- Proceedings of the American Control Conference San Diego, California June 1999 Optimal Robust Disturbance Attenuation for MIMO Uncertain Systems in H” M.S. Djouadi School of Aerospace Engineering Georgia Institute of Technology Atlanta, Georgia 30332-0150 seddik .djouadi@ae.gatech. edu Abstract Banach space, LP(X), (or ILP(X)), 1 5 p 5 03, stands for the Lebesgue-Bochner space of measurable, and p In this paper, we consider the optimal robust disturbance th power absolutely integrable X-valued functions on the attenuation problem (ORDAP) for multi-input multi- unit circle 8 9 under the norm output (MIMO) uncertain plants. Duality theory is used < Ilf1lEP(x)= ro.zT) I l f ( e ” ) l l ~ d m , for to show existence of optimal feedback laws. Next a key ., , multiplication operator acting on particular vector-valued Ilf(eae)llx, for p = m (4) I l f l l L ( X ) - ess 1 3sup Hardy spaces is introduced. It is then proved that OR-.~ 1 0 . 2 ~ ) DAP-for-MIMO systems is equal to the operator induced where rn is the normalized Lebesgue measure, f E norm of a specific operator. The latter is shown to be a ILp(X), and 11 . I I x denotes the norm on X . HP(X), (or combination of multiplication and Toeplitz operators. An W ( X ) ) , 1 5 p 5 CO, is the Hardy space of X-valued an“infinite matrix” representation with respect to a canonialytic functions on the unit disc 9 , viewed as a closed cal basis is derived, and the norm of the relevant operator subspace of Lp(X) (or resp. U’(X)). “Tr(A)” denotes is approximated by special matrix norms. the trace of the matrix A, and “ess sup” the essential supremum. 1 I I Not ation C stands for the field of complex numbers, If z E C then 1 Introduction Z denotes the complex conjugate of 2. For an n-vector C E C,,, where C,, denotes the n-dimensional complex The optimal robust disturbance attenuation problem denotes the space (ORDAP) was initiated by Zames in 1979 (see [3]), and space, IC1 is the Euclidean norm. G X n of n x n matrices A with entries in C, and IAl is the largest considered by Bird, Francis, Owen and Zames [4,5, 6, 21. and q::n denote the complex In the single-input single-output (SISO) case, Bird and singular value of A. Gnxn Francis proposed a two-block H m approximate solution. A2 A1 A1, Owen and Zames in [6] showed that the approximations Banach space of 2n x n matrices A , A = obtained this way can be arbitrary poor. Duality was A2 E G x nwith , respectively the following norms then applied to show existence of optimal solutions, and IlAll = (IAlCI + IAZCI) (1) approximations based on convex programming were proIC1 5 1 posed. In this paper, we consider ORDAP for MIMO uncertain systems. First, we give the duality structure of llAllnuc= i n f { x ICkllWklmax, C k E G , (2) the problem showing existence of optimal solutions. A key k multiplication operator is then introduced, and its norm W k E eF : AU = C ( < f U ) W k , VU E G} is shown to coincide with the norm involved in ORDAP. k Next, we prove that optimal performance for ORDAP in where 11 . [Inuc is the nuclear norm, and qfl denotes the the MIMO case is equal to the induced norm of a specific operator. The latter is a Banach space projection of a Cl, complex Banach space of 2n-vectors C, C = multiplication operator. The projection is then explicitly C2 E C,, with the norm computed showing that our operator is a combination of multiplication and Toeplitz operators. This leads natuIClrnax = maX(lClI,lel) (3) rally to an infinite matrix representation of the operator is the dual space of C?n::n [l, 21, and vise-versa with respect to a canonical basis, and its norm is apsince Gnxnand q ; k n are finite dimensional and there- proximated by special matrix norms. Part of this paper fore reflexive. If X denotes a finite dimensional complex extends our scalar case results [7, 8, 91. ( ), ( ;); 0-7803-4990-6199$10.00 0 1999 AACC 1856 2 v Problem Formulation In ORDAP a stable uncertain plant is subjected to disturbances at the output. The objective is the synthesis of a feedback control law which provides the best possible uniform attenuation of output disturbances in spite of uncertainty in the plant model [3, 4, 5, 61. More precisely, let P be a stable LTI uncertain plant, lying in a set of uncertainty described by the expression + x B(Po,V) = {(I VX)Po : E Hrn(Gxn), IlXllW < 1,Po E H r n ( G x n ) , V*' E H r n ( G x n ) } (5) where Pois the nominal plant, V is a fixed stable invertible transfer function, and X is a variable stable transfer function. The output disturbances d are assumed to lie in the set We seek a feedback control law C for the set %(Po, V) which minimizes the weighted sensitivity norm IlW1(I PC)-lIla. This is equivalent to solving the optimization problem given by + po = inf C stabilizing all P E B(Po,V) IIW1('+PC)-'llm sup (7) ~EB(Po~V) + = inf ~ ) IlWi(I - poQ)(I+ VXpoQ)-'Qllm (8) The optimization (8) is non-linear and non-convex. However, Owen in [2] showed that (8) is equal to the smallest fixed-point of the following function in a parameter T , which is convex in Q QEHm(Cnxn) Then (11) can be rewritten as (13) is the distance from ( 7 ) to the subspace S = (F ' ) H W ( G x nof) W'(Gn,n). According to [2, 61, the subspace S has an equiva!ent description given by S RH"(Cxn) where R = ( 1: ) is inner. where 3 SUP Q E Hw(Gxn) x E H ~ ( G X IlVpoQll- 5 1 lIXIlm < 1 inf C E G = where W1 is outer in H ' ( G x n ) . Letting Q = C ( I P,C)-l, (6) can be expressed in the form [3, 21 Po Here W , T&', E H a ( G X n ) are outer functions, and U E H'(@nxn) is inner. Following [2, 61, we assume throughout: (Al) (T&'*r?l+ V*8)(eie)> 0, ve E [o, 2 ~ ) . Define lLa(&.,) (resp. Hm(Gnxn)) to be the Banach space consisting of 2 n x n essentially bounded (resp. and analytic) functions defined in the unit disc D, with values in the space Gnxn, endowed with the norm for ess sup eqo,zr) max 1 ~ 51 1 C E G (IWl(1- PoQ)W(e")CI + rlPoQV(e'e)CI) Let lL1(q:;*) denote the Banach space of absolutely Lebesgue integrable q:;,-valued functions defined on the unit circle aD, under the norm lzr1 ( llGllL1(c~~Gn) = Cnxn) ess sup ea0.2~) (IWl(1- PoQ)W(e")CI (9) ) (I nuc dm (14) { F E IHI' (q:En) : /d'" F(eis)dm = O) (15) and $,(q:;n) denotes the space obtained by taking complex conjugate of all functions in @(q::n). Next, let E$(q:Gn) be the quotient max 1 1 15 1 66G + IPo&V(e")CI) GI (eie) Gz(e'e) It has been shown in [2] that lLM(Gnxn) is isometrically isomorphic to the dual space of lL1(qzcn). Define @(q,&) to be the subspace of the Hardy space Ell1 (@2n,Ukn) given by Therefore, we consider ORDAP for MIMO systems, as represented by the performance index (9), specifically po = QEHiyf Duality Structure of the Problem (10) space L1(q:;n)/$(qz;n), under the quotient norm Furthermore, the optimization (10) can be shown to be equivalent to [2] 1857 II[Glllw = inf g€mq:Gn) IIG + gllL1(q:;n), (16) (G,,,) in L1( q::,),then W"O (G,X,) is isometrically isomorphic to the dual space of (@2",":,). Therefore, the pre-orthogonal of S is given by [lo, 111 Hm where @ denotes the direct sum of two subspaces, and t = ((I - RR*)IL' (G;;:, e E Z : ( G , ~ ,1) ( 2 ). ( : ) W(cznxn), then @ may be viewed as where F = Let @ = E cisely n-:(GZn) (18) Clearly, MQ is a bounded linear operator. We show in The quotient norm in by 2 is denoted by 11 II [FIIIs = k$IIF+ ~IIL~(C;:;,,), the next Proposition that ' @(q::,) 11s and is given the operator induced norm is equal to V [F]E 3 (19) Theorem 2 ([12], p.121) implies the following theorem, which shows existence of optimal feedback control laws for ORDAP of MIMO systems. N 11@11-. (@(G,,,))* Proposition 1 [lo] Let @ and MQ be defined as above. Then 1. llMd@(G,";,) N ( @ ( ~ 2 n x n ) ) * ~ ~ = 5 SUP ~ ~ g ~ ~ L z ( c n x n1 ) l l ~ * ~ l l L z ( ~ z n x n= ) ll*Il-, and 2. 9E~"(cxn) llgllLz(cnxn) < - 1 IIM@gIILZ(eznXn) = ll@ll-. llM*ll = S U P g E L2(Cx,) + Theorem 1 Under assumption ( A l ) (W*W and v * v ) ( e i e ) > 0, V8 E [ 0 , 2 ~ ) . There exists at least The dual space of L'(Gnxn) is given by L2(q::,) one Q o E H m ( G x , ) (i.e., a n optimal controller CO) vise-versa, hence I L' (Gnxn) is reflexive [lo]. In the next such that Proposition we characterize the dual space of @ (Gnxn). = 1(7) -j j ~ ~ I / N and @ ( q::,) defined (20) Proposition 2 Let I@(Gnxn) as above. Then 1. @ ( C z n x n ) N (p(q:;,))*, and 2. 4 (G,,,,) and @ ( q::,) Hence spaces. IF1 E s Exact Solution via Operator Theory are reflexive Banach Proof The proposition follows essentially since q::n are finite dimensional [lo]. Gnx, and In this section we give a solution based on operator theory for the ORDAP of MIMO systems. The key observation 4.2 Operator Theoretic Solution here is to interpret the norm (12) as an operator induced Let n' be the orthogonal projection on the closed subnorm in special vector-valued Hardy spaces. of I@(Gnxn), where space I@(Gznxn) e @ 0 f i H 2 ( G x n )is the orthogonal complement of R H 2 ( G X n )(,f i H 2 ( G x n ) ) l . Here orthogonality is 4.1 A Multiplication Operator understood to be with respect to the inner product of Let L z ( G x n )( H 2 ( G x n ) )denote the Banach space of matrices, i.e., A , B E w(GnXn) are orthogonal if and Lebesgue square integrable (and analytic) GX n-valued only if functions on the unit disc under the norm (Gnxn) 1 2T JO where I . I denotes the largest singular value. ' (Gnxn) (@ to be the Banach Likewise define L space of Gnxn-valued and Lebesgue square integrable (and analytic) functions on the unit disc endowed with the norm (G,,,,)) Tr{A*B}(eie)dm =0 Then 11' is a bounded linear operator on @ Next, define the following operator 1858 2' by :~ (25) (G,,,). ' ( + 6 w~~ ~ ( C)e~T ~~H ~~ ( ~G ~), ) Z' =rI'M( ) (26) T h e o r e m 2 Under assumptions ( A l ) , and (AZ) W , R continuous on the unit circle, if po where > poo, to H m ( G x n ) ,then they all admit power series expansions in the unit disc, for instance i.e., when the open unit analyticity constraint is removed, the following hold: E', i ) po is equal to the operator induced norm of namely Then k m = Il='ll (27) ai) There exists a maximal vector f o r E', i.e., f k = 0,fl,f2,.. . E m (31) M k Proof The proof is similar to the proof of theorem 3 [ l l ] . It should be noted that assumption (A2) implies that the Hence straightforward computations yield the following optimal solution is allpass, inferring existence of a maxi- recursive algorithm mal vector (see [lo, 111 for details). The orthogonal projection n' is also given by proposition 3 [ll],i.e., IT = I - R P + k , where this time I is the P+ is the positive Ftiesz proidentity map on into @ (G,,,,,), roughly_speakjection from L2 ing, P+ associates to each vector function in IL2((@2,,,,) its positive Fourier coefficients. The operator Z' is then a combination of Toeplitz and multiplication operators, @(enxn), (e,,,,) that is 2 = ( I - R P + k ) 5 The operator Z' has then the following infinite matrix representation with respect to the basis { z k e j } Computation of the Operator Norm In this section we show how it is possible to compute the norm of the operator Z', and hence the optimal performance index po. First, note that for f E H2(Gxn), f(z) = CEO z k f k , for some fk E G,,, the action of 8' on f is determined by 6 k=O Approximation by Norms of Matrices k=O In what follows we will show that it is possible to apTherefore with respect to the canonical basis { z k e j } ,k = proximate the norm of 8' by the norms of finite dimen0 , 1 , - . . , j = 1 , 2 , - . - , n ,where { e j } is a basis for G x n ; sional operators or matrices. Indeed, if we restrict Z' to E' has the following representation finite dimensional subspaces of its domain, then we will a sequence of finite rank operators which norms obtain E' = (E'l,E'z,. . . , 2 z n , . ..) (30) are "nearly" optimal, in the sense that they approach The k l-th column corresponding to the representa- the norm of E' as the dimension of the initial spaces intion of E'zk with respect to the basis { z k e j } ; is then creases. More explicitly, let P, denote the orthogonal formed by the coefficients of Z'zk in with re- projection from If2(&,,,) to the subspace spanned by spect to { z k e j } . To compute these coefficients consider, { z k e j } ,k = 0 , 1 , . . . ,m; and P, the orthogonal projection @ onto the span of { z k e j } ,k = 0 , 1 , . . . ,m. Z' = , since W , R I and R 2 belong from Clearly, IIP,II 5 1; and likewise it is not very hard to -RI P+Rr W + G,,,,, ( (G,,,,) 1859 show that IIPmll 5 1. The rank of Pm, rank(P,), defined as the dimension of the image space of P,, is such that r a n k ( P m ) 5 m2n2. The operators EL = PmZ’Pm, m = 0 , 1 , 2 , .. .; have rank less than or equal m2n2.Next, we show that in fact, as m tends to infinity, EL converges to Z‘ in the strong operator topology, i.e., for f E f f 2 ( G x n ) , I l f IILZ(C,,) I 1, we have that mlim +w 1 1 2 f - Pm6‘Pmf IILZ(Gnxn) =O (34) Indeed, by adding and subtracting Pm2‘f , the triangle inequality implies lP’f - ~ m ~ ’ ~ m f I I L Z ( E z , xI , ) IP’f - P m w L z ( C , , x , ) References [l]R. Schatten. N o r m Ideals of Completely Continu- ous Operators. Springer-Verlag, Berlin, Gottingen, Heidelberg, 1960. + IIPmE’f - PmE‘PmfllL2(~z,x,) I + ll~mllII=’llllf - P m f 1 1 L 2 ( c n x n ) (35) IP’f - Pm=’fllLz(cz,,,) It is therefore possible to approximate as closely as desired the norm of E‘, and hence pa, by computing the special matrix norms defined by (38) or (39). From expression (36), we see that explicit computation of the maximal vectors of Z L , (or EL*),will shed light on the form of the maximal vector of Z’, and therefore the optimal controller Co. Gn,, Since is finite dimensional, by the Riesz-Fischer Theorem (c.f. Lecture VI, [13]) the two terms on the right-hand side of (35) tend to zero, and thus (34) must hold. Therefore (34) implies IIE’II - mlim - + mI I P ~ E ’ P ~ ~ I I ~ ~ ~ = ~ ~o , , , ) (36) It follows then IIPmE’PmII + 112’11 as m + 00, or equivalently, I IPmz’ Pm f m IIn,zie,,,, + IIE’f II L 2 x. , . Moreover, as noted earlier with respect to the canonical basis { z k e j } ,PmE’Pm is represented by a matrix obtained by truncating the infinite matrix representation (33) of E’, i.e., (e, [2] J.G. Owen. Performance Optimization of Highly Uncertain Systems in Hw. PhD thesis, McGill University, 1993. [3] G. Zames. Feedback and optimal sensitivity: Model reference transformations, multiplicative seminorms, and approximate inverses. I E E E Transactions on Automatic Control, AC-26(2):301-320, April 1981. [4] J.F Bird and B.A. Francis. On the robust disturbance attenuation. Proceeding of I E E E Conference o n Decision and Control, pages 1804-1809, 1986. [5] B.A. Francis. On disturbance attenuation with plant uncertainty. Workshop o n New Perspectives in Industrial Control System Design, 1986. [6] J.G. Owen and G. Zames. Duality theory of robust disturbance attenuation. Automatica, 29(3):695-705, 1993. (37) The computation of the norm of the matrices EL can be further simplified, for e.g., in the skalar case (n=l), EL is a 2(m 1) x ( m l), (each entry ( a k j l ) in (37) is a complex 2 x 1 vector), the norm of EL can be shown to reduce to [7] small + + fm 1pm ’ *11= E Gn + max max(lhlml, lhzmI) = 1 hlm,h2m I=lm hlm + =;m*h2ml [ll] M.S. Djouadi. Exact solution to the non-standard H”’ problem. Proceedings of I E E E Conference on Decision and Control, Tampa, Florida, 1998. Clearly, such norms are bounded above and below, by respectively lELml and max(lEiml , lZbm1), i.e. + 7 [lo] M.S. Djouadi. Optimization of Highly Uncertain Feedback Systems in Hm. PhD thesis, Dept. of Electrical Eng., McGill University, 1998. E @, (39) mm(IEimI [8] M.S. Djouadi and G. Zames. On optimal robust disturbance minimization. Proceedings of,the American Control Conference, June 1998. [9] M.S. Djouadi and G. Zames. On optimal robust disturbance attenuation. Submitted to Systems & Control Letters. where I . I is the Euclidean norm, and ELrn, k = 1,2; are (m 1) x (m 1) matrices. Or equivalently, by passing to the adjoint of E&, a&*,we have -I * + [7] M.S. Djouadi. Banach space optimization of uncertain systems in Hm. PhD Research Proposal, Department of Electrical Engineering, McGill University, 1997. I%mI) I II%II I IEimI + IELml [12] D.G. Luenberger. Optimization by Vector Space Methods. John Wiley, 1968. [13] H. Helson. Lectures o n Invariant Subspaces. Academic Press, New York and London, 1964. (40) 1860