708 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, OCTOBER 1976

For our purposes, a passive system can be defined as a system which

always dissipates energy, provided the energy input to the system is such
that the product of system input and output represents input power.
(This makes sense physically if the system is an electrical network [SI; in
other cases, there mightbe no simplephysical interpretation.) Several
useful properties of passive systems have been noted in [2] for the linear
case and in [l], [3] for nonlinear systems.
The notion of passivity was extended by Willems [4] to allow a more
general definition of input power. A "dissipative system" is defined in [4]
to be one for which a supply rate (input power) and storage function
which contradicts the requirement that ek be bounded. It is not difficult (stored energy) can be found, with the property that (in a sense made
to show that there exist no i> cop, and P that ensure stability for all more precise in [4])energy is always dissipated. As one might expect,
allowable parameter variations. Q.E.D. there are some useful connections between dissipativeness (in Willems'
sense)' and Lyapunov stability.
Thepresentshortpaperextendsthe passivityresults of [I] using
REFERENCES concepts very similar to those of Willems [4]. Our approach differs from
that of Willemsprimarily because we treat bsipativeness as an in-
1. M. Horowiu, Synrhesis of Feedbuck Sysrems. New York: Academic, 1963.
J. A. Roluik and I. M. Horowitz, "Feedback control system synthesis for plants put-output property; that is,we do not postulatethe existence of an
with large parameter variations,- I E E E T r m . Auromr. Conrr. (Short Papers), vol. internal storage function. However, we deduce the existence of such a
AC-l4,-pp. 714-718, Dec. 1969.
D. D. Siljak, Nonlinear Systems. New York: Wiley,1969. storage function (in general nonunique), so it could be argued that this
D. M. Salmon, "Minimax controllerdesign," I E E E Trm. Auromor. Conrr., vol. difference is unimportant. A morenoticeabledifferencebetween OUT
AC-13. pp. 369-376, Aug. 1968. results and those of [4] is that, by sacrificing some generality, we obtain
H. P.Horisberger,"Controlof linear systems withlargeparametervariations."
Ph.D. dissertation. McGill Univ., Montreal, P.Q., Canada. Aug. 1973. results which are considerably more explicit. The central result of this
E. W. Cheney and A. A. Goldstein, "Newton's method of convex progamming and short paper is an algebraic criterion, in terms of functions of the system
Chebyscheff approximation," Numer. " a h . . pp. 253-259. 1959.
E. Pol&, Compurnrional Merhods in Oprimizarion. New York Academic, 1971. state, for the input-output property of dissipativeness. This result is then
T. L. Johnson and M. Athans "On the design of optimal constrained dynamic used to derive stability criteria.
compensators for linear constant systems,'' IEEE T r m . Auromr. Conrr. (Short
Papers), vol. AC-15, pp. 658-660, Dec. 1970. The structure of the short paper is as follows. In Section I1 we derive
H. P.Horisberger and P. R. Bilanger, "Solution of the optimal constant output the algebraic criterionfor dissipativeness,which leads to computable
feedbackproblembyconjugategradients," I E E E T r m . Auromor. Conrr. (Tech. storagefunctions.Thesefunctionshave the properties of Lyapunov
Notes and Corresp.), vol. AC-19, pp. 434435. Aug. 1974.
J. M. Danskm. The %oy of Max-Min. New York: Springer,1967. functions and in Section I11 are used to get s t a b ~ t yresults. In Section
IV, the special case of passive systems is considered.

11. ALGEBRAIC CRITERION

FOR DISSIPATIVENESS
The Stabilityof Nonlinear DissipativeSystems
The systems to be studied are described by the equations
DAVID HILL, MEMBER, IEEE,AVD
PETER MOYLAN, MEMBER,IEEE i=f(x)+G(x)u
y=h(x)+J(x)u. (1)
Abstract--This short paper presents technique
a for
generating
Lyapunov functions for a broad class of nonlinear systems represented by The values of x, u, and y lie in R n , R m , and RP, respectively. The
state eqnations. The system, for which a Lyapunov function is required, is
admissable controlsare taken tobe locally square integrable. T h e
assumed to have a property called dissipativeness. Roughly speaking, this
functionsf:R"-tR", G : R " + R n X m ,h:R"-tRP of the state vector x
means that the system absorbsmore energy from the external world than it satisfy f(0) = 0,h(0) =0, and the following assumption.
supplies. Differenttypes of dissipativeness can be considered dependingon
Assumption I : The functions appearing in ( I ) have sufficient smooth-
how one chooses to define "power input." Dissipativeness is shorn to be
ness to make the systemwell defined; that is,for any x ( t J E R" and
characterized by the existence of acomputablefunctionwhich can be
admissible u ( .), there exists a unique solution on [to. 00)such that y ( .) is
interpreted as the "stored energy"of lbe system. Under certain conditions,
locally square integrable.
this energy function is a Lyapunov function which establishes stability, and
Theaboveformulation includes abroad class of systems. In fact,
in some cases asymptotic stability, of the isolated system.
Balakrishnan [7] has claimed that, under quite unrestrictive conditions,
any time-invariant finite dimensional system can be represented by ( I )
I. INTRODUCTION with an appropriate choice of the state vector. However, it is not known
In the study of a physical system, such as an electrical network or a whether the transformation in [7] wil preserve our Assumption 1 and
mechanicalmachine,theconcept of stored energy is often useful in finite-dimensionality of the state space. Rather than explore these diffi-
deducing the behavior of the system. In many control problems, how- culties, we will adopt the attitudethattherepresentation ( I ) is
ever, one is dealing with an abstract mathematical model where it may sufficiently general to warrant study in its own right.
be difficult or even impossible to find some property of the model which Motivated by Wdlems [4], we associate with ( I ) a sup& rate
corresponds to physical energy. In this short paper we show, for a certain
class of nonlinear systems, that an ''energy" approach can still be useful w(u,y)=y'9+2y'Su+u'Ru (2)
in stability analysis, despite the fact that the "energy" might not have
any physical meaning. In effect, OUT technique is a method for generat- where Q E R P X p , S E R P X m ,and R E R m X " are constant matrices with
ing Lyapunov functions. Q and R symmetric. The supply rate is an abstraction of the concept of
The theorydescribed herehas its origins in workbyMoylan and input power. In physicalsystems, input power is associatedwith the
Anderson [ 1]-[3] and Willems [4] on the properties of passive systems. concept of stored energy. For moregeneral abstract systems,such as
given by (I), physical reasoning fails us, but we can define a possible
Manuscript received August 19, 1975; revised May 18, 1976. Paper recommended by G.
L. Blankenship, Chairman of the IEEE S C S Stability. Nonlinear, and Distributed
Systems Committee. This work was supported by the Australian Research Grants Com- 'The word "dissipative" is used in an entirely different sense in some of the literature
mlttee. on differential equations (see, for example, [6D. Despite the coincidence in terminology,
The authors are with the Department of Electrical Engineering. University of Newas- there IS no connection between the two Qpes of dissipativenessdefined in [4] and [6]. Our
tle, Newcastle. N.S.W., Australia. definition is essentially equivalent to that of Willems [4].
SHORT PAPERS 709

candidateforthename stored energy.Consider thefunctionand tl > to, andany x(ro), straightforward use of (5) gives

subject to (1) and x(O)= xW This is called the available storage in [4]; it + i O t 1 [ I ( x ) +W ( x ) u l [ I ( x )+ W ( x ) u l d t . (6)
can be interpreted as the maximum amount of energy which may be
extracted from the system (1). Note that @=(x)> 0 for all x .
Setting x(ta)= 0, we have condition (4).
In the sequel, we shall impose the following assumptions on the system
For necessity, we proceed to show that +,(-) given by (3) is a solution
(1) and the supply rate (2).
of (5) for some appropriate functionsI ( . ) and W ( . ) .
Assumprion 2: The state space of the system (1) is reachable from the
For any state x. at r=O, there exists by Assumption 2 a time r - < O
origin. More precisely, given any x , and I , , there exists a to< t , and an
and an admissible control u ( . ) defined on [t-l,O] such that x(t-,)=O
admissible control u( .) such that the state can bedriven from x(ta)=O to
and x(0)=xW From (4), then,
x(tl)=xl.
Assumption 3: The available storage +,(x), when it exists, is a differen-
tiable function of x .
Assumption 4: For any y # 0 there exists some u such that the supply
rate (2) satisfies w(u,y)< 0. The right-hand side of this inequality depends only on x @ whereas u ( . )
Assumptions 1-3 are not overly restrictive; in fact, for linear systems, canbechosenarbitrarily on [O,T]. Hence, there exists afunction .
Assumptions 1 and 3 are trivially satisfied. Assumption4 is simply a C :R + R of x such that
restriction on the class of matrices Q, S, and R that will be considered. It
is important to notice that u and y are considered to be independent
(7)
variables-that is, they are not necessarily related by the state equations
(])-for the purposes of Assumption 4.
We now consider the concept of dissipativeness, which can be inter- whenever x ( 0 ) = x W From (9, we have @,(x)< co for all x . Also, dis-
pretedas saying thatthe initially unexcited system can only absorb sipativeness implies that $1~(0)
= 0.

energy. In this connection, it is worth noting that Assumption4 ensures Now it is shown in [4]that @a satisfies
that the definition does not collapse to triviality.
Definition I: The system (1) with supply rate (2) is said to be dissipa-
tive if for all admissible u ( . ) and all t l > t2, we have
for all t , > to and all admissible u ( + ) , where x ( t o ) = x o and x ( t l ) = x l .
~ o ~ l w ( t )>
d0r (4) Assumption 3 then gives

with x(to)=O and w ( t ) = w[u(t), y ( t ) ] evaluated along the trajectory of

(1).
Two important special cases of dissipative systems are the following. along any trajectory of (1). To turn this inequality into an equality, we
introduce a functiond : R X Rm+R via
passive systems w = uy
finite gain systems w = k%u -yy, k being a fixed scalar.

A trivial, but useful, observation is that if (1) is dissipative with respect

to I supply rates wi, i = 1,. . . ,I, then it is also dissipative with respect to = - V @ = ( x ) [ f ( ~ ) + G ( x ) u l + ~ [ ~ , h ( x ) + J ( x ) (9)
ul.

From (8), d ( x , u) > 0 for all x and u. In addition, it is clear from (9)
that d ( x , u ) is quadraticin u. Combining thesetwo observations, it
follows that d ( x , u ) may be factored as
for any set { q} of nonnegative coefficients.
Notice that dissipativeness, as just defined, is an input-output prop- d ( x , u ) = [ I ( xW) +( x ) u l [ I ( x ) + W(x)ul ( 10)
erty of the system. The following theorem, which is the central result of
this short paper, shows that dissipativeness can also be characterized in for some functions I : Rn+Rq, W : R n + R q X m , and some integer q.
terms of the coefficients in the state equation(1). A restricted version of (Notice, however, thatthe choice of q, I , and W is farfrom being
this theorem, for passive systems only, appears in [I]. unique.)
Theorem I : A necessary and sufficient condition for (1) to be dissipa- Substituting (10) into (9) gives
tivewith respect to supply rate (2) is that there exist real functions
@: R - r R , I : R-+Rq, and W :R + R q X m (for some integer q) satisfying -V@,(x)f(~)-V@,(x)G(x)u+h(x)Qh(x)+2h(x)S (x)u+ud (x)u
=I(x)I(x)+2I(x)W(x)u+uW(x)W(x)u
@ ( x ) > 0, d o ) =0
V@(~)f(x)=h(x)Qh(~)-I(x)l(~) (5) for all x and u. Equating coefficients of like powers of u, we obtain (5)
with @ = 9., n
$ G ( x ) V @ ( x )= S ( x ) h ( x ) - W ( x ) I ( x )
Equation (6) can be interpreted as expressing an energy balance for
i (x)=W(x)W(x) system (l),and shows thatthefunctions arestoragefunctions
@(e)

satisfying Willems definition [4]. The proof of Theorem 1 uses a dif-

for all x, where ferent approach to that adopted in [I] for the special case of passivity.
There the necessity part of the proof relied uponHamilton-Jacobi
+ ( x ) + J J ( ~ ) Q(Jx )
d ( x )= R + J ( ~ ) sSJ theory. For a more complete discussion of dissipativeness along the lines
and of this short paper, the report [8] can be consulted. In particular, it is
shown in [8] thatthealgebraicequations (5) possess maximum and
s (x)=QJ(x)+S. minimum solutions which correspond to therequired suppb and aoailabk
srorage, defined in [4]. (Equation (3) provides the minimum solution.
Proof: To prove sufficiency we suppose that I ( - ) , and E(.)
+(a), This observation will be useful inalater section.) -
are given such that (5) is satisfied. Then for any admissible u( any to ,
)
e We can now give characterizations of dissipativeness, with respect to
 710 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, O a O B E R 1976

particular supply rates, by substituting the appropriate Q,R, S into (5). to do this is to assume that (1) is uniformly zero-state detectable, in the
Example I : For finite gain, we set Q = - I , S =0, and R = k21 where sense that there exists a strictly monotone increasing continuous function
k is a scalar. This gives ( 5 ) as /3 (.) defined on [0,w), with /3 (O)=O, and

V + ( x ) f ( x )= - h ( x ) h ( x )- [ ( x ) [ ( x )
f G ( x ) V @ ( x ) =- J ( x ) h ( x ) - W ( X ) ~ ( X ) and a constant T > 0 such that for any x, and to with u(t)=O

kZI-J(x)J(x)= W ( x ) W ( x )

which,whenspecialized to linear systems,is a generalization of the

Bounded Real Lemma [5]. This is similar to the definition of uniform observability used in [lo], and
A differential version of (6)is given by the following result. it is easy to show that it implies that
Corollaly: If system (1) is dissipative with respect to supply rate (2),
then there exists a real function +(.) satisfying + ( x ) > 0, +(O) = 0, such
that I
for all solutions e(-) of (5).
N. STASILITY SYSTEMS
OF PASSIVE
for the system (1).
Proof: This is simply a restatement of (9), with d ( x , u ) defined as in As an illustration of the use of our theoryindevelopingstability
results forparticular supplyrates,wenow confine our attention to
( 10). A
passivestystems. As observed aboveand in [l], passivesystems are
stable. To achieve asymptotic stability, we introduce strong passivity-a
111. STASILITYOF DISSIPATIVE
SYSTEMS property which ensures that no nontrivial trajectory is free of dissipa-
Willems [4] suggests the usefulness of the theory of dissipative systems tion of energy.
inthe investigation of systemstabilityvia Lyapunov methods. The Definition 3: The system (1) is said to be 1) U-strongly passive (USP)
Lyapunov functions are genertllized energy functions, corresponding to if it is dissipative with respect to
I$(.) in Theorem 1. This leads us to consider conditions for which +(.) is
positive definite, in the sense that + ( x ) > 0 for all x#O. w(u,y)=uy-auu forsomec>O,
Definition 2: The system (1) is zero-state detectable if, for any trajec- 2) Y-strongly passive (YSP) if it is dissipative with respect to
tory such that u ( t ) = O , y ( t ) = O implies x ( t ) = O .
Lemma I: If the system (1) is dissipative with respect to supply rate w(u,y) = uy - ~ y y for some z > 0,
(2) and zero-state detectable, then all solutions +(.) of ( 5 ) are positive
definite. and 3) very-strongly passive (VSP) if it is dissipative with respect to
Proof: The minimum solution of (5), givenby (3), is positive defi-
nite if .there exists a control such that w(t) < 0 on [r,, w), with strict w(u,~)=uy-zIuu-aE2yy forsomezl>0,z2>0.
inequality on asubset of positive meashe. Now from Assumption 4,
there certainly exists u such that w(b,y)<O for any y#O. In fact,. the One way to interpret these definitions is as follows. A YSP system is a
quadratic natxe of w(-,.) gives us sufficient freedom to choose such a u passive system for which a small amount of positive feedback does not
under the additional constraint that the chosen u and y be compatible destroy the passivity property; a USP system can be similarly interpreted
with thestateequations (1). (Thatis, if thereexistsany u such that in terms of feedforward. (The usual definition [lo], [ 111 of strict passivity
w(u,y)<O, then-because w ( . . .) isquacjraticin its arguments-there corresponds to our definition of USP.) A VSP system is, of course, one
actually exists a wide choice of such us. We could for example choose whichis both USP and YSP. Another example of the possibility of
u = K ( x ) y , where there is enough freedom in choosing K ( x ) to ensure combining supply rates in the manner mentioned in Section I1 is pro-
that [ I - J ( x ) K ( x ) ] is nonsingular for all x.) Accordingly, we can choose vided by obsewing that a USP and finite gain system is VSP.
a feedback law u*(.) such that w(u*(y), y ) < O for y#O, and u*(O)=O. We now summarize stability results for passive systems which follow
Wenow havethe desiredresult,provided that we can exclude the immediately from Theorem 2.
situation in whichy(t)=O for almost all t. However this last possibility is Theorem 3: For systems of the form (l), passive and USP systems are
excluded by zero-state detectability. n stable, while YSP and VSP systems are asymptotically stable.
Now we easily obtain our main stability theorem. The following example serves to illustrate the ideas of tlus short paper.
Theorem 2: Let system (1) be dissipative with respect to supply rate Example 2: We consider the equation
(2) and zero-state detectable. Thenthe free
system i = f ( x ) is
i+f(x)i+g(x)=u (12)
(Lyapunov) stable if Q < 0 and asymptotically stable if Q < 0.
Praoj From the corollary to Thebrem 1 and Lemma 1, there exists wheref(.), g ( . ) are functions of the scalar variable x and u is a forcing
a positive definite +( .) which satisfies term. Setting u = 0 gives the Lienard equation [9].
Letting F ( x ) = J G f ( u ) d u , a set of first-order equations equivalent to
(12) is

along the trajectories of x =f(x). The result then follows from standard
x, = - F ( x , ) + x2
Lyapunov stability [9] (for the case of Q < 0, asymptotic stability follows x2= - g ( x , ) + u
byusinga contradiction argument, basedonthe LaSalle Invariance
where x, = x .
[9]Principle anddetectability).
zero-state a We combine (13) with the system output equation defined by
We see immediately that passive systems ( Q = R = 0, S = I ) are stable
and finitegainsystems ( Q = - I , S=O, R = k21) are asymptotically y=au,-PF(x,), a>B>O
stable. That a finite gain system with a minimal state space is asymptoti-
cally stable has been shown in [ 101 using a different approach. and consider the passivity of thissystem (of course we could find
Theorem 2 has given conditionsfor the
local
stability of the conditions for other forms of dissipativeness). It is convenient to define
equilibrium at x = O . It is evident that to achieveglobal asymptotic G ( x ) = JGg(u)du.
stability, we need to impose stronger conditions on the system. One way The calculations are straightforward and involve substitution into (5)
SHORT PAPERS 711

and determinationof conditions for theexistence of a solution. We only Identification of Linear Systems with Time-Delay
present here the results for two special cases. Operating in a Closed Loop in the Presence of Noise
For a = 1/2, B = 0, passivity follows if
E. GABAY AND S. J. MERHAV, MEMBER, IEEE
G ( x )> 0
Abstruct-Xbe subject of this short paper is on-line i d e n t i f i i n of the
and parameter vector defining a linear dynamical system which operates in a
clased loop in the presence of noise, and incorporates a time delay. The
g ( x ) F ( x ) 0.
method is based on the equation error. Other known subsystems in the
The storage function is closed-loop system increase the dimension of the closed-loop parameter
vector whichtends to degrade the estimationconvergenmprocess. By
means of composite state variablq introduced in this short paper, this
+,(x)=t[x+F(x)I2+G(x).
increase is prevented and the open-loop parameters are directly identified
from dosed-loop input-output data. The pure time delay in the closed loop
For a = p = $, passivity follows if causes a representation problem inthe equation error formulation. Tbis is
G(x)>O overcome by composite delayed state variables. The value of the time
delay is determined by means of ex- parameters provided by a higher
and ordermodelandasimple on-line search procednre. The method is
illustrated by simulated examples.
f (x) 0.

The storage function is I. IhTRODUCnON ANm STATEMENT OF THE PROBLEM

+2(x)= $x2+G ( x ) . The subsystem G(s) of unknown structure and order to be identified
includes a pure time delay and is part of a closed-loop system incorpo-
The functions and +2 are standard Lyapunov functions used for the rating other known dynamicalsubsystems (Fig. 1). The system is excited
study of the stability of theLienard equation: choice of +2 as a by a given stationary random input, and Gaussian uncorrelated additive
Lyapunov function is motivated by its interpretation as the sum of the noise is assumed to be present in the loop. In terms of Laplace trans-
kinetic and potential energies of (12), and + I = + 2 + i F ( x ) + i F ( ~ )i2s forms, G ( s ) is given by
called the modified energy function [9]. We have now shown that they m

can both be interpreted as stored energy functions (depending on how e- bpi

one defines the system output) arising from a study of the passivity of G(s)=e-- N ( s b) = i=O
n
(13).
D (s.a)
1+ c.
J =1
ajs/

V. CONCLUSION
where rn, n denotethe highest numeratoranddenominator powers,
Themain result presented here is Theorem 2,which relates the respectively, and rn < n. G(s) operates in a closed loop (Fig. 1) with
stability of abroad class of nonlinear systems to the input-output other dynamical elements G,(s), G2(s) and G,(s) having known parame-
property of dissipativeness. We can characterize dissipativeness by the ters. Tracking or regulating tasks in manual control are special cases of
existence of a computable function +(.) of the state. This function is the system in Fig. 1. The following assumptions regarding the input,
thought of as the stored energy of the system and under certain condi- system, and noise are made.
tions is a Lyapunov function. Assumption I: The input ~ ( tis) a sample of a given random stationary
Our approach can also be applied profitably to interconnected sys- mean square bounded ergodic process. Its spectral distribution guaran-
tems, in particular feedback systems. One can, for example, derive tees a persistent excitation of all the modes of G(s).
Lyapunov versions of the stability criteria of Zames [ l I], using a Assumption 2: G ( s ) and the closed-loop system denoted by T ( s ) is
Lyapunovfunction which is the sum of the storage functions for the stable and time invariant.
individual subsystems. Details of these and other results are currently in Assumption 3: The noise n l ( t ) (Fig. 1 ) is a zero-mean stationary
preparation and will be reported separately. ergodic Gaussian process uncorrelated with x ( t ) .
It is required to provide unbiased estimates of Q and b from ~ ( t and
) a
suitably chosen closed-loop system output. Since the sensitivity of the
REFERENCES closed-loop system output[output of G2(s)] to variationsin a, b is
reduced by the loop gain, the system error,denoted by z ( t ) (Fig. 1)
P. J. Moylan, Implications of passivityin a class of nonlinear systems, IEEE which retains this sensitivity, is chosen as the appropriate system output.
Trans. Automur. Contr., vol. AC-19, pp. 373-381, Aug. 1974.
B. D. Anderson, A system theory criterion for positive real matrices, S I A M 1. The corresponding closed-loop transfer function relating x(t) to r ( t ) is
Contr, vol. 5, pp. 171-182.May 1967.
P. J. Moylan and B. D. Anderson, Nonlinear regulator theory and aninverse
optimal control problem, IEEE Truns. Automnr. Conrr., vol. AC-18, pp. 460-465, T(s)=G,(s)[l+Gl(s)G2(s)Gj(s)G(s)1-. (2)
Dec. 1973.
J. C. Willems, Dissipative dynamical systems Part I: General theory; Part 11:
Linear systems with quadratic supply rates, Arch. RafioMi Mechanics and Ana!vsis, In-time domain,z ( t ) is given by
vol. 45, no. 5 , pp. 321-393, 1972.
B. D. Anderson and S. Vongpanitlerd, Nerwork Analysis and Synrhesis. Englewood ~z(t)=T(p)x(t)-fn(t)=y(~)+n(t) (3)
Cliffs, NJ: Prentice-Hall, 1973.
J. P. LaSalle, Dissipative systems, in OrdiMly Differential Epuntiom 1971 NRL-
MRC Conference, L. Weiss, Ed. New York: Academic,1972. wherep d/dt, i=O, 1,2;.., and
A. V. Balakrishnan, On the controllability of a nonlinear System Proc. Nur.
A d Sci. U.S.A., vol. 55, pp. 4 6 5 4 6 8 , 1966.
D. J. Hill and P. J. Moylan, Cyclo-dissipativeness,dissipativenessand lorslessness n ( t ) = - G2 (zJ)G, ( p ) G ( p ) T ( p ) n ~ ( t ) . (4)
for nonlinear dynamical systems, Dep. Elec. h g . , Univ. Newcastle, Newcastle,
N e w South Wales,Australia, Tech. Rep. EE-7526, Nov. 1975. y ( t ) is the system error in the absence of noise. The known subsystems
J. LaSaIle and S. Lefschea Smbi/ity by Lyapnms DirectMefhod. New Y a k :
Academic, 1961.
J. C.WiUems,The generation of Lyapunov functions for input-output stable
systems, S I A M 1. Contr., vol. 9, pp. 105-134, Feb. 1971. Manuscript received August 12, 1975; revised April I , 1976. Paper recommended by E.
G. Z a m q On the input-output stabilityoftime-varying nonlinear feedback Tse, Past Chairman of the IEEE S C S Stochastic Control Committee.
systems; Part I: Conditions derived using concepts of loop gain, conicity and The authors are withthe Depamnent of Aeronautical Engineering, Technion-Israel
positivity, IEEE Trans. Automat. Conrr., vol. AGI I, pp. 228-238, Apr. 1966. Institute of Technology, Haifa, Israel.