FM9 14:40 -

Dissipative Nonlinear Systems: Basic Properties and Stability Analysis

David J. Hill
Departmentof Electrical& Computer Engineering, The University of Newcastle,
Universlty Drive. Caiiaghan. N W 2308.Australia
Abstract circle criterion can all be generalised to com-
Energy-relatedconcepts like passivity, dissipa- putationally handle systems with nonlinear dy-
tion and energy storage are widely appreciated namics and multiple equilibria. The results have
in the study of physical systems. It turns out that evolved in work by Willems [ 11, 12) and Hill and
recognition of such properties in abstract sys- Moylan [ 13-191.
tems can leadtpsignificant progress in the study Notation
of dynamic propertiesand stabilisation.The key R + is the positive real line [0, a); R, is the ex-
result inthe developmentof such a theory for lin- tended real line Ru(- = ) U ( + =]; Z, is the
ear dynamics was the Kalman-Yakubovich positive integers (0,1,2. . 1; &;(R+) is the Hilbert
lemma. This paper reviews a theory which ele- space of square integrable Rmvalued functions
vates this result to nonlinear systems and pro- on R +; %(Y) denotesthe Banachalgebraof con-
vides a framework for general stability theory of tinuous linear operators on Hilbert space Y; A'
interconnectedsystems. denotes the adjoint operator of A; a function
1 INTRODUCTION
f : R + -R+isinclass%iff(O) = Oanditisstrictly
monotone increasing.
Consider the analysis and control of qualitative
behaviour in nonlinear (and possibly intercon- 2 DYNAMICAL SYSTEMS
nected) systems.Through the 1960'sto early 80's. We take it as self-evidentthat a basic description
the stability theory approach to analysis based of a system should mathematically express its in-
on Lyapunov theory and/or input-output meth- teraction with the environment. Without an em-
ods was developed. A central role was played phasis on the external variables, there is no satis-
by passivity, finite gain and related concepts. factory way to keep the system description
The computational developments were mainly invariant under embeddlng within larger system
limitedto linear dynamics and memoryless nonli- descriptions. The definitions adopted here are
neariiies. Consequently, emphasis came to posi- taken from [19] and include ideas from Willems
tive and bounded-realness in the frequency do- [20], Zames [21] and Hill and Moylan [17].
main and the well-known Kalman Yakubovich Let w,X be abstract sets called the signalrange
Lemma. For the next ten years or so, these con- and state space respectively. W has a zero ele-
cepts played some role in advancements of ment denoted by e. t c R denotesthe time axis.
adaptiveand robust control.While small gain ar-
guments are commonly used in convergence
proofs of nonlinear algorithms, again the major
-
The signal space W ,has elements which are
functions w : t w. (Thesignificance of the sub
script 'e' will be clarified later). The value of o at
focus has been on linear plant dynamics. Re- t E t i s denoted by w(t). Let B denote the space
cently, considerable interest has emerged for with elementsx :t x called state trajectories.
--+

the use of energy and gain concepts for adap- The quantity X ( I ) E x is called the State af time
tive and robust stabilisationof smooth nonlinear I . The slgnal space for both external and internal
systems [ 1-51, This work relies on nonlinear gen- system variables is SGi x W,. Let 93, c % x W , con-
eralisations of the above mentionedlinear char- sist of those signals which are compatible with
acterisationsof passivity and gain concepts.The the dynamics of the system. %,is called the total
extensionof computationalaspects to nonlinear behaviou A dynamical sysfem in State space
dynamics is made possible by techniques of form is defined by Zs= ( r ,W,x,93B,) if 93, satisfies
geometric analysis and control [6,7]. This paper the axiom of state [20]. The axiom of state says
reviews a theoretical framework within which that any trajectory leading to a given state can
such developments are proceeding. be combinedwith anyfuturetrajectoryemanat-
The outstanding feature of stability theory -see ing from that state to give a feasibletrajectory in
for instance [8-101-is the facility to combine sub- ScGiven the system 2,,we can define its exter-
system propertiesinorder to conclude an overall nal behaviour as
system property. However, much literature on 93, = ( w E W.13X E 96 S.t. ( X , O ) E 93,)
stability theory provides the impression that it is
limitedto study of a single equilibrium in systems 2, = ( r ,W ,93,) is the external dynamical system
with inear dynamics and memoryless nonlinea- induced by Zs.r = R + and r = 2, are preferred
riiies (whichcan be confinedto sectors). Further, choices of the time axis for continuous-timeand
there may appear to be only a limitedset of rath- discrete-time systems. External descriptions
er Inflexible stability conditions. By contrast, the based on t = R have been used. However, as-
modern theories aim directly to solve problems sumptionsaboutbehaviour at i 4 - = are gen-
of nonlinear dynamical systems possibly with erally unappealing and especially problematic
multipleequilibria. This paper shows that the ba- for nonlinear systems. A finite time origin forces
sic tools of stability theory includingthe passivity considerationof the role of initialconditions ~ ( 0 ) .
theorem, Kalman-Yakubovich Lemma and For a fixed x0 := do), the dynamical equations
CH3229-2/92/0000-3259$1.OO Q 1992 IEEE 3259

Authorized licensed use limited to: Peking University. Downloaded on December 28,2021 at 13:53:04 UTC from IEEE Xplore. Restrictions apply.
 can be solved for a subset of the external beha- ular, we denote by X,(Q) the set of all states
viour $,(x0) S $.. Suppose,forsome particularin- reachable from $2.
vestigation,the externalsignals have been clas- A further concept needed is that of an equilibrk
sified as inputs or outputs. The signal space um state, i.e. a state X* s.t. X* = tp(t,x', 0)Vr E T.
W , = CU, x SJ, where 21. and SJ. are called the 3 DISSIPATIVESYSTEMS
input signal space and output signal space re-
spectively. The external signal w can be written The concept of dissipativenessexpresses the no-
as (u,~)where U E CU, is the input and y E SJ. is tion that a system dissipates energy in some a b
the output. Clearly, the external behaviour 9. stract sense. The 'energy' may be derived from
becomes a relation (? on CU. x SJ.. 6 can be some underlyingphysics. However, this connec-
tion is not required. Suppose the energysupply
used as an input-output system representation functional is a predetermined functional
[21]. However, this representation is slightly too
E : W e x t -. R such that for a given w E W,and
general for extracting internal behaviour in-
formation, i.e. properties of statetrajectories.We T E r , the number E(w,T) E R represents ab-
stract energy input to the system on interval
e
assume hasthe propertythatfor a given x0 E X
[O, U.
and U E 21, ,there exists a unique y E SJ,.. Then
Definition 1. Dynamical system E. is weakly dissi-
%.(x0) corresponds to an operator pative with respect to energy supply E and a
e
G(xo):2 1 , d sj, . The relation has beendecom-
class e of subsets of 9, if and only if 381 E e and
posed into a family of operators indexed by
x0 E X .This family is denoted GQ. A dynamical p E R s.1.
system input-outputrepresentationis a family of E(w,T) + p 2 0 (5)
operators
VW E 9, and VT E t.With /3 = 0, we call G dissk
G(Q) := (G(x0)LxoE 52 5 XI (1) pative.
P is some prefened subset of internalstates. We Since - E represents energy extracted from the
will see later that it is sensible to include 'system system, inequality(5) places an upper bound on
equilibria' in SE as 'rest states'. If SZ = {x'), then the energy extracted over [ O , q for any signal in
G(P) becomes the usual single operator repre- a subset of the behaviour.For this definitionto be
sentation used in input-output system theory [a]. meaningful, e will be chosen to be nontrivial.
Inthat case, the dependence on X' istraditional- Suppose that the dynamicalsystem Z, has an in-
ly ignored. put-output representation G(P) corresponding
Now introduce the operator family PT,T E t, to W , = 21, x 91.. Just for notational conve-
which truncates signals at time T , i.e. nience, the dependence on SZ can be sup
pressedwhen only input-outputcalculationsare
required. Also we use the notation
E(U,Y,T) := E((u,y),T).
The most useful general class of energy supply
It is convenient to introduce the notation functionals are quadratics on inner product
oT:= P , a . This enables us to impose a causality spaces. Suppose that '%. , 91. are extensions of
structure on system G(Q) [a]. Hilbert spaces CU, 91 in the usual sense of input-
Forthecasewhere the externalsignals are classi- output stability theory [a]. The causal truncation
fied as Inputs and outputs, we will make use of a operators now map according to P ; :CU. + 9,
more refined system representation than Zs. and P; : 91. -. 9. The inner productson '%, SJ sat-
Note that there is a mapping of (xo,u) E X x CU, isfy some sensible consistency properties. The
to y E SJc In order to monitor the internal state commonly used examples are &;(R +) and e;(Z+)
x(t), we consider mappings ly : t x x x CU. -. X for continuous and discrete-time respectively.
and r :X x U --c Y called the state-transitionand Elements of CU, 91 and W = CU x SJ are called
readout function respectively. These describe smallsignals. CThey are bounded in the norms in-
the time evolution of the state and output ac- duced by the respective inner products).
cording to
Let Q:sJ.-%,, S:CU.-,SJ. and R:CU,-CU,
x(0 = w,m, 4 (2) be predeterminedcausal linear operators and
YO) = r(x(t),W (3) E(u,y,T)= y,Qy > T + 2 ~ , S U> T + < U,RU >,(6)
If tp, rsatisfysomestandardaxioms[22],theybe- Taking the first term in the right side, it denotes
come a reasonable description of a causal < YSQYT > = < Y.(QYT)~> = C Y ~ Q > Y . mese
(time-invanapt) dynamical system. For instance, equalities follow from causality and properties of
q is such that x(tl) depends only on ~ ( t )for the inner product). Within (6). we regard Q,S,R
t E [O,r,]. (ly, r) is called a state-spacerealization as operators on 91, CU. Suppose they are contin-
of G(P) if uous with Q,R self-adjoint.It is easy to check that
y(0 = (G(xO)u)(O = r[ly(~.xo.U), u(t)l (4) E is well-defined and satisfies the properties of
Definition 1. When E has the quadratic form giv-
Vt E t,Vxo E 52 and VU E CU,. en by (6).we refer to (Q. S,R) dissipativeness. Cer-
A variety of system properties can be defined tain special choices of Q, S and R give common-
such as reachability,detectability [ 171. In partic- ly used input-output properties. The symbols 0
3280

Authorized licensed use limited to: Peking University. Downloaded on December 28,2021 at 13:53:04 UTC from IEEE Xplore. Restrictions apply.
and I are used to denote the zero and identity era1abstract dissipative systems, it is possible to
operators respectively. define energy-like functions.
Examples 1. A range of so-calleddissipation lemmas can be
We assume y = Gu throughout. For each proper- given which equate external disslpativeness
ty, (Q,S.R) are glven. The qualifier 'weak' is not properties to existence of internal energy stor-
used. age functions. One such result follows.
Theorem 1. Assume that XIG X is reachable
(a) Pseudo strict passivity:Q = - df, S = 51I, from 9 S x and detectable with respect to P.
R = - dwhere d , ~ ER
<~~Y>,+B~EIl~II:+~IlYll: (7)
-
Then G(S2) is dissipative with respect to E if 3 a
function # :X R.+ s.t. #(x) < ~ V EXX,,#(x) >
0 VX e Q, #(x) = 0 Vx E 9 and
Some terminology for strict passivity is as follows. 44x1) + E(u,G(xAu.T) 2 $(xd (1 1)
If E = 0, b = 0, the property is passivity. E > 0
VT E t, VU E Cue where x2 = q(T,x,. U).
gives input strict passivity (ISP) and 6 > 0 gives
output strict passivity (OSP). If E > 0 and 6 > 0, The sufficiencyhalf of Theorem 1 is obvious: since
property (7) is called very strict passivity (VSP). $(x3 2 0,(1 1) clearly implies (9)with @(x) = #(x).
(b) Finite-gain (FG): Q = - I , S = 0, R = k21 Thus, the function B(.) has been associatedwith
where k E R + initial stored energy. The proof of necessity fol-
lows by identifyinga function #which satisfies in-
II Y 11s: k2 II It: +B (8) equality (1 1 ). One such function is

The above special cases will serve our current

cl &(x) := - rEQ,
inf E(u, G(x)u,T )
TrO
(12)

needs. Numerous other concepts are useful in- A #(.) satisfying the conditions of Theorem 1 is
cludingsectors, exterior conicity, interior conicity called a storage function. It can be shown [ 1 1,
etc. [12,14,16,211. 151that the set of all storage functions is a con-
Pure input-outputanalysis generally takes G as a vex set with minimal element #a given by (12).#a
single operator or a relation and ignores the un- is called the available Storage; it can be inter-
derlying internal behaviour. Our framework al- preted as the maximum amount of energy
lowsthis (on Hilbert spaces). However,all but the which may be extractedfromthe system at state
most ardent believer in input-output methods x. If # is defined on all x.the system is called inter-
would ask what the implicationsof externaldissi- nally dissipative.We notethat an externallydissi-
pativeness are on an internalsystem representa- pative system with x,(Q) = Xis internblly dissipa-
tion. Suppose now that G ( 9 )has a time-invariant tive.
realisation ( q , ~ )and 9 EX. Take e as It is easy to generate simple examples showing
($,(xo)lxo E 53 G X ) . Definition2 then says that a the distinction between external and internal,
dynamical system G(Q) with energy supply E is weakand nonweak etc. propertiesfornonlinear
weakly dissipative if and only if 3 a function systems with multiple equilibria[ 15,171. In simple
/3 : Q + R S.t. examples, it is ofteneasy to guess suitablecandi-
dates for #(.). However, for special classes of dy-
W.
G(xd)u, T ) + B(x0) 2 0 (9) namical systems, systemmatic ways exist to
computethe #(.). Consldertheclassof nonlinear
Vu E 'U,, VT E t and Vxo E 9. If B(x0) = 0 V finite-dimensionalsystems which are affine Inthe
xo E Q, C(Q) is dissipative. The dynamical system control
2,is called externally (weakly) dissipative if G ( 9 )
is(weakly) dissipative. Let 'U = SJ = &;(R +).Then i = f ( x ) + G(x)u (13)
(9)for passivity becomes y = h(x) + J(x)u (14)
The signal spaces are CU, = &;(R+) and
SJ. = k;(R +); the state xtakes its values in Rn.The
usual smoothness for unique solutions is as-
sumed. Uhis class of systems is wldely used in
VU E e;@+), VT 2 0 and Vx(0) = X, E n. modernnonlinear controltheory[ 1-4,6].) Further,
Typically inequality(9)with B = 0 can only be sa- suppose xr(9)= Ra for 9 G Rn.
tisfiedfor Q a finite set of isolatedpoints.(We see Then quadratic dissipativeness can bewriien as
later that under certain conditions these points
are stable equilibria). However, with # 0, in-
equality (9)can hold Vxo E X (if XX9) = x). In-
equaliy (lo)is familiar from literature on defining
E(U,Y,T) =
I,&(t),y(t))dr 2 0 (15)

passivity concepts for nonlinear circuits [23]. for x(0) E 9,all T 2 0 and all U E &m,(R+)where
Physical reasoning associates a passive electri-
cal circuit with a unique stored energy function w(u,y) = y'Qy + 2y'Su + u'Ru (16)
(to within an additive constant) which has cer- and Q E RpYp,SE Rpxmand Rmxm are constant
tain positivityproperties.It turns out that for gen- matrices with Q.R symmetric.
3261

Authorized licensed use limited to: Peking University. Downloaded on December 28,2021 at 13:53:04 UTC from IEEE Xplore. Restrictions apply.
 Following [13], we note that for q5 E C1 the dis- VT E t, VU E U,. If p = 0, the system is calledfi-
sipation inequalityis equivalentto the partialdif- nite gain stable. It can be shown that (weak)
ferential inequality(PDI) FGS 5 (weak) FG, but not necessarilywitht = p.
VqY(f(x) + G(x)u) 5 w(u, h(x) + J(x)u) (1 7)
For study of input-output stability theory, we
adopt the setting of quadratic dissipativeness on
for all x E R", U E R". We then have an algebraic Hilbert spaces. This basic input-output stability
characterisationof dissipativeness [ 131. tool places a sign definiteness property on oper-
Theorem 2 A sufficient condition for system ator Q.These properties are the standard ones
(13)-( 14) to be dissipative with respect to supply for elements of the Banach algebra 93(Qo.
rate (16) is that there exist real functions lemma 1. G is weakly (Q,S,R) dissipative with Q
q5 : Rn R, : Rn Rq and W : Rn-.. Rqx* (for
-+ -+
strictly negative definite if and only if G is WFGS.
some integer 4) satisfying #(x) 2 O,$(X) = 0 for For study of internal stability concepts, we a-
x E Q and sume that the state space Is a normed space.
vq5(xyy(x) = h l ( x ) ~ h ( x )- eyx)e(x) (18) Note the dissipation inequality (1 1) implies that
internal dissipativeness is a natural generalisa-
+yx)v#(x) = M#(X)~(~) - wwe(x) (19) tion of the concept of a Lyapunov function to
systems with inputs. (It is actually a weaker ver-
N(x) = W(x)W(x) (20) sion of the concept of a controtlyapunovfunc-
for all X , where tion [28] used in recent nonlinear stabilisation
theory.) However, a strong statement about in-
M(x) = R + J'(x)S + S'J(x) + J'(x)QJ(x) (21) ternal stability can be made without use of stan-
N(x) = QJ(x) +S (22) dard Lyapuov theorems or smoothness
constraints on the storage functions. Let
These conditions are necessary if we assume
that some storage function, say $a at (12). is C'
d(x,Q) = $5 II x - xo I
(131. The solution of the equations (18)-(20) can Theorem 3. Suppose G(Q) is ( Q , S , R ) dissipative
be approachedvia Hamilton-Jacobiequations with Q strictly negative definite. Let
relatedto the nonlinear optimalcontrolproblem D , = ( x : d(x,Q) 5 d , ) , for some dl > 0, be uni-
defined by (12)-(14) [3, 241. A major difficulty is formly reachable from Q and detectable with
that smooth systems (13)-(14) may not have respectto Q. Then 3some d, > 0 (dependent on
smooth optimal costs. So,ingeneral,the storage d , ) such that, with zero input, all state trajectories
functions are not necessarily differentiable. starting in o2= (x : d(x,Q) 5 d,), remain in Dl
James [25] has recentlyshown that, for the non- and asymptotically approach 52.
linear systems considered here, storage func- Lemma 1 and Theorem 3 give that, under suit-
tions can be assumed to be lower semicontinu- able reachability and detectability properties,
ouswithout loss of generality. Further,the storage FGS implies local asymptotic stability.With 8 not
functions can be completely characterised by a connected region in X, then D , and D , need
the POI (1 7) hterpreted in a weak sense valid for not be connected.It can be shown that a weak
semi continuousfunctions [26]. FGS system is not necessarily Internallystable. In
For linearfinite-dimensionalsystems, a complete fact, it may have unbounded trajectories with
theory can be presented:using Parseval's theo- zero input [ 171 or exhibit more exotic behaviour.
rem, external dissipativeness (1 5) can be char- Example 2
acterisedinthe frequencydomain by a general- Consider the modified Duffing equation
isation of positive-real and bounded-real
functions[27]; the distinctionbetweenweak and j; + 6 i - px + x3 = ycoswt (24)
non-weak properties disappears; the storage where 6 E R , , E €4, y E R (61. This is a special
functions take the form #(x) = ix'Px, p E Wax"; case of a forced Lienard equation
and the equations (18)-(20) reduce to those i + f ( x ) i + g(x) = U (25)
well-known for the Kalman-Yakubovich (or Posi-

1.'
tive-Real) and Bounded-Real Lemmas [ 1 1,271.
Thus, the diqipationlemmas are substantial gen- Let $(x) = g(u)da and assume:
eralisations of the Kalman-Yakubovich Lemma
to general dynamicalsystems. (i) f(u) 2 6 > 0 Vu
4 STABILITY THEORY (ii) $(x) + C 2 0 VX and some C E R
The framework established in Section 3 enables These properties are satisfied for the Duffing
study of input-output and Lyapunov stability
properties for dynamical systems with multiple equation with C = B2 Consider system G(Q)
equilibria. mapping U to y = i with Q unspecified. A real-
The basic input-outputstability concept requires ization of G(f2) is
that y = GUis a small signal for all small signal in-
puts.A stronger propertyis weak finite guinstabil-
ify(FGS) where 3 y E R+, p E R S.I.
II Y llr= Y I U IIT + cr (23)
3262

Authorized licensed use limited to: Peking University. Downloaded on December 28,2021 at 13:53:04 UTC from IEEE Xplore. Restrictions apply.
It is easy to show G(R) is weak OSP with the simple development of Theorem 4. Lemma
2 connected dissipativeness parametersof sub
$(x) = 21 2 + $(x) + c (26) systems Giand the overall 6.Suppose now each
The system is weak FGS. Thus L2 inputs produce subsystem Gi(Ri) is dissipative and sets Dl, are unC
formly reachable and detectable with respect
de2 outputs.
to 53, as in Theorem 2. Clearly, we need to make
For the Duffing equation case, take

-
a connection between sets Dli,Qi and their
8,= f [ ,@I. G(8,) is OSP and 8, is locally counterparts for the overall feedback system.
There exists storage functions $, :Dli R, which
asymptotically stable, i.e. autonomous trajecto- are positivsdefinitewith respectto f2,and satisfy
ries are attracted to Q, However, in [6] we see (1 1). Consider the product space X = X, x Xz
that for U(C) = ycosorandcertainvaluesof 6 , w , y and the total storage function
the trajectories exhibit chaos.
Dissipativesystemst heory provides a strong basis
cl $(XI = Ql(X1) + P h ( X - 3 (28)
for analysis of stabiiityof interconnectedsystems. Then it is easy to show that $(-)is a storage func-
To illustrate, consiUer the special case of a feed-
back system shown in Figure 1,
tion for e(&)
which corresponds to ($,i.k) dissi-
pativeness given by Lemma 2 where
h = .RI x Q2 For zero external inputs, we have
5 < Y , &Y > r + +(MO) (29)
The reachability and detectability assumptions
ensure $(.) is bounded above and below by
Figure 1, Feedback system
class 36 functions. Then the conclusion that b is
The subsystems are represented by operators asymptoticallystablecan be obtained using the
G1 : Y8 Y82and Gz : ‘V82 Vel where ‘V81,‘V8z
-+

same steps as for Theorem 2. If D , = X i , d is glob-

are edended signal spaces. Signals rl,rZrepre- ally asymptotically stable.
sent external reference and/or disturbance in-
puts. Assume the feedback system can be rep- Results along these lines have been presentedin
[ 1 1, 14, 161. (The discussion here allows more re-

-.-.
resented by a well-defined operator
laxed $ smoothness and stability of equilibrium
6 :V81x V82 V,, x Q‘81 which maps external sets.)
inputs (r,, r3 (y,,y& :,= y . This can be presented Example 3.
as an operator array in the obvious way.
Consider coupled Lienard equations (of the
lemma 2. Suppose that the subsystems form (25))
G,, i = 1,2, are weakly ( e i ,Si,R,)dissipative. Then
the feedback system 6 is weakly (6,i,k) dissipa- i 1 +fI(Xl)k + gdxd = - k 2 2 + rl (30)
tive with
i 2 + g2(x& = k l i l + rz
+f2(X&2 (31)
where kl,kzE R , and A{.), gd.) satisfy the as-
sumptions made in Example 2. The coupling is
through velocity terms. This system clearly can
where p E R,. be viewed as a feedback system of the form in
From Lemmas 1,2, we havethe input-outputsta- Figure 1 with a realisationof each G, of the form
bility result. given in Example 2 except y i = k;,.
Theorem4.Undertheconditionsof Lemma 2, the Let $ i ( X i , i , ) = k,{$if + $,{xi) + CJ. Then from EX-
feedbacksystemis weak FGS if the operator is ample 2 it followsthat Giis weak OSP withstorage
strictly negative definite for some p E W +. function $> The total storage function is
This result contains the well-known passivity, $J = + $> Corollary 1 gives that the system is
small-gain and conic operator theorems of in- WFGS. (We get E,. = &and d, = 0).
put-outputstability analysis [8] as simple special ki
cases. To illustrate, consider pseudo strictly p a s - Now restrict the problem to the special cas8 of
sive subsystems [a]. coupled Duffing equations.Then
Corollary 1. Suppose the two subsystems of a
feedbacksystemare weakly pseudo strictly pas-
sive. Then the feedback system is WFGS if
where
E 1 + 62 > 0, E 2 + 6, > 0

Proof: Set p = 1 and substitute dissipativeness

parametersinto 6 at (27). Ll
We now turn to the study of internal stability.
Clearly, the generalizationof Theorem 3 to inter- W(X,,X~can be interpreted as the potential en-
connected systems is far more technical than ergy of the system. The theory says that
3263

Authorized licensed use limited to: Peking University. Downloaded on December 28,2021 at 13:53:04 UTC from IEEE Xplore. Restrictions apply.
 52 = ( ( x ~ , x & x ~= f &,x2 = f ,/&I
is globally
[9] K.S. NarendraandJ.H. Taylor,&auencv Dorrlpln
Criteria for Absolute StabllUy, Academic Press.
New York, 1973.
asymptotically stable. 0 [lo] A.N. Mlcheland R.K. M l i l e r , ~ a t i v e A ~
Much more can be said about stabilitytheory of Scale Dmmicai Svsterns, Academic
interconnected systems than space here per- %New York, 1977.
mits. The flexibility of the dissiputiveness property [11] J.C. willems, 'Dissipative Dynamlcal Systems,
Part I: General Theory, Part II: Linear Systems wlth
and the facility to use input-outputand/or inter- Quadratic Supply Rates'. Archive for Rational
nal analysis leads to a powerful stability 'tool Mechanics and Analysis, Vol. 45, NoS. 1972, pp.
box'. The experienced user can tailor results to 32 1-393.
suit the occasion. The use of multiplier tech [12] J.C. Willems, 'Mechanismsfor the Stabllltyand In-
niques captures the Popov type criteria (gen- stability In Feedback Systems", Proc. /€€€,%I. 64.
eralised to nonlinear dynamics) [ 141; instability No.1, January 1976, pp. 24-35.
[13] D.J. Hill and P.J. Moylan, 'Stabilltyof nonlineardis-
results can be given without the usual assump- sipative systems', /€E€ Tram on Auto Contrd.Vol.
tion of a linear unstable subsystem [18]. AC-23, No.2, pp. 143-149,pp. 706-711.
Results on Input-outputstability are perhaps the [14] D.J. Hill and P.J. Moylan, 'Stability Results for Non-
linear Feedback Systems-, Automatica, Vd.13.
cleanest In the theory of nonlinear system quali- N0.4, July 1977, pp.377-382.
tative behaviour.The lines of argument are clear, [15] D.J. Hill and P.J. Moyian, 'Dissipative Dynamical
intuitive and relatively simple. The main results Systems: Basic Input-Output and State Proper-
are uncluttered with details of initial conditions ties", J. Franklin institute, Vd.309. NOS.
and existence of equilibria. The conclusions are pp.327-357,May 1980.
clean stability statements for very general sys- [16] P.J. Moylan and D.J. Hili, 'Stablllfy Crlterla for
tems. However, Examples 2 and 3 indicate that Large-ScaleSystems-, I€€€ TrunsonAuto Control.
Vol. AC-23, No.2, pp. 143-149,April 1978.
these conclusions may not reflect some very in- [17] D.J. Hill and P.J. Moyian, "Connections Between
t eresthg system behaviour. Finite Gain and Asymptotic Stability'. /€E€ Tram
Thus conditions for input-output stability, while onAuto Control,Vol. AC-25. No.5,pp.931-36,Oc-
tober 1980.
not giving a complete answer to the stability [18] D.J. Hili and P.J. Moylan, 'General instability re-
problem, are valuable as a conceptualbasis for sults for interconnectedsystems', SIAMJourndof
a large set of stability conditions. Control and Optim/zat/on,Vol. 21, No.2. March
5 CONCLUSIONS 1983, pp,256-279.
[19] D.J. Hili, 'Dlsslpaliveness, Stability Theory and
This paper has reviewedthe theory of dissipative some RemainingProblems'. In and Con-
systems with emphasis on establishing general- ,- ed.C ~ M C.F. . Mar-
ised Kalman-Yakubovich Lemmas for nonlinear tln and R.E. Sueks, North Holland, Amsterdam.
systems and its application to stability analysis of 1988.
[20] J.C. Willems, 'System Theoretic Models for the
interconnected multiple equilibria systems. The Analysis of Physical Systems", Ricerche di Auto-
basic tools of stabilitytheory are thus seen to be matica,Vol.10.No.2, December 1979, pp.7 1-106.
applicable to general questions in qualitative [21] G. Zames, 'On the Input-OutputStabilityof Time-
behaviour of nonlinear systems. The recent re- Vatylng Nonlinear Feedback Systems. Part I:
sults in [ 1-51confirm their role. Conditions Derived Using Concepts of Loop
Gain, Conicity and Positivity', /E€€ Trans. Auto-
REFERENCES matic Control, Vol. AC- 11, pp. 228-238, April
C.I.Bymnes, A, lsidori and J.C. Willems, 'Passivity, 1966.
Feedback Equivalenceand the Global Stabiliza- [22] R.E. Kalman, P.L. Falb and M.A. Arbib.
tion of Mlnlmwn Phase Nonlinear Systems'. IEEE Mathematical Svstem Theorv, McGraw-Hill,New
Trans. Automatic Contrd, Vo1.36. pp.1228-1240, York, 1969.
1991 a [23] R.A. Rohrer, 'Lumped Network PassiVltyCriteria'.
A. Rodriguezand R. Ortega, "Adaptive Stabiiiza- I€€€ Trans. Circuit Theory, Vol. CT-15, No.1, March
tlon of Nonlinear Systems: the Non-feedback II- 1968.
nearizabb Case", Preprints of the 11th IFAC [24] P.J. Moytanand B.D. Anderson, 'Nonlinear Regu-
World Congress, Vo1.4, pp.121-124. lator Theory and an Inverse Optimal Control
AJ. van der Schaft, 'L,- gain Analysis of Nonlin- Problem", IEEE Trans. Auto. Control, Vol. AC-18,
ear Systems and Nonlinear State Feedback H DD October 1973, pp.460-464.
Control', IEEE Trans. on Automatic Control, (251 M.R. James, "A Partial Differential Inequality for
Vo1.37, Noh, June 1992, pp.770-784. Dissipative Nonlinear Systems-, submitted for
P.V. Kokotovic, 'Joy of Feedback:Nonlinearand publication.
Adaptive', 1991 Bode Prize Lecture, University of [26] M.G. Crandall, L.C. Evans and P.L. Lions, "Some
California, Santa Barbara, Technical Report Properties of Viscosity Solutions of Hamilton-Ja-
CCED-92-0207. cobi Equations'. Trans. AMS, Vd.282. No.2. April
A.L. Fradkov and D.J. Hili, 'Nonlinear Adaptive 1984. pp.487-502.
Control: An East-West View'". submitted to 12th [27] B.D. Anderson and S. Vongpanitlerd. pletworlc
World Congress of IFAC. Analvsis and Svnthesls. Prentice-Hail, Englewood
J. Guckenheimer and P. Holmes, plonlinear Os- Cliffs, N.J. 1973.
mica1Svstems.and Bifurcationsof I281 E.D. Sontag, "A 'Universal' Constructlon of Art-
Vector F W , Springer-Verlag,New York, 1983. stein's Theorem on Nonlinear Stabilization', Sys-
A. l s i d o r l , f l tems and Control Letters, Vo1.13, No.2,1989.
m, Second Edltion. Springer-Verlag. 1989. [29] M. Vidyasagar, L,-stabillty of interconnected
CA. Desoer and M. Vidyasagar, &@back Svs- Systems using a Reformulation of the Passlvity
Inaut-OupJt ProDerties, Academlc Press, Theorem", IEEE Trans, Circults and Systems, Vd.
New York, 1975. CAS-24, pp.637445, November 1977.

Authorized licensed use limited to: Peking University. Downloaded on December 28,2021 at 13:53:04 UTC from IEEE Xplore. Restrictions apply.