LIDS-P-1545 April 1986 An Interpolation Problem Associated With Hf Optimal Design In Systems With Distributed Input Lags by Gilead Tadmor! Abstract A variety of Hf optimal design problems reduce to interpolation of compressed multiplication operators, f(s) ->rk(w(s)f(s)), where w(s) is a given rational function and the subspace K is of the form K = H2 00(s)H 2 . Here we consider 0(s) = (1-ea-S)/(s-a), which stands for a distributed delay in a system's input. The interpolation scheme we develop, adapts to a broader class of distributed lags, namely, those determined by transfer functions of the form B(e S)/b(s), where B(z) and b(s) are polynomials and b(s) = 0 implies B(e- s ) = 0. Interpolation, HP Key Words. eigenvalues, Time domain analysis. spaces, Distributed delays, Maximum /Dr. Gilead Tadmor, Massachusetts Institute of Technology, Laboratory for Information and Decision Systems, Room 35-312, 77 Massachusetts Ave., Cambridge, MA 02139 (617)253-6173. 2 0. Notations H2 (resp. Ha) is the space of L2 (resp. L,) functions on the imaginary axis which admit analytic continuation in the right half plane. source on HP spaces is [7]. A good By abuse of notations we shall not distinguish between a function in H2 and its inverse Lapace transform in L2 [0,-] (e.g. by "^" or avdz). presentation. Hopefully, this will simplify, rather than obscure, the A star will denote the adjoint of an operator (e.g. T*). 3 1. Introduction A variety of H a design problems (see e.g. the surveys [6,8]) give rise to "interpolation" problems of the following form: rational function. (For simplicity we assume throughout this note that w(s) Let 0(s) be an inner function in H a and is both stable and minimum phase.) let K be the subspace K =: H280(s)H 2 . multiplication by w(s). interpolating Consider now the compression to K of That is, the operator T:f(s) -4nK(w(s)f(s)) An Let w(s) be a proper for f(s)eK function for T is q(s) e HF having these two properties. (i) 11TIh (ii) Tf(s) = TK(T(s)f(s)). (= the operator norm of T) = IVl¥Iq , Note that, necessarily, if ¥(s) (1.1) interpolates T, it is of the form w(s) + 0(s)h(s) for some h(s) in Ha. Our setup is slightly, but not significantly, different: Instead of an inner function 0(s), we have the H a function (1-e'-S)/(s-o), with Rea>O. One easily observes that the outer part of this function is He-invertible; hence, our subspace K =: H2 0(1-ea- S)/(s-a)H2 is the same as that obtained by substituting (1-ea-S)/(s-a) by its inner part. The advantage of the present setup is that it facilitates the time domain analysis, as most of ours will be. The treatment of the present particular examples is motivated by a larger class of functions 0(s); namely those of the form 0(s) = B(e- s)/b (s), where both B(z) and b(s) are polynomials and b(s) = 0 implies B(e-S ) = 0. 4 Functions 0(s) in this class correspond to delay operations formed (in the time domain) by cascades of elementary operators of those next three types (i) f(t) -)f(t-1) (ii) f(t) -)f(t) - eaf(t-1) (iii) f(t) - fealf(t-v)dz. Operators of the first type were treated in [1,2,3] [4,5]. and Inspired by those works, the present author developed the theory in [10], to suit combination of operators of both types (i) and (ii) (i.e., the case b(s) a 1), solution. and suggested also a simple computational scheme for the Our aim here is to adapt those theory and scheme to suit non constant b(s). For simplicity, the analysis is done explicitly only for the elementary building stone of type (iii). Using the methodology of [10], the reader will then be able to obtain the more general result. For future use, let us give w(s) two specific forms. The first is P w(s) 1+ i Bi. +s i=1 Since w(s) is stable, Repi , i=1,...,p, all belong to the open right half plane. The number q = lim w(s) could be scaled to be either 0 or 1. The s-)300 trivial case w(s) i a is excluded. The second form is W(s) =- (s) dTs) 5 where n(s) and d(s) are coprime polynomials. Merely for simplicity in notations, we assume that all our system's parameters (i.e., a, A, ¥i and pi) are real numbers. For the same reason, we also treat only the case where AI- -.,ip are distinct. can be easily dispensed. Both assumptions 6 2. An Outline of the Solution The underlying idea behind the solution (following [1,2,3] and [4,5]) is this next observation of Sarason [9, Proposition 5.1]: Lemma 2.1 (Sarason). If the operator T has a maximal function, say f(s), then the unique interpolating function for T is ¥(s) =: Tf(s)/f(s). Furthermore, that q(s) is a constant multiple of an inner function. Thereby, one is lead to a search for a maximal function for T, or equivalently, for T*T. Since (it is easy to see) T*T is the sum of R2 I (I, being the identity operator) and a compact, infinite dimensional operator, the spectrum of T*T consists of countably many eigenvalues, each of a finite multiplicity, and of their only accumulation point, which is q2 . follows from Weyl's Lemma (see e.g. Ill, p.32 5 Theorem 1]). This Our main effort will go thereby into the search for a maximal eigenvalue and eigenfunction for T*T. In some cases, however, a maximal eigenvalue does not exist. may yet occur that w(s) itself is an interpolating function. Then it (Note that by definition, w(s) satisfies condition (l.1.ii), but mostly (1.1.i) fails and IlW(S) I > I ITI].) Here are sufficient conditions for either of the two cases. Proposition 2.1. If lw(jo) I<] H] The spectral point t 2 is not an eigenvalue of T*T. for all weR, then a maximal eigenvalue does not exist, yet w(s) is an interpolating function for T. If, on the other hand, there exist 00>0 such that Iw(jw)I > HI eigenvalues are larger than exist. for all weR with ItI>wo0 q2. then countably many In particular, a maximal eigenvalue does 7 This statement sumnerizes Propositions 3.1 and 3.2 Observation 3.4 in [10], where one can also find the proof. and part of (The present generalization requires no significant changes.) Henceforth we work eigenfunctions of T*T. on characterization of eigenvalues and The analysis will be carried out in the time domain. So our first task is to find out the exact form of the subspace K and of the operators T and T*, following an inversed Laplace transform. This is done in the following section 3. In section 4 we then construct (in terms of the number a and ({(X2 ):X2 >0}. the function w(s)) a parameterized family of matrices It has the property that X2 is an eigenvalue if and only if det G(X2 ) = 0. Given an eigenvalue X2 , a corresponding eigenfunction is then obtained from a right annihilating vector for W(X2 ). One thus finds the solution to the interpolation problem after a search for the maximal X2 which solves det D(X2) = 0. This next simple observation (see [10, Observation 3.4]) helps in limiting the domain of that search. Proposition 2.2. the interval (q2, The maximal eigenvalue of T*T (if it exists) lies in cIw(s)II2]. 8 3. The Time Domain Setup. A function fl in K = (1-ea-S)/(s-a)H2 has this next time domain representation f1 (t) =- eag(t-), for t-O where g(t) belongs to L 2 [0,w) and is understood as the zero function for negative values of the argument. Thereby, a function fo belongs to K if and only if for each g(t) in L 2 [0,~) the following holds 0= I <f0 (t), = J <f 0 0 eacg(t-T)dv> dt eaIf 0 (t+,)dT, g(t)>dt. The subspace K is characterized, therefore, by the requirement f(t+)d = for all t>O. (3.1) Differentiation of (3.1) yields a difference equation eaf0 (t+l) - f0 (t) = 0 Observation 3.1. for all t>0. A function fo(t) (3.2) in L 2 [O,-) belongs to K (i.e., 9 satisfies Eq. (3.1) for all t0>) if and only if Eq. (3.1) is satisfied for t=0 and Eq. (3.2) holds for all positive t. The proof is immediate. The observation tells us that a function in K is determined by its restriction to [0,1], and that restriction belongs to the subspace of L2 10,1] which is orthogonal to eat. We also obtain the projection onto K along L 2 10,-) (denoted Kr .) Observation 3.2. f0(t) = (nKf)(t). (i) fo(t) Given a function f(t) in L 2 (0,o), set Then = (1-e 2) e-ia(f(t+i) 2ae t eaf(T+i)d,) i=0 for te[0,1), and (ii) f0(t+1) = e af (t) for all t>1. Proof. A short, yet indirect proof goes as follows: First one observes that indeed, (i) and (ii) force fo(t) to satisfy the conditions in Observtion 3.1. Second, it easily follows that when f(t) belongs to K (i.e., it satisfies Eqs. (3.1) and (3.2)) then the function fo(t) defined in (i) is just f(t) itself. A constructive proof which sheds light on the way wK was obtained in the first place, can be found in [10, section 2]. Corollary 3.3. Let X c L2 [0,1] be the subspace of functions which are orthogonal to eat, with its inner product weighted by the scalar 1/(l-e(i.e. <fg>x = 1/(1-e- 2 )<fg>L [0 ,1].) Then K is isometric to X. 2a ), 10 Proof. Since a function in K is determined by its restriction to the interval [0,1] (i.e., by its representative in X), one only has to check the claim concerning the inner product, which follows from condition (3.2). Now let us find the representation of T as an operator on X (which is well defined, by the corollary): become convolutions. Under inverse Laplace transform, products Hence the uncompressed operator of multiplication by w(s) turns into T¥i f(t) -)g(t) = if(t) + e (3.3) i=1 In order to obtain Tf(t) we have to project the right hand side of (3.3) onto K. By Observation 3.2, this goes in two steps: First compute (1-e-2a ) 2 e-iag(t+i) i=O and then project the sum (as a function on [0,1]) onto the subspace X. Since the domain of T is the subspace K, we have nK(Tqf) = if. computation is thus needed only for terms of the form Using Eq. (3.2), one gets The 0eA(T-t)f(t)dt. +i e (T-t-i) f(z)dz = i-1 i-! e(q-i)P-q CL = -eO q=0= qe( -(+e-ia eP(r-t)f(T)dr fe' f )d + e-ia< eP(r-t)f( O +e the other, a=P, being very (Here we chose to treat only the case aif, similar.) eS(~-t)f(r)d~ The following summation yields co (1-e 2a) ~ e i a+e i=0O e-(a+) (,-t-i)f(T)d, = a (- +J t)f(T)d e + . Now comes the projection onto X (i.e., -subtraction of (3.4) 12 2A<k(), eat>e a t 2al where k(t) stands for the right hand side of (3.4)). It yields (collecting all terms of Tf(t)) - P Tf(t) = if(t) + ¥ 0p. (r-t) ii e1 f( )dT i=1 p ~ + -(a+8i) i(s-t) e e ¥i e- i=1 f(e)dT (3.5) 1-e + e2aei=1 i a.ii=1 1 e if(T)dT i -a e -e (Since f(t)I[o,1] belongs to X we could drop terms involving fleatf(t)dt.) As is born out from Observation 3.2 and Corollary 3.3, one can compute T* by first taking the L2 [O,1] adjoint of the operators described in Eq. (3.4), and then project onto X. In our example, these straightforward computations yield a very simple form p T*f(t)= If(t) + Yi i=1 P (t-) e f(T)dT. (3.6) 13 4. The Eigenvalue/Eigenfunction Problem Here are system representations for the operators T and T*: Let xi(t) = -.ixi(t) + Tif(t) p+ (t) = axp+1 (t) p+1 g(t) = If(t) + (4.1) ~ x i(t) i=l i+p+l(t) = 0iXi+p+l(t) - yig(t) i=l,. p 2p+1 xi (t ) , h(t) = ng(t) + ~ i=p+2 and impose the boundary constraints xi(1) = eaxi(0) P xp+l = (0) i- xi(1) = 0 Observation 4.1. i -a i (4.2) x (0) 1 i=p+2, ...,2p+1 Suppose f(t) belongs to X. Then g(t) = Tf(t) and 14 h(t) = if T*g(t) and only if f(t), g(t), h(t) and functions xl(t),...,x2p+l(t) satisfy Eq. (4.1) and the boundary conditions (4.2). Proof. The "only if" direction is trivial. (4.2) are met. Then by simple algebra -(a+fi ) xi(O) = e 1-e for i=l,...,p. Assume Eqs. (4.1) and ) J f ()dS This insures, in turn, that g(t) = Tf(t) and, indeed, h(t) = T*g(t). The following nice and useful observation was made by Flammn in a somewhat different setup [2,3], yet his ideas carry over to our case as well. Observation 4.2. Suppose a positive number x2 is an eigenvalue of T*T, and that f(t) e X is an associated eigenfunction. the set of functions, each of the form Let tl(t),...,42p(t) be (t) = tqe t , where the numbers p is a root of order (at least) q+1 of this next equation X d(s)d(-s) - n(s)n(-s) = 0. (Recall the notation w(s) = n(s)/d(s).) (4.3) Then f(t) is a linear combination of the functions gi(t), i=1,...,2p and of eat. Proof. If f(t) is an eigenfunction of T*T then we can substitute 15 P+l X2 f(t) for h(t), and nf(t) + E xi(t) for g(t) in Eq. (4.1). xi(t) and f(t) are analytic on [0,1]. Necessarily, Taking the Laplace transforms of these functions' analytic extensions to [0, ], the following equation is obtained 'I' (A -2 )f(s) = cT(sI - 0 :a :-1 0 ,-]---- 81. , CT + ( ) B X2p+l f(s)) (4.4) (Here CT is the (2 p+l)-raw vector (1,...,1).) Further, direct computation shows that the coefficient of f(s) on the right hand side of (4.4) is w(s)w(-s)-n 2 . f(s) = Thus Eq. (4.4) is equivalent to 2 X2-w(s)w(-s) [ ])1 (45) x.1(0) l 1( 2Xp+l (0) + Finally, one notices that poles of the inverse matrix, other than at s=a, cancel with poles of w(s) or with those of w(-s). Thus, the poles of f(s) belong to the set of zeros of Eq. (4.3), and the point a. Now, given X2 and the question: is k2 an eigenvalue?, the observation 16 tells us that we can restrict our search to the 2p-dimensional subspace l (t ) ,...,t)2p(X. } V=:sp{eat, since if f(t) (Indeed, this subspace is 2p dimensional v0 eat + viti(t) E belongs to X then vO = - 2a/(e 2 a-1) E <ti(I), eaU>.) Let us thus consider the restriction of T to V. It ranges in the 3pdimensional subspace U =: sp(eat, tl(t),...,2p(t), e-I t,...,e pt) n X. The restriction w =: sp{eat, of T* to {l(t),...,42p(t) ' U, e-0 in turn, takes ) X. t,...,epp t} values These in restricted operators can be represented by 2px3p and 3px4p matrices, denoted 2p P 1 1 12 2p ' and . P 11 ....... T1 2 31 In these representations, 32J the coefficients of (in that order) of e±it,...,efPt. 2p, 3p and l1 (t),...-2p(t), 1 4p vectors describe the e-1t,...,ef-Pt and then The coefficient of eat does not appear since it is always determined by the others. The following are important facts. Observation 4.3. 2px2p identity. Proof. The matrix T22 is invertible. Let us consider only that case where Eq. (4.3) has 2p distinct roots pl,... ,2p- exposition.) The product T*1 T 1 is equal to X2I, where I is the (The other case requires just a little more complicated Then T1 and T*1 are these next diagonal matrices 17 TI = diag{w(jl),,...-,w(R2p)} and T1i = diag{w(-1),...,w(-p) (This follows from substituting {i(t) for f(t) in Eq. (3.5).) Since pi are roots of Eq. (4.3), the first statement is true. Similarly, one finds that T22 (the block which is responsible for the transformation of the subspace sp{e- P t,...,ee-Pt} into itself, by T*) is this next T22 diaglw(P 1),...,w()p) } . By assumption, w(s) is minimum phase and P1 ...,fp are positive numbers. Hence the second statement in the observation. Theorem 4.4. T1 and T*2. From these facts we get Fix X2 and the corresponding matrices T1, T2, Tl1, T22, Then X2 is an eigenvalue of T*T if and only if this next matrix 231TI + T32T2 is singular (i.e., det 0(X2 ) = 0). If indeed, X2 is an eigenvalue and 18 V2=p is a right annihilating vector for a(X2) (i.e., C(X2 )v = 0), then the function 2p f(t) = S 2p vii(t) - eat i=1 t (.i(,) 2a e2a- ea> (4.6) - e i=1 is an associated eigenfunction for X2. Proof. Suppose T*Tf(t) = X2 f(t). 0(X2 )v = 0. Then following the previous proof, Suppose, on the other hand, that x2 is an eigenvalue and f(t), a corresponding eigenfunction. Observation 4.2 tells us that f(t) has the representation (4.6), for some vector v. Since the coefficients of e-pat, i=l,...,p in T*Tf(t) are all zero, it follows that both the vector T22T2 v and (T31 T 1 + T32T 2 )v must vanish. Finally, since T22 is invertible, the condition a(X2 )v = 0 is met. In the theorem's statement we did not give the specific formulas for the various components in 0(X2 ). The interested reader will readily obtain them from Eqs. (3.5) and (3.6) The theorem gives rise to the main part in the interpolation scheme: One computes det 0a( 2) for a decreasing sequence of X2 in (12, I1w(s)112], until he hits the highest solution of det 0(X2 ) = 0. This will be the highest eigenvalue of T*T. The corresponding eigenfunction's values in 19 [0,1] are then given by Eq. (4.6) (for v such that f(X2 )v = 0). The values of f(t) for t>1 are thereby determined via f(t+i) = e-iaf(t), which is Eq. (3.2). In order to find the interpolating function one has to compute also Tf(t) (via Eq. 3.5) and the Laplace transforms of f(t) and Tf(t): Observation 4.5. Suppose A2 is the maximal eigenvalue for T*T, let ,1"',~t2p be the roots of Eq. (4.3) (for simplicity we assume these are 2p distinct numbers) and let v be a vector satisfying D(X2 )v = 0. Then the unique interpolating function, ¥(s), for T, is given by 2p p w(i)viti(s) + . 2p T(s) =: i= -(P.+s) zi + vo Z e S-1 . . ea-s-1 . va. viti() + -e I a-s i=1 where the p-vector z is z = T 2 v, and the numbers zo and vo are obtained so as to ensure that f(t) and Tf(t) be in X, i.e. 2p vn a O and e2a-1 vi e2a_ ! i=1 3 y0 i(t)eatdt 20 2a 2a-- Zo = W(Fi)vit ti(t)eatdt ziJ e + i dt. i=1 The functions ti(s) are the Laplace transforms of 4i(t), i.e. o ti(s) .1 e ts t=-i ( t)dt i-s)t = [ ei dt '0 = -e ei - s Ai- s By Sarason's Lemma 2.1, above, q(s) = Tf(s)/f(s), where f(s) is Proof. the maximal eigenfunction. By Eq. (3.2), the transform of a function f(t) in K is f(s) = J e-tSf(t)dt e i ( sa+ ) J e-tsf(s)ds i=0 _e-(a+s) e-tf(s)ds . So, it suffices to compute the transforms along [0,1]. The term l/1-e-( a+S ) appears in both Tf(s) and f(s), so it is canceled by devision. The rest of 21 the claim follows by direct computation. 22 5. Concluding Remarks Via the relatively simple example of 0(s) = (1-ea- s )/(s- a) we demonstrated the main steps in an interpolation scheme for a class of transfer functions with distributed delays. Key steps in the analysis were the computation of the projection mK' hence of T and T*, in the time domain, and characterization of maximal eigenvalues of T*T as zeros of Q(X2 ). The main difference of our case from the corresponding non distributed one (i.e., 0(s) = 1-ea S) is that here one has to look for the maximal eigenfunction within the subspace X, instead of within the whole of L2[0,1]. The methods developed in [10], bearing some corrections due to the introduction of the counterpart of X, enable one to handle the more general case of multiple distributed lags (i.e., 0(s) = B(e-s)/b(s), as described in section 2). Indeed, then life becomes more complicated with heavier formulas and computations. This is due, in part, to the fact that the role played in the present example, by some scalars (e.g., e-a ) is taken, in the more general case, by matrices. The unfortunate lack of cormutativity of matrix multiplication does not help. Acknowledgement I wish to thank Professor Sanjoy K. Mitter and Dr. David S. Flamm for interesting discussions and enlightening remarks concerning this work. 23 References [11 D.S. Flanmmn and S.K. Mitter, "Progress on H a Optimal Sensitivity for Delay Systems I: Minimum Phase Plant with Input Delay, 1 Pole/Zero Weighting Function", LIDS Technical Report P-1513, MIT, Novenber 1985. [2] D.S. Flamm and S.K. Mitter, "Progress on H a Optimal Sensitivity for Delay Systems II: Minimum Phase Plant with Input Delay, Genera Rational Weighting Function", preprint, LIDS-MIT, January 1986. [31 D.S. Flamm, Ph.D. Thesis, LIDS-MIT, to appear. [4] C. Foias, A. Tannenbaum, and G. Zames, "Weighted Sensitivity Technical Report, Dept. of Minimization for Delay Systems," Electrical Engineering, McGill University, September [1985]. [5] C. Foias, A. Tannenbaum, and G. Zames, "On the H1-Optimal Sensitivity Problem for Systems with Delays," Technical Report, Dept. of Electrical Engineering, McGill University, December 1985. [6] B.A. Francis and J. Doyle, "Linear Control Theory with an H a Optimality Criterion", Systems Control Group Report #8501, University of Toronto, October 1985. 17] J.B. Garnett, Bounded Analytical Functions, Academic Press, New York, 1981. [8] J.W. Helton, "Worst Case Analysis in the Frequency Domain: The H a Approach to Control", IEEE Trans. Auto Control, AC-30 (1985), pp. 1154-1170. [91 D. Sarason, "Generalized Interpolation in Ho," Trans. AMS, 127 (1967), pp. 170-203. [10] G. Tadmor, "Hi Weighted Sensitivity Minimization for Systems with Commensurate Input Lags", LIDS-Technical Report P-1546, MIT, March 1986. [11] K. Yosida, Functional Analysis, Springer, New York 1965.