1963 Corsespondence 263

This derivation is accomplished by as- where X* and 2 represent the complex con- which for all z on the unit circle becomes
suming the control loop open, the input of jugate of x* and uc respectively. Xow it may 1
the sampler to be a unitimpulse,andall be said that (1 1)is satisfied for all values of a‘X(z) - ~ <-
initialconditions to be zero. Thenthe z x* and u* which do not vanish simultane- 2 k
transform’ of equation (1) is obtained. ous]>-. In particular, it is satisfied when or
The inequality of Tsypkin takes the fol-
lowing form. If 1
u* = -
k
x* = X(Z). (12) and
for all values of z on the unit circle in the z Therefore, on the basis of (2) and (12), (11) 1
may be written as Re [ I P ( z ) ] > - -
plane, then the system is a.s.i.1. Here k is a k
numberassociatedwiththesectorwithin
which the nonlinearity must lie. This func- I t h a s therefore been shown that for a
- X(S)‘BX(z)
[SX(i)]’B[zX(z)]
tion is assumed to be restricted as follows: Liapunovfunction of thequadratic-form
type to exist,it is necessary that (3) be satis-
O<-<k
R f

m(0) = 0.
(u)

(4)
+ ~ ’ X ( -Z )k ‘I
- < 0. (13) fied for all z on the unit circle and k <p.
CosCLUsIoNs
Rewriting (13), Lve have
I t has been shown t h a t satisfaction of
RESULTS Tsypkin’s inequality is a necessary- condition
NAIK
(I s 12 - l)X(S)’BX(a) for the existence of a quadratic-form Lia-
Following the procedure of Popox-, con- punov function for the system (1). I t is of
sider the following quadratic form:
= X‘BX, ( 3
+ a’X(s) - -
k ‘I < 0, (14)
theoreticalinterest to know if theinverse
problem alxvays has a s ~ l u t i o n . ~If. ~(3) is
satisfied, is it possible to construct a quad-
where B is a constant x X n symmetric posi- which for all valuesof z on the unit circle in ratic-form Liapunol- function for the system
tive [Link] a trajectory of the system the z plane reduces t o ( l ) ?For a certain class of second-order sys-
( l ) , the T; difference may be written as 1 tems, the answer is affirmative.X procedure
AV = l’[x(k + l)] V [ x ( k ) ]
- Re [ a ‘ X ( z ) ] < -I
k
for constructingan“optimum‘quadratic
form for the general problem can no doubt
= [ G x ( k ) + h n z ( k ) ] ’ B [ G x ( k )f hnz(k)] be obtained, but may be of little practical
or
- x(kj’Bx(k). (6) [Link]’s inequalit!. appearstobe
1 the most powerful approach presently avail-
In order to make the sign of (6) dependent Re [-a‘X(z)]
= Re [Il’(x)] > - -. (15)
able for the class of systems discussed here.
k
upon a gain sector, it is convenient to add For a more restricted class of nonlinear ele-
For the case of a single pole at z = 1 in the ments, a Lur’e-typestabilityfunctionhas
-WZ(k)[U(k) - a’x(k)] = 0
zplane, it is convenient to represent thesys- been recently proposed.5 Preliminary in-
to both sides of ( 6 ) . Then tem (1) by vestigations with this function are quite en-
AI7 = (Cx + hnz)’B(Gx + hnz) x(k + 1) = G*x(kj +- h*tn(k) couraging. Theanalogywithcontinuous-
time systems is not complete a t present, al-
y(k + 1) = y ( k ) + Tnr(k) though future work could possibly lead to a
u(k) = a’x(k) + a,y(k) moregeneralfrequenq-domaininequality
than that of Tsypkin.
I t is now possible to consider ( 7 ) as a quad- d k ) =~ [ ~ ( k j ] , (16) J. B. PEARSON,
JR.
ratic function of the n+l variables x and nz. Control and Information Systems Lab.
I n this manner, it has been demonstrated5 where School of Electrical Engineering
that it is always possible to determinea gain Purdue University
sector, Lafayette, Ind.
n 1 (u)
O<-<q,
U ’. A. Yakubovich,“Thesolution
6 1 of certain
matrixinequalities in automatic control theory.”
Dokl. A k a d . A-auk. L‘SSR, vol.
3, pp. 620-623;
within M-hich ( 7 j remains negative definite, a‘ = (a, a? . . . March, 1962. (English translation.)
pro\-ided the eigenvalues of G are less than 7 E. R. Kalman *LiapunovfunctFns for the
unity in magnitude. Therefore if (7) is to be and G“ is a constant TI - 1X z - 1 matrix, problem of Lur’e in ‘automaticcontrol PFOC. Natl.
Acad. Sci. G S A , pp. 201-205; February. 1963.
negative definite for all m / u satisfying (8), whose eigenvaluesare less thanunity in
it must be negative definite for magnitude. In this case
rn = ku, (9) K ( a ) = - a’X(z) - &I-(z)
where 0 < k <q. Then combining ( 7)eand (9)
AI,’ = ( G x +
hku)’B(Gx hku) +
+
- X’BX kua’x - uku < 0. (10) where Increasing the Sampling Efficiency
If in place of the real variables x and u, X ( Z ) ( d - G*)-’h*. for a Control System*
complex variables x* and u* are substituted, Consider the 1,’ function to be Control systems where data is sampled
then(10) becomes equivalent t o are encountered in practice quite frequently.
Re [(G* +
hktP)’B(Gx* hku”) - F*’Bx* + 1: = X’B * x - - yan2 (a, < 0). (18) The aim always is to construct the curves of
2T output or error as close to the continuous
+
k 8 a ‘ x * - a*ku“] < 0, (11) output or error,respectively, as possible.
The analogous equation to (14) is The usual sampling techniques which have
4 J . T; Tou. ‘Digital and Sampled-Data Control been used, use fixed period sampling. Vary-
Systems,McGrau-Hill
S . U.:3959.
Book Co.. Inc.,h-ew Tork. 2 - + a‘~(;)
I)x(z)’B*x(z) ing s1mpling rate techniques have been used
5 J . B. Pearson and J. E. Gibson. “On the asymp-
totic stability of a class o i saturating sampled-data
systems,” Trans. A I E E (Applications & Industry). * Received February 19, 1963; revised manuscript
(To be published.) received March 4, 1963.
 264 T R A N S AICETEIEO N S ON A
CUOTNOTM
ROAT
LIC July

t o a certaindegreebyHufnagel,',' this gives can assume that the samplingperiod should

and Bergen' in the study of sampled data - always be less than about 10 per cent to 25
control systems. S o n e of these have consid- T = --2.303(i,./en) for minimum y2. (5) per cent of the time constant of the linear
eredthe idea of increase in sampling effi- - system. Let
~2 vs T,, is plotted in Fig. 3.
ciency. Dorf, Farren, and Phillip? are the T p5 ( 0 . 1 ~- 0.257), (6)
first to consider such a n idea and they have The mean-square value of difference be-
come out with an experimental study show- tween actual error curve and sampled error 7 being the predominant time constant of the
ing that there is increase in sampling effi- curveis minimized over each
sampling system, so that one has
ciency if the sampling period is a function of period for T,= --2.303(Cn/?,). Thisonly
has meaning if P, and e,, are of opposite sign T, 5 Tz. ( 2
error and error [Link] theoretical justilica-
tion except from the point of view of stabil- because then only T , is positive. In a given -4. further consideration has to he made in
ity has been given. In this communicationa situation 8,, and 8. arenot necessarily of case &-to, because then T',-tO Ivhich might
theoretical technique for increasing the effi- opposite sign, and therefore the value of T , lead alsoto difficulties because any sampling
cienq- of sampling is given6 withcertain is no longer valid. To avoid this difficult): device might only be able to have a mini-
practicallimitations. I t shouldbeempha- it is proposed thatduringthesampling mum sampling period TI, so that one must
sized that the goal of this technique is the period when e, and F,, are of the same sign, have
same as that suggested by Dorf, Farren, and then the value of T,, should be taken to be
Phillips,j but while they have justified their -
0.5(P,,/Enjasshon-ninFig. 3. Thiscorresponds T n L TI. (8)
choice of sampling purely on experimental to the same ~ ~ i .which2 is obtained when 8, Hence the criterion for efficient sampling is
grounds, this technique suggests a theoreti- and 2, are of opposite sign. This seems to be that each sampling period is given by
cal basisfor the choice of thesampling quite logical.
Furthermore, one can encounter an addi-
T , = - 2.3030, Q < 0
period making it a function of error rate and
second derivative of errorratherthanof tional difficulty if ?,-O, xvhich can lead t o = 0.5, a 2 0, (9)
error rate alone. X-ery erroneousresults. To avoidthisone where
Consider the system shown in Fig. 1. If
the error is sampled then one xvould like to a=-
en
have actual error curve as close to sampled .?"
error curve over each period for maximum
efficiency.
Consider Fig. 2. To achievethis,one
I SAYPLEX
defines
AN3 i3L3
I
Consider Fig. 1 again. One notices t h a t
31 = e - e,, (1) if T,, is variable over each period, so is the
- Fig. I-General control system. stability, because stability is very much a
and then T? is minimized ox-er each period. function of T,,. In order t o avoid instability
This is expressed as in the system a gain element, K , is inserted
in the system and is varied over each period
as T,, changes to insure stability.6 K is obvi-
ously a function of Tn and time constantsof
the sJ-stem. A block diagram t o achieve the
Since the sampling period is much less than above criterion is shown in Fig. 4.
the time constants of linear system, we can A detailed digital computer study for a
writeovereachsampling period second-ordersystem for deterministicand
probabilistic inputs has been made and will
tt I
e = e, + +P,f .'
EnZ. (3)
1
0
1
1
1
2
1
3 n-: n TIME
t
t be presented shortly.' Some other models,
including minimization of error by using this
Fig. ?-.Actual error curve and sampled adaptivetechnique of samplingandgain
Eq. ( 3 ) is anexpansion of e in form of error curve. control, are also under study.
Taylor series at Tnneglecting higher derivi- S. C. GUPTA
tives of e a t e , than second. This is permissa- Electrical Engineering Dept.
ble because of small sampling period com- .Arizona State Cniversity
pared to the time constant of the system. Tempe, Ariz.
This leads to
7 s. C . Gupta, 'TheoreticalandExperimental
Study for Increasing the Sampling Efficiency of Con-
trol Systems,' to be published.

If we now make

I R. E. Hufnagel. "Analysis of cyclic rate sampled A General Classification of

data feedback systems.' Tram. A I E E (Applicalions
and I n d u s t r y ) , vel. 77, pp; 421-425; I ovember. 1958. Amplitude-ModulatedCarrier
2 R. E. Hutnagel. Analysls ot Aperiodically
Sampled Data Feedback System," Ph.D. dissertation. Control Systems*
Cornell University, Ithaca. K.Y.: June. 1959.
a E. I . J u r y and F. J. [Link] of ISTRODL-CTIOX
sampled-data control systems with periodically time-
varying sampling rate." I R E [Link] ACTOJMTIC There are several types of control sys-
COSTROL. vol. BC-4. pp. 15-21; M a y , 1959.
4 A . [Link]. "On $atistical desi,w of linear tems whichin\-olve amplitude-modulation
random sampling systems. Proc. IFAC, Butrenvorth processes. Theseinclude single-sinusoid
[Link]. 126-
131. 1960. carrier-signal
systems,
sampled-data sys-
R. C . Dori. XI. C: Farren. and C. X. Phillips. tems, finite-pulse-width systems, and modu-
"Adaptive sampling
trequencyfor
sampled-data
control systems,' I R E T R ~ S S .os ALTOXLATICCos- latedsystems whichuse square-wave car-
TROL. vel. .\c-7, pp. 38-47: January. 1962.
6 S. [Link],'AdaptiveGainand-Adaptive a=(:), +. ='#me p n s ' a r l s of GIs! riersignals. Althoughall of these systems
Sampling Sampled Data Control Systems. presented
a t IEEE Winter General Meeting, Kew T o r k , K. X-.;
Zi-February
January 1, 1963. - --- 4-\
Fin. - model iar efficient samoling. * Received February 7. 1963.