256 IEEE TRANSACTIONS ON ALTOMATIC CONTROL. VOL. AC-27. KO. I.

FEBRUARY 1982

are introduced. The problem considered below is the robustness of the The case of 3'? f O is treated by noting that it is equivalent to the situation
configuration of Fig. I(b) when the plant is subjected to linear feedback when r E Yo is replaced by r - !V21.'E I,. Stability is therefore unaffected
perturbations of the typediscussedaboveandthe linear feedback f is by ..V, and (20) holds nith x replaced by I . - LC?.'?I.'. i t . . we obtain
simultaneously replaced by F + X where jV is a memoFless map of Y into
itself of the form

where ,VIand 3; map Yo into itself, N , has finite incremental gain k l such
that

which is simply (14). T h s completes the proof of the theorem.

RemurX 5: If. as is usuallythecaseinnonlinearstudies.acausality
and h':,O=O and ,V2 is bounded in the sense that there exists a scalar 4 a 0 structure exists, the result holds with all vectors and operators replaced by
such that .VI?- C Yo and their truncated counterparts. The proof is trivial.
In conclusion. a partial answer to the problem of robustness of feed-
backsystems to unstableplantperturbationshasbeenprovidedby
Considering stable linear feedback perturbations to the plant. The results
Theorem 2: If thefeedbacksystem of Fig. l(b) is stable.then the are a generalization of those underlying kno\vn design techniques [ 141 and
system is alsostable inthepresence of simultaneouslinearfeedback also enable the effect of nonlinear feedback perturbations on both stabil-
perturbations of the plant and nonlinear perturbations of thefeedback ity and response characteristics to be assessed.
dynamics if conditions i) and ii) of Theorem 1 are satisfied and also

Moreover, under these conditions. the responses J' and I , ' of the real and
perturbed feedback systems are related by the inequality

Proqf.. The feedback relations of the perturbed system are

which can be written as

or. bearinginmindthefactthat (4)implies that ( I -L<.K ' exists

in Ib.

This equation has a unique solution ).'E Yo for each choice of r E l h if

+
( I f., K - ' H ) - 'L,.N is a contraction. Taking initially the case of ,V2 = O
(ie., 4 =O). this is clearly the case nith contraction constant p as

A Note on the Stability of Large Scale

Systems with Delays
using (4). We can in factverify (14) inthiscase bl- choosing the first
iterate .,6 = O in the successive approximationsequenceandnotethat IL HONG SUH AND ZEUNGYALI BlEN
!;=(I+L,K-'H)-'I.,~=(I-L,K-'H)-'?.~~~~~
Ahstrocr -A stabilio test is proposed for large scale systems with delays
by employing both the aggregation technique based on a Lyapunov function
and the stricti? quasi-diagonal dominance.
Using the relationship ( I + L, K I N ) - ' ,I G ( 1 - A ) -
~ and the triangle
inequality then yields I. INTRODUCTION
Recently. there are a number of research works on the stability of the
largescalesystems.Callier er ul. obtaineda necessayand sufficient
conditionforthelargescalelineardistributedsystems to be stable by

Manuscript recelred March In. 19x1: rexked lune 26. I981 and August I?. I981
The authors arc with rhc Dcpartmcnt of Electncal Sclcnce. Korea .Ad\anccd Institute of
Scwnce and Technologr. Seoul. Korea

0 0 1 8-9~X6/82/0200-0256So0.75 c1982 IEEE

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. AC-27, NO. 1, 257
FEBRUARY 1982

usingthe
concept of inputoutput stability incorporated with the W I J = - ( ~ ~ , - ~ , , ~ , ) ~ , J + ? I , , E ~ J ( I - ~ , / ) ( 10)
hierarchicaldecompositiontechnique [l]. Siljak [2]derived a sufficient
condition for the large scale systems with memoryless nonlinear interac- where 6,, denotes Kronecker delta symbol, and q,,,9,:. n,, and E,, are
tionsto be stable by employingtheaggregationmethodbased on the given in (5)-(9), respectively. If w,, < O and W is strictly quasi-diagonal
Lyapunov function.wheretheaggregatedsystem was assumed to be a dominant, then ( I ) is stable.
linear autonomous system with no delay. In [3],Moylan and Hill tried to For theProof of the Theorem I . we first consider thestability of a
unify the approach of input-output stability and the approach based on specific class of linear autonomous system with delays whose dynamics is
the Lyapunov theory by developing the concept of dissipativeness. In [4] given by
and [5]. i t was found that the time delays causeno particulardifficulties in ,A'
determining the stability of a specifictype of largescalesystemswith
delays in interactions.
i,(t)=-u,lz,(r)+h,,z,(r-T,)+ 2 a,,zJ(t-hiJ). i=l.2;..,N
j = l
In this note, a stability test is proposed for large scale linear systems Jf,
with delays, which is applicable for a more general class of large scale
systemsthanthoseconsideredin [4]and 151. For this, theaggregated
system considered in the note is chosen to be a linear autonomous system where a , , >O, h,, and a , , are real constants, and 7; and h,, are positive real
with delays. constants which denote the time delay. We propose a sufficient condition
Throughout the note. A' and ! j a II will denote the transpose of a matrix for the stability of (1 1) in the following:
A and the Euclidian norm of a finite-dimensional vector a , respectively. Lemma I: If there exist positive scalers d, i = 1,2.. . . X . such that for
I .
all i
11.STABILITY OF A LARGES C A L E SYSTEM WITH DELAYS
Consider a large scale system with delays in state and interconnections
whose dynamics is given by
.v
2
x ' , ( r ) = A , s , ( r ) l a , s I ( t - 7 ; ) +A,,x,(r-h,,), r=l I 2 ,... , x
then (10) is stable.
The proof of Lemma 1 is straightforward if the Nyquist array technique
/=I
/#I in[7] IS employed.Hence,theproofis omitted. It is remarked that in
[4],[5], a sufficient condition for the stability of the case when hii = O in
(1) (1 1) was given as in Lemma 1.
where X,E R", is a state vector, A , is a stable 7zi X n, constant matrix. is a, Proof of Theorem 1: To determine the stability of (I). let a positive
an T I , X n , constant matrix, A,, is an n , X n, constant interconnection scaler function Y, in (2) be Lyapunov functions for ( I ) . Then it is easily
matrix, and h,, PO and T, P O denote timedelay.Let G, be a positive s h o w n t h a t f o r i = 1 , 2 ;... N a n d f o r a l l [ E R .
definite symmetric matrix, and let H, be a positive symmetric matrix such
that ~ , ( r ) = - 2~ v ~ l ( r ) l , ~ ~ l l i ( ~ ~ ( f ) G , . ~ , ( t ) )
A:H,+H,A,=-G,. (2)
+ ( ~ ~ ~ ( t ) ~ ~ H ~ ~ ~ ~ ~ ~ s , ( r ) ) ~ ~ , s ( r - ~
It is noted that since A; is assumed to be stable, for any positive definite M
matrix G , , there always exist a positive definite matrix H, satisfying (2) [2]. +(V;1(f)IIH~1iIH,X,(t))7 2 A,,.xJ(r-hi,) (13)
As in [X], let v, be a positive scaler function given by j=l
J f l

v , ( t ) = ( ~ ~ ~ ; ' ~ l . ~ ~ ( ( t ) ~ , x , ( t )i )=' 1' ,*2 . - . . , ~ . (3) N

~ - q , , ~ , ( t ) + q , ~ ~ ~ k , ? r , ( r~ -~ ~ 2
~ ,, ,). ~x ,~( r+- ~~ ,~~ ,) ~~
From ( I ) , consider the interaction-free system with 2, = O given by 1'1
jfr
ii(t)=A,x,(r), i=1,2;..,N. (4)
( 14)
Then it is easily shown from (3) and (4)that, for any t N
~-q,*Y,(f)+q,,~,l~.~,(r-~)Il+q,, 2 E,/ll.Y,(f-h,,)ll (15)
llx,(t)il~vi(~)~q,,ll.~,(~)ll. (5) j=1
J'I

II H,-' I1 -, 'V
2 VI (t)(.l'(t)Glx,(t))Pq,zY,(f)r (6) ~-'li,Yi(t)fl),l",Y,(r-7;)f9,1 2
J=I
tljYJ(r-htJ). (I6)
and /#I

llgradv,(t)li~q,, (7) From (16). let the aggregated system of (1) be given by
.v
where q,,= I#'H,)/uf1'(
( H , ) and q i 2 = u,,,(G,)/20,~(H,). Here unz(H , )
and o,~,(H , ) denote the minimum and maximum eigenvalueof H , , respec- ci,(t>=-q,24,(t)+q,17,4r(r--.)+tl,l 2 CiJ4j(t-h,,),
J=I
tively. Let t,, and 7,be defined by J f i

E!, = '.v2( (8) i=1,2;..,N. (17)

and I t is shown in [2]that by the comparison principle, if (17) is stable. then

~l)isstable.In(17).byletting~,,=~,,.b,,=q,,~~.anda,,=q,~~,,.andby
7,= U , ; p ( a;a,>. (9) Lemma 1, we can conclude that if W is strictly quasi-diagonal dominant,
( I ) is stable. This completes the proof.
It is now sholvn that thetime-delay actions cause no difficultiesin It is remarked that stability condition given in Theorem 1 is the same as
determining the stability of the system in (1). the stability condition for thecase xvhen h,, =O and A i = O for i . 1 =
Theorem I : Consider a large scale system described by (I). Let A , be a 1.2, . ., in ( I ) [2].Thus, we can conclude that time delays cause no
.?.I,

stable matrix for i = 1.2:. ., .Ar. Let 12' X rl; constant matrix W = ( w , , ) be particular difficulties in determining the stability of large scale systems
given by Nith delays.
258 IEEE TRANSACTIONS ON ALTOMATIC CONTROL. VOL. AC-27, NO. 1. FEBRUARY 1982

111. CONCLUDING REMARKS FIXED

11. CHARACTERIZING THE DECENTRALIZED MODES
-ROUGH BLOCKDIAGONALLY
DOhllNANT MATRICES
The results in this note may be further applicable for the synthesis of
decentralizedstabilizing controllerfor largescale control system \vith Consider the set of interconnected linear dynamical subsystems
time delays xvhich was considered in [7].
.\-
.~,=A,,.~,+B,u,- 2 A,,.Y] i = ..~. . . K
REFERENCES j = l
J'I

F. M. Callier. W. S. Chan. and C. A. Desoer. .'Input-output stahilltg of Interconnected j; = C,.x,

system u i n z decomDositions: an improvedformulation." I E E E T r u m Auror1~11.
eonrr.. vol. LC-23. pp. 150-162. 1978.
D. D.Sdiak. Lum-Scale D&rtanirc Swerns. Srnhrbrr und Stn,<rure. New York: where x,€ R",. u , E Rr,. y,E R m t . The equations ( I ) can be written in
Elsevier. 1978. matrix form as
P. J. Molyan and D. J. Hill. "Stabihty criteria for large s a l e s>stcms." I E E E Tram.
Auronlar. Conrr.. vol. AC-23. pp. 143-149, 1978.
B. D. 0.Anderson. "Time delays in l u g e scale systems." to be puhllshed 1=A s + Bu
P. J. Molyan. "A connective stability result for interconnected passive systems." I E E E
Tram. Auronlar. Cnnrr.. vol. AC-25, pp. 813-813. 1980 J = CX
.
0. D. I. Nwokah. "A note on the stability of multivariable system.." l t x 1.Comr b o l .
31. pp. 587-592. 1980. where
M. Ikeda and D . D.Siljak. "Decentralized stabilization of large scale systems u i t h
time delay," in Pruc. I7rh Allerron Cot$ Circuir Swr. Theory, Unib illinois. 1979. pp.
64-72
A = { A , , , j = l . ...,:V.j = I , ....: V } E R " " "
J Bernussou and J. C. Geromel. "Stability approach to rohustcontrol for intercon-
nected systems.'' in Large Scnle S ~ ~ r e n rEngineering
s Apphcarrons. M Singh and A.
Titli. Fds New York: North-Holland. 1979.
B=blockdiag(B,.B 2.....Bv)ER""

C=blockdiag(C,.C:.....C,)ERn'X"
s ?X' .v

Procedure to Eliminate Decentralized

n= 2 n, r= 2 r, M = 2 m,.
,=I ,=I i=l
Fixed Modes with Reduced Information Exchange
The triple ( C . A . B ) in ( 2 ) is assumed to becontrollableand
V. A. AREvLENTANO AND M. G . SINGH obsen;able.Consider also the set

Abstract-When decentralized fixed modes occur in a set of intercon- ~={K~K=b1ockdiag(K,,.K,,:~-,K,,~).K,,€R'~'*"'~,~=

nected subsystems then complete pole assignment or stabilization (if they
By applying the decentralized output feedback law
are unstable) is not possible. In this brief paper we presenta way of
choosing a new structure for the feedbackmatrix such that the fixedmodes u t = K,,?,
are eliminated and such that the exchange of information among subsjs-
tems isreduced. The fixed modes are characterized here by means of block Lve obtain the closed-loop system given by
diagonally dominant matrices. The procedure is illustrated on hvo e l m -
ples. .t = ( A - BKC)\.

= Cx. (3)
I. INTRODUCTION
The set of decentralized fixed modes of ( C . .A, B ) is defined as
The concept of fixed modesin decentralized control systems \vas
introduced by Wang and Davison [ I ] and they correspond to the system
eigenvalues that are not affected by decentralized feedback.
A(C.B.B.K)= n c[A-BKC] (4)
KE I;
The fixed modesconstitute a very importantconcept since the! de-
termine \vhether a system can be stabilized by decentralized control. The where 0[.4 + BKC] denotes the set of eigenvalues of A + BKC. Let the
presence of unstable fixed modes prevents stabilization. Also. the presence submatrices in the diagonal of ;1+ BKC be denoted by .i,l.
where
of any fixed modes does not allow complete pole assignment.
The fixed modes also play an important role in determining whether a
strongly connected system [2] can be made controllable and obsenable
from a local controller and a local output.
Then. .4 - BKC can be represented ac
Noxv. we know that the fixed modes are completely determined by the
triple (C, A . B ) and the structure of the feedback matrix. We call the fixed
modes decentralized if they arise as a result of decentralized feedback.
In largescalesystems.thedecomposition intosubsystems is often
arbitrary. Clearly. a decomposition w-hch yields fixed modes under de-
centralized feedback is not desirable. In this paper \ve present a construc-
tive approach to changing the original decomposition in order to remove
the fixed modes. This is done by changing the pattern of communication
of information. i.e., by changing the structure of the feedback matrix. If the diagonal submatrices A , , are nonsingular and if
The basic criteria used in defining the new decomposition is to rcmove
fixed modes and. if possible. also reduce the information exchange.
The characterization of fixed modes is done hereinterms of block
diagonally dominant matrices [3]. The characterization provides a way of
choosing a new feedback structure.Examplesarealso provided to il-
then .4 + BRC is strictly block diagonally dominant.
lustrate the procedure.
The following theorem is taken from [3].
+
Theorertl I : If the matrix .1 BKC is strictly block diagonally domi-
Manuscript received December 15. 1980: rcvwd >la?. 12. l Y X l and Jul\ X. 1981. nant. then .A + BKC is nonsingular. The connection of this theorem with
The authors are with the Uniwrsity of hlancheztcr In4itutc d S c ~ n c cand Tcyhncdos>.
klanchcstsr. England the decentralized fixed modes is made through the following corollary.