1 *> \brief <b> SSYEVR computes the eigenvalues and, optionally, the left and/or right eigenvectors for SY matrices</b>
3 * =========== DOCUMENTATION ===========
5 * Online html documentation available at
6 * http://www.netlib.org/lapack/explore-html/
9 *> Download SSYEVR + dependencies
10 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/ssyevr.f">
12 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/ssyevr.f">
14 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/ssyevr.f">
21 * SUBROUTINE SSYEVR( JOBZ, RANGE, UPLO, N, A, LDA, VL, VU, IL, IU,
22 * ABSTOL, M, W, Z, LDZ, ISUPPZ, WORK, LWORK,
23 * IWORK, LIWORK, INFO )
25 * .. Scalar Arguments ..
26 * CHARACTER JOBZ, RANGE, UPLO
27 * INTEGER IL, INFO, IU, LDA, LDZ, LIWORK, LWORK, M, N
30 * .. Array Arguments ..
31 * INTEGER ISUPPZ( * ), IWORK( * )
32 * REAL A( LDA, * ), W( * ), WORK( * ), Z( LDZ, * )
41 *> SSYEVR computes selected eigenvalues and, optionally, eigenvectors
42 *> of a real symmetric matrix A. Eigenvalues and eigenvectors can be
43 *> selected by specifying either a range of values or a range of
44 *> indices for the desired eigenvalues.
46 *> SSYEVR first reduces the matrix A to tridiagonal form T with a call
47 *> to SSYTRD. Then, whenever possible, SSYEVR calls SSTEMR to compute
48 *> the eigenspectrum using Relatively Robust Representations. SSTEMR
49 *> computes eigenvalues by the dqds algorithm, while orthogonal
50 *> eigenvectors are computed from various "good" L D L^T representations
51 *> (also known as Relatively Robust Representations). Gram-Schmidt
52 *> orthogonalization is avoided as far as possible. More specifically,
53 *> the various steps of the algorithm are as follows.
55 *> For each unreduced block (submatrix) of T,
56 *> (a) Compute T - sigma I = L D L^T, so that L and D
57 *> define all the wanted eigenvalues to high relative accuracy.
58 *> This means that small relative changes in the entries of D and L
59 *> cause only small relative changes in the eigenvalues and
60 *> eigenvectors. The standard (unfactored) representation of the
61 *> tridiagonal matrix T does not have this property in general.
62 *> (b) Compute the eigenvalues to suitable accuracy.
63 *> If the eigenvectors are desired, the algorithm attains full
64 *> accuracy of the computed eigenvalues only right before
65 *> the corresponding vectors have to be computed, see steps c) and d).
66 *> (c) For each cluster of close eigenvalues, select a new
67 *> shift close to the cluster, find a new factorization, and refine
68 *> the shifted eigenvalues to suitable accuracy.
69 *> (d) For each eigenvalue with a large enough relative separation compute
70 *> the corresponding eigenvector by forming a rank revealing twisted
71 *> factorization. Go back to (c) for any clusters that remain.
73 *> The desired accuracy of the output can be specified by the input
76 *> For more details, see SSTEMR's documentation and:
77 *> - Inderjit S. Dhillon and Beresford N. Parlett: "Multiple representations
78 *> to compute orthogonal eigenvectors of symmetric tridiagonal matrices,"
79 *> Linear Algebra and its Applications, 387(1), pp. 1-28, August 2004.
80 *> - Inderjit Dhillon and Beresford Parlett: "Orthogonal Eigenvectors and
81 *> Relative Gaps," SIAM Journal on Matrix Analysis and Applications, Vol. 25,
82 *> 2004. Also LAPACK Working Note 154.
83 *> - Inderjit Dhillon: "A new O(n^2) algorithm for the symmetric
84 *> tridiagonal eigenvalue/eigenvector problem",
85 *> Computer Science Division Technical Report No. UCB/CSD-97-971,
86 *> UC Berkeley, May 1997.
89 *> Note 1 : SSYEVR calls SSTEMR when the full spectrum is requested
90 *> on machines which conform to the ieee-754 floating point standard.
91 *> SSYEVR calls SSTEBZ and SSTEIN on non-ieee machines and
92 *> when partial spectrum requests are made.
94 *> Normal execution of SSTEMR may create NaNs and infinities and
95 *> hence may abort due to a floating point exception in environments
96 *> which do not handle NaNs and infinities in the ieee standard default
105 *> JOBZ is CHARACTER*1
106 *> = 'N': Compute eigenvalues only;
107 *> = 'V': Compute eigenvalues and eigenvectors.
112 *> RANGE is CHARACTER*1
113 *> = 'A': all eigenvalues will be found.
114 *> = 'V': all eigenvalues in the half-open interval (VL,VU]
116 *> = 'I': the IL-th through IU-th eigenvalues will be found.
117 *> For RANGE = 'V' or 'I' and IU - IL < N - 1, SSTEBZ and
123 *> UPLO is CHARACTER*1
124 *> = 'U': Upper triangle of A is stored;
125 *> = 'L': Lower triangle of A is stored.
131 *> The order of the matrix A. N >= 0.
136 *> A is REAL array, dimension (LDA, N)
137 *> On entry, the symmetric matrix A. If UPLO = 'U', the
138 *> leading N-by-N upper triangular part of A contains the
139 *> upper triangular part of the matrix A. If UPLO = 'L',
140 *> the leading N-by-N lower triangular part of A contains
141 *> the lower triangular part of the matrix A.
142 *> On exit, the lower triangle (if UPLO='L') or the upper
143 *> triangle (if UPLO='U') of A, including the diagonal, is
150 *> The leading dimension of the array A. LDA >= max(1,N).
156 *> If RANGE='V', the lower bound of the interval to
157 *> be searched for eigenvalues. VL < VU.
158 *> Not referenced if RANGE = 'A' or 'I'.
164 *> If RANGE='V', the upper bound of the interval to
165 *> be searched for eigenvalues. VL < VU.
166 *> Not referenced if RANGE = 'A' or 'I'.
172 *> If RANGE='I', the index of the
173 *> smallest eigenvalue to be returned.
174 *> 1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0.
175 *> Not referenced if RANGE = 'A' or 'V'.
181 *> If RANGE='I', the index of the
182 *> largest eigenvalue to be returned.
183 *> 1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0.
184 *> Not referenced if RANGE = 'A' or 'V'.
190 *> The absolute error tolerance for the eigenvalues.
191 *> An approximate eigenvalue is accepted as converged
192 *> when it is determined to lie in an interval [a,b]
193 *> of width less than or equal to
195 *> ABSTOL + EPS * max( |a|,|b| ) ,
197 *> where EPS is the machine precision. If ABSTOL is less than
198 *> or equal to zero, then EPS*|T| will be used in its place,
199 *> where |T| is the 1-norm of the tridiagonal matrix obtained
200 *> by reducing A to tridiagonal form.
202 *> See "Computing Small Singular Values of Bidiagonal Matrices
203 *> with Guaranteed High Relative Accuracy," by Demmel and
204 *> Kahan, LAPACK Working Note #3.
206 *> If high relative accuracy is important, set ABSTOL to
207 *> SLAMCH( 'Safe minimum' ). Doing so will guarantee that
208 *> eigenvalues are computed to high relative accuracy when
209 *> possible in future releases. The current code does not
210 *> make any guarantees about high relative accuracy, but
211 *> future releases will. See J. Barlow and J. Demmel,
212 *> "Computing Accurate Eigensystems of Scaled Diagonally
213 *> Dominant Matrices", LAPACK Working Note #7, for a discussion
214 *> of which matrices define their eigenvalues to high relative
221 *> The total number of eigenvalues found. 0 <= M <= N.
222 *> If RANGE = 'A', M = N, and if RANGE = 'I', M = IU-IL+1.
227 *> W is REAL array, dimension (N)
228 *> The first M elements contain the selected eigenvalues in
234 *> Z is REAL array, dimension (LDZ, max(1,M))
235 *> If JOBZ = 'V', then if INFO = 0, the first M columns of Z
236 *> contain the orthonormal eigenvectors of the matrix A
237 *> corresponding to the selected eigenvalues, with the i-th
238 *> column of Z holding the eigenvector associated with W(i).
239 *> If JOBZ = 'N', then Z is not referenced.
240 *> Note: the user must ensure that at least max(1,M) columns are
241 *> supplied in the array Z; if RANGE = 'V', the exact value of M
242 *> is not known in advance and an upper bound must be used.
243 *> Supplying N columns is always safe.
249 *> The leading dimension of the array Z. LDZ >= 1, and if
250 *> JOBZ = 'V', LDZ >= max(1,N).
253 *> \param[out] ISUPPZ
255 *> ISUPPZ is INTEGER array, dimension ( 2*max(1,M) )
256 *> The support of the eigenvectors in Z, i.e., the indices
257 *> indicating the nonzero elements in Z. The i-th eigenvector
258 *> is nonzero only in elements ISUPPZ( 2*i-1 ) through
259 *> ISUPPZ( 2*i ). This is an output of SSTEMR (tridiagonal
260 *> matrix). The support of the eigenvectors of A is typically
261 *> 1:N because of the orthogonal transformations applied by SORMTR.
262 *> Implemented only for RANGE = 'A' or 'I' and IU - IL = N - 1
267 *> WORK is REAL array, dimension (MAX(1,LWORK))
268 *> On exit, if INFO = 0, WORK(1) returns the optimal LWORK.
274 *> The dimension of the array WORK. LWORK >= max(1,26*N).
275 *> For optimal efficiency, LWORK >= (NB+6)*N,
276 *> where NB is the max of the blocksize for SSYTRD and SORMTR
277 *> returned by ILAENV.
279 *> If LWORK = -1, then a workspace query is assumed; the routine
280 *> only calculates the optimal sizes of the WORK and IWORK
281 *> arrays, returns these values as the first entries of the WORK
282 *> and IWORK arrays, and no error message related to LWORK or
283 *> LIWORK is issued by XERBLA.
288 *> IWORK is INTEGER array, dimension (MAX(1,LIWORK))
289 *> On exit, if INFO = 0, IWORK(1) returns the optimal LWORK.
295 *> The dimension of the array IWORK. LIWORK >= max(1,10*N).
297 *> If LIWORK = -1, then a workspace query is assumed; the
298 *> routine only calculates the optimal sizes of the WORK and
299 *> IWORK arrays, returns these values as the first entries of
300 *> the WORK and IWORK arrays, and no error message related to
301 *> LWORK or LIWORK is issued by XERBLA.
307 *> = 0: successful exit
308 *> < 0: if INFO = -i, the i-th argument had an illegal value
309 *> > 0: Internal error
315 *> \author Univ. of Tennessee
316 *> \author Univ. of California Berkeley
317 *> \author Univ. of Colorado Denver
322 *> \ingroup realSYeigen
324 *> \par Contributors:
327 *> Inderjit Dhillon, IBM Almaden, USA \n
328 *> Osni Marques, LBNL/NERSC, USA \n
329 *> Ken Stanley, Computer Science Division, University of
330 *> California at Berkeley, USA \n
331 *> Jason Riedy, Computer Science Division, University of
332 *> California at Berkeley, USA \n
334 * =====================================================================
335 SUBROUTINE SSYEVR( JOBZ, RANGE, UPLO, N, A, LDA, VL, VU, IL, IU,
336 $ ABSTOL, M, W, Z, LDZ, ISUPPZ, WORK, LWORK,
337 $ IWORK, LIWORK, INFO )
339 * -- LAPACK driver routine (version 3.6.1) --
340 * -- LAPACK is a software package provided by Univ. of Tennessee, --
341 * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
344 * .. Scalar Arguments ..
345 CHARACTER JOBZ, RANGE, UPLO
346 INTEGER IL, INFO, IU, LDA, LDZ, LIWORK, LWORK, M, N
349 * .. Array Arguments ..
350 INTEGER ISUPPZ( * ), IWORK( * )
351 REAL A( LDA, * ), W( * ), WORK( * ), Z( LDZ, * )
354 * =====================================================================
358 PARAMETER ( ZERO = 0.0E+0, ONE = 1.0E+0, TWO = 2.0E+0 )
360 * .. Local Scalars ..
361 LOGICAL ALLEIG, INDEIG, LOWER, LQUERY, TEST, VALEIG,
364 INTEGER I, IEEEOK, IINFO, IMAX, INDD, INDDD, INDE,
365 $ INDEE, INDIBL, INDIFL, INDISP, INDIWO, INDTAU,
366 $ INDWK, INDWKN, ISCALE, J, JJ, LIWMIN,
367 $ LLWORK, LLWRKN, LWKOPT, LWMIN, NB, NSPLIT
368 REAL ABSTLL, ANRM, BIGNUM, EPS, RMAX, RMIN, SAFMIN,
369 $ SIGMA, SMLNUM, TMP1, VLL, VUU
371 * .. External Functions ..
375 EXTERNAL LSAME, ILAENV, SLAMCH, SLANSY
377 * .. External Subroutines ..
378 EXTERNAL SCOPY, SORMTR, SSCAL, SSTEBZ, SSTEMR, SSTEIN,
379 $ SSTERF, SSWAP, SSYTRD, XERBLA
381 * .. Intrinsic Functions ..
382 INTRINSIC MAX, MIN, SQRT
384 * .. Executable Statements ..
386 * Test the input parameters.
388 IEEEOK = ILAENV( 10, 'SSYEVR', 'N', 1, 2, 3, 4 )
390 LOWER = LSAME( UPLO, 'L' )
391 WANTZ = LSAME( JOBZ, 'V' )
392 ALLEIG = LSAME( RANGE, 'A' )
393 VALEIG = LSAME( RANGE, 'V' )
394 INDEIG = LSAME( RANGE, 'I' )
396 LQUERY = ( ( LWORK.EQ.-1 ) .OR. ( LIWORK.EQ.-1 ) )
398 LWMIN = MAX( 1, 26*N )
399 LIWMIN = MAX( 1, 10*N )
402 IF( .NOT.( WANTZ .OR. LSAME( JOBZ, 'N' ) ) ) THEN
404 ELSE IF( .NOT.( ALLEIG .OR. VALEIG .OR. INDEIG ) ) THEN
406 ELSE IF( .NOT.( LOWER .OR. LSAME( UPLO, 'U' ) ) ) THEN
408 ELSE IF( N.LT.0 ) THEN
410 ELSE IF( LDA.LT.MAX( 1, N ) ) THEN
414 IF( N.GT.0 .AND. VU.LE.VL )
416 ELSE IF( INDEIG ) THEN
417 IF( IL.LT.1 .OR. IL.GT.MAX( 1, N ) ) THEN
419 ELSE IF( IU.LT.MIN( N, IL ) .OR. IU.GT.N ) THEN
425 IF( LDZ.LT.1 .OR. ( WANTZ .AND. LDZ.LT.N ) ) THEN
431 NB = ILAENV( 1, 'SSYTRD', UPLO, N, -1, -1, -1 )
432 NB = MAX( NB, ILAENV( 1, 'SORMTR', UPLO, N, -1, -1, -1 ) )
433 LWKOPT = MAX( ( NB+1 )*N, LWMIN )
437 IF( LWORK.LT.LWMIN .AND. .NOT.LQUERY ) THEN
439 ELSE IF( LIWORK.LT.LIWMIN .AND. .NOT.LQUERY ) THEN
445 CALL XERBLA( 'SSYEVR', -INFO )
447 ELSE IF( LQUERY ) THEN
451 * Quick return if possible
461 IF( ALLEIG .OR. INDEIG ) THEN
465 IF( VL.LT.A( 1, 1 ) .AND. VU.GE.A( 1, 1 ) ) THEN
478 * Get machine constants.
480 SAFMIN = SLAMCH( 'Safe minimum' )
481 EPS = SLAMCH( 'Precision' )
482 SMLNUM = SAFMIN / EPS
483 BIGNUM = ONE / SMLNUM
484 RMIN = SQRT( SMLNUM )
485 RMAX = MIN( SQRT( BIGNUM ), ONE / SQRT( SQRT( SAFMIN ) ) )
487 * Scale matrix to allowable range, if necessary.
495 ANRM = SLANSY( 'M', UPLO, N, A, LDA, WORK )
496 IF( ANRM.GT.ZERO .AND. ANRM.LT.RMIN ) THEN
499 ELSE IF( ANRM.GT.RMAX ) THEN
503 IF( ISCALE.EQ.1 ) THEN
506 CALL SSCAL( N-J+1, SIGMA, A( J, J ), 1 )
510 CALL SSCAL( J, SIGMA, A( 1, J ), 1 )
514 $ ABSTLL = ABSTOL*SIGMA
521 * Initialize indices into workspaces. Note: The IWORK indices are
522 * used only if SSTERF or SSTEMR fail.
524 * WORK(INDTAU:INDTAU+N-1) stores the scalar factors of the
525 * elementary reflectors used in SSYTRD.
527 * WORK(INDD:INDD+N-1) stores the tridiagonal's diagonal entries.
529 * WORK(INDE:INDE+N-1) stores the off-diagonal entries of the
530 * tridiagonal matrix from SSYTRD.
532 * WORK(INDDD:INDDD+N-1) is a copy of the diagonal entries over
533 * -written by SSTEMR (the SSTERF path copies the diagonal to W).
535 * WORK(INDEE:INDEE+N-1) is a copy of the off-diagonal entries over
536 * -written while computing the eigenvalues in SSTERF and SSTEMR.
538 * INDWK is the starting offset of the left-over workspace, and
539 * LLWORK is the remaining workspace size.
541 LLWORK = LWORK - INDWK + 1
543 * IWORK(INDIBL:INDIBL+M-1) corresponds to IBLOCK in SSTEBZ and
544 * stores the block indices of each of the M<=N eigenvalues.
546 * IWORK(INDISP:INDISP+NSPLIT-1) corresponds to ISPLIT in SSTEBZ and
547 * stores the starting and finishing indices of each block.
549 * IWORK(INDIFL:INDIFL+N-1) stores the indices of eigenvectors
550 * that corresponding to eigenvectors that fail to converge in
551 * SSTEIN. This information is discarded; if any fail, the driver
554 * INDIWO is the offset of the remaining integer workspace.
558 * Call SSYTRD to reduce symmetric matrix to tridiagonal form.
560 CALL SSYTRD( UPLO, N, A, LDA, WORK( INDD ), WORK( INDE ),
561 $ WORK( INDTAU ), WORK( INDWK ), LLWORK, IINFO )
563 * If all eigenvalues are desired
564 * then call SSTERF or SSTEMR and SORMTR.
568 IF( IL.EQ.1 .AND. IU.EQ.N ) THEN
572 IF( ( ALLEIG.OR.TEST ) .AND. ( IEEEOK.EQ.1 ) ) THEN
573 IF( .NOT.WANTZ ) THEN
574 CALL SCOPY( N, WORK( INDD ), 1, W, 1 )
575 CALL SCOPY( N-1, WORK( INDE ), 1, WORK( INDEE ), 1 )
576 CALL SSTERF( N, W, WORK( INDEE ), INFO )
578 CALL SCOPY( N-1, WORK( INDE ), 1, WORK( INDEE ), 1 )
579 CALL SCOPY( N, WORK( INDD ), 1, WORK( INDDD ), 1 )
581 IF (ABSTOL .LE. TWO*N*EPS) THEN
586 CALL SSTEMR( JOBZ, 'A', N, WORK( INDDD ), WORK( INDEE ),
587 $ VL, VU, IL, IU, M, W, Z, LDZ, N, ISUPPZ,
588 $ TRYRAC, WORK( INDWK ), LWORK, IWORK, LIWORK,
593 * Apply orthogonal matrix used in reduction to tridiagonal
594 * form to eigenvectors returned by SSTEMR.
596 IF( WANTZ .AND. INFO.EQ.0 ) THEN
598 LLWRKN = LWORK - INDWKN + 1
599 CALL SORMTR( 'L', UPLO, 'N', N, M, A, LDA,
600 $ WORK( INDTAU ), Z, LDZ, WORK( INDWKN ),
607 * Everything worked. Skip SSTEBZ/SSTEIN. IWORK(:) are
615 * Otherwise, call SSTEBZ and, if eigenvectors are desired, SSTEIN.
616 * Also call SSTEBZ and SSTEIN if SSTEMR fails.
624 CALL SSTEBZ( RANGE, ORDER, N, VLL, VUU, IL, IU, ABSTLL,
625 $ WORK( INDD ), WORK( INDE ), M, NSPLIT, W,
626 $ IWORK( INDIBL ), IWORK( INDISP ), WORK( INDWK ),
627 $ IWORK( INDIWO ), INFO )
630 CALL SSTEIN( N, WORK( INDD ), WORK( INDE ), M, W,
631 $ IWORK( INDIBL ), IWORK( INDISP ), Z, LDZ,
632 $ WORK( INDWK ), IWORK( INDIWO ), IWORK( INDIFL ),
635 * Apply orthogonal matrix used in reduction to tridiagonal
636 * form to eigenvectors returned by SSTEIN.
639 LLWRKN = LWORK - INDWKN + 1
640 CALL SORMTR( 'L', UPLO, 'N', N, M, A, LDA, WORK( INDTAU ), Z,
641 $ LDZ, WORK( INDWKN ), LLWRKN, IINFO )
644 * If matrix was scaled, then rescale eigenvalues appropriately.
646 * Jump here if SSTEMR/SSTEIN succeeded.
648 IF( ISCALE.EQ.1 ) THEN
654 CALL SSCAL( IMAX, ONE / SIGMA, W, 1 )
657 * If eigenvalues are not in order, then sort them, along with
658 * eigenvectors. Note: We do not sort the IFAIL portion of IWORK.
659 * It may not be initialized (if SSTEMR/SSTEIN succeeded), and we do
660 * not return this detailed information to the user.
667 IF( W( JJ ).LT.TMP1 ) THEN
676 CALL SSWAP( N, Z( 1, I ), 1, Z( 1, J ), 1 )
681 * Set WORK(1) to optimal workspace size.