SB02SD

Estimating conditioning and forward error bound for the solution of discrete-time Riccati equation

[Specification] [Arguments] [Method] [References] [Comments] [Example]

Purpose

  To estimate the conditioning and compute an error bound on the
  solution of the real discrete-time matrix algebraic Riccati
  equation (see FURTHER COMMENTS)
                              -1
      X = op(A)'*X*(I_n + G*X)  *op(A) + Q,                      (1)

  where op(A) = A or A' (A**T) and Q, G are symmetric (Q = Q**T,
  G = G**T). The matrices A, Q and G are N-by-N and the solution X
  is N-by-N.

Specification
      SUBROUTINE SB02SD( JOB, FACT, TRANA, UPLO, LYAPUN, N, A, LDA, T,
     $                   LDT, U, LDU, G, LDG, Q, LDQ, X, LDX, SEPD,
     $                   RCOND, FERR, IWORK, DWORK, LDWORK, INFO )
C     .. Scalar Arguments ..
      CHARACTER          FACT, JOB, LYAPUN, TRANA, UPLO
      INTEGER            INFO, LDA, LDG, LDQ, LDT, LDU, LDWORK, LDX, N
      DOUBLE PRECISION   FERR, RCOND, SEPD
C     .. Array Arguments ..
      INTEGER            IWORK( * )
      DOUBLE PRECISION   A( LDA, * ), DWORK( * ),  G( LDG, * ),
     $                   Q( LDQ, * ), T( LDT, * ), U( LDU, * ),
     $                   X( LDX, * )

Arguments

Mode Parameters

  JOB     CHARACTER*1
          Specifies the computation to be performed, as follows:
          = 'C':  Compute the reciprocal condition number only;
          = 'E':  Compute the error bound only;
          = 'B':  Compute both the reciprocal condition number and
                  the error bound.

  FACT    CHARACTER*1
          Specifies whether or not the real Schur factorization of
          the matrix Ac = inv(I_n + G*X)*A (if TRANA = 'N'), or
          Ac = A*inv(I_n + X*G) (if TRANA = 'T' or 'C'), is supplied
          on entry, as follows:
          = 'F':  On entry, T and U (if LYAPUN = 'O') contain the
                  factors from the real Schur factorization of the
                  matrix Ac;
          = 'N':  The Schur factorization of Ac will be computed
                  and the factors will be stored in T and U (if
                  LYAPUN = 'O').

  TRANA   CHARACTER*1
          Specifies the form of op(A) to be used, as follows:
          = 'N':  op(A) = A    (No transpose);
          = 'T':  op(A) = A**T (Transpose);
          = 'C':  op(A) = A**T (Conjugate transpose = Transpose).

  UPLO    CHARACTER*1
          Specifies which part of the symmetric matrices Q and G is
          to be used, as follows:
          = 'U':  Upper triangular part;
          = 'L':  Lower triangular part.

  LYAPUN  CHARACTER*1
          Specifies whether or not the original Lyapunov equations
          should be solved in the iterative estimation process,
          as follows:
          = 'O':  Solve the original Lyapunov equations, updating
                  the right-hand sides and solutions with the
                  matrix U, e.g., RHS <-- U'*RHS*U;
          = 'R':  Solve reduced Lyapunov equations only, without
                  updating the right-hand sides and solutions.

Input/Output Parameters
  N       (input) INTEGER
          The order of the matrices A, X, Q, and G.  N >= 0.

  A       (input) DOUBLE PRECISION array, dimension (LDA,N)
          If FACT = 'N' or LYAPUN = 'O', the leading N-by-N part of
          this array must contain the matrix A.
          If FACT = 'F' and LYAPUN = 'R', A is not referenced.

  LDA     INTEGER
          The leading dimension of the array A.
          LDA >= max(1,N), if FACT = 'N' or  LYAPUN = 'O';
          LDA >= 1,        if FACT = 'F' and LYAPUN = 'R'.

  T       (input or output) DOUBLE PRECISION array, dimension
          (LDT,N)
          If FACT = 'F', then T is an input argument and on entry,
          the leading N-by-N upper Hessenberg part of this array
          must contain the upper quasi-triangular matrix T in Schur
          canonical form from a Schur factorization of Ac (see
          argument FACT).
          If FACT = 'N', then T is an output argument and on exit,
          if INFO = 0 or INFO = N+1, the leading N-by-N upper
          Hessenberg part of this array contains the upper quasi-
          triangular matrix T in Schur canonical form from a Schur
          factorization of Ac (see argument FACT).

  LDT     INTEGER
          The leading dimension of the array T.  LDT >= max(1,N).

  U       (input or output) DOUBLE PRECISION array, dimension
          (LDU,N)
          If LYAPUN = 'O' and FACT = 'F', then U is an input
          argument and on entry, the leading N-by-N part of this
          array must contain the orthogonal matrix U from a real
          Schur factorization of Ac (see argument FACT).
          If LYAPUN = 'O' and FACT = 'N', then U is an output
          argument and on exit, if INFO = 0 or INFO = N+1, it
          contains the orthogonal N-by-N matrix from a real Schur
          factorization of Ac (see argument FACT).
          If LYAPUN = 'R', the array U is not referenced.

  LDU     INTEGER
          The leading dimension of the array U.
          LDU >= 1,        if LYAPUN = 'R';
          LDU >= MAX(1,N), if LYAPUN = 'O'.

  G       (input) DOUBLE PRECISION array, dimension (LDG,N)
          If UPLO = 'U', the leading N-by-N upper triangular part of
          this array must contain the upper triangular part of the
          matrix G.
          If UPLO = 'L', the leading N-by-N lower triangular part of
          this array must contain the lower triangular part of the
          matrix G.                     _
          Matrix G should correspond to G in the "reduced" Riccati
          equation (with matrix T, instead of A), if LYAPUN = 'R'.
          See METHOD.

  LDG     INTEGER
          The leading dimension of the array G.  LDG >= max(1,N).

  Q       (input) DOUBLE PRECISION array, dimension (LDQ,N)
          If UPLO = 'U', the leading N-by-N upper triangular part of
          this array must contain the upper triangular part of the
          matrix Q.
          If UPLO = 'L', the leading N-by-N lower triangular part of
          this array must contain the lower triangular part of the
          matrix Q.                     _
          Matrix Q should correspond to Q in the "reduced" Riccati
          equation (with matrix T, instead of A), if LYAPUN = 'R'.
          See METHOD.

  LDQ     INTEGER
          The leading dimension of the array Q.  LDQ >= max(1,N).

  X       (input) DOUBLE PRECISION array, dimension (LDX,N)
          The leading N-by-N part of this array must contain the
          symmetric solution matrix of the original Riccati
          equation (with matrix A), if LYAPUN = 'O', or of the
          "reduced" Riccati equation (with matrix T), if
          LYAPUN = 'R'. See METHOD.

  LDX     INTEGER
          The leading dimension of the array X.  LDX >= max(1,N).

  SEPD    (output) DOUBLE PRECISION
          If JOB = 'C' or JOB = 'B', the estimated quantity
          sepd(op(Ac),op(Ac)').
          If N = 0, or X = 0, or JOB = 'E', SEPD is not referenced.

  RCOND   (output) DOUBLE PRECISION
          If JOB = 'C' or JOB = 'B', an estimate of the reciprocal
          condition number of the discrete-time Riccati equation.
          If N = 0 or X = 0, RCOND is set to 1 or 0, respectively.
          If JOB = 'E', RCOND is not referenced.

  FERR    (output) DOUBLE PRECISION
          If JOB = 'E' or JOB = 'B', an estimated forward error
          bound for the solution X. If XTRUE is the true solution,
          FERR bounds the magnitude of the largest entry in
          (X - XTRUE) divided by the magnitude of the largest entry
          in X.
          If N = 0 or X = 0, FERR is set to 0.
          If JOB = 'C', FERR is not referenced.

Workspace
  IWORK   INTEGER array, dimension (N*N)

  DWORK   DOUBLE PRECISION array, dimension (LDWORK)
          On exit, if INFO = 0 or INFO = N+1, DWORK(1) returns the
          optimal value of LDWORK.

  LDWORK  INTEGER
          The dimension of the array DWORK.
          Let LWA = N*N, if LYAPUN = 'O';
              LWA = 0,   otherwise,
          and LWN = N,   if LYAPUN = 'R' and JOB = 'E' or 'B';
              LWN = 0,   otherwise.
          If FACT = 'N', then
             LDWORK  = MAX(LWA + 5*N, MAX(3,2*N*N) + N*N),
                                              if JOB = 'C';
             LDWORK  = MAX(LWA + 5*N, MAX(3,2*N*N) + 2*N*N + LWN),
                                              if JOB = 'E' or 'B'.
          If FACT = 'F', then
             LDWORK  = MAX(3,2*N*N) + N*N,    if JOB = 'C';
             LDWORK  = MAX(3,2*N*N) + 2*N*N + LWN,
                                              if JOB = 'E' or 'B'.
          For good performance, LDWORK must generally be larger.

          If LDWORK = -1, then a workspace query is assumed;
          the routine only calculates the optimal size of the
          DWORK array, returns this value as the first entry of
          the DWORK array, and no error message related to LDWORK
          is issued by XERBLA.

Error Indicator
  INFO    INTEGER
          = 0:  successful exit;
          < 0:  if INFO = -i, the i-th argument had an illegal
                value;
          > 0:  if INFO = i, i <= N, the QR algorithm failed to
                complete the reduction of the matrix Ac to Schur
                canonical form (see LAPACK Library routine DGEES);
                on exit, the matrix T(i+1:N,i+1:N) contains the
                partially converged Schur form, and DWORK(i+1:N) and
                DWORK(N+i+1:2*N) contain the real and imaginary
                parts, respectively, of the converged eigenvalues;
                this error is unlikely to appear;
          = N+1:  if T has almost reciprocal eigenvalues; perturbed
                values were used to solve Lyapunov equations, but
                the matrix T, if given (for FACT = 'F'), is
                unchanged.

Method
  The condition number of the Riccati equation is estimated as

  cond = ( norm(Theta)*norm(A) + norm(inv(Omega))*norm(Q) +
              norm(Pi)*norm(G) ) / norm(X),

  where Omega, Theta and Pi are linear operators defined by

  Omega(W) = op(Ac)'*W*op(Ac) - W,
  Theta(W) = inv(Omega(op(W)'*X*op(Ac) + op(Ac)'X*op(W))),
     Pi(W) = inv(Omega(op(Ac)'*X*W*X*op(Ac))),

  and Ac = inv(I_n + G*X)*A (if TRANA = 'N'), or
      Ac = A*inv(I_n + X*G) (if TRANA = 'T' or 'C').

  Note that the Riccati equation (1) is equivalent to

      X = op(Ac)'*X*op(Ac) + op(Ac)'*X*G*X*op(Ac) + Q,           (2)

  and to
      _          _                _ _ _         _
      X = op(T)'*X*op(T) + op(T)'*X*G*X*op(T) + Q,               (3)
        _           _               _
  where X = U'*X*U, Q = U'*Q*U, and G = U'*G*U, with U the
  orthogonal matrix reducing Ac to a real Schur form, T = U'*Ac*U.

  The routine estimates the quantities

  sepd(op(Ac),op(Ac)') = 1 / norm(inv(Omega)),

  norm(Theta) and norm(Pi) using 1-norm condition estimator.

  The forward error bound is estimated using a practical error bound
  similar to the one proposed in [2].

References
  [1] Ghavimi, A.R. and Laub, A.J.
      Backward error, sensitivity, and refinement of computed
      solutions of algebraic Riccati equations.
      Numerical Linear Algebra with Applications, vol. 2, pp. 29-49,
      1995.

  [2] Higham, N.J.
      Perturbation theory and backward error for AX-XB=C.
      BIT, vol. 33, pp. 124-136, 1993.

  [3] Petkov, P.Hr., Konstantinov, M.M., and Mehrmann, V.
      DGRSVX and DMSRIC: Fortran 77 subroutines for solving
      continuous-time matrix algebraic Riccati equations with
      condition and accuracy estimates.
      Preprint SFB393/98-16, Fak. f. Mathematik, Tech. Univ.
      Chemnitz, May 1998.

Numerical Aspects
                            3
  The algorithm requires 0(N ) operations.
  The accuracy of the estimates obtained depends on the solution
  accuracy and on the properties of the 1-norm estimator.

Further Comments
  The option LYAPUN = 'R' may occasionally produce slightly worse
  or better estimates, and it is much faster than the option 'O'.
  When SEPD is computed and it is zero, the routine returns
  immediately, with RCOND and FERR (if requested) set to 0 and 1,
  respectively. In this case, the equation is singular.

  Let B be an N-by-M matrix (if TRANA = 'N') or an M-by-N matrix
  (if TRANA = 'T' or 'C'), let R be an M-by-M symmetric positive
  definite matrix (R = R**T), and denote G = op(B)*inv(R)*op(B)'.
  Then, the Riccati equation (1) is equivalent to the standard
  discrete-time matrix algebraic Riccati equation

      X = op(A)'*X*op(A) -                                       (4)
                                             -1
          op(A)'*X*op(B)*(R + op(B)'*X*op(B))  *op(B)'*X*op(A) + Q.

  By symmetry, the equation (1) is also equivalent to
                            -1
      X = op(A)'*(I_n + X*G)  *X*op(A) + Q.

Example

Program Text

*     SB02SD EXAMPLE PROGRAM TEXT
*     Copyright (c) 2002-2017 NICONET e.V.
*
*     .. Parameters ..
      INTEGER          NIN, NOUT
      PARAMETER        ( NIN = 5, NOUT = 6 )
      INTEGER          NMAX
      PARAMETER        ( NMAX = 20 )
      INTEGER          LDA, LDG, LDQ, LDT, LDU, LDX
      PARAMETER        ( LDA = NMAX, LDG = NMAX, LDQ = NMAX, LDT = NMAX,
     $                   LDU = NMAX, LDX = NMAX )
      INTEGER          LIWORK
      PARAMETER        ( LIWORK = NMAX*NMAX )
      INTEGER          LDWORK
      PARAMETER        ( LDWORK = 8*NMAX*NMAX + 10*NMAX )
      DOUBLE PRECISION ZERO, ONE
      PARAMETER        ( ZERO = 0.0D+0, ONE = 1.0D+0 )
*     .. Local Scalars ..
      DOUBLE PRECISION FERR, RCND, RCOND, SEPD
      INTEGER          I, INFO1, INFO2, INFO3, IS, IU, IW, J, N, N2,
     $                 SDIM
      CHARACTER*1      FACT, JOB, JOBS, LYAPUN, TRANA, TRANAT, UPLO
*     .. Local Arrays ..
      LOGICAL          BWORK(2*NMAX)
      INTEGER          IWORK(LIWORK)
      DOUBLE PRECISION A(LDA,NMAX), AS(LDA,NMAX), DWORK(LDWORK),
     $                 G(LDG,NMAX), Q(LDQ,NMAX), T(LDT,NMAX),
     $                 U(LDU,NMAX), X(LDX,NMAX)
*     .. External Functions ..
      LOGICAL          LSAME, SELECT
      EXTERNAL         LSAME, SELECT
*     .. External Subroutines ..
      EXTERNAL         DGEES, DGESV, DLACPY, DLASET, DSWAP, DSYMM,
     $                 MA02AD, MA02ED, MB01RU, SB02MD, SB02SD
*     .. Executable Statements ..
*
      WRITE ( NOUT, FMT = 99999 )
*     Skip the heading in the data file and read the data.
      READ ( NIN, FMT = '()' )
      READ ( NIN, FMT = * ) N, JOB, FACT, TRANA, UPLO, LYAPUN
      IF ( N.LT.0 .OR. N.GT.NMAX ) THEN
         WRITE ( NOUT, FMT = 99993 ) N
      ELSE
         READ ( NIN, FMT = * ) ( ( A(I,J), J = 1,N ), I = 1,N )
         READ ( NIN, FMT = * ) ( ( Q(I,J), J = 1,N ), I = 1,N )
         READ ( NIN, FMT = * ) ( ( G(I,J), J = 1,N ), I = 1,N )
         CALL DLACPY( 'Full', N, N, A, LDA, AS, LDA )
         CALL DLACPY( UPLO, N, N, Q, LDQ, X, LDX )
         N2 = 2*N
         IS = 2*N2 + 1
         IU = IS + N2*N2
         IW = IU + N2*N2
*        Solve the discrete-time Riccati equation.
         CALL SB02MD( 'discrete', 'direct', UPLO, 'no scaling',
     $                'stable', N, AS, LDA, G, LDG, X, LDX, RCND,
     $                DWORK(1), DWORK(N2+1), DWORK(IS), N2, DWORK(IU),
     $                N2, IWORK, DWORK(IW), LDWORK-IW+1, BWORK, INFO1 )
*
         IF ( INFO1.EQ.0 ) THEN
            WRITE ( NOUT, FMT = 99995 )
            DO 10 I = 1, N
               WRITE ( NOUT, FMT = 99994 ) ( X(I,J), J = 1,N )
   10       CONTINUE
            IF ( LSAME( FACT, 'F' ) .OR. LSAME( LYAPUN, 'R' ) ) THEN
               CALL DLASET( 'Full', N, N, ZERO, ONE, DWORK, N )
               CALL DSYMM( 'Left', UPLO, N, N, ONE, G, LDG, X, LDX,
     $                     ONE, DWORK, N )
               IF ( LSAME( TRANA, 'N' ) ) THEN
*                 Compute Ac = inv(I_n + G*X)*A.
                  CALL DLACPY( 'Full', N, N, A, LDA, T, LDT )
                  CALL DGESV( N, N, DWORK, N, IWORK, T, LDT, INFO3 )
               ELSE
*                 Compute Ac = A*inv(I_n + X*G)
                  CALL MA02AD( 'Full', N, N, A, LDA, T, LDT )
                  CALL DGESV( N, N, DWORK, N, IWORK, T, LDT, INFO3 )
                  DO 20 J = 2, N
                     CALL DSWAP( J-1, T(1,J), 1, T(J,1), LDT )
   20             CONTINUE
               END IF
*              Compute the Schur factorization of Ac.
               JOBS = 'V'
               CALL DGEES( JOBS, 'Not ordered', SELECT, N, T, LDT, SDIM,
     $                     DWORK(1), DWORK(N+1), U, LDU, DWORK(2*N+1),
     $                     LDWORK-2*N, BWORK, INFO3 )
               IF( INFO3.NE.0 ) THEN
                  WRITE ( NOUT, FMT = 99996 ) INFO3
                  STOP
               END IF
            END IF
*
            IF ( LSAME( LYAPUN, 'R' ) ) THEN
               IF( LSAME( TRANA, 'N' )  ) THEN
                  TRANAT = 'T'
               ELSE
                  TRANAT = 'N'
               END IF
*
               CALL MB01RU( UPLO, TRANAT, N, N, ZERO, ONE, X, LDX,
     $                      U, LDU, X, LDX, DWORK, N*N, INFO2 )
               CALL MA02ED( UPLO, N, X, LDX )
               CALL MB01RU( UPLO, TRANAT, N, N, ZERO, ONE, G, LDG,
     $                      U, LDU, G, LDG, DWORK, N*N, INFO2 )
               CALL MB01RU( UPLO, TRANAT, N, N, ZERO, ONE, Q, LDQ,
     $                      U, LDU, Q, LDQ, DWORK, N*N, INFO2 )
            END IF
*           Estimate the condition and error bound on the solution.
            CALL SB02SD( JOB, FACT, TRANA, UPLO, LYAPUN, N, A, LDA, T,
     $                   LDT, U, LDU, G, LDG, Q, LDQ, X, LDX, SEPD,
     $                   RCOND, FERR, IWORK, DWORK, LDWORK, INFO2 )
*
            IF ( INFO2.NE.0 ) THEN
               WRITE ( NOUT, FMT = 99997 ) INFO2
            END IF
            IF ( INFO2.EQ.0 .OR. INFO2.EQ.N+1 ) THEN
               WRITE ( NOUT, FMT = 99992 ) SEPD
               WRITE ( NOUT, FMT = 99991 ) RCOND
               WRITE ( NOUT, FMT = 99990 ) FERR
            END IF
         ELSE
            WRITE ( NOUT, FMT = 99998 ) INFO1
         END IF
      END IF
      STOP
*
99999 FORMAT (' SB02SD EXAMPLE PROGRAM RESULTS',/1X)
99998 FORMAT (' INFO on exit from SB02MD =',I2)
99997 FORMAT (' INFO on exit from SB02SD =',I2)
99996 FORMAT (' INFO on exit from DGEES  =',I2)
99995 FORMAT (' The solution matrix X is')
99994 FORMAT (20(1X,F8.4))
99993 FORMAT (/' N is out of range.',/' N = ',I5)
99992 FORMAT (/' Estimated separation = ',F8.4)
99991 FORMAT (/' Estimated reciprocal condition number = ',F8.4)
99990 FORMAT (/' Estimated error bound = ',F8.4)
      END
Program Data
 SB02SD EXAMPLE PROGRAM DATA
   2     B     N     N     U     O
   2.0 -1.0
   1.0  0.0
   0.0  0.0
   0.0  1.0
   1.0  0.0
   0.0  0.0
Program Results
 SB02SD EXAMPLE PROGRAM RESULTS

 The solution matrix X is
  -0.7691   1.2496
   1.2496  -2.3306

 Estimated separation =   0.4456

 Estimated reciprocal condition number =   0.1445

 Estimated error bound =   0.0000

Return to index