1      SUBROUTINE ZSTEGR( JOBZ, RANGE, N, D, E, VL, VU, IL, IU, ABSTOL,
2     $                   M, W, Z, LDZ, ISUPPZ, WORK, LWORK, IWORK,
3     $                   LIWORK, INFO )
4*
5*  -- LAPACK computational routine (instru to count ops, version 3.0) --
6*     Univ. of Tennessee, Univ. of California Berkeley, NAG Ltd.,
7*     Courant Institute, Argonne National Lab, and Rice University
8*     June 30, 1999
9*
10*     .. Scalar Arguments ..
11      CHARACTER          JOBZ, RANGE
12      INTEGER            IL, INFO, IU, LDZ, LIWORK, LWORK, M, N
13      DOUBLE PRECISION   ABSTOL, VL, VU
14*     ..
15*     .. Array Arguments ..
16      INTEGER            ISUPPZ( * ), IWORK( * )
17      DOUBLE PRECISION   D( * ), E( * ), W( * ), WORK( * )
18      COMPLEX*16         Z( LDZ, * )
19*     ..
20*     Common block to return operation count ..
21*     .. Common blocks ..
22      COMMON             / LATIME / OPS, ITCNT
23*     ..
24*     .. Scalars in Common ..
25      DOUBLE PRECISION   ITCNT, OPS
26*     ..
27*
28*  Purpose
29*  =======
30*
31*  ZSTEGR computes eigenvalues by the dqds algorithm, while
32*  orthogonal eigenvectors are computed from various "good" L D L^T
33*  representations (also known as Relatively Robust Representations).
34*  Gram-Schmidt orthogonalization is avoided as far as possible. More
35*  specifically, the various steps of the algorithm are as follows.
36*  For the i-th unreduced block of T,
37*     (a) Compute T - sigma_i = L_i D_i L_i^T, such that L_i D_i L_i^T
38*         is a relatively robust representation,
39*     (b) Compute the eigenvalues, lambda_j, of L_i D_i L_i^T to high
40*         relative accuracy by the dqds algorithm,
41*     (c) If there is a cluster of close eigenvalues, "choose" sigma_i
42*         close to the cluster, and go to step (a),
43*     (d) Given the approximate eigenvalue lambda_j of L_i D_i L_i^T,
44*         compute the corresponding eigenvector by forming a
45*         rank-revealing twisted factorization.
46*  The desired accuracy of the output can be specified by the input
47*  parameter ABSTOL.
48*
49*  For more details, see "A new O(n^2) algorithm for the symmetric
50*  tridiagonal eigenvalue/eigenvector problem", by Inderjit Dhillon,
51*  Computer Science Division Technical Report No. UCB/CSD-97-971,
52*  UC Berkeley, May 1997.
53*
54*  Note 1 : Currently ZSTEGR is only set up to find ALL the n
55*  eigenvalues and eigenvectors of T in O(n^2) time
56*  Note 2 : Currently the routine ZSTEIN is called when an appropriate
57*  sigma_i cannot be chosen in step (c) above. ZSTEIN invokes modified
58*  Gram-Schmidt when eigenvalues are close.
59*  Note 3 : ZSTEGR works only on machines which follow ieee-754
60*  floating-point standard in their handling of infinities and NaNs.
61*  Normal execution of ZSTEGR may create NaNs and infinities and hence
62*  may abort due to a floating point exception in environments which
63*  do not conform to the ieee standard.
64*
65*  Arguments
66*  =========
67*
68*  JOBZ    (input) CHARACTER*1
69*          = 'N':  Compute eigenvalues only;
70*          = 'V':  Compute eigenvalues and eigenvectors.
71*
72*  RANGE   (input) CHARACTER*1
73*          = 'A': all eigenvalues will be found.
74*          = 'V': all eigenvalues in the half-open interval (VL,VU]
75*                 will be found.
76*          = 'I': the IL-th through IU-th eigenvalues will be found.
77********** Only RANGE = 'A' is currently supported *********************
78*
79*  N       (input) INTEGER
80*          The order of the matrix.  N >= 0.
81*
82*  D       (input/output) DOUBLE PRECISION array, dimension (N)
83*          On entry, the n diagonal elements of the tridiagonal matrix
84*          T. On exit, D is overwritten.
85*
86*  E       (input/output) DOUBLE PRECISION array, dimension (N)
87*          On entry, the (n-1) subdiagonal elements of the tridiagonal
88*          matrix T in elements 1 to N-1 of E; E(N) need not be set.
89*          On exit, E is overwritten.
90*
91*  VL      (input) DOUBLE PRECISION
92*  VU      (input) DOUBLE PRECISION
93*          If RANGE='V', the lower and upper bounds of the interval to
94*          be searched for eigenvalues. VL < VU.
95*          Not referenced if RANGE = 'A' or 'I'.
96*
97*  IL      (input) INTEGER
98*  IU      (input) INTEGER
99*          If RANGE='I', the indices (in ascending order) of the
100*          smallest and largest eigenvalues to be returned.
101*          1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0.
102*          Not referenced if RANGE = 'A' or 'V'.
103*
104*  ABSTOL  (input) DOUBLE PRECISION
105*          The absolute error tolerance for the
106*          eigenvalues/eigenvectors. IF JOBZ = 'V', the eigenvalues and
107*          eigenvectors output have residual norms bounded by ABSTOL,
108*          and the dot products between different eigenvectors are
109*          bounded by ABSTOL. If ABSTOL is less than N*EPS*|T|, then
110*          N*EPS*|T| will be used in its place, where EPS is the
111*          machine precision and |T| is the 1-norm of the tridiagonal
112*          matrix. The eigenvalues are computed to an accuracy of
113*          EPS*|T| irrespective of ABSTOL. If high relative accuracy
114*          is important, set ABSTOL to DLAMCH( 'Safe minimum' ).
115*          See Barlow and Demmel "Computing Accurate Eigensystems of
116*          Scaled Diagonally Dominant Matrices", LAPACK Working Note #7
117*          for a discussion of which matrices define their eigenvalues
118*          to high relative accuracy.
119*
120*  M       (output) INTEGER
121*          The total number of eigenvalues found.  0 <= M <= N.
122*          If RANGE = 'A', M = N, and if RANGE = 'I', M = IU-IL+1.
123*
124*  W       (output) DOUBLE PRECISION array, dimension (N)
125*          The first M elements contain the selected eigenvalues in
126*          ascending order.
127*
128*  Z       (output) COMPLEX*16 array, dimension (LDZ, max(1,M) )
129*          If JOBZ = 'V', then if INFO = 0, the first M columns of Z
130*          contain the orthonormal eigenvectors of the matrix T
131*          corresponding to the selected eigenvalues, with the i-th
132*          column of Z holding the eigenvector associated with W(i).
133*          If JOBZ = 'N', then Z is not referenced.
134*          Note: the user must ensure that at least max(1,M) columns are
135*          supplied in the array Z; if RANGE = 'V', the exact value of M
136*          is not known in advance and an upper bound must be used.
137*
138*  LDZ     (input) INTEGER
139*          The leading dimension of the array Z.  LDZ >= 1, and if
140*          JOBZ = 'V', LDZ >= max(1,N).
141*
142*  ISUPPZ  (output) INTEGER ARRAY, dimension ( 2*max(1,M) )
143*          The support of the eigenvectors in Z, i.e., the indices
144*          indicating the nonzero elements in Z. The i-th eigenvector
145*          is nonzero only in elements ISUPPZ( 2*i-1 ) through
146*          ISUPPZ( 2*i ).
147*
148*  WORK    (workspace/output) DOUBLE PRECISION array, dimension (LWORK)
149*          On exit, if INFO = 0, WORK(1) returns the optimal
150*          (and minimal) LWORK.
151*
152*  LWORK   (input) INTEGER
153*          The dimension of the array WORK.  LWORK >= max(1,18*N)
154*
155*          If LWORK = -1, then a workspace query is assumed; the routine
156*          only calculates the optimal size of the WORK array, returns
157*          this value as the first entry of the WORK array, and no error
158*          message related to LWORK is issued by XERBLA.
159*
160*  IWORK   (workspace/output) INTEGER array, dimension (LIWORK)
161*          On exit, if INFO = 0, IWORK(1) returns the optimal LIWORK.
162*
163*  LIWORK  (input) INTEGER
164*          The dimension of the array IWORK.  LIWORK >= max(1,10*N)
165*
166*          If LIWORK = -1, then a workspace query is assumed; the
167*          routine only calculates the optimal size of the IWORK array,
168*          returns this value as the first entry of the IWORK array, and
169*          no error message related to LIWORK is issued by XERBLA.
170*
171*  INFO    (output) INTEGER
172*          = 0:  successful exit
173*          < 0:  if INFO = -i, the i-th argument had an illegal value
174*          > 0:  if INFO = 1, internal error in DLARRE,
175*                if INFO = 2, internal error in ZLARRV.
176*
177*  Further Details
178*  ===============
179*
180*  Based on contributions by
181*     Inderjit Dhillon, IBM Almaden, USA
182*     Osni Marques, LBNL/NERSC, USA
183*     Ken Stanley, Computer Science Division, University of
184*       California at Berkeley, USA
185*
186*  =====================================================================
187*
188*     .. Parameters ..
189      DOUBLE PRECISION   ZERO, ONE
190      PARAMETER          ( ZERO = 0.0D0, ONE = 1.0D0 )
191      COMPLEX*16         CZERO
192      PARAMETER          ( CZERO = ( 0.0D0, 0.0D0 ) )
193*     ..
194*     .. Local Scalars ..
195      LOGICAL            ALLEIG, INDEIG, LQUERY, VALEIG, WANTZ
196      INTEGER            I, IBEGIN, IEND, IINDBL, IINDWK, IINFO, IINSPL,
197     $                   INDGRS, INDWOF, INDWRK, ITMP, J, JJ, LIWMIN,
198     $                   LWMIN, NSPLIT
199      DOUBLE PRECISION   BIGNUM, EPS, RMAX, RMIN, SAFMIN, SCALE, SMLNUM,
200     $                   THRESH, TMP, TNRM, TOL
201*     ..
202*     .. External Functions ..
203      LOGICAL            LSAME
204      DOUBLE PRECISION   DLAMCH, DLANST
205      EXTERNAL           LSAME, DLAMCH, DLANST
206*     ..
207*     .. External Subroutines ..
208      EXTERNAL           DLARRE, DSCAL, XERBLA, ZLARRV, ZLASET, ZSWAP
209*     ..
210*     .. Intrinsic Functions ..
211      INTRINSIC          DBLE, MAX, MIN, SQRT
212*     ..
213*     .. Executable Statements ..
214*
215*     Test the input parameters.
216*
217      WANTZ = LSAME( JOBZ, 'V' )
218      ALLEIG = LSAME( RANGE, 'A' )
219      VALEIG = LSAME( RANGE, 'V' )
220      INDEIG = LSAME( RANGE, 'I' )
221*
222      LQUERY = ( ( LWORK.EQ.-1 ) .OR. ( LIWORK.EQ.-1 ) )
223      LWMIN = 18*N
224      LIWMIN = 10*N
225*
226      INFO = 0
227      IF( .NOT.( WANTZ .OR. LSAME( JOBZ, 'N' ) ) ) THEN
228         INFO = -1
229      ELSE IF( .NOT.( ALLEIG .OR. VALEIG .OR. INDEIG ) ) THEN
230         INFO = -2
231*
232*     The following two lines need to be removed once the
233*     RANGE = 'V' and RANGE = 'I' options are provided.
234*
235      ELSE IF( VALEIG .OR. INDEIG ) THEN
236         INFO = -2
237      ELSE IF( N.LT.0 ) THEN
238         INFO = -3
239      ELSE IF( VALEIG .AND. N.GT.0 .AND. VU.LE.VL ) THEN
240         INFO = -7
241      ELSE IF( INDEIG .AND. IL.LT.1 ) THEN
242         INFO = -8
243*     The following change should be made in DSTEVX also, otherwise
244*     IL can be specified as N+1 and IU as N.
245*     ELSE IF( INDEIG .AND. ( IU.LT.MIN( N, IL ) .OR. IU.GT.N ) ) THEN
246      ELSE IF( INDEIG .AND. ( IU.LT.IL .OR. IU.GT.N ) ) THEN
247         INFO = -9
248      ELSE IF( LDZ.LT.1 .OR. ( WANTZ .AND. LDZ.LT.N ) ) THEN
249         INFO = -14
250      ELSE IF( LWORK.LT.LWMIN .AND. .NOT.LQUERY ) THEN
251         INFO = -17
252      ELSE IF( LIWORK.LT.LIWMIN .AND. .NOT.LQUERY ) THEN
253         INFO = -19
254      END IF
255      IF( INFO.EQ.0 ) THEN
256         WORK( 1 ) = LWMIN
257         IWORK( 1 ) = LIWMIN
258      END IF
259*
260      IF( INFO.NE.0 ) THEN
261         CALL XERBLA( 'DSTEGR', -INFO )
262         RETURN
263      ELSE IF( LQUERY ) THEN
264         RETURN
265      END IF
266*
267*     Quick return if possible
268*
269      M = 0
270      IF( N.EQ.0 )
271     $   RETURN
272*
273      IF( N.EQ.1 ) THEN
274         IF( ALLEIG .OR. INDEIG ) THEN
275            M = 1
276            W( 1 ) = D( 1 )
277         ELSE
278            IF( VL.LT.D( 1 ) .AND. VU.GE.D( 1 ) ) THEN
279               M = 1
280               W( 1 ) = D( 1 )
281            END IF
282         END IF
283         IF( WANTZ )
284     $      Z( 1, 1 ) = ONE
285         RETURN
286      END IF
287*
288*     Get machine constants.
289*
290      OPS = OPS + DBLE( 7 )
291      SAFMIN = DLAMCH( 'Safe minimum' )
292      EPS = DLAMCH( 'Precision' )
293      SMLNUM = SAFMIN / EPS
294      BIGNUM = ONE / SMLNUM
295      RMIN = SQRT( SMLNUM )
296      RMAX = MIN( SQRT( BIGNUM ), ONE / SQRT( SQRT( SAFMIN ) ) )
297*
298*     Scale matrix to allowable range, if necessary.
299*
300      SCALE = ONE
301      TNRM = DLANST( 'M', N, D, E )
302      IF( TNRM.GT.ZERO .AND. TNRM.LT.RMIN ) THEN
303         OPS = OPS + DBLE( 1 )
304         SCALE = RMIN / TNRM
305      ELSE IF( TNRM.GT.RMAX ) THEN
306         OPS = OPS + DBLE( 1 )
307         SCALE = RMAX / TNRM
308      END IF
309      IF( SCALE.NE.ONE ) THEN
310         OPS = OPS + DBLE( 2*N )
311         CALL DSCAL( N, SCALE, D, 1 )
312         CALL DSCAL( N-1, SCALE, E, 1 )
313         TNRM = TNRM*SCALE
314      END IF
315      INDGRS = 1
316      INDWOF = 2*N + 1
317      INDWRK = 3*N + 1
318*
319      IINSPL = 1
320      IINDBL = N + 1
321      IINDWK = 2*N + 1
322*
323      CALL ZLASET( 'Full', N, N, CZERO, CZERO, Z, LDZ )
324*
325*     Compute the desired eigenvalues of the tridiagonal after splitting
326*     into smaller subblocks if the corresponding of-diagonal elements
327*     are small
328*
329      OPS = OPS + DBLE( 1 )
330      THRESH = EPS*TNRM
331      CALL DLARRE( N, D, E, THRESH, NSPLIT, IWORK( IINSPL ), M, W,
332     $             WORK( INDWOF ), WORK( INDGRS ), WORK( INDWRK ),
333     $             IINFO )
334      IF( IINFO.NE.0 ) THEN
335         INFO = 1
336         RETURN
337      END IF
338*
339      IF( WANTZ ) THEN
340*
341*        Compute the desired eigenvectors corresponding to the computed
342*        eigenvalues
343*
344         OPS = OPS + DBLE( 1 )
345         TOL = MAX( ABSTOL, DBLE( N )*THRESH )
346         IBEGIN = 1
347         DO 20 I = 1, NSPLIT
348            IEND = IWORK( IINSPL+I-1 )
349            DO 10 J = IBEGIN, IEND
350               IWORK( IINDBL+J-1 ) = I
351   10       CONTINUE
352            IBEGIN = IEND + 1
353   20    CONTINUE
354*
355         CALL ZLARRV( N, D, E, IWORK( IINSPL ), M, W, IWORK( IINDBL ),
356     $                WORK( INDGRS ), TOL, Z, LDZ, ISUPPZ,
357     $                WORK( INDWRK ), IWORK( IINDWK ), IINFO )
358         IF( IINFO.NE.0 ) THEN
359            INFO = 2
360            RETURN
361         END IF
362*
363      END IF
364*
365      IBEGIN = 1
366      DO 40 I = 1, NSPLIT
367         IEND = IWORK( IINSPL+I-1 )
368         DO 30 J = IBEGIN, IEND
369            OPS = OPS + DBLE( 1 )
370            W( J ) = W( J ) + WORK( INDWOF+I-1 )
371   30    CONTINUE
372         IBEGIN = IEND + 1
373   40 CONTINUE
374*
375*     If matrix was scaled, then rescale eigenvalues appropriately.
376*
377      IF( SCALE.NE.ONE ) THEN
378         CALL DSCAL( M, ONE / SCALE, W, 1 )
379      END IF
380*
381*     If eigenvalues are not in order, then sort them, along with
382*     eigenvectors.
383*
384      IF( NSPLIT.GT.1 ) THEN
385         DO 60 J = 1, M - 1
386            I = 0
387            TMP = W( J )
388            DO 50 JJ = J + 1, M
389               IF( W( JJ ).LT.TMP ) THEN
390                  I = JJ
391                  TMP = W( JJ )
392               END IF
393   50       CONTINUE
394            IF( I.NE.0 ) THEN
395               W( I ) = W( J )
396               W( J ) = TMP
397               IF( WANTZ ) THEN
398                  CALL ZSWAP( N, Z( 1, I ), 1, Z( 1, J ), 1 )
399                  ITMP = ISUPPZ( 2*I-1 )
400                  ISUPPZ( 2*I-1 ) = ISUPPZ( 2*J-1 )
401                  ISUPPZ( 2*J-1 ) = ITMP
402                  ITMP = ISUPPZ( 2*I )
403                  ISUPPZ( 2*I ) = ISUPPZ( 2*J )
404                  ISUPPZ( 2*J ) = ITMP
405               END IF
406            END IF
407   60    CONTINUE
408      END IF
409*
410      WORK( 1 ) = LWMIN
411      IWORK( 1 ) = LIWMIN
412      RETURN
413*
414*     End of ZSTEGR
415*
416      END
417