On the Stability of Nonlinear

Sampled-Data Feedback Systems

b y CHI-TSONG CHEN
Electronics Research Laboratory
University of California, Berkeley, California

SUMMARY:7'his work is a generalization of Tsypkin's stability criterion for a class of

time-varying nonlinear sampled-data feedback systems. Some sufficient conditions for
the response to any bounded input sequence to be bounded are preserded. No assumptions
are made concerning the internal dynamics of the linear subsystem, except that its input-
output relation is of the form of convolution. The essence of the proof is to consider the
nonlinear system as a perturbation of a stable linear sys~m.

Description of the S y s t e m
The system considered is a sampled-data feedback system of the type
shown in Fig. 1. It is assumed that its linear and nonlinear parts, G and N,
are separated by an ideal sampler (6). The sampling period, T, is assumed,
without loss of generality, to be 1. Since the theory of distributions is well-
established, there is no difficulty in interpreting expressions involving Dirac-~
functions. Thus, if the input to the sampler is x(t), its output will be
x*(t) = ~ x(n)~(t -- n). For brevity, {x(n)} will be used to denote the se-
nT0
quence {x (0), x (1), . . . , x (n), . . - }. The behavior of the sampled-data system
is completely characterized by considering it only at the sampling instants.
This is usually enough in practical applications, because the hidden oscillation
can be easily predicted and avoided by changing the sampling period. There-
fore, we shall consider only the performance at the sampling instants.
N is a memoryless, time-varying nonlinear element, continuous with
respect to its two arguments; furthermore there are two constants f~l and f~2
such that ~ > [/~1[ and

(N.1) f~l < ~(a, n) < f12, for a ~ 0

O"

~(0, n) = 0, for all n (the set of natural numbers).

Fie.. I. Time-varying nonlinear sampled-data Fic~. 2. Linearized sampled-data

feedback system. feedback system.

316
 Stability of Nonlinear Sampled-Data Feedback Systems

G is a nonanticipative, linear, time-invariant subsystem; it usually consists

of a zero-order hold (7) and a linear plant. Let g(t) be its unit impulse response,
g*(t) its sampled impulse response; then g*(t) = ~ g(n)5(t - n). 1 Thus the
n =0

input-output relation of G is given b y

(G.1) y(n) = z(n) ~- ~ g(n - m)x(m)

m~O

where Ix(n)} is the input sequence, {y(n)} the output sequence, {z(n)} the
sampled zero-input response.
It is assumed that {z(n)} satisfies the following condition:

(G.2) for all initial states, {z(n)}d 2, i.e., ~ ]z(n) l 2 < ~ .

n~0

Note that z(t) ei2(o,~) does not imply {z(n) }el2. However, we have {z(n) }el2
if z (t) eL2(0,~ )t3 L ~ (0,~) and is piecewise c ontinu ous.

Definition

a(z) ~ ~[g(n)] = ~ g(n)z_,,

n~0

H(z) A ~Fh(n)J A $G (z)

1 + ~G(z)
(~, n) ~ ~ (~, n) - ~

Ilhll, = nE~ O Ih(~)l

][zllp ~ (n~~ 0 Iz(nDIP) 'zp, 1 __<p< ~,

I[zll~ & suplz(n) I.

n

Sufficient Conditions f o r S t a b i l i t y
Recently, Tsypkin (1, 2) applied the Popov method to establish the suffi-
cient conditions for the stability of the system shown in Fig. 1. Jury and Lee (3)
extended his result to obtain less conservative sufficient conditions by adding
restrictions on the slope of the nonlinearity. Although not stated explicitly,
Tsypkin considered only the linear subsystem with rational transfer function.
In this paper we remove this restriction and present a different proof which is
analogous to that of Sandberg (4).
1 If g(t) has discontinuities, then the presence of the zero-order hold implies t h a t in the
definition of 9", g(n) should be replaced by g(n~-) whenever g is discontinuous at n.

Vol. 280, No. 4, October 1965 317

Chi-Tsong Chen

Theorem I
In the sampled-data system shown in Fig. 1, N is a memoryless, time-vary-
ing nonlinear element satisfying (N.1), G is a nonanticipative linear time-
invariant subsystem satisfying (G.1)and (G.2). If
(i) g(n) = r + gl(n) for n = 0, 1, 2, . . . ,
where r is a nonnegative constant and {g1(n)}d ~, i.e., ~ {gl(n) I < o¢
n =0

(ii) inf ]1 + f~G(z){ > 0, for /3~ < ~ < /3..,

}~I>_t

(iii) .y A /32 - ~1 max G(e~)

2 -~<-_<- l+~G(e,~ ) <1

then for any input {u(n)}d:, and for any initial state, the output sequence
{y(n)} is an element of l 2, hence is a bounded sequence which tends to zero as
n --~ ~ . A bound on the norm of {y(n) } is given in Eq. 10.

Remarks
If r = 0, then this theorem still holds if we relax /32 >_ {/31[ to /32 > 0,
~2 >_/31.2
In general, G(z) = ~ g(n)z -~ with {g(n) }d 1 cannot be put in a closed form,
n =0
neither is it a rational function of z. Although G(z) is analytic for {z{ > 1,
nothing can be said about the singularities inside the unit circle. Also, the
singularities need not be isolated, therefore the residue theorem is generally
not applicable.
It is interesting to note, if (ii) is not satisfied, a nonanticipative linear part
G m a y result in an anticipative closed loop system. For example, let the non-
linear part N be linear, let the slope of its characteristic f] be 1, and let G(z)
= - 1 + z-l; then the closed loop transfer function is H(z) = -- z + 1, i.e.,
H (z) is anticipative.
If the system is linear, then ~ = 0 and the condition (iii) drops out; condi-
tions (i) and (ii) are then the necessary and sufficient conditions for the linear
system to be stable.

Proof of T h e o r e m I
Consider the linearized sampled-data feedback system shown in Fig. 2;
let H (z) be the transfer function of the linearized feedback system, then

tirz ~ ( z - 1)
+ G1(z)
H(z) = ~V(z) = (1 + ~r)z -- 1 ( 1 % ~r)z -- 1
l + $G(z) 1+ ~(z-1) G,(z)
(1 + ~r)z - 1
After completion of this work, Dr. Sandberg advised the author t h a t the theorem for this
case is given in (12).

318 Journal of The Franklin Institute

 Stability of Nonlinear Sampled-Data Feedb~k Systems

Since ~2 >_ lfll[ implies ~r >__ 0, therefore~(z 1) is the z-transform

- -

(1 + f3r)z -- 1
of an l' sequence. Because the convolution of two 11 sequences is an l' sequence,
it follows t h a t G2(z) A t~(z -- 1) Gl(z) is the z-transform of an l 1sequence.
(1 + ~r)z - 1
Since the z-transform of an 11 sequence is analytic for t z[ :> 1 and continuous
for [z[ = 1, so is 1 + G2(z). Condition (ii) implies that 1 + G2(z) is bounded
away from zero for [z[ >_ 1, therefore (1 ~- G._,(z))-~ is analytic for tzl > 1
and continuous for Iz[ = 1. It follows that (1 -F G2(z)) -1 has a Laurent expan-
sion of the form ~ b(n)z -'~ valid for Iz] > 1. Wiener's result (5) implies
n =0

that { b ( n ) } d 1, i.e., (1 ~- G~(z)) -I is the z-transform of an l' sequence. Since

~rz
(1 -t- f3r)z - 1 + G2(z) is the z-transform of an 11 sequence that is the sum of
two l' sequences, therefore H(z) is the z-transform of the convolution of two
l' sequences. Again by the fact t h a t the convolution of two l ~ sequences is
an l 1 sequence, we conclude that { h ( n ) } d 1.
Let ~(n) be the output of the linearized system; then

~(n) = z(~) + ~ h(~ - m)Eu(m) - z(m)J. (~)

m~0

Since {h (n) }d 1, if {z (n) }d ~ and {u (n) }el~ then {/~(n) }d -~.

Introducing ~(n) and ~, we see t h a t (G.1) can be written as
1
y(n) = ~(n) + ~,,~o h ( n - m)(oEu(m ) - y(m), m~. (2)

Let z(n) ~= u(n) - y(n), #(n) A u(n) - ~(n), then Eq. 2 becomes
1
a(n) = d(n) -- ~ m-o
~ h(n -- m)(oVz(m), m-1. (3)

Given the sequence {d (n) }, {z (n) } m a y be calculated through the recursivc

Eq. 3. The more interesting interpretation of Eq. 3 is to consider it as the

equation of a feedback system shown in Fig. 3: l i b ( n ) } istheimpulsere -

sponse of the linear subsystem, {d(n)} is its zero-input response and {z(n)}
is the output of the entire system.

A
H o-

FIG. 3. System interpretation of Eq. 3.

Vol. 280, No. 4, October 1965 319

Chi-Tsong Chert
We show that if condition (ii) is satisfied, the sequence {a 0i')} is such that
¢ (n) is finite for all n. The only possibility for ¢ (n) to be infinite for some n is
f~G(z) &(o)
that -- h (0) ~r¢ (n), n~ = ¢ (n). Since h (0) = ~-~limH (z) = lim~_+.1 q-fiG (z) - 1 +3g (0)'

it is equivalent to g(0) ~[a(n), n~ = - z(n), or g(0) = --1

1 + t~g(0) ~ + ~l-o (n), hi"
~(n)
--1 --1
Since ~[¢(n),¢(n) n~ ~ 32--2 3~, it implies ~_< g(0) _< ~ - for 3~ > 0;
--i 1 --I
g(0) _<~-2 for3~ = 0 ; a n d g ( 0 ) > _ ~ , g ( 0 ) _ < - ~ 2 for3~ < 0 . But these cases
are ruled out by condition (ii) when z --~ ~ (lim I 1 + 3G(z) I = ]1 + 3g(O)] > 0
z---~ :¢
for 31 _< ~ _< fl~), therefore I¢(n) I < ~ for all finite n.
Define {aN(n)} as the sequence {¢(n)} truncated after its N th term, and
Y~N(Z) its corresponding z-transform, i.e.,
aN (n) = ¢ (n) if n < N
=0 if n>N
and

Zu(z) ~ ~ ( n ) z - " = E ~(n)z-".

n~O n~O

Similarly for d~(n) and ~ ( z ) . Multiplying z-~ and taking the summation
from 0 to some fixed N, Eq. 3 becomes

n=O
a(n)z-"= n--O
~ d(n)z-" -- a1 [ ~o
= ( ~"o h ( n - m)(o(aN(m),m))z-"

or

Y~N(Z) = ~N(Z) -- ~ H(Z)~N(Z) + V(z) (4)

where

~ ( z ) = ~ ~(~(~), n)z_.
nz0

V(z) = ~1 ~ ( 2 h(n -- m)~(~(m), m))z_..

n=N+l m=0

Then V (z) is the z-transform of the zero-state response Estarting from (N + 1)th
sampling instantJ as a result of { ~ (ag (m), m) } being applied to a system whose
impulse response is •{1~ h(n) •I Since {aN(n)} is of finite length, Z~(z), ~(z),
and ~N(z) are polynomials, hence are all convergent for [z I > 1. V(z) is also

320 Journal of The Franklin Institute

 Stability of Nonlinear Sampled-DataFeedbackSystems

convergent for [z I >__ 1, since it is the z-transform of the convolution of two l ~

sequences. It is convenient here to change the variable z to e ~~ and apply
Minkowski's inequality to Eq. 4; t h e n

(// , ] z ~ ( e ~) - V(e':'gl'-'do~) <_ (L [2N(ei~')["dw

+ . (5)

Using Parseval's equality, the fact t h a t XN(Z) and V(z) have no power of
z-1 in common, condition (iii), and the fact t h a t I¢(~, n) < ~/~ - /31 l ol , we
obtain

(n~O
~ 1,~0,)1~) "~= (~n=O I,~N(~)I~) "~ ~ (~n=O Io'~(~*)1= -4- n~O
~ I~(~)1~) '~

= (~.
,,=o
IO-N(n) -- v(n) l~) '/~ --< ,8~ 2fi-- ~1 max,~ IH(e"~)l (~.,=o [°~(~)1~)'/~

N N
-t- ( ~ I,~(n)l~) "~ ~ ~' ( E lo(n)t~) "~ + ('~. I,~(~)1~)'~2. (6)
n~O n~O n=O

Since ~ (n) = u (n) - ~ (n), therefore {u (n) }el2 and {~ (n) }el'-'imply {~ (n) }el2.
thus

( E2v Io(n)l~) "~ < 1 ( E~r I~-(n) l~) "~ < 1 ( E I,~(n) i .~)~/,, < oo (7)

i.e., {~(n)}d 2. B u t y(n)= u ( n ) - ~(n), it follows t h a t {y(n)}d2; hence

{y(n)} is a b o u n d e d sequence which tends to zero as n --~ ~ . Q.E.D.
I n case G(z) is a rational function of z, the condition (i) in the t h e o r e m
can be relaxed.

Corollary I
In the theorem, we replace (i) b y (i');

(i') G(z) is a rational function of z

t h e n the same conclusion holds.

Proof

(i') and (ii) imply t h a t the rational function H (z) - 1 +t~G~G

(z) (z) is analytic

for ] z I > a where a is a n u m b e r < 1 ; t h e n {h (n) }d 1, as can easily be seen from

the residue theorem. T h e rest of the proof is the same as in the theorem. Q.E.D.

Vol. 280, No. 4, October 1965 321

Chi-Tsong Chen

Remarks
It is easy to see that if G(z) is a rational function of z, condition (ii) is
equivalent to the Nyquist criterion. More precisely, if G(z) has q poles
outside the unit circle, then 1-t-~G(z) ~ 0 , f~_<~_<~2 for Izf > 1, if
and only if the Nyquist plot of G(z) (8) does not go through the critical in-
terval [ ( - - ~1, 0), ( - - ~,1 0). ~] and encircles it q times in a clockwise direction.
Let G(z) in Fig. 1 be a rational function of z with q poles outside the unit
circle, and let t32 > t31 > 0, then according to Corollary I: if the Nyquist plot
of G(z) (8) lies outside the disk with center at (-½(~1 -~ + ¢~-1), 0) and
radius ½(~f~ - ¢~2-~), and encircles the disk q times in a clockwise direction,
then the system is stable, i.e., the output sequence is bounded and tends to
zero as n--* ~ .
The boundedness of Jlyll2 can be given in terms of the norms of the input
sequence and the zero-input sampled response. From Eq. 7, we obtain
1
I1~11= -< ]---__, II~ll~. (8)

Taking the z-transform of Eq. 1 and using Minkowski's inequality and

Parseval's equality, we obtain

I[~{[~-< llzll~+ max IH(e'~)l(lIull~-4- [[zl[~). (9)

Since [[o-I1,, _> Ilyl[~ - [lull:,, II~.lr~,<. ilu]l~ + little, and max IS(e':")[
- - T < to< ,s-

< -/~2-~-/~1' Eq. 8 becomes

2~ -1
rlyJJ~ - Ff~rr~ < ~ ll~ll~ + rrzrr2 + - - (llull~+ ll~ll=)l
J
or

2~2 (llul]2 Jr ][zl12) -t- J]u[12. (10)

Theorem II
In the sampled-data system shown in Fig. 1, N is a memoryless, time-vary-
ing nonlinear element satisfying (N.1), G is a nonanticipative linear time-
invariant subsystem satisfying (G.1). If for all initial states, {z(n)}d ~, i.e.,
Ilz[[® < ~ , and if
(i) g(n) = r + g1(n) for n = 0, 1, 2, . . . ,
where r is a nonnegative constant and {gl(n)}el ~.
(ii) inf I1 + f~G(z) l > 0, for /31 _< 13 < 132
lzl>l
A ~ 2 - - ~1
(iii) ~ = l[h[[1 < 1
2t~

322 Journal of The Franklin Institute

 Stability of Nonlinear Sampled-Data Feedback Systems

then for any {u(n)}d ~, and any initial state, the output sequence {y(n)~ is
an element of l ~ (the response to any bounded input sequence is bounded).

Remarks
This theorem is the discrete analog to the one in (9).
Except for some special cases, it is not easy to obtain the frequency do-
main interpretation for condition (iii).
This theorem may be stated more generally, i.e., if {z(n)Jd ~', 1 <_ p <_ ~ ,
for all initial states, and if (i) (ii) (iii) hold, then for any {u (n)}d v, and any
initial state, the output {y(n) }dp.

Proof of T h e o r e m H
Since condition (ii) implies that {h(n)} is nonanticipative, and since ~ is
memoryless, from Eq. 3, we have

1 ~ h(n-m)(o[a~(m),m]
a ~ ( n ) = ~ ¢ ( n ) -- ~,.~o n<N. (11)

S i n c e {h (n) }d ~, {u (n) }d = and {z (n) }d =, it follows that {d (n) }d% We have

shown that a(n) < oo for all finite n, therefore we m a y take the norm and
apply Minkowski's inequality to Eq. 11,

1
(12)
Then, from

llh*~,vl]® = sup 1 ~ h(n -- m)(o(aN(m), m)[

n m--O

_< sup l- ~ ]h (n -- m) ] ] ~ (aN (m), m) ]

n m~o

< /33 -- Ol sup /- ~ I h ( n - - m ) tlaN(m)]

2 n m--0

< ~ - 81 II~ll®llhlll (13)

- 2
and condition (iii), we obtain
1 1

Since Hail® > ]]ylJ~ -- Ilu[[~, I[~Jl~ ~ Ilul[~ ÷ IFflf~, and from Eq. 1
[[0[[~ < [[zll~ + Ilhfllllull~ ÷ Ilhl[,llzlr~, it follows that

1 (1 + Ifhlll)(llul[~ + I]z]]~) + [rull~,

[[urr~ -< 1 -.
< 2~ Q.E.D.
(1 -- a)(/~2 --/~x) (llur[~ + IJ~ll~) + llull~,

Vol. 280, No. 4, October 1965 323

Chi-Tsong Chen

Corollary H
In Theorem II, we replace (i) by (i')

(i') G(z) is ,~ rational function of z

then the same conclusion holds.

Conclusions
Stability criteria for a class of time-varying nonlinear sampled-data sys-
tems are obtained. In the l 2 case, because of Parseval's equality, we are able
to obtain the stability criterion in the frequency domain. The impulse transfer
function of the linear subsystem is not assumed to be a rational function;
therefore these criteria are applicable to a very broad class of systems.
In the 1~ case, if G (z) is a rational function of z, the stability region in the
G(z)-plane is the same as in (2), although the analytic form is different. In
this paper we use the linearized closed-loop sampled impulse response and
some well-known facts about linear systems. Therefore, we can apply Minkow-
ski's inequality directly to the z-transform of the convolution equation which
characterizes the system.

Acknowledgments
The author wishes to express his gratitude to Professor C. A. Desoer for his guidance and
encouragement. Thanks are also due to Professor T. Kato for his helpful discussion.
This work was supported by the Joint Services Electronics Program (U. S. Army, U. S.
Navy and U. S. Air Force) under Grant, AF-AFOSR-139-64.

References
(1) Y. Z. Tsypkin, "On the Stability in the Large of Nonlinear Sampled-Data Systems,"
Dokl. Aka~l. Nauk., Vol. 145, pp. 52-55, July, 1962.
(2) Y. Z. Tsypkin, "Frequency Criteria for the Absolute Stability ~of Nonlinear Sampled-
Data Systems," Avtomatika i Telemekhanika, Vol. 25, No. 3, pp. 281-289, March, 1964.
(3) E. L Jury and B.W. Lee, "On the Stability of a Certain Class of Nonlinear Sam-
pled Data Systems," I E E E Trans. on Automatic Control, Vol. AC-9, pp. 51-61, Jan., 1964.
(4) I. W. Sandberg, "Frequency-Domain Criteria for the Stability of Nonlinear Feedback
Systems," Proc. 1964 National Electronics Conference, pp. 737-741.
(5) N. Wiener, "The Fourier Integral and Certain of its Application," Mass., Cambridge
University Press, p. 91, 1933.
(0) E. I. Jury, "Sampled-Data Control Systems," New York, John Wiley~and Sons, Inc.,
p. 6, 1958.
(7) E. I. Jury, op. cit. (p. 48).
(8) E. I. Jury, op. cit. (p. 44, Fig. 1.33).
(9) I. W. Sandberg, "A Condition for the £~-Stability of Feedback Systems Containing a
Single Time-Varying Nonlinear Element," Bell Syst. Tech. J., Vol. 43, pp. 1815-1817;
July, 1964.
(10) C. A. Desoer, "A Generalization of the Popov Criterion," I E E E Trans. on Automatic
Control, Vol. AC-10, No. 2, pp. 182-185, April, 1965.
(11) C. A. Desoer, "On the General Formulation of the Nyquist Criterion," I E E E Trans. on
Circuit Theory, Vol. CT-12; June, 1965.
(12) I. W. Sandberg, "On the Boundedness of Solutions of Nonlinear Integral Equations,"
Bell Syst. Tech. J., Vol. 44, No. 3, pp. 439-453, March, 1965.

324 Journal of The FranklinInstitute