1*> \brief \b DLARRE given the tridiagonal matrix T, sets small off-diagonal elements to zero and for each unreduced block Ti, finds base representations and eigenvalues.
2*
3*  =========== DOCUMENTATION ===========
4*
5* Online html documentation available at
6*            http://www.netlib.org/lapack/explore-html/
7*
8*> \htmlonly
9*> Download DLARRE + dependencies
10*> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/dlarre.f">
11*> [TGZ]</a>
12*> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/dlarre.f">
13*> [ZIP]</a>
14*> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/dlarre.f">
15*> [TXT]</a>
16*> \endhtmlonly
17*
18*  Definition:
19*  ===========
20*
21*       SUBROUTINE DLARRE( RANGE, N, VL, VU, IL, IU, D, E, E2,
22*                           RTOL1, RTOL2, SPLTOL, NSPLIT, ISPLIT, M,
23*                           W, WERR, WGAP, IBLOCK, INDEXW, GERS, PIVMIN,
24*                           WORK, IWORK, INFO )
25*
26*       .. Scalar Arguments ..
27*       CHARACTER          RANGE
28*       INTEGER            IL, INFO, IU, M, N, NSPLIT
29*       DOUBLE PRECISION  PIVMIN, RTOL1, RTOL2, SPLTOL, VL, VU
30*       ..
31*       .. Array Arguments ..
32*       INTEGER            IBLOCK( * ), ISPLIT( * ), IWORK( * ),
33*      $                   INDEXW( * )
34*       DOUBLE PRECISION   D( * ), E( * ), E2( * ), GERS( * ),
35*      $                   W( * ),WERR( * ), WGAP( * ), WORK( * )
36*       ..
37*
38*
39*> \par Purpose:
40*  =============
41*>
42*> \verbatim
43*>
44*> To find the desired eigenvalues of a given real symmetric
45*> tridiagonal matrix T, DLARRE sets any "small" off-diagonal
46*> elements to zero, and for each unreduced block T_i, it finds
47*> (a) a suitable shift at one end of the block's spectrum,
48*> (b) the base representation, T_i - sigma_i I = L_i D_i L_i^T, and
49*> (c) eigenvalues of each L_i D_i L_i^T.
50*> The representations and eigenvalues found are then used by
51*> DSTEMR to compute the eigenvectors of T.
52*> The accuracy varies depending on whether bisection is used to
53*> find a few eigenvalues or the dqds algorithm (subroutine DLASQ2) to
54*> conpute all and then discard any unwanted one.
55*> As an added benefit, DLARRE also outputs the n
56*> Gerschgorin intervals for the matrices L_i D_i L_i^T.
57*> \endverbatim
58*
59*  Arguments:
60*  ==========
61*
62*> \param[in] RANGE
63*> \verbatim
64*>          RANGE is CHARACTER*1
65*>          = 'A': ("All")   all eigenvalues will be found.
66*>          = 'V': ("Value") all eigenvalues in the half-open interval
67*>                           (VL, VU] will be found.
68*>          = 'I': ("Index") the IL-th through IU-th eigenvalues (of the
69*>                           entire matrix) will be found.
70*> \endverbatim
71*>
72*> \param[in] N
73*> \verbatim
74*>          N is INTEGER
75*>          The order of the matrix. N > 0.
76*> \endverbatim
77*>
78*> \param[in,out] VL
79*> \verbatim
80*>          VL is DOUBLE PRECISION
81*> \endverbatim
82*>
83*> \param[in,out] VU
84*> \verbatim
85*>          VU is DOUBLE PRECISION
86*>          If RANGE='V', the lower and upper bounds for the eigenvalues.
87*>          Eigenvalues less than or equal to VL, or greater than VU,
88*>          will not be returned.  VL < VU.
89*>          If RANGE='I' or ='A', DLARRE computes bounds on the desired
90*>          part of the spectrum.
91*> \endverbatim
92*>
93*> \param[in] IL
94*> \verbatim
95*>          IL is INTEGER
96*> \endverbatim
97*>
98*> \param[in] IU
99*> \verbatim
100*>          IU is INTEGER
101*>          If RANGE='I', the indices (in ascending order) of the
102*>          smallest and largest eigenvalues to be returned.
103*>          1 <= IL <= IU <= N.
104*> \endverbatim
105*>
106*> \param[in,out] D
107*> \verbatim
108*>          D is DOUBLE PRECISION array, dimension (N)
109*>          On entry, the N diagonal elements of the tridiagonal
110*>          matrix T.
111*>          On exit, the N diagonal elements of the diagonal
112*>          matrices D_i.
113*> \endverbatim
114*>
115*> \param[in,out] E
116*> \verbatim
117*>          E is DOUBLE PRECISION array, dimension (N)
118*>          On entry, the first (N-1) entries contain the subdiagonal
119*>          elements of the tridiagonal matrix T; E(N) need not be set.
120*>          On exit, E contains the subdiagonal elements of the unit
121*>          bidiagonal matrices L_i. The entries E( ISPLIT( I ) ),
122*>          1 <= I <= NSPLIT, contain the base points sigma_i on output.
123*> \endverbatim
124*>
125*> \param[in,out] E2
126*> \verbatim
127*>          E2 is DOUBLE PRECISION array, dimension (N)
128*>          On entry, the first (N-1) entries contain the SQUARES of the
129*>          subdiagonal elements of the tridiagonal matrix T;
130*>          E2(N) need not be set.
131*>          On exit, the entries E2( ISPLIT( I ) ),
132*>          1 <= I <= NSPLIT, have been set to zero
133*> \endverbatim
134*>
135*> \param[in] RTOL1
136*> \verbatim
137*>          RTOL1 is DOUBLE PRECISION
138*> \endverbatim
139*>
140*> \param[in] RTOL2
141*> \verbatim
142*>          RTOL2 is DOUBLE PRECISION
143*>           Parameters for bisection.
144*>           An interval [LEFT,RIGHT] has converged if
145*>           RIGHT-LEFT.LT.MAX( RTOL1*GAP, RTOL2*MAX(|LEFT|,|RIGHT|) )
146*> \endverbatim
147*>
148*> \param[in] SPLTOL
149*> \verbatim
150*>          SPLTOL is DOUBLE PRECISION
151*>          The threshold for splitting.
152*> \endverbatim
153*>
154*> \param[out] NSPLIT
155*> \verbatim
156*>          NSPLIT is INTEGER
157*>          The number of blocks T splits into. 1 <= NSPLIT <= N.
158*> \endverbatim
159*>
160*> \param[out] ISPLIT
161*> \verbatim
162*>          ISPLIT is INTEGER array, dimension (N)
163*>          The splitting points, at which T breaks up into blocks.
164*>          The first block consists of rows/columns 1 to ISPLIT(1),
165*>          the second of rows/columns ISPLIT(1)+1 through ISPLIT(2),
166*>          etc., and the NSPLIT-th consists of rows/columns
167*>          ISPLIT(NSPLIT-1)+1 through ISPLIT(NSPLIT)=N.
168*> \endverbatim
169*>
170*> \param[out] M
171*> \verbatim
172*>          M is INTEGER
173*>          The total number of eigenvalues (of all L_i D_i L_i^T)
174*>          found.
175*> \endverbatim
176*>
177*> \param[out] W
178*> \verbatim
179*>          W is DOUBLE PRECISION array, dimension (N)
180*>          The first M elements contain the eigenvalues. The
181*>          eigenvalues of each of the blocks, L_i D_i L_i^T, are
182*>          sorted in ascending order ( DLARRE may use the
183*>          remaining N-M elements as workspace).
184*> \endverbatim
185*>
186*> \param[out] WERR
187*> \verbatim
188*>          WERR is DOUBLE PRECISION array, dimension (N)
189*>          The error bound on the corresponding eigenvalue in W.
190*> \endverbatim
191*>
192*> \param[out] WGAP
193*> \verbatim
194*>          WGAP is DOUBLE PRECISION array, dimension (N)
195*>          The separation from the right neighbor eigenvalue in W.
196*>          The gap is only with respect to the eigenvalues of the same block
197*>          as each block has its own representation tree.
198*>          Exception: at the right end of a block we store the left gap
199*> \endverbatim
200*>
201*> \param[out] IBLOCK
202*> \verbatim
203*>          IBLOCK is INTEGER array, dimension (N)
204*>          The indices of the blocks (submatrices) associated with the
205*>          corresponding eigenvalues in W; IBLOCK(i)=1 if eigenvalue
206*>          W(i) belongs to the first block from the top, =2 if W(i)
207*>          belongs to the second block, etc.
208*> \endverbatim
209*>
210*> \param[out] INDEXW
211*> \verbatim
212*>          INDEXW is INTEGER array, dimension (N)
213*>          The indices of the eigenvalues within each block (submatrix);
214*>          for example, INDEXW(i)= 10 and IBLOCK(i)=2 imply that the
215*>          i-th eigenvalue W(i) is the 10-th eigenvalue in block 2
216*> \endverbatim
217*>
218*> \param[out] GERS
219*> \verbatim
220*>          GERS is DOUBLE PRECISION array, dimension (2*N)
221*>          The N Gerschgorin intervals (the i-th Gerschgorin interval
222*>          is (GERS(2*i-1), GERS(2*i)).
223*> \endverbatim
224*>
225*> \param[out] PIVMIN
226*> \verbatim
227*>          PIVMIN is DOUBLE PRECISION
228*>          The minimum pivot in the Sturm sequence for T.
229*> \endverbatim
230*>
231*> \param[out] WORK
232*> \verbatim
233*>          WORK is DOUBLE PRECISION array, dimension (6*N)
234*>          Workspace.
235*> \endverbatim
236*>
237*> \param[out] IWORK
238*> \verbatim
239*>          IWORK is INTEGER array, dimension (5*N)
240*>          Workspace.
241*> \endverbatim
242*>
243*> \param[out] INFO
244*> \verbatim
245*>          INFO is INTEGER
246*>          = 0:  successful exit
247*>          > 0:  A problem occured in DLARRE.
248*>          < 0:  One of the called subroutines signaled an internal problem.
249*>                Needs inspection of the corresponding parameter IINFO
250*>                for further information.
251*>
252*>          =-1:  Problem in DLARRD.
253*>          = 2:  No base representation could be found in MAXTRY iterations.
254*>                Increasing MAXTRY and recompilation might be a remedy.
255*>          =-3:  Problem in DLARRB when computing the refined root
256*>                representation for DLASQ2.
257*>          =-4:  Problem in DLARRB when preforming bisection on the
258*>                desired part of the spectrum.
259*>          =-5:  Problem in DLASQ2.
260*>          =-6:  Problem in DLASQ2.
261*> \endverbatim
262*
263*  Authors:
264*  ========
265*
266*> \author Univ. of Tennessee
267*> \author Univ. of California Berkeley
268*> \author Univ. of Colorado Denver
269*> \author NAG Ltd.
270*
271*> \date September 2012
272*
273*> \ingroup auxOTHERauxiliary
274*
275*> \par Further Details:
276*  =====================
277*>
278*> \verbatim
279*>
280*>  The base representations are required to suffer very little
281*>  element growth and consequently define all their eigenvalues to
282*>  high relative accuracy.
283*> \endverbatim
284*
285*> \par Contributors:
286*  ==================
287*>
288*>     Beresford Parlett, University of California, Berkeley, USA \n
289*>     Jim Demmel, University of California, Berkeley, USA \n
290*>     Inderjit Dhillon, University of Texas, Austin, USA \n
291*>     Osni Marques, LBNL/NERSC, USA \n
292*>     Christof Voemel, University of California, Berkeley, USA \n
293*>
294*  =====================================================================
295      SUBROUTINE DLARRE( RANGE, N, VL, VU, IL, IU, D, E, E2,
296     $                    RTOL1, RTOL2, SPLTOL, NSPLIT, ISPLIT, M,
297     $                    W, WERR, WGAP, IBLOCK, INDEXW, GERS, PIVMIN,
298     $                    WORK, IWORK, INFO )
299*
300*  -- LAPACK auxiliary routine (version 3.4.2) --
301*  -- LAPACK is a software package provided by Univ. of Tennessee,    --
302*  -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
303*     September 2012
304*
305*     .. Scalar Arguments ..
306      CHARACTER          RANGE
307      INTEGER            IL, INFO, IU, M, N, NSPLIT
308      DOUBLE PRECISION  PIVMIN, RTOL1, RTOL2, SPLTOL, VL, VU
309*     ..
310*     .. Array Arguments ..
311      INTEGER            IBLOCK( * ), ISPLIT( * ), IWORK( * ),
312     $                   INDEXW( * )
313      DOUBLE PRECISION   D( * ), E( * ), E2( * ), GERS( * ),
314     $                   W( * ),WERR( * ), WGAP( * ), WORK( * )
315*     ..
316*
317*  =====================================================================
318*
319*     .. Parameters ..
320      DOUBLE PRECISION   FAC, FOUR, FOURTH, FUDGE, HALF, HNDRD,
321     $                   MAXGROWTH, ONE, PERT, TWO, ZERO
322      PARAMETER          ( ZERO = 0.0D0, ONE = 1.0D0,
323     $                     TWO = 2.0D0, FOUR=4.0D0,
324     $                     HNDRD = 100.0D0,
325     $                     PERT = 8.0D0,
326     $                     HALF = ONE/TWO, FOURTH = ONE/FOUR, FAC= HALF,
327     $                     MAXGROWTH = 64.0D0, FUDGE = 2.0D0 )
328      INTEGER            MAXTRY, ALLRNG, INDRNG, VALRNG
329      PARAMETER          ( MAXTRY = 6, ALLRNG = 1, INDRNG = 2,
330     $                     VALRNG = 3 )
331*     ..
332*     .. Local Scalars ..
333      LOGICAL            FORCEB, NOREP, USEDQD
334      INTEGER            CNT, CNT1, CNT2, I, IBEGIN, IDUM, IEND, IINFO,
335     $                   IN, INDL, INDU, IRANGE, J, JBLK, MB, MM,
336     $                   WBEGIN, WEND
337      DOUBLE PRECISION   AVGAP, BSRTOL, CLWDTH, DMAX, DPIVOT, EABS,
338     $                   EMAX, EOLD, EPS, GL, GU, ISLEFT, ISRGHT, RTL,
339     $                   RTOL, S1, S2, SAFMIN, SGNDEF, SIGMA, SPDIAM,
340     $                   TAU, TMP, TMP1
341
342
343*     ..
344*     .. Local Arrays ..
345      INTEGER            ISEED( 4 )
346*     ..
347*     .. External Functions ..
348      LOGICAL            LSAME
349      DOUBLE PRECISION            DLAMCH
350      EXTERNAL           DLAMCH, LSAME
351
352*     ..
353*     .. External Subroutines ..
354      EXTERNAL           DCOPY, DLARNV, DLARRA, DLARRB, DLARRC, DLARRD,
355     $                   DLASQ2
356*     ..
357*     .. Intrinsic Functions ..
358      INTRINSIC          ABS, MAX, MIN
359
360*     ..
361*     .. Executable Statements ..
362*
363
364      INFO = 0
365
366*
367*     Decode RANGE
368*
369      IF( LSAME( RANGE, 'A' ) ) THEN
370         IRANGE = ALLRNG
371      ELSE IF( LSAME( RANGE, 'V' ) ) THEN
372         IRANGE = VALRNG
373      ELSE IF( LSAME( RANGE, 'I' ) ) THEN
374         IRANGE = INDRNG
375      END IF
376
377      M = 0
378
379*     Get machine constants
380      SAFMIN = DLAMCH( 'S' )
381      EPS = DLAMCH( 'P' )
382
383*     Set parameters
384      RTL = SQRT(EPS)
385      BSRTOL = SQRT(EPS)
386
387*     Treat case of 1x1 matrix for quick return
388      IF( N.EQ.1 ) THEN
389         IF( (IRANGE.EQ.ALLRNG).OR.
390     $       ((IRANGE.EQ.VALRNG).AND.(D(1).GT.VL).AND.(D(1).LE.VU)).OR.
391     $       ((IRANGE.EQ.INDRNG).AND.(IL.EQ.1).AND.(IU.EQ.1)) ) THEN
392            M = 1
393            W(1) = D(1)
394*           The computation error of the eigenvalue is zero
395            WERR(1) = ZERO
396            WGAP(1) = ZERO
397            IBLOCK( 1 ) = 1
398            INDEXW( 1 ) = 1
399            GERS(1) = D( 1 )
400            GERS(2) = D( 1 )
401         ENDIF
402*        store the shift for the initial RRR, which is zero in this case
403         E(1) = ZERO
404         RETURN
405      END IF
406
407*     General case: tridiagonal matrix of order > 1
408*
409*     Init WERR, WGAP. Compute Gerschgorin intervals and spectral diameter.
410*     Compute maximum off-diagonal entry and pivmin.
411      GL = D(1)
412      GU = D(1)
413      EOLD = ZERO
414      EMAX = ZERO
415      E(N) = ZERO
416      DO 5 I = 1,N
417         WERR(I) = ZERO
418         WGAP(I) = ZERO
419         EABS = ABS( E(I) )
420         IF( EABS .GE. EMAX ) THEN
421            EMAX = EABS
422         END IF
423         TMP1 = EABS + EOLD
424         GERS( 2*I-1) = D(I) - TMP1
425         GL =  MIN( GL, GERS( 2*I - 1))
426         GERS( 2*I ) = D(I) + TMP1
427         GU = MAX( GU, GERS(2*I) )
428         EOLD  = EABS
429 5    CONTINUE
430*     The minimum pivot allowed in the Sturm sequence for T
431      PIVMIN = SAFMIN * MAX( ONE, EMAX**2 )
432*     Compute spectral diameter. The Gerschgorin bounds give an
433*     estimate that is wrong by at most a factor of SQRT(2)
434      SPDIAM = GU - GL
435
436*     Compute splitting points
437      CALL DLARRA( N, D, E, E2, SPLTOL, SPDIAM,
438     $                    NSPLIT, ISPLIT, IINFO )
439
440*     Can force use of bisection instead of faster DQDS.
441*     Option left in the code for future multisection work.
442      FORCEB = .FALSE.
443
444*     Initialize USEDQD, DQDS should be used for ALLRNG unless someone
445*     explicitly wants bisection.
446      USEDQD = (( IRANGE.EQ.ALLRNG ) .AND. (.NOT.FORCEB))
447
448      IF( (IRANGE.EQ.ALLRNG) .AND. (.NOT. FORCEB) ) THEN
449*        Set interval [VL,VU] that contains all eigenvalues
450         VL = GL
451         VU = GU
452      ELSE
453*        We call DLARRD to find crude approximations to the eigenvalues
454*        in the desired range. In case IRANGE = INDRNG, we also obtain the
455*        interval (VL,VU] that contains all the wanted eigenvalues.
456*        An interval [LEFT,RIGHT] has converged if
457*        RIGHT-LEFT.LT.RTOL*MAX(ABS(LEFT),ABS(RIGHT))
458*        DLARRD needs a WORK of size 4*N, IWORK of size 3*N
459         CALL DLARRD( RANGE, 'B', N, VL, VU, IL, IU, GERS,
460     $                    BSRTOL, D, E, E2, PIVMIN, NSPLIT, ISPLIT,
461     $                    MM, W, WERR, VL, VU, IBLOCK, INDEXW,
462     $                    WORK, IWORK, IINFO )
463         IF( IINFO.NE.0 ) THEN
464            INFO = -1
465            RETURN
466         ENDIF
467*        Make sure that the entries M+1 to N in W, WERR, IBLOCK, INDEXW are 0
468         DO 14 I = MM+1,N
469            W( I ) = ZERO
470            WERR( I ) = ZERO
471            IBLOCK( I ) = 0
472            INDEXW( I ) = 0
473 14      CONTINUE
474      END IF
475
476
477***
478*     Loop over unreduced blocks
479      IBEGIN = 1
480      WBEGIN = 1
481      DO 170 JBLK = 1, NSPLIT
482         IEND = ISPLIT( JBLK )
483         IN = IEND - IBEGIN + 1
484
485*        1 X 1 block
486         IF( IN.EQ.1 ) THEN
487            IF( (IRANGE.EQ.ALLRNG).OR.( (IRANGE.EQ.VALRNG).AND.
488     $         ( D( IBEGIN ).GT.VL ).AND.( D( IBEGIN ).LE.VU ) )
489     $        .OR. ( (IRANGE.EQ.INDRNG).AND.(IBLOCK(WBEGIN).EQ.JBLK))
490     $        ) THEN
491               M = M + 1
492               W( M ) = D( IBEGIN )
493               WERR(M) = ZERO
494*              The gap for a single block doesn't matter for the later
495*              algorithm and is assigned an arbitrary large value
496               WGAP(M) = ZERO
497               IBLOCK( M ) = JBLK
498               INDEXW( M ) = 1
499               WBEGIN = WBEGIN + 1
500            ENDIF
501*           E( IEND ) holds the shift for the initial RRR
502            E( IEND ) = ZERO
503            IBEGIN = IEND + 1
504            GO TO 170
505         END IF
506*
507*        Blocks of size larger than 1x1
508*
509*        E( IEND ) will hold the shift for the initial RRR, for now set it =0
510         E( IEND ) = ZERO
511*
512*        Find local outer bounds GL,GU for the block
513         GL = D(IBEGIN)
514         GU = D(IBEGIN)
515         DO 15 I = IBEGIN , IEND
516            GL = MIN( GERS( 2*I-1 ), GL )
517            GU = MAX( GERS( 2*I ), GU )
518 15      CONTINUE
519         SPDIAM = GU - GL
520
521         IF(.NOT. ((IRANGE.EQ.ALLRNG).AND.(.NOT.FORCEB)) ) THEN
522*           Count the number of eigenvalues in the current block.
523            MB = 0
524            DO 20 I = WBEGIN,MM
525               IF( IBLOCK(I).EQ.JBLK ) THEN
526                  MB = MB+1
527               ELSE
528                  GOTO 21
529               ENDIF
530 20         CONTINUE
531 21         CONTINUE
532
533            IF( MB.EQ.0) THEN
534*              No eigenvalue in the current block lies in the desired range
535*              E( IEND ) holds the shift for the initial RRR
536               E( IEND ) = ZERO
537               IBEGIN = IEND + 1
538               GO TO 170
539            ELSE
540
541*              Decide whether dqds or bisection is more efficient
542               USEDQD = ( (MB .GT. FAC*IN) .AND. (.NOT.FORCEB) )
543               WEND = WBEGIN + MB - 1
544*              Calculate gaps for the current block
545*              In later stages, when representations for individual
546*              eigenvalues are different, we use SIGMA = E( IEND ).
547               SIGMA = ZERO
548               DO 30 I = WBEGIN, WEND - 1
549                  WGAP( I ) = MAX( ZERO,
550     $                        W(I+1)-WERR(I+1) - (W(I)+WERR(I)) )
551 30            CONTINUE
552               WGAP( WEND ) = MAX( ZERO,
553     $                     VU - SIGMA - (W( WEND )+WERR( WEND )))
554*              Find local index of the first and last desired evalue.
555               INDL = INDEXW(WBEGIN)
556               INDU = INDEXW( WEND )
557            ENDIF
558         ENDIF
559         IF(( (IRANGE.EQ.ALLRNG) .AND. (.NOT. FORCEB) ).OR.USEDQD) THEN
560*           Case of DQDS
561*           Find approximations to the extremal eigenvalues of the block
562            CALL DLARRK( IN, 1, GL, GU, D(IBEGIN),
563     $               E2(IBEGIN), PIVMIN, RTL, TMP, TMP1, IINFO )
564            IF( IINFO.NE.0 ) THEN
565               INFO = -1
566               RETURN
567            ENDIF
568            ISLEFT = MAX(GL, TMP - TMP1
569     $               - HNDRD * EPS* ABS(TMP - TMP1))
570
571            CALL DLARRK( IN, IN, GL, GU, D(IBEGIN),
572     $               E2(IBEGIN), PIVMIN, RTL, TMP, TMP1, IINFO )
573            IF( IINFO.NE.0 ) THEN
574               INFO = -1
575               RETURN
576            ENDIF
577            ISRGHT = MIN(GU, TMP + TMP1
578     $                 + HNDRD * EPS * ABS(TMP + TMP1))
579*           Improve the estimate of the spectral diameter
580            SPDIAM = ISRGHT - ISLEFT
581         ELSE
582*           Case of bisection
583*           Find approximations to the wanted extremal eigenvalues
584            ISLEFT = MAX(GL, W(WBEGIN) - WERR(WBEGIN)
585     $                  - HNDRD * EPS*ABS(W(WBEGIN)- WERR(WBEGIN) ))
586            ISRGHT = MIN(GU,W(WEND) + WERR(WEND)
587     $                  + HNDRD * EPS * ABS(W(WEND)+ WERR(WEND)))
588         ENDIF
589
590
591*        Decide whether the base representation for the current block
592*        L_JBLK D_JBLK L_JBLK^T = T_JBLK - sigma_JBLK I
593*        should be on the left or the right end of the current block.
594*        The strategy is to shift to the end which is "more populated"
595*        Furthermore, decide whether to use DQDS for the computation of
596*        the eigenvalue approximations at the end of DLARRE or bisection.
597*        dqds is chosen if all eigenvalues are desired or the number of
598*        eigenvalues to be computed is large compared to the blocksize.
599         IF( ( IRANGE.EQ.ALLRNG ) .AND. (.NOT.FORCEB) ) THEN
600*           If all the eigenvalues have to be computed, we use dqd
601            USEDQD = .TRUE.
602*           INDL is the local index of the first eigenvalue to compute
603            INDL = 1
604            INDU = IN
605*           MB =  number of eigenvalues to compute
606            MB = IN
607            WEND = WBEGIN + MB - 1
608*           Define 1/4 and 3/4 points of the spectrum
609            S1 = ISLEFT + FOURTH * SPDIAM
610            S2 = ISRGHT - FOURTH * SPDIAM
611         ELSE
612*           DLARRD has computed IBLOCK and INDEXW for each eigenvalue
613*           approximation.
614*           choose sigma
615            IF( USEDQD ) THEN
616               S1 = ISLEFT + FOURTH * SPDIAM
617               S2 = ISRGHT - FOURTH * SPDIAM
618            ELSE
619               TMP = MIN(ISRGHT,VU) -  MAX(ISLEFT,VL)
620               S1 =  MAX(ISLEFT,VL) + FOURTH * TMP
621               S2 =  MIN(ISRGHT,VU) - FOURTH * TMP
622            ENDIF
623         ENDIF
624
625*        Compute the negcount at the 1/4 and 3/4 points
626         IF(MB.GT.1) THEN
627            CALL DLARRC( 'T', IN, S1, S2, D(IBEGIN),
628     $                    E(IBEGIN), PIVMIN, CNT, CNT1, CNT2, IINFO)
629         ENDIF
630
631         IF(MB.EQ.1) THEN
632            SIGMA = GL
633            SGNDEF = ONE
634         ELSEIF( CNT1 - INDL .GE. INDU - CNT2 ) THEN
635            IF( ( IRANGE.EQ.ALLRNG ) .AND. (.NOT.FORCEB) ) THEN
636               SIGMA = MAX(ISLEFT,GL)
637            ELSEIF( USEDQD ) THEN
638*              use Gerschgorin bound as shift to get pos def matrix
639*              for dqds
640               SIGMA = ISLEFT
641            ELSE
642*              use approximation of the first desired eigenvalue of the
643*              block as shift
644               SIGMA = MAX(ISLEFT,VL)
645            ENDIF
646            SGNDEF = ONE
647         ELSE
648            IF( ( IRANGE.EQ.ALLRNG ) .AND. (.NOT.FORCEB) ) THEN
649               SIGMA = MIN(ISRGHT,GU)
650            ELSEIF( USEDQD ) THEN
651*              use Gerschgorin bound as shift to get neg def matrix
652*              for dqds
653               SIGMA = ISRGHT
654            ELSE
655*              use approximation of the first desired eigenvalue of the
656*              block as shift
657               SIGMA = MIN(ISRGHT,VU)
658            ENDIF
659            SGNDEF = -ONE
660         ENDIF
661
662
663*        An initial SIGMA has been chosen that will be used for computing
664*        T - SIGMA I = L D L^T
665*        Define the increment TAU of the shift in case the initial shift
666*        needs to be refined to obtain a factorization with not too much
667*        element growth.
668         IF( USEDQD ) THEN
669*           The initial SIGMA was to the outer end of the spectrum
670*           the matrix is definite and we need not retreat.
671            TAU = SPDIAM*EPS*N + TWO*PIVMIN
672            TAU = MAX( TAU,TWO*EPS*ABS(SIGMA) )
673         ELSE
674            IF(MB.GT.1) THEN
675               CLWDTH = W(WEND) + WERR(WEND) - W(WBEGIN) - WERR(WBEGIN)
676               AVGAP = ABS(CLWDTH / DBLE(WEND-WBEGIN))
677               IF( SGNDEF.EQ.ONE ) THEN
678                  TAU = HALF*MAX(WGAP(WBEGIN),AVGAP)
679                  TAU = MAX(TAU,WERR(WBEGIN))
680               ELSE
681                  TAU = HALF*MAX(WGAP(WEND-1),AVGAP)
682                  TAU = MAX(TAU,WERR(WEND))
683               ENDIF
684            ELSE
685               TAU = WERR(WBEGIN)
686            ENDIF
687         ENDIF
688*
689         DO 80 IDUM = 1, MAXTRY
690*           Compute L D L^T factorization of tridiagonal matrix T - sigma I.
691*           Store D in WORK(1:IN), L in WORK(IN+1:2*IN), and reciprocals of
692*           pivots in WORK(2*IN+1:3*IN)
693            DPIVOT = D( IBEGIN ) - SIGMA
694            WORK( 1 ) = DPIVOT
695            DMAX = ABS( WORK(1) )
696            J = IBEGIN
697            DO 70 I = 1, IN - 1
698               WORK( 2*IN+I ) = ONE / WORK( I )
699               TMP = E( J )*WORK( 2*IN+I )
700               WORK( IN+I ) = TMP
701               DPIVOT = ( D( J+1 )-SIGMA ) - TMP*E( J )
702               WORK( I+1 ) = DPIVOT
703               DMAX = MAX( DMAX, ABS(DPIVOT) )
704               J = J + 1
705 70         CONTINUE
706*           check for element growth
707            IF( DMAX .GT. MAXGROWTH*SPDIAM ) THEN
708               NOREP = .TRUE.
709            ELSE
710               NOREP = .FALSE.
711            ENDIF
712            IF( USEDQD .AND. .NOT.NOREP ) THEN
713*              Ensure the definiteness of the representation
714*              All entries of D (of L D L^T) must have the same sign
715               DO 71 I = 1, IN
716                  TMP = SGNDEF*WORK( I )
717                  IF( TMP.LT.ZERO ) NOREP = .TRUE.
718 71            CONTINUE
719            ENDIF
720            IF(NOREP) THEN
721*              Note that in the case of IRANGE=ALLRNG, we use the Gerschgorin
722*              shift which makes the matrix definite. So we should end up
723*              here really only in the case of IRANGE = VALRNG or INDRNG.
724               IF( IDUM.EQ.MAXTRY-1 ) THEN
725                  IF( SGNDEF.EQ.ONE ) THEN
726*                    The fudged Gerschgorin shift should succeed
727                     SIGMA =
728     $                    GL - FUDGE*SPDIAM*EPS*N - FUDGE*TWO*PIVMIN
729                  ELSE
730                     SIGMA =
731     $                    GU + FUDGE*SPDIAM*EPS*N + FUDGE*TWO*PIVMIN
732                  END IF
733               ELSE
734                  SIGMA = SIGMA - SGNDEF * TAU
735                  TAU = TWO * TAU
736               END IF
737            ELSE
738*              an initial RRR is found
739               GO TO 83
740            END IF
741 80      CONTINUE
742*        if the program reaches this point, no base representation could be
743*        found in MAXTRY iterations.
744         INFO = 2
745         RETURN
746
747 83      CONTINUE
748*        At this point, we have found an initial base representation
749*        T - SIGMA I = L D L^T with not too much element growth.
750*        Store the shift.
751         E( IEND ) = SIGMA
752*        Store D and L.
753         CALL DCOPY( IN, WORK, 1, D( IBEGIN ), 1 )
754         CALL DCOPY( IN-1, WORK( IN+1 ), 1, E( IBEGIN ), 1 )
755
756
757         IF(MB.GT.1 ) THEN
758*
759*           Perturb each entry of the base representation by a small
760*           (but random) relative amount to overcome difficulties with
761*           glued matrices.
762*
763            DO 122 I = 1, 4
764               ISEED( I ) = 1
765 122        CONTINUE
766
767            CALL DLARNV(2, ISEED, 2*IN-1, WORK(1))
768            DO 125 I = 1,IN-1
769               D(IBEGIN+I-1) = D(IBEGIN+I-1)*(ONE+EPS*PERT*WORK(I))
770               E(IBEGIN+I-1) = E(IBEGIN+I-1)*(ONE+EPS*PERT*WORK(IN+I))
771 125        CONTINUE
772            D(IEND) = D(IEND)*(ONE+EPS*FOUR*WORK(IN))
773*
774         ENDIF
775*
776*        Don't update the Gerschgorin intervals because keeping track
777*        of the updates would be too much work in DLARRV.
778*        We update W instead and use it to locate the proper Gerschgorin
779*        intervals.
780
781*        Compute the required eigenvalues of L D L' by bisection or dqds
782         IF ( .NOT.USEDQD ) THEN
783*           If DLARRD has been used, shift the eigenvalue approximations
784*           according to their representation. This is necessary for
785*           a uniform DLARRV since dqds computes eigenvalues of the
786*           shifted representation. In DLARRV, W will always hold the
787*           UNshifted eigenvalue approximation.
788            DO 134 J=WBEGIN,WEND
789               W(J) = W(J) - SIGMA
790               WERR(J) = WERR(J) + ABS(W(J)) * EPS
791 134        CONTINUE
792*           call DLARRB to reduce eigenvalue error of the approximations
793*           from DLARRD
794            DO 135 I = IBEGIN, IEND-1
795               WORK( I ) = D( I ) * E( I )**2
796 135        CONTINUE
797*           use bisection to find EV from INDL to INDU
798            CALL DLARRB(IN, D(IBEGIN), WORK(IBEGIN),
799     $                  INDL, INDU, RTOL1, RTOL2, INDL-1,
800     $                  W(WBEGIN), WGAP(WBEGIN), WERR(WBEGIN),
801     $                  WORK( 2*N+1 ), IWORK, PIVMIN, SPDIAM,
802     $                  IN, IINFO )
803            IF( IINFO .NE. 0 ) THEN
804               INFO = -4
805               RETURN
806            END IF
807*           DLARRB computes all gaps correctly except for the last one
808*           Record distance to VU/GU
809            WGAP( WEND ) = MAX( ZERO,
810     $           ( VU-SIGMA ) - ( W( WEND ) + WERR( WEND ) ) )
811            DO 138 I = INDL, INDU
812               M = M + 1
813               IBLOCK(M) = JBLK
814               INDEXW(M) = I
815 138        CONTINUE
816         ELSE
817*           Call dqds to get all eigs (and then possibly delete unwanted
818*           eigenvalues).
819*           Note that dqds finds the eigenvalues of the L D L^T representation
820*           of T to high relative accuracy. High relative accuracy
821*           might be lost when the shift of the RRR is subtracted to obtain
822*           the eigenvalues of T. However, T is not guaranteed to define its
823*           eigenvalues to high relative accuracy anyway.
824*           Set RTOL to the order of the tolerance used in DLASQ2
825*           This is an ESTIMATED error, the worst case bound is 4*N*EPS
826*           which is usually too large and requires unnecessary work to be
827*           done by bisection when computing the eigenvectors
828            RTOL = LOG(DBLE(IN)) * FOUR * EPS
829            J = IBEGIN
830            DO 140 I = 1, IN - 1
831               WORK( 2*I-1 ) = ABS( D( J ) )
832               WORK( 2*I ) = E( J )*E( J )*WORK( 2*I-1 )
833               J = J + 1
834  140       CONTINUE
835            WORK( 2*IN-1 ) = ABS( D( IEND ) )
836            WORK( 2*IN ) = ZERO
837            CALL DLASQ2( IN, WORK, IINFO )
838            IF( IINFO .NE. 0 ) THEN
839*              If IINFO = -5 then an index is part of a tight cluster
840*              and should be changed. The index is in IWORK(1) and the
841*              gap is in WORK(N+1)
842               INFO = -5
843               RETURN
844            ELSE
845*              Test that all eigenvalues are positive as expected
846               DO 149 I = 1, IN
847                  IF( WORK( I ).LT.ZERO ) THEN
848                     INFO = -6
849                     RETURN
850                  ENDIF
851 149           CONTINUE
852            END IF
853            IF( SGNDEF.GT.ZERO ) THEN
854               DO 150 I = INDL, INDU
855                  M = M + 1
856                  W( M ) = WORK( IN-I+1 )
857                  IBLOCK( M ) = JBLK
858                  INDEXW( M ) = I
859 150           CONTINUE
860            ELSE
861               DO 160 I = INDL, INDU
862                  M = M + 1
863                  W( M ) = -WORK( I )
864                  IBLOCK( M ) = JBLK
865                  INDEXW( M ) = I
866 160           CONTINUE
867            END IF
868
869            DO 165 I = M - MB + 1, M
870*              the value of RTOL below should be the tolerance in DLASQ2
871               WERR( I ) = RTOL * ABS( W(I) )
872 165        CONTINUE
873            DO 166 I = M - MB + 1, M - 1
874*              compute the right gap between the intervals
875               WGAP( I ) = MAX( ZERO,
876     $                          W(I+1)-WERR(I+1) - (W(I)+WERR(I)) )
877 166        CONTINUE
878            WGAP( M ) = MAX( ZERO,
879     $           ( VU-SIGMA ) - ( W( M ) + WERR( M ) ) )
880         END IF
881*        proceed with next block
882         IBEGIN = IEND + 1
883         WBEGIN = WEND + 1
884 170  CONTINUE
885*
886
887      RETURN
888*
889*     end of DLARRE
890*
891      END
892