1 *> \brief \b ZLAQR4 computes the eigenvalues of a Hessenberg matrix, and optionally the matrices from the Schur decomposition.
3 * =========== DOCUMENTATION ===========
5 * Online html documentation available at
6 * http://www.netlib.org/lapack/explore-html/
9 *> Download ZLAQR4 + dependencies
10 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/zlaqr4.f">
12 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/zlaqr4.f">
14 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/zlaqr4.f">
21 * SUBROUTINE ZLAQR4( WANTT, WANTZ, N, ILO, IHI, H, LDH, W, ILOZ,
22 * IHIZ, Z, LDZ, WORK, LWORK, INFO )
24 * .. Scalar Arguments ..
25 * INTEGER IHI, IHIZ, ILO, ILOZ, INFO, LDH, LDZ, LWORK, N
26 * LOGICAL WANTT, WANTZ
28 * .. Array Arguments ..
29 * COMPLEX*16 H( LDH, * ), W( * ), WORK( * ), Z( LDZ, * )
38 *> ZLAQR4 implements one level of recursion for ZLAQR0.
39 *> It is a complete implementation of the small bulge multi-shift
40 *> QR algorithm. It may be called by ZLAQR0 and, for large enough
41 *> deflation window size, it may be called by ZLAQR3. This
42 *> subroutine is identical to ZLAQR0 except that it calls ZLAQR2
45 *> ZLAQR4 computes the eigenvalues of a Hessenberg matrix H
46 *> and, optionally, the matrices T and Z from the Schur decomposition
47 *> H = Z T Z**H, where T is an upper triangular matrix (the
48 *> Schur form), and Z is the unitary matrix of Schur vectors.
50 *> Optionally Z may be postmultiplied into an input unitary
51 *> matrix Q so that this routine can give the Schur factorization
52 *> of a matrix A which has been reduced to the Hessenberg form H
53 *> by the unitary matrix Q: A = Q*H*Q**H = (QZ)*H*(QZ)**H.
62 *> = .TRUE. : the full Schur form T is required;
63 *> = .FALSE.: only eigenvalues are required.
69 *> = .TRUE. : the matrix of Schur vectors Z is required;
70 *> = .FALSE.: Schur vectors are not required.
76 *> The order of the matrix H. N .GE. 0.
87 *> It is assumed that H is already upper triangular in rows
88 *> and columns 1:ILO-1 and IHI+1:N and, if ILO.GT.1,
89 *> H(ILO,ILO-1) is zero. ILO and IHI are normally set by a
90 *> previous call to ZGEBAL, and then passed to ZGEHRD when the
91 *> matrix output by ZGEBAL is reduced to Hessenberg form.
92 *> Otherwise, ILO and IHI should be set to 1 and N,
93 *> respectively. If N.GT.0, then 1.LE.ILO.LE.IHI.LE.N.
94 *> If N = 0, then ILO = 1 and IHI = 0.
99 *> H is COMPLEX*16 array, dimension (LDH,N)
100 *> On entry, the upper Hessenberg matrix H.
101 *> On exit, if INFO = 0 and WANTT is .TRUE., then H
102 *> contains the upper triangular matrix T from the Schur
103 *> decomposition (the Schur form). If INFO = 0 and WANT is
104 *> .FALSE., then the contents of H are unspecified on exit.
105 *> (The output value of H when INFO.GT.0 is given under the
106 *> description of INFO below.)
108 *> This subroutine may explicitly set H(i,j) = 0 for i.GT.j and
109 *> j = 1, 2, ... ILO-1 or j = IHI+1, IHI+2, ... N.
115 *> The leading dimension of the array H. LDH .GE. max(1,N).
120 *> W is COMPLEX*16 array, dimension (N)
121 *> The computed eigenvalues of H(ILO:IHI,ILO:IHI) are stored
122 *> in W(ILO:IHI). If WANTT is .TRUE., then the eigenvalues are
123 *> stored in the same order as on the diagonal of the Schur
124 *> form returned in H, with W(i) = H(i,i).
135 *> Specify the rows of Z to which transformations must be
136 *> applied if WANTZ is .TRUE..
137 *> 1 .LE. ILOZ .LE. ILO; IHI .LE. IHIZ .LE. N.
142 *> Z is COMPLEX*16 array, dimension (LDZ,IHI)
143 *> If WANTZ is .FALSE., then Z is not referenced.
144 *> If WANTZ is .TRUE., then Z(ILO:IHI,ILOZ:IHIZ) is
145 *> replaced by Z(ILO:IHI,ILOZ:IHIZ)*U where U is the
146 *> orthogonal Schur factor of H(ILO:IHI,ILO:IHI).
147 *> (The output value of Z when INFO.GT.0 is given under
148 *> the description of INFO below.)
154 *> The leading dimension of the array Z. if WANTZ is .TRUE.
155 *> then LDZ.GE.MAX(1,IHIZ). Otherwize, LDZ.GE.1.
160 *> WORK is COMPLEX*16 array, dimension LWORK
161 *> On exit, if LWORK = -1, WORK(1) returns an estimate of
162 *> the optimal value for LWORK.
168 *> The dimension of the array WORK. LWORK .GE. max(1,N)
169 *> is sufficient, but LWORK typically as large as 6*N may
170 *> be required for optimal performance. A workspace query
171 *> to determine the optimal workspace size is recommended.
173 *> If LWORK = -1, then ZLAQR4 does a workspace query.
174 *> In this case, ZLAQR4 checks the input parameters and
175 *> estimates the optimal workspace size for the given
176 *> values of N, ILO and IHI. The estimate is returned
177 *> in WORK(1). No error message related to LWORK is
178 *> issued by XERBLA. Neither H nor Z are accessed.
184 *> = 0: successful exit
185 *> .GT. 0: if INFO = i, ZLAQR4 failed to compute all of
186 *> the eigenvalues. Elements 1:ilo-1 and i+1:n of WR
187 *> and WI contain those eigenvalues which have been
188 *> successfully computed. (Failures are rare.)
190 *> If INFO .GT. 0 and WANT is .FALSE., then on exit,
191 *> the remaining unconverged eigenvalues are the eigen-
192 *> values of the upper Hessenberg matrix rows and
193 *> columns ILO through INFO of the final, output
196 *> If INFO .GT. 0 and WANTT is .TRUE., then on exit
198 *> (*) (initial value of H)*U = U*(final value of H)
200 *> where U is a unitary matrix. The final
201 *> value of H is upper Hessenberg and triangular in
202 *> rows and columns INFO+1 through IHI.
204 *> If INFO .GT. 0 and WANTZ is .TRUE., then on exit
206 *> (final value of Z(ILO:IHI,ILOZ:IHIZ)
207 *> = (initial value of Z(ILO:IHI,ILOZ:IHIZ)*U
209 *> where U is the unitary matrix in (*) (regard-
210 *> less of the value of WANTT.)
212 *> If INFO .GT. 0 and WANTZ is .FALSE., then Z is not
219 *> \author Univ. of Tennessee
220 *> \author Univ. of California Berkeley
221 *> \author Univ. of Colorado Denver
224 *> \date September 2012
226 *> \ingroup complex16OTHERauxiliary
228 *> \par Contributors:
231 *> Karen Braman and Ralph Byers, Department of Mathematics,
232 *> University of Kansas, USA
237 *> K. Braman, R. Byers and R. Mathias, The Multi-Shift QR
238 *> Algorithm Part I: Maintaining Well Focused Shifts, and Level 3
239 *> Performance, SIAM Journal of Matrix Analysis, volume 23, pages
242 *> K. Braman, R. Byers and R. Mathias, The Multi-Shift QR
243 *> Algorithm Part II: Aggressive Early Deflation, SIAM Journal
244 *> of Matrix Analysis, volume 23, pages 948--973, 2002.
246 * =====================================================================
247 SUBROUTINE ZLAQR4( WANTT, WANTZ, N, ILO, IHI, H, LDH, W, ILOZ,
248 $ IHIZ, Z, LDZ, WORK, LWORK, INFO )
250 * -- LAPACK auxiliary routine (version 3.4.2) --
251 * -- LAPACK is a software package provided by Univ. of Tennessee, --
252 * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
255 * .. Scalar Arguments ..
256 INTEGER IHI, IHIZ, ILO, ILOZ, INFO, LDH, LDZ, LWORK, N
259 * .. Array Arguments ..
260 COMPLEX*16 H( LDH, * ), W( * ), WORK( * ), Z( LDZ, * )
263 * ================================================================
267 * ==== Matrices of order NTINY or smaller must be processed by
268 * . ZLAHQR because of insufficient subdiagonal scratch space.
269 * . (This is a hard limit.) ====
271 PARAMETER ( NTINY = 11 )
273 * ==== Exceptional deflation windows: try to cure rare
274 * . slow convergence by varying the size of the
275 * . deflation window after KEXNW iterations. ====
277 PARAMETER ( KEXNW = 5 )
279 * ==== Exceptional shifts: try to cure rare slow convergence
280 * . with ad-hoc exceptional shifts every KEXSH iterations.
283 PARAMETER ( KEXSH = 6 )
285 * ==== The constant WILK1 is used to form the exceptional
287 DOUBLE PRECISION WILK1
288 PARAMETER ( WILK1 = 0.75d0 )
290 PARAMETER ( ZERO = ( 0.0d0, 0.0d0 ),
291 $ ONE = ( 1.0d0, 0.0d0 ) )
293 PARAMETER ( TWO = 2.0d0 )
295 * .. Local Scalars ..
296 COMPLEX*16 AA, BB, CC, CDUM, DD, DET, RTDISC, SWAP, TR2
298 INTEGER I, INF, IT, ITMAX, K, KACC22, KBOT, KDU, KS,
299 $ KT, KTOP, KU, KV, KWH, KWTOP, KWV, LD, LS,
300 $ LWKOPT, NDEC, NDFL, NH, NHO, NIBBLE, NMIN, NS,
301 $ NSMAX, NSR, NVE, NW, NWMAX, NWR, NWUPBD
305 * .. External Functions ..
310 COMPLEX*16 ZDUM( 1, 1 )
312 * .. External Subroutines ..
313 EXTERNAL ZLACPY, ZLAHQR, ZLAQR2, ZLAQR5
315 * .. Intrinsic Functions ..
316 INTRINSIC ABS, DBLE, DCMPLX, DIMAG, INT, MAX, MIN, MOD,
319 * .. Statement Functions ..
320 DOUBLE PRECISION CABS1
322 * .. Statement Function definitions ..
323 CABS1( CDUM ) = ABS( DBLE( CDUM ) ) + ABS( DIMAG( CDUM ) )
325 * .. Executable Statements ..
328 * ==== Quick return for N = 0: nothing to do. ====
335 IF( N.LE.NTINY ) THEN
337 * ==== Tiny matrices must use ZLAHQR. ====
341 $ CALL ZLAHQR( WANTT, WANTZ, N, ILO, IHI, H, LDH, W, ILOZ,
342 $ IHIZ, Z, LDZ, INFO )
345 * ==== Use small bulge multi-shift QR with aggressive early
346 * . deflation on larger-than-tiny matrices. ====
348 * ==== Hope for the best. ====
352 * ==== Set up job flags for ILAENV. ====
365 * ==== NWR = recommended deflation window size. At this
366 * . point, N .GT. NTINY = 11, so there is enough
367 * . subdiagonal workspace for NWR.GE.2 as required.
368 * . (In fact, there is enough subdiagonal space for
371 NWR = ILAENV( 13, 'ZLAQR4', JBCMPZ, N, ILO, IHI, LWORK )
373 NWR = MIN( IHI-ILO+1, ( N-1 ) / 3, NWR )
375 * ==== NSR = recommended number of simultaneous shifts.
376 * . At this point N .GT. NTINY = 11, so there is at
377 * . enough subdiagonal workspace for NSR to be even
378 * . and greater than or equal to two as required. ====
380 NSR = ILAENV( 15, 'ZLAQR4', JBCMPZ, N, ILO, IHI, LWORK )
381 NSR = MIN( NSR, ( N+6 ) / 9, IHI-ILO )
382 NSR = MAX( 2, NSR-MOD( NSR, 2 ) )
384 * ==== Estimate optimal workspace ====
386 * ==== Workspace query call to ZLAQR2 ====
388 CALL ZLAQR2( WANTT, WANTZ, N, ILO, IHI, NWR+1, H, LDH, ILOZ,
389 $ IHIZ, Z, LDZ, LS, LD, W, H, LDH, N, H, LDH, N, H,
392 * ==== Optimal workspace = MAX(ZLAQR5, ZLAQR2) ====
394 LWKOPT = MAX( 3*NSR / 2, INT( WORK( 1 ) ) )
396 * ==== Quick return in case of workspace query. ====
398 IF( LWORK.EQ.-1 ) THEN
399 WORK( 1 ) = DCMPLX( LWKOPT, 0 )
403 * ==== ZLAHQR/ZLAQR0 crossover point ====
405 NMIN = ILAENV( 12, 'ZLAQR4', JBCMPZ, N, ILO, IHI, LWORK )
406 NMIN = MAX( NTINY, NMIN )
408 * ==== Nibble crossover point ====
410 NIBBLE = ILAENV( 14, 'ZLAQR4', JBCMPZ, N, ILO, IHI, LWORK )
411 NIBBLE = MAX( 0, NIBBLE )
413 * ==== Accumulate reflections during ttswp? Use block
414 * . 2-by-2 structure during matrix-matrix multiply? ====
416 KACC22 = ILAENV( 16, 'ZLAQR4', JBCMPZ, N, ILO, IHI, LWORK )
417 KACC22 = MAX( 0, KACC22 )
418 KACC22 = MIN( 2, KACC22 )
420 * ==== NWMAX = the largest possible deflation window for
421 * . which there is sufficient workspace. ====
423 NWMAX = MIN( ( N-1 ) / 3, LWORK / 2 )
426 * ==== NSMAX = the Largest number of simultaneous shifts
427 * . for which there is sufficient workspace. ====
429 NSMAX = MIN( ( N+6 ) / 9, 2*LWORK / 3 )
430 NSMAX = NSMAX - MOD( NSMAX, 2 )
432 * ==== NDFL: an iteration count restarted at deflation. ====
436 * ==== ITMAX = iteration limit ====
438 ITMAX = MAX( 30, 2*KEXSH )*MAX( 10, ( IHI-ILO+1 ) )
440 * ==== Last row and column in the active block ====
444 * ==== Main Loop ====
448 * ==== Done when KBOT falls below ILO ====
453 * ==== Locate active block ====
455 DO 10 K = KBOT, ILO + 1, -1
456 IF( H( K, K-1 ).EQ.ZERO )
463 * ==== Select deflation window size:
465 * . If possible and advisable, nibble the entire
466 * . active block. If not, use size MIN(NWR,NWMAX)
467 * . or MIN(NWR+1,NWMAX) depending upon which has
468 * . the smaller corresponding subdiagonal entry
471 * . Exceptional Case:
472 * . If there have been no deflations in KEXNW or
473 * . more iterations, then vary the deflation window
474 * . size. At first, because, larger windows are,
475 * . in general, more powerful than smaller ones,
476 * . rapidly increase the window to the maximum possible.
477 * . Then, gradually reduce the window size. ====
480 NWUPBD = MIN( NH, NWMAX )
481 IF( NDFL.LT.KEXNW ) THEN
482 NW = MIN( NWUPBD, NWR )
484 NW = MIN( NWUPBD, 2*NW )
486 IF( NW.LT.NWMAX ) THEN
487 IF( NW.GE.NH-1 ) THEN
490 KWTOP = KBOT - NW + 1
491 IF( CABS1( H( KWTOP, KWTOP-1 ) ).GT.
492 $ CABS1( H( KWTOP-1, KWTOP-2 ) ) )NW = NW + 1
495 IF( NDFL.LT.KEXNW ) THEN
497 ELSE IF( NDEC.GE.0 .OR. NW.GE.NWUPBD ) THEN
504 * ==== Aggressive early deflation:
505 * . split workspace under the subdiagonal into
506 * . - an nw-by-nw work array V in the lower
507 * . left-hand-corner,
508 * . - an NW-by-at-least-NW-but-more-is-better
509 * . (NW-by-NHO) horizontal work array along
511 * . - an at-least-NW-but-more-is-better (NHV-by-NW)
512 * . vertical work array along the left-hand-edge.
517 NHO = ( N-NW-1 ) - KT + 1
519 NVE = ( N-NW ) - KWV + 1
521 * ==== Aggressive early deflation ====
523 CALL ZLAQR2( WANTT, WANTZ, N, KTOP, KBOT, NW, H, LDH, ILOZ,
524 $ IHIZ, Z, LDZ, LS, LD, W, H( KV, 1 ), LDH, NHO,
525 $ H( KV, KT ), LDH, NVE, H( KWV, 1 ), LDH, WORK,
528 * ==== Adjust KBOT accounting for new deflations. ====
532 * ==== KS points to the shifts. ====
536 * ==== Skip an expensive QR sweep if there is a (partly
537 * . heuristic) reason to expect that many eigenvalues
538 * . will deflate without it. Here, the QR sweep is
539 * . skipped if many eigenvalues have just been deflated
540 * . or if the remaining active block is small.
542 IF( ( LD.EQ.0 ) .OR. ( ( 100*LD.LE.NW*NIBBLE ) .AND. ( KBOT-
543 $ KTOP+1.GT.MIN( NMIN, NWMAX ) ) ) ) THEN
545 * ==== NS = nominal number of simultaneous shifts.
546 * . This may be lowered (slightly) if ZLAQR2
547 * . did not provide that many shifts. ====
549 NS = MIN( NSMAX, NSR, MAX( 2, KBOT-KTOP ) )
550 NS = NS - MOD( NS, 2 )
552 * ==== If there have been no deflations
553 * . in a multiple of KEXSH iterations,
554 * . then try exceptional shifts.
555 * . Otherwise use shifts provided by
556 * . ZLAQR2 above or from the eigenvalues
557 * . of a trailing principal submatrix. ====
559 IF( MOD( NDFL, KEXSH ).EQ.0 ) THEN
561 DO 30 I = KBOT, KS + 1, -2
562 W( I ) = H( I, I ) + WILK1*CABS1( H( I, I-1 ) )
567 * ==== Got NS/2 or fewer shifts? Use ZLAHQR
568 * . on a trailing principal submatrix to
569 * . get more. (Since NS.LE.NSMAX.LE.(N+6)/9,
570 * . there is enough space below the subdiagonal
571 * . to fit an NS-by-NS scratch array.) ====
573 IF( KBOT-KS+1.LE.NS / 2 ) THEN
576 CALL ZLACPY( 'A', NS, NS, H( KS, KS ), LDH,
578 CALL ZLAHQR( .false., .false., NS, 1, NS,
579 $ H( KT, 1 ), LDH, W( KS ), 1, 1, ZDUM,
583 * ==== In case of a rare QR failure use
584 * . eigenvalues of the trailing 2-by-2
585 * . principal submatrix. Scale to avoid
586 * . overflows, underflows and subnormals.
587 * . (The scale factor S can not be zero,
588 * . because H(KBOT,KBOT-1) is nonzero.) ====
590 IF( KS.GE.KBOT ) THEN
591 S = CABS1( H( KBOT-1, KBOT-1 ) ) +
592 $ CABS1( H( KBOT, KBOT-1 ) ) +
593 $ CABS1( H( KBOT-1, KBOT ) ) +
594 $ CABS1( H( KBOT, KBOT ) )
595 AA = H( KBOT-1, KBOT-1 ) / S
596 CC = H( KBOT, KBOT-1 ) / S
597 BB = H( KBOT-1, KBOT ) / S
598 DD = H( KBOT, KBOT ) / S
599 TR2 = ( AA+DD ) / TWO
600 DET = ( AA-TR2 )*( DD-TR2 ) - BB*CC
601 RTDISC = SQRT( -DET )
602 W( KBOT-1 ) = ( TR2+RTDISC )*S
603 W( KBOT ) = ( TR2-RTDISC )*S
609 IF( KBOT-KS+1.GT.NS ) THEN
611 * ==== Sort the shifts (Helps a little) ====
614 DO 50 K = KBOT, KS + 1, -1
619 IF( CABS1( W( I ) ).LT.CABS1( W( I+1 ) ) )
632 * ==== If there are only two shifts, then use
635 IF( KBOT-KS+1.EQ.2 ) THEN
636 IF( CABS1( W( KBOT )-H( KBOT, KBOT ) ).LT.
637 $ CABS1( W( KBOT-1 )-H( KBOT, KBOT ) ) ) THEN
638 W( KBOT-1 ) = W( KBOT )
640 W( KBOT ) = W( KBOT-1 )
644 * ==== Use up to NS of the the smallest magnatiude
645 * . shifts. If there aren't NS shifts available,
646 * . then use them all, possibly dropping one to
647 * . make the number of shifts even. ====
649 NS = MIN( NS, KBOT-KS+1 )
650 NS = NS - MOD( NS, 2 )
653 * ==== Small-bulge multi-shift QR sweep:
654 * . split workspace under the subdiagonal into
655 * . - a KDU-by-KDU work array U in the lower
656 * . left-hand-corner,
657 * . - a KDU-by-at-least-KDU-but-more-is-better
658 * . (KDU-by-NHo) horizontal work array WH along
660 * . - and an at-least-KDU-but-more-is-better-by-KDU
661 * . (NVE-by-KDU) vertical work WV arrow along
662 * . the left-hand-edge. ====
667 NHO = ( N-KDU+1-4 ) - ( KDU+1 ) + 1
669 NVE = N - KDU - KWV + 1
671 * ==== Small-bulge multi-shift QR sweep ====
673 CALL ZLAQR5( WANTT, WANTZ, KACC22, N, KTOP, KBOT, NS,
674 $ W( KS ), H, LDH, ILOZ, IHIZ, Z, LDZ, WORK,
675 $ 3, H( KU, 1 ), LDH, NVE, H( KWV, 1 ), LDH,
676 $ NHO, H( KU, KWH ), LDH )
679 * ==== Note progress (or the lack of it). ====
687 * ==== End of main loop ====
690 * ==== Iteration limit exceeded. Set INFO to show where
691 * . the problem occurred and exit. ====
697 * ==== Return the optimal value of LWORK. ====
699 WORK( 1 ) = DCMPLX( LWKOPT, 0 )
701 * ==== End of ZLAQR4 ====