1       SUBROUTINE DLARRV( N, VL, VU, D, L, PIVMIN,
  2      $                   ISPLIT, M, DOL, DOU, MINRGP,
  3      $                   RTOL1, RTOL2, W, WERR, WGAP,
  4      $                   IBLOCK, INDEXW, GERS, Z, LDZ, ISUPPZ,
  5      $                   WORK, IWORK, INFO )
  6 *
  7 *  -- LAPACK auxiliary routine (version 3.3.1) --
  8 *  -- LAPACK is a software package provided by Univ. of Tennessee,    --
  9 *  -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
 10 *  -- April 2011                                                      --
 11 *
 12 *     .. Scalar Arguments ..
 13       INTEGER            DOL, DOU, INFO, LDZ, M, N
 14       DOUBLE PRECISION   MINRGP, PIVMIN, RTOL1, RTOL2, VL, VU
 15 *     ..
 16 *     .. Array Arguments ..
 17       INTEGER            IBLOCK( * ), INDEXW( * ), ISPLIT( * ),
 18      $                   ISUPPZ( * ), IWORK( * )
 19       DOUBLE PRECISION   D( * ), GERS( * ), L( * ), W( * ), WERR( * ),
 20      $                   WGAP( * ), WORK( * )
 21       DOUBLE PRECISION  Z( LDZ, * )
 22 *     ..
 23 *
 24 *  Purpose
 25 *  =======
 26 *
 27 *  DLARRV computes the eigenvectors of the tridiagonal matrix
 28 *  T = L D L**T given L, D and APPROXIMATIONS to the eigenvalues of L D L**T.
 29 *  The input eigenvalues should have been computed by DLARRE.
 30 *
 31 *  Arguments
 32 *  =========
 33 *
 34 *  N       (input) INTEGER
 35 *          The order of the matrix.  N >= 0.
 36 *
 37 *  VL      (input) DOUBLE PRECISION
 38 *  VU      (input) DOUBLE PRECISION
 39 *          Lower and upper bounds of the interval that contains the desired
 40 *          eigenvalues. VL < VU. Needed to compute gaps on the left or right
 41 *          end of the extremal eigenvalues in the desired RANGE.
 42 *
 43 *  D       (input/output) DOUBLE PRECISION array, dimension (N)
 44 *          On entry, the N diagonal elements of the diagonal matrix D.
 45 *          On exit, D may be overwritten.
 46 *
 47 *  L       (input/output) DOUBLE PRECISION array, dimension (N)
 48 *          On entry, the (N-1) subdiagonal elements of the unit
 49 *          bidiagonal matrix L are in elements 1 to N-1 of L
 50 *          (if the matrix is not splitted.) At the end of each block
 51 *          is stored the corresponding shift as given by DLARRE.
 52 *          On exit, L is overwritten.
 53 *
 54 *  PIVMIN  (input) DOUBLE PRECISION
 55 *          The minimum pivot allowed in the Sturm sequence.
 56 *
 57 *  ISPLIT  (input) INTEGER array, dimension (N)
 58 *          The splitting points, at which T breaks up into blocks.
 59 *          The first block consists of rows/columns 1 to
 60 *          ISPLIT( 1 ), the second of rows/columns ISPLIT( 1 )+1
 61 *          through ISPLIT( 2 ), etc.
 62 *
 63 *  M       (input) INTEGER
 64 *          The total number of input eigenvalues.  0 <= M <= N.
 65 *
 66 *  DOL     (input) INTEGER
 67 *  DOU     (input) INTEGER
 68 *          If the user wants to compute only selected eigenvectors from all
 69 *          the eigenvalues supplied, he can specify an index range DOL:DOU.
 70 *          Or else the setting DOL=1, DOU=M should be applied.
 71 *          Note that DOL and DOU refer to the order in which the eigenvalues
 72 *          are stored in W.
 73 *          If the user wants to compute only selected eigenpairs, then
 74 *          the columns DOL-1 to DOU+1 of the eigenvector space Z contain the
 75 *          computed eigenvectors. All other columns of Z are set to zero.
 76 *
 77 *  MINRGP  (input) DOUBLE PRECISION
 78 *
 79 *  RTOL1   (input) DOUBLE PRECISION
 80 *  RTOL2   (input) DOUBLE PRECISION
 81 *           Parameters for bisection.
 82 *           An interval [LEFT,RIGHT] has converged if
 83 *           RIGHT-LEFT.LT.MAX( RTOL1*GAP, RTOL2*MAX(|LEFT|,|RIGHT|) )
 84 *
 85 *  W       (input/output) DOUBLE PRECISION array, dimension (N)
 86 *          The first M elements of W contain the APPROXIMATE eigenvalues for
 87 *          which eigenvectors are to be computed.  The eigenvalues
 88 *          should be grouped by split-off block and ordered from
 89 *          smallest to largest within the block ( The output array
 90 *          W from DLARRE is expected here ). Furthermore, they are with
 91 *          respect to the shift of the corresponding root representation
 92 *          for their block. On exit, W holds the eigenvalues of the
 93 *          UNshifted matrix.
 94 *
 95 *  WERR    (input/output) DOUBLE PRECISION array, dimension (N)
 96 *          The first M elements contain the semiwidth of the uncertainty
 97 *          interval of the corresponding eigenvalue in W
 98 *
 99 *  WGAP    (input/output) DOUBLE PRECISION array, dimension (N)
100 *          The separation from the right neighbor eigenvalue in W.
101 *
102 *  IBLOCK  (input) INTEGER array, dimension (N)
103 *          The indices of the blocks (submatrices) associated with the
104 *          corresponding eigenvalues in W; IBLOCK(i)=1 if eigenvalue
105 *          W(i) belongs to the first block from the top, =2 if W(i)
106 *          belongs to the second block, etc.
107 *
108 *  INDEXW  (input) INTEGER array, dimension (N)
109 *          The indices of the eigenvalues within each block (submatrix);
110 *          for example, INDEXW(i)= 10 and IBLOCK(i)=2 imply that the
111 *          i-th eigenvalue W(i) is the 10-th eigenvalue in the second block.
112 *
113 *  GERS    (input) DOUBLE PRECISION array, dimension (2*N)
114 *          The N Gerschgorin intervals (the i-th Gerschgorin interval
115 *          is (GERS(2*i-1), GERS(2*i)). The Gerschgorin intervals should
116 *          be computed from the original UNshifted matrix.
117 *
118 *  Z       (output) DOUBLE PRECISION array, dimension (LDZ, max(1,M) )
119 *          If INFO = 0, the first M columns of Z contain the
120 *          orthonormal eigenvectors of the matrix T
121 *          corresponding to the input eigenvalues, with the i-th
122 *          column of Z holding the eigenvector associated with W(i).
123 *          Note: the user must ensure that at least max(1,M) columns are
124 *          supplied in the array Z.
125 *
126 *  LDZ     (input) INTEGER
127 *          The leading dimension of the array Z.  LDZ >= 1, and if
128 *          JOBZ = 'V', LDZ >= max(1,N).
129 *
130 *  ISUPPZ  (output) INTEGER array, dimension ( 2*max(1,M) )
131 *          The support of the eigenvectors in Z, i.e., the indices
132 *          indicating the nonzero elements in Z. The I-th eigenvector
133 *          is nonzero only in elements ISUPPZ( 2*I-1 ) through
134 *          ISUPPZ( 2*I ).
135 *
136 *  WORK    (workspace) DOUBLE PRECISION array, dimension (12*N)
137 *
138 *  IWORK   (workspace) INTEGER array, dimension (7*N)
139 *
140 *  INFO    (output) INTEGER
141 *          = 0:  successful exit
142 *
143 *          > 0:  A problem occured in DLARRV.
144 *          < 0:  One of the called subroutines signaled an internal problem.
145 *                Needs inspection of the corresponding parameter IINFO
146 *                for further information.
147 *
148 *          =-1:  Problem in DLARRB when refining a child's eigenvalues.
149 *          =-2:  Problem in DLARRF when computing the RRR of a child.
150 *                When a child is inside a tight cluster, it can be difficult
151 *                to find an RRR. A partial remedy from the user's point of
152 *                view is to make the parameter MINRGP smaller and recompile.
153 *                However, as the orthogonality of the computed vectors is
154 *                proportional to 1/MINRGP, the user should be aware that
155 *                he might be trading in precision when he decreases MINRGP.
156 *          =-3:  Problem in DLARRB when refining a single eigenvalue
157 *                after the Rayleigh correction was rejected.
158 *          = 5:  The Rayleigh Quotient Iteration failed to converge to
159 *                full accuracy in MAXITR steps.
160 *
161 *  Further Details
162 *  ===============
163 *
164 *  Based on contributions by
165 *     Beresford Parlett, University of California, Berkeley, USA
166 *     Jim Demmel, University of California, Berkeley, USA
167 *     Inderjit Dhillon, University of Texas, Austin, USA
168 *     Osni Marques, LBNL/NERSC, USA
169 *     Christof Voemel, University of California, Berkeley, USA
170 *
171 *  =====================================================================
172 *
173 *     .. Parameters ..
174       INTEGER            MAXITR
175       PARAMETER          ( MAXITR = 10 )
176       DOUBLE PRECISION   ZERO, ONE, TWO, THREE, FOUR, HALF
177       PARAMETER          ( ZERO = 0.0D0, ONE = 1.0D0,
178      $                     TWO = 2.0D0, THREE = 3.0D0,
179      $                     FOUR = 4.0D0, HALF = 0.5D0)
180 *     ..
181 *     .. Local Scalars ..
182       LOGICAL            ESKIP, NEEDBS, STP2II, TRYRQC, USEDBS, USEDRQ
183       INTEGER            DONE, I, IBEGIN, IDONE, IEND, II, IINDC1,
184      $                   IINDC2, IINDR, IINDWK, IINFO, IM, IN, INDEIG,
185      $                   INDLD, INDLLD, INDWRK, ISUPMN, ISUPMX, ITER,
186      $                   ITMP1, J, JBLK, K, MINIWSIZE, MINWSIZE, NCLUS,
187      $                   NDEPTH, NEGCNT, NEWCLS, NEWFST, NEWFTT, NEWLST,
188      $                   NEWSIZ, OFFSET, OLDCLS, OLDFST, OLDIEN, OLDLST,
189      $                   OLDNCL, P, PARITY, Q, WBEGIN, WEND, WINDEX,
190      $                   WINDMN, WINDPL, ZFROM, ZTO, ZUSEDL, ZUSEDU,
191      $                   ZUSEDW
192       DOUBLE PRECISION   BSTRES, BSTW, EPS, FUDGE, GAP, GAPTOL, GL, GU,
193      $                   LAMBDA, LEFT, LGAP, MINGMA, NRMINV, RESID,
194      $                   RGAP, RIGHT, RQCORR, RQTOL, SAVGAP, SGNDEF,
195      $                   SIGMA, SPDIAM, SSIGMA, TAU, TMP, TOL, ZTZ
196 *     ..
197 *     .. External Functions ..
198       DOUBLE PRECISION   DLAMCH
199       EXTERNAL           DLAMCH
200 *     ..
201 *     .. External Subroutines ..
202       EXTERNAL           DCOPY, DLAR1V, DLARRB, DLARRF, DLASET,
203      $                   DSCAL
204 *     ..
205 *     .. Intrinsic Functions ..
206       INTRINSIC ABSDBLEMAXMIN
207 *     ..
208 *     .. Executable Statements ..
209 *     ..
210 
211 *     The first N entries of WORK are reserved for the eigenvalues
212       INDLD = N+1
213       INDLLD= 2*N+1
214       INDWRK= 3*N+1
215       MINWSIZE = 12 * N
216 
217       DO 5 I= 1,MINWSIZE
218          WORK( I ) = ZERO
219  5    CONTINUE
220 
221 *     IWORK(IINDR+1:IINDR+N) hold the twist indices R for the
222 *     factorization used to compute the FP vector
223       IINDR = 0
224 *     IWORK(IINDC1+1:IINC2+N) are used to store the clusters of the current
225 *     layer and the one above.
226       IINDC1 = N
227       IINDC2 = 2*N
228       IINDWK = 3*+ 1
229 
230       MINIWSIZE = 7 * N
231       DO 10 I= 1,MINIWSIZE
232          IWORK( I ) = 0
233  10   CONTINUE
234 
235       ZUSEDL = 1
236       IF(DOL.GT.1THEN
237 *        Set lower bound for use of Z
238          ZUSEDL = DOL-1
239       ENDIF
240       ZUSEDU = M
241       IF(DOU.LT.M) THEN
242 *        Set lower bound for use of Z
243          ZUSEDU = DOU+1
244       ENDIF
245 *     The width of the part of Z that is used
246       ZUSEDW = ZUSEDU - ZUSEDL + 1
247 
248 
249       CALL DLASET( 'Full', N, ZUSEDW, ZERO, ZERO,
250      $                    Z(1,ZUSEDL), LDZ )
251 
252       EPS = DLAMCH( 'Precision' )
253       RQTOL = TWO * EPS
254 *
255 *     Set expert flags for standard code.
256       TRYRQC = .TRUE.
257 
258       IF((DOL.EQ.1).AND.(DOU.EQ.M)) THEN
259       ELSE
260 *        Only selected eigenpairs are computed. Since the other evalues
261 *        are not refined by RQ iteration, bisection has to compute to full
262 *        accuracy.
263          RTOL1 = FOUR * EPS
264          RTOL2 = FOUR * EPS
265       ENDIF
266 
267 *     The entries WBEGIN:WEND in W, WERR, WGAP correspond to the
268 *     desired eigenvalues. The support of the nonzero eigenvector
269 *     entries is contained in the interval IBEGIN:IEND.
270 *     Remark that if k eigenpairs are desired, then the eigenvectors
271 *     are stored in k contiguous columns of Z.
272 
273 *     DONE is the number of eigenvectors already computed
274       DONE = 0
275       IBEGIN = 1
276       WBEGIN = 1
277       DO 170 JBLK = 1, IBLOCK( M )
278          IEND = ISPLIT( JBLK )
279          SIGMA = L( IEND )
280 *        Find the eigenvectors of the submatrix indexed IBEGIN
281 *        through IEND.
282          WEND = WBEGIN - 1
283  15      CONTINUE
284          IF( WEND.LT.M ) THEN
285             IF( IBLOCK( WEND+1 ).EQ.JBLK ) THEN
286                WEND = WEND + 1
287                GO TO 15
288             END IF
289          END IF
290          IF( WEND.LT.WBEGIN ) THEN
291             IBEGIN = IEND + 1
292             GO TO 170
293          ELSEIF( (WEND.LT.DOL).OR.(WBEGIN.GT.DOU) ) THEN
294             IBEGIN = IEND + 1
295             WBEGIN = WEND + 1
296             GO TO 170
297          END IF
298 
299 *        Find local spectral diameter of the block
300          GL = GERS( 2*IBEGIN-1 )
301          GU = GERS( 2*IBEGIN )
302          DO 20 I = IBEGIN+1 , IEND
303             GL = MIN( GERS( 2*I-1 ), GL )
304             GU = MAX( GERS( 2*I ), GU )
305  20      CONTINUE
306          SPDIAM = GU - GL
307 
308 *        OLDIEN is the last index of the previous block
309          OLDIEN = IBEGIN - 1
310 *        Calculate the size of the current block
311          IN = IEND - IBEGIN + 1
312 *        The number of eigenvalues in the current block
313          IM = WEND - WBEGIN + 1
314 
315 *        This is for a 1x1 block
316          IF( IBEGIN.EQ.IEND ) THEN
317             DONE = DONE+1
318             Z( IBEGIN, WBEGIN ) = ONE
319             ISUPPZ( 2*WBEGIN-1 ) = IBEGIN
320             ISUPPZ( 2*WBEGIN ) = IBEGIN
321             W( WBEGIN ) = W( WBEGIN ) + SIGMA
322             WORK( WBEGIN ) = W( WBEGIN )
323             IBEGIN = IEND + 1
324             WBEGIN = WBEGIN + 1
325             GO TO 170
326          END IF
327 
328 *        The desired (shifted) eigenvalues are stored in W(WBEGIN:WEND)
329 *        Note that these can be approximations, in this case, the corresp.
330 *        entries of WERR give the size of the uncertainty interval.
331 *        The eigenvalue approximations will be refined when necessary as
332 *        high relative accuracy is required for the computation of the
333 *        corresponding eigenvectors.
334          CALL DCOPY( IM, W( WBEGIN ), 1,
335      $                   WORK( WBEGIN ), 1 )
336 
337 *        We store in W the eigenvalue approximations w.r.t. the original
338 *        matrix T.
339          DO 30 I=1,IM
340             W(WBEGIN+I-1= W(WBEGIN+I-1)+SIGMA
341  30      CONTINUE
342 
343 
344 *        NDEPTH is the current depth of the representation tree
345          NDEPTH = 0
346 *        PARITY is either 1 or 0
347          PARITY = 1
348 *        NCLUS is the number of clusters for the next level of the
349 *        representation tree, we start with NCLUS = 1 for the root
350          NCLUS = 1
351          IWORK( IINDC1+1 ) = 1
352          IWORK( IINDC1+2 ) = IM
353 
354 *        IDONE is the number of eigenvectors already computed in the current
355 *        block
356          IDONE = 0
357 *        loop while( IDONE.LT.IM )
358 *        generate the representation tree for the current block and
359 *        compute the eigenvectors
360    40    CONTINUE
361          IF( IDONE.LT.IM ) THEN
362 *           This is a crude protection against infinitely deep trees
363             IF( NDEPTH.GT.M ) THEN
364                INFO = -2
365                RETURN
366             ENDIF
367 *           breadth first processing of the current level of the representation
368 *           tree: OLDNCL = number of clusters on current level
369             OLDNCL = NCLUS
370 *           reset NCLUS to count the number of child clusters
371             NCLUS = 0
372 *
373             PARITY = 1 - PARITY
374             IFPARITY.EQ.0 ) THEN
375                OLDCLS = IINDC1
376                NEWCLS = IINDC2
377             ELSE
378                OLDCLS = IINDC2
379                NEWCLS = IINDC1
380             END IF
381 *           Process the clusters on the current level
382             DO 150 I = 1, OLDNCL
383                J = OLDCLS + 2*I
384 *              OLDFST, OLDLST = first, last index of current cluster.
385 *                               cluster indices start with 1 and are relative
386 *                               to WBEGIN when accessing W, WGAP, WERR, Z
387                OLDFST = IWORK( J-1 )
388                OLDLST = IWORK( J )
389                IF( NDEPTH.GT.0 ) THEN
390 *                 Retrieve relatively robust representation (RRR) of cluster
391 *                 that has been computed at the previous level
392 *                 The RRR is stored in Z and overwritten once the eigenvectors
393 *                 have been computed or when the cluster is refined
394 
395                   IF((DOL.EQ.1).AND.(DOU.EQ.M)) THEN
396 *                    Get representation from location of the leftmost evalue
397 *                    of the cluster
398                      J = WBEGIN + OLDFST - 1
399                   ELSE
400                      IF(WBEGIN+OLDFST-1.LT.DOL) THEN
401 *                       Get representation from the left end of Z array
402                         J = DOL - 1
403                      ELSEIF(WBEGIN+OLDFST-1.GT.DOU) THEN
404 *                       Get representation from the right end of Z array
405                         J = DOU
406                      ELSE
407                         J = WBEGIN + OLDFST - 1
408                      ENDIF
409                   ENDIF
410                   CALL DCOPY( IN, Z( IBEGIN, J ), 1, D( IBEGIN ), 1 )
411                   CALL DCOPY( IN-1, Z( IBEGIN, J+1 ), 1, L( IBEGIN ),
412      $               1 )
413                   SIGMA = Z( IEND, J+1 )
414 
415 *                 Set the corresponding entries in Z to zero
416                   CALL DLASET( 'Full'IN2, ZERO, ZERO,
417      $                         Z( IBEGIN, J), LDZ )
418                END IF
419 
420 *              Compute DL and DLL of current RRR
421                DO 50 J = IBEGIN, IEND-1
422                   TMP = D( J )*L( J )
423                   WORK( INDLD-1+J ) = TMP
424                   WORK( INDLLD-1+J ) = TMP*L( J )
425    50          CONTINUE
426 
427                IF( NDEPTH.GT.0 ) THEN
428 *                 P and Q are index of the first and last eigenvalue to compute
429 *                 within the current block
430                   P = INDEXW( WBEGIN-1+OLDFST )
431                   Q = INDEXW( WBEGIN-1+OLDLST )
432 *                 Offset for the arrays WORK, WGAP and WERR, i.e., the P-OFFSET
433 *                 through the Q-OFFSET elements of these arrays are to be used.
434 *                  OFFSET = P-OLDFST
435                   OFFSET = INDEXW( WBEGIN ) - 1
436 *                 perform limited bisection (if necessary) to get approximate
437 *                 eigenvalues to the precision needed.
438                   CALL DLARRB( IN, D( IBEGIN ),
439      $                         WORK(INDLLD+IBEGIN-1),
440      $                         P, Q, RTOL1, RTOL2, OFFSET,
441      $                         WORK(WBEGIN),WGAP(WBEGIN),WERR(WBEGIN),
442      $                         WORK( INDWRK ), IWORK( IINDWK ),
443      $                         PIVMIN, SPDIAM, IN, IINFO )
444                   IF( IINFO.NE.0 ) THEN
445                      INFO = -1
446                      RETURN
447                   ENDIF
448 *                 We also recompute the extremal gaps. W holds all eigenvalues
449 *                 of the unshifted matrix and must be used for computation
450 *                 of WGAP, the entries of WORK might stem from RRRs with
451 *                 different shifts. The gaps from WBEGIN-1+OLDFST to
452 *                 WBEGIN-1+OLDLST are correctly computed in DLARRB.
453 *                 However, we only allow the gaps to become greater since
454 *                 this is what should happen when we decrease WERR
455                   IF( OLDFST.GT.1THEN
456                      WGAP( WBEGIN+OLDFST-2 ) =
457      $             MAX(WGAP(WBEGIN+OLDFST-2),
458      $                 W(WBEGIN+OLDFST-1)-WERR(WBEGIN+OLDFST-1)
459      $                 - W(WBEGIN+OLDFST-2)-WERR(WBEGIN+OLDFST-2) )
460                   ENDIF
461                   IF( WBEGIN + OLDLST -1 .LT. WEND ) THEN
462                      WGAP( WBEGIN+OLDLST-1 ) =
463      $               MAX(WGAP(WBEGIN+OLDLST-1),
464      $                   W(WBEGIN+OLDLST)-WERR(WBEGIN+OLDLST)
465      $                   - W(WBEGIN+OLDLST-1)-WERR(WBEGIN+OLDLST-1) )
466                   ENDIF
467 *                 Each time the eigenvalues in WORK get refined, we store
468 *                 the newly found approximation with all shifts applied in W
469                   DO 53 J=OLDFST,OLDLST
470                      W(WBEGIN+J-1= WORK(WBEGIN+J-1)+SIGMA
471  53               CONTINUE
472                END IF
473 
474 *              Process the current node.
475                NEWFST = OLDFST
476                DO 140 J = OLDFST, OLDLST
477                   IF( J.EQ.OLDLST ) THEN
478 *                    we are at the right end of the cluster, this is also the
479 *                    boundary of the child cluster
480                      NEWLST = J
481                   ELSE IF ( WGAP( WBEGIN + J -1).GE.
482      $                    MINRGP* ABS( WORK(WBEGIN + J -1) ) ) THEN
483 *                    the right relative gap is big enough, the child cluster
484 *                    (NEWFST,..,NEWLST) is well separated from the following
485                      NEWLST = J
486                    ELSE
487 *                    inside a child cluster, the relative gap is not
488 *                    big enough.
489                      GOTO 140
490                   END IF
491 
492 *                 Compute size of child cluster found
493                   NEWSIZ = NEWLST - NEWFST + 1
494 
495 *                 NEWFTT is the place in Z where the new RRR or the computed
496 *                 eigenvector is to be stored
497                   IF((DOL.EQ.1).AND.(DOU.EQ.M)) THEN
498 *                    Store representation at location of the leftmost evalue
499 *                    of the cluster
500                      NEWFTT = WBEGIN + NEWFST - 1
501                   ELSE
502                      IF(WBEGIN+NEWFST-1.LT.DOL) THEN
503 *                       Store representation at the left end of Z array
504                         NEWFTT = DOL - 1
505                      ELSEIF(WBEGIN+NEWFST-1.GT.DOU) THEN
506 *                       Store representation at the right end of Z array
507                         NEWFTT = DOU
508                      ELSE
509                         NEWFTT = WBEGIN + NEWFST - 1
510                      ENDIF
511                   ENDIF
512 
513                   IF( NEWSIZ.GT.1THEN
514 *
515 *                    Current child is not a singleton but a cluster.
516 *                    Compute and store new representation of child.
517 *
518 *
519 *                    Compute left and right cluster gap.
520 *
521 *                    LGAP and RGAP are not computed from WORK because
522 *                    the eigenvalue approximations may stem from RRRs
523 *                    different shifts. However, W hold all eigenvalues
524 *                    of the unshifted matrix. Still, the entries in WGAP
525 *                    have to be computed from WORK since the entries
526 *                    in W might be of the same order so that gaps are not
527 *                    exhibited correctly for very close eigenvalues.
528                      IF( NEWFST.EQ.1 ) THEN
529                         LGAP = MAX( ZERO,
530      $                       W(WBEGIN)-WERR(WBEGIN) - VL )
531                     ELSE
532                         LGAP = WGAP( WBEGIN+NEWFST-2 )
533                      ENDIF
534                      RGAP = WGAP( WBEGIN+NEWLST-1 )
535 *
536 *                    Compute left- and rightmost eigenvalue of child
537 *                    to high precision in order to shift as close
538 *                    as possible and obtain as large relative gaps
539 *                    as possible
540 *
541                      DO 55 K =1,2
542                         IF(K.EQ.1THEN
543                            P = INDEXW( WBEGIN-1+NEWFST )
544                         ELSE
545                            P = INDEXW( WBEGIN-1+NEWLST )
546                         ENDIF
547                         OFFSET = INDEXW( WBEGIN ) - 1
548                         CALL DLARRB( IN, D(IBEGIN),
549      $                       WORK( INDLLD+IBEGIN-1 ),P,P,
550      $                       RQTOL, RQTOL, OFFSET,
551      $                       WORK(WBEGIN),WGAP(WBEGIN),
552      $                       WERR(WBEGIN),WORK( INDWRK ),
553      $                       IWORK( IINDWK ), PIVMIN, SPDIAM,
554      $                       IN, IINFO )
555  55                  CONTINUE
556 *
557                      IF((WBEGIN+NEWLST-1.LT.DOL).OR.
558      $                  (WBEGIN+NEWFST-1.GT.DOU)) THEN
559 *                       if the cluster contains no desired eigenvalues
560 *                       skip the computation of that branch of the rep. tree
561 *
562 *                       We could skip before the refinement of the extremal
563 *                       eigenvalues of the child, but then the representation
564 *                       tree could be different from the one when nothing is
565 *                       skipped. For this reason we skip at this place.
566                         IDONE = IDONE + NEWLST - NEWFST + 1
567                         GOTO 139
568                      ENDIF
569 *
570 *                    Compute RRR of child cluster.
571 *                    Note that the new RRR is stored in Z
572 *
573 *                    DLARRF needs LWORK = 2*N
574                      CALL DLARRF( IN, D( IBEGIN ), L( IBEGIN ),
575      $                         WORK(INDLD+IBEGIN-1),
576      $                         NEWFST, NEWLST, WORK(WBEGIN),
577      $                         WGAP(WBEGIN), WERR(WBEGIN),
578      $                         SPDIAM, LGAP, RGAP, PIVMIN, TAU,
579      $                         Z(IBEGIN, NEWFTT),Z(IBEGIN, NEWFTT+1),
580      $                         WORK( INDWRK ), IINFO )
581                      IF( IINFO.EQ.0 ) THEN
582 *                       a new RRR for the cluster was found by DLARRF
583 *                       update shift and store it
584                         SSIGMA = SIGMA + TAU
585                         Z( IEND, NEWFTT+1 ) = SSIGMA
586 *                       WORK() are the midpoints and WERR() the semi-width
587 *                       Note that the entries in W are unchanged.
588                         DO 116 K = NEWFST, NEWLST
589                            FUDGE =
590      $                          THREE*EPS*ABS(WORK(WBEGIN+K-1))
591                            WORK( WBEGIN + K - 1 ) =
592      $                          WORK( WBEGIN + K - 1- TAU
593                            FUDGE = FUDGE +
594      $                          FOUR*EPS*ABS(WORK(WBEGIN+K-1))
595 *                          Fudge errors
596                            WERR( WBEGIN + K - 1 ) =
597      $                          WERR( WBEGIN + K - 1 ) + FUDGE
598 *                          Gaps are not fudged. Provided that WERR is small
599 *                          when eigenvalues are close, a zero gap indicates
600 *                          that a new representation is needed for resolving
601 *                          the cluster. A fudge could lead to a wrong decision
602 *                          of judging eigenvalues 'separated' which in
603 *                          reality are not. This could have a negative impact
604 *                          on the orthogonality of the computed eigenvectors.
605  116                    CONTINUE
606 
607                         NCLUS = NCLUS + 1
608                         K = NEWCLS + 2*NCLUS
609                         IWORK( K-1 ) = NEWFST
610                         IWORK( K ) = NEWLST
611                      ELSE
612                         INFO = -2
613                         RETURN
614                      ENDIF
615                   ELSE
616 *
617 *                    Compute eigenvector of singleton
618 *
619                      ITER = 0
620 *
621                      TOL = FOUR * LOG(DBLE(IN)) * EPS
622 *
623                      K = NEWFST
624                      WINDEX = WBEGIN + K - 1
625                      WINDMN = MAX(WINDEX - 1,1)
626                      WINDPL = MIN(WINDEX + 1,M)
627                      LAMBDA = WORK( WINDEX )
628                      DONE = DONE + 1
629 *                    Check if eigenvector computation is to be skipped
630                      IF((WINDEX.LT.DOL).OR.
631      $                  (WINDEX.GT.DOU)) THEN
632                         ESKIP = .TRUE.
633                         GOTO 125
634                      ELSE
635                         ESKIP = .FALSE.
636                      ENDIF
637                      LEFT = WORK( WINDEX ) - WERR( WINDEX )
638                      RIGHT = WORK( WINDEX ) + WERR( WINDEX )
639                      INDEIG = INDEXW( WINDEX )
640 *                    Note that since we compute the eigenpairs for a child,
641 *                    all eigenvalue approximations are w.r.t the same shift.
642 *                    In this case, the entries in WORK should be used for
643 *                    computing the gaps since they exhibit even very small
644 *                    differences in the eigenvalues, as opposed to the
645 *                    entries in W which might "look" the same.
646 
647                      IF( K .EQ. 1THEN
648 *                       In the case RANGE='I' and with not much initial
649 *                       accuracy in LAMBDA and VL, the formula
650 *                       LGAP = MAX( ZERO, (SIGMA - VL) + LAMBDA )
651 *                       can lead to an overestimation of the left gap and
652 *                       thus to inadequately early RQI 'convergence'.
653 *                       Prevent this by forcing a small left gap.
654                         LGAP = EPS*MAX(ABS(LEFT),ABS(RIGHT))
655                      ELSE
656                         LGAP = WGAP(WINDMN)
657                      ENDIF
658                      IF( K .EQ. IM) THEN
659 *                       In the case RANGE='I' and with not much initial
660 *                       accuracy in LAMBDA and VU, the formula
661 *                       can lead to an overestimation of the right gap and
662 *                       thus to inadequately early RQI 'convergence'.
663 *                       Prevent this by forcing a small right gap.
664                         RGAP = EPS*MAX(ABS(LEFT),ABS(RIGHT))
665                      ELSE
666                         RGAP = WGAP(WINDEX)
667                      ENDIF
668                      GAP = MIN( LGAP, RGAP )
669                      IF(( K .EQ. 1).OR.(K .EQ. IM)) THEN
670 *                       The eigenvector support can become wrong
671 *                       because significant entries could be cut off due to a
672 *                       large GAPTOL parameter in LAR1V. Prevent this.
673                         GAPTOL = ZERO
674                      ELSE
675                         GAPTOL = GAP * EPS
676                      ENDIF
677                      ISUPMN = IN
678                      ISUPMX = 1
679 *                    Update WGAP so that it holds the minimum gap
680 *                    to the left or the right. This is crucial in the
681 *                    case where bisection is used to ensure that the
682 *                    eigenvalue is refined up to the required precision.
683 *                    The correct value is restored afterwards.
684                      SAVGAP = WGAP(WINDEX)
685                      WGAP(WINDEX) = GAP
686 *                    We want to use the Rayleigh Quotient Correction
687 *                    as often as possible since it converges quadratically
688 *                    when we are close enough to the desired eigenvalue.
689 *                    However, the Rayleigh Quotient can have the wrong sign
690 *                    and lead us away from the desired eigenvalue. In this
691 *                    case, the best we can do is to use bisection.
692                      USEDBS = .FALSE.
693                      USEDRQ = .FALSE.
694 *                    Bisection is initially turned off unless it is forced
695                      NEEDBS =  .NOT.TRYRQC
696  120                 CONTINUE
697 *                    Check if bisection should be used to refine eigenvalue
698                      IF(NEEDBS) THEN
699 *                       Take the bisection as new iterate
700                         USEDBS = .TRUE.
701                         ITMP1 = IWORK( IINDR+WINDEX )
702                         OFFSET = INDEXW( WBEGIN ) - 1
703                         CALL DLARRB( IN, D(IBEGIN),
704      $                       WORK(INDLLD+IBEGIN-1),INDEIG,INDEIG,
705      $                       ZERO, TWO*EPS, OFFSET,
706      $                       WORK(WBEGIN),WGAP(WBEGIN),
707      $                       WERR(WBEGIN),WORK( INDWRK ),
708      $                       IWORK( IINDWK ), PIVMIN, SPDIAM,
709      $                       ITMP1, IINFO )
710                         IF( IINFO.NE.0 ) THEN
711                            INFO = -3
712                            RETURN
713                         ENDIF
714                         LAMBDA = WORK( WINDEX )
715 *                       Reset twist index from inaccurate LAMBDA to
716 *                       force computation of true MINGMA
717                         IWORK( IINDR+WINDEX ) = 0
718                      ENDIF
719 *                    Given LAMBDA, compute the eigenvector.
720                      CALL DLAR1V( IN1IN, LAMBDA, D( IBEGIN ),
721      $                    L( IBEGIN ), WORK(INDLD+IBEGIN-1),
722      $                    WORK(INDLLD+IBEGIN-1),
723      $                    PIVMIN, GAPTOL, Z( IBEGIN, WINDEX ),
724      $                    .NOT.USEDBS, NEGCNT, ZTZ, MINGMA,
725      $                    IWORK( IINDR+WINDEX ), ISUPPZ( 2*WINDEX-1 ),
726      $                    NRMINV, RESID, RQCORR, WORK( INDWRK ) )
727                      IF(ITER .EQ. 0THEN
728                         BSTRES = RESID
729                         BSTW = LAMBDA
730                      ELSEIF(RESID.LT.BSTRES) THEN
731                         BSTRES = RESID
732                         BSTW = LAMBDA
733                      ENDIF
734                      ISUPMN = MIN(ISUPMN,ISUPPZ( 2*WINDEX-1 ))
735                      ISUPMX = MAX(ISUPMX,ISUPPZ( 2*WINDEX ))
736                      ITER = ITER + 1
737 
738 *                    sin alpha <= |resid|/gap
739 *                    Note that both the residual and the gap are
740 *                    proportional to the matrix, so ||T|| doesn't play
741 *                    a role in the quotient
742 
743 *
744 *                    Convergence test for Rayleigh-Quotient iteration
745 *                    (omitted when Bisection has been used)
746 *
747                      IF( RESID.GT.TOL*GAP .AND. ABS( RQCORR ).GT.
748      $                    RQTOL*ABS( LAMBDA ) .AND. .NOT. USEDBS)
749      $                    THEN
750 *                       We need to check that the RQCORR update doesn't
751 *                       move the eigenvalue away from the desired one and
752 *                       towards a neighbor. -> protection with bisection
753                         IF(INDEIG.LE.NEGCNT) THEN
754 *                          The wanted eigenvalue lies to the left
755                            SGNDEF = -ONE
756                         ELSE
757 *                          The wanted eigenvalue lies to the right
758                            SGNDEF = ONE
759                         ENDIF
760 *                       We only use the RQCORR if it improves the
761 *                       the iterate reasonably.
762                         IF( ( RQCORR*SGNDEF.GE.ZERO )
763      $                       .AND.( LAMBDA + RQCORR.LE. RIGHT)
764      $                       .AND.( LAMBDA + RQCORR.GE. LEFT)
765      $                       ) THEN
766                            USEDRQ = .TRUE.
767 *                          Store new midpoint of bisection interval in WORK
768                            IF(SGNDEF.EQ.ONE) THEN
769 *                             The current LAMBDA is on the left of the true
770 *                             eigenvalue
771                               LEFT = LAMBDA
772 *                             We prefer to assume that the error estimate
773 *                             is correct. We could make the interval not
774 *                             as a bracket but to be modified if the RQCORR
775 *                             chooses to. In this case, the RIGHT side should
776 *                             be modified as follows:
777 *                              RIGHT = MAX(RIGHT, LAMBDA + RQCORR)
778                            ELSE
779 *                             The current LAMBDA is on the right of the true
780 *                             eigenvalue
781                               RIGHT = LAMBDA
782 *                             See comment about assuming the error estimate is
783 *                             correct above.
784 *                              LEFT = MIN(LEFT, LAMBDA + RQCORR)
785                            ENDIF
786                            WORK( WINDEX ) =
787      $                       HALF * (RIGHT + LEFT)
788 *                          Take RQCORR since it has the correct sign and
789 *                          improves the iterate reasonably
790                            LAMBDA = LAMBDA + RQCORR
791 *                          Update width of error interval
792                            WERR( WINDEX ) =
793      $                             HALF * (RIGHT-LEFT)
794                         ELSE
795                            NEEDBS = .TRUE.
796                         ENDIF
797                         IF(RIGHT-LEFT.LT.RQTOL*ABS(LAMBDA)) THEN
798 *                             The eigenvalue is computed to bisection accuracy
799 *                             compute eigenvector and stop
800                            USEDBS = .TRUE.
801                            GOTO 120
802                         ELSEIF( ITER.LT.MAXITR ) THEN
803                            GOTO 120
804                         ELSEIF( ITER.EQ.MAXITR ) THEN
805                            NEEDBS = .TRUE.
806                            GOTO 120
807                         ELSE
808                            INFO = 5
809                            RETURN
810                         END IF
811                      ELSE
812                         STP2II = .FALSE.
813         IF(USEDRQ .AND. USEDBS .AND.
814      $                     BSTRES.LE.RESID) THEN
815                            LAMBDA = BSTW
816                            STP2II = .TRUE.
817                         ENDIF
818                         IF (STP2II) THEN
819 *                          improve error angle by second step
820                            CALL DLAR1V( IN1IN, LAMBDA,
821      $                          D( IBEGIN ), L( IBEGIN ),
822      $                          WORK(INDLD+IBEGIN-1),
823      $                          WORK(INDLLD+IBEGIN-1),
824      $                          PIVMIN, GAPTOL, Z( IBEGIN, WINDEX ),
825      $                          .NOT.USEDBS, NEGCNT, ZTZ, MINGMA,
826      $                          IWORK( IINDR+WINDEX ),
827      $                          ISUPPZ( 2*WINDEX-1 ),
828      $                          NRMINV, RESID, RQCORR, WORK( INDWRK ) )
829                         ENDIF
830                         WORK( WINDEX ) = LAMBDA
831                      END IF
832 *
833 *                    Compute FP-vector support w.r.t. whole matrix
834 *
835                      ISUPPZ( 2*WINDEX-1 ) = ISUPPZ( 2*WINDEX-1 )+OLDIEN
836                      ISUPPZ( 2*WINDEX ) = ISUPPZ( 2*WINDEX )+OLDIEN
837                      ZFROM = ISUPPZ( 2*WINDEX-1 )
838                      ZTO = ISUPPZ( 2*WINDEX )
839                      ISUPMN = ISUPMN + OLDIEN
840                      ISUPMX = ISUPMX + OLDIEN
841 *                    Ensure vector is ok if support in the RQI has changed
842                      IF(ISUPMN.LT.ZFROM) THEN
843                         DO 122 II = ISUPMN,ZFROM-1
844                            Z( II, WINDEX ) = ZERO
845  122                    CONTINUE
846                      ENDIF
847                      IF(ISUPMX.GT.ZTO) THEN
848                         DO 123 II = ZTO+1,ISUPMX
849                            Z( II, WINDEX ) = ZERO
850  123                    CONTINUE
851                      ENDIF
852                      CALL DSCAL( ZTO-ZFROM+1, NRMINV,
853      $                       Z( ZFROM, WINDEX ), 1 )
854  125                 CONTINUE
855 *                    Update W
856                      W( WINDEX ) = LAMBDA+SIGMA
857 *                    Recompute the gaps on the left and right
858 *                    But only allow them to become larger and not
859 *                    smaller (which can only happen through "bad"
860 *                    cancellation and doesn't reflect the theory
861 *                    where the initial gaps are underestimated due
862 *                    to WERR being too crude.)
863                      IF(.NOT.ESKIP) THEN
864                         IF( K.GT.1THEN
865                            WGAP( WINDMN ) = MAX( WGAP(WINDMN),
866      $                          W(WINDEX)-WERR(WINDEX)
867      $                          - W(WINDMN)-WERR(WINDMN) )
868                         ENDIF
869                         IF( WINDEX.LT.WEND ) THEN
870                            WGAP( WINDEX ) = MAX( SAVGAP,
871      $                          W( WINDPL )-WERR( WINDPL )
872      $                          - W( WINDEX )-WERR( WINDEX) )
873                         ENDIF
874                      ENDIF
875                      IDONE = IDONE + 1
876                   ENDIF
877 *                 here ends the code for the current child
878 *
879  139              CONTINUE
880 *                 Proceed to any remaining child nodes
881                   NEWFST = J + 1
882  140           CONTINUE
883  150        CONTINUE
884             NDEPTH = NDEPTH + 1
885             GO TO 40
886          END IF
887          IBEGIN = IEND + 1
888          WBEGIN = WEND + 1
889  170  CONTINUE
890 *
891 
892       RETURN
893 *
894 *     End of DLARRV
895 *
896       END