80 FEBRUARY IEEE TRANSACTIONS CONTROL] ON AUTOMATIC 1973 Time-Domain Criteria for the Lz-Stability of Nonstationary Feedback Systems &I. K. SUNDARESHAN AND hi. A. L. THATHACHAR Fig. 1. The feedbacksystemunderconsideration. of linear andnonlinear timeAbstract-Criteria for the L-stability varying feedback systems are given. These are conditions in the time domain involving the solution of certain associatedmatrix Riccati equations and permitting the use of a very general class of Lroperators as multipliers. I. INTRODKCTIOX The problem of deriving criteria for the &-stability of systems containing a linear time-varying operator and a memoryless nonlinearity in cascade, i n a negative feedback loop, is not amenable to simple treatment, owing tothe difficulty in obtaining positivity conditions for rime-varying operators. I<ecently, n’illems [ l ] , [2], a-5vie11 R X Estrada and newer [3], have obtained conditions for the positivity of systems with a statespace description. in terms of an associated matrix lliccati equation. Here, we follow this approach and give ronditions for the positivity and stability of time-varying systems. The poaitivity conditi~~n obtailled here is the same as that given ill [ 3 ] ,but the method of derivation is felt to be simpler and more direct and 1135 been inspired by the work of \Tillem:: [ l ] , [2]. Although the altalyris iscollcerlled with finite-dimensional singleinput single-outputsystems, the extension to the general case is straightforward. tctIT,andzerootherwise;~(.)EL.=)~~(.)EL~,’dTTE+]. Note that H E EA =) H has finit.e gain and HI,Hz E EA =) H = HIHz has finite gain. H E Q E n 81is said to be positive ( e ) [strongly positive ( e ) ] if the and inequality (UT(. 1, YHT(. 1) 2 (UHT(.), U H T ( . )), ‘plu H ( . E Ls ‘pl T E R+holds with e = 0 [ e > 01. System: The syst.em (Fig. 1) is described by t,he input-out,put relations e l ( . ) = u l ( . ) - w2( . ), e!( . ) = UZ(.) WI(.) wit,h w l ( . ) = Gel( .), G E 6~ n 6 r and WP(. ) = Se?( .), :V E %o.w, i.e., S:Lnc + L?. 3 S T ( . ) = n ( x ( . I), 0 I x ( . ) n ( ~ ( . )I ) s x 2 ( .1, V x ( . 1 E L e , and n(.) is odd and monotone nondecreasing. Kote that. S = E, the identit,!. operator in I,,,, yields a linear system more general than that considered in [ 4 ] . Problem: Given that u1(. ), z c n ( . ) E LPand el(. ), e ? ( .) E LZe, find conditions on G which ensure e l ( . ), e ? ( .) E L. The method of solution to this problem by factoring the open loop into two positive ( r ) operators, one of which is strongly positive ( e ) and has finite gain, is by now well established, and the introd u d o n of “multipliers” to render flexibility to this approach is well known [ 4 ] . + 111. 11. PROBLEN FORMULATIOX ,\‘otation and Definitions: Here, detailed definitions will be omitted as these can be found in an earlier paper by the authors [4]. Some new notation will be introduced. Let $ E denote the class of linear causal operators H in L?, with an ext.ernal (input-output) description, i.e., H E ZPE =) 3 a map h : K+ X R + K I s w h t h a t y ~ ( f )= H ~ c x ( t ) h ( t , i ) z c ~ ( dr )i , V U H ( . ) E L e , where U H ( . ) : R + -+ R is the input to H and y ~ .): ( R + + K is t.he output of H and h ( t , ~ = ) 0, V T > t . =io” + &EA C 6E 3 H E &Ed -}J= - L= (h.(t,r)(2dt d l is finite.2 Kote that H E =) H : Lz + L? [’i]. Let. 91denote the class of linear causal operators H in Lr, aith an internal (statespace) description, i.e., H E 61=) 3 AH( ) : R+ R n X K”, ~ H ( . ) : R+- Rn, C H ( . ) : R + I En,and d ~ ( . ) : & ++ R, such that H is described by t.he dynanlical equations --t Lemma: An operator H E dH(t) MAIN QIM RESULTS is positive ( e ) if: a) > 0, ‘d t E R+, (3.1) and b) there exists a real symmetric bounded nonnegative-definite n X n matrix RH(^) satisfying the Riccat.i equation kH(t) = ;dH-’(t)[cH(t) - RH(t)bH(t)I [ b H ’ ( t ) R H ( t )- CH’(t)] + - [ R H ( ~ ) A H ( ~Aa’(t)Rx(t)l. ) (3.2) H E Qr,w is strongly posit.ive ( e ) if there exists an e > 0 such that (3.1) and ( 3 . 2 ) are satisfied wit.h d a ( t ) replaced by [ d a ( t ) - E]. Proof: a ) Positiuify ( e ) of H: It is required t.0 prove that ( u d . ) ,Y H ~ . ) )2 0, v UH(-) E Ln. and ‘pl E R+. (3.3) + i H ( t ) = A H ( ~ ) z H ! t )b H ( t j t t H ( f ) ; y H (=t j c H ’ ( f ) z H ( f ) + where XH(.):R+ + R n i s the state of H and U H ( . ) , dH(t)ieH(t) IJH( (2.1) .) defined s above. c 61 3 H E t p I y = ) (2.1) is a minimal representation of H. I t is aell known [ 5 ] that H is uniformly reachable and uniformly observable =) H E 61.u. H E QE is said to have finite gain if + + ( $ z ~ ‘ ( t ) [ k ~ (Rt a) ( t ) A d f ) A ~ ’ ( t ) R ~ ( t ) l z ~ ( t ) = SOT isfinite [ET(.) is the truncation of x ( .) defined b>- z ~ ( t = ) z(t), + ra‘(t)RH(t)be(t)uH(t)- U H ( ~ ) C H ’ ( ~ ) T H (~ )d H ( t ) U z r 2 ( t ) } dt, (3.5) Manuscript reeeiwd July 3, 1972: revised September 13, 1972. l l l e authorsarewiththeDepartment of ElectricalEngineering.Indian Institute o f Sclence. llangalore 12. India. 1 K denotes t i l e real nrlml~ers;K - ,the nonnegarive real numbers: and Rm. the ~r-dirnen~i~nal Eurlidean space. 2 A s aninterestingaside. i t may I!e nutedtllatsuchoperators are termed “€lill~ert-8cl1mi~l~ operators” in the matllematics literature [ i .p. 541. since H E 61.u. If the RHS 5 0, then (3.4) is satisfied. Wewill show this by proving that, if RH(!)is a solution of (3.2), t.he integrand on the RHS I 0. Kow, the integrand on t.he RHS of (3.5) is 81 TECHNICAL NOTES AND CORRESPONDENCE <&H‘(t)[kia(t) + + R ~ ( t ) A a ( t+) A ~ ’ ( t W ~ ( t ) l z ~ ( t ) [zH’(t)Rrr(t)bH(t)UH(t)- U H ( ~ ) C H ’ ( ~ ~ H ( ~ ) sup a~(.)ELzc - dH(thLH2(t)l. (3.6) Now evaluate the supremum on the RHS by differentiating this with respect to u H ( f )and setting it equal t,o zero. Thus, on solving, we get. u x ( t ) = $dH-’(t)[sH’(t)RH(t)bH(t)- (3.7) CH‘(t)TH(t)l and, because of (3.1), this U H ( ~ ) ,for which the supremum is attained, exist,s. Substit,utingin (3.6) and simplifying, we have RHS of (3.6) = + + & ~ ’ ( t ) ( f ? ~ ( t ) RH(f)AB(t) AH’(t)Ra(t) REFERENCES + %dH-l(t)[RH(t)bH(f)- cx(t)l [ b x ’ ( t ) R d t ) J. C. WiUems. “Least squares stationary optimal control and the algebraic Riccati equation,” I E E E Trans. Automal. Contr. (Special Issue a n L i n e a r Quad:ptic-Guussian Problem). vol. .4C16, pp. 621-634. Dec. 1971. DissiDativedvnamicalsystems-Partsand I 11,”Electron.Syst. Lab.; Mass.-Inst. Technol., CambLidge. Tech. Rep., Nov. 1971. stabi1it.y of systems with a R. F. Estrada and C. a. Desoer Passivity and state representation,” Int. J. Cohr., vol. 13. pp. 1-26, 197.1. K. Sundareshan and XI. A. L. Thathachar, “L?-stablllty of l i n y timvarying systems-Conditions involving noncausal multipliers, IEEE Trans. Automat. Conlr.. vol. .4C17, pp. 501-510, +ug,. 1972. R. E. Kalman, P. L. Falb. and M . A . Arbib, Toptcs In Mathematical System Theory. ?Jew York:McGraw-Hill. 1?,69. H. D. Albertsonand R. F . \Vomack Minimum-stat.erealizations of linear t.ime-varying systems.” I E E E Trans.’Aufomat.Confr. (Corresp.), vol. AC-13. pp. 305-309,June 1968. N. I. hkhiezerand I. M. Glazman. Theory of LinearOperatorsinHilbert Space, vol. 1. .New York: Ungar, 1961. -. - cx’(t)l J S H ( t ) = 0, M. because of (3.2). Hence, the desired result follows. b ) Strong posiiivity ( e ) of H : It,is required to prove that (WIT(. 1, Y H ~)). v u H (.) E LZr, LHS = 1, w - 1) 2 0, r T E R+, and for some +d (uH~ - ),( C H ’ ( * ) ~ a r )( . ~ )UET( ( )) - = (UHT(.1, BHT(. )X where @x(.) = CH’(.)ZH(.) E > 0. E(UHT(. (3.8) 1, U H T ( . 1) + [ d ~ ( --) E]uH(.). Hence, (3.8) holds if H is positive ( e ) with respect to the new output g H ( . ), Le., with d ~ () .replaced by [dx(.) - e ] . Q.E.D. Using t,his lemma,the proofs of the following stability theorems are straightforward. Theorem 1 (Linear System): If t.here exists anoperator flri E n QrM such that ~ l f - 1 E 68.4, the composition L = M G E 8l.U and if the following condit,ions: a) dL(t) > E > 0, v t E R’, ddt) > 0, v t E R+; (3.9) b) t,here exist real symmetric bounded nonnegat,ive-definite matrices R L ( ~and ) R.u(t), solut.ions of the Riccati equations, B.L(~)= $ [ d ~ ( t) - R s ( f ) b ~ . ( t ) l [ b ~ ’ ( t) Rc~~ (’ (f t)) l E]-~[CL(~) + A ~ ’ ( t ) R d t ) l (3.10) - [Rdt)Adt) I t,his doesnot pose a serious problem,since, L being a linear operator, startingwithanarbitrary realization of L(i.e., L E O r ) , it is possible to arrive at a minimal realizat,ion t.hrough well-established comput,ational algorithms 161. However, if one were to follow t.his approach,certain precautionary measures need be taken.Note that, since a minimal realization would result in certain ext.ra st,ates to be removed, which would not, consequently, influence t.he stabilit.y conditions, it is necessary t o ensure a priori that these statesdo not contribute,by themselves, toinstability. One sufficient. condit.ion guarant.eeing this, for example, would be t,o require that f l f and G are globally asymptotically st.able, which implies limt,, s,w(f) = 0 and limt,, m ( t ) = 0. It. is to be emphasized t.hat t,he stability theorems should be applied only after reducing L to a minimal form. and B s f ( t ) = ;d.w-’(t)[cAr(t) - R.w(t)b~(t)l [b.w’(t)Rx(t) - ~.w’(t)l + A~r‘(t)R.ri(t)l (3.11) - [R.ri(t)A.dt) are satisfied, then t.he syst,em described by Fig. 1 is LTstable for all G E EA n 6r.w and N = E. Theorem 2 (AVonZinear System): If there exists an operat.or L M E 89~4n bru such t,hat,X - 1 E $ E A , L = MG E 8 r u , and the following conditions: Comments on “A Simplified Irreducible Realization Algorithm” R. D. GUPTA .4ND F. W. FAIRMAN In the above paper,’ Chen and Mital have provided a theorem t,hat enables a reduct.ion in the size of the matrices used in Ho’s algorithmforminimal realizat,ion. This is achieved by utilizing theadditional informationcontained in the degres of t.he least common denominators of the rows and columns of thetransfer function matrix G(s). Theorem 1’ is shown here tobe asimple consequence of known resu1t.s for t.he realizat.ion of single-input multi-output. and single-outputmulti-inputsystems. The proof following. given here isbased on Lemma 4 [I], which is restated as the L a m : If G(s) is a st,rictly proper rational matrix and t.he monic polynomial y ( s ) is the least. common denominator of t.he entries, gij(s), in G(s), t,hen ~ ( s is ) the minimal polynomial of the minimal realization of G(s). Consider a single-output multi-input. system. The transfer function matrix G(s) can then be written as(using the aut,hors’’ numbering of equations where possible) 0 G(s) = [ql(s),gz(s); . .,gp] = L ~ ~ L I S - ~ (1) i=l a) d ~ ( t > ) E > 0, tt t E R+; b)there exists a realsymmet,ric bounded nonnegative-definite matrix R L ( ~ solution ), of (3.10); and c) J!I = p~ z, E R’, z E 3 &7) = z(t I 4 v ) l dzJ < P , + are satisfied, then the syst.em described by Fig. 1 is &stable for all G E EA n ~ I and M M E XOU. Remurk: The condition of the theorems that L = MG E 6rJr Manuscript received July 31, 1972. This work was supported by the National needs some explanation. Alt.hough M,G E 61 =) L = MG E 8 1 , ~Research Council of Canada. R. D. Gupta is a CommonwealthScholaratQueen‘sUniversity,Kingston. it does not follow, in general, that, M,G E 6 r . v =) L E Q r u . But Ont., Canada, on leave from lladhav Engineering Col!ege, Gwalior,.India. A simple realizaticn of L may be obtained by defining r L ( t ) = [rG(t)zx(t)l’ vcW = u,w(l), ~ ( t =) u d ) , and YL(O = u d t ) . J F. \V. FairmaniswiththeDepartment of Llectrlcal Engineerlng,Queen’s University.Kingston,Ont..,Canada. 1 C.-T. Chen and D. P. Mltal, I B E E T r a n . Automat. Conlr. (Short Papers), vol. AC-17, pp. 535-537, Aug. 1972.