TP06 18:OO Proceedingsof the 37th IEEE Conference on Decision & Control Tampa, Florida USA- December 1998 Exact Solution to the Non-Standard Hm Problem Mohamed Seddik Djouadi Center for Intelligent Machines, and Dept. of Electrical Eng., McGill University Montreal, Quebec, Canada Corresp. address Dept. of Electrical Eng. Wayne State University 5050 Anthony Wayne Dr. # 3100 Detroit, Michigan, USA 48202 djouadi@ece.eng.wayne.edu, djouadi@cim.mcgill.ca functions defined on dD and D with values in the spaces defined above. Abstract The non-standard H w problem po = infQEHm 11 IW(1 - poQ)I + IVpoQI IIw Plays a fundamental role in robust feedback control. Here we develop a new operator theoretic framework to characterize explicitly its solution. We also establish absolute continuity of a dual extremal measure showing that the optimal performance/controller satisfy an extremal identity, which provides a test of optimality. Moreover via Banach space duality theory we show existence of a maximal vector leading t o an explicit formula for the optimal controller. 1 Introduction In this paper we consider the non-standard H w optimization known as the “two-disk” problem defined as Notation C stands for the field of complex numbers, G x n denotes the space of n x n matrices with entries in C. q,,, is the space of 2n x n-matrices with the norm IIAJJ= lAll + lAzl where A = (2) Al,2 n x n matrices, and IAll is the largest singular value of Al. C J , ~ ,is the same as GnXn but with the norm IlAll = rnax(STr(Al),STr(Az)), where ~ )Cj”=, oj(Ai),and STr(Ai) = T r a ~ e ( { A ~ A i }= aj(Ai) is a singular value of Ai. q, stands for 2n x 1 with the norm IIEllq, = l&l I&I, where + ( Et ) E U?, and 1&l is the Euclidean [= , norm. Q, is the dual space of q,. D is the open unit disk, aD the unit circle, and m stands for the normalized Lebesgue measure. * denotes either the complex conjugate transpose of a vector or an operator depending on the context, and “Tr” is the abbreviation used for trace. We will use freely the standard spaces ILP and W , 1 5 p 5 00, of 0-7803-4394-8198 $1 0.000 1998 IEEE 2843 where Po E G x nrepresents the plant, W , V are outer functions in H w ( G x n ) weighting functions, all defined in the unit disk D (see [l,2, 31). It is shown in [4, 5, 31 that under certain conditions the optimal robust disturbance attenuation problem (ORDAP) reduces to (1) as well as problems involving questions of well posedness of feedback systems [6].Zames and Owen in [2, 31 used Banach space duality theory t o show that under specified conditions there exists a solution t o (l),which satisfies a flatness or “allpass” condition and is unique in the SISO case. However the operator theoretic structure of the solution has still not been described and there has been no closed form solution reported in the literature essentially because (1) is more difficult than the standard two-block H w problem. We will exhibit an explicit solution via operator theory for the optimal performance and the optimal controller. Moreover these are shown t o satisfy an extremal identity, which can be used as a test of optimality . Part of this paper extends our scalar case results [7, 8, 91. Following [3, 21, the following assumptions will be made throughout: (Al) (W*W+V*V)(eZ‘)> 0, V8 E [0,2n), and (A2) the outer part of Po is invertible in H W ( G x , ) . The optimization (1) has been shown in [l, 2, 31 t o be where [GIE @.(Gnxn) is the equivalent class of G, @,(Gnxn) denotes the subspace of the Hardy ' (G,,,) given by space W equivalent to c where U is inner, W and are outer in H w ( G x , ) . Let IHIw(C&,,) denote the Banach space of bounded analytic functions in D with values in ,&a , with the norm IJKII= ~ ess. + Ilc2(eie)I) SUP (IKl(eie)I e~[0,2x) ( 2 ) E W@'(C&,,). where K = (3) Then (2) is (4)is the distance from ( K(Gnxn) denotes ( ) t o the subspace ) H'(@nxn) of W(C&,,). As- sumptions ( A l ) and (A2) ensure closedness of 8 . According to [l,2, 31 with a slight modification 8 has an equivalent description given by JWGraXrl). It has been shown in [ l l , 1, 2 , 3 ] that Lm(U&,,,) is isometrically isomorphic t o the dual space of IL1 (G,,,). Since Woo (q,,,) may be viewed as a subspace of lLW(C$,,,), and is the preannihilator of WDO (G,,,), then a standard result from Banach space duality theory asserts that MIm (q,,,) is isometrically isomorphic t o the dual and ) we write: space of Q ( G n x n W""(G,,,) Then, every K = =(@(Gnxn))* ( ) defines a linear bounded Tr(K;Gl Let I S be the subspace of where R*R = I almost everywhere (a.e.). Existence of solutions is next obtained using duality, and some of the ideas in [2, 31. 2 Duality Structure of the Problem and Existence of Optimal Solutions Let IL1 (CA,,,) denote the Banach space of weakly Lebesgue measurable [lo], and absolutely inte,,,-valued functions defined on the unit grable CA, circle d D under the norm where G = max(STr(G1), STr(Gz))(eie)dm z: ( ) E IL1 (G,,,), (8) functional +K defined on H;(@znxn) by +K([G]) = IlGllel = the space obtained by E(&,,,) equivalent to S = and taking complex conjugate of all functions in ls = + KlGz)(e")dm (9) (CA,,,) defined by ((I- RR*)IL~(C~,,,) CB RE(G,,))/x (10) where X = E ( G n x n )n((I-RR*)lL1 (C~nxn)@RE(Gxn)) (11) It follows from lemma 1 of [2, 11 modulo minor details that l s is the preannihilator of 8 in @(&,,,). Theorem 2 ([12], page 121), implies the following Theorem 1 Under assumptions ( A I ) and (Ai?)), there exists at least one optimal Q, E H w ( G x , ) such that: (6) and STr(Gi) = Tr({G:Gi}*) = Cj"=laj(Gi), i = 1, 2; where "Tr" denotes the trace. aj(Gi) denotes the singular value of Gi [l, 2, 31. Next,&t ilKf,(C~,~,) be the quotient space IL1 ( C A ~ ~ ~ ) / E @ , un-( C A , ~ ~ ) der the quotient norm In the next section, we show that if we assume further (A3) W , W ,RI, R2 are continuous, as is the outer spectral factor of W*W+c*c,and p, > p,,, where max (STr(G1 + gl), STr(G2 + g ~ )( e) i e ) d m 2044 i.e., when the open unit disk analyticity constraint is the space of G x n is removed, and C!(Gxn) valued continuous functions on d D , (in the scalar case ( n = I ) , po0 = II min(lW(eie)l, Iv(eie)l)IIm), the optimal solution is flat [3, 1,2]),then the supremum in (12) is always achieved. 3 Absolute Continuity and Existence of an Extrema1 Identity for the Optimal Performance and Controller Lemma 1 Under assumptioms ( A l ) , (A2) and (A3), the following hold: Let C!(C:nxn) denote the space of continuous functions on d D with values in Q., Iff E we set Proof: the first equality follows from Lemma 2 [l,21, the second from Theorem 1 ([12], page 121). It turns out that the maximum in (22) is actually achieved by a measure U, absolutely continuous with respect to the Lebesgue measure. Then it will follow by the Radon-Nikodym Theorem that the supremum in (12) is achieved in 'B. The dual space of is given by the space M ( @ a n x n ) of Ginxn-functions of bounded variations on d D under the norm [l,2, 31 Lemma 2 Under assumptions (Al), (A3), the following holds lo,inl ( zi ) ll4llw = where max(STr(Gy,l),S T r ( G y , l ) ) ( e i e ) d w y E M ( @ i n x n ) , and wy denote the sum of the total variations on [0,2n) of all entries of V I and UZ. Let be the space of Gnxnvalued functions which are continuous on d D and is known as the disk alanalytic in D. a(@;nxn) gebra. The annihilator of a(@;,,,) is completely described by the vector-valued version of the F. and M. Resz Theorem [13] which yields A(Gnxn)l = {P E C(Gnx,) : /I = H m , (16) E $&zxn)) It follows that the dual space ( A ( G n x n ) )= * M(@znxn)/w:(Gnxn) (17) Next, define the subspaces 8 , and by 8, = ([ld], chap. IV), we say that ( ",",$' ) is a dual extrema1 function. Proof by lemma 1 there exists [U,] E S i , II[v,]lllw 5 1 such that (22) holds. If we let wo to be the total variation of the entries of v,, then a similar argument to the proof of Theorem 2 in [l,21 shows that w, is such that V E Bore1 set, E C d D , w,(E) = 0 implies m ( E ) = 0. Hence by the Radon-Nikodym Theorem there exists F E IL' (w,, G2,xn) s.t. 8 ne ( q n X n ) d m = Fdw, (20) In the following lemma we establish that the dis- ( ) t o 6, is the same as t o 6. 3F, E IL1(Gdnxn) (25) (24) and (25) imply dw, = 0, and therefore U, is absolutely continuous w.r.t. m and hence (23) holds. A well known fact in the theory of HP spaces of scalar valued functions is that every coset in L' / H A contains one representative of the least poscible coset norm [16]. It is not very hard t o show this dw, = dw,+dw, 6* = {U E M ( C z n x n ) : dU = ( I - RR*)dP @ RG, E W @ z n x n ) , G E $AGxnl/Y (24) However by the Lebesgue decomposition Theorem for vector measures [15], there is a measure w, absolutely continuous with respect to m, and a singular measure w, s.t. (18) It follows then from lemma 3 [l, 21 that the annihilator of 8, is given by tance from Following V respectively ( ( I - R R * ) M ( @ i n x n )@ R E i ( G x n ) ) P (A2) and 2845 = F,dm+dw,, 4 Exact Solution via Operator Theory property holds for @(C2nxn)and is inherited by ‘6. This is summarized in the following lemma which proof is omitted (see [17]). Lemma 3 0 4.1 A Key Multiplication Operator Let L2(C&) stand for the Banach space of C&,valued functions defined on dD under the norm: Given [F] E @(C,,nxn), there exists h, E m;(Gnxn) s . t . where G = ( “,: ) . . (q,). Note that L2(qn) E L2 is not a Hilbert space. Let CP = Given [ F ]E‘ 9, there exists xo E X s.t. ( $:) E l”(C&xn), the multiplica- tion operator associated with CP and mapping the standard Lebesgue space L2(Cn)into IL2(qn) is denoted by M a . More precisely Ma f = CP f for f E L 2 . The next proposition, is a generalization of Proposition 2 obtained for the scalar case in [8, 91. in this case we say that F + x , is an extremal kernel for PI. Combining lemma 2 and 3, we obtain the following Theorem which looks familiar in the theory of extremal problems, and provides a test of optimality. Theorem 2 Under assumptions ( A l ) , (AZ) and (A3), Fo = ( 2: ) kernel for [ F ] , and and only if E Qo E [F] E’ 8 is an extremal H m ( @ n x n ) i s optimal if Proof omitted (see [17]). ( 2; Tr{((W*,O)+ QP*) )}(~Ze) I(W - RiQo)(eie)J+ 1R2Qo(eie)I max(STr(Fol),STr(Fo2))(eie), ax. Proof llFoll~1 Next we characterize the dual space of IL2 (C&) and ndL (q,)where , nd2(@2*,) to be the closed subspace of L2(q,) consisting of 11 . Ilazcs;,) bounded of analytic functions in the unit disk D . = “Only if” by assumption 3F0 E’ 1, and Qo E H M ( G x n ) s.t. (28) 9, Proposition 2 Let I L 2 ( Q n ) stand for the Banach space off&, -valued functions defined on d D under the n o m : + where G = but the integrand is “5” (IW - RieQol 1 R2QoI) (eie) max(STr( Fol),STr( F o 2 ) ) (e ) a.e. which is 5 po max(STr(Fol),STr(Fo2))(eie) a.e. Integrating implies equality must hold throughout. This combined with “flatness” imply max(STr(Fol), STr(Foz))(eie) = 1 a.e. “If” follows by integrating (as), (see [17] for details). zl ( ) E !L2(C2n). Define l@(C,ln)to be the analogue to @(C&) but in IL2(C2n).Then we have the following: 1. L2(CJn) N (lL2 (U&))* ) (IL2(C2,))* 2. JL2(q,= 3. W C 2 n )21 (W@2*,))* 4.IHf2(qn) N (nd2(C,2n))* Hence all these Banach spaces are reflexive. 2846 For the reverse inequality, by lemma 2 and 3 and Proof 1. and 2. follow from [ll].3. and 4. follow since Q,, is finite dimensional (see [17]). In the next section we characterize the optimal solution in terms of an operator which is analogue to the Sarason operator for the standard HM problem. This generalizes the scalar case results obtained in [7, 8, 91. Theorem 2, there 3 F = H2(Cn) €1 B ofL1-norm 1 s.t. 4.2 Operator Theoretic Solution Let II be the orthogonal projection operator on the closed subspace @(C&) 8 RH2(Cn)of HL(C&). Then II is a linear bounded operator on w(C&). Now we define the following key operator [7, 8, 91: Z: (2) +@ (C&) e RH2(Cn) by the proof of theorem 2, = 1 a.e. m, then max(STr(Fl), STr(FJ))(eie) there 3h E H 2 s.t. lhI2 = max(STr(Fl), STr(&)) a.e., and llhllHz = 1 [HI. But then F h E L 2 ( Q n x n ) , and IIF~IIL~(c~,.~) = 1, , orthogonal moreover F h E ( R H 2 ( G x n ) ) Ithe complement of R H 2 ( G x n )in L 2 ( 2 ~ n x n )Now . let I, be the n x n identity matrix, then Jd (34) In the following proposition the form of the orthogonal projection II is computed explicitly. 27T 27T Tr(hI,(W*, 0 ) F h ) d m Tr((W*,0)F)dm = (41) Exploiting an idea of Young [19], the LHS of (41) is equal t o Proposition 3 II = I - RP+R*, where P+ is the standard Riesz projection, and I the identity map 1171. I(PM( 7 )hIn,Fh)l L llPM The operator E is then a combination of Topelitz and multiplication operators, more precisely Z = ( I - RP+R*) (V (35) where P = II@ I,, M The following Theorem generalizes the results obtained for the scalar case in [7, 8, 91. Theorem 3 Let p o be the performance index defined by expression (1). Under assumptions ( A l ) , (A2) and (A3), p o is equal to the operator induced norm of E,namely denotes the tensor product, M the multi- plication operator associated to ( 7 ) , and (., .) denotes the inner product in lL2 (GnX,). Then, P o = 11=.11 (36) Moreover there exist at least a maximal vector f E H 2 ( C n ) of L2((cCn)-norm1 such that PM( =.( :)@In (43) llPM f W ) II = IlEll (44) therefore IlEf l l L z ( q n ) = 11=.11 (37) ( 0 1 Proof: VQ E H w ( G x n ) ,and all g E H 2 ( C n )we have Q g E H 2 ( C n ) and then IIRQg = 0. Let f E H2((Cn) with norm at most 1, then the first inequality follows from (see [7, 91 for the details) 11E.11 = sup IIfIILz(,n min )<I 1(7) Expressions (391, (42) and (44) imply that (36) and (37) hold, for some vector f E L 2 ( C n ) ,I l f l l ~ z p )= 1. The following corollary follows then from theorem f - Rg/l (38) (3) and ProPosition (3). Lz(qn 1 Corollary 1 The optimal performance index po is given by the following expression (39) 2847 [4] J.F Bird and B.A. Francis. On the robust disturbance attenuation. Proceeding of IEEE Conference on Decision and Control, pages 1804-1809, 1986. . and there exists f E H'((@"), I l f l l L z p , = 1, such that [5] B.A. Francis. On disturbance attenuation with plant uncertainty. Workshop on New Perspectives in Industrial Control System Design, 1986. 4.3 Optimal Controller Theorem 1 implies that there exists a vector function Q E Woo ( G n x n such ) that: 1 1 Q 1 1 ~ = 11Z.11 [6] M.C. Smith. Well posedness of Hw optimal control problems. SIAM, 28(2):342-358, 1988. [7] M.S. Djouadi. Banach space optimization of uncertain systems in Hm. Ph. D. Research Proposal, Department of Electrical Engineering, McGill University, May 1997. (47) and by Theorem 3 there exists f E H 2 ( C n )of norm 1 such that, IlEf 11 = llEllllf l l H a = 11Zl). By Proposition 1, 4 can be viewed as a multiplication operator n/r, acting from H 2 ( C n ) into fl(Gn)such that, ~~MQII = Ilqllw-. Then [8] M.S. Djouadi and G. Zames. Operator theoretic solution to the optimal robust disturbance attenuation problem. Proceedings of the World Multiconference on Systemics, Cybernetics and Informatics, 3:25-29, July 1997. Il=llllf l l H 2 = 1 1 3IIw(qn)= llIIQfllEP(qn) I llQf llw(qn)5 11~11w- Ilf = Il~llllf 11H2 IIHZ [9] M.S. Djouadi and G. Zames. On optimal robust disturbance minimization. Proceedings of the the American Control Conference, June 1998. since the orthogonal projection II acting from fl(@, ") onto W (C&)0 RH2( G x nhas ) induced norm less than or equal 1 [7, 91. It follows that, Qf = Sf, m a.e. In the scalar case we can divide through, therefore we have \E = m a.e., and hence by Theorem 3: [lo] E. Hille and R.S. Phillips. Functional Analysis and Semi-groups. AMS, Providence, R.I., 1957. y, [ll] J . Dieudonnge. Sur le thQorkmede Lebesgue Nikodym V. Canadian Journal of Mathematics, 3:129-139, 1951. [12] D.G. Luenberger. Optimization by Vector Space Methods. John Wiley, 1968. [13] R. Ryan. The F. and M. Riesz theorem for vector measures. Indag. Math., 25:408-412, 1963. [14] J.B. Garnett. Bounded Analytic Functions. Academic Press, San Diego, New York, Boston, 1981. [15] J . Diestel and J.J. Uhl. Vector Measures. AMS, Providence, RI, 1977. and the optimal Qo is given by the following expression ( Y ) - R Q o = -Sf f + Q~ = R* ( ) - R Zf* ~ m, a.e. thus the controller CO= Q o V - ' ( I - QoV-lPo)-' achieves optimal robust performance. [16] V.P. Havin. Spaces Hm and L1/HL. Journal of Soviet Mathematics, 39:120-148, 1974. References [l] J.G. Owen and G. Zames. Robust disturbance minimization by duality. Systems and Control Letters, 19:255-263, 1992. [17] M.S. Djouadi. Optimization of Highly Uncertain Feedback Systems in Hw. PhD thesis, Dept. of Electrical Eng., McGill University, 1998. [18] K . Hoffman. Bounded Spaces of Analytic Functions. Dover, New York, 1988. [19] N.J. Young. The Nevanlinna-Pick problem for matrix-valued functions. Journal of Operator Theory, 15:239-265, 1986. [2] 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. [3] J.G. Owen. Performance Optimization of Highly Uncertain Systems in Hm. PhD thesis, Dept. of Electrical Eng., McGill University, 1993. 2848