Proceedings of the 3 P Conference on Decision & Control Phoenix,Arizona USA December 1999 MIMO Robust Disturbance Rejection for Uncertain Plants M. Seddik Djouadi School of Aerospace Engineering Georgia Institute of Technology Atlanta, Georgia 30332-0150 seddik .djouadi@ae.gatech. edu measurable, and p t h power absolutely integrable X valued functions on the unit circle 3 9 under the norm Abstract In this paper, we investigate in a unified framework two basic feedback optimization problems, rejection of output disturbances in presence of plant uncertainty, and contraction of radius of uncertainty for MIMO plants. Duality theory is provided showing existence of optimal feedback laws. Optimal, nearly optimal solutions are characterized, and flatness conditions are derived. The theory developed here allows previous convex programming based numerical methods to be applied to this class of problems Notation C stands for the field of complex numbers, If x E C then Z denotes the complex conjugate of x. For an n-vector E &, where @, denotes the n-dimensional complex space, I<l is the Euclidean norm. G x ndenotes the space of n x n matrices A with entries in C, and IAI is the largest singular value of A. A* denotes the complex conjugate transpose of A. and @2“,“: denote the complex Banach space of 2n x n where m is the normalized Lebesgue measure, f E ILP(X), and 11 Ilx denotes the norm on X . H p ( X ) , (or I”(X)),1 5 p 5 CO, is the Hardy space of X valued analytic functions on the unit disc 9,viewed as a closed subspace of L P ( X ) (or resp. lLP(X)). “Tr(A)” denotes the trace of the matrix A, and “ess sup” the essential supremum. If B is a Banach space, then B* denotes the dual space of B. < cnxn spectively the following norms where 11 . llnuc is the nuclear norm, and Gnudenotes the complex Banach space of 2n-vectors <, = ( ); <I, <Z < E G with the norm IClmax = m4lC11, IC211 1.1 Rejection of Output Disturbances in Presence of Plant Uncertainty In this problem a stable LTI plant P known to belong to a weighted ball in H w ( G x n ) + V ) = ((1 V x ) p o : X E H w ( 6 x n ) , IIxIlw < 1, po E Hm(G,xn), v*’E (2) Hm(Gxn)} (1) [l, 21, and viseversa since GnXn y d qz;,, are finite dimensional and therefore reflexive. If X denotes a finite dimensional complex Banach space, L P ( X ) , (or ILP(X)), 1 5 p 5 CO, stands for the LebesgueBochner space of o 1999~ E E E It is well known that there are mainly two reasons for using feedback, the first is to reduce the effect of any unmeasured disturbances acting on the system, the second is to reduce the effect of dynamic uncertainty. The objective of this paper is to investigate the ability of feedback to reduce uncertainty by analyzing two optimization problems. The first is the MIMO version of the optimal robust disturbance attenuation (ORDAP) [3, 2, 41, and the second is a similar problem which captures the potential of feedback to contact radius of uncertainty [3]. It may be worth recalling these two problems in some details. B(po>o) cnxn is the dual space of qz:n 0-7803-5250-5/99/$tO.00 1 Introduction 4050 is subject to disturbances at the output (see Figure (1)). The objective is to synthetize a robustly stabilizing feedback control law C for the set B(P,,V), which minimizes the weighted sensitivity norm IlW(l+ PC)-’Ilm over all P E %(Po, V ) , where W E Hm(Gxn) is outer [5]. This problem is equivalent to the opti- lu @- uncertain Plant \ stable Filter . controller Figure 2: TwwDegree of F'reedom Feedback Control Scheme Figure 1: Feedback Control in Presence of Plant and Dis turbance Uncertainty mization [3, 21 po = inf Q E H"(@nxn) IIVpoQllm 5 1 x SUP E~ ~ ( IlW(1- P,Q)(I + VXpoQ)-'Qll~ ~ ~ < ~1 ) l Following [3, 21, the potential of feedback to reduce plant uncertainty c m now be formally stated as the minimization of the "worst-case" Wlweighted closed-loop multiplicative uncertainty radius ~ ~ p p ~ S ( p ,11,Wvl)Amllm over all robustly stabilizing nominal plant invariant control laws. Mathematically, l this ~ problem l l ~can be translated into the following minimization [3] (3) In [2] it was shown that p0 is equal to the smallest fixed point of the function x : IO, 00) e [0,00)defined as a MIMO version of the "two-disk" problem po = inf Q E H"(Cxn) IlVpoQllm 5 1 x SUP E Hm(Cxn) l l ~ l<l1~ IlW:(I- poQ)(I+ VXpoQ)-'VXIIm (7) It was shown in (Lemma 2.2 [2]) that expression (7) is equal to Po = inf Q E Hw(Cxn) IlVpoQllm I1 SUP x EH~(C,.~,) llxllm< 1 + IlWi(I- K Q ) ( I VXpoQ)-'Vllm 1.2 Attenuation of Plant Uncertainty In his seminal paper [3], Zames posed the following question regarding the ability of feedback to reduce the effect of plant uncertainty: if a plant lies in some uncertainty set, say %(Po, V), what is the smallest radius of any set of closed-loop uncertainty that can be achieved by a single feedback control law ? To avoid the trivial answer to this question that zero closed-loop uncertainty can be achieved by disconnecting the system from the input (see Figure (2)), it is suggested in [3] to adopt the invariant plant scheme, i.e., the nominal closed-loop system is normalized to be equal to the open-loop system Po,where Po is the nominal plant. From Figure (2), we have then ( I + cPo)-'u= I (5) The closed-loop map K of the two-degree of freedom control scheme above belongs to Hw(GXn), and can be expressed as K - P, (I + P,Q)(I + AZJQ)-'AP = AmPo, where Q = C(I + PoC)-' (6) Expression (8) is then mathematically identical to (3). We henceforth consider the two-disk optimization (4). In a similar vein to [6], we develop a duality theory for (4), showing existence of optimal feedback laws. We then characterize the optimal and nearly optimal solutions. We also give a flatness condition for the optimum. Finally, we demonstrate that the optimal controller satisfies an extremal identity. The theory developed here allows the convex programming based numerical methods discussed in [7, 21 to be applied to the MIMO extension of ORDAP (4). 2 Duality Results The optimization problem (4) . _can be shown to be equivaient to 121 po =QEHiqf C X n ) 405 1 ess sup eE[O,2n) max 51 C E C = where A P = P- Po, and Am E Hw(@nxn) represents the closed-loop multiplicative plant uncertainty. (8) (IW - ~ ~ Q ) ( e " ) C+I IvQ(eie)CI) (9) where t?l and E H m ( @ n x n ) are outer functions, and U E Hw(@n,,) is inner [5]. Hw(Gnx,)) Define I L w ( ~ n x n (resp. ) to be the Banach space consisting of 2n x n essentially bounded (resp. analytic) functions defined in the unit disc 9, with values in the space Gnxn, endowed with the norm K=( 2) Theorem 2 ([8],p.121) implies that there exists at least one optimal parameter Q, E Ha(&,,) (.i.e. one optimal controller CO)such that Then (9) can be rewritten as Expression (11) is the distance from subspace S = ('. ) Hm(Gxn) of ( ) to the W(Gnxn). We po = ess inf QEHOD(Cnxm) sup eE[o,zr) max 1 ~ 51 1 C E G (IW - kQ)(e")CI + IfizQ(e"kI) assume throughout: (Al) (W*W + p v ) ( e i e ) > 0, VO E [O, 27r). Assumption (Al) ensures closedness of S. According to [2, 71, the subspace S has an equivalent description given by where R is inner. q;:,)denote the Banach space of absolutely Let lL1 ( Lebesgue integrable q;;,-valued functions defined on the unit circle 89,under the norm Remark 1 This predual description shows that the convex optimization based techniques discussed in [2, 71 can also be applied to the MIMO extension of ORDAP, but with diflerent matrix norms. Namely, the matrix norm IlAll = m= IiI 5 1 (IAlCI + lAZC1) (21) 6 E G It has been shown in [2, 71 that lLm(GnXn) is isometrically isomorphic to the dual space of lLi (q;;,). Again, define @ (q;:,) to be the subspace of the Hardy space Hi (q;;,) given by { F E E' (G:En) : /a, F(e")dm = 0) 0 (14) and @(Cj:;,) denotes the space obtained by taking complex conjugate of all functions in l€$,(Q:;,). Next, let @(q;",) be the quotient space lLi (q;;,)/@,(Cj;;,), under the quotient norm V[GI E L' (GEn)E(GZn) CB R z i ( G x n ) ) / X ( ii ). Since the norm (21) is difler- entiable on the unit sphere of C,,, it can be computed using calculus. 3 Allpass Property of the Optimum: Alignement in the Dual and Extremal Identity for the Optimum In this section we assume (A2) W, 21and RZare continuous_onthe p i t circle, as is the outer spectral factor of +w + v*v. Let C!(q;;,) denote the Banach space of continuous &,xn-vdued functions defined on the unit circle under the norm (15) may be viewed as a subspace of Since WO(&,,) Loo(G,,,), and $ ( ~ ~ ~ , is , ) the pre-orthogonal of (G ),,, in IL1(q;;,)then , W O ( & x n ) is is* metrically isomorphic to the dual space of q(q;;,). Therefore, the pre-orthogonal of S is given by [4, 61 S = ( ( I - RR*)L1(G;:n) where A = F=( 2) It will be shown that the dual space C'(q:;,)* is isometrically isomorphic to a space M(c:znx,) consisting (16) 4052 of qn:gn-valued complex bounded measures defined on the unit circle. Accordingly, let U = (2 ) Lemma 2 Under assumptions ( A l ) and (A2) thew ezists at least one optimal Q o E H W ( G x n )such that E M(cznxn)l and in- x troduce the following bilinear form on M(&xn) e ( q S n1 < v , >=/ ~ Tr{K;dvl(o) +~ , * d v ~ ( o ) } , W K ) K I , ZE e(&,,) (23) This form has the following equivalent representation: let w y be the sum of the total variations on [Ole)of all entries of v1,2. By the Radon-Nikodym Theorem, there ,wU), r = 1, 2, exists a vector function GY,PE L1(GXn such that (23) is reduced to < v,K >= Next define poo= + Tr{K;GV,i K,*Gv,z}dwv(B) (24) IWr) inf Q € e ( Gx n ) [I ( ) - AQll (33) N i.e., when the open unit analyticity constraint is removed. We show that the optimal solution in this case is flat or allpass, and an approximate flatness condition holds. The norm on M(Cznxn)is now defined to be Lemma 1 Proof: Since GnX,is the dual space of CjE;,, by Lemma 3.3 [2] and Theorem 2.10 191. Gnxnis finite dimensional, hence reflexive and it follows that G;;,, is the dual space of-Gnxn. Then by Singer’s Theorem (p.398 Theorem 1 Under assumptions; ( A l ) (W*W + v * Q ) ( e i e ) > 0, Vo E [0,_2?r); an! W , f i 1 , fiz, and the outer spectral factor of w*W + V*V are continuous on the unit circle. If p, > p,, then i. Any optimal Q, E Hm(@,.,) in (11) satisfies the flatness or “allpass” condition [lo]) W G n x n ) = (e(G%n))*. Assumption (A2) means we are working in the space of &,xn-valued functions which are continuous on the closure of the unit disc 3,and analytic inside D. Call this space A(C:znxn).The F. and M. Riesz Theorem [ll]characterizes the orthogonal of A(b(c2nxn) as A(&nxn ii. If {Qn}ris any sequence HW(C,,,,) ={A E ~ ( t 2 n x n ): x = Hm, $(GZn)} H E The dual space of A(C,,,,,> such that (35) (27) is then the quotient space A(Gnxn)* = W C Z n x n > / N G n x n ) ’ then (28) under the quotient norm for v E M(C2nxn) The annihilator of S, = S n e(G,&) is then given by & = {U E M(&nxn) :dv p E ~ ( G n x n ) ,G E R i ( c x n ) ) / * where *= ( ( I - + = ( I - RR*)dP RG, (30) where 1.i.m is the limit in quadratic mean. The condition po > poo is sharp. Proof: i. Let 4, be the extremal functional corresponding to the extremal measure v, which achieves the maximum in (32), i.e., =< U,,. >, where $,(a) +RE;(Cxn))n$(C?Zn)(31) dpo = Gdw(B),G = E L*(G,%,w). Let be The following Lemma is then a direct consequence of the continuous linear extension of qj0 to Lw ( G n X n ) Theorem 1 (p. 121 [8]). 4053 RR*)iW(bnxn) ( zt ) 4, Theorem 2 Under assumptions ( A l ) (WW + v * v ) ( e i e ) > 0, VB E [0,2?r), (A2) W , R I , Rz are continuous on the unit circle, as is the outer spectral factor which vanishes on S,. Then fie= II( ) -fiQeII ( 2; ) of W*W+pv, and p, > poo,F, = E [Fl E 9, II[qjl~=. 1 is an extrema1 kernel for [ F ] , and Q , is {T+(W - ii1C?,)'G1 + T r ( f i z Q o ) * G z } d w 4 0 ) - Ilom 5 1 optamal af and only if + I{Tr(W - i % Q o ) * G ~ Tr(RzQe)'G,}dwu(0)l [0,2=) Ld max IC1 51 (I(W - filQO)(eie)CI + li%Q.(eiB)CI) T ~ { ( ( w * ,+ oQ ) W)( 2; )}(eie) = C E C Proof: The proof is similar to the proof of Theorem 2 [Cl that - 4 Properties of the Optimal Behavior of MIMO OIWDAP Suppose now that E is a Bore1 subset of 853 such that $ ( E ) = 0 and rn(E) > 0. Define the following essentially bounded matrix valued function R* = Giii*(eis), =o, Then &(k)= 0 = for for 0 E E eE / Tr{GiG3(eie)}drn (40) In this section we generalize properties of the optimal behavior of the ORDAP obtained in [2] for SISO systems to MIMO systems. Note that for the MIMO extension of ORDAP, V is scaled by a parameter T , then assumption (A2) is strengthened to (A2') The outer spectral factor of W*W r2Q*Qis continuous for any positive r , and the outer part of Po is invertible in Hm( G x n ) . Recall that in the MIMO case ORDAP assumes the following form + (41) E which implies that G3(eie) = 0 rn a.e. in E , but since m ( E ) > 0, G3 must be identically zero. Therefore the maximum in (32) is achieved on the set {U E M(c2nxn) : u(0) =. J[i,e)(I- RR*)du'(B), U' E ~ ( ~ 2 n x n ) }which / ~ is the annihilator of &'(@nxn) implying by duality po = poo and then contradicting our assumption, . and thus i. must hold. ii. Follows from the same argument used to prove ii. Theorem 2 [7]. (43) Theorem 3 1. Under assumptions (Al), @*I@ + Q*v)(ei6e) > " E [o,2.rr)), and above. If p, > 0, then there exists an optimal feedback control V ) such that law stabilazing every system in %?(Po, lXloO Remark 2 The argument used in the proof of i. above shows that the extrema1 measure U, is absolutely continuous with respect to the Lebesgue measure. Hence the supremum in (20) is achieved at some function F E ( ) $, II[F]llg = 1. Therefore the optimum -RQo satisfies the following Theorem analogoui to Theorem 2 [6], and which provides a test of optimality for feedback controllers. 4054 IlWl(I - P o Q ) ( I + XVJ'oQ)-'llm SUP X E Hw(C,n) 1 =po (44) 2. If in addition po > boo,then there exasts at least one optimal controller which stabilizes evew system in %?(Po: V ) and achieves (43). I n this case, t h e magnitude of the weighted sensitivity function for the nominal plant Po satisfies + I W ( I PoCO)-'(eaB)I = po - p,1vp,~,(1 (eie)l, rn a.e. + poco)-' (45) application of Theorem 1 are met. Thus, Q , satisfies max IC1 I 1 (I(W - &Qo>(eis)CI + lhQo(eie)CI)=PO CEG V8 E [0,2n) proving that (44) holds, and we are done. References [l] R. Schatten. Norm Ideals of Completely Continuous Operators. Springer-Verlag, Berlin, Gottingen, Heidelberg, 1960. [2] J.G. Owen. Performance Optimization of Highly Uncertain Systems in HO". PhD thesis, Dept. of Electrical Eng., McGill University, 1993. [3] G. Zames. Feedback and optimal sensitivity: Model reference transformations, multiplicative seminorms, and approximate inverses. IEEE l'hnsactions on Automatic Control, AC-26(2):301-320, April 1981. [4] M.S. Djouadi. Optimization of Highly Uncertain Feedback Systems in Hm. PhD thesis, Dept. of Electrical Eng., McGill University, 1998. [5] H. Helson. Lectures on Invariant Subspaces. Academic Press, New York and London, 1964. [6] M.S. Djouadi. Exact solution to the nonstandard Hw problem. Proceedings of the IEEE CDC, December 1998. [7] G. Zames and J.G. Owen. Duality theory for MIMO robust disturbance rejection. IEEE Transactions on Automatic Control, AC-38(5):743-752, May 1993. [SI D.G. Luenberger. Optimization bv Vector Space Methods. John Wiley, 1968. 191 H. Chapellat and M. Dahleh. Analysis of time varying control strategies for optimal disturbance rejection and robustness. IEEE lhnsactions on Automatic Control, 37(11):1734-1746,1992. 2. By 1. there exists at least one optimal controller which stabilizes every system in B(P,, V ) . Therefore x ( p o )= p o = inf QEHm(Gxn) ess sup max ee[o,zn) IcI 5 1 C E G (IWI- PoQ)(e")CI + polVPoQ(eie)CI) implies x(po) = p, = ess SUP max worn 161 5 1 (IW(I- PoQo)(eie)CI C E G + poIVPoQO(e")Cl I Assumption (A27 implies that (A2) holds for r?l and r v for each positive r . Hence all conditions for the 4055 [lo] I. Singer. Sur les applications lineaires integrales des espaces de fonctions continues. Revue Roumaine de Math. puws et appl., 4:391-401, 1959. [ll] R. Ryan. The F. and M. Riesz theorem for vector measures. Indag. Math., 25:408412, 1963.