forked from Reference-LAPACK/lapack
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdgejsv.f
1780 lines (1771 loc) · 70.8 KB
/
dgejsv.f
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
*> \brief \b DGEJSV
*
* =========== DOCUMENTATION ===========
*
* Online html documentation available at
* http://www.netlib.org/lapack/explore-html/
*
*> \htmlonly
*> Download DGEJSV + dependencies
*> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/dgejsv.f">
*> [TGZ]</a>
*> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/dgejsv.f">
*> [ZIP]</a>
*> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/dgejsv.f">
*> [TXT]</a>
*> \endhtmlonly
*
* Definition:
* ===========
*
* SUBROUTINE DGEJSV( JOBA, JOBU, JOBV, JOBR, JOBT, JOBP,
* M, N, A, LDA, SVA, U, LDU, V, LDV,
* WORK, LWORK, IWORK, INFO )
*
* .. Scalar Arguments ..
* IMPLICIT NONE
* INTEGER INFO, LDA, LDU, LDV, LWORK, M, N
* ..
* .. Array Arguments ..
* DOUBLE PRECISION A( LDA, * ), SVA( N ), U( LDU, * ), V( LDV, * ),
* $ WORK( LWORK )
* INTEGER IWORK( * )
* CHARACTER*1 JOBA, JOBP, JOBR, JOBT, JOBU, JOBV
* ..
*
*
*> \par Purpose:
* =============
*>
*> \verbatim
*>
*> DGEJSV computes the singular value decomposition (SVD) of a real M-by-N
*> matrix [A], where M >= N. The SVD of [A] is written as
*>
*> [A] = [U] * [SIGMA] * [V]^t,
*>
*> where [SIGMA] is an N-by-N (M-by-N) matrix which is zero except for its N
*> diagonal elements, [U] is an M-by-N (or M-by-M) orthonormal matrix, and
*> [V] is an N-by-N orthogonal matrix. The diagonal elements of [SIGMA] are
*> the singular values of [A]. The columns of [U] and [V] are the left and
*> the right singular vectors of [A], respectively. The matrices [U] and [V]
*> are computed and stored in the arrays U and V, respectively. The diagonal
*> of [SIGMA] is computed and stored in the array SVA.
*> DGEJSV can sometimes compute tiny singular values and their singular vectors much
*> more accurately than other SVD routines, see below under Further Details.
*> \endverbatim
*
* Arguments:
* ==========
*
*> \param[in] JOBA
*> \verbatim
*> JOBA is CHARACTER*1
*> Specifies the level of accuracy:
*> = 'C': This option works well (high relative accuracy) if A = B * D,
*> with well-conditioned B and arbitrary diagonal matrix D.
*> The accuracy cannot be spoiled by COLUMN scaling. The
*> accuracy of the computed output depends on the condition of
*> B, and the procedure aims at the best theoretical accuracy.
*> The relative error max_{i=1:N}|d sigma_i| / sigma_i is
*> bounded by f(M,N)*epsilon* cond(B), independent of D.
*> The input matrix is preprocessed with the QRF with column
*> pivoting. This initial preprocessing and preconditioning by
*> a rank revealing QR factorization is common for all values of
*> JOBA. Additional actions are specified as follows:
*> = 'E': Computation as with 'C' with an additional estimate of the
*> condition number of B. It provides a realistic error bound.
*> = 'F': If A = D1 * C * D2 with ill-conditioned diagonal scalings
*> D1, D2, and well-conditioned matrix C, this option gives
*> higher accuracy than the 'C' option. If the structure of the
*> input matrix is not known, and relative accuracy is
*> desirable, then this option is advisable. The input matrix A
*> is preprocessed with QR factorization with FULL (row and
*> column) pivoting.
*> = 'G': Computation as with 'F' with an additional estimate of the
*> condition number of B, where A=D*B. If A has heavily weighted
*> rows, then using this condition number gives too pessimistic
*> error bound.
*> = 'A': Small singular values are the noise and the matrix is treated
*> as numerically rank deficient. The error in the computed
*> singular values is bounded by f(m,n)*epsilon*||A||.
*> The computed SVD A = U * S * V^t restores A up to
*> f(m,n)*epsilon*||A||.
*> This gives the procedure the licence to discard (set to zero)
*> all singular values below N*epsilon*||A||.
*> = 'R': Similar as in 'A'. Rank revealing property of the initial
*> QR factorization is used do reveal (using triangular factor)
*> a gap sigma_{r+1} < epsilon * sigma_r in which case the
*> numerical RANK is declared to be r. The SVD is computed with
*> absolute error bounds, but more accurately than with 'A'.
*> \endverbatim
*>
*> \param[in] JOBU
*> \verbatim
*> JOBU is CHARACTER*1
*> Specifies whether to compute the columns of U:
*> = 'U': N columns of U are returned in the array U.
*> = 'F': full set of M left sing. vectors is returned in the array U.
*> = 'W': U may be used as workspace of length M*N. See the description
*> of U.
*> = 'N': U is not computed.
*> \endverbatim
*>
*> \param[in] JOBV
*> \verbatim
*> JOBV is CHARACTER*1
*> Specifies whether to compute the matrix V:
*> = 'V': N columns of V are returned in the array V; Jacobi rotations
*> are not explicitly accumulated.
*> = 'J': N columns of V are returned in the array V, but they are
*> computed as the product of Jacobi rotations. This option is
*> allowed only if JOBU .NE. 'N', i.e. in computing the full SVD.
*> = 'W': V may be used as workspace of length N*N. See the description
*> of V.
*> = 'N': V is not computed.
*> \endverbatim
*>
*> \param[in] JOBR
*> \verbatim
*> JOBR is CHARACTER*1
*> Specifies the RANGE for the singular values. Issues the licence to
*> set to zero small positive singular values if they are outside
*> specified range. If A .NE. 0 is scaled so that the largest singular
*> value of c*A is around DSQRT(BIG), BIG=SLAMCH('O'), then JOBR issues
*> the licence to kill columns of A whose norm in c*A is less than
*> DSQRT(SFMIN) (for JOBR = 'R'), or less than SMALL=SFMIN/EPSLN,
*> where SFMIN=SLAMCH('S'), EPSLN=SLAMCH('E').
*> = 'N': Do not kill small columns of c*A. This option assumes that
*> BLAS and QR factorizations and triangular solvers are
*> implemented to work in that range. If the condition of A
*> is greater than BIG, use DGESVJ.
*> = 'R': RESTRICTED range for sigma(c*A) is [DSQRT(SFMIN), DSQRT(BIG)]
*> (roughly, as described above). This option is recommended.
*> ~~~~~~~~~~~~~~~~~~~~~~~~~~~
*> For computing the singular values in the FULL range [SFMIN,BIG]
*> use DGESVJ.
*> \endverbatim
*>
*> \param[in] JOBT
*> \verbatim
*> JOBT is CHARACTER*1
*> If the matrix is square then the procedure may determine to use
*> transposed A if A^t seems to be better with respect to convergence.
*> If the matrix is not square, JOBT is ignored. This is subject to
*> changes in the future.
*> The decision is based on two values of entropy over the adjoint
*> orbit of A^t * A. See the descriptions of WORK(6) and WORK(7).
*> = 'T': transpose if entropy test indicates possibly faster
*> convergence of Jacobi process if A^t is taken as input. If A is
*> replaced with A^t, then the row pivoting is included automatically.
*> = 'N': do not speculate.
*> This option can be used to compute only the singular values, or the
*> full SVD (U, SIGMA and V). For only one set of singular vectors
*> (U or V), the caller should provide both U and V, as one of the
*> matrices is used as workspace if the matrix A is transposed.
*> The implementer can easily remove this constraint and make the
*> code more complicated. See the descriptions of U and V.
*> \endverbatim
*>
*> \param[in] JOBP
*> \verbatim
*> JOBP is CHARACTER*1
*> Issues the licence to introduce structured perturbations to drown
*> denormalized numbers. This licence should be active if the
*> denormals are poorly implemented, causing slow computation,
*> especially in cases of fast convergence (!). For details see [1,2].
*> For the sake of simplicity, this perturbations are included only
*> when the full SVD or only the singular values are requested. The
*> implementer/user can easily add the perturbation for the cases of
*> computing one set of singular vectors.
*> = 'P': introduce perturbation
*> = 'N': do not perturb
*> \endverbatim
*>
*> \param[in] M
*> \verbatim
*> M is INTEGER
*> The number of rows of the input matrix A. M >= 0.
*> \endverbatim
*>
*> \param[in] N
*> \verbatim
*> N is INTEGER
*> The number of columns of the input matrix A. M >= N >= 0.
*> \endverbatim
*>
*> \param[in,out] A
*> \verbatim
*> A is DOUBLE PRECISION array, dimension (LDA,N)
*> On entry, the M-by-N matrix A.
*> \endverbatim
*>
*> \param[in] LDA
*> \verbatim
*> LDA is INTEGER
*> The leading dimension of the array A. LDA >= max(1,M).
*> \endverbatim
*>
*> \param[out] SVA
*> \verbatim
*> SVA is DOUBLE PRECISION array, dimension (N)
*> On exit,
*> - For WORK(1)/WORK(2) = ONE: The singular values of A. During the
*> computation SVA contains Euclidean column norms of the
*> iterated matrices in the array A.
*> - For WORK(1) .NE. WORK(2): The singular values of A are
*> (WORK(1)/WORK(2)) * SVA(1:N). This factored form is used if
*> sigma_max(A) overflows or if small singular values have been
*> saved from underflow by scaling the input matrix A.
*> - If JOBR='R' then some of the singular values may be returned
*> as exact zeros obtained by "set to zero" because they are
*> below the numerical rank threshold or are denormalized numbers.
*> \endverbatim
*>
*> \param[out] U
*> \verbatim
*> U is DOUBLE PRECISION array, dimension ( LDU, N ) or ( LDU, M )
*> If JOBU = 'U', then U contains on exit the M-by-N matrix of
*> the left singular vectors.
*> If JOBU = 'F', then U contains on exit the M-by-M matrix of
*> the left singular vectors, including an ONB
*> of the orthogonal complement of the Range(A).
*> If JOBU = 'W' .AND. (JOBV = 'V' .AND. JOBT = 'T' .AND. M = N),
*> then U is used as workspace if the procedure
*> replaces A with A^t. In that case, [V] is computed
*> in U as left singular vectors of A^t and then
*> copied back to the V array. This 'W' option is just
*> a reminder to the caller that in this case U is
*> reserved as workspace of length N*N.
*> If JOBU = 'N' U is not referenced, unless JOBT='T'.
*> \endverbatim
*>
*> \param[in] LDU
*> \verbatim
*> LDU is INTEGER
*> The leading dimension of the array U, LDU >= 1.
*> IF JOBU = 'U' or 'F' or 'W', then LDU >= M.
*> \endverbatim
*>
*> \param[out] V
*> \verbatim
*> V is DOUBLE PRECISION array, dimension ( LDV, N )
*> If JOBV = 'V', 'J' then V contains on exit the N-by-N matrix of
*> the right singular vectors;
*> If JOBV = 'W', AND (JOBU = 'U' AND JOBT = 'T' AND M = N),
*> then V is used as workspace if the procedure
*> replaces A with A^t. In that case, [U] is computed
*> in V as right singular vectors of A^t and then
*> copied back to the U array. This 'W' option is just
*> a reminder to the caller that in this case V is
*> reserved as workspace of length N*N.
*> If JOBV = 'N' V is not referenced, unless JOBT='T'.
*> \endverbatim
*>
*> \param[in] LDV
*> \verbatim
*> LDV is INTEGER
*> The leading dimension of the array V, LDV >= 1.
*> If JOBV = 'V' or 'J' or 'W', then LDV >= N.
*> \endverbatim
*>
*> \param[out] WORK
*> \verbatim
*> WORK is DOUBLE PRECISION array, dimension (MAX(7,LWORK))
*> On exit, if N > 0 .AND. M > 0 (else not referenced),
*> WORK(1) = SCALE = WORK(2) / WORK(1) is the scaling factor such
*> that SCALE*SVA(1:N) are the computed singular values
*> of A. (See the description of SVA().)
*> WORK(2) = See the description of WORK(1).
*> WORK(3) = SCONDA is an estimate for the condition number of
*> column equilibrated A. (If JOBA = 'E' or 'G')
*> SCONDA is an estimate of DSQRT(||(R^t * R)^(-1)||_1).
*> It is computed using DPOCON. It holds
*> N^(-1/4) * SCONDA <= ||R^(-1)||_2 <= N^(1/4) * SCONDA
*> where R is the triangular factor from the QRF of A.
*> However, if R is truncated and the numerical rank is
*> determined to be strictly smaller than N, SCONDA is
*> returned as -1, thus indicating that the smallest
*> singular values might be lost.
*>
*> If full SVD is needed, the following two condition numbers are
*> useful for the analysis of the algorithm. They are provided for
*> a developer/implementer who is familiar with the details of
*> the method.
*>
*> WORK(4) = an estimate of the scaled condition number of the
*> triangular factor in the first QR factorization.
*> WORK(5) = an estimate of the scaled condition number of the
*> triangular factor in the second QR factorization.
*> The following two parameters are computed if JOBT = 'T'.
*> They are provided for a developer/implementer who is familiar
*> with the details of the method.
*>
*> WORK(6) = the entropy of A^t*A :: this is the Shannon entropy
*> of diag(A^t*A) / Trace(A^t*A) taken as point in the
*> probability simplex.
*> WORK(7) = the entropy of A*A^t.
*> \endverbatim
*>
*> \param[in] LWORK
*> \verbatim
*> LWORK is INTEGER
*> Length of WORK to confirm proper allocation of work space.
*> LWORK depends on the job:
*>
*> If only SIGMA is needed (JOBU = 'N', JOBV = 'N') and
*> -> .. no scaled condition estimate required (JOBE = 'N'):
*> LWORK >= max(2*M+N,4*N+1,7). This is the minimal requirement.
*> ->> For optimal performance (blocked code) the optimal value
*> is LWORK >= max(2*M+N,3*N+(N+1)*NB,7). Here NB is the optimal
*> block size for DGEQP3 and DGEQRF.
*> In general, optimal LWORK is computed as
*> LWORK >= max(2*M+N,N+LWORK(DGEQP3),N+LWORK(DGEQRF), 7).
*> -> .. an estimate of the scaled condition number of A is
*> required (JOBA='E', 'G'). In this case, LWORK is the maximum
*> of the above and N*N+4*N, i.e. LWORK >= max(2*M+N,N*N+4*N,7).
*> ->> For optimal performance (blocked code) the optimal value
*> is LWORK >= max(2*M+N,3*N+(N+1)*NB, N*N+4*N, 7).
*> In general, the optimal length LWORK is computed as
*> LWORK >= max(2*M+N,N+LWORK(DGEQP3),N+LWORK(DGEQRF),
*> N+N*N+LWORK(DPOCON),7).
*>
*> If SIGMA and the right singular vectors are needed (JOBV = 'V'),
*> -> the minimal requirement is LWORK >= max(2*M+N,4*N+1,7).
*> -> For optimal performance, LWORK >= max(2*M+N,3*N+(N+1)*NB,7),
*> where NB is the optimal block size for DGEQP3, DGEQRF, DGELQF,
*> DORMLQ. In general, the optimal length LWORK is computed as
*> LWORK >= max(2*M+N,N+LWORK(DGEQP3), N+LWORK(DPOCON),
*> N+LWORK(DGELQF), 2*N+LWORK(DGEQRF), N+LWORK(DORMLQ)).
*>
*> If SIGMA and the left singular vectors are needed
*> -> the minimal requirement is LWORK >= max(2*M+N,4*N+1,7).
*> -> For optimal performance:
*> if JOBU = 'U' :: LWORK >= max(2*M+N,3*N+(N+1)*NB,7),
*> if JOBU = 'F' :: LWORK >= max(2*M+N,3*N+(N+1)*NB,N+M*NB,7),
*> where NB is the optimal block size for DGEQP3, DGEQRF, DORMQR.
*> In general, the optimal length LWORK is computed as
*> LWORK >= max(2*M+N,N+LWORK(DGEQP3),N+LWORK(DPOCON),
*> 2*N+LWORK(DGEQRF), N+LWORK(DORMQR)).
*> Here LWORK(DORMQR) equals N*NB (for JOBU = 'U') or
*> M*NB (for JOBU = 'F').
*>
*> If the full SVD is needed: (JOBU = 'U' or JOBU = 'F') and
*> -> if JOBV = 'V'
*> the minimal requirement is LWORK >= max(2*M+N,6*N+2*N*N).
*> -> if JOBV = 'J' the minimal requirement is
*> LWORK >= max(2*M+N, 4*N+N*N,2*N+N*N+6).
*> -> For optimal performance, LWORK should be additionally
*> larger than N+M*NB, where NB is the optimal block size
*> for DORMQR.
*> \endverbatim
*>
*> \param[out] IWORK
*> \verbatim
*> IWORK is INTEGER array, dimension (MAX(3,M+3*N)).
*> On exit,
*> IWORK(1) = the numerical rank determined after the initial
*> QR factorization with pivoting. See the descriptions
*> of JOBA and JOBR.
*> IWORK(2) = the number of the computed nonzero singular values
*> IWORK(3) = if nonzero, a warning message:
*> If IWORK(3) = 1 then some of the column norms of A
*> were denormalized floats. The requested high accuracy
*> is not warranted by the data.
*> \endverbatim
*>
*> \param[out] INFO
*> \verbatim
*> INFO is INTEGER
*> < 0: if INFO = -i, then the i-th argument had an illegal value.
*> = 0: successful exit;
*> > 0: DGEJSV did not converge in the maximal allowed number
*> of sweeps. The computed values may be inaccurate.
*> \endverbatim
*
* Authors:
* ========
*
*> \author Univ. of Tennessee
*> \author Univ. of California Berkeley
*> \author Univ. of Colorado Denver
*> \author NAG Ltd.
*
*> \ingroup doubleGEsing
*
*> \par Further Details:
* =====================
*>
*> \verbatim
*>
*> DGEJSV implements a preconditioned Jacobi SVD algorithm. It uses DGEQP3,
*> DGEQRF, and DGELQF as preprocessors and preconditioners. Optionally, an
*> additional row pivoting can be used as a preprocessor, which in some
*> cases results in much higher accuracy. An example is matrix A with the
*> structure A = D1 * C * D2, where D1, D2 are arbitrarily ill-conditioned
*> diagonal matrices and C is well-conditioned matrix. In that case, complete
*> pivoting in the first QR factorizations provides accuracy dependent on the
*> condition number of C, and independent of D1, D2. Such higher accuracy is
*> not completely understood theoretically, but it works well in practice.
*> Further, if A can be written as A = B*D, with well-conditioned B and some
*> diagonal D, then the high accuracy is guaranteed, both theoretically and
*> in software, independent of D. For more details see [1], [2].
*> The computational range for the singular values can be the full range
*> ( UNDERFLOW,OVERFLOW ), provided that the machine arithmetic and the BLAS
*> & LAPACK routines called by DGEJSV are implemented to work in that range.
*> If that is not the case, then the restriction for safe computation with
*> the singular values in the range of normalized IEEE numbers is that the
*> spectral condition number kappa(A)=sigma_max(A)/sigma_min(A) does not
*> overflow. This code (DGEJSV) is best used in this restricted range,
*> meaning that singular values of magnitude below ||A||_2 / DLAMCH('O') are
*> returned as zeros. See JOBR for details on this.
*> Further, this implementation is somewhat slower than the one described
*> in [1,2] due to replacement of some non-LAPACK components, and because
*> the choice of some tuning parameters in the iterative part (DGESVJ) is
*> left to the implementer on a particular machine.
*> The rank revealing QR factorization (in this code: DGEQP3) should be
*> implemented as in [3]. We have a new version of DGEQP3 under development
*> that is more robust than the current one in LAPACK, with a cleaner cut in
*> rank deficient cases. It will be available in the SIGMA library [4].
*> If M is much larger than N, it is obvious that the initial QRF with
*> column pivoting can be preprocessed by the QRF without pivoting. That
*> well known trick is not used in DGEJSV because in some cases heavy row
*> weighting can be treated with complete pivoting. The overhead in cases
*> M much larger than N is then only due to pivoting, but the benefits in
*> terms of accuracy have prevailed. The implementer/user can incorporate
*> this extra QRF step easily. The implementer can also improve data movement
*> (matrix transpose, matrix copy, matrix transposed copy) - this
*> implementation of DGEJSV uses only the simplest, naive data movement.
*> \endverbatim
*
*> \par Contributors:
* ==================
*>
*> Zlatko Drmac (Zagreb, Croatia) and Kresimir Veselic (Hagen, Germany)
*
*> \par References:
* ================
*>
*> \verbatim
*>
*> [1] Z. Drmac and K. Veselic: New fast and accurate Jacobi SVD algorithm I.
*> SIAM J. Matrix Anal. Appl. Vol. 35, No. 2 (2008), pp. 1322-1342.
*> LAPACK Working note 169.
*> [2] Z. Drmac and K. Veselic: New fast and accurate Jacobi SVD algorithm II.
*> SIAM J. Matrix Anal. Appl. Vol. 35, No. 2 (2008), pp. 1343-1362.
*> LAPACK Working note 170.
*> [3] Z. Drmac and Z. Bujanovic: On the failure of rank-revealing QR
*> factorization software - a case study.
*> ACM Trans. Math. Softw. Vol. 35, No 2 (2008), pp. 1-28.
*> LAPACK Working note 176.
*> [4] Z. Drmac: SIGMA - mathematical software library for accurate SVD, PSV,
*> QSVD, (H,K)-SVD computations.
*> Department of Mathematics, University of Zagreb, 2008.
*> \endverbatim
*
*> \par Bugs, examples and comments:
* =================================
*>
*> Please report all bugs and send interesting examples and/or comments to
*> [email protected]. Thank you.
*>
* =====================================================================
SUBROUTINE DGEJSV( JOBA, JOBU, JOBV, JOBR, JOBT, JOBP,
$ M, N, A, LDA, SVA, U, LDU, V, LDV,
$ WORK, LWORK, IWORK, INFO )
*
* -- LAPACK computational routine --
* -- LAPACK is a software package provided by Univ. of Tennessee, --
* -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
*
* .. Scalar Arguments ..
IMPLICIT NONE
INTEGER INFO, LDA, LDU, LDV, LWORK, M, N
* ..
* .. Array Arguments ..
DOUBLE PRECISION A( LDA, * ), SVA( N ), U( LDU, * ), V( LDV, * ),
$ WORK( LWORK )
INTEGER IWORK( * )
CHARACTER*1 JOBA, JOBP, JOBR, JOBT, JOBU, JOBV
* ..
*
* ===========================================================================
*
* .. Local Parameters ..
DOUBLE PRECISION ZERO, ONE
PARAMETER ( ZERO = 0.0D0, ONE = 1.0D0 )
* ..
* .. Local Scalars ..
DOUBLE PRECISION AAPP, AAQQ, AATMAX, AATMIN, BIG, BIG1, COND_OK,
$ CONDR1, CONDR2, ENTRA, ENTRAT, EPSLN, MAXPRJ, SCALEM,
$ SCONDA, SFMIN, SMALL, TEMP1, USCAL1, USCAL2, XSC
INTEGER IERR, N1, NR, NUMRANK, p, q, WARNING
LOGICAL ALMORT, DEFR, ERREST, GOSCAL, JRACC, KILL, LSVEC,
$ L2ABER, L2KILL, L2PERT, L2RANK, L2TRAN,
$ NOSCAL, ROWPIV, RSVEC, TRANSP
* ..
* .. Intrinsic Functions ..
INTRINSIC DABS, DLOG, MAX, MIN, DBLE, IDNINT, DSIGN, DSQRT
* ..
* .. External Functions ..
DOUBLE PRECISION DLAMCH, DNRM2
INTEGER IDAMAX
LOGICAL LSAME
EXTERNAL IDAMAX, LSAME, DLAMCH, DNRM2
* ..
* .. External Subroutines ..
EXTERNAL DCOPY, DGELQF, DGEQP3, DGEQRF, DLACPY, DLASCL,
$ DLASET, DLASSQ, DLASWP, DORGQR, DORMLQ,
$ DORMQR, DPOCON, DSCAL, DSWAP, DTRSM, XERBLA
*
EXTERNAL DGESVJ
* ..
*
* Test the input arguments
*
LSVEC = LSAME( JOBU, 'U' ) .OR. LSAME( JOBU, 'F' )
JRACC = LSAME( JOBV, 'J' )
RSVEC = LSAME( JOBV, 'V' ) .OR. JRACC
ROWPIV = LSAME( JOBA, 'F' ) .OR. LSAME( JOBA, 'G' )
L2RANK = LSAME( JOBA, 'R' )
L2ABER = LSAME( JOBA, 'A' )
ERREST = LSAME( JOBA, 'E' ) .OR. LSAME( JOBA, 'G' )
L2TRAN = LSAME( JOBT, 'T' )
L2KILL = LSAME( JOBR, 'R' )
DEFR = LSAME( JOBR, 'N' )
L2PERT = LSAME( JOBP, 'P' )
*
IF ( .NOT.(ROWPIV .OR. L2RANK .OR. L2ABER .OR.
$ ERREST .OR. LSAME( JOBA, 'C' ) )) THEN
INFO = - 1
ELSE IF ( .NOT.( LSVEC .OR. LSAME( JOBU, 'N' ) .OR.
$ LSAME( JOBU, 'W' )) ) THEN
INFO = - 2
ELSE IF ( .NOT.( RSVEC .OR. LSAME( JOBV, 'N' ) .OR.
$ LSAME( JOBV, 'W' )) .OR. ( JRACC .AND. (.NOT.LSVEC) ) ) THEN
INFO = - 3
ELSE IF ( .NOT. ( L2KILL .OR. DEFR ) ) THEN
INFO = - 4
ELSE IF ( .NOT. ( L2TRAN .OR. LSAME( JOBT, 'N' ) ) ) THEN
INFO = - 5
ELSE IF ( .NOT. ( L2PERT .OR. LSAME( JOBP, 'N' ) ) ) THEN
INFO = - 6
ELSE IF ( M .LT. 0 ) THEN
INFO = - 7
ELSE IF ( ( N .LT. 0 ) .OR. ( N .GT. M ) ) THEN
INFO = - 8
ELSE IF ( LDA .LT. M ) THEN
INFO = - 10
ELSE IF ( LSVEC .AND. ( LDU .LT. M ) ) THEN
INFO = - 13
ELSE IF ( RSVEC .AND. ( LDV .LT. N ) ) THEN
INFO = - 15
ELSE IF ( (.NOT.(LSVEC .OR. RSVEC .OR. ERREST).AND.
& (LWORK .LT. MAX(7,4*N+1,2*M+N))) .OR.
& (.NOT.(LSVEC .OR. RSVEC) .AND. ERREST .AND.
& (LWORK .LT. MAX(7,4*N+N*N,2*M+N))) .OR.
& (LSVEC .AND. (.NOT.RSVEC) .AND. (LWORK .LT. MAX(7,2*M+N,4*N+1)))
& .OR.
& (RSVEC .AND. (.NOT.LSVEC) .AND. (LWORK .LT. MAX(7,2*M+N,4*N+1)))
& .OR.
& (LSVEC .AND. RSVEC .AND. (.NOT.JRACC) .AND.
& (LWORK.LT.MAX(2*M+N,6*N+2*N*N)))
& .OR. (LSVEC .AND. RSVEC .AND. JRACC .AND.
& LWORK.LT.MAX(2*M+N,4*N+N*N,2*N+N*N+6)))
& THEN
INFO = - 17
ELSE
* #:)
INFO = 0
END IF
*
IF ( INFO .NE. 0 ) THEN
* #:(
CALL XERBLA( 'DGEJSV', - INFO )
RETURN
END IF
*
* Quick return for void matrix (Y3K safe)
* #:)
IF ( ( M .EQ. 0 ) .OR. ( N .EQ. 0 ) ) THEN
IWORK(1:3) = 0
WORK(1:7) = 0
RETURN
ENDIF
*
* Determine whether the matrix U should be M x N or M x M
*
IF ( LSVEC ) THEN
N1 = N
IF ( LSAME( JOBU, 'F' ) ) N1 = M
END IF
*
* Set numerical parameters
*
*! NOTE: Make sure DLAMCH() does not fail on the target architecture.
*
EPSLN = DLAMCH('Epsilon')
SFMIN = DLAMCH('SafeMinimum')
SMALL = SFMIN / EPSLN
BIG = DLAMCH('O')
* BIG = ONE / SFMIN
*
* Initialize SVA(1:N) = diag( ||A e_i||_2 )_1^N
*
*(!) If necessary, scale SVA() to protect the largest norm from
* overflow. It is possible that this scaling pushes the smallest
* column norm left from the underflow threshold (extreme case).
*
SCALEM = ONE / DSQRT(DBLE(M)*DBLE(N))
NOSCAL = .TRUE.
GOSCAL = .TRUE.
DO 1874 p = 1, N
AAPP = ZERO
AAQQ = ONE
CALL DLASSQ( M, A(1,p), 1, AAPP, AAQQ )
IF ( AAPP .GT. BIG ) THEN
INFO = - 9
CALL XERBLA( 'DGEJSV', -INFO )
RETURN
END IF
AAQQ = DSQRT(AAQQ)
IF ( ( AAPP .LT. (BIG / AAQQ) ) .AND. NOSCAL ) THEN
SVA(p) = AAPP * AAQQ
ELSE
NOSCAL = .FALSE.
SVA(p) = AAPP * ( AAQQ * SCALEM )
IF ( GOSCAL ) THEN
GOSCAL = .FALSE.
CALL DSCAL( p-1, SCALEM, SVA, 1 )
END IF
END IF
1874 CONTINUE
*
IF ( NOSCAL ) SCALEM = ONE
*
AAPP = ZERO
AAQQ = BIG
DO 4781 p = 1, N
AAPP = MAX( AAPP, SVA(p) )
IF ( SVA(p) .NE. ZERO ) AAQQ = MIN( AAQQ, SVA(p) )
4781 CONTINUE
*
* Quick return for zero M x N matrix
* #:)
IF ( AAPP .EQ. ZERO ) THEN
IF ( LSVEC ) CALL DLASET( 'G', M, N1, ZERO, ONE, U, LDU )
IF ( RSVEC ) CALL DLASET( 'G', N, N, ZERO, ONE, V, LDV )
WORK(1) = ONE
WORK(2) = ONE
IF ( ERREST ) WORK(3) = ONE
IF ( LSVEC .AND. RSVEC ) THEN
WORK(4) = ONE
WORK(5) = ONE
END IF
IF ( L2TRAN ) THEN
WORK(6) = ZERO
WORK(7) = ZERO
END IF
IWORK(1) = 0
IWORK(2) = 0
IWORK(3) = 0
RETURN
END IF
*
* Issue warning if denormalized column norms detected. Override the
* high relative accuracy request. Issue licence to kill columns
* (set them to zero) whose norm is less than sigma_max / BIG (roughly).
* #:(
WARNING = 0
IF ( AAQQ .LE. SFMIN ) THEN
L2RANK = .TRUE.
L2KILL = .TRUE.
WARNING = 1
END IF
*
* Quick return for one-column matrix
* #:)
IF ( N .EQ. 1 ) THEN
*
IF ( LSVEC ) THEN
CALL DLASCL( 'G',0,0,SVA(1),SCALEM, M,1,A(1,1),LDA,IERR )
CALL DLACPY( 'A', M, 1, A, LDA, U, LDU )
* computing all M left singular vectors of the M x 1 matrix
IF ( N1 .NE. N ) THEN
CALL DGEQRF( M, N, U,LDU, WORK, WORK(N+1),LWORK-N,IERR )
CALL DORGQR( M,N1,1, U,LDU,WORK,WORK(N+1),LWORK-N,IERR )
CALL DCOPY( M, A(1,1), 1, U(1,1), 1 )
END IF
END IF
IF ( RSVEC ) THEN
V(1,1) = ONE
END IF
IF ( SVA(1) .LT. (BIG*SCALEM) ) THEN
SVA(1) = SVA(1) / SCALEM
SCALEM = ONE
END IF
WORK(1) = ONE / SCALEM
WORK(2) = ONE
IF ( SVA(1) .NE. ZERO ) THEN
IWORK(1) = 1
IF ( ( SVA(1) / SCALEM) .GE. SFMIN ) THEN
IWORK(2) = 1
ELSE
IWORK(2) = 0
END IF
ELSE
IWORK(1) = 0
IWORK(2) = 0
END IF
IWORK(3) = 0
IF ( ERREST ) WORK(3) = ONE
IF ( LSVEC .AND. RSVEC ) THEN
WORK(4) = ONE
WORK(5) = ONE
END IF
IF ( L2TRAN ) THEN
WORK(6) = ZERO
WORK(7) = ZERO
END IF
RETURN
*
END IF
*
TRANSP = .FALSE.
L2TRAN = L2TRAN .AND. ( M .EQ. N )
*
AATMAX = -ONE
AATMIN = BIG
IF ( ROWPIV .OR. L2TRAN ) THEN
*
* Compute the row norms, needed to determine row pivoting sequence
* (in the case of heavily row weighted A, row pivoting is strongly
* advised) and to collect information needed to compare the
* structures of A * A^t and A^t * A (in the case L2TRAN.EQ..TRUE.).
*
IF ( L2TRAN ) THEN
DO 1950 p = 1, M
XSC = ZERO
TEMP1 = ONE
CALL DLASSQ( N, A(p,1), LDA, XSC, TEMP1 )
* DLASSQ gets both the ell_2 and the ell_infinity norm
* in one pass through the vector
WORK(M+N+p) = XSC * SCALEM
WORK(N+p) = XSC * (SCALEM*DSQRT(TEMP1))
AATMAX = MAX( AATMAX, WORK(N+p) )
IF (WORK(N+p) .NE. ZERO) AATMIN = MIN(AATMIN,WORK(N+p))
1950 CONTINUE
ELSE
DO 1904 p = 1, M
WORK(M+N+p) = SCALEM*DABS( A(p,IDAMAX(N,A(p,1),LDA)) )
AATMAX = MAX( AATMAX, WORK(M+N+p) )
AATMIN = MIN( AATMIN, WORK(M+N+p) )
1904 CONTINUE
END IF
*
END IF
*
* For square matrix A try to determine whether A^t would be better
* input for the preconditioned Jacobi SVD, with faster convergence.
* The decision is based on an O(N) function of the vector of column
* and row norms of A, based on the Shannon entropy. This should give
* the right choice in most cases when the difference actually matters.
* It may fail and pick the slower converging side.
*
ENTRA = ZERO
ENTRAT = ZERO
IF ( L2TRAN ) THEN
*
XSC = ZERO
TEMP1 = ONE
CALL DLASSQ( N, SVA, 1, XSC, TEMP1 )
TEMP1 = ONE / TEMP1
*
ENTRA = ZERO
DO 1113 p = 1, N
BIG1 = ( ( SVA(p) / XSC )**2 ) * TEMP1
IF ( BIG1 .NE. ZERO ) ENTRA = ENTRA + BIG1 * DLOG(BIG1)
1113 CONTINUE
ENTRA = - ENTRA / DLOG(DBLE(N))
*
* Now, SVA().^2/Trace(A^t * A) is a point in the probability simplex.
* It is derived from the diagonal of A^t * A. Do the same with the
* diagonal of A * A^t, compute the entropy of the corresponding
* probability distribution. Note that A * A^t and A^t * A have the
* same trace.
*
ENTRAT = ZERO
DO 1114 p = N+1, N+M
BIG1 = ( ( WORK(p) / XSC )**2 ) * TEMP1
IF ( BIG1 .NE. ZERO ) ENTRAT = ENTRAT + BIG1 * DLOG(BIG1)
1114 CONTINUE
ENTRAT = - ENTRAT / DLOG(DBLE(M))
*
* Analyze the entropies and decide A or A^t. Smaller entropy
* usually means better input for the algorithm.
*
TRANSP = ( ENTRAT .LT. ENTRA )
*
* If A^t is better than A, transpose A.
*
IF ( TRANSP ) THEN
* In an optimal implementation, this trivial transpose
* should be replaced with faster transpose.
DO 1115 p = 1, N - 1
DO 1116 q = p + 1, N
TEMP1 = A(q,p)
A(q,p) = A(p,q)
A(p,q) = TEMP1
1116 CONTINUE
1115 CONTINUE
DO 1117 p = 1, N
WORK(M+N+p) = SVA(p)
SVA(p) = WORK(N+p)
1117 CONTINUE
TEMP1 = AAPP
AAPP = AATMAX
AATMAX = TEMP1
TEMP1 = AAQQ
AAQQ = AATMIN
AATMIN = TEMP1
KILL = LSVEC
LSVEC = RSVEC
RSVEC = KILL
IF ( LSVEC ) N1 = N
*
ROWPIV = .TRUE.
END IF
*
END IF
* END IF L2TRAN
*
* Scale the matrix so that its maximal singular value remains less
* than DSQRT(BIG) -- the matrix is scaled so that its maximal column
* has Euclidean norm equal to DSQRT(BIG/N). The only reason to keep
* DSQRT(BIG) instead of BIG is the fact that DGEJSV uses LAPACK and
* BLAS routines that, in some implementations, are not capable of
* working in the full interval [SFMIN,BIG] and that they may provoke
* overflows in the intermediate results. If the singular values spread
* from SFMIN to BIG, then DGESVJ will compute them. So, in that case,
* one should use DGESVJ instead of DGEJSV.
*
BIG1 = DSQRT( BIG )
TEMP1 = DSQRT( BIG / DBLE(N) )
*
CALL DLASCL( 'G', 0, 0, AAPP, TEMP1, N, 1, SVA, N, IERR )
IF ( AAQQ .GT. (AAPP * SFMIN) ) THEN
AAQQ = ( AAQQ / AAPP ) * TEMP1
ELSE
AAQQ = ( AAQQ * TEMP1 ) / AAPP
END IF
TEMP1 = TEMP1 * SCALEM
CALL DLASCL( 'G', 0, 0, AAPP, TEMP1, M, N, A, LDA, IERR )
*
* To undo scaling at the end of this procedure, multiply the
* computed singular values with USCAL2 / USCAL1.
*
USCAL1 = TEMP1
USCAL2 = AAPP
*
IF ( L2KILL ) THEN
* L2KILL enforces computation of nonzero singular values in
* the restricted range of condition number of the initial A,
* sigma_max(A) / sigma_min(A) approx. DSQRT(BIG)/DSQRT(SFMIN).
XSC = DSQRT( SFMIN )
ELSE
XSC = SMALL
*
* Now, if the condition number of A is too big,
* sigma_max(A) / sigma_min(A) .GT. DSQRT(BIG/N) * EPSLN / SFMIN,
* as a precaution measure, the full SVD is computed using DGESVJ
* with accumulated Jacobi rotations. This provides numerically
* more robust computation, at the cost of slightly increased run
* time. Depending on the concrete implementation of BLAS and LAPACK
* (i.e. how they behave in presence of extreme ill-conditioning) the
* implementor may decide to remove this switch.
IF ( ( AAQQ.LT.DSQRT(SFMIN) ) .AND. LSVEC .AND. RSVEC ) THEN
JRACC = .TRUE.
END IF
*
END IF
IF ( AAQQ .LT. XSC ) THEN
DO 700 p = 1, N
IF ( SVA(p) .LT. XSC ) THEN
CALL DLASET( 'A', M, 1, ZERO, ZERO, A(1,p), LDA )
SVA(p) = ZERO
END IF
700 CONTINUE
END IF
*
* Preconditioning using QR factorization with pivoting
*
IF ( ROWPIV ) THEN
* Optional row permutation (Bjoerck row pivoting):
* A result by Cox and Higham shows that the Bjoerck's
* row pivoting combined with standard column pivoting
* has similar effect as Powell-Reid complete pivoting.
* The ell-infinity norms of A are made nonincreasing.
DO 1952 p = 1, M - 1
q = IDAMAX( M-p+1, WORK(M+N+p), 1 ) + p - 1
IWORK(2*N+p) = q
IF ( p .NE. q ) THEN
TEMP1 = WORK(M+N+p)
WORK(M+N+p) = WORK(M+N+q)
WORK(M+N+q) = TEMP1
END IF
1952 CONTINUE
CALL DLASWP( N, A, LDA, 1, M-1, IWORK(2*N+1), 1 )
END IF
*
* End of the preparation phase (scaling, optional sorting and
* transposing, optional flushing of small columns).
*
* Preconditioning
*
* If the full SVD is needed, the right singular vectors are computed
* from a matrix equation, and for that we need theoretical analysis
* of the Businger-Golub pivoting. So we use DGEQP3 as the first RR QRF.
* In all other cases the first RR QRF can be chosen by other criteria
* (eg speed by replacing global with restricted window pivoting, such
* as in SGEQPX from TOMS # 782). Good results will be obtained using
* SGEQPX with properly (!) chosen numerical parameters.
* Any improvement of DGEQP3 improves overall performance of DGEJSV.
*
* A * P1 = Q1 * [ R1^t 0]^t:
DO 1963 p = 1, N
* .. all columns are free columns
IWORK(p) = 0
1963 CONTINUE
CALL DGEQP3( M,N,A,LDA, IWORK,WORK, WORK(N+1),LWORK-N, IERR )
*
* The upper triangular matrix R1 from the first QRF is inspected for
* rank deficiency and possibilities for deflation, or possible
* ill-conditioning. Depending on the user specified flag L2RANK,
* the procedure explores possibilities to reduce the numerical
* rank by inspecting the computed upper triangular factor. If
* L2RANK or L2ABER are up, then DGEJSV will compute the SVD of
* A + dA, where ||dA|| <= f(M,N)*EPSLN.
*
NR = 1
IF ( L2ABER ) THEN
* Standard absolute error bound suffices. All sigma_i with
* sigma_i < N*EPSLN*||A|| are flushed to zero. This is an
* aggressive enforcement of lower numerical rank by introducing a
* backward error of the order of N*EPSLN*||A||.
TEMP1 = DSQRT(DBLE(N))*EPSLN
DO 3001 p = 2, N
IF ( DABS(A(p,p)) .GE. (TEMP1*DABS(A(1,1))) ) THEN
NR = NR + 1
ELSE
GO TO 3002
END IF
3001 CONTINUE
3002 CONTINUE
ELSE IF ( L2RANK ) THEN
* .. similarly as above, only slightly more gentle (less aggressive).
* Sudden drop on the diagonal of R1 is used as the criterion for
* close-to-rank-deficient.
TEMP1 = DSQRT(SFMIN)
DO 3401 p = 2, N
IF ( ( DABS(A(p,p)) .LT. (EPSLN*DABS(A(p-1,p-1))) ) .OR.
$ ( DABS(A(p,p)) .LT. SMALL ) .OR.
$ ( L2KILL .AND. (DABS(A(p,p)) .LT. TEMP1) ) ) GO TO 3402
NR = NR + 1
3401 CONTINUE
3402 CONTINUE
*
ELSE
* The goal is high relative accuracy. However, if the matrix
* has high scaled condition number the relative accuracy is in
* general not feasible. Later on, a condition number estimator
* will be deployed to estimate the scaled condition number.
* Here we just remove the underflowed part of the triangular
* factor. This prevents the situation in which the code is
* working hard to get the accuracy not warranted by the data.
TEMP1 = DSQRT(SFMIN)
DO 3301 p = 2, N
IF ( ( DABS(A(p,p)) .LT. SMALL ) .OR.
$ ( L2KILL .AND. (DABS(A(p,p)) .LT. TEMP1) ) ) GO TO 3302
NR = NR + 1
3301 CONTINUE
3302 CONTINUE
*
END IF
*
ALMORT = .FALSE.
IF ( NR .EQ. N ) THEN
MAXPRJ = ONE
DO 3051 p = 2, N
TEMP1 = DABS(A(p,p)) / SVA(IWORK(p))
MAXPRJ = MIN( MAXPRJ, TEMP1 )