International Mathematical Forum, 5, 2010, no. 33, 1637-1644 On Solving Large Algebraic Riccati Matrix Equations Amer Kaabi Department of Basic Science Khoramshahr Marine Science and Technology University Khoramshahr, Iran kaabi amer@yahoo.com kaabi amer@kmsu.ac.ir Abstract In this paper, we present a numerical method for solving large continuoustime algebraic Riccati equations. This method is based on the global FOM algorithm and we call it by global FOM-Riccati-Like (GFRL) algorithm. Keywords: Argebraic Riccati equations; Modified global Arnoldi; Krylov subspaces; FOM 1 Introduction In this paper, we are concerned with the numerical solution of large continuoustime algebraic Riccati equations (CAREs) A T X + XA XBB T X + CC T = 0 (1) where A, X IR n n,b IR n p, C IR n s, with s n, and p n. we present a numerical method for solving this kind of Riccati matrix equation. This kind of matrix equation arise in control theory, linear quadratic regular
1638 A. Kaabi problems, H or H 2 control, model reduction problems and many other cases; see, e. g., [1, 2, 4, 5, 9]. In addition Newton method and its variants, the methods which are based on eigenvector approaches and consist in computing the Lagrange invariant subspaces of a Hamiltonian matrix are used for small dimensional problem [3, 9, 10]. Recently, Jbilou et al. [8] have proposed the global FOM and GMRES algorithms for solving matrix equations AX = B, where A IR n n, and B IR n s. Also, In [6, 7], he proposed methods for approximating solution of large and sparse continuous-time algebraic Riccati matrix equations. His proposed methods are methods of projection onto a matrix Krylov subspace. He used a matrix Arnoldi process to construct an orthonormal basis. In this paper, we present two new algorithms based on block Krylov subspace methods for solving large and sparse continuous-time algebraic Riccati matrix equations. Throughout this paper, we use the following notations. Let IE = IR n n. For X and Y in IE, we define the inner product <X,Y > F = tr(x T Y ), where tr(z) denotes the trace of the square matrix Z, and X T denotes the transpose of the matrix X. The associated norm is the well-known Frobenius norm denoted by. F. For V IE the matrix Krylov subspace K m (A, V ) is defined as follows: K m (A, V )=span{v, AV,..., A m V }, which is a subspace of IE. A set of members of IE is said to be F-orthonormal if it is orthonormal with respect to scaler product <.,.> F. This paper is organized as follows. In Section 2 we give a brief discussion of Arnoldi algorithm. The new algorithm is given in Section 3. Some concluding remarks are given in Section 4. 2 Global Arnoldi process. Let V IE, as we see in [8] global Arnoldi process construct an F-orthonormal basis V 1,V 2,..., V m, of Krylov subspace K m (A, V ). It is clear that, each U in
On solving large algebraic Riccati matrix equations 1639 K m (A, V ) can be written by the following formula U = m 1 i=0 α i A i V, where α i, i =0, 1,..., m 1 are scalers. The modified global Arnoldi algorithm for constructing an F-orthogonal basis of this Krylov subspace is as following: Algorithm 1. Modified Global Arnoldi algorithm 1. Choose an n n matrix V 1 such that V F =1. 2. For j =1,..., m Do: 3. Compute W := AV j 4. For i =1,..., j Do: 5. h i, j = tr(vi T W ) 6. W = W h i, j V i 7. EndDo 8. h j+1, j = W F. If h j+1, j = 0 then stop. 9. V j+1 = W/h j+1, j. 10. EndDo. Let us collect the matrices V i constructed by the algorithm 1 in the n mn and n (m +1)n orthonormal matrices V m =[V 1,..., V m ] and V m+1 =[V m,v m+1 ] and also we denote by H m the upper m m Hessenberg matrix whose entries are the scalers h ij and the (m +1) m matrix H m is the same as H m except for an additional row whose only nonzero element is h m+1, m in the (m +1,m) position. Next, we use the notation for the following product: m V m y = α i V i, (2) i=1 where α =(α 1,α 2,..., α m ) is a vector in IR m and, by same way, we set V m H m =[V m Hm 1,..., V m Hm m ], (3) where H j m denote the jth column of the matrix H m. It can be easily seen that V m (α + β) =V m α + V m β, (V m H m ) α = V m (H m α) (4)
1640 A. Kaabi where α and β are two vectors in IR m. The following propositions and theorem have proved in [8]. Proposition 1. If the global Arnoldi algorithm does not before the mth step, then the block system {V 1,..., V m } is an F-orthogonal basis of the matrix Krylov subspace K m (A, V 1 ). Proposition 2. Let V m =[V 1,..., V m ] where the n n matrices V i,i=1,..., m are defined by the global Arnoldi algorithm. Then we have V m α F = α 2, (5) where α IR m. Theorem 1 Let V m,h m and H m as defined before, then using the product, the following relations hold: AV m = V m H m + Z m+1, (6) where Z m+1 = h m+1,m [0 n n,..., 0 n n,v m+1 ], and AV m = V m+1 H m. (7) In [8], Jbilou et al. proposed the global FOM and GMRES algorithms, which are based on Algorithm 1, for solving linear system with multiple right hand sides. In this paper we use this idea for proposing two global methods for solving large algebraic Riccati matrix equations. 3 Theoretical Description of Global FOM-Riccati- Like (GFRL) Algorithm. As in [8], we define the operator T: IR n n IR n n, X T(X) =A T X + XA XBB T X. Hence Eq(1) can be expressed as T(X) = CC T. (8)
On solving large algebraic Riccati matrix equations 1641 In this section we proposed a new method for the computation of approximate solution to (8). This method is based on modified global Arnoldi algorithm. Given an initial guess X 0 and the corresponding residual R 0 = CC T T(X 0 ), the (GFRL) seeks an approximate solution X m from the affine subspace X 0 + K m (T,R 0 ) of dimension m by imposing the Galerkin condition R m = CC T T(X m ) F K m (T,R 0 ). (9) Here note that T k (R 0 ) is defined recursively as T(T (k 1) R 0 ). Now we recall the following proposition. Proposition 3. Let T be the operator described before and assume, That R 0 is of full rank. Then K m (T,R 0 )=K m (A T,R 0 ). Proof. See [8]. From this proposition and by the fact that X m X 0 + K m (T,R 0 ), it is easily seen that X m X 0 = Z m K m (A T,R 0 ), (10) and Relation (10) implies that R m = CC T T(X m ) F K m (A T,R 0 ). (11) X m = X 0 + V m y m for some y m IR m. Hence from the orthogonality relation (11) we have <R 0 A T (V m y m ) (V m y m )A+(V m y m )BB T (V m y m ),V i > F =0, i =1,..., m, (12) or m <A T V j +V j A V j B T BV j y m (j),v i > F y m (j) =< R 0,V i > F, i =1,..., m. (13) j=1
1642 A. Kaabi where y m =(y m (1),..., y(m) m )T. Now, if we let V 1 = R 0 / R 0 F, then it can be easily seen that the relation (13) is equivalent to the following system: (H T m + P m m k=1 y (k) m L(k) m )y m = R 0 F e 1 (14) where e 1 is the first canonical basis vector in IR m and H m is obtained by modified global Arnoldi algorithm and the entries of matrix P m is obtained by the following p i,j =<V j A, V i > F i, j =1,..., m. Also the entries of matrices L (k) m,k=1,..., m is obtained by the following l (k) i,j =< V k BB T V j,v i > F i, j =1,..., m, k =1,..., m. Thus, by These relations we can define the GFRL algorithm summarized as follows. Algorithm 2. Restarted global FOM Riccati-like algorithm (GFRL) for the solution of (1) 1. Choose an initial approximate solution X 0 and a tolerance ɛ. 2. Compute R 0 = A T X 0 + X 0 A X 0 BB T X 0 + CC T and V 1 = R 0 / R 0 F. 3. If R 0 F <ɛthen exit. 4. For j=1,...,m apply Algorithm 1 to compute the F- Orthonormal basis V 1,V 2,..., V m of K m (A T,R 0 ). 5. Define the m m matrices L m (k) and P m with l (k) i,j =< V k BB T V j,v i > F, p i,j =<V i B,V j > F, i,j,k =1,..., m, entries. 6. Solve (H m + P m m k=1 y m (k) L(k) m )y m = R 0 F e 1 for y m. 7. Compute X m = X 0 + m y m, and set X 0 = X m. 8. Go to 2. As FOM algorithm, GFRL is impractical when m is large because the of the growth of memory and computational requirements as m increases. One way to overcome this problem is to restart the Algorithm GFRL every m iterations. In the GFRL algorithm a nonlinear system of dimension m should be solved.
On solving large algebraic Riccati matrix equations 1643 4 Conclusion. We have presented global FOM Riccati-like algorithm for solving large scale Riccati matrix equations. The algorithm use the global Arnoldi process to generate F-Orthonormal bases of certain Krylov subspaces and reduce the Riccati matrix equations to a small nonlinear system. Thus, the computer storage and arithmetic work required are reduced. References [1] B. D. O. Anderson, J. B. Moore, Linear Optimal control, Prentice- Hall, Englewood Cliffs, Nj, 1971. [2] A. C. Antoulas. Approximation of large-scale Dynamical Systems. Advances in Design and Control. SIAM, Philadelphia, 2005. [3] P. Benner, R. Byers, An exact line search method for solving generalized continuous algebraic Riccati equations, IEEE Trans. Automat. Control 43(1) (1998) 101-107. [4] B. N. Datta and K.Datta, Theoretical and computationalaspects of some linear algebra problems in control theory, in Computational and Combinatorical Methods in Systems Theory (C. I. Byrnes and A. Lindquist,Eds.), Elsevier, Amsterdam,1986,pp. 201-212 [5] J. C. Doyle, K. Glover, P. P. Khargonekar, B. A. Francis, State-space solution to standard H 2 and H control problems, IEEE Trans. Automat. Control 34(8) (1989) 831-864. [6] K. Jbilou, An Arnoldi based algorithm for large algebraic Riccati equations, Appl. Math. Lett. 19 (2006) 437-444. [7] K. Jbilou, Block Krylov subspace methods for large algebraic Riccati equations Numer. Algorithms 34 (2003) 339-353. [8] K. Jbilou, A. Messaoudi, H. Sadok, Global FOM and GMRES algorithms for matrix equations, Appl. Numer. Math. 31 (1999) 97-109.
1644 A. Kaabi [9] P. Lancaster, L. Rodman, The Algebraic Riccati Equations, Clarendon Press, Oxford, 1995. [10] P. Van Dooren, A generalized eigenvalue approach for solving Riccati equations, SIAM J. Sci. Statist. Comput. 2 (1981) 121-135. Received: January, 2010