427 SHORT PAPERS - “On the input-output stahi1it.y of time-yaryingnonlinearfeedback systems-Part 11: Conditions Involving clrcles In the frerluencs plane and sector nonlinearities,” I E E E Trans. Automaf. Contr., 1.01. AC-11. pp. 465476, July 1966. I. Is7. Sandherg, “On theLrhoundedness of solutions of nonlinear functional equations.” Bell S y s t . Tech. J . , vol. 43, part 2, pp. 1581-1599, July 1964. - “Someresults on thetheory of physicalsystemsgoverned by nonlinear functional equations,” Bejl Sgst. Tech. J., vol. 44, pp. 871-898, MayJune 1965. E. I. Jury an:, B. IV. Lee ”The absolute st.abilitv of svstems with many nonlinearities Azltomat. iemote Contr.. vol. 26, pp. 9451966, .June 1966. 13. D. 0.r\ndkrson. “Stabihty of control systems with multiple nonlinearities,” J . Franklin Inat., vol. 282. pp. 15.5160. Sept. 1966. V. A. Yacubovich “Frequency conditions for the absolute shhility of control systems with ‘several nonlinear or linear nonstationary blocks.” Automat. Remote Contr.. vol. 28. pp. &30, June 1967. F. X. I3ailey, “Theapplication of Lyapunov’ssecondmethodtointerconnected systems ” S I d X J . Contr.. ser. .‘ vol. 3. pp. 143-462. 1966. R. F. Estrada On the arahility of multiloop feedback systems.” I E E E Trans. Automitic Control. r o ~ ~. ~ - 1 pp. 7 .781-791, Dec. 1972. F . R. Gantmacher. M a t r i x Theory. New York: Chelsea, 1959. VIJI. CONCLUSIONS The stabilityof multjloop systems is investigated here in a fashion which directlyincorporates the ‘Lstnwt.ure” of the system in the analysis. This can be accomplished because system stability can often be interpreted in terms of margins within which subloops are stable. The method of analysis is straight.forward, and a general analysis procedure is given. The theory is applied to t.wo specific multiloop systems. A system composed of two single-loop systems, each having a linear and a nonlinear element, in cascade with anouter feedback loop is examined. Conditions are derived which as the out.er feedback loop gain approaches zero reduce to the familiar circle criterion for each of the single-loop systems. &o a system is analyzed which consists of an interconnect.ion of three lineartime invariant elements, a linear t h e varying element,, a piecewise linear nonlinearity, and a hysterisis nonlinearity. Average Variation L2-Stability Criteria for Time-Varying Feedback Systems-A Unified Approach APPENDIX C o h l P L E T I O N OF THE P R O O F O F THEOREM 1 In order to show that thehypotheses of Theorem 1 are sufficient to I - [(bijIg(Hj)]has an inverse whose guaranteethatthematrix elements are all nonnegative, the following preliminary results (see, e.g., [ l o ] pp. 66 and 71 of vol. 11) are employed. Theorem A : 9 mat,rix A havingall elements nonnegativehas always a nonnegative eigenvalue r such that the moduli of all the eigenvalues of A do not exceed r. To this “maximal” eigenvalue r there corresponds an eigenvect.or y such that y 2 0 and y # 0 (refer t o Sect.ion I1 for notation). Moreover, the adjoint matrix B(X) = (AI - A)-llXZ - 81 has all elements nonnegative for X 2 r. Theorem B: If a mat.rix G has all off diagonal elements negative or zero and the successive principalminors are posit.ive, then all principal minors are positive. As in Theorem 1, assume that the successive principal minors of I - [Ih,,Ig(Hj)] are all positive. Since the last, successive principal minor is the det.erminant of t.his matrix, the matrix is nonsingular and has an inverse. From Theorem A, it is clear that. mat.rix [Ibi,lg(Hj)] has a “maximal” eigenvalue r. Further, the matrix (I [lbijlg(Hj)])-1!I- [lbiilg(Hj)]lhas all elements nonnegat.ive if r 5 1 . But [ I - [/bij1g(H;)]Iis t.he last successive principal minor of I [Ibijlg(Hj)]and is positive. Hence, if r 5 1 then Z - [!bijlg(Hj)]has an inverse with all nonnegat,ive e1ement.s. Iiow it only need be shown that r 5 1 . Since r is an eigenvalue, it is found that 31.K. SUNDARESH-U’T A N D 1.1.A . L. THATHACHAR. Abstract-The problem of developing L-stability criteriafor feedback systems with a single time-varying gain, which impose on the gain, is treated. A unified averagevariationconstraints approach is presented which facilitates the development of such averagevariationcriteria for both linear and nonlinearsystems. The stability criteria derived here are shown to be more general than the existing results. I. IKTRODUCTIOS Following the work of Freedman and Za.mes [ l ] in developing Lnstability criteria for linear systems containing a time-varying gain k ( t ) in an otherwise t.ime-invariant. feedback loop, by imposing averaging constraints on the 1ogarit.hmicvariat.ion of k ( t ) , there has been a considerable act.ivity in obtaining similar results for linear and nonlinearsystems. 3Iajor resultsin t.his direction are t.he absolute stabilit,y criteria of Venkatesh [ 2 ] for linear aystenw and of [3] for nonlinear systems. Venkatesh’s criKarendraandTaylor terion imposesaconstraint on k ( t ) more general t.han [ l ] in considering only the positive lobes of d log k ( t ) / d t for averaging. Hoaever, both [ 2 ] and [ 3 ] make use of Corduneanu’s ext.ension of t.he Lyapunov method for the derivat,ion of results and hence, do not o = 1r1 - [ I b i j l g ( ~ j ) l I= - [ I b i j l g ( ~ j ) l ( r - 1111. facilit,ate a straightforward extension t.o the L-stability problem. This means ( 1 - r ) is an eigenvalue of t.he matrix I - [Ibi,lg(Hj)]. I n a recent paper, the authors[4] have presented Ln-stability criteria Now the charact,erist,ic equat.ion for an n X n mat.rix B can be for linear systems which impose instantaneous bounds on l / k ( t ) writ.ten as dk(t)ldt, by employing the t.heory of positivity of compositions of causal operatorsandtime-vaving gains. The mainaim of the present paper is t o emphasize that [ 4 ] contains t.he core of a proIB - X I 1 = (-X)” Sp(-X)n-k = 0 cedure which inconjunction with asuit.ablefactorization of k ( t ) E=l yields anextension to averagevariation stability criteria. This is where each S k denotes the sum of all principal minors of order k of demonstrated here by deriving within the framework developed in the matrix B. Let.ting B = I - [lbijlg(Hj)]and X = 1 - r results in [4],stability crit.eria which are moregeneral than [ 2 ] forlinear n systems and [ 3 ] for non1inea.rsystems. (r - 1)” ~ k ( r 1)f-k = o + 1 + 2 + L=l where each Sk represents the sum of all principal minors of order k of I - [ l b i j / g ( H j ) ] .But from Theorem B it is clear t.hat. all principal minors of I - [Ibij!g(Hj)]are positive. Hence, each Sk > 0. Thus, the above charact.erist.ic equationcannot Hence r, 5 1. be satisfied for 7 > 1. REFEREWES [l] G.Zames, “On the inpuboutput stability of time-varying nonlinear feed- back systems-Part I : Conditions derived using the concepts of loop gain, conicity.andpositivity,” I B E E Trans. Automat. Contr., vol. AC-11, pp. 228-238, Apr. 1966. A. :vOtUtiQ,lS urd Dejinitiuns While t.he notations employed in [ 4 ] nil1 be folloxed, a few of the import,ant ones are briefly recapit,ulat.ed. Let,, R, R’, and J + denote; respectively, the reak, the nonnegative reds, and the nonnegat.ive integers. An operat.or H in L? (L?,) is a single-valued mapping of Ln Manuscriptreceived May 21973:revisedOctober12, 19i3.Paper recommended by R. A . S l t o o ~ .Chairman of t h e I E E E S C SStability, Nonlinear, and Distributed Systems Committee. The authors are with the Department of Electrical Engineering. Indian Institute of Science, Bangalore, India. 428 IEEE TR.LYSACTIONS CONTROL, O N AUTOMATIC (,he) into itself. If H is an operator in L e , then H is said t o be "positive(e)" [st.rongly positive(e)] if bhe inequalit.y ( x T ( .), ( H z ( .))T) 2 B ( z T ( . 1, TT(.)) holds wit.h 6 = 0[6 > 01, V x(. ) E I?,and V T E' R+, TT(.) being the truncation of z( .) defined by, z T ( t ) = ~ ( t V) t E [0,T]and zero otherwise. Letdenotethe Banachalgebrawith an ident.ity E, oflinear time-invariant causal convolution operators in that, are defined by where { r i ) ,i E J is a sequence inR +,{ h i ) ,i E J is an Zl-sequence in R and h ( . ) E L. H ( j w ) = h i esp(-jwr,) f?: h ( t ) esp(-jot) dt. An operator H E ( B c issaid to be "regular in (Be'' if H-3 E Cat. X is the clasof time-varying operators K in 4,defined by K z ( t ) = k(t)z(t)V T( . ) E I,,, 0 < inf k ( t ) 6 k ( t ) sup k(t) < m V t E R+. Let. I denote the identity of K . X @ c X 3 K E X@=) dkct)ldt 6 28k(t) for some 8 E K + V t E R'. &cE 3 K E X, =)dk(t):'dt 3 --2ak(t) for some 0: E R' V t E R+. With this notation, it is simple t o obtain the following result. Propositiolnl: If K E X@(Xs)has a decomposition K = e I I?, 0 < e < inf k ( t ) , then there exists a scalar @' > @ such that I? E J;@'(xa'jand the difference (6' - 8 ) ran be made arbitrarily small. ciEJ+ + + + < + B. System Descriptio11 The system (Fig. 1) is described by the input-output relations, e l ( . ) = ul(.) - w ( . ) , e?(.) = 7 4 . ) .al(.), w ( . ) = Gel(.),and w2(. ) = Ke,( . ) where G E 63,and K E 3;. + C . The JIai1t Problem Given that u 1 ( . ) , 7 1 1 ( . ) E L . 2 and e l ( . ) , e,(-) E I,?,, find conditions on G and K which ensure that the system is &stable, i.e., el( .), e?(.) E L . 111. PRELIMIKARIRESULTS The following lenunas give conditions for thepositivity(e) of and time-varying gains. compositions of causal operators in I, (I2?<) The proofs of these may be found in [4]. Lemma 1: Let P be a causal operator in I/>,and Q be an operator in L?, defined by & x ( t ) = q(t)r(t)V T(.) E L?,, q ( . ) : R ++ R is absolutely continuous on R+ and is bounded for all t E R+.Then QP is positive(e) if 1 ) P is positive(e) and 3) q ( . ) is nonnegative and monotone nonincreasing. Lemma 8: Let. P and Q be defined as inlemma 1 . Then PQ is positive(e) if 1 ) P is posit.ive(e) and2) q ( . ) is nonnegat.ive and monotone nondecreasing. Lemma 3: Let H E C a C sat,isfy Re H ( j w - 8 ) 2 0 V w E K and some 8 E R-. Then, KH is positive(e) for all K E X@ and H K is positive for all h- E KO. If H further satisfies He H ( j w ) 2 e > 0 V w E K , then KH is strongly positive(e) for all K E X@" and HK is strongly positive(ej for all K E Xy,where 8" < 8, the difference (0 - a")being arbitrarily small. ;Tote: The first part of Lemma 3, i.e., positivity(e) of KH and Hi? is proved in [4](refer to t.he steps from ( 5 . 3 ) to ( 5 . 5 ) in [ 4 ] ) . The proof of the strong posit,ivity(e) also follows similarly, after making use of the proposition in Section 11-A. IV. AIAIN RKSULTS The lemmas enunciated in the previous section will be crucially used in the derivation of the stabi1it.y criteria for the system under consideration. Stability Criteria for Liltear System Fig. 1. AUGUST The linear feedback system under consideration. for some T > 0 and some scalars a$ E R +,'d t denoting only the posit.ive lobes of the argument. %XaB 1974 E R +, with [ .] + c x 3 K E anx,@=) for some T > 0 and some scalal-s a,@ E R+, V t E X + wit.h [ . ] denoting only t.he negative lobe5 of the argument. The following lemmas depict certainfactorization properties of the operators belonging to these classes. Lennmma 3 : If h-E @,Xa$, t.hen K admits a factorization h-= K l K , such that K I E X,q and K , E X,. Proof: The proof follow along the same lines a s the proof of a similar factorizationlemma in Freedmanand Zanles [ I ] , by deadditive decomposition of fining I(t) = log k(t) andobtainingan I(t). This, hoaever, is contained in Ranlarajan and Thathachar. Lemma 5: If K E @nX,p, then K admits afactorization K = KlK, such that K ZE X, and K , E X@. Proof: The proof can be obtainedfroma slight, variat.ion of that of Lemma 4, by defining / ( t ) = -log k ( t ) . The following theorem gives a criterion for the Lstability of t.he system under consideration. a ,and constants ( ~ $ 1 Theorem 1: If there esists an operator X E C E R+ such that, ' 1) M is regular i n bc 2) Re M ( j w ) GGw) 3) I<c ,IrCjw - 2 (4.3) 6 >0V 8') G ( j w - 8 ' ) ER w 2 0 V (4.4) w ER (4.5) and 4) Re X C j w - a) 20V w E R, (4.6) then the system (Fig. 1) is &stable for all K E @,X,@ U &Jia@, where @ < p', the difference (8' - 8 ) being arbitrarily small. Proof: The proofwill be outlinedfor the case K E a p X a g only, since the other follows fromsimilararguments. Since K E =) K = K&,, Kl E 6 8 and K , E P , make the system transformations as shoKn in Fig. 2. An application of lemma 3 ensures the st.rong positivity(e) of X G K l from (4.4)and (4.5) and the positivity(e) of K J - ' from (4.6). The finiteness of gain of XGKI (for the definition of the gain of an operator, see [ 4 ] ) is also ensured since X , G E aCand I C ! E 3;. Hence, invoking a basic result due t o Zanles [6], Lrstability of the system ensues. Stability Criteria f o r :Yonliltear Systems In this section, we will consider a syst.enl having a different configuration (Fig. 3) from the one described earlier in Section I1 B, in having a nonlinear operator F in the feedback path. Let '5, denote the rlas of nonlinear operat.ors F in I,, defined by, F z ( . ) = f ( r ( .) ) V ) E L e , f(0) = 0, zf(r) 2 0 V r E R and [xl - x ? ] [ ~ ( I I-) f(+z)] 2 0 V rI,Z? E R. Let .&.II c 3.tf 3 F E &.v =)f(. ) is odd. Let. us introduce the follom-ing subclas of 3;. D e h e the following subclases of X : QpX& c x 3 K E @,X,@ =) 1' - ~ ( CU @) dr < 26 (4.7) (4.1) for some scalars 8 , E~ R + ++ t E R +. 429 SHORT PAPERS G M-1 w; Fig. 2. in (0 ) e;(*) u;(+MU~ Transformations of the system for Theorem 1. Fig. 4. Transformat.ions of the system for Theorem 2. time-varying systems with monotone nonlinearities. It. is of interest to observe in this context that, in spite of the ample atkention this problem has received ever since the appearance in the literature of O'Shea's [9]results for t.ime-invariant syst.ems, it remained unsolved even for the simpler and less general case of the imposition of instantaneous bounds on the logarithmic variation of the gain. [n] have also obtained C) Recently, Ramarajan and Thathachar Fig. 3. The nonlinear feedback system under considerat.ion. results similar t o Theorem 1 by folloning the approach of Freedman and Zames [l]. However, the present. met.hod of derivation is It may be shown, following Freedman [7], t.hatthis averaging different and has followed from an ext.ension of the techniques decondition is sufficient to ensure a useful fact.orization of K. veloped in [4] for the inst.ant.aneous bounds case, thus providing a Lemma 6: If K E @X@,then K admits a factorization K = K I K , unified approach for thederivcztion of aoerage variation criferia with such t.hat K CE X6 and K , E Xo, 3 i o being the class [Xa]6=0. various types of aLleraging constraints. In this context., it is interesting This lemma' lends itself useful in proving the following stability to notet.hat.the earlier invest.igators had toresort to entirelydifferent criterion. methods for t.he derivation of stability criteria wit.h bounds of either Theorem 2: If there exists an operat,or X E aCand a constant type on t.he gain (for exa.nlple compare the results of Karendra and p' E R + such t,hat,, Taylor in [3] and [8]; while the criteria in [SI which impose instanregular 1) M is aC, (4.8) taneous bounds were derived using the Popov approach, generalisation of these results with averaging bounds in [3] could be obt.ained only by employing Corduneanu's method). 2) = E z, E ac3 11z11= :zi; iz(t); dt < 1 D) It may be noted that thehypotheses of Theorem 1 differ isJ+ from Freedman andZames [ l ] only in t.he c0nstraint.s on k ( t ) (which (4'9) are more general). Hence, it is possible t,o state this result in purely 3) Re M ( j o ) G ( j w ) 3 6 > 0, 'd w E R, (4.10) geometrical t.enns, involving certain "shifted" Nyquist diagrams and dispensing wit.h t.he elrplicit w e of multipliers, following Freedand man [TI. x + z 4) Re M ( j w - p')G(jw + - 8') 3 ++ w 0, E R, (4.11) V. CONCLUSIONS then the system (Fig. 3) is &stable for all F E and all K E U X ~ , New criteria for the of containing a single where fl < B', the difference (0' - p ) being a r b i h r i l y small. time-varying gain k ( t ) in an otherwise time-invariant negative feedcOrollury: If in addition to the hypotheses Of t'he theorem, back loop, are developed bJ, inlposing averaging constraints on the E '+ and z ( t ) E R C t then the system gain, A unified approach for thederivation of such average variation is L s t a b l e for all F E 5.w and K E u X ~ . criteria forboth linear and nonlinear systems is delineat.ed. The Since E =) = KzK"9K 2 E xB.and K u E stabilitycriteriapresented here permitavery general c l a s of causal make thesystemtransformations as shown in Fig. 4. It is now operators in L2 to be used as multipliers and are more general than sufficientt.0 prove that. K2'N-l is positive(e) and MGKz is st.rongly the existing results. However, t,he results of the present findings posit,ive(e) wit.h finit.e gain. The former follows from (4.9) and can be improved by further investigation in developing similar Lemma 2, the latter folloTVsfrom (4.10), (4.11)7 and h n m a 3. average variation crite& which the of nonThe proof of the corollary mas' also be obtained from similar argucausal multipliers, analogous to the results of [4]for the instantaments. neous bounds case. '" " A Few Remarks A) In comparison criteria the with of the earlier investigat,ors, it is easy to realize that Theorem 1 is more general than t.he results of ~~~~d~~~and Zames Ill a,nd Venkat,esh [21,while Theorem is genera,l than Narendra and ~~~l~~ 131. It, should especially be noted that the technique employed in [2] and [3] has been to me ~ ~ extension of ~ the L d ~met,hod,~~which requires ,~ the multiplier ;Jf(s)to be a rational funct,ion of L ' s ) J and the sJ7st,em t o be described by a differential equation. Hence the present results aTe improve men^ upon a,nd [31in the follon,ing specis: 1) general multipliers employed, 2 ) a broader c l a s of systems considered, 3 ) less restrictive averaging conditions on t.he gain, 4)a stronger stabilit,y concept and 3 ) elegance of the met,hod of deriva- REFERENCES [I] M . I. Freedmanand if?;;;, ;;;;,stems G . Zames."Logarithmicvariation criteria for the withtime-varying s i n s . " SIA.II J Contr., vel. 5. PP. [z] T . V.Venkatesh."Glohalvariation criteria for the stahi1it.y of linear t i m e varying syst.ems." S I A M J . Contr;: vol. 9. pp. 431-440. 19i1. [B] J. H. Taylorand K. S. Narendra, The Corduneanu-Popovapproach t.0 the ~ stability ~ ~of nonlinear ~ ~timevarying ~~systems,'' ~S I A~M J . A p p l . Math.. ~ vel. 18, pp. 267-281, 1970. 141 &I. K . Sundareshanand M . .\. I,. Thathachar. "L?stal>ilit.y of linear timevarying systems-Conditions involving noncausal multipliers," IEEE Trans.Automat. C o d r . . vol. AC-17,pp. 5 0 S 2 1 0 , hua. 1972. [ti] 8. Ramarajan and M . .I\. L. Thathachar. L1-stal)ilits of time-=w'ing sysrems with glohal enndidone on the rime-varying gain." I n l . J . SWt. S e i . , [sl ~ ~ ~ ~ ~ ~ ~ ~ ~of time-varying ~ ~ l ~ nonlinear ~ ~ l ieedl,ack ~ s ~ g19';,3hz; wstems-PartI:Conditionsderivedueineconceptsof loop gain.conicity, and positivity," I E B E Trans. - 4 ~ O m Q l .COntr.. v.01. A G 1 1 . PP. 2 2 s 2 3 8 , ,\pr.1966. - "On the inpuc-output staldity oftime-varyingnonlinearieedhack tion. syste'ms-Part 11: Conditions involving circles in the frequency plane and B) An additionalnovel feature of Theorem 2 is the use of sector nonlinearities." ibid., vol. .\C-11, pp. 4 6 5 4 7 5 , July 1966. OK&ea t,vpe mult,ipliers wit,h integral constraints on t,he kernel, for [i] &x. I.Freedman,"Lrstahi1it.yoftimevaryingsystems-Constmction of multiplierswith prescrihed phasecharacteristics," SI.4.V J . Contr.. V O ~ . 6. DD. 559-5i8 1968. (81 K. $. Narenhra and J. 15. Taylor. "Lyapunov functions for nonlinear time1 I t shouldbenotedthatFreedman [7] hasobtainedaslightlydifferent varying systems:" Inform. Contr., vol. 12. pp. 378-393, 1968. = [ 3 & 9 ] ~ - ~Hoaever, ). [9]R. P. O'Shea, h combinedfrequency-timedomainstability criterion for factorioat.ion. K = K z K uK~I 5 Xoand K&XB (where 6 alsofolloas from (4.7>, the proof heinsautonommlscontinuolls systems." I R E 6 Trans. A u t o m i . Contr.. vol..\C-ll. thefactorizationgiveninlemma 171. after defining Z(t) = -log L(1). pp. 177184,July 1966. similarthat given to in r a i