1 *> \brief \b SLAQR0 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 SLAQR0 + dependencies
10 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/slaqr0.f">
12 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/slaqr0.f">
14 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/slaqr0.f">
21 * SUBROUTINE SLAQR0( WANTT, WANTZ, N, ILO, IHI, H, LDH, WR, WI,
22 * ILOZ, 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 * REAL H( LDH, * ), WI( * ), WORK( * ), WR( * ),
39 *> SLAQR0 computes the eigenvalues of a Hessenberg matrix H
40 *> and, optionally, the matrices T and Z from the Schur decomposition
41 *> H = Z T Z**T, where T is an upper quasi-triangular matrix (the
42 *> Schur form), and Z is the orthogonal matrix of Schur vectors.
44 *> Optionally Z may be postmultiplied into an input orthogonal
45 *> matrix Q so that this routine can give the Schur factorization
46 *> of a matrix A which has been reduced to the Hessenberg form H
47 *> by the orthogonal matrix Q: A = Q*H*Q**T = (QZ)*T*(QZ)**T.
56 *> = .TRUE. : the full Schur form T is required;
57 *> = .FALSE.: only eigenvalues are required.
63 *> = .TRUE. : the matrix of Schur vectors Z is required;
64 *> = .FALSE.: Schur vectors are not required.
70 *> The order of the matrix H. N .GE. 0.
81 *> It is assumed that H is already upper triangular in rows
82 *> and columns 1:ILO-1 and IHI+1:N and, if ILO.GT.1,
83 *> H(ILO,ILO-1) is zero. ILO and IHI are normally set by a
84 *> previous call to SGEBAL, and then passed to SGEHRD when the
85 *> matrix output by SGEBAL is reduced to Hessenberg form.
86 *> Otherwise, ILO and IHI should be set to 1 and N,
87 *> respectively. If N.GT.0, then 1.LE.ILO.LE.IHI.LE.N.
88 *> If N = 0, then ILO = 1 and IHI = 0.
93 *> H is REAL array, dimension (LDH,N)
94 *> On entry, the upper Hessenberg matrix H.
95 *> On exit, if INFO = 0 and WANTT is .TRUE., then H contains
96 *> the upper quasi-triangular matrix T from the Schur
97 *> decomposition (the Schur form); 2-by-2 diagonal blocks
98 *> (corresponding to complex conjugate pairs of eigenvalues)
99 *> are returned in standard form, with H(i,i) = H(i+1,i+1)
100 *> and H(i+1,i)*H(i,i+1).LT.0. If INFO = 0 and WANTT is
101 *> .FALSE., then the contents of H are unspecified on exit.
102 *> (The output value of H when INFO.GT.0 is given under the
103 *> description of INFO below.)
105 *> This subroutine may explicitly set H(i,j) = 0 for i.GT.j and
106 *> j = 1, 2, ... ILO-1 or j = IHI+1, IHI+2, ... N.
112 *> The leading dimension of the array H. LDH .GE. max(1,N).
117 *> WR is REAL array, dimension (IHI)
122 *> WI is REAL array, dimension (IHI)
123 *> The real and imaginary parts, respectively, of the computed
124 *> eigenvalues of H(ILO:IHI,ILO:IHI) are stored in WR(ILO:IHI)
125 *> and WI(ILO:IHI). If two eigenvalues are computed as a
126 *> complex conjugate pair, they are stored in consecutive
127 *> elements of WR and WI, say the i-th and (i+1)th, with
128 *> WI(i) .GT. 0 and WI(i+1) .LT. 0. If WANTT is .TRUE., then
129 *> the eigenvalues are stored in the same order as on the
130 *> diagonal of the Schur form returned in H, with
131 *> WR(i) = H(i,i) and, if H(i:i+1,i:i+1) is a 2-by-2 diagonal
132 *> block, WI(i) = sqrt(-H(i+1,i)*H(i,i+1)) and
144 *> Specify the rows of Z to which transformations must be
145 *> applied if WANTZ is .TRUE..
146 *> 1 .LE. ILOZ .LE. ILO; IHI .LE. IHIZ .LE. N.
151 *> Z is REAL array, dimension (LDZ,IHI)
152 *> If WANTZ is .FALSE., then Z is not referenced.
153 *> If WANTZ is .TRUE., then Z(ILO:IHI,ILOZ:IHIZ) is
154 *> replaced by Z(ILO:IHI,ILOZ:IHIZ)*U where U is the
155 *> orthogonal Schur factor of H(ILO:IHI,ILO:IHI).
156 *> (The output value of Z when INFO.GT.0 is given under
157 *> the description of INFO below.)
163 *> The leading dimension of the array Z. if WANTZ is .TRUE.
164 *> then LDZ.GE.MAX(1,IHIZ). Otherwize, LDZ.GE.1.
169 *> WORK is REAL array, dimension LWORK
170 *> On exit, if LWORK = -1, WORK(1) returns an estimate of
171 *> the optimal value for LWORK.
177 *> The dimension of the array WORK. LWORK .GE. max(1,N)
178 *> is sufficient, but LWORK typically as large as 6*N may
179 *> be required for optimal performance. A workspace query
180 *> to determine the optimal workspace size is recommended.
182 *> If LWORK = -1, then SLAQR0 does a workspace query.
183 *> In this case, SLAQR0 checks the input parameters and
184 *> estimates the optimal workspace size for the given
185 *> values of N, ILO and IHI. The estimate is returned
186 *> in WORK(1). No error message related to LWORK is
187 *> issued by XERBLA. Neither H nor Z are accessed.
193 *> = 0: successful exit
194 *> .GT. 0: if INFO = i, SLAQR0 failed to compute all of
195 *> the eigenvalues. Elements 1:ilo-1 and i+1:n of WR
196 *> and WI contain those eigenvalues which have been
197 *> successfully computed. (Failures are rare.)
199 *> If INFO .GT. 0 and WANT is .FALSE., then on exit,
200 *> the remaining unconverged eigenvalues are the eigen-
201 *> values of the upper Hessenberg matrix rows and
202 *> columns ILO through INFO of the final, output
205 *> If INFO .GT. 0 and WANTT is .TRUE., then on exit
207 *> (*) (initial value of H)*U = U*(final value of H)
209 *> where U is an orthogonal matrix. The final
210 *> value of H is upper Hessenberg and quasi-triangular
211 *> in rows and columns INFO+1 through IHI.
213 *> If INFO .GT. 0 and WANTZ is .TRUE., then on exit
215 *> (final value of Z(ILO:IHI,ILOZ:IHIZ)
216 *> = (initial value of Z(ILO:IHI,ILOZ:IHIZ)*U
218 *> where U is the orthogonal matrix in (*) (regard-
219 *> less of the value of WANTT.)
221 *> If INFO .GT. 0 and WANTZ is .FALSE., then Z is not
228 *> \author Univ. of Tennessee
229 *> \author Univ. of California Berkeley
230 *> \author Univ. of Colorado Denver
233 *> \date September 2012
235 *> \ingroup realOTHERauxiliary
237 *> \par Contributors:
240 *> Karen Braman and Ralph Byers, Department of Mathematics,
241 *> University of Kansas, USA
246 *> K. Braman, R. Byers and R. Mathias, The Multi-Shift QR
247 *> Algorithm Part I: Maintaining Well Focused Shifts, and Level 3
248 *> Performance, SIAM Journal of Matrix Analysis, volume 23, pages
251 *> K. Braman, R. Byers and R. Mathias, The Multi-Shift QR
252 *> Algorithm Part II: Aggressive Early Deflation, SIAM Journal
253 *> of Matrix Analysis, volume 23, pages 948--973, 2002.
255 * =====================================================================
256 SUBROUTINE SLAQR0( WANTT, WANTZ, N, ILO, IHI, H, LDH, WR, WI,
257 $ ILOZ, IHIZ, Z, LDZ, WORK, LWORK, INFO )
259 * -- LAPACK auxiliary routine (version 3.4.2) --
260 * -- LAPACK is a software package provided by Univ. of Tennessee, --
261 * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
264 * .. Scalar Arguments ..
265 INTEGER IHI, IHIZ, ILO, ILOZ, INFO, LDH, LDZ, LWORK, N
268 * .. Array Arguments ..
269 REAL H( LDH, * ), WI( * ), WORK( * ), WR( * ),
273 * ================================================================
276 * ==== Matrices of order NTINY or smaller must be processed by
277 * . SLAHQR because of insufficient subdiagonal scratch space.
278 * . (This is a hard limit.) ====
280 PARAMETER ( NTINY = 11 )
282 * ==== Exceptional deflation windows: try to cure rare
283 * . slow convergence by varying the size of the
284 * . deflation window after KEXNW iterations. ====
286 PARAMETER ( KEXNW = 5 )
288 * ==== Exceptional shifts: try to cure rare slow convergence
289 * . with ad-hoc exceptional shifts every KEXSH iterations.
292 PARAMETER ( KEXSH = 6 )
294 * ==== The constants WILK1 and WILK2 are used to form the
295 * . exceptional shifts. ====
297 PARAMETER ( WILK1 = 0.75e0, WILK2 = -0.4375e0 )
299 PARAMETER ( ZERO = 0.0e0, ONE = 1.0e0 )
301 * .. Local Scalars ..
302 REAL AA, BB, CC, CS, DD, SN, SS, SWAP
303 INTEGER I, INF, IT, ITMAX, K, KACC22, KBOT, KDU, KS,
304 $ KT, KTOP, KU, KV, KWH, KWTOP, KWV, LD, LS,
305 $ LWKOPT, NDEC, NDFL, NH, NHO, NIBBLE, NMIN, NS,
306 $ NSMAX, NSR, NVE, NW, NWMAX, NWR, NWUPBD
310 * .. External Functions ..
317 * .. External Subroutines ..
318 EXTERNAL SLACPY, SLAHQR, SLANV2, SLAQR3, SLAQR4, SLAQR5
320 * .. Intrinsic Functions ..
321 INTRINSIC ABS, INT, MAX, MIN, MOD, REAL
323 * .. Executable Statements ..
326 * ==== Quick return for N = 0: nothing to do. ====
333 IF( N.LE.NTINY ) THEN
335 * ==== Tiny matrices must use SLAHQR. ====
339 $ CALL SLAHQR( WANTT, WANTZ, N, ILO, IHI, H, LDH, WR, WI,
340 $ ILOZ, IHIZ, Z, LDZ, INFO )
343 * ==== Use small bulge multi-shift QR with aggressive early
344 * . deflation on larger-than-tiny matrices. ====
346 * ==== Hope for the best. ====
350 * ==== Set up job flags for ILAENV. ====
363 * ==== NWR = recommended deflation window size. At this
364 * . point, N .GT. NTINY = 11, so there is enough
365 * . subdiagonal workspace for NWR.GE.2 as required.
366 * . (In fact, there is enough subdiagonal space for
369 NWR = ILAENV( 13, 'SLAQR0', JBCMPZ, N, ILO, IHI, LWORK )
371 NWR = MIN( IHI-ILO+1, ( N-1 ) / 3, NWR )
373 * ==== NSR = recommended number of simultaneous shifts.
374 * . At this point N .GT. NTINY = 11, so there is at
375 * . enough subdiagonal workspace for NSR to be even
376 * . and greater than or equal to two as required. ====
378 NSR = ILAENV( 15, 'SLAQR0', JBCMPZ, N, ILO, IHI, LWORK )
379 NSR = MIN( NSR, ( N+6 ) / 9, IHI-ILO )
380 NSR = MAX( 2, NSR-MOD( NSR, 2 ) )
382 * ==== Estimate optimal workspace ====
384 * ==== Workspace query call to SLAQR3 ====
386 CALL SLAQR3( WANTT, WANTZ, N, ILO, IHI, NWR+1, H, LDH, ILOZ,
387 $ IHIZ, Z, LDZ, LS, LD, WR, WI, H, LDH, N, H, LDH,
388 $ N, H, LDH, WORK, -1 )
390 * ==== Optimal workspace = MAX(SLAQR5, SLAQR3) ====
392 LWKOPT = MAX( 3*NSR / 2, INT( WORK( 1 ) ) )
394 * ==== Quick return in case of workspace query. ====
396 IF( LWORK.EQ.-1 ) THEN
397 WORK( 1 ) = REAL( LWKOPT )
401 * ==== SLAHQR/SLAQR0 crossover point ====
403 NMIN = ILAENV( 12, 'SLAQR0', JBCMPZ, N, ILO, IHI, LWORK )
404 NMIN = MAX( NTINY, NMIN )
406 * ==== Nibble crossover point ====
408 NIBBLE = ILAENV( 14, 'SLAQR0', JBCMPZ, N, ILO, IHI, LWORK )
409 NIBBLE = MAX( 0, NIBBLE )
411 * ==== Accumulate reflections during ttswp? Use block
412 * . 2-by-2 structure during matrix-matrix multiply? ====
414 KACC22 = ILAENV( 16, 'SLAQR0', JBCMPZ, N, ILO, IHI, LWORK )
415 KACC22 = MAX( 0, KACC22 )
416 KACC22 = MIN( 2, KACC22 )
418 * ==== NWMAX = the largest possible deflation window for
419 * . which there is sufficient workspace. ====
421 NWMAX = MIN( ( N-1 ) / 3, LWORK / 2 )
424 * ==== NSMAX = the Largest number of simultaneous shifts
425 * . for which there is sufficient workspace. ====
427 NSMAX = MIN( ( N+6 ) / 9, 2*LWORK / 3 )
428 NSMAX = NSMAX - MOD( NSMAX, 2 )
430 * ==== NDFL: an iteration count restarted at deflation. ====
434 * ==== ITMAX = iteration limit ====
436 ITMAX = MAX( 30, 2*KEXSH )*MAX( 10, ( IHI-ILO+1 ) )
438 * ==== Last row and column in the active block ====
442 * ==== Main Loop ====
446 * ==== Done when KBOT falls below ILO ====
451 * ==== Locate active block ====
453 DO 10 K = KBOT, ILO + 1, -1
454 IF( H( K, K-1 ).EQ.ZERO )
461 * ==== Select deflation window size:
463 * . If possible and advisable, nibble the entire
464 * . active block. If not, use size MIN(NWR,NWMAX)
465 * . or MIN(NWR+1,NWMAX) depending upon which has
466 * . the smaller corresponding subdiagonal entry
469 * . Exceptional Case:
470 * . If there have been no deflations in KEXNW or
471 * . more iterations, then vary the deflation window
472 * . size. At first, because, larger windows are,
473 * . in general, more powerful than smaller ones,
474 * . rapidly increase the window to the maximum possible.
475 * . Then, gradually reduce the window size. ====
478 NWUPBD = MIN( NH, NWMAX )
479 IF( NDFL.LT.KEXNW ) THEN
480 NW = MIN( NWUPBD, NWR )
482 NW = MIN( NWUPBD, 2*NW )
484 IF( NW.LT.NWMAX ) THEN
485 IF( NW.GE.NH-1 ) THEN
488 KWTOP = KBOT - NW + 1
489 IF( ABS( H( KWTOP, KWTOP-1 ) ).GT.
490 $ ABS( H( KWTOP-1, KWTOP-2 ) ) )NW = NW + 1
493 IF( NDFL.LT.KEXNW ) THEN
495 ELSE IF( NDEC.GE.0 .OR. NW.GE.NWUPBD ) THEN
502 * ==== Aggressive early deflation:
503 * . split workspace under the subdiagonal into
504 * . - an nw-by-nw work array V in the lower
505 * . left-hand-corner,
506 * . - an NW-by-at-least-NW-but-more-is-better
507 * . (NW-by-NHO) horizontal work array along
509 * . - an at-least-NW-but-more-is-better (NHV-by-NW)
510 * . vertical work array along the left-hand-edge.
515 NHO = ( N-NW-1 ) - KT + 1
517 NVE = ( N-NW ) - KWV + 1
519 * ==== Aggressive early deflation ====
521 CALL SLAQR3( WANTT, WANTZ, N, KTOP, KBOT, NW, H, LDH, ILOZ,
522 $ IHIZ, Z, LDZ, LS, LD, WR, WI, H( KV, 1 ), LDH,
523 $ NHO, H( KV, KT ), LDH, NVE, H( KWV, 1 ), LDH,
526 * ==== Adjust KBOT accounting for new deflations. ====
530 * ==== KS points to the shifts. ====
534 * ==== Skip an expensive QR sweep if there is a (partly
535 * . heuristic) reason to expect that many eigenvalues
536 * . will deflate without it. Here, the QR sweep is
537 * . skipped if many eigenvalues have just been deflated
538 * . or if the remaining active block is small.
540 IF( ( LD.EQ.0 ) .OR. ( ( 100*LD.LE.NW*NIBBLE ) .AND. ( KBOT-
541 $ KTOP+1.GT.MIN( NMIN, NWMAX ) ) ) ) THEN
543 * ==== NS = nominal number of simultaneous shifts.
544 * . This may be lowered (slightly) if SLAQR3
545 * . did not provide that many shifts. ====
547 NS = MIN( NSMAX, NSR, MAX( 2, KBOT-KTOP ) )
548 NS = NS - MOD( NS, 2 )
550 * ==== If there have been no deflations
551 * . in a multiple of KEXSH iterations,
552 * . then try exceptional shifts.
553 * . Otherwise use shifts provided by
554 * . SLAQR3 above or from the eigenvalues
555 * . of a trailing principal submatrix. ====
557 IF( MOD( NDFL, KEXSH ).EQ.0 ) THEN
559 DO 30 I = KBOT, MAX( KS+1, KTOP+2 ), -2
560 SS = ABS( H( I, I-1 ) ) + ABS( H( I-1, I-2 ) )
561 AA = WILK1*SS + H( I, I )
565 CALL SLANV2( AA, BB, CC, DD, WR( I-1 ), WI( I-1 ),
566 $ WR( I ), WI( I ), CS, SN )
568 IF( KS.EQ.KTOP ) THEN
569 WR( KS+1 ) = H( KS+1, KS+1 )
571 WR( KS ) = WR( KS+1 )
572 WI( KS ) = WI( KS+1 )
576 * ==== Got NS/2 or fewer shifts? Use SLAQR4 or
577 * . SLAHQR on a trailing principal submatrix to
578 * . get more. (Since NS.LE.NSMAX.LE.(N+6)/9,
579 * . there is enough space below the subdiagonal
580 * . to fit an NS-by-NS scratch array.) ====
582 IF( KBOT-KS+1.LE.NS / 2 ) THEN
585 CALL SLACPY( 'A', NS, NS, H( KS, KS ), LDH,
587 IF( NS.GT.NMIN ) THEN
588 CALL SLAQR4( .false., .false., NS, 1, NS,
589 $ H( KT, 1 ), LDH, WR( KS ),
590 $ WI( KS ), 1, 1, ZDUM, 1, WORK,
593 CALL SLAHQR( .false., .false., NS, 1, NS,
594 $ H( KT, 1 ), LDH, WR( KS ),
595 $ WI( KS ), 1, 1, ZDUM, 1, INF )
599 * ==== In case of a rare QR failure use
600 * . eigenvalues of the trailing 2-by-2
601 * . principal submatrix. ====
603 IF( KS.GE.KBOT ) THEN
604 AA = H( KBOT-1, KBOT-1 )
605 CC = H( KBOT, KBOT-1 )
606 BB = H( KBOT-1, KBOT )
608 CALL SLANV2( AA, BB, CC, DD, WR( KBOT-1 ),
609 $ WI( KBOT-1 ), WR( KBOT ),
610 $ WI( KBOT ), CS, SN )
615 IF( KBOT-KS+1.GT.NS ) THEN
617 * ==== Sort the shifts (Helps a little)
618 * . Bubble sort keeps complex conjugate
619 * . pairs together. ====
622 DO 50 K = KBOT, KS + 1, -1
627 IF( ABS( WR( I ) )+ABS( WI( I ) ).LT.
628 $ ABS( WR( I+1 ) )+ABS( WI( I+1 ) ) ) THEN
644 * ==== Shuffle shifts into pairs of real shifts
645 * . and pairs of complex conjugate shifts
646 * . assuming complex conjugate shifts are
647 * . already adjacent to one another. (Yes,
650 DO 70 I = KBOT, KS + 2, -2
651 IF( WI( I ).NE.-WI( I-1 ) ) THEN
655 WR( I-1 ) = WR( I-2 )
660 WI( I-1 ) = WI( I-2 )
666 * ==== If there are only two shifts and both are
667 * . real, then use only one. ====
669 IF( KBOT-KS+1.EQ.2 ) THEN
670 IF( WI( KBOT ).EQ.ZERO ) THEN
671 IF( ABS( WR( KBOT )-H( KBOT, KBOT ) ).LT.
672 $ ABS( WR( KBOT-1 )-H( KBOT, KBOT ) ) ) THEN
673 WR( KBOT-1 ) = WR( KBOT )
675 WR( KBOT ) = WR( KBOT-1 )
680 * ==== Use up to NS of the the smallest magnatiude
681 * . shifts. If there aren't NS shifts available,
682 * . then use them all, possibly dropping one to
683 * . make the number of shifts even. ====
685 NS = MIN( NS, KBOT-KS+1 )
686 NS = NS - MOD( NS, 2 )
689 * ==== Small-bulge multi-shift QR sweep:
690 * . split workspace under the subdiagonal into
691 * . - a KDU-by-KDU work array U in the lower
692 * . left-hand-corner,
693 * . - a KDU-by-at-least-KDU-but-more-is-better
694 * . (KDU-by-NHo) horizontal work array WH along
696 * . - and an at-least-KDU-but-more-is-better-by-KDU
697 * . (NVE-by-KDU) vertical work WV arrow along
698 * . the left-hand-edge. ====
703 NHO = ( N-KDU+1-4 ) - ( KDU+1 ) + 1
705 NVE = N - KDU - KWV + 1
707 * ==== Small-bulge multi-shift QR sweep ====
709 CALL SLAQR5( WANTT, WANTZ, KACC22, N, KTOP, KBOT, NS,
710 $ WR( KS ), WI( KS ), H, LDH, ILOZ, IHIZ, Z,
711 $ LDZ, WORK, 3, H( KU, 1 ), LDH, NVE,
712 $ H( KWV, 1 ), LDH, NHO, H( KU, KWH ), LDH )
715 * ==== Note progress (or the lack of it). ====
723 * ==== End of main loop ====
726 * ==== Iteration limit exceeded. Set INFO to show where
727 * . the problem occurred and exit. ====
733 * ==== Return the optimal value of LWORK. ====
735 WORK( 1 ) = REAL( LWKOPT )
737 * ==== End of SLAQR0 ====