1 *> \brief <b> SPBSVX computes the solution to system of linear equations A * X = B for OTHER matrices</b>
3 * =========== DOCUMENTATION ===========
5 * Online html documentation available at
6 * http://www.netlib.org/lapack/explore-html/
9 *> Download SPBSVX + dependencies
10 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/spbsvx.f">
12 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/spbsvx.f">
14 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/spbsvx.f">
21 * SUBROUTINE SPBSVX( FACT, UPLO, N, KD, NRHS, AB, LDAB, AFB, LDAFB,
22 * EQUED, S, B, LDB, X, LDX, RCOND, FERR, BERR,
25 * .. Scalar Arguments ..
26 * CHARACTER EQUED, FACT, UPLO
27 * INTEGER INFO, KD, LDAB, LDAFB, LDB, LDX, N, NRHS
30 * .. Array Arguments ..
32 * REAL AB( LDAB, * ), AFB( LDAFB, * ), B( LDB, * ),
33 * $ BERR( * ), FERR( * ), S( * ), WORK( * ),
43 *> SPBSVX uses the Cholesky factorization A = U**T*U or A = L*L**T to
44 *> compute the solution to a real system of linear equations
46 *> where A is an N-by-N symmetric positive definite band matrix and X
47 *> and B are N-by-NRHS matrices.
49 *> Error bounds on the solution and a condition estimate are also
58 *> The following steps are performed:
60 *> 1. If FACT = 'E', real scaling factors are computed to equilibrate
62 *> diag(S) * A * diag(S) * inv(diag(S)) * X = diag(S) * B
63 *> Whether or not the system will be equilibrated depends on the
64 *> scaling of the matrix A, but if equilibration is used, A is
65 *> overwritten by diag(S)*A*diag(S) and B by diag(S)*B.
67 *> 2. If FACT = 'N' or 'E', the Cholesky decomposition is used to
68 *> factor the matrix A (after equilibration if FACT = 'E') as
69 *> A = U**T * U, if UPLO = 'U', or
70 *> A = L * L**T, if UPLO = 'L',
71 *> where U is an upper triangular band matrix, and L is a lower
72 *> triangular band matrix.
74 *> 3. If the leading i-by-i principal minor is not positive definite,
75 *> then the routine returns with INFO = i. Otherwise, the factored
76 *> form of A is used to estimate the condition number of the matrix
77 *> A. If the reciprocal of the condition number is less than machine
78 *> precision, INFO = N+1 is returned as a warning, but the routine
79 *> still goes on to solve for X and compute error bounds as
82 *> 4. The system of equations is solved for X using the factored form
85 *> 5. Iterative refinement is applied to improve the computed solution
86 *> matrix and calculate error bounds and backward error estimates
89 *> 6. If equilibration was used, the matrix X is premultiplied by
90 *> diag(S) so that it solves the original system before
99 *> FACT is CHARACTER*1
100 *> Specifies whether or not the factored form of the matrix A is
101 *> supplied on entry, and if not, whether the matrix A should be
102 *> equilibrated before it is factored.
103 *> = 'F': On entry, AFB contains the factored form of A.
104 *> If EQUED = 'Y', the matrix A has been equilibrated
105 *> with scaling factors given by S. AB and AFB will not
107 *> = 'N': The matrix A will be copied to AFB and factored.
108 *> = 'E': The matrix A will be equilibrated if necessary, then
109 *> copied to AFB and factored.
114 *> UPLO is CHARACTER*1
115 *> = 'U': Upper triangle of A is stored;
116 *> = 'L': Lower triangle of A is stored.
122 *> The number of linear equations, i.e., the order of the
129 *> The number of superdiagonals of the matrix A if UPLO = 'U',
130 *> or the number of subdiagonals if UPLO = 'L'. KD >= 0.
136 *> The number of right-hand sides, i.e., the number of columns
137 *> of the matrices B and X. NRHS >= 0.
142 *> AB is REAL array, dimension (LDAB,N)
143 *> On entry, the upper or lower triangle of the symmetric band
144 *> matrix A, stored in the first KD+1 rows of the array, except
145 *> if FACT = 'F' and EQUED = 'Y', then A must contain the
146 *> equilibrated matrix diag(S)*A*diag(S). The j-th column of A
147 *> is stored in the j-th column of the array AB as follows:
148 *> if UPLO = 'U', AB(KD+1+i-j,j) = A(i,j) for max(1,j-KD)<=i<=j;
149 *> if UPLO = 'L', AB(1+i-j,j) = A(i,j) for j<=i<=min(N,j+KD).
150 *> See below for further details.
152 *> On exit, if FACT = 'E' and EQUED = 'Y', A is overwritten by
153 *> diag(S)*A*diag(S).
159 *> The leading dimension of the array A. LDAB >= KD+1.
162 *> \param[in,out] AFB
164 *> AFB is REAL array, dimension (LDAFB,N)
165 *> If FACT = 'F', then AFB is an input argument and on entry
166 *> contains the triangular factor U or L from the Cholesky
167 *> factorization A = U**T*U or A = L*L**T of the band matrix
168 *> A, in the same storage format as A (see AB). If EQUED = 'Y',
169 *> then AFB is the factored form of the equilibrated matrix A.
171 *> If FACT = 'N', then AFB is an output argument and on exit
172 *> returns the triangular factor U or L from the Cholesky
173 *> factorization A = U**T*U or A = L*L**T.
175 *> If FACT = 'E', then AFB is an output argument and on exit
176 *> returns the triangular factor U or L from the Cholesky
177 *> factorization A = U**T*U or A = L*L**T of the equilibrated
178 *> matrix A (see the description of A for the form of the
179 *> equilibrated matrix).
185 *> The leading dimension of the array AFB. LDAFB >= KD+1.
188 *> \param[in,out] EQUED
190 *> EQUED is CHARACTER*1
191 *> Specifies the form of equilibration that was done.
192 *> = 'N': No equilibration (always true if FACT = 'N').
193 *> = 'Y': Equilibration was done, i.e., A has been replaced by
194 *> diag(S) * A * diag(S).
195 *> EQUED is an input argument if FACT = 'F'; otherwise, it is an
201 *> S is REAL array, dimension (N)
202 *> The scale factors for A; not accessed if EQUED = 'N'. S is
203 *> an input argument if FACT = 'F'; otherwise, S is an output
204 *> argument. If FACT = 'F' and EQUED = 'Y', each element of S
210 *> B is REAL array, dimension (LDB,NRHS)
211 *> On entry, the N-by-NRHS right hand side matrix B.
212 *> On exit, if EQUED = 'N', B is not modified; if EQUED = 'Y',
213 *> B is overwritten by diag(S) * B.
219 *> The leading dimension of the array B. LDB >= max(1,N).
224 *> X is REAL array, dimension (LDX,NRHS)
225 *> If INFO = 0 or INFO = N+1, the N-by-NRHS solution matrix X to
226 *> the original system of equations. Note that if EQUED = 'Y',
227 *> A and B are modified on exit, and the solution to the
228 *> equilibrated system is inv(diag(S))*X.
234 *> The leading dimension of the array X. LDX >= max(1,N).
240 *> The estimate of the reciprocal condition number of the matrix
241 *> A after equilibration (if done). If RCOND is less than the
242 *> machine precision (in particular, if RCOND = 0), the matrix
243 *> is singular to working precision. This condition is
244 *> indicated by a return code of INFO > 0.
249 *> FERR is REAL array, dimension (NRHS)
250 *> The estimated forward error bound for each solution vector
251 *> X(j) (the j-th column of the solution matrix X).
252 *> If XTRUE is the true solution corresponding to X(j), FERR(j)
253 *> is an estimated upper bound for the magnitude of the largest
254 *> element in (X(j) - XTRUE) divided by the magnitude of the
255 *> largest element in X(j). The estimate is as reliable as
256 *> the estimate for RCOND, and is almost always a slight
257 *> overestimate of the true error.
262 *> BERR is REAL array, dimension (NRHS)
263 *> The componentwise relative backward error of each solution
264 *> vector X(j) (i.e., the smallest relative change in
265 *> any element of A or B that makes X(j) an exact solution).
270 *> WORK is REAL array, dimension (3*N)
275 *> IWORK is INTEGER array, dimension (N)
281 *> = 0: successful exit
282 *> < 0: if INFO = -i, the i-th argument had an illegal value
283 *> > 0: if INFO = i, and i is
284 *> <= N: the leading minor of order i of A is
285 *> not positive definite, so the factorization
286 *> could not be completed, and the solution has not
287 *> been computed. RCOND = 0 is returned.
288 *> = N+1: U is nonsingular, but RCOND is less than machine
289 *> precision, meaning that the matrix is singular
290 *> to working precision. Nevertheless, the
291 *> solution and error bounds are computed because
292 *> there are a number of situations where the
293 *> computed solution can be more accurate than the
294 *> value of RCOND would suggest.
300 *> \author Univ. of Tennessee
301 *> \author Univ. of California Berkeley
302 *> \author Univ. of Colorado Denver
307 *> \ingroup realOTHERsolve
309 *> \par Further Details:
310 * =====================
314 *> The band storage scheme is illustrated by the following example, when
315 *> N = 6, KD = 2, and UPLO = 'U':
317 *> Two-dimensional storage of the symmetric matrix A:
324 *> (aij=conjg(aji)) a66
326 *> Band storage of the upper triangle of A:
328 *> * * a13 a24 a35 a46
329 *> * a12 a23 a34 a45 a56
330 *> a11 a22 a33 a44 a55 a66
332 *> Similarly, if UPLO = 'L' the format of A is as follows:
334 *> a11 a22 a33 a44 a55 a66
335 *> a21 a32 a43 a54 a65 *
336 *> a31 a42 a53 a64 * *
338 *> Array elements marked * are not used by the routine.
341 * =====================================================================
342 SUBROUTINE SPBSVX( FACT, UPLO, N, KD, NRHS, AB, LDAB, AFB, LDAFB,
343 $ EQUED, S, B, LDB, X, LDX, RCOND, FERR, BERR,
344 $ WORK, IWORK, INFO )
346 * -- LAPACK driver routine (version 3.4.1) --
347 * -- LAPACK is a software package provided by Univ. of Tennessee, --
348 * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
351 * .. Scalar Arguments ..
352 CHARACTER EQUED, FACT, UPLO
353 INTEGER INFO, KD, LDAB, LDAFB, LDB, LDX, N, NRHS
356 * .. Array Arguments ..
358 REAL AB( LDAB, * ), AFB( LDAFB, * ), B( LDB, * ),
359 $ BERR( * ), FERR( * ), S( * ), WORK( * ),
363 * =====================================================================
367 PARAMETER ( ZERO = 0.0E+0, ONE = 1.0E+0 )
369 * .. Local Scalars ..
370 LOGICAL EQUIL, NOFACT, RCEQU, UPPER
371 INTEGER I, INFEQU, J, J1, J2
372 REAL AMAX, ANORM, BIGNUM, SCOND, SMAX, SMIN, SMLNUM
374 * .. External Functions ..
377 EXTERNAL LSAME, SLAMCH, SLANSB
379 * .. External Subroutines ..
380 EXTERNAL SCOPY, SLACPY, SLAQSB, SPBCON, SPBEQU, SPBRFS,
381 $ SPBTRF, SPBTRS, XERBLA
383 * .. Intrinsic Functions ..
386 * .. Executable Statements ..
389 NOFACT = LSAME( FACT, 'N' )
390 EQUIL = LSAME( FACT, 'E' )
391 UPPER = LSAME( UPLO, 'U' )
392 IF( NOFACT .OR. EQUIL ) THEN
396 RCEQU = LSAME( EQUED, 'Y' )
397 SMLNUM = SLAMCH( 'Safe minimum' )
398 BIGNUM = ONE / SMLNUM
401 * Test the input parameters.
403 IF( .NOT.NOFACT .AND. .NOT.EQUIL .AND. .NOT.LSAME( FACT, 'F' ) )
406 ELSE IF( .NOT.UPPER .AND. .NOT.LSAME( UPLO, 'L' ) ) THEN
408 ELSE IF( N.LT.0 ) THEN
410 ELSE IF( KD.LT.0 ) THEN
412 ELSE IF( NRHS.LT.0 ) THEN
414 ELSE IF( LDAB.LT.KD+1 ) THEN
416 ELSE IF( LDAFB.LT.KD+1 ) THEN
418 ELSE IF( LSAME( FACT, 'F' ) .AND. .NOT.
419 $ ( RCEQU .OR. LSAME( EQUED, 'N' ) ) ) THEN
426 SMIN = MIN( SMIN, S( J ) )
427 SMAX = MAX( SMAX, S( J ) )
429 IF( SMIN.LE.ZERO ) THEN
431 ELSE IF( N.GT.0 ) THEN
432 SCOND = MAX( SMIN, SMLNUM ) / MIN( SMAX, BIGNUM )
438 IF( LDB.LT.MAX( 1, N ) ) THEN
440 ELSE IF( LDX.LT.MAX( 1, N ) ) THEN
447 CALL XERBLA( 'SPBSVX', -INFO )
453 * Compute row and column scalings to equilibrate the matrix A.
455 CALL SPBEQU( UPLO, N, KD, AB, LDAB, S, SCOND, AMAX, INFEQU )
456 IF( INFEQU.EQ.0 ) THEN
458 * Equilibrate the matrix.
460 CALL SLAQSB( UPLO, N, KD, AB, LDAB, S, SCOND, AMAX, EQUED )
461 RCEQU = LSAME( EQUED, 'Y' )
465 * Scale the right-hand side.
470 B( I, J ) = S( I )*B( I, J )
475 IF( NOFACT .OR. EQUIL ) THEN
477 * Compute the Cholesky factorization A = U**T *U or A = L*L**T.
482 CALL SCOPY( J-J1+1, AB( KD+1-J+J1, J ), 1,
483 $ AFB( KD+1-J+J1, J ), 1 )
488 CALL SCOPY( J2-J+1, AB( 1, J ), 1, AFB( 1, J ), 1 )
492 CALL SPBTRF( UPLO, N, KD, AFB, LDAFB, INFO )
494 * Return if INFO is non-zero.
502 * Compute the norm of the matrix A.
504 ANORM = SLANSB( '1', UPLO, N, KD, AB, LDAB, WORK )
506 * Compute the reciprocal of the condition number of A.
508 CALL SPBCON( UPLO, N, KD, AFB, LDAFB, ANORM, RCOND, WORK, IWORK,
511 * Compute the solution matrix X.
513 CALL SLACPY( 'Full', N, NRHS, B, LDB, X, LDX )
514 CALL SPBTRS( UPLO, N, KD, NRHS, AFB, LDAFB, X, LDX, INFO )
516 * Use iterative refinement to improve the computed solution and
517 * compute error bounds and backward error estimates for it.
519 CALL SPBRFS( UPLO, N, KD, NRHS, AB, LDAB, AFB, LDAFB, B, LDB, X,
520 $ LDX, FERR, BERR, WORK, IWORK, INFO )
522 * Transform the solution matrix X to a solution of the original
528 X( I, J ) = S( I )*X( I, J )
532 FERR( J ) = FERR( J ) / SCOND
536 * Set INFO = N+1 if the matrix is singular to working precision.
538 IF( RCOND.LT.SLAMCH( 'Epsilon' ) )