1 *> \brief <b> CPOSVXX computes the solution to system of linear equations A * X = B for PO matrices</b>
3 * =========== DOCUMENTATION ===========
5 * Online html documentation available at
6 * http://www.netlib.org/lapack/explore-html/
9 *> Download CPOSVXX + dependencies
10 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/cposvxx.f">
12 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/cposvxx.f">
14 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/cposvxx.f">
21 * SUBROUTINE CPOSVXX( FACT, UPLO, N, NRHS, A, LDA, AF, LDAF, EQUED,
22 * S, B, LDB, X, LDX, RCOND, RPVGRW, BERR,
23 * N_ERR_BNDS, ERR_BNDS_NORM, ERR_BNDS_COMP,
24 * NPARAMS, PARAMS, WORK, RWORK, INFO )
26 * .. Scalar Arguments ..
27 * CHARACTER EQUED, FACT, UPLO
28 * INTEGER INFO, LDA, LDAF, LDB, LDX, N, NRHS, NPARAMS,
32 * .. Array Arguments ..
33 * COMPLEX A( LDA, * ), AF( LDAF, * ), B( LDB, * ),
34 * $ WORK( * ), X( LDX, * )
35 * REAL S( * ), PARAMS( * ), BERR( * ), RWORK( * ),
36 * $ ERR_BNDS_NORM( NRHS, * ),
37 * $ ERR_BNDS_COMP( NRHS, * )
46 *> CPOSVXX uses the Cholesky factorization A = U**T*U or A = L*L**T
47 *> to compute the solution to a complex system of linear equations
48 *> A * X = B, where A is an N-by-N symmetric positive definite matrix
49 *> and X and B are N-by-NRHS matrices.
51 *> If requested, both normwise and maximum componentwise error bounds
52 *> are returned. CPOSVXX will return a solution with a tiny
53 *> guaranteed error (O(eps) where eps is the working machine
54 *> precision) unless the matrix is very ill-conditioned, in which
55 *> case a warning is returned. Relevant condition numbers also are
56 *> calculated and returned.
58 *> CPOSVXX accepts user-provided factorizations and equilibration
59 *> factors; see the definitions of the FACT and EQUED options.
60 *> Solving with refinement and using a factorization from a previous
61 *> CPOSVXX call will also produce a solution with either O(eps)
62 *> errors or warnings, but we cannot make that claim for general
63 *> user-provided factorizations and equilibration factors if they
64 *> differ from what CPOSVXX would itself produce.
72 *> The following steps are performed:
74 *> 1. If FACT = 'E', real scaling factors are computed to equilibrate
77 *> diag(S)*A*diag(S) *inv(diag(S))*X = diag(S)*B
79 *> Whether or not the system will be equilibrated depends on the
80 *> scaling of the matrix A, but if equilibration is used, A is
81 *> overwritten by diag(S)*A*diag(S) and B by diag(S)*B.
83 *> 2. If FACT = 'N' or 'E', the Cholesky decomposition is used to
84 *> factor the matrix A (after equilibration if FACT = 'E') as
85 *> A = U**T* U, if UPLO = 'U', or
86 *> A = L * L**T, if UPLO = 'L',
87 *> where U is an upper triangular matrix and L is a lower triangular
90 *> 3. If the leading i-by-i principal minor is not positive definite,
91 *> then the routine returns with INFO = i. Otherwise, the factored
92 *> form of A is used to estimate the condition number of the matrix
93 *> A (see argument RCOND). If the reciprocal of the condition number
94 *> is less than machine precision, the routine still goes on to solve
95 *> for X and compute error bounds as described below.
97 *> 4. The system of equations is solved for X using the factored form
100 *> 5. By default (unless PARAMS(LA_LINRX_ITREF_I) is set to zero),
101 *> the routine will use iterative refinement to try to get a small
102 *> error and error bounds. Refinement calculates the residual to at
103 *> least twice the working precision.
105 *> 6. If equilibration was used, the matrix X is premultiplied by
106 *> diag(S) so that it solves the original system before
114 *> Some optional parameters are bundled in the PARAMS array. These
115 *> settings determine how refinement is performed, but often the
116 *> defaults are acceptable. If the defaults are acceptable, users
117 *> can pass NPARAMS = 0 which prevents the source code from accessing
118 *> the PARAMS argument.
123 *> FACT is CHARACTER*1
124 *> Specifies whether or not the factored form of the matrix A is
125 *> supplied on entry, and if not, whether the matrix A should be
126 *> equilibrated before it is factored.
127 *> = 'F': On entry, AF contains the factored form of A.
128 *> If EQUED is not 'N', the matrix A has been
129 *> equilibrated with scaling factors given by S.
130 *> A and AF are not modified.
131 *> = 'N': The matrix A will be copied to AF and factored.
132 *> = 'E': The matrix A will be equilibrated if necessary, then
133 *> copied to AF and factored.
138 *> UPLO is CHARACTER*1
139 *> = 'U': Upper triangle of A is stored;
140 *> = 'L': Lower triangle of A is stored.
146 *> The number of linear equations, i.e., the order of the
153 *> The number of right hand sides, i.e., the number of columns
154 *> of the matrices B and X. NRHS >= 0.
159 *> A is COMPLEX array, dimension (LDA,N)
160 *> On entry, the symmetric matrix A, except if FACT = 'F' and EQUED =
161 *> 'Y', then A must contain the equilibrated matrix
162 *> diag(S)*A*diag(S). If UPLO = 'U', the leading N-by-N upper
163 *> triangular part of A contains the upper triangular part of the
164 *> matrix A, and the strictly lower triangular part of A is not
165 *> referenced. If UPLO = 'L', the leading N-by-N lower triangular
166 *> part of A contains the lower triangular part of the matrix A, and
167 *> the strictly upper triangular part of A is not referenced. A is
168 *> not modified if FACT = 'F' or 'N', or if FACT = 'E' and EQUED =
171 *> On exit, if FACT = 'E' and EQUED = 'Y', A is overwritten by
172 *> diag(S)*A*diag(S).
178 *> The leading dimension of the array A. LDA >= max(1,N).
183 *> AF is COMPLEX array, dimension (LDAF,N)
184 *> If FACT = 'F', then AF is an input argument and on entry
185 *> contains the triangular factor U or L from the Cholesky
186 *> factorization A = U**T*U or A = L*L**T, in the same storage
187 *> format as A. If EQUED .ne. 'N', then AF is the factored
188 *> form of the equilibrated matrix diag(S)*A*diag(S).
190 *> If FACT = 'N', then AF is an output argument and on exit
191 *> returns the triangular factor U or L from the Cholesky
192 *> factorization A = U**T*U or A = L*L**T of the original
195 *> If FACT = 'E', then AF is an output argument and on exit
196 *> returns the triangular factor U or L from the Cholesky
197 *> factorization A = U**T*U or A = L*L**T of the equilibrated
198 *> matrix A (see the description of A for the form of the
199 *> equilibrated matrix).
205 *> The leading dimension of the array AF. LDAF >= max(1,N).
208 *> \param[in,out] EQUED
210 *> EQUED is CHARACTER*1
211 *> Specifies the form of equilibration that was done.
212 *> = 'N': No equilibration (always true if FACT = 'N').
213 *> = 'Y': Both row and column equilibration, i.e., A has been
214 *> replaced by diag(S) * A * diag(S).
215 *> EQUED is an input argument if FACT = 'F'; otherwise, it is an
221 *> S is REAL array, dimension (N)
222 *> The row scale factors for A. If EQUED = 'Y', A is multiplied on
223 *> the left and right by diag(S). S is an input argument if FACT =
224 *> 'F'; otherwise, S is an output argument. If FACT = 'F' and EQUED
225 *> = 'Y', each element of S must be positive. If S is output, each
226 *> element of S is a power of the radix. If S is input, each element
227 *> of S should be a power of the radix to ensure a reliable solution
228 *> and error estimates. Scaling by powers of the radix does not cause
229 *> rounding errors unless the result underflows or overflows.
230 *> Rounding errors during scaling lead to refining with a matrix that
231 *> is not equivalent to the input matrix, producing error estimates
232 *> that may not be reliable.
237 *> B is COMPLEX array, dimension (LDB,NRHS)
238 *> On entry, the N-by-NRHS right hand side matrix B.
240 *> if EQUED = 'N', B is not modified;
241 *> if EQUED = 'Y', B is overwritten by diag(S)*B;
247 *> The leading dimension of the array B. LDB >= max(1,N).
252 *> X is COMPLEX array, dimension (LDX,NRHS)
253 *> If INFO = 0, the N-by-NRHS solution matrix X to the original
254 *> system of equations. Note that A and B are modified on exit if
255 *> EQUED .ne. 'N', and the solution to the equilibrated system is
262 *> The leading dimension of the array X. LDX >= max(1,N).
268 *> Reciprocal scaled condition number. This is an estimate of the
269 *> reciprocal Skeel condition number of the matrix A after
270 *> equilibration (if done). If this is less than the machine
271 *> precision (in particular, if it is zero), the matrix is singular
272 *> to working precision. Note that the error may still be small even
273 *> if this number is very small and the matrix appears ill-
277 *> \param[out] RPVGRW
280 *> Reciprocal pivot growth. On exit, this contains the reciprocal
281 *> pivot growth factor norm(A)/norm(U). The "max absolute element"
282 *> norm is used. If this is much less than 1, then the stability of
283 *> the LU factorization of the (equilibrated) matrix A could be poor.
284 *> This also means that the solution X, estimated condition numbers,
285 *> and error bounds could be unreliable. If factorization fails with
286 *> 0<INFO<=N, then this contains the reciprocal pivot growth factor
287 *> for the leading INFO columns of A.
292 *> BERR is REAL array, dimension (NRHS)
293 *> Componentwise relative backward error. This is the
294 *> componentwise relative backward error of each solution vector X(j)
295 *> (i.e., the smallest relative change in any element of A or B that
296 *> makes X(j) an exact solution).
299 *> \param[in] N_ERR_BNDS
301 *> N_ERR_BNDS is INTEGER
302 *> Number of error bounds to return for each right hand side
303 *> and each type (normwise or componentwise). See ERR_BNDS_NORM and
304 *> ERR_BNDS_COMP below.
307 *> \param[out] ERR_BNDS_NORM
309 *> ERR_BNDS_NORM is REAL array, dimension (NRHS, N_ERR_BNDS)
310 *> For each right-hand side, this array contains information about
311 *> various error bounds and condition numbers corresponding to the
312 *> normwise relative error, which is defined as follows:
314 *> Normwise relative error in the ith solution vector:
315 *> max_j (abs(XTRUE(j,i) - X(j,i)))
316 *> ------------------------------
319 *> The array is indexed by the type of error information as described
320 *> below. There currently are up to three pieces of information
323 *> The first index in ERR_BNDS_NORM(i,:) corresponds to the ith
326 *> The second index in ERR_BNDS_NORM(:,err) contains the following
328 *> err = 1 "Trust/don't trust" boolean. Trust the answer if the
329 *> reciprocal condition number is less than the threshold
330 *> sqrt(n) * slamch('Epsilon').
332 *> err = 2 "Guaranteed" error bound: The estimated forward error,
333 *> almost certainly within a factor of 10 of the true error
334 *> so long as the next entry is greater than the threshold
335 *> sqrt(n) * slamch('Epsilon'). This error bound should only
336 *> be trusted if the previous boolean is true.
338 *> err = 3 Reciprocal condition number: Estimated normwise
339 *> reciprocal condition number. Compared with the threshold
340 *> sqrt(n) * slamch('Epsilon') to determine if the error
341 *> estimate is "guaranteed". These reciprocal condition
342 *> numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some
343 *> appropriately scaled matrix Z.
344 *> Let Z = S*A, where S scales each row by a power of the
345 *> radix so all absolute row sums of Z are approximately 1.
347 *> See Lapack Working Note 165 for further details and extra
351 *> \param[out] ERR_BNDS_COMP
353 *> ERR_BNDS_COMP is REAL array, dimension (NRHS, N_ERR_BNDS)
354 *> For each right-hand side, this array contains information about
355 *> various error bounds and condition numbers corresponding to the
356 *> componentwise relative error, which is defined as follows:
358 *> Componentwise relative error in the ith solution vector:
359 *> abs(XTRUE(j,i) - X(j,i))
360 *> max_j ----------------------
363 *> The array is indexed by the right-hand side i (on which the
364 *> componentwise relative error depends), and the type of error
365 *> information as described below. There currently are up to three
366 *> pieces of information returned for each right-hand side. If
367 *> componentwise accuracy is not requested (PARAMS(3) = 0.0), then
368 *> ERR_BNDS_COMP is not accessed. If N_ERR_BNDS .LT. 3, then at most
369 *> the first (:,N_ERR_BNDS) entries are returned.
371 *> The first index in ERR_BNDS_COMP(i,:) corresponds to the ith
374 *> The second index in ERR_BNDS_COMP(:,err) contains the following
376 *> err = 1 "Trust/don't trust" boolean. Trust the answer if the
377 *> reciprocal condition number is less than the threshold
378 *> sqrt(n) * slamch('Epsilon').
380 *> err = 2 "Guaranteed" error bound: The estimated forward error,
381 *> almost certainly within a factor of 10 of the true error
382 *> so long as the next entry is greater than the threshold
383 *> sqrt(n) * slamch('Epsilon'). This error bound should only
384 *> be trusted if the previous boolean is true.
386 *> err = 3 Reciprocal condition number: Estimated componentwise
387 *> reciprocal condition number. Compared with the threshold
388 *> sqrt(n) * slamch('Epsilon') to determine if the error
389 *> estimate is "guaranteed". These reciprocal condition
390 *> numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some
391 *> appropriately scaled matrix Z.
392 *> Let Z = S*(A*diag(x)), where x is the solution for the
393 *> current right-hand side and S scales each row of
394 *> A*diag(x) by a power of the radix so all absolute row
395 *> sums of Z are approximately 1.
397 *> See Lapack Working Note 165 for further details and extra
401 *> \param[in] NPARAMS
403 *> NPARAMS is INTEGER
404 *> Specifies the number of parameters set in PARAMS. If .LE. 0, the
405 *> PARAMS array is never referenced and default values are used.
408 *> \param[in,out] PARAMS
410 *> PARAMS is REAL array, dimension NPARAMS
411 *> Specifies algorithm parameters. If an entry is .LT. 0.0, then
412 *> that entry will be filled with default value used for that
413 *> parameter. Only positions up to NPARAMS are accessed; defaults
414 *> are used for higher-numbered parameters.
416 *> PARAMS(LA_LINRX_ITREF_I = 1) : Whether to perform iterative
417 *> refinement or not.
419 *> = 0.0 : No refinement is performed, and no error bounds are
421 *> = 1.0 : Use the double-precision refinement algorithm,
422 *> possibly with doubled-single computations if the
423 *> compilation environment does not support DOUBLE
425 *> (other values are reserved for future use)
427 *> PARAMS(LA_LINRX_ITHRESH_I = 2) : Maximum number of residual
428 *> computations allowed for refinement.
430 *> Aggressive: Set to 100 to permit convergence using approximate
431 *> factorizations or factorizations other than LU. If
432 *> the factorization uses a technique other than
433 *> Gaussian elimination, the guarantees in
434 *> err_bnds_norm and err_bnds_comp may no longer be
437 *> PARAMS(LA_LINRX_CWISE_I = 3) : Flag determining if the code
438 *> will attempt to find a solution with small componentwise
439 *> relative error in the double-precision algorithm. Positive
440 *> is true, 0.0 is false.
441 *> Default: 1.0 (attempt componentwise convergence)
446 *> WORK is COMPLEX array, dimension (2*N)
451 *> RWORK is REAL array, dimension (2*N)
457 *> = 0: Successful exit. The solution to every right-hand side is
459 *> < 0: If INFO = -i, the i-th argument had an illegal value
460 *> > 0 and <= N: U(INFO,INFO) is exactly zero. The factorization
461 *> has been completed, but the factor U is exactly singular, so
462 *> the solution and error bounds could not be computed. RCOND = 0
464 *> = N+J: The solution corresponding to the Jth right-hand side is
465 *> not guaranteed. The solutions corresponding to other right-
466 *> hand sides K with K > J may not be guaranteed as well, but
467 *> only the first such right-hand side is reported. If a small
468 *> componentwise error is not requested (PARAMS(3) = 0.0) then
469 *> the Jth right-hand side is the first with a normwise error
470 *> bound that is not guaranteed (the smallest J such
471 *> that ERR_BNDS_NORM(J,1) = 0.0). By default (PARAMS(3) = 1.0)
472 *> the Jth right-hand side is the first with either a normwise or
473 *> componentwise error bound that is not guaranteed (the smallest
474 *> J such that either ERR_BNDS_NORM(J,1) = 0.0 or
475 *> ERR_BNDS_COMP(J,1) = 0.0). See the definition of
476 *> ERR_BNDS_NORM(:,1) and ERR_BNDS_COMP(:,1). To get information
477 *> about all of the right-hand sides check ERR_BNDS_NORM or
484 *> \author Univ. of Tennessee
485 *> \author Univ. of California Berkeley
486 *> \author Univ. of Colorado Denver
491 *> \ingroup complexPOsolve
493 * =====================================================================
494 SUBROUTINE CPOSVXX( FACT, UPLO, N, NRHS, A, LDA, AF, LDAF, EQUED,
495 $ S, B, LDB, X, LDX, RCOND, RPVGRW, BERR,
496 $ N_ERR_BNDS, ERR_BNDS_NORM, ERR_BNDS_COMP,
497 $ NPARAMS, PARAMS, WORK, RWORK, INFO )
499 * -- LAPACK driver routine (version 3.4.1) --
500 * -- LAPACK is a software package provided by Univ. of Tennessee, --
501 * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
504 * .. Scalar Arguments ..
505 CHARACTER EQUED, FACT, UPLO
506 INTEGER INFO, LDA, LDAF, LDB, LDX, N, NRHS, NPARAMS,
510 * .. Array Arguments ..
511 COMPLEX A( LDA, * ), AF( LDAF, * ), B( LDB, * ),
512 $ WORK( * ), X( LDX, * )
513 REAL S( * ), PARAMS( * ), BERR( * ), RWORK( * ),
514 $ ERR_BNDS_NORM( NRHS, * ),
515 $ ERR_BNDS_COMP( NRHS, * )
518 * ==================================================================
522 PARAMETER ( ZERO = 0.0E+0, ONE = 1.0E+0 )
523 INTEGER FINAL_NRM_ERR_I, FINAL_CMP_ERR_I, BERR_I
524 INTEGER RCOND_I, NRM_RCOND_I, NRM_ERR_I, CMP_RCOND_I
525 INTEGER CMP_ERR_I, PIV_GROWTH_I
526 PARAMETER ( FINAL_NRM_ERR_I = 1, FINAL_CMP_ERR_I = 2,
528 PARAMETER ( RCOND_I = 4, NRM_RCOND_I = 5, NRM_ERR_I = 6 )
529 PARAMETER ( CMP_RCOND_I = 7, CMP_ERR_I = 8,
532 * .. Local Scalars ..
533 LOGICAL EQUIL, NOFACT, RCEQU
535 REAL AMAX, BIGNUM, SMIN, SMAX, SCOND, SMLNUM
537 * .. External Functions ..
538 EXTERNAL LSAME, SLAMCH, CLA_PORPVGRW
540 REAL SLAMCH, CLA_PORPVGRW
542 * .. External Subroutines ..
543 EXTERNAL CPOCON, CPOEQUB, CPOTRF, CPOTRS, CLACPY,
544 $ CLAQHE, XERBLA, CLASCL2, CPORFSX
546 * .. Intrinsic Functions ..
549 * .. Executable Statements ..
552 NOFACT = LSAME( FACT, 'N' )
553 EQUIL = LSAME( FACT, 'E' )
554 SMLNUM = SLAMCH( 'Safe minimum' )
555 BIGNUM = ONE / SMLNUM
556 IF( NOFACT .OR. EQUIL ) THEN
560 RCEQU = LSAME( EQUED, 'Y' )
563 * Default is failure. If an input parameter is wrong or
564 * factorization fails, make everything look horrible. Only the
565 * pivot growth is set here, the rest is initialized in CPORFSX.
569 * Test the input parameters. PARAMS is not tested until CPORFSX.
571 IF( .NOT.NOFACT .AND. .NOT.EQUIL .AND. .NOT.
572 $ LSAME( FACT, 'F' ) ) THEN
574 ELSE IF( .NOT.LSAME( UPLO, 'U' ) .AND.
575 $ .NOT.LSAME( UPLO, 'L' ) ) THEN
577 ELSE IF( N.LT.0 ) THEN
579 ELSE IF( NRHS.LT.0 ) THEN
581 ELSE IF( LDA.LT.MAX( 1, N ) ) THEN
583 ELSE IF( LDAF.LT.MAX( 1, N ) ) THEN
585 ELSE IF( LSAME( FACT, 'F' ) .AND. .NOT.
586 $ ( RCEQU .OR. LSAME( EQUED, 'N' ) ) ) THEN
593 SMIN = MIN( SMIN, S( J ) )
594 SMAX = MAX( SMAX, S( J ) )
596 IF( SMIN.LE.ZERO ) THEN
598 ELSE IF( N.GT.0 ) THEN
599 SCOND = MAX( SMIN, SMLNUM ) / MIN( SMAX, BIGNUM )
605 IF( LDB.LT.MAX( 1, N ) ) THEN
607 ELSE IF( LDX.LT.MAX( 1, N ) ) THEN
614 CALL XERBLA( 'CPOSVXX', -INFO )
620 * Compute row and column scalings to equilibrate the matrix A.
622 CALL CPOEQUB( N, A, LDA, S, SCOND, AMAX, INFEQU )
623 IF( INFEQU.EQ.0 ) THEN
625 * Equilibrate the matrix.
627 CALL CLAQHE( UPLO, N, A, LDA, S, SCOND, AMAX, EQUED )
628 RCEQU = LSAME( EQUED, 'Y' )
632 * Scale the right-hand side.
634 IF( RCEQU ) CALL CLASCL2( N, NRHS, S, B, LDB )
636 IF( NOFACT .OR. EQUIL ) THEN
638 * Compute the Cholesky factorization of A.
640 CALL CLACPY( UPLO, N, N, A, LDA, AF, LDAF )
641 CALL CPOTRF( UPLO, N, AF, LDAF, INFO )
643 * Return if INFO is non-zero.
647 * Pivot in column INFO is exactly 0
648 * Compute the reciprocal pivot growth factor of the
649 * leading rank-deficient INFO columns of A.
651 RPVGRW = CLA_PORPVGRW( UPLO, N, A, LDA, AF, LDAF, RWORK )
656 * Compute the reciprocal pivot growth factor RPVGRW.
658 RPVGRW = CLA_PORPVGRW( UPLO, N, A, LDA, AF, LDAF, RWORK )
660 * Compute the solution matrix X.
662 CALL CLACPY( 'Full', N, NRHS, B, LDB, X, LDX )
663 CALL CPOTRS( UPLO, N, NRHS, AF, LDAF, X, LDX, INFO )
665 * Use iterative refinement to improve the computed solution and
666 * compute error bounds and backward error estimates for it.
668 CALL CPORFSX( UPLO, EQUED, N, NRHS, A, LDA, AF, LDAF,
669 $ S, B, LDB, X, LDX, RCOND, BERR, N_ERR_BNDS, ERR_BNDS_NORM,
670 $ ERR_BNDS_COMP, NPARAMS, PARAMS, WORK, RWORK, INFO )
676 CALL CLASCL2( N, NRHS, S, X, LDX )