The solution of the equation AXDBXC=E is discussed, partly in terms of the generalized eigenproblem. Useful applications arise in connection with the numerical solution of implicit differential equations.
The solution of the equation AXDBXC=E is discussed, partly in terms of the generalized eigenproblem. Useful applications arise in connection with the numerical solution of implicit differential equations.
The solution of the equation AXDBXC=E is discussed, partly in terms of the generalized eigenproblem. Useful applications arise in connection with the numerical solution of implicit differential equations.
AND ITS APPLICATION IN THE NUMERICAL SOLUTION OF IMPLICIT ORDINARY DIFFERENTIAL EQUATIONS MICHAEL A. EPTON Abstract. The solution of the equation AXD- BXC=E is discussed, partly in terms of the generalized eigenproblem. Useful applications arise in connection with the numerical solution of implicit differential equations. Introduction. We consider here the linear equation for X (1) AXD- BXC = E where A, B~R mxm, C, De R ' ~xnand X, E~R mxn By using tensor products together with the appropriate definition of ~ and 5, equation (1) can be written (1)' [ ( ADT) - (BCT)]3c = ~. This equation is a generalization of the system (2) AX- XC = E studied by Gantmacher [1], Bartels and Stewart [2] and others. Gantmacher has shown that (2) has a unique solution if spectrum (A) N spectrum (C) = ~, and has further given an explicit solution provided the Jordan decompositions are given for A and C. Bartels and Stewart have shown how (2) may be solved if just a Schur type decomposition is available for A and C. Recently, Enright [3] has observed that (2) may be quite efficiently solved if one matrix, say C, is reduced to Schur form while A is reduced to Hessenberg. The essential business of this paper is to extend these ideas to the system (1). Received May 16, 1979. Revised August 1, 1980. 342 MICHAEL A. EPTON Discussion. The existence and uniqueness of a solution to (1) is determined by the spectral decompositions of the matrix pencils A- 2B, C- 2D. In fact, (1) may be transformed into (2) using a shift technique (cf. [1], V.2, p. 28). E = AXD - BXC = (A - )~B + )~B)XD - BXC = ( A- 2B) XD- BX( C- ) ~D) . Assuming that A- 2B and C- 2D are regular pencils and that 2 is not an eigenvalue of either, we may pre-multiply by ( A- 2B) -1, post-multiply by ( C- 2D)- 1 and obtain (3) ( A- 2B) - I E( C- 2D) -1 = X[ D( C- ) ~D) - a] - [ ( A- 2B) - I B] X which has essentially the same form as equation (2). The transformation indicated by equation (3) represents a perfectly acceptable approach to the solution of (1). However, its success depends upon a choice for the shift 2 that makes the matrices ( A- ) ~B) and ( C- 2D) well conditioned. In what follows we shall show how equation (1) may be approached directly, using techniques associated with the numerical treatment of the generalized eigenproblem. Without loss of generality we assume that the row dimension m of X and E exceeds or equals the column dimension n. If this is not originally true, it can be achieved simply by transposing equation (1). When this assumption is made, the algorithm to be described presently is essentially optimal with respect to arithmetic operations. Using the theory for the generalized eigenproblem we know that there exist transformations Q and Z ~ R "" such that QCZ and QDZ are upper triangular. (Such transformations are provided by the Q- z code of Stewart and Moler [4] and its refinement by Ward [5]; also pertinent is the L- Z algorithm of Kaufmann [6]). Indeed, Gantmacher shows how to find transformations Q and Z such that QCZ and QDZ are in Jordan normal form, but we usually avoid this transformation because its computation tends to be a numerically poorly conditioned process*. Setting (4.a), (4.b) QCZ = M, QDZ = A and defining X', E' by (5.a), (5.b) X' = x o- 1, E' = EZ we see that (1) takes the form AXQ- 1QDZ- BXQ- 1QCZ = EZ or (6) AX' A- BX' M = E' . * Counterexamples to this advice do exist. See Applications below. METHODS FOR THE SOLUTI ON OF AXD- BXC=E . . . 343 Denoting the elements of A and M by -;~ij,/~ij, the columns of X', E' by x~, e~ where i = l(1)n, and using the fact that A and M are upper triangular, the j t h column of equation (6) reads J J (7) A ~ x'i2~j-B ~ x'~pij = e~. i =1 i =l These equations may be solved sequentially for x'~ i =1(1)n by solving the n systems (8) (,~jjA - #jiB)x) = e' j- ~, (2oA - I~,jB)xl. i <j Moreover, systems of equations of the form (8) may be quite efficiently solved if transformations R and S are found such that RAS and RBS are both upper triangular or even upper Hessenberg. The simultaneous reduction of A and B to upper Hessenberg form may be performed most efficiently by the preliminary routines in Kaufmann' s L- Z generalized eigenvalue package [6], and less efficiently but more stably by the preliminary routines for Stewart and Moler' s ( 2- Z generalized eigenvalue package. With R and S computed such that RAS and RBS are Hessenberg, the resulting systems (9) (2~jRAS-t~j~RBS)(S-lx' ~)= RI e~- ~ ()~jA-p~fl)x'~l i <j may be solved for S-~x~ quite readily. Because RAS and RBS are Hessenberg, so is 2~jRAS-I~jjRBS, and such matrices can be factored in about (n2/2) multiplies. Since S is generally of simple form (in the L- Z algorithm, S is a product of elementary permutations and elementary lower triangular transformations), xj is readily computed by (10) x} = S( S- I x ) . Applications. The applications of most immediate interest to the author and ultimately the stimulus of these remarks arise in the implementation of implicit Runge-Kutta integration formulae [73 and block multistep formulae [8] for the numerical solution of implicit differential equations (i.e., equations of the form g(2, x)=0, [93.) Because implicit Runge-Kutta formulae are a subcase of the block multistep formulae, we treat this latter case in detail and conclude with some remarks about the simplifications possible for I RK formulae. As used to obtain a solution x(t) of g( 2, x) =0, the block multistep methods work as follows. Suppose one has available quantities (2j,,_p, xj, n_~), j = 1(1)v, p = 1 (1)k that approximate 2(t), x(t) at past times tj, n_ p = tn_p_ ~ + hcj. (Usually the numbers c~ satisfy O~=cj~_ 1.) The essential idea here is that for fixed p, time points 344 MICHAEL A. EPTON t~,._p, lie in the interval [t._p_ 1, t._p] and xj , . _p 5%. _p represent the solution x( t ) and its derivative ~(t) at these time points. The integer v, called the block size of the method, is the number of estimates of x and ~ generated at each integration step. To advance the method from time t._a to time t., one then requires that quantities x~.., ~j.. associated with the interval [t. _ 1, t.] satisfy the differential equation, (11) g(5:j,., xj , . ) = 0 and simultaneously, the block muitistep discretization formulae = 0, , =l , , v. p=0 To solve the pair of equations (11) and (12) for x~.,, 2~.,, a variant of Newton' s method is used that at stage l of the iteration, enforces upon perturbations .s,n - - Xj , n " ' Ln - - j , n the conditions (13) g(2}t,)., x j,.,) ~ + [Og/~Yc] 65:(],)., + [~g/Ox] ~x}l). = 0 j = 1 . . . . . v (14) rl t) + ~, e r A ) ~ ) -- ht~ ()~a) ~ = 0 x-' i j ~oj , n " r ' i j ~j , n, t 3=1 where r ~ is defined as the amount by which x},,, x(~), fail to satisfy (12): r~t) ~ t.(o)y0) _ hfl(ob?(o = ~' i j ~j , n "- r i j --j, nJ j =l (15) + p~l [j =~ ( ~x J ' " - P- hf l l z ] ) ~J ' " - P) l " In (13), [Og/OS:] and [0g/0x] denote the Jacobians of g with' respect to its first and second argument evaluated at some characteristic value of (2, x). In general, the solution of the pair (13), (14) may be accomplished by solving two systems of the form (1), as follows. Define matrices A, B, C, D, G (z), R (, 6X ( and 6X a) by A = (1/h)[Og/c35c], C = - r M) l T Lt ' l J J ' GO) = Lgr :~(/)~,., x~,))], Then (13) and (14) can be written (13)' (14)' B = [ag/c3x], D (o) T = ] R"' = G a) + hA3X (0 + B6X a) = 0 R ) + bXt OD + h6X ) C = 0. METHODS FOR THE SOLUTION OF AXD- BXC=E. . . 345 Postmultiplying (13)' by C and D, respectively, and substituting (14)', then yields for f i X (l) and 61~ (I), (16) A 5 X, (t)D- B jX(t)C = - AR ( + G(I)C (17) A 5fifa)D - B 6~( C = (l/h) BR (0 - G(D both of which are identical in form to (1). Note that because both of the matrices (o) O = [eij ], C = - [ill )] may be singular, it is generally necessary to solve both (16) and (17). If one of these is nonsi.ngular, it is better to solve for just one of 5X (1), 5Jf(~) and obtain the other from (14)'. In the special case that the method is an implicit Runge-Kutta method, we have e(o) =Stj (Kronecker delta) and only one system needs to be solved. It is ij conventional to solve (17). If the I RK method is also a collocation method, then the matrix pencil C- 2D C - )~D = (o) T " - U~i j ] - ~[61~3 = - (/~(O)T + ; , 1) can be reduced to Jordan canonical form by means of a similarity transformation that is explicitly computable. Thus, in this particular instance, the remark made above advising against the use of Jordan normal forms is irrelevant. REFERENCES 1. F. R. Gantmacher, Theory of Matrices, Chelsea Publishing Co., New York, N.Y. (1977). 2. R. H. Bartels and G. W. Stewart, Algorithm 432, Solution of the matrix equation AX + XB= C, ACM, 15 (1972), 214-235. 3. W. H. Enright, Improving the efficiency of matrix operations in the numerical solution of stiff ordinary differential equations. ACM Trans. Math. Software, 4, No. 2 (1978), 127-136. 4. C. B. Moler and G. W. Stewart, An algorithm for generalized matrix eigenvalue problems, Sl AM J. Num. Anal., 10, No. 2 (1973), 241-256. 5. R. C. Ward, The combination shift QZ algorithm, SIAM J. Num. Anal., 12, No. 6 (1975), 835-853. 6. L. Kaufmann, The LZ algorithm to solve the generalized eigenvalue problem, 11, No. 5 (1974), 997- 1024. 7. J. C. Butcher, Implicit Runge-Kutta processes, Math. Comp, 18, (1964), 50-64. 8. C. W. Gear, Simultaneous numerical solutions of differential-algebraic equations, IEEE Trans. Circuit Theory, CT-18, No. 1 (1971), 89-95. 9. T. A. Bickart and Z. Picel, High order stiffly stable composite multistep methods for numerical integration of stiff differential equations, BIT 13, (1973), 272-286. BOEING COMPUTER SERVICES COMPANY 565 ANDOVER PARK WEST, M/S 9C--01 TUKWILA, WASHINGTON 98188 U.S,A. BIT 20--23