IB01QD

Estimating initial state and system matrices B and D, given A, C, and input-output trajectories

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

Purpose

  To estimate the initial state and the system matrices  B  and  D
  of a linear time-invariant (LTI) discrete-time system, given the
  matrix pair  (A,C)  and the input and output trajectories of the
  system. The model structure is :

        x(k+1) = Ax(k) + Bu(k),   k >= 0,
        y(k)   = Cx(k) + Du(k),

  where  x(k)  is the  n-dimensional state vector (at time k),
         u(k)  is the  m-dimensional input vector,
         y(k)  is the  l-dimensional output vector,
  and  A, B, C, and D  are real matrices of appropriate dimensions.
  Matrix  A  is assumed to be in a real Schur form.

Specification
      SUBROUTINE IB01QD( JOBX0, JOB, N, M, L, NSMP, A, LDA, C, LDC, U,
     $                   LDU, Y, LDY, X0, B, LDB, D, LDD, TOL, IWORK,
     $                   DWORK, LDWORK, IWARN, INFO )
C     .. Scalar Arguments ..
      DOUBLE PRECISION   TOL
      INTEGER            INFO, IWARN, L, LDA, LDB, LDC, LDD, LDU,
     $                   LDWORK, LDY, M, N, NSMP
      CHARACTER          JOB, JOBX0
C     .. Array Arguments ..
      DOUBLE PRECISION   A(LDA, *), B(LDB, *), C(LDC, *), D(LDD, *),
     $                   DWORK(*),  U(LDU, *), X0(*), Y(LDY, *)
      INTEGER            IWORK(*)

Arguments

Mode Parameters

  JOBX0   CHARACTER*1
          Specifies whether or not the initial state should be
          computed, as follows:
          = 'X':  compute the initial state x(0);
          = 'N':  do not compute the initial state (x(0) is known
                  to be zero).

  JOB     CHARACTER*1
          Specifies which matrices should be computed, as follows:
          = 'B':  compute the matrix B only (D is known to be zero);
          = 'D':  compute the matrices B and D.

Input/Output Parameters
  N       (input) INTEGER
          The order of the system.  N >= 0.

  M       (input) INTEGER
          The number of system inputs.  M >= 0.

  L       (input) INTEGER
          The number of system outputs.  L > 0.

  NSMP    (input) INTEGER
          The number of rows of matrices  U  and  Y  (number of
          samples,  t).
          NSMP >= N*M + a + e,  where
          a = 0,  if  JOBX0 = 'N';
          a = N,  if  JOBX0 = 'X';
          e = 0,  if  JOBX0 = 'X'  and  JOB = 'B';
          e = 1,  if  JOBX0 = 'N'  and  JOB = 'B';
          e = M,  if  JOB   = 'D'.

  A       (input) DOUBLE PRECISION array, dimension (LDA,N)
          The leading N-by-N part of this array must contain the
          system state matrix  A  in a real Schur form.

  LDA     INTEGER
          The leading dimension of the array A.  LDA >= MAX(1,N).

  C       (input) DOUBLE PRECISION array, dimension (LDC,N)
          The leading L-by-N part of this array must contain the
          system output matrix  C  (corresponding to the real Schur
          form of  A).

  LDC     INTEGER
          The leading dimension of the array C.  LDC >= L.

  U       (input/output) DOUBLE PRECISION array, dimension (LDU,M)
          On entry, the leading NSMP-by-M part of this array must
          contain the t-by-m input-data sequence matrix  U,
          U = [u_1 u_2 ... u_m].  Column  j  of  U  contains the
          NSMP  values of the j-th input component for consecutive
          time increments.
          On exit, if  JOB = 'D',  the leading NSMP-by-M part of
          this array contains details of the QR factorization of
          the t-by-m matrix  U, possibly computed sequentially
          (see METHOD).
          If  JOB = 'B',  this array is unchanged on exit.
          If M = 0, this array is not referenced.

  LDU     INTEGER
          The leading dimension of the array U.
          LDU >= MAX(1,NSMP),  if M > 0;
          LDU >= 1,            if M = 0.

  Y       (input) DOUBLE PRECISION array, dimension (LDY,L)
          The leading NSMP-by-L part of this array must contain the
          t-by-l output-data sequence matrix  Y,
          Y = [y_1 y_2 ... y_l].  Column  j  of  Y  contains the
          NSMP  values of the j-th output component for consecutive
          time increments.

  LDY     INTEGER
          The leading dimension of the array Y.  LDY >= MAX(1,NSMP).

  X0      (output) DOUBLE PRECISION array, dimension (N)
          If  JOBX0 = 'X',  the estimated initial state of the
          system,  x(0).
          If  JOBX0 = 'N',  x(0)  is set to zero without any
          calculations.

  B       (output) DOUBLE PRECISION array, dimension (LDB,M)
          If  N > 0,  M > 0,  and  INFO = 0,  the leading N-by-M
          part of this array contains the system input matrix  B
          in the coordinates corresponding to the real Schur form
          of  A.
          If  N = 0  or  M = 0,  this array is not referenced.

  LDB     INTEGER
          The leading dimension of the array B.
          LDB >= N,  if  N > 0  and  M > 0;
          LDB >= 1,  if  N = 0  or   M = 0.

  D       (output) DOUBLE PRECISION array, dimension (LDD,M)
          If  M > 0,  JOB = 'D',  and  INFO = 0,  the leading
          L-by-M part of this array contains the system input-output
          matrix  D.
          If  M = 0  or  JOB = 'B',  this array is not referenced.

  LDD     INTEGER
          The leading dimension of the array D.
          LDD >= L,  if  M > 0  and  JOB = 'D';
          LDD >= 1,  if  M = 0  or   JOB = 'B'.

Tolerances
  TOL     DOUBLE PRECISION
          The tolerance to be used for estimating the rank of
          matrices. If the user sets  TOL > 0,  then the given value
          of  TOL  is used as a lower bound for the reciprocal
          condition number;  a matrix whose estimated condition
          number is less than  1/TOL  is considered to be of full
          rank.  If the user sets  TOL <= 0,  then  EPS  is used
          instead, where  EPS  is the relative machine precision
          (see LAPACK Library routine DLAMCH).  TOL <= 1.

Workspace
  IWORK   INTEGER array, dimension (LIWORK), where
          LIWORK >= N*M + a,            if  JOB = 'B',
          LIWORK >= max( N*M + a, M ),  if  JOB = 'D',
          with  a = 0,  if  JOBX0 = 'N';
                a = N,  if  JOBX0 = 'X'.

  DWORK   DOUBLE PRECISION array, dimension (LDWORK)
          On exit, if  INFO = 0,  DWORK(1) returns the optimal value
          of LDWORK;  DWORK(2)  contains the reciprocal condition
          number of the triangular factor of the QR factorization of
          the matrix  W2  (see METHOD); if  M > 0  and  JOB = 'D',
          DWORK(3)  contains the reciprocal condition number of the
          triangular factor of the QR factorization of  U.
          On exit, if  INFO = -23,  DWORK(1)  returns the minimum
          value of LDWORK.

  LDWORK  INTEGER
          The length of the array DWORK.
          LDWORK >= max( LDW1, min( LDW2, LDW3 ) ),  where
          LDW1 = 2,          if  M = 0  or   JOB = 'B',
          LDW1 = 3,          if  M > 0  and  JOB = 'D',
          LDWa = t*L*(r + 1) + max( N + max( d, f ), 6*r ),
          LDW2 = LDWa,       if  M = 0  or  JOB = 'B',
          LDW2 = max( LDWa, t*L*(r + 1) + 2*M*M + 6*M ),
                             if  M > 0  and JOB = 'D',
          LDWb = (b + r)*(r + 1) +
                  max( q*(r + 1) + N*N*M + c + max( d, f ), 6*r ),
          LDW3 = LDWb,       if  M = 0  or  JOB = 'B',
          LDW3 = max( LDWb, (b + r)*(r + 1) + 2*M*M + 6*M ),
                             if  M > 0  and JOB = 'D',
             r = N*M + a,
             a = 0,                  if  JOBX0 = 'N',
             a = N,                  if  JOBX0 = 'X';
             b = 0,                  if  JOB   = 'B',
             b = L*M,                if  JOB   = 'D';
             c = 0,                  if  JOBX0 = 'N',
             c = L*N,                if  JOBX0 = 'X';
             d = 0,                  if  JOBX0 = 'N',
             d = 2*N*N + N,          if  JOBX0 = 'X';
             f = 2*r,                if  JOB   = 'B'   or  M = 0,
             f = M + max( 2*r, M ),  if  JOB   = 'D'  and  M > 0;
             q = b + r*L.
          For good performance,  LDWORK  should be larger.
          If  LDWORK >= LDW2  or
              LDWORK >= t*L*(r + 1) + (b + r)*(r + 1) + N*N*M + c +
                        max( d, f ),
          then standard QR factorizations of the matrices  U  and/or
          W2  (see METHOD) are used.
          Otherwise, the QR factorizations are computed sequentially
          by performing  NCYCLE  cycles, each cycle (except possibly
          the last one) processing  s < t  samples, where  s  is
          chosen from the equation
            LDWORK = s*L*(r + 1) + (b + r)*(r + 1) + N*N*M + c +
                     max( d, f ).
          (s  is at least  N*M+a+e,  the minimum value of  NSMP.)
          The computational effort may increase and the accuracy may
          decrease with the decrease of  s.  Recommended value is
          LDWORK = LDW2,  assuming a large enough cache size, to
          also accommodate  A,  C,  U,  and  Y.

Warning Indicator
  IWARN   INTEGER
          = 0:  no warning;
          = 4:  the least squares problem to be solved has a
                rank-deficient coefficient matrix.

Error Indicator
  INFO    INTEGER
          = 0:  successful exit;
          < 0:  if INFO = -i, the i-th argument had an illegal
                value;
          = 2:  the singular value decomposition (SVD) algorithm did
                not converge.

Method
  An extension and refinement of the method in [1,2] is used.
  Specifically, denoting

        X = [ vec(D')' vec(B)' x0' ]',

  where  vec(M)  is the vector obtained by stacking the columns of
  the matrix  M,  then  X  is the least squares solution of the
  system  S*X = vec(Y),  with the matrix  S = [ diag(U)  W ],
  defined by

        ( U         |     | ... |     |     | ... |     |         )
        (   U       |  11 | ... |  n1 |  12 | ... |  nm |         )
    S = (     :     | y   | ... | y   | y   | ... | y   | P*Gamma ),
        (       :   |     | ... |     |     | ... |     |         )
        (         U |     | ... |     |     | ... |     |         )
                                                                  ij
  diag(U)  having  L  block rows and columns.  In this formula,  y
  are the outputs of the system for zero initial state computed
  using the following model, for j = 1:m, and for i = 1:n,
         ij          ij                    ij
        x  (k+1) = Ax  (k) + e_i u_j(k),  x  (0) = 0,

         ij          ij
        y  (k)   = Cx  (k),

  where  e_i  is the i-th n-dimensional unit vector,  Gamma  is
  given by

             (     C     )
             (    C*A    )
     Gamma = (   C*A^2   ),
             (     :     )
             ( C*A^(t-1) )

  and  P  is a permutation matrix that groups together the rows of
  Gamma  depending on the same row of  C,  namely
  [ c_j;  c_j*A;  c_j*A^2; ...  c_j*A^(t-1) ],  for j = 1:L.
  The first block column,  diag(U),  is not explicitly constructed,
  but its structure is exploited. The last block column is evaluated
  using powers of A with exponents 2^k. No interchanges are applied.
  A special QR decomposition of the matrix  S  is computed. Let
  U = q*[ r' 0 ]'  be the QR decomposition of  U,  if  M > 0,  where
  r  is  M-by-M.   Then,  diag(q')  is applied to  W  and  vec(Y).
  The block-rows of  S  and  vec(Y)  are implicitly permuted so that
  matrix  S  becomes

     ( diag(r)  W1 )
     (    0     W2 ),

  where  W1  has L*M rows. Then, the QR decomposition of  W2 is
  computed (sequentially, if  M > 0) and used to obtain  B  and  x0.
  The intermediate results and the QR decomposition of  U  are
  needed to find  D.  If a triangular factor is too ill conditioned,
  then singular value decomposition (SVD) is employed. SVD is not
  generally needed if the input sequence is sufficiently
  persistently exciting and  NSMP  is large enough.
  If the matrix  W  cannot be stored in the workspace (i.e.,
  LDWORK < LDW2),  the QR decompositions of  W2  and  U  are
  computed sequentially.

References
  [1] Verhaegen M., and Varga, A.
      Some Experience with the MOESP Class of Subspace Model
      Identification Methods in Identifying the BO105 Helicopter.
      Report TR R165-94, DLR Oberpfaffenhofen, 1994.

  [2] Sima, V., and Varga, A.
      RASP-IDENT : Subspace Model Identification Programs.
      Deutsche Forschungsanstalt fur Luft- und Raumfahrt e. V.,
      Report TR R888-94, DLR Oberpfaffenhofen, Oct. 1994.

Numerical Aspects
  The implemented method is numerically stable.

Further Comments
  The algorithm for computing the system matrices  B  and  D  is
  less efficient than the MOESP or N4SID algorithms implemented in
  SLICOT Library routine IB01PD, because a large least squares
  problem has to be solved, but the accuracy is better, as the
  computed matrices  B  and  D  are fitted to the input and output
  trajectories. However, if matrix  A  is unstable, the computed
  matrices  B  and  D  could be inaccurate.

Example

Program Text

  None
Program Data
  None
Program Results
  None

Return to Supporting Routines index