1 | SUBROUTINE CSTEMR( JOBZ, RANGE, N, D, E, VL, VU, IL, IU,
|
---|
2 | $ M, W, Z, LDZ, NZC, ISUPPZ, TRYRAC, WORK, LWORK,
|
---|
3 | $ IWORK, LIWORK, INFO )
|
---|
4 | IMPLICIT NONE
|
---|
5 | *
|
---|
6 | * -- LAPACK computational routine (version 3.1) --
|
---|
7 | * Univ. of Tennessee, Univ. of California Berkeley and NAG Ltd..
|
---|
8 | * November 2006
|
---|
9 | *
|
---|
10 | * .. Scalar Arguments ..
|
---|
11 | CHARACTER JOBZ, RANGE
|
---|
12 | LOGICAL TRYRAC
|
---|
13 | INTEGER IL, INFO, IU, LDZ, NZC, LIWORK, LWORK, M, N
|
---|
14 | REAL VL, VU
|
---|
15 | * ..
|
---|
16 | * .. Array Arguments ..
|
---|
17 | INTEGER ISUPPZ( * ), IWORK( * )
|
---|
18 | REAL D( * ), E( * ), W( * ), WORK( * )
|
---|
19 | COMPLEX Z( LDZ, * )
|
---|
20 | * ..
|
---|
21 | *
|
---|
22 | * Purpose
|
---|
23 | * =======
|
---|
24 | *
|
---|
25 | * CSTEMR computes selected eigenvalues and, optionally, eigenvectors
|
---|
26 | * of a real symmetric tridiagonal matrix T. Any such unreduced matrix has
|
---|
27 | * a well defined set of pairwise different real eigenvalues, the corresponding
|
---|
28 | * real eigenvectors are pairwise orthogonal.
|
---|
29 | *
|
---|
30 | * The spectrum may be computed either completely or partially by specifying
|
---|
31 | * either an interval (VL,VU] or a range of indices IL:IU for the desired
|
---|
32 | * eigenvalues.
|
---|
33 | *
|
---|
34 | * Depending on the number of desired eigenvalues, these are computed either
|
---|
35 | * by bisection or the dqds algorithm. Numerically orthogonal eigenvectors are
|
---|
36 | * computed by the use of various suitable L D L^T factorizations near clusters
|
---|
37 | * of close eigenvalues (referred to as RRRs, Relatively Robust
|
---|
38 | * Representations). An informal sketch of the algorithm follows.
|
---|
39 | *
|
---|
40 | * For each unreduced block (submatrix) of T,
|
---|
41 | * (a) Compute T - sigma I = L D L^T, so that L and D
|
---|
42 | * define all the wanted eigenvalues to high relative accuracy.
|
---|
43 | * This means that small relative changes in the entries of D and L
|
---|
44 | * cause only small relative changes in the eigenvalues and
|
---|
45 | * eigenvectors. The standard (unfactored) representation of the
|
---|
46 | * tridiagonal matrix T does not have this property in general.
|
---|
47 | * (b) Compute the eigenvalues to suitable accuracy.
|
---|
48 | * If the eigenvectors are desired, the algorithm attains full
|
---|
49 | * accuracy of the computed eigenvalues only right before
|
---|
50 | * the corresponding vectors have to be computed, see steps c) and d).
|
---|
51 | * (c) For each cluster of close eigenvalues, select a new
|
---|
52 | * shift close to the cluster, find a new factorization, and refine
|
---|
53 | * the shifted eigenvalues to suitable accuracy.
|
---|
54 | * (d) For each eigenvalue with a large enough relative separation compute
|
---|
55 | * the corresponding eigenvector by forming a rank revealing twisted
|
---|
56 | * factorization. Go back to (c) for any clusters that remain.
|
---|
57 | *
|
---|
58 | * For more details, see:
|
---|
59 | * - Inderjit S. Dhillon and Beresford N. Parlett: "Multiple representations
|
---|
60 | * to compute orthogonal eigenvectors of symmetric tridiagonal matrices,"
|
---|
61 | * Linear Algebra and its Applications, 387(1), pp. 1-28, August 2004.
|
---|
62 | * - Inderjit Dhillon and Beresford Parlett: "Orthogonal Eigenvectors and
|
---|
63 | * Relative Gaps," SIAM Journal on Matrix Analysis and Applications, Vol. 25,
|
---|
64 | * 2004. Also LAPACK Working Note 154.
|
---|
65 | * - Inderjit Dhillon: "A new O(n^2) algorithm for the symmetric
|
---|
66 | * tridiagonal eigenvalue/eigenvector problem",
|
---|
67 | * Computer Science Division Technical Report No. UCB/CSD-97-971,
|
---|
68 | * UC Berkeley, May 1997.
|
---|
69 | *
|
---|
70 | * Notes:
|
---|
71 | * 1.CSTEMR works only on machines which follow IEEE-754
|
---|
72 | * floating-point standard in their handling of infinities and NaNs.
|
---|
73 | * This permits the use of efficient inner loops avoiding a check for
|
---|
74 | * zero divisors.
|
---|
75 | *
|
---|
76 | * 2. LAPACK routines can be used to reduce a complex Hermitean matrix to
|
---|
77 | * real symmetric tridiagonal form.
|
---|
78 | *
|
---|
79 | * (Any complex Hermitean tridiagonal matrix has real values on its diagonal
|
---|
80 | * and potentially complex numbers on its off-diagonals. By applying a
|
---|
81 | * similarity transform with an appropriate diagonal matrix
|
---|
82 | * diag(1,e^{i \phy_1}, ... , e^{i \phy_{n-1}}), the complex Hermitean
|
---|
83 | * matrix can be transformed into a real symmetric matrix and complex
|
---|
84 | * arithmetic can be entirely avoided.)
|
---|
85 | *
|
---|
86 | * While the eigenvectors of the real symmetric tridiagonal matrix are real,
|
---|
87 | * the eigenvectors of original complex Hermitean matrix have complex entries
|
---|
88 | * in general.
|
---|
89 | * Since LAPACK drivers overwrite the matrix data with the eigenvectors,
|
---|
90 | * CSTEMR accepts complex workspace to facilitate interoperability
|
---|
91 | * with CUNMTR or CUPMTR.
|
---|
92 | *
|
---|
93 | * Arguments
|
---|
94 | * =========
|
---|
95 | *
|
---|
96 | * JOBZ (input) CHARACTER*1
|
---|
97 | * = 'N': Compute eigenvalues only;
|
---|
98 | * = 'V': Compute eigenvalues and eigenvectors.
|
---|
99 | *
|
---|
100 | * RANGE (input) CHARACTER*1
|
---|
101 | * = 'A': all eigenvalues will be found.
|
---|
102 | * = 'V': all eigenvalues in the half-open interval (VL,VU]
|
---|
103 | * will be found.
|
---|
104 | * = 'I': the IL-th through IU-th eigenvalues will be found.
|
---|
105 | *
|
---|
106 | * N (input) INTEGER
|
---|
107 | * The order of the matrix. N >= 0.
|
---|
108 | *
|
---|
109 | * D (input/output) REAL array, dimension (N)
|
---|
110 | * On entry, the N diagonal elements of the tridiagonal matrix
|
---|
111 | * T. On exit, D is overwritten.
|
---|
112 | *
|
---|
113 | * E (input/output) REAL array, dimension (N)
|
---|
114 | * On entry, the (N-1) subdiagonal elements of the tridiagonal
|
---|
115 | * matrix T in elements 1 to N-1 of E. E(N) need not be set on
|
---|
116 | * input, but is used internally as workspace.
|
---|
117 | * On exit, E is overwritten.
|
---|
118 | *
|
---|
119 | * VL (input) REAL
|
---|
120 | * VU (input) REAL
|
---|
121 | * If RANGE='V', the lower and upper bounds of the interval to
|
---|
122 | * be searched for eigenvalues. VL < VU.
|
---|
123 | * Not referenced if RANGE = 'A' or 'I'.
|
---|
124 | *
|
---|
125 | * IL (input) INTEGER
|
---|
126 | * IU (input) INTEGER
|
---|
127 | * If RANGE='I', the indices (in ascending order) of the
|
---|
128 | * smallest and largest eigenvalues to be returned.
|
---|
129 | * 1 <= IL <= IU <= N, if N > 0.
|
---|
130 | * Not referenced if RANGE = 'A' or 'V'.
|
---|
131 | *
|
---|
132 | * M (output) INTEGER
|
---|
133 | * The total number of eigenvalues found. 0 <= M <= N.
|
---|
134 | * If RANGE = 'A', M = N, and if RANGE = 'I', M = IU-IL+1.
|
---|
135 | *
|
---|
136 | * W (output) REAL array, dimension (N)
|
---|
137 | * The first M elements contain the selected eigenvalues in
|
---|
138 | * ascending order.
|
---|
139 | *
|
---|
140 | * Z (output) COMPLEX array, dimension (LDZ, max(1,M) )
|
---|
141 | * If JOBZ = 'V', and if INFO = 0, then the first M columns of Z
|
---|
142 | * contain the orthonormal eigenvectors of the matrix T
|
---|
143 | * corresponding to the selected eigenvalues, with the i-th
|
---|
144 | * column of Z holding the eigenvector associated with W(i).
|
---|
145 | * If JOBZ = 'N', then Z is not referenced.
|
---|
146 | * Note: the user must ensure that at least max(1,M) columns are
|
---|
147 | * supplied in the array Z; if RANGE = 'V', the exact value of M
|
---|
148 | * is not known in advance and can be computed with a workspace
|
---|
149 | * query by setting NZC = -1, see below.
|
---|
150 | *
|
---|
151 | * LDZ (input) INTEGER
|
---|
152 | * The leading dimension of the array Z. LDZ >= 1, and if
|
---|
153 | * JOBZ = 'V', then LDZ >= max(1,N).
|
---|
154 | *
|
---|
155 | * NZC (input) INTEGER
|
---|
156 | * The number of eigenvectors to be held in the array Z.
|
---|
157 | * If RANGE = 'A', then NZC >= max(1,N).
|
---|
158 | * If RANGE = 'V', then NZC >= the number of eigenvalues in (VL,VU].
|
---|
159 | * If RANGE = 'I', then NZC >= IU-IL+1.
|
---|
160 | * If NZC = -1, then a workspace query is assumed; the
|
---|
161 | * routine calculates the number of columns of the array Z that
|
---|
162 | * are needed to hold the eigenvectors.
|
---|
163 | * This value is returned as the first entry of the Z array, and
|
---|
164 | * no error message related to NZC is issued by XERBLA.
|
---|
165 | *
|
---|
166 | * ISUPPZ (output) INTEGER ARRAY, dimension ( 2*max(1,M) )
|
---|
167 | * The support of the eigenvectors in Z, i.e., the indices
|
---|
168 | * indicating the nonzero elements in Z. The i-th computed eigenvector
|
---|
169 | * is nonzero only in elements ISUPPZ( 2*i-1 ) through
|
---|
170 | * ISUPPZ( 2*i ). This is relevant in the case when the matrix
|
---|
171 | * is split. ISUPPZ is only accessed when JOBZ is 'V' and N > 0.
|
---|
172 | *
|
---|
173 | * TRYRAC (input/output) LOGICAL
|
---|
174 | * If TRYRAC.EQ..TRUE., indicates that the code should check whether
|
---|
175 | * the tridiagonal matrix defines its eigenvalues to high relative
|
---|
176 | * accuracy. If so, the code uses relative-accuracy preserving
|
---|
177 | * algorithms that might be (a bit) slower depending on the matrix.
|
---|
178 | * If the matrix does not define its eigenvalues to high relative
|
---|
179 | * accuracy, the code can uses possibly faster algorithms.
|
---|
180 | * If TRYRAC.EQ..FALSE., the code is not required to guarantee
|
---|
181 | * relatively accurate eigenvalues and can use the fastest possible
|
---|
182 | * techniques.
|
---|
183 | * On exit, a .TRUE. TRYRAC will be set to .FALSE. if the matrix
|
---|
184 | * does not define its eigenvalues to high relative accuracy.
|
---|
185 | *
|
---|
186 | * WORK (workspace/output) REAL array, dimension (LWORK)
|
---|
187 | * On exit, if INFO = 0, WORK(1) returns the optimal
|
---|
188 | * (and minimal) LWORK.
|
---|
189 | *
|
---|
190 | * LWORK (input) INTEGER
|
---|
191 | * The dimension of the array WORK. LWORK >= max(1,18*N)
|
---|
192 | * if JOBZ = 'V', and LWORK >= max(1,12*N) if JOBZ = 'N'.
|
---|
193 | * If LWORK = -1, then a workspace query is assumed; the routine
|
---|
194 | * only calculates the optimal size of the WORK array, returns
|
---|
195 | * this value as the first entry of the WORK array, and no error
|
---|
196 | * message related to LWORK is issued by XERBLA.
|
---|
197 | *
|
---|
198 | * IWORK (workspace/output) INTEGER array, dimension (LIWORK)
|
---|
199 | * On exit, if INFO = 0, IWORK(1) returns the optimal LIWORK.
|
---|
200 | *
|
---|
201 | * LIWORK (input) INTEGER
|
---|
202 | * The dimension of the array IWORK. LIWORK >= max(1,10*N)
|
---|
203 | * if the eigenvectors are desired, and LIWORK >= max(1,8*N)
|
---|
204 | * if only the eigenvalues are to be computed.
|
---|
205 | * If LIWORK = -1, then a workspace query is assumed; the
|
---|
206 | * routine only calculates the optimal size of the IWORK array,
|
---|
207 | * returns this value as the first entry of the IWORK array, and
|
---|
208 | * no error message related to LIWORK is issued by XERBLA.
|
---|
209 | *
|
---|
210 | * INFO (output) INTEGER
|
---|
211 | * On exit, INFO
|
---|
212 | * = 0: successful exit
|
---|
213 | * < 0: if INFO = -i, the i-th argument had an illegal value
|
---|
214 | * > 0: if INFO = 1X, internal error in SLARRE,
|
---|
215 | * if INFO = 2X, internal error in CLARRV.
|
---|
216 | * Here, the digit X = ABS( IINFO ) < 10, where IINFO is
|
---|
217 | * the nonzero error code returned by SLARRE or
|
---|
218 | * CLARRV, respectively.
|
---|
219 | *
|
---|
220 | *
|
---|
221 | * Further Details
|
---|
222 | * ===============
|
---|
223 | *
|
---|
224 | * Based on contributions by
|
---|
225 | * Beresford Parlett, University of California, Berkeley, USA
|
---|
226 | * Jim Demmel, University of California, Berkeley, USA
|
---|
227 | * Inderjit Dhillon, University of Texas, Austin, USA
|
---|
228 | * Osni Marques, LBNL/NERSC, USA
|
---|
229 | * Christof Voemel, University of California, Berkeley, USA
|
---|
230 | *
|
---|
231 | * =====================================================================
|
---|
232 | *
|
---|
233 | * .. Parameters ..
|
---|
234 | REAL ZERO, ONE, FOUR, MINRGP
|
---|
235 | PARAMETER ( ZERO = 0.0E0, ONE = 1.0E0,
|
---|
236 | $ FOUR = 4.0E0,
|
---|
237 | $ MINRGP = 3.0E-3 )
|
---|
238 | * ..
|
---|
239 | * .. Local Scalars ..
|
---|
240 | LOGICAL ALLEIG, INDEIG, LQUERY, VALEIG, WANTZ, ZQUERY
|
---|
241 | INTEGER I, IBEGIN, IEND, IFIRST, IIL, IINDBL, IINDW,
|
---|
242 | $ IINDWK, IINFO, IINSPL, IIU, ILAST, IN, INDD,
|
---|
243 | $ INDE2, INDERR, INDGP, INDGRS, INDWRK, ITMP,
|
---|
244 | $ ITMP2, J, JBLK, JJ, LIWMIN, LWMIN, NSPLIT,
|
---|
245 | $ NZCMIN, OFFSET, WBEGIN, WEND
|
---|
246 | REAL BIGNUM, CS, EPS, PIVMIN, R1, R2, RMAX, RMIN,
|
---|
247 | $ RTOL1, RTOL2, SAFMIN, SCALE, SMLNUM, SN,
|
---|
248 | $ THRESH, TMP, TNRM, WL, WU
|
---|
249 | * ..
|
---|
250 | * ..
|
---|
251 | * .. External Functions ..
|
---|
252 | LOGICAL LSAME
|
---|
253 | REAL SLAMCH, SLANST
|
---|
254 | EXTERNAL LSAME, SLAMCH, SLANST
|
---|
255 | * ..
|
---|
256 | * .. External Subroutines ..
|
---|
257 | EXTERNAL CLARRV, CSWAP, SCOPY, SLAE2, SLAEV2, SLARRC,
|
---|
258 | $ SLARRE, SLARRJ, SLARRR, SLASRT, SSCAL, XERBLA
|
---|
259 | * ..
|
---|
260 | * .. Intrinsic Functions ..
|
---|
261 | INTRINSIC MAX, MIN, SQRT
|
---|
262 |
|
---|
263 |
|
---|
264 | * ..
|
---|
265 | * .. Executable Statements ..
|
---|
266 | *
|
---|
267 | * Test the input parameters.
|
---|
268 | *
|
---|
269 | WANTZ = LSAME( JOBZ, 'V' )
|
---|
270 | ALLEIG = LSAME( RANGE, 'A' )
|
---|
271 | VALEIG = LSAME( RANGE, 'V' )
|
---|
272 | INDEIG = LSAME( RANGE, 'I' )
|
---|
273 | *
|
---|
274 | LQUERY = ( ( LWORK.EQ.-1 ).OR.( LIWORK.EQ.-1 ) )
|
---|
275 | ZQUERY = ( NZC.EQ.-1 )
|
---|
276 | TRYRAC = ( INFO.NE.0 )
|
---|
277 |
|
---|
278 | * SSTEMR needs WORK of size 6*N, IWORK of size 3*N.
|
---|
279 | * In addition, SLARRE needs WORK of size 6*N, IWORK of size 5*N.
|
---|
280 | * Furthermore, CLARRV needs WORK of size 12*N, IWORK of size 7*N.
|
---|
281 | IF( WANTZ ) THEN
|
---|
282 | LWMIN = 18*N
|
---|
283 | LIWMIN = 10*N
|
---|
284 | ELSE
|
---|
285 | * need less workspace if only the eigenvalues are wanted
|
---|
286 | LWMIN = 12*N
|
---|
287 | LIWMIN = 8*N
|
---|
288 | ENDIF
|
---|
289 |
|
---|
290 | WL = ZERO
|
---|
291 | WU = ZERO
|
---|
292 | IIL = 0
|
---|
293 | IIU = 0
|
---|
294 |
|
---|
295 | IF( VALEIG ) THEN
|
---|
296 | * We do not reference VL, VU in the cases RANGE = 'I','A'
|
---|
297 | * The interval (WL, WU] contains all the wanted eigenvalues.
|
---|
298 | * It is either given by the user or computed in SLARRE.
|
---|
299 | WL = VL
|
---|
300 | WU = VU
|
---|
301 | ELSEIF( INDEIG ) THEN
|
---|
302 | * We do not reference IL, IU in the cases RANGE = 'V','A'
|
---|
303 | IIL = IL
|
---|
304 | IIU = IU
|
---|
305 | ENDIF
|
---|
306 | *
|
---|
307 | INFO = 0
|
---|
308 | IF( .NOT.( WANTZ .OR. LSAME( JOBZ, 'N' ) ) ) THEN
|
---|
309 | INFO = -1
|
---|
310 | ELSE IF( .NOT.( ALLEIG .OR. VALEIG .OR. INDEIG ) ) THEN
|
---|
311 | INFO = -2
|
---|
312 | ELSE IF( N.LT.0 ) THEN
|
---|
313 | INFO = -3
|
---|
314 | ELSE IF( VALEIG .AND. N.GT.0 .AND. WU.LE.WL ) THEN
|
---|
315 | INFO = -7
|
---|
316 | ELSE IF( INDEIG .AND. ( IIL.LT.1 .OR. IIL.GT.N ) ) THEN
|
---|
317 | INFO = -8
|
---|
318 | ELSE IF( INDEIG .AND. ( IIU.LT.IIL .OR. IIU.GT.N ) ) THEN
|
---|
319 | INFO = -9
|
---|
320 | ELSE IF( LDZ.LT.1 .OR. ( WANTZ .AND. LDZ.LT.N ) ) THEN
|
---|
321 | INFO = -13
|
---|
322 | ELSE IF( LWORK.LT.LWMIN .AND. .NOT.LQUERY ) THEN
|
---|
323 | INFO = -17
|
---|
324 | ELSE IF( LIWORK.LT.LIWMIN .AND. .NOT.LQUERY ) THEN
|
---|
325 | INFO = -19
|
---|
326 | END IF
|
---|
327 | *
|
---|
328 | * Get machine constants.
|
---|
329 | *
|
---|
330 | SAFMIN = SLAMCH( 'Safe minimum' )
|
---|
331 | EPS = SLAMCH( 'Precision' )
|
---|
332 | SMLNUM = SAFMIN / EPS
|
---|
333 | BIGNUM = ONE / SMLNUM
|
---|
334 | RMIN = SQRT( SMLNUM )
|
---|
335 | RMAX = MIN( SQRT( BIGNUM ), ONE / SQRT( SQRT( SAFMIN ) ) )
|
---|
336 | *
|
---|
337 | IF( INFO.EQ.0 ) THEN
|
---|
338 | WORK( 1 ) = LWMIN
|
---|
339 | IWORK( 1 ) = LIWMIN
|
---|
340 | *
|
---|
341 | IF( WANTZ .AND. ALLEIG ) THEN
|
---|
342 | NZCMIN = N
|
---|
343 | ELSE IF( WANTZ .AND. VALEIG ) THEN
|
---|
344 | CALL SLARRC( 'T', N, VL, VU, D, E, SAFMIN,
|
---|
345 | $ NZCMIN, ITMP, ITMP2, INFO )
|
---|
346 | ELSE IF( WANTZ .AND. INDEIG ) THEN
|
---|
347 | NZCMIN = IIU-IIL+1
|
---|
348 | ELSE
|
---|
349 | * WANTZ .EQ. FALSE.
|
---|
350 | NZCMIN = 0
|
---|
351 | ENDIF
|
---|
352 | IF( ZQUERY .AND. INFO.EQ.0 ) THEN
|
---|
353 | Z( 1,1 ) = NZCMIN
|
---|
354 | ELSE IF( NZC.LT.NZCMIN .AND. .NOT.ZQUERY ) THEN
|
---|
355 | INFO = -14
|
---|
356 | END IF
|
---|
357 | END IF
|
---|
358 |
|
---|
359 | IF( INFO.NE.0 ) THEN
|
---|
360 | *
|
---|
361 | CALL XERBLA( 'CSTEMR', -INFO )
|
---|
362 | *
|
---|
363 | RETURN
|
---|
364 | ELSE IF( LQUERY .OR. ZQUERY ) THEN
|
---|
365 | RETURN
|
---|
366 | END IF
|
---|
367 | *
|
---|
368 | * Handle N = 0, 1, and 2 cases immediately
|
---|
369 | *
|
---|
370 | M = 0
|
---|
371 | IF( N.EQ.0 )
|
---|
372 | $ RETURN
|
---|
373 | *
|
---|
374 | IF( N.EQ.1 ) THEN
|
---|
375 | IF( ALLEIG .OR. INDEIG ) THEN
|
---|
376 | M = 1
|
---|
377 | W( 1 ) = D( 1 )
|
---|
378 | ELSE
|
---|
379 | IF( WL.LT.D( 1 ) .AND. WU.GE.D( 1 ) ) THEN
|
---|
380 | M = 1
|
---|
381 | W( 1 ) = D( 1 )
|
---|
382 | END IF
|
---|
383 | END IF
|
---|
384 | IF( WANTZ.AND.(.NOT.ZQUERY) ) THEN
|
---|
385 | Z( 1, 1 ) = ONE
|
---|
386 | ISUPPZ(1) = 1
|
---|
387 | ISUPPZ(2) = 1
|
---|
388 | END IF
|
---|
389 | RETURN
|
---|
390 | END IF
|
---|
391 | *
|
---|
392 | IF( N.EQ.2 ) THEN
|
---|
393 | IF( .NOT.WANTZ ) THEN
|
---|
394 | CALL SLAE2( D(1), E(1), D(2), R1, R2 )
|
---|
395 | ELSE IF( WANTZ.AND.(.NOT.ZQUERY) ) THEN
|
---|
396 | CALL SLAEV2( D(1), E(1), D(2), R1, R2, CS, SN )
|
---|
397 | END IF
|
---|
398 | IF( ALLEIG.OR.
|
---|
399 | $ (VALEIG.AND.(R2.GT.WL).AND.
|
---|
400 | $ (R2.LE.WU)).OR.
|
---|
401 | $ (INDEIG.AND.(IIL.EQ.1)) ) THEN
|
---|
402 | M = M+1
|
---|
403 | W( M ) = R2
|
---|
404 | IF( WANTZ.AND.(.NOT.ZQUERY) ) THEN
|
---|
405 | Z( 1, M ) = -SN
|
---|
406 | Z( 2, M ) = CS
|
---|
407 | * Note: At most one of SN and CS can be zero.
|
---|
408 | IF (SN.NE.ZERO) THEN
|
---|
409 | IF (CS.NE.ZERO) THEN
|
---|
410 | ISUPPZ(2*M-1) = 1
|
---|
411 | ISUPPZ(2*M-1) = 2
|
---|
412 | ELSE
|
---|
413 | ISUPPZ(2*M-1) = 1
|
---|
414 | ISUPPZ(2*M-1) = 1
|
---|
415 | END IF
|
---|
416 | ELSE
|
---|
417 | ISUPPZ(2*M-1) = 2
|
---|
418 | ISUPPZ(2*M) = 2
|
---|
419 | END IF
|
---|
420 | ENDIF
|
---|
421 | ENDIF
|
---|
422 | IF( ALLEIG.OR.
|
---|
423 | $ (VALEIG.AND.(R1.GT.WL).AND.
|
---|
424 | $ (R1.LE.WU)).OR.
|
---|
425 | $ (INDEIG.AND.(IIU.EQ.2)) ) THEN
|
---|
426 | M = M+1
|
---|
427 | W( M ) = R1
|
---|
428 | IF( WANTZ.AND.(.NOT.ZQUERY) ) THEN
|
---|
429 | Z( 1, M ) = CS
|
---|
430 | Z( 2, M ) = SN
|
---|
431 | * Note: At most one of SN and CS can be zero.
|
---|
432 | IF (SN.NE.ZERO) THEN
|
---|
433 | IF (CS.NE.ZERO) THEN
|
---|
434 | ISUPPZ(2*M-1) = 1
|
---|
435 | ISUPPZ(2*M-1) = 2
|
---|
436 | ELSE
|
---|
437 | ISUPPZ(2*M-1) = 1
|
---|
438 | ISUPPZ(2*M-1) = 1
|
---|
439 | END IF
|
---|
440 | ELSE
|
---|
441 | ISUPPZ(2*M-1) = 2
|
---|
442 | ISUPPZ(2*M) = 2
|
---|
443 | END IF
|
---|
444 | ENDIF
|
---|
445 | ENDIF
|
---|
446 | RETURN
|
---|
447 | END IF
|
---|
448 |
|
---|
449 | * Continue with general N
|
---|
450 |
|
---|
451 | INDGRS = 1
|
---|
452 | INDERR = 2*N + 1
|
---|
453 | INDGP = 3*N + 1
|
---|
454 | INDD = 4*N + 1
|
---|
455 | INDE2 = 5*N + 1
|
---|
456 | INDWRK = 6*N + 1
|
---|
457 | *
|
---|
458 | IINSPL = 1
|
---|
459 | IINDBL = N + 1
|
---|
460 | IINDW = 2*N + 1
|
---|
461 | IINDWK = 3*N + 1
|
---|
462 | *
|
---|
463 | * Scale matrix to allowable range, if necessary.
|
---|
464 | * The allowable range is related to the PIVMIN parameter; see the
|
---|
465 | * comments in SLARRD. The preference for scaling small values
|
---|
466 | * up is heuristic; we expect users' matrices not to be close to the
|
---|
467 | * RMAX threshold.
|
---|
468 | *
|
---|
469 | SCALE = ONE
|
---|
470 | TNRM = SLANST( 'M', N, D, E )
|
---|
471 | IF( TNRM.GT.ZERO .AND. TNRM.LT.RMIN ) THEN
|
---|
472 | SCALE = RMIN / TNRM
|
---|
473 | ELSE IF( TNRM.GT.RMAX ) THEN
|
---|
474 | SCALE = RMAX / TNRM
|
---|
475 | END IF
|
---|
476 | IF( SCALE.NE.ONE ) THEN
|
---|
477 | CALL SSCAL( N, SCALE, D, 1 )
|
---|
478 | CALL SSCAL( N-1, SCALE, E, 1 )
|
---|
479 | TNRM = TNRM*SCALE
|
---|
480 | IF( VALEIG ) THEN
|
---|
481 | * If eigenvalues in interval have to be found,
|
---|
482 | * scale (WL, WU] accordingly
|
---|
483 | WL = WL*SCALE
|
---|
484 | WU = WU*SCALE
|
---|
485 | ENDIF
|
---|
486 | END IF
|
---|
487 | *
|
---|
488 | * Compute the desired eigenvalues of the tridiagonal after splitting
|
---|
489 | * into smaller subblocks if the corresponding off-diagonal elements
|
---|
490 | * are small
|
---|
491 | * THRESH is the splitting parameter for SLARRE
|
---|
492 | * A negative THRESH forces the old splitting criterion based on the
|
---|
493 | * size of the off-diagonal. A positive THRESH switches to splitting
|
---|
494 | * which preserves relative accuracy.
|
---|
495 | *
|
---|
496 | IF( TRYRAC ) THEN
|
---|
497 | * Test whether the matrix warrants the more expensive relative approach.
|
---|
498 | CALL SLARRR( N, D, E, IINFO )
|
---|
499 | ELSE
|
---|
500 | * The user does not care about relative accurately eigenvalues
|
---|
501 | IINFO = -1
|
---|
502 | ENDIF
|
---|
503 | * Set the splitting criterion
|
---|
504 | IF (IINFO.EQ.0) THEN
|
---|
505 | THRESH = EPS
|
---|
506 | ELSE
|
---|
507 | THRESH = -EPS
|
---|
508 | * relative accuracy is desired but T does not guarantee it
|
---|
509 | TRYRAC = .FALSE.
|
---|
510 | ENDIF
|
---|
511 | *
|
---|
512 | IF( TRYRAC ) THEN
|
---|
513 | * Copy original diagonal, needed to guarantee relative accuracy
|
---|
514 | CALL SCOPY(N,D,1,WORK(INDD),1)
|
---|
515 | ENDIF
|
---|
516 | * Store the squares of the offdiagonal values of T
|
---|
517 | DO 5 J = 1, N-1
|
---|
518 | WORK( INDE2+J-1 ) = E(J)**2
|
---|
519 | 5 CONTINUE
|
---|
520 |
|
---|
521 | * Set the tolerance parameters for bisection
|
---|
522 | IF( .NOT.WANTZ ) THEN
|
---|
523 | * SLARRE computes the eigenvalues to full precision.
|
---|
524 | RTOL1 = FOUR * EPS
|
---|
525 | RTOL2 = FOUR * EPS
|
---|
526 | ELSE
|
---|
527 | * SLARRE computes the eigenvalues to less than full precision.
|
---|
528 | * CLARRV will refine the eigenvalue approximations, and we only
|
---|
529 | * need less accurate initial bisection in SLARRE.
|
---|
530 | * Note: these settings do only affect the subset case and SLARRE
|
---|
531 | RTOL1 = MAX( SQRT(EPS)*5.0E-2, FOUR * EPS )
|
---|
532 | RTOL2 = MAX( SQRT(EPS)*5.0E-3, FOUR * EPS )
|
---|
533 | ENDIF
|
---|
534 | CALL SLARRE( RANGE, N, WL, WU, IIL, IIU, D, E,
|
---|
535 | $ WORK(INDE2), RTOL1, RTOL2, THRESH, NSPLIT,
|
---|
536 | $ IWORK( IINSPL ), M, W, WORK( INDERR ),
|
---|
537 | $ WORK( INDGP ), IWORK( IINDBL ),
|
---|
538 | $ IWORK( IINDW ), WORK( INDGRS ), PIVMIN,
|
---|
539 | $ WORK( INDWRK ), IWORK( IINDWK ), IINFO )
|
---|
540 | IF( IINFO.NE.0 ) THEN
|
---|
541 | INFO = 10 + ABS( IINFO )
|
---|
542 | RETURN
|
---|
543 | END IF
|
---|
544 | * Note that if RANGE .NE. 'V', SLARRE computes bounds on the desired
|
---|
545 | * part of the spectrum. All desired eigenvalues are contained in
|
---|
546 | * (WL,WU]
|
---|
547 |
|
---|
548 |
|
---|
549 | IF( WANTZ ) THEN
|
---|
550 | *
|
---|
551 | * Compute the desired eigenvectors corresponding to the computed
|
---|
552 | * eigenvalues
|
---|
553 | *
|
---|
554 | CALL CLARRV( N, WL, WU, D, E,
|
---|
555 | $ PIVMIN, IWORK( IINSPL ), M,
|
---|
556 | $ 1, M, MINRGP, RTOL1, RTOL2,
|
---|
557 | $ W, WORK( INDERR ), WORK( INDGP ), IWORK( IINDBL ),
|
---|
558 | $ IWORK( IINDW ), WORK( INDGRS ), Z, LDZ,
|
---|
559 | $ ISUPPZ, WORK( INDWRK ), IWORK( IINDWK ), IINFO )
|
---|
560 | IF( IINFO.NE.0 ) THEN
|
---|
561 | INFO = 20 + ABS( IINFO )
|
---|
562 | RETURN
|
---|
563 | END IF
|
---|
564 | ELSE
|
---|
565 | * SLARRE computes eigenvalues of the (shifted) root representation
|
---|
566 | * CLARRV returns the eigenvalues of the unshifted matrix.
|
---|
567 | * However, if the eigenvectors are not desired by the user, we need
|
---|
568 | * to apply the corresponding shifts from SLARRE to obtain the
|
---|
569 | * eigenvalues of the original matrix.
|
---|
570 | DO 20 J = 1, M
|
---|
571 | ITMP = IWORK( IINDBL+J-1 )
|
---|
572 | W( J ) = W( J ) + E( IWORK( IINSPL+ITMP-1 ) )
|
---|
573 | 20 CONTINUE
|
---|
574 | END IF
|
---|
575 | *
|
---|
576 |
|
---|
577 | IF ( TRYRAC ) THEN
|
---|
578 | * Refine computed eigenvalues so that they are relatively accurate
|
---|
579 | * with respect to the original matrix T.
|
---|
580 | IBEGIN = 1
|
---|
581 | WBEGIN = 1
|
---|
582 | DO 39 JBLK = 1, IWORK( IINDBL+M-1 )
|
---|
583 | IEND = IWORK( IINSPL+JBLK-1 )
|
---|
584 | IN = IEND - IBEGIN + 1
|
---|
585 | WEND = WBEGIN - 1
|
---|
586 | * check if any eigenvalues have to be refined in this block
|
---|
587 | 36 CONTINUE
|
---|
588 | IF( WEND.LT.M ) THEN
|
---|
589 | IF( IWORK( IINDBL+WEND ).EQ.JBLK ) THEN
|
---|
590 | WEND = WEND + 1
|
---|
591 | GO TO 36
|
---|
592 | END IF
|
---|
593 | END IF
|
---|
594 | IF( WEND.LT.WBEGIN ) THEN
|
---|
595 | IBEGIN = IEND + 1
|
---|
596 | GO TO 39
|
---|
597 | END IF
|
---|
598 |
|
---|
599 | OFFSET = IWORK(IINDW+WBEGIN-1)-1
|
---|
600 | IFIRST = IWORK(IINDW+WBEGIN-1)
|
---|
601 | ILAST = IWORK(IINDW+WEND-1)
|
---|
602 | RTOL2 = FOUR * EPS
|
---|
603 | CALL SLARRJ( IN,
|
---|
604 | $ WORK(INDD+IBEGIN-1), WORK(INDE2+IBEGIN-1),
|
---|
605 | $ IFIRST, ILAST, RTOL2, OFFSET, W(WBEGIN),
|
---|
606 | $ WORK( INDERR+WBEGIN-1 ),
|
---|
607 | $ WORK( INDWRK ), IWORK( IINDWK ), PIVMIN,
|
---|
608 | $ TNRM, IINFO )
|
---|
609 | IBEGIN = IEND + 1
|
---|
610 | WBEGIN = WEND + 1
|
---|
611 | 39 CONTINUE
|
---|
612 | ENDIF
|
---|
613 | *
|
---|
614 | * If matrix was scaled, then rescale eigenvalues appropriately.
|
---|
615 | *
|
---|
616 | IF( SCALE.NE.ONE ) THEN
|
---|
617 | CALL SSCAL( M, ONE / SCALE, W, 1 )
|
---|
618 | END IF
|
---|
619 | *
|
---|
620 | * If eigenvalues are not in increasing order, then sort them,
|
---|
621 | * possibly along with eigenvectors.
|
---|
622 | *
|
---|
623 | IF( NSPLIT.GT.1 ) THEN
|
---|
624 | IF( .NOT. WANTZ ) THEN
|
---|
625 | CALL SLASRT( 'I', M, W, IINFO )
|
---|
626 | IF( IINFO.NE.0 ) THEN
|
---|
627 | INFO = 3
|
---|
628 | RETURN
|
---|
629 | END IF
|
---|
630 | ELSE
|
---|
631 | DO 60 J = 1, M - 1
|
---|
632 | I = 0
|
---|
633 | TMP = W( J )
|
---|
634 | DO 50 JJ = J + 1, M
|
---|
635 | IF( W( JJ ).LT.TMP ) THEN
|
---|
636 | I = JJ
|
---|
637 | TMP = W( JJ )
|
---|
638 | END IF
|
---|
639 | 50 CONTINUE
|
---|
640 | IF( I.NE.0 ) THEN
|
---|
641 | W( I ) = W( J )
|
---|
642 | W( J ) = TMP
|
---|
643 | IF( WANTZ ) THEN
|
---|
644 | CALL CSWAP( N, Z( 1, I ), 1, Z( 1, J ), 1 )
|
---|
645 | ITMP = ISUPPZ( 2*I-1 )
|
---|
646 | ISUPPZ( 2*I-1 ) = ISUPPZ( 2*J-1 )
|
---|
647 | ISUPPZ( 2*J-1 ) = ITMP
|
---|
648 | ITMP = ISUPPZ( 2*I )
|
---|
649 | ISUPPZ( 2*I ) = ISUPPZ( 2*J )
|
---|
650 | ISUPPZ( 2*J ) = ITMP
|
---|
651 | END IF
|
---|
652 | END IF
|
---|
653 | 60 CONTINUE
|
---|
654 | END IF
|
---|
655 | ENDIF
|
---|
656 | *
|
---|
657 | *
|
---|
658 | WORK( 1 ) = LWMIN
|
---|
659 | IWORK( 1 ) = LIWMIN
|
---|
660 | RETURN
|
---|
661 | *
|
---|
662 | * End of CSTEMR
|
---|
663 | *
|
---|
664 | END
|
---|