=l (2’)(j)(i!)(m)’ I
+p ’ (A,T)4 +1 (ACT)’ + . . . (16)
T < (2)(~)(m)lllA,II (12j) 23(3) 24(7.5)

Our improved discrete state equation approximation to Comparing equations (Isa), (15b), (15c), and (16) we
the continuous system becomes: observe that the first three terms in all four equations are
identical and that each of the weighting factors for the
X,(k + 1) = G,X,(k) + L&U&) (13a) other terms (ACT)’ in equation (15) approaches the corre-
X,(O) = Xd(0) = X,(O) (13b) sponding terms in equation (16) as the value of m increases.
Therefore, the discrete-time model of equation (13) is
where X,(k) E X,(k) F X(t) at t = kT. Also: better than that of equation (6).
It has been reported4 that the simplest Gdl in equation
H, = [G, - Z,] A,‘B, (13c)
(lOa) or Ggl in equation (15a) has been successfully

m
applied to solve a stiff state-space equation for which the
For comparing equations (8) and (12) we list as follows:

2(m)I[1,+2&G&T)
fourth-order Runge-Kutta method fails. This is due to the

II
-1
fact that the Runge-Kutta method6 approximates the
Ggz In- &ACT)
infinite series expansion of eAcT by taking only the first
five terms, i.e.,
4 G,, for i= 1 (14a)

eAcT = In t ACT + ~(A,T)Z +

&) (AcT)3

t & WT4 t & WY5 + ... (17a)

2 G, for j=2 (14c)

z ... s 1, + ACT t f(Ac~)2 t &j Qm3

When m = 1 in equation (14a), we observe that Ggl in

equation (14a) is identical to Gdl in equation (8a), or: W’b)
G,, = ~~~ = [r, - ~A,T]-~[I~ t $A,T] T< 2/11-4,ll
Comparing equations (15a) and (17b) we observe that the
= Z,, t ACT t f(A,T)’ +& (AcT)3 + &j (AcT)4 first five dominant terms of Ggl in equation (15a) are nearly
equal to those of equation (17b), while Ggl contains an
additional infinite series which apparently Improves the
t & (AcT)5
+... (lsa> accuracy of the numerical integration of a stiff differential
equation. Therefore, the other more sophisticated models
When m = 2, equation (14a) becomes: will be more effective in solving these stiff differential
equations.

II
2
A matrix inversion has also been introduced in the pro-
T<4/11A,ll
posed method. However, a matrix inversion is a common
exercise for control system applications such as identifica-
=I,+A,T++(A,T)2+ - ’ tion, estimation, and optimal control problems.
(AcT)3
22(9 From equation (12j) we also observe that a more
sophisticated model, which uses more terms, enables the
+& use of a larger sampling period. This can be verified as
(AcT)4 + &) (ACT)5 + . . . Usb)
5 follows. Letting Ti designate the sampling periods if the

Appl. Math. Modelling, 1980, Vol 4, December 451

Discretecontinuous model conversion: L. S. Shieh et al.

model G, for j = 1 and 2 in equation (13) we have: equivalent continuous-time model so that well developed
theorems and methods in the continuous-time and
frequency domains’ can be effectively applied. The corre-
sponding continuous state equation is:
forj= 1 (18a) J&(t) =-W&(t) +&U,(t) (21a)
and Z(O) = J&(O) (21b)
where Xd(kT) SCXc(t) at t = kT and Ud(kT) = U,(t) at
t=kTand:
in U,(kT + T) - U,(kT)
UC(t) = Ud(kT) + (t - kT)
T

forj=2 (18b) k=O,1,2 ,..., kT<t<(k+l)T (21~)

From equation (6) we can solve for the equivalent system
Consider that A, is a diagonal matrix with all eigenvalues
matrix A, and input vector B, as:
$ located on the diagonal. The absolute value of the largest
eigenvalue, A,, is often used to evaluate the minimal
sampling period. Now we can compare the sampling periods A, = k lnG, (21d)
T, and T2 in equations (18a) and (18b) by solving the
following equation : and
B, = A, [Cd - ZJIHd (2le)
The logarithmic matrix function in equation (2 1d) can be
expressed by :
=
N
1-;h,T,+-
1
16m2
1
h&T:
r’
)) m n
-h&T; (194 1 n+2 +
1 n+4 +
16m2 t -R -R . . . GQ)
nt2 nt4 I
or
where :
1
T, = Tz< Tz (19b) R = [Gd - Z,] [G, + Z,]-’ (22b)
1+ $-+Q2 The matrix R is expressed by a rational matrix function in
a bilinear expansion form.’
from equation (19b) we conclude that a larger sampling Three methods are proposed for approximating 1nGd
period T2 can be used if Gsa instead of Ggl is used to formu- in determining A, as follows.
late the equivalent discrete-time model. As a result, we have
greatly increased our flexibility in selecting the minimal Direct truncation method
common sampling period among the various subsystems of The approximate A, can be determined by taking the
a large composite system. first several dominant terms of the expansion in equation
Given a unit-feedback continuous control system with an (22a) and truncating others. Some approximate models
input function r(t) and error function U,(t) it is often are :
necessary to determine the equivalent discrete model with
a sample and hold acting on the Uo(t) for digital control.’ A,=‘lnGd Wa)
A solution to this problem is to first convert the open-loop T
continuous system to a discrete-time model via our new
method. Then, the piecewise constant function Ud(t) will =;,R+jR”+$R’++R’t...] (23b)
be a good approximation of the Uo(t). As a result, the out-
put response of a closed-loop digital system to an input 2
r(t) will closely follow the original continuous closed-loop R (23~)
-T
system. If the continuous closed-loop system is directly con-
verted to an equivalent closed-loop discrete model with a
sample and hold acting on r(t), the discrete model gives Ej,, tfR”, (234
good output response only if the input frequency of r(t)
is sufficiently low in comparison with the sampling
+R tfR3t &KS] WeI
frequency.

-...
Discrete-to-continuous conversion
The equivalent B, can be obtained by substituting equation
Consider a discrete state equation: (23) into equation (21e). If the model in equation (23~) is
X,(/c + 1) = GdX,(k) •t H&,(k) (20a) used, we have :

&(O) = X,(O) (20b)

A, z ; (G, - Zn)(Gd + In)-’ Pa)
It is necessary to convert the discrete-time model to an

452 Appl. Math. Modelling, 1980, Vol 4, December

 Discretecontinuous model conversion: L. S. Shieh et al.

and

B, s ; (r, - +A,T) Hd (24b)

1

= ; R [(Kz + Kz,) N + (KZK3K4)] [N2

The approximate models in equation (24) are identical to
those developed from the discrete-time model Gdl in
equation (8a) and &r in equation (9a). Since R is a bilinear +(KIK2+K1Kq+K3K4)N+(K1K2K3K4)]-1
form, the numerical errors due to matrix multiplications in
equation (23) are minimized. =+R(-?R~ + $Z,)(R4 - ioR + Tr,)-l (274

Matrix continued fraction method Note that A, in equation (27a) is identical to that in
A method using a matrix continued fraction approxima- equation (23~). The B, can be obtained by substituting A,
tion’ has been shown to be effective in approximating high- of equation (27) into equation (21e).
order multivariable systems by a low-order one. The same
approach can be applied to approximate the logarithmic
matrix function in equation (22a). Rewriting equations
Geometric-series method
(21 d) and (22a) and expanding into a matrix continued
fraction results in: Rewriting equations (21) and (22) yields:

A,=$lG, (25a) A, = ilnGd

2 1
Wb) =r t$R3t...t1R"t -Rn+a
n n+2

1
1 1
n+4
t---R t-Rnt6+ ... (284
nt4 nt6

where N=R’: 1 1
n+2
. . . t--R” t -R
R = [G, - Znl[G, + ZJ’ n nt2

(28b)
The matrices Ki (=kiZn) are diagonal matrices and are the

1
1
matrix quotients in the matrix continued fraction expan-
+c R nt2i
sion. The scalars ki can be determined from the following i=2
scalar Routh algorithm”:
al,l= l,aI,j=O; j=2,3, ... (26a) The weighting factor of the term R”t2i in the infmite
1 series in equation (28b) can be approximated by:
a2,j= -. j=l,2,... (26b)
2/- 1’
1 1
ai,i=ai-2,j+l-ki-2ai-l,itl; (29)
j=l,2 ,..., i=3,4 ,... (26~) n (ltt): (It:!

ki=ai,l/ai+l,l; i= 1,2, . . . (26d)

Thus the logarithmic matrix function in equation (28b) can
The ki, i = 1, 2, . . . so obtained are kl= 1, k2 = - 3,
be approximated by using a geometric series, that is:
k, = S/4, k4 = - 28/9, etc. The approximate models can be
obtained by keeping the first several dominant matrix
1 1
quotients Ki and discarding others, that is: AC=; R t$R3t . . . +-R”+ -Rn+2
[ n n+2

(30a)
A,-~R[K,]-~=~R

1
(27a) 1
R ni2i

=;R,K, +N[K2]-‘]-1=;R[K2],N+KlKJ1

1
=3 R(R2 - 3Z,J1 tfR3+...+$‘+_~“+Z
n n+2

s ;R [K, +N[Kz + N[K,]-‘]-‘I-’

t3Ob)
=~R[N+K2K1][(K,+KI)N+KlK2K~I-’

1
= CR (R2 - yZ,)($R’ - $Z,)-’ (27~) = R+$R3+...+-R”
n

Appt. Math. Modelling, 1980, Vol 4, December 453

Discretecontinuous model conversion: L. S. Shieh et al.

domains can be effectively used for discrete systems. The

rapid conversion of the mixed models of continuous and
discrete subsystems will greatly simplify design procedures.

1 Acknowledgements
(3Oc)
This work was supported in part by US Army Missile
Research and Development Command, DAAHOl-80-C-0323,
and US Army Research Office, DAAG29-80-K-0077.

References
Harris, E. L. Automatica 1979, 1.5, 91
2 Shieh, L. S. et al. Int. J. Syst. Sci. 1979, 10, 1415
for llR211< 1 +” (30d) Shieh, L. S. et al. ht. J. Control 1978, 28, 383
i n1 4 Shieh, L. S. et al. Comput. Elec. Eng. 1979,6, 3
5 Chen, C. F. and Hsiao, C. H.Znt. J. Syst. Sci. 1975,6, 833
Observe that [In - (l/(1 t 2/n)) R*]-’ is a geometric series. 6 Vidyasagar, M. ‘Nonlinear systems analysis’, Prentice-Hall,
Comparing equations (30a), (30b), and (23) we note that Inc., Englewood Cliffs, USA, 1978, p. 113.
all equations have the same first several dominant terms, Jury, E. I., ‘Theory and application of the Z-transform
method’, John Wiley, New York, 1964
but other terms differ by weighting factors; I/n(l + 2i/n) 8 Thaler, G. J. ‘Design of feedback systems’, Dowden, Hutchinson
in equation (30a), l/n(l + 2/n)’ in equation (30b) and and Ross, Inc., Stroudsburg, USA, 1973
zero in equation (23). The A, for n = 1, 3, and 5 are: 9 Shieh, L. S. and Gaudiano, F. F. IEEE Trans. Circuits Syst.
1975, CAS22, (9), 721
1 10 Jury, E. I. ‘Inners and stability of dynamic systems’, John
___ Wiley, New York, 1974
R3 + (1 + 32 R5 + . . .
I

2
=TR[I,-$R’]-’ asn=l (31a)
Appendix

1
1 1 An illustrative example
-RR f R7t...
3(1 t 3) 3(1 + 5)” Consider a continuous-time state equation:
2 Xo(t) = &X,(t) + BJJ&) (33a)
=TR[I,-~R’][I,,-:R’]-’ asn=3 (3lb)
X,(O) = 0 (33b)
2 1 where :
A,=? R t$R3+iR5+ -RI
L S(1 t 1)
A~=[: _J, BC=[f (‘1 and X(t)=[$j

1
1
+ R9t...
5(1 t $)” (33c)
2 When T = 0.25 s (where T = $< 2/11Acll = 2/7 asi = 1 and
=,R[I,-$R2-$~R4][In-+R2]-1 m = 1 in equation (12j)) and .Ye(t) is a unit-step function,
asn=5 (31~) we have the exact discrete-time model from equation (6):

The corresponding B, can be obtained by substituting the X& + 1) = G&(k) + HdUd(k) (34a)
A, from equation (31) into the following equation: Xd(O) =X,(O) = 0 (34b)
B, z A, [Gd -I,]-’ Hd (32) where X,(t) = X,(kT) as t = kT
Thus, we obtain the equivalent continuous-time models in
equation (21).

and
Gd=
1.45412
0.58381 0.38920
0.48111 1 (34c)

1
Conclusions
Hd =
A geometric-series method has been presented for accurate 0.64872
0.32437 0.05190
0.16865
representation of a continuous-time state equation by an For on-line real-time use, it is desirable to use an approxi-
equivalent discrete-time state equation. This allows the use mate model. If the approximate models Ggl in equation
of well-developed theorems and algorithms in the discrete- (14a) for m = 1,2, and 3 are chosen, then we have:
time domain to solve the continuous-time problems in-
1.4615 0.4102
directly. It has also been shown that if a more accurate
asm=1
model is used, a larger sampling period can be used. Thus,
Gd%GGgl=
[ 0.6154 0.4359 1
the difficulty of selecting a common sampling period
among subsystems of a composite system can be greatly 1.4561 0.3939
LY
- asm=2
reduced. Three methods have been presented for develop- 0.5909 0.4713 I
ing equivalent continuous-time state equations from

1
discrete-time state equations. Therefore, well-developed 1.4550 0.3913
asm=3 (35a)
design methods in the continuous-time and frequency 0.5869 0.4769

454 Appt. Math. Modelling, 1980, Vol 4, December

 Discretecontinuous model conversion: L. S. Shieh et al.

1 [ 1
and 1.01419 1.96861 1.99850 0.00598
A, = 2.95293 -3.90733 B, r 0.99927 0.98132
H(fgHsl= asm=l

1.99991
0.6666
0.3334 0.0513
0.1795 I and

1 [ I
1.0034 1 1.99299 0.00138
-[
- 0.6530
0.3265 0.0518
0.1711 I asm=2 A,= B, r
2.98951 -3.97907 0.99998 0.99581

(37a)
-[ 0.6507
0.3254 0.1697 I
0.0519 asm=3 (35b)
The approximate models in equations (27b), (27c), and

1.999298
If the approximate models Ggz in equation (14b) are used, (27d) via the matrix continued fraction method are :
we have :

I [
1.007460 1.983676 0.003124

1
1.4561 0.3939
Gd=Gg2z asm=l A,? 2.975556 -3.951780 B, = 0.999666 0.990289 I
0.5909 0.4713

I [ I
1.000747 1.998432 1.999963 0.000306

1
1.4546 0.3903
asm=2 A,= 2.997691 -3.995378 B, = 1 .000018 0.999076
0.5856 0.4788
and

I
1.4538 0.3894

I
asm=3 (36a) 1.000061 1.999824 1.999973 0.000029
0.5844 0.4800

W’b)
A,% 2.999780 -3.999546 1.000007 0.999911
and
0.6531 0.0518
Hd=HHgz= asm=I
0.3266 0.1711 I The approximate models in equations (31a), (31 b), and
(31 c) via the geometric series method are:

I [ I
-I 0.3250
0.6498 0.1692 1
0.0519 asm=2 1.00746 1.98368 1.99930 0.003 13
A,% B,Z
2.97556 -3.95178 0.99967 0.99029

I
0.6485 0.0518
-I
-
0.3244 0.1689
asm=3 (36b)
A,=
1.00075
2.99772
1.99844
-3.99538 I [
B,G
2.00672
1.00003
0.00166
0.99908 I
Comparing equations (35), (36), and (34c), the approxima-
tion is satisfactory. Observe that the results in equation (36) and

I [ I
are better than those in equation (35).
1.00009 1.99977 1.999975 0.00004
On the other hand, if Cd and Hd in equation (34) are A,= B,E
given and the equivalent A, and B, of equation (33) are 2.99965 -3.99937 0.99997 0.99987
required, the approximate models in equations (23c), (23d),
(37c)
and (23e) via the direct truncation method are :
Comparing the results in equation (37) with the exact

I [ I
1.04560 1.82744 1.95932 0.02630
A,= B Cz results in equation (33), illustrates that the approximation
2.74120 -3.52304 0.97968 0.90054 is very satisfactory.

Appl. Math. Modelling, 1980, Vol 4, December 455