LAPACK: SRC/dpbrfs.f Source File

Go to the documentation of this file.
00001       SUBROUTINE DPBRFS( UPLO, N, KD, NRHS, AB, LDAB, AFB, LDAFB, B,
00002      $                   LDB, X, LDX, FERR, BERR, WORK, IWORK, INFO )
00003 *
00004 *  -- LAPACK routine (version 3.2) --
00005 *  -- LAPACK is a software package provided by Univ. of Tennessee,    --
00006 *  -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
00007 *     November 2006
00008 *
00009 *     Modified to call DLACN2 in place of DLACON, 5 Feb 03, SJH.
00010 *
00011 *     .. Scalar Arguments ..
00012       CHARACTER          UPLO
00013       INTEGER            INFO, KD, LDAB, LDAFB, LDB, LDX, N, NRHS
00014 *     ..
00015 *     .. Array Arguments ..
00016       INTEGER            IWORK( * )
00017       DOUBLE PRECISION   AB( LDAB, * ), AFB( LDAFB, * ), B( LDB, * ),
00018      $                   BERR( * ), FERR( * ), WORK( * ), X( LDX, * )
00019 *     ..
00020 *
00021 *  Purpose
00022 *  =======
00023 *
00024 *  DPBRFS improves the computed solution to a system of linear
00025 *  equations when the coefficient matrix is symmetric positive definite
00026 *  and banded, and provides error bounds and backward error estimates
00027 *  for the solution.
00028 *
00029 *  Arguments
00030 *  =========
00031 *
00032 *  UPLO    (input) CHARACTER*1
00033 *          = 'U':  Upper triangle of A is stored;
00034 *          = 'L':  Lower triangle of A is stored.
00035 *
00036 *  N       (input) INTEGER
00037 *          The order of the matrix A.  N >= 0.
00038 *
00039 *  KD      (input) INTEGER
00040 *          The number of superdiagonals of the matrix A if UPLO = 'U',
00041 *          or the number of subdiagonals if UPLO = 'L'.  KD >= 0.
00042 *
00043 *  NRHS    (input) INTEGER
00044 *          The number of right hand sides, i.e., the number of columns
00045 *          of the matrices B and X.  NRHS >= 0.
00046 *
00047 *  AB      (input) DOUBLE PRECISION array, dimension (LDAB,N)
00048 *          The upper or lower triangle of the symmetric band matrix A,
00049 *          stored in the first KD+1 rows of the array.  The j-th column
00050 *          of A is stored in the j-th column of the array AB as follows:
00051 *          if UPLO = 'U', AB(kd+1+i-j,j) = A(i,j) for max(1,j-kd)<=i<=j;
00052 *          if UPLO = 'L', AB(1+i-j,j)    = A(i,j) for j<=i<=min(n,j+kd).
00053 *
00054 *  LDAB    (input) INTEGER
00055 *          The leading dimension of the array AB.  LDAB >= KD+1.
00056 *
00057 *  AFB     (input) DOUBLE PRECISION array, dimension (LDAFB,N)
00058 *          The triangular factor U or L from the Cholesky factorization
00059 *          A = U**T*U or A = L*L**T of the band matrix A as computed by
00060 *          DPBTRF, in the same storage format as A (see AB).
00061 *
00062 *  LDAFB   (input) INTEGER
00063 *          The leading dimension of the array AFB.  LDAFB >= KD+1.
00064 *
00065 *  B       (input) DOUBLE PRECISION array, dimension (LDB,NRHS)
00066 *          The right hand side matrix B.
00067 *
00068 *  LDB     (input) INTEGER
00069 *          The leading dimension of the array B.  LDB >= max(1,N).
00070 *
00071 *  X       (input/output) DOUBLE PRECISION array, dimension (LDX,NRHS)
00072 *          On entry, the solution matrix X, as computed by DPBTRS.
00073 *          On exit, the improved solution matrix X.
00074 *
00075 *  LDX     (input) INTEGER
00076 *          The leading dimension of the array X.  LDX >= max(1,N).
00077 *
00078 *  FERR    (output) DOUBLE PRECISION array, dimension (NRHS)
00079 *          The estimated forward error bound for each solution vector
00080 *          X(j) (the j-th column of the solution matrix X).
00081 *          If XTRUE is the true solution corresponding to X(j), FERR(j)
00082 *          is an estimated upper bound for the magnitude of the largest
00083 *          element in (X(j) - XTRUE) divided by the magnitude of the
00084 *          largest element in X(j).  The estimate is as reliable as
00085 *          the estimate for RCOND, and is almost always a slight
00086 *          overestimate of the true error.
00087 *
00088 *  BERR    (output) DOUBLE PRECISION array, dimension (NRHS)
00089 *          The componentwise relative backward error of each solution
00090 *          vector X(j) (i.e., the smallest relative change in
00091 *          any element of A or B that makes X(j) an exact solution).
00092 *
00093 *  WORK    (workspace) DOUBLE PRECISION array, dimension (3*N)
00094 *
00095 *  IWORK   (workspace) INTEGER array, dimension (N)
00096 *
00097 *  INFO    (output) INTEGER
00098 *          = 0:  successful exit
00099 *          < 0:  if INFO = -i, the i-th argument had an illegal value
00100 *
00101 *  Internal Parameters
00102 *  ===================
00103 *
00104 *  ITMAX is the maximum number of steps of iterative refinement.
00105 *
00106 *  =====================================================================
00107 *
00108 *     .. Parameters ..
00109       INTEGER            ITMAX
00110       PARAMETER          ( ITMAX = 5 )
00111       DOUBLE PRECISION   ZERO
00112       PARAMETER          ( ZERO = 0.0D+0 )
00113       DOUBLE PRECISION   ONE
00114       PARAMETER          ( ONE = 1.0D+0 )
00115       DOUBLE PRECISION   TWO
00116       PARAMETER          ( TWO = 2.0D+0 )
00117       DOUBLE PRECISION   THREE
00118       PARAMETER          ( THREE = 3.0D+0 )
00119 *     ..
00120 *     .. Local Scalars ..
00121       LOGICAL            UPPER
00122       INTEGER            COUNT, I, J, K, KASE, L, NZ
00123       DOUBLE PRECISION   EPS, LSTRES, S, SAFE1, SAFE2, SAFMIN, XK
00124 *     ..
00125 *     .. Local Arrays ..
00126       INTEGER            ISAVE( 3 )
00127 *     ..
00128 *     .. External Subroutines ..
00129       EXTERNAL           DAXPY, DCOPY, DLACN2, DPBTRS, DSBMV, XERBLA
00130 *     ..
00131 *     .. Intrinsic Functions ..
00132       INTRINSIC          ABS, MAX, MIN
00133 *     ..
00134 *     .. External Functions ..
00135       LOGICAL            LSAME
00136       DOUBLE PRECISION   DLAMCH
00137       EXTERNAL           LSAME, DLAMCH
00138 *     ..
00139 *     .. Executable Statements ..
00140 *
00141 *     Test the input parameters.
00142 *
00143       INFO = 0
00144       UPPER = LSAME( UPLO, 'U' )
00145       IF( .NOT.UPPER .AND. .NOT.LSAME( UPLO, 'L' ) ) THEN
00146          INFO = -1
00147       ELSE IF( N.LT.0 ) THEN
00148          INFO = -2
00149       ELSE IF( KD.LT.0 ) THEN
00150          INFO = -3
00151       ELSE IF( NRHS.LT.0 ) THEN
00152          INFO = -4
00153       ELSE IF( LDAB.LT.KD+1 ) THEN
00154          INFO = -6
00155       ELSE IF( LDAFB.LT.KD+1 ) THEN
00156          INFO = -8
00157       ELSE IF( LDB.LT.MAX( 1, N ) ) THEN
00158          INFO = -10
00159       ELSE IF( LDX.LT.MAX( 1, N ) ) THEN
00160          INFO = -12
00161       END IF
00162       IF( INFO.NE.0 ) THEN
00163          CALL XERBLA( 'DPBRFS', -INFO )
00164          RETURN
00165       END IF
00166 *
00167 *     Quick return if possible
00168 *
00169       IF( N.EQ.0 .OR. NRHS.EQ.0 ) THEN
00170          DO 10 J = 1, NRHS
00171             FERR( J ) = ZERO
00172             BERR( J ) = ZERO
00173    10    CONTINUE
00174          RETURN
00175       END IF
00176 *
00177 *     NZ = maximum number of nonzero elements in each row of A, plus 1
00178 *
00179       NZ = MIN( N+1, 2*KD+2 )
00180       EPS = DLAMCH( 'Epsilon' )
00181       SAFMIN = DLAMCH( 'Safe minimum' )
00182       SAFE1 = NZ*SAFMIN
00183       SAFE2 = SAFE1 / EPS
00184 *
00185 *     Do for each right hand side
00186 *
00187       DO 140 J = 1, NRHS
00188 *
00189          COUNT = 1
00190          LSTRES = THREE
00191    20    CONTINUE
00192 *
00193 *        Loop until stopping criterion is satisfied.
00194 *
00195 *        Compute residual R = B - A * X
00196 *
00197          CALL DCOPY( N, B( 1, J ), 1, WORK( N+1 ), 1 )
00198          CALL DSBMV( UPLO, N, KD, -ONE, AB, LDAB, X( 1, J ), 1, ONE,
00199      $               WORK( N+1 ), 1 )
00200 *
00201 *        Compute componentwise relative backward error from formula
00202 *
00203 *        max(i) ( abs(R(i)) / ( abs(A)*abs(X) + abs(B) )(i) )
00204 *
00205 *        where abs(Z) is the componentwise absolute value of the matrix
00206 *        or vector Z.  If the i-th component of the denominator is less
00207 *        than SAFE2, then SAFE1 is added to the i-th components of the
00208 *        numerator and denominator before dividing.
00209 *
00210          DO 30 I = 1, N
00211             WORK( I ) = ABS( B( I, J ) )
00212    30    CONTINUE
00213 *
00214 *        Compute abs(A)*abs(X) + abs(B).
00215 *
00216          IF( UPPER ) THEN
00217             DO 50 K = 1, N
00218                S = ZERO
00219                XK = ABS( X( K, J ) )
00220                L = KD + 1 - K
00221                DO 40 I = MAX( 1, K-KD ), K - 1
00222                   WORK( I ) = WORK( I ) + ABS( AB( L+I, K ) )*XK
00223                   S = S + ABS( AB( L+I, K ) )*ABS( X( I, J ) )
00224    40          CONTINUE
00225                WORK( K ) = WORK( K ) + ABS( AB( KD+1, K ) )*XK + S
00226    50       CONTINUE
00227          ELSE
00228             DO 70 K = 1, N
00229                S = ZERO
00230                XK = ABS( X( K, J ) )
00231                WORK( K ) = WORK( K ) + ABS( AB( 1, K ) )*XK
00232                L = 1 - K
00233                DO 60 I = K + 1, MIN( N, K+KD )
00234                   WORK( I ) = WORK( I ) + ABS( AB( L+I, K ) )*XK
00235                   S = S + ABS( AB( L+I, K ) )*ABS( X( I, J ) )
00236    60          CONTINUE
00237                WORK( K ) = WORK( K ) + S
00238    70       CONTINUE
00239          END IF
00240          S = ZERO
00241          DO 80 I = 1, N
00242             IF( WORK( I ).GT.SAFE2 ) THEN
00243                S = MAX( S, ABS( WORK( N+I ) ) / WORK( I ) )
00244             ELSE
00245                S = MAX( S, ( ABS( WORK( N+I ) )+SAFE1 ) /
00246      $             ( WORK( I )+SAFE1 ) )
00247             END IF
00248    80    CONTINUE
00249          BERR( J ) = S
00250 *
00251 *        Test stopping criterion. Continue iterating if
00252 *           1) The residual BERR(J) is larger than machine epsilon, and
00253 *           2) BERR(J) decreased by at least a factor of 2 during the
00254 *              last iteration, and
00255 *           3) At most ITMAX iterations tried.
00256 *
00257          IF( BERR( J ).GT.EPS .AND. TWO*BERR( J ).LE.LSTRES .AND.
00258      $       COUNT.LE.ITMAX ) THEN
00259 *
00260 *           Update solution and try again.
00261 *
00262             CALL DPBTRS( UPLO, N, KD, 1, AFB, LDAFB, WORK( N+1 ), N,
00263      $                   INFO )
00264             CALL DAXPY( N, ONE, WORK( N+1 ), 1, X( 1, J ), 1 )
00265             LSTRES = BERR( J )
00266             COUNT = COUNT + 1
00267             GO TO 20
00268          END IF
00269 *
00270 *        Bound error from formula
00271 *
00272 *        norm(X - XTRUE) / norm(X) .le. FERR =
00273 *        norm( abs(inv(A))*
00274 *           ( abs(R) + NZ*EPS*( abs(A)*abs(X)+abs(B) ))) / norm(X)
00275 *
00276 *        where
00277 *          norm(Z) is the magnitude of the largest component of Z
00278 *          inv(A) is the inverse of A
00279 *          abs(Z) is the componentwise absolute value of the matrix or
00280 *             vector Z
00281 *          NZ is the maximum number of nonzeros in any row of A, plus 1
00282 *          EPS is machine epsilon
00283 *
00284 *        The i-th component of abs(R)+NZ*EPS*(abs(A)*abs(X)+abs(B))
00285 *        is incremented by SAFE1 if the i-th component of
00286 *        abs(A)*abs(X) + abs(B) is less than SAFE2.
00287 *
00288 *        Use DLACN2 to estimate the infinity-norm of the matrix
00289 *           inv(A) * diag(W),
00290 *        where W = abs(R) + NZ*EPS*( abs(A)*abs(X)+abs(B) )))
00291 *
00292          DO 90 I = 1, N
00293             IF( WORK( I ).GT.SAFE2 ) THEN
00294                WORK( I ) = ABS( WORK( N+I ) ) + NZ*EPS*WORK( I )
00295             ELSE
00296                WORK( I ) = ABS( WORK( N+I ) ) + NZ*EPS*WORK( I ) + SAFE1
00297             END IF
00298    90    CONTINUE
00299 *
00300          KASE = 0
00301   100    CONTINUE
00302          CALL DLACN2( N, WORK( 2*N+1 ), WORK( N+1 ), IWORK, FERR( J ),
00303      $                KASE, ISAVE )
00304          IF( KASE.NE.0 ) THEN
00305             IF( KASE.EQ.1 ) THEN
00306 *
00307 *              Multiply by diag(W)*inv(A').
00308 *
00309                CALL DPBTRS( UPLO, N, KD, 1, AFB, LDAFB, WORK( N+1 ), N,
00310      $                      INFO )
00311                DO 110 I = 1, N
00312                   WORK( N+I ) = WORK( N+I )*WORK( I )
00313   110          CONTINUE
00314             ELSE IF( KASE.EQ.2 ) THEN
00315 *
00316 *              Multiply by inv(A)*diag(W).
00317 *
00318                DO 120 I = 1, N
00319                   WORK( N+I ) = WORK( N+I )*WORK( I )
00320   120          CONTINUE
00321                CALL DPBTRS( UPLO, N, KD, 1, AFB, LDAFB, WORK( N+1 ), N,
00322      $                      INFO )
00323             END IF
00324             GO TO 100
00325          END IF
00326 *
00327 *        Normalize error.
00328 *
00329          LSTRES = ZERO
00330          DO 130 I = 1, N
00331             LSTRES = MAX( LSTRES, ABS( X( I, J ) ) )
00332   130    CONTINUE
00333          IF( LSTRES.NE.ZERO )
00334      $      FERR( J ) = FERR( J ) / LSTRES
00335 *
00336   140 CONTINUE
00337 *
00338       RETURN
00339 *
00340 *     End of DPBRFS
00341 *
00342       END