3 * =========== DOCUMENTATION ===========
5 * Online html documentation available at
6 * http://www.netlib.org/lapack/explore-html/
9 *> Download CTGSEN + dependencies
10 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/ctgsen.f">
12 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/ctgsen.f">
14 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/ctgsen.f">
21 * SUBROUTINE CTGSEN( IJOB, WANTQ, WANTZ, SELECT, N, A, LDA, B, LDB,
22 * ALPHA, BETA, Q, LDQ, Z, LDZ, M, PL, PR, DIF,
23 * WORK, LWORK, IWORK, LIWORK, INFO )
25 * .. Scalar Arguments ..
26 * LOGICAL WANTQ, WANTZ
27 * INTEGER IJOB, INFO, LDA, LDB, LDQ, LDZ, LIWORK, LWORK,
31 * .. Array Arguments ..
35 * COMPLEX A( LDA, * ), ALPHA( * ), B( LDB, * ),
36 * $ BETA( * ), Q( LDQ, * ), WORK( * ), Z( LDZ, * )
45 *> CTGSEN reorders the generalized Schur decomposition of a complex
46 *> matrix pair (A, B) (in terms of an unitary equivalence trans-
47 *> formation Q**H * (A, B) * Z), so that a selected cluster of eigenvalues
48 *> appears in the leading diagonal blocks of the pair (A,B). The leading
49 *> columns of Q and Z form unitary bases of the corresponding left and
50 *> right eigenspaces (deflating subspaces). (A, B) must be in
51 *> generalized Schur canonical form, that is, A and B are both upper
54 *> CTGSEN also computes the generalized eigenvalues
56 *> w(j)= ALPHA(j) / BETA(j)
58 *> of the reordered matrix pair (A, B).
60 *> Optionally, the routine computes estimates of reciprocal condition
61 *> numbers for eigenvalues and eigenspaces. These are Difu[(A11,B11),
62 *> (A22,B22)] and Difl[(A11,B11), (A22,B22)], i.e. the separation(s)
63 *> between the matrix pairs (A11, B11) and (A22,B22) that correspond to
64 *> the selected cluster and the eigenvalues outside the cluster, resp.,
65 *> and norms of "projections" onto left and right eigenspaces w.r.t.
66 *> the selected cluster in the (1,1)-block.
76 *> Specifies whether condition numbers are required for the
77 *> cluster of eigenvalues (PL and PR) or the deflating subspaces
79 *> =0: Only reorder w.r.t. SELECT. No extras.
80 *> =1: Reciprocal of norms of "projections" onto left and right
81 *> eigenspaces w.r.t. the selected cluster (PL and PR).
82 *> =2: Upper bounds on Difu and Difl. F-norm-based estimate
84 *> =3: Estimate of Difu and Difl. 1-norm-based estimate
86 *> About 5 times as expensive as IJOB = 2.
87 *> =4: Compute PL, PR and DIF (i.e. 0, 1 and 2 above): Economic
88 *> version to get it all.
89 *> =5: Compute PL, PR and DIF (i.e. 0, 1 and 3 above)
95 *> .TRUE. : update the left transformation matrix Q;
96 *> .FALSE.: do not update Q.
102 *> .TRUE. : update the right transformation matrix Z;
103 *> .FALSE.: do not update Z.
108 *> SELECT is LOGICAL array, dimension (N)
109 *> SELECT specifies the eigenvalues in the selected cluster. To
110 *> select an eigenvalue w(j), SELECT(j) must be set to
117 *> The order of the matrices A and B. N >= 0.
122 *> A is COMPLEX array, dimension(LDA,N)
123 *> On entry, the upper triangular matrix A, in generalized
124 *> Schur canonical form.
125 *> On exit, A is overwritten by the reordered matrix A.
131 *> The leading dimension of the array A. LDA >= max(1,N).
136 *> B is COMPLEX array, dimension(LDB,N)
137 *> On entry, the upper triangular matrix B, in generalized
138 *> Schur canonical form.
139 *> On exit, B is overwritten by the reordered matrix B.
145 *> The leading dimension of the array B. LDB >= max(1,N).
150 *> ALPHA is COMPLEX array, dimension (N)
155 *> BETA is COMPLEX array, dimension (N)
157 *> The diagonal elements of A and B, respectively,
158 *> when the pair (A,B) has been reduced to generalized Schur
159 *> form. ALPHA(i)/BETA(i) i=1,...,N are the generalized
165 *> Q is COMPLEX array, dimension (LDQ,N)
166 *> On entry, if WANTQ = .TRUE., Q is an N-by-N matrix.
167 *> On exit, Q has been postmultiplied by the left unitary
168 *> transformation matrix which reorder (A, B); The leading M
169 *> columns of Q form orthonormal bases for the specified pair of
170 *> left eigenspaces (deflating subspaces).
171 *> If WANTQ = .FALSE., Q is not referenced.
177 *> The leading dimension of the array Q. LDQ >= 1.
178 *> If WANTQ = .TRUE., LDQ >= N.
183 *> Z is COMPLEX array, dimension (LDZ,N)
184 *> On entry, if WANTZ = .TRUE., Z is an N-by-N matrix.
185 *> On exit, Z has been postmultiplied by the left unitary
186 *> transformation matrix which reorder (A, B); The leading M
187 *> columns of Z form orthonormal bases for the specified pair of
188 *> left eigenspaces (deflating subspaces).
189 *> If WANTZ = .FALSE., Z is not referenced.
195 *> The leading dimension of the array Z. LDZ >= 1.
196 *> If WANTZ = .TRUE., LDZ >= N.
202 *> The dimension of the specified pair of left and right
203 *> eigenspaces, (deflating subspaces) 0 <= M <= N.
215 *> If IJOB = 1, 4 or 5, PL, PR are lower bounds on the
216 *> reciprocal of the norm of "projections" onto left and right
217 *> eigenspace with respect to the selected cluster.
219 *> If M = 0 or M = N, PL = PR = 1.
220 *> If IJOB = 0, 2 or 3 PL, PR are not referenced.
225 *> DIF is REAL array, dimension (2).
226 *> If IJOB >= 2, DIF(1:2) store the estimates of Difu and Difl.
227 *> If IJOB = 2 or 4, DIF(1:2) are F-norm-based upper bounds on
228 *> Difu and Difl. If IJOB = 3 or 5, DIF(1:2) are 1-norm-based
229 *> estimates of Difu and Difl, computed using reversed
230 *> communication with CLACN2.
231 *> If M = 0 or N, DIF(1:2) = F-norm([A, B]).
232 *> If IJOB = 0 or 1, DIF is not referenced.
237 *> WORK is COMPLEX array, dimension (MAX(1,LWORK))
238 *> On exit, if INFO = 0, WORK(1) returns the optimal LWORK.
244 *> The dimension of the array WORK. LWORK >= 1
245 *> If IJOB = 1, 2 or 4, LWORK >= 2*M*(N-M)
246 *> If IJOB = 3 or 5, LWORK >= 4*M*(N-M)
248 *> If LWORK = -1, then a workspace query is assumed; the routine
249 *> only calculates the optimal size of the WORK array, returns
250 *> this value as the first entry of the WORK array, and no error
251 *> message related to LWORK is issued by XERBLA.
256 *> IWORK is INTEGER array, dimension (MAX(1,LIWORK))
257 *> On exit, if INFO = 0, IWORK(1) returns the optimal LIWORK.
263 *> The dimension of the array IWORK. LIWORK >= 1.
264 *> If IJOB = 1, 2 or 4, LIWORK >= N+2;
265 *> If IJOB = 3 or 5, LIWORK >= MAX(N+2, 2*M*(N-M));
267 *> If LIWORK = -1, then a workspace query is assumed; the
268 *> routine only calculates the optimal size of the IWORK array,
269 *> returns this value as the first entry of the IWORK array, and
270 *> no error message related to LIWORK is issued by XERBLA.
276 *> =0: Successful exit.
277 *> <0: If INFO = -i, the i-th argument had an illegal value.
278 *> =1: Reordering of (A, B) failed because the transformed
279 *> matrix pair (A, B) would be too far from generalized
280 *> Schur form; the problem is very ill-conditioned.
281 *> (A, B) may have been partially reordered.
282 *> If requested, 0 is returned in DIF(*), PL and PR.
288 *> \author Univ. of Tennessee
289 *> \author Univ. of California Berkeley
290 *> \author Univ. of Colorado Denver
295 *> \ingroup complexOTHERcomputational
297 *> \par Further Details:
298 * =====================
302 *> CTGSEN first collects the selected eigenvalues by computing unitary
303 *> U and W that move them to the top left corner of (A, B). In other
304 *> words, the selected eigenvalues are the eigenvalues of (A11, B11) in
306 *> U**H*(A, B)*W = (A11 A12) (B11 B12) n1
307 *> ( 0 A22),( 0 B22) n2
310 *> where N = n1+n2 and U**H means the conjugate transpose of U. The first
311 *> n1 columns of U and W span the specified pair of left and right
312 *> eigenspaces (deflating subspaces) of (A, B).
314 *> If (A, B) has been obtained from the generalized real Schur
315 *> decomposition of a matrix pair (C, D) = Q*(A, B)*Z', then the
316 *> reordered generalized Schur form of (C, D) is given by
318 *> (C, D) = (Q*U)*(U**H *(A, B)*W)*(Z*W)**H,
320 *> and the first n1 columns of Q*U and Z*W span the corresponding
321 *> deflating subspaces of (C, D) (Q and Z store Q*U and Z*W, resp.).
323 *> Note that if the selected eigenvalue is sufficiently ill-conditioned,
324 *> then its value may differ significantly from its value before
327 *> The reciprocal condition numbers of the left and right eigenspaces
328 *> spanned by the first n1 columns of U and W (or Q*U and Z*W) may
329 *> be returned in DIF(1:2), corresponding to Difu and Difl, resp.
331 *> The Difu and Difl are defined as:
333 *> Difu[(A11, B11), (A22, B22)] = sigma-min( Zu )
335 *> Difl[(A11, B11), (A22, B22)] = Difu[(A22, B22), (A11, B11)],
337 *> where sigma-min(Zu) is the smallest singular value of the
338 *> (2*n1*n2)-by-(2*n1*n2) matrix
340 *> Zu = [ kron(In2, A11) -kron(A22**H, In1) ]
341 *> [ kron(In2, B11) -kron(B22**H, In1) ].
343 *> Here, Inx is the identity matrix of size nx and A22**H is the
344 *> conjuguate transpose of A22. kron(X, Y) is the Kronecker product between
345 *> the matrices X and Y.
347 *> When DIF(2) is small, small changes in (A, B) can cause large changes
348 *> in the deflating subspace. An approximate (asymptotic) bound on the
349 *> maximum angular error in the computed deflating subspaces is
351 *> EPS * norm((A, B)) / DIF(2),
353 *> where EPS is the machine precision.
355 *> The reciprocal norm of the projectors on the left and right
356 *> eigenspaces associated with (A11, B11) may be returned in PL and PR.
357 *> They are computed as follows. First we compute L and R so that
358 *> P*(A, B)*Q is block diagonal, where
360 *> P = ( I -L ) n1 Q = ( I R ) n1
361 *> ( 0 I ) n2 and ( 0 I ) n2
364 *> and (L, R) is the solution to the generalized Sylvester equation
366 *> A11*R - L*A22 = -A12
367 *> B11*R - L*B22 = -B12
369 *> Then PL = (F-norm(L)**2+1)**(-1/2) and PR = (F-norm(R)**2+1)**(-1/2).
370 *> An approximate (asymptotic) bound on the average absolute error of
371 *> the selected eigenvalues is
373 *> EPS * norm((A, B)) / PL.
375 *> There are also global error bounds which valid for perturbations up
376 *> to a certain restriction: A lower bound (x) on the smallest
377 *> F-norm(E,F) for which an eigenvalue of (A11, B11) may move and
378 *> coalesce with an eigenvalue of (A22, B22) under perturbation (E,F),
379 *> (i.e. (A + E, B + F), is
381 *> x = min(Difu,Difl)/((1/(PL*PL)+1/(PR*PR))**(1/2)+2*max(1/PL,1/PR)).
383 *> An approximate bound on x can be computed from DIF(1:2), PL and PR.
385 *> If y = ( F-norm(E,F) / x) <= 1, the angles between the perturbed
386 *> (L', R') and unperturbed (L, R) left and right deflating subspaces
387 *> associated with the selected cluster in the (1,1)-blocks can be
390 *> max-angle(L, L') <= arctan( y * PL / (1 - y * (1 - PL * PL)**(1/2))
391 *> max-angle(R, R') <= arctan( y * PR / (1 - y * (1 - PR * PR)**(1/2))
393 *> See LAPACK User's Guide section 4.11 or the following references
394 *> for more information.
396 *> Note that if the default method for computing the Frobenius-norm-
397 *> based estimate DIF is not wanted (see CLATDF), then the parameter
398 *> IDIFJB (see below) should be changed from 3 to 4 (routine CLATDF
399 *> (IJOB = 2 will be used)). See CTGSYL for more details.
402 *> \par Contributors:
405 *> Bo Kagstrom and Peter Poromaa, Department of Computing Science,
406 *> Umea University, S-901 87 Umea, Sweden.
411 *> [1] B. Kagstrom; A Direct Method for Reordering Eigenvalues in the
412 *> Generalized Real Schur Form of a Regular Matrix Pair (A, B), in
413 *> M.S. Moonen et al (eds), Linear Algebra for Large Scale and
414 *> Real-Time Applications, Kluwer Academic Publ. 1993, pp 195-218.
416 *> [2] B. Kagstrom and P. Poromaa; Computing Eigenspaces with Specified
417 *> Eigenvalues of a Regular Matrix Pair (A, B) and Condition
418 *> Estimation: Theory, Algorithms and Software, Report
419 *> UMINF - 94.04, Department of Computing Science, Umea University,
420 *> S-901 87 Umea, Sweden, 1994. Also as LAPACK Working Note 87.
421 *> To appear in Numerical Algorithms, 1996.
423 *> [3] B. Kagstrom and P. Poromaa, LAPACK-Style Algorithms and Software
424 *> for Solving the Generalized Sylvester Equation and Estimating the
425 *> Separation between Regular Matrix Pairs, Report UMINF - 93.23,
426 *> Department of Computing Science, Umea University, S-901 87 Umea,
427 *> Sweden, December 1993, Revised April 1994, Also as LAPACK working
428 *> Note 75. To appear in ACM Trans. on Math. Software, Vol 22, No 1,
431 * =====================================================================
432 SUBROUTINE CTGSEN( IJOB, WANTQ, WANTZ, SELECT, N, A, LDA, B, LDB,
433 $ ALPHA, BETA, Q, LDQ, Z, LDZ, M, PL, PR, DIF,
434 $ WORK, LWORK, IWORK, LIWORK, INFO )
436 * -- LAPACK computational routine (version 3.6.1) --
437 * -- LAPACK is a software package provided by Univ. of Tennessee, --
438 * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
441 * .. Scalar Arguments ..
443 INTEGER IJOB, INFO, LDA, LDB, LDQ, LDZ, LIWORK, LWORK,
447 * .. Array Arguments ..
451 COMPLEX A( LDA, * ), ALPHA( * ), B( LDB, * ),
452 $ BETA( * ), Q( LDQ, * ), WORK( * ), Z( LDZ, * )
455 * =====================================================================
459 PARAMETER ( IDIFJB = 3 )
461 PARAMETER ( ZERO = 0.0E+0, ONE = 1.0E+0 )
463 * .. Local Scalars ..
464 LOGICAL LQUERY, SWAP, WANTD, WANTD1, WANTD2, WANTP
465 INTEGER I, IERR, IJB, K, KASE, KS, LIWMIN, LWMIN, MN2,
467 REAL DSCALE, DSUM, RDSCAL, SAFMIN
473 * .. External Subroutines ..
475 EXTERNAL CLACN2, CLACPY, CLASSQ, CSCAL, CTGEXC, CTGSYL,
478 * .. Intrinsic Functions ..
479 INTRINSIC ABS, CMPLX, CONJG, MAX, SQRT
481 * .. Executable Statements ..
483 * Decode and test the input parameters
486 LQUERY = ( LWORK.EQ.-1 .OR. LIWORK.EQ.-1 )
488 IF( IJOB.LT.0 .OR. IJOB.GT.5 ) THEN
490 ELSE IF( N.LT.0 ) THEN
492 ELSE IF( LDA.LT.MAX( 1, N ) ) THEN
494 ELSE IF( LDB.LT.MAX( 1, N ) ) THEN
496 ELSE IF( LDQ.LT.1 .OR. ( WANTQ .AND. LDQ.LT.N ) ) THEN
498 ELSE IF( LDZ.LT.1 .OR. ( WANTZ .AND. LDZ.LT.N ) ) THEN
503 CALL XERBLA( 'CTGSEN', -INFO )
509 WANTP = IJOB.EQ.1 .OR. IJOB.GE.4
510 WANTD1 = IJOB.EQ.2 .OR. IJOB.EQ.4
511 WANTD2 = IJOB.EQ.3 .OR. IJOB.EQ.5
512 WANTD = WANTD1 .OR. WANTD2
514 * Set M to the dimension of the specified pair of deflating
518 IF( .NOT.LQUERY .OR. IJOB.NE.0 ) THEN
520 ALPHA( K ) = A( K, K )
521 BETA( K ) = B( K, K )
532 IF( IJOB.EQ.1 .OR. IJOB.EQ.2 .OR. IJOB.EQ.4 ) THEN
533 LWMIN = MAX( 1, 2*M*(N-M) )
534 LIWMIN = MAX( 1, N+2 )
535 ELSE IF( IJOB.EQ.3 .OR. IJOB.EQ.5 ) THEN
536 LWMIN = MAX( 1, 4*M*(N-M) )
537 LIWMIN = MAX( 1, 2*M*(N-M), N+2 )
546 IF( LWORK.LT.LWMIN .AND. .NOT.LQUERY ) THEN
548 ELSE IF( LIWORK.LT.LIWMIN .AND. .NOT.LQUERY ) THEN
553 CALL XERBLA( 'CTGSEN', -INFO )
555 ELSE IF( LQUERY ) THEN
559 * Quick return if possible.
561 IF( M.EQ.N .OR. M.EQ.0 ) THEN
570 CALL CLASSQ( N, A( 1, I ), 1, DSCALE, DSUM )
571 CALL CLASSQ( N, B( 1, I ), 1, DSCALE, DSUM )
573 DIF( 1 ) = DSCALE*SQRT( DSUM )
579 * Get machine constant
581 SAFMIN = SLAMCH( 'S' )
583 * Collect the selected blocks at the top-left corner of (A, B).
591 * Swap the K-th block to position KS. Compute unitary Q
592 * and Z that will swap adjacent diagonal blocks in (A, B).
595 $ CALL CTGEXC( WANTQ, WANTZ, N, A, LDA, B, LDB, Q, LDQ, Z,
600 * Swap is rejected: exit.
617 * Solve generalized Sylvester equation for R and L:
618 * A11 * R - L * A22 = A12
619 * B11 * R - L * B22 = B12
624 CALL CLACPY( 'Full', N1, N2, A( 1, I ), LDA, WORK, N1 )
625 CALL CLACPY( 'Full', N1, N2, B( 1, I ), LDB, WORK( N1*N2+1 ),
628 CALL CTGSYL( 'N', IJB, N1, N2, A, LDA, A( I, I ), LDA, WORK,
629 $ N1, B, LDB, B( I, I ), LDB, WORK( N1*N2+1 ), N1,
630 $ DSCALE, DIF( 1 ), WORK( N1*N2*2+1 ),
631 $ LWORK-2*N1*N2, IWORK, IERR )
633 * Estimate the reciprocal of norms of "projections" onto
634 * left and right eigenspaces
638 CALL CLASSQ( N1*N2, WORK, 1, RDSCAL, DSUM )
639 PL = RDSCAL*SQRT( DSUM )
640 IF( PL.EQ.ZERO ) THEN
643 PL = DSCALE / ( SQRT( DSCALE*DSCALE / PL+PL )*SQRT( PL ) )
647 CALL CLASSQ( N1*N2, WORK( N1*N2+1 ), 1, RDSCAL, DSUM )
648 PR = RDSCAL*SQRT( DSUM )
649 IF( PR.EQ.ZERO ) THEN
652 PR = DSCALE / ( SQRT( DSCALE*DSCALE / PR+PR )*SQRT( PR ) )
657 * Compute estimates Difu and Difl.
665 * Frobenius norm-based Difu estimate.
667 CALL CTGSYL( 'N', IJB, N1, N2, A, LDA, A( I, I ), LDA, WORK,
668 $ N1, B, LDB, B( I, I ), LDB, WORK( N1*N2+1 ),
669 $ N1, DSCALE, DIF( 1 ), WORK( N1*N2*2+1 ),
670 $ LWORK-2*N1*N2, IWORK, IERR )
672 * Frobenius norm-based Difl estimate.
674 CALL CTGSYL( 'N', IJB, N2, N1, A( I, I ), LDA, A, LDA, WORK,
675 $ N2, B( I, I ), LDB, B, LDB, WORK( N1*N2+1 ),
676 $ N2, DSCALE, DIF( 2 ), WORK( N1*N2*2+1 ),
677 $ LWORK-2*N1*N2, IWORK, IERR )
680 * Compute 1-norm-based estimates of Difu and Difl using
681 * reversed communication with CLACN2. In each step a
682 * generalized Sylvester equation or a transposed variant
692 * 1-norm-based estimate of Difu.
695 CALL CLACN2( MN2, WORK( MN2+1 ), WORK, DIF( 1 ), KASE,
700 * Solve generalized Sylvester equation
702 CALL CTGSYL( 'N', IJB, N1, N2, A, LDA, A( I, I ), LDA,
703 $ WORK, N1, B, LDB, B( I, I ), LDB,
704 $ WORK( N1*N2+1 ), N1, DSCALE, DIF( 1 ),
705 $ WORK( N1*N2*2+1 ), LWORK-2*N1*N2, IWORK,
709 * Solve the transposed variant.
711 CALL CTGSYL( 'C', IJB, N1, N2, A, LDA, A( I, I ), LDA,
712 $ WORK, N1, B, LDB, B( I, I ), LDB,
713 $ WORK( N1*N2+1 ), N1, DSCALE, DIF( 1 ),
714 $ WORK( N1*N2*2+1 ), LWORK-2*N1*N2, IWORK,
719 DIF( 1 ) = DSCALE / DIF( 1 )
721 * 1-norm-based estimate of Difl.
724 CALL CLACN2( MN2, WORK( MN2+1 ), WORK, DIF( 2 ), KASE,
729 * Solve generalized Sylvester equation
731 CALL CTGSYL( 'N', IJB, N2, N1, A( I, I ), LDA, A, LDA,
732 $ WORK, N2, B( I, I ), LDB, B, LDB,
733 $ WORK( N1*N2+1 ), N2, DSCALE, DIF( 2 ),
734 $ WORK( N1*N2*2+1 ), LWORK-2*N1*N2, IWORK,
738 * Solve the transposed variant.
740 CALL CTGSYL( 'C', IJB, N2, N1, A( I, I ), LDA, A, LDA,
741 $ WORK, N2, B, LDB, B( I, I ), LDB,
742 $ WORK( N1*N2+1 ), N2, DSCALE, DIF( 2 ),
743 $ WORK( N1*N2*2+1 ), LWORK-2*N1*N2, IWORK,
748 DIF( 2 ) = DSCALE / DIF( 2 )
752 * If B(K,K) is complex, make it real and positive (normalization
753 * of the generalized Schur form) and Store the generalized
754 * eigenvalues of reordered pair (A, B)
757 DSCALE = ABS( B( K, K ) )
758 IF( DSCALE.GT.SAFMIN ) THEN
759 TEMP1 = CONJG( B( K, K ) / DSCALE )
760 TEMP2 = B( K, K ) / DSCALE
762 CALL CSCAL( N-K, TEMP1, B( K, K+1 ), LDB )
763 CALL CSCAL( N-K+1, TEMP1, A( K, K ), LDA )
765 $ CALL CSCAL( N, TEMP2, Q( 1, K ), 1 )
767 B( K, K ) = CMPLX( ZERO, ZERO )
770 ALPHA( K ) = A( K, K )
771 BETA( K ) = B( K, K )