Theorem Let X be a symmetric solution of DR(X) = 0. Let X b be a symmetric approximation to Xand set V := X? X. b If R b = R + B T XB b and I + V B R

Transcription

1 Initializing Newton's method for discrete-time algebraic Riccati equations using the buttery SZ algorithm Heike Fabender Zentrum fur Technomathematik Fachbereich - Mathematik und Informatik Universitat Bremen D-84 Bremen, Germany heikemath.uni-bremen.de Peter Benner Zentrum fur Technomathematik Fachbereich - Mathematik und Informatik Universitat Bremen D-84 Bremen, Germany bennermath.uni-bremen.de Abstract The numerical solution of discrete-time algebraic Riccati equations is discussed. We propose to compute an approximate solution of the discrete-time algebraic Riccati equation by the (buttery) SZ algorithm. This solution is then rened by a defect correction method based on Newton's method. The resulting method is very ecient and produces highly accurate results. Keywords. discrete-time algebraic Riccati equation, Newton's method, SZ algorithm, symplectic matrix pencil, defect correction Introduction The standard (discrete-time) linear-quadratic optimization problem consists in nding a control trajectory fu(t) t 0g, minimizing the cost functional J (x 0 u) = X t=0 [x(t) T Qx(t) + u(t) T Ru(t)] in terms of u subject to the dynamical constraint x(t + ) = Ax(t) + Bu(t) x(0) := x 0 : Under certain conditions there is a unique control law, u(t) = Hx(t) H :=?(R + B T XB)? B T XA minimizing J in terms of u subject to the dynamical constraint. The matrix X is the unique symmetric stabilizing solution of the algebraic matrix Riccati equation 0 = DR(X) = Q? X + A T XA?A T XB(R + B T XB)? B T XA: () The last equation is usually referred to as discrete-time algebraic Riccati equation (DARE). It appears not only in the context presented, but also in numerous procedures for analysis, synthesis, and design of control and estimation systems with H or H performance criteria, as well as in other branches of applied mathematics. The DARE () can be considered as a nonlinear set of equations. Therefore, Newton's method has been one of the rst methods proposed to solve DAREs [8]. Given a symmetric matrix X 0, the method can be given in algorithmic form as follows: FOR k = 0 : : : A k A? B(R + B T X k B)? B T X k A: Solve for N k in the Stein equation A T k N ka k? N k =?DR(X k ): X k+ X k + N k END FOR It is well known that (under certain reasonable assumptions) if X 0 is a stabilizing starting guess, then all iterates are stabilizing and converge globally quadratic to the desired solution X (see, e.g., [8, 9, ]). Despite the ultimate rapid convergence, the iteration may initially converge slowly. This can be due to a large initial error jjx? X 0 jj or a disastrously large rst Newton step resulting in a large error jjx? X jj. In both cases, it is possible that many iterations are required to nd the region of rapid convergence. An ill-conditioned Stein equation makes it dicult to compute an accurate Newton step. An inaccurately computed Newton step can cause the usual convergence theory to break down in practise. Sometimes rounding errors or a poor choice of X 0 cause Newton's method to converge to a non-stabilizing solution. For these reasons, Newton's method is usually not used by itself to solve DAREs. Usually it is used as a defect correction method or for iterative renement of an approximate solution obtained by a more robust method. A defect correction method can be based on the following result of Mehrmann and Tan [, Theorem.7] (see also [, Chapter 0]).

2 Theorem Let X be a symmetric solution of DR(X) = 0. Let X b be a symmetric approximation to Xand set V := X? X. b If R b = R + B T XB b and I + V B R b? B T are nonsingular, then V satises the DARE ba T V b A? V? b A T V B( b R + B T V B)? B T V b A + b Q = 0 where b Q = DR( b X) and b A = (I? B b R? B T b X)A: Hence, the error V fullls a discrete time algebraic Riccati equation just like the desired solution X. The defect discrete-time algebraic Riccati equation may be solved by any method for discrete-time algebraic Riccati equations. Most frequently, in situations like this, Newton's method is used [8,, 9, ]. Then V can be used to correct X. b Iterating this process until DR(X) is suitably small is called a defect correction method. Using this approach, it is often possible to squeeze out the maximum possible accuracy [] after only a few iterations. Here we propose to compute an approximate solution bx of DR(X) by the (buttery) SZ algorithm. Combined with the defect correction method, the resulting method for solving () is a very ecient method and produces highly accurate results. Now assume R to be positive denite and dene A 0 I?BR K? N =?? B T Q I 0 A T : () Using the standard control-theoretic assumptions that (A B) is stabilizable, (Q A) is detectable, Q is positive semidenite, then K? N has no eigenvalues on the unit circle and there exists a unique stabilizing solution X of the DARE () see, e.g., [9]. It is then easily seen that K?N has precisely n eigenvalues in the open unit disk and n outside. Moreover, the Riccati solution X can be given in terms of the deating subspace of K?N corresponding to the n eigenvalues : : : n inside the unit circle using the relation A 0 I I?BR =? B T I Q I?X?X 0 A T where IR nn, () = f : : : n g. Therefore, if we can compute Y Z IR nn such that the columns of [ Y Z ] span the desired deating subspace of K? N, then X =?ZY? is the desired solution of the Riccati equation (). See, e.g., [9, 0, ], and the references therein. It is worthwhile to note that K? N of the form () is a symplectic matrix pencil. A symplectic matrix pencil K? N K N IR nn is dened by the property where KJK T = NJN T J = 0 I n?i n 0 and I n is the nn identity matrix. The nonzero eigenvalues of a symplectic matrix pencil occur in reciprocal pairs: If is an eigenvalue of K? N with left eigenvector x, then? is an eigenvalue of K? N with right eigenvector (Jx) T. The numerical computation of a deating subspace of a (symplectic) matrix pencil K? N is usually carried out by an iterative procedure like the QZ algorithm. The QZ algorithm is numerically backward stable but it ignores the symplectic structure. Applying the QZ algorithm to a symplectic matrix pencil results in a general nn matrix pencil in generalized Schur form from which the eigenvalues and deating subspaces can be read o. Due to roundo errors unavoidable in nite-precision arithmetic, the computed eigenvalues will in general not come in pairs f? g, although the exact eigenvalues have this property. Even worse, small perturbations may cause eigenvalues close to the unit circle to cross the unit circle such that the number of true and computed eigenvalues inside the open unit disk may dier. Moreover, the application of the QZ algorithm to K? N is computationally quite expensive. The usual initial reduction to Hessenberg-triangular form requires about 70n ops plus 4n for accumulating the Z matrix each iteration step requires about 88n ops for the transformations and 6n ops for accumulating Z see, e.g., []. An estimated 40n ops are necessary for ordering the generalized Schur form. This results in a total cost of roughly 4n ops for solving a DARE using the QZ algorithm, using standard assumptions about convergence of the QZ iteration (see, e.g., [7]). Here we propose to use the buttery SZ algorithm for computing the deating subspace of K? N. The buttery SZ algorithm [4, 6] is a structure-preserving algorithm. It makes use of the fact that any symplectic matrix pencil can be reduced to a matrix pencil of the form ? which is determined by just 4n? parameters. By exploiting this special reduced form, the SZ algorithm is 7

3 fast and ecient in each iteration step only O(n) arithmetic operations are required instead of O(n ) arithmetic operations for a QZ step. We thus save a significant amount of work. Of course, the accumulation of the Z matrix is O(n ) as in the QZ step. Moreover, by forcing the symplectic structure the above mentioned problems of the QZ algorithm are avoided. Combined with a defect correction method, the resulting method for solving discrete-time algebraic Riccati equations is a very ecient method and produces highly accurate results. The buttery SZ algorithm For simplicity let us assume at the moment that A I 0 is nonsingular. Premultiplying K? N by 0 A?T results in a symplectic matrix pencil K 0? N 0 A 0 I?BR = A?T Q A?T?? B T 0 I where K 0 N 0 are both symplectic. In [4, 6] it is shown that for every symplectic matrix pencil L? M with LJL T = MJM T = J there exist numerous symplectic matrices Z and nonsingular matrices S such that C F 0?I S(L? M)Z =? 0 C? I T where C and F are diagonal matrices, and T is a symmetric tridiagonal matrix. Such a symplectic matrix pencil is called a symplectic buttery pencil. If T is an unreduced tridiagonal matrix, then the buttery pencil is called unreduced. If any of the n? subdiagonal elements of T are zero, the problem can be split into at least two problems of smaller dimension, but with the same symplectic buttery structure. Once the reduction to a symplectic buttery pencil is achieved, the SZ algorithm is a suitable tool for computing the eigenvalues/deating subspaces of the symplectic pencil. The SZ algorithm preserves the symplectic buttery form in its iterations. It is the analogue of the SR algorithm for the generalized eigenproblem, just as the QZ algorithm is the analogue of the QR algorithm for the generalized eigenproblem. Both are instances of the GZ algorithm [4]. Each iteration step begins with L and M such that the buttery pencil L? M is unreduced. Choose a spectral transformation function q and compute a symplectic matrix Z such that Z? q(m? N)e = e for some scalar. Then transform the pencil to fm? e N = (M? N)Z : This introduces a bulge into the matrices f M and e N. Now transform the pencil to cm? b N = S? (f M? e N) e Z where M c and N b are of symplectic buttery form. S and ez are symplectic, and Ze e = e. This concludes the iteration. Under certain assumption, it can be shown that the buttery SZ algorithm converges cubically. For a detailed discussion of the buttery SZ algorithm see [4, 6]. Given a symplectic pencil L? M, where LJL T = MJM T = J, rst symplectic matrices Z 0 and S 0 are computed such that bl? c M := S? 0 LZ 0? S? 0 MZ 0 is a symplectic buttery pencil. Using the buttery SZ algorithm, symplectic matrices Z and S are computed such that S? b LZ? S? c MZ is a symplectic buttery pencil and the symmetric tridiagonal matrix b T in the lower right block of S? c MZ is reduced to quasi-diagonal form with and blocks on the diagonal. The eigenproblem decouples into a number of simple or 4 4 generalized symplectic eigenproblems. Solving these subproblems, nally symplectic matrices Z S are computed such that S b? S? LZ Z = S? S? c MZ Z = 0 0 where the eigenvalues of the matrix pencil? are precisely the n stable generalized eigenvalues. Let Z = Z 0 Z Z. Partitioning Z conformably, Z = Z Z () Z Z the Riccati solution X is found by solving a system of linear equations: X =?Z Z? : (4) Instead of generating the symplectic matrix Z as in (), one can work with n n matrices X Y and T such that nally X =?Z Z? Y = Z? Z, and T = Z?. Starting from X = Y = 0 T = I, this can be implemented without accumulating the intermediate symplectic transformations used in the buttery SZ algorithm, just using the parameters that determine these transformations. As for every symplectic matrix Z written in the form (), Z Z? is symmetric, this approach guarantees that all intermediate (and

4 the nal) X are symmetric. Such an approach, called symmetric updating was rst proposed by Byers and Mehrmann [] in the context of solving continuous-time algebraic Riccati equations via the Hamiltonian SR algorithm and has also been proposed for solving DAREs with an SR algorithm in []. If the so computed approximate solution of the DARE is rened using Newton's method, usually the same number of iterations is required as when rening an approximation computed by the QZ algorithm. Even if one or two iterations more are necessary due to the loss of accuracy caused by using non-orthogonal transformations, this is well compensated by the cheaper SZ iteration. A hybrid method for DAREs Combined with a strategy to deate zero and innity eigenvalues from the symplectic pencil in order to deal with discrete-time algebraic Riccati equations with singular A matrix, the hybrid method consisting of the SZ algorithm followed by a few Newton iteration steps results in an ecient and accurate method. Altogether, we propose the following algorithm to solve the DARE (). Algorithm Input: The coecient matrices A IR nn, B IR nm, Q = Q T IR nn, and R IR mm. Output: An approximation X ~ = X ~ T IR nn to the stabilizing solution of the DARE.. Form the symplectic pencil K? N as in ().. Use Algorithm.6 of [] to deate all zero and innite eigenvalues of K? N. That is, compute a nonsingular transformation matrix T and a symplectic matrix S such that T (K? N)S = 4 0 A ~ A ~ I n?k 0? 0 ~ Q 0 I k 4 I n?k 0? G ~? G ~ 0 I n?k? G ~ T? G ~ ? A ~ T? A ~ T and rst n? k columns of S span the deating subspace of K? N corresponding to the zero eigenvalues.. Apply the buttery SZ algorithm described Section (and in detail in [4]) to the symplectic pencil ~K? N ~ := Ik 0 ~A 0 0 A ~?T? ~Q I k Ik? ~ G 0? ~ A T such that ~T ( K ~? N) ~ S ~ =? 0 0 where the eigenvalues of? are the stable nonzero eigenvalues of K? N. 4. Partition S ~ S = S S S where Sjj IR kk, j =. Set " In?k # S Z := S 0 S 0 0 I n?k : 0 0 S 0 S Then the rst n columns of Z span the stable deating subspace of K?N and an approximate solution b X of the DARE can be computed as in (4).. Use Newton's method endowed with a line search strategy as proposed in [] and starting guess X 0 = b X in order to iteratively rene the solution of the DARE to the highest achievable accuracy. Note that all left transformation matrices need not be accumulated. The accumulation of the right transformation matrices and the computation of b X via (4) can be avoided using the symmetric updating technique as mentioned at the end of Section. More details of the algorithm and its implementation as well as a thorough numerical study regarding performance and accuracy will be reported in []. 4 Concluding remarks We have discussed the numerical solution of discretetime Riccati equations. By initializing Newton's method with a starting guess computed by applying the buttery SZ algorithm combined with a method for deating zero and innite eigenvalues to the corresponding symplectic matrix pencil, an ecient and accurate hybrid method is derived. References [] G. Banse. Symplektische Eigenwertverfahren zur Losung zeitdiskreter optimaler Steuerungsprobleme. Dissertation, Fachbereich { Mathematik und Informatik, Universitat Bremen, Bremen, FRG, June 99. In German. [] P. Benner. Accelerating Newton's method for discrete-time algebraic Riccati equations. In Proc. MTNS 98, Padova, Italy, 998. To appear. [] P. Benner and H. Fabender. A hybrid method for the numerical solution of discrete-time algebraic Riccati equations. In preparation.

5 [4] P. Benner, H. Fabender, and D.S. Watkins. SR and SZ algorithms for the symplectic (buttery) eigenproblem. Linear Algebra Appl., 87:4{76, 999. [] R. Byers and V. Mehrmann. Symmetric updating of the solution of the algebraic Riccati equation. Methods of Operations Research, 4:7{, 98. [6] H. Fabender. Symplectic Methods for Symplectic Eigenproblems. Habilitationsschrift, Fachbereich { Mathematik und Informatik, Universitat Bremen, 84 Bremen, (Germany), 998. [7] G.H. Golub and C.F. Van Loan. Matrix Computations. Johns Hopkins University Press, Baltimore, third edition, 996. [8] G.A. Hewer. An iterative technique for the computation of steady state gains for the discrete optimal regulator. IEEE Trans. Automat. Control, AC-6:8{ 84, 97. [9] P. Lancaster and L. Rodman. The Algebraic Riccati Equation. Oxford University Press, Oxford, 99. [0] A.J. Laub. Algebraic aspects of generalized eigenvalue problems for solving Riccati equations. In C.I. Byrnes and A. Lindquist, editors, Computational and Combinatorial Methods in Systems Theory, pages { 7. Elsevier (North-Holland), 986. [] V. Mehrmann. The Autonomous Linear Quadratic Control Problem, Theory and Numerical Solution. Number 6 in Lecture Notes in Control and Information Sciences. Springer-Verlag, Heidelberg, July 99. [] V. Mehrmann and E. Tan. Defect correction methods for the solution of algebraic Riccati equations. IEEE Trans. Automat. Control, :69{698, 988. [] V. Sima. Algorithms for Linear-Quadratic Optimization, volume 00 of Pure and Applied Mathematics. Marcel Dekker, Inc., New York, NY, 996. [4] D.S. Watkins and L. Elsner. Theory of decomposition and bulge chasing algorithms for the generalized eigenvalue problem. SIAM J. Matrix Anal. Appl., :94{967, 994.