LAPACK 3.3.0

dlaed8.f

Go to the documentation of this file.
00001       SUBROUTINE DLAED8( ICOMPQ, K, N, QSIZ, D, Q, LDQ, INDXQ, RHO,
00002      $                   CUTPNT, Z, DLAMDA, Q2, LDQ2, W, PERM, GIVPTR,
00003      $                   GIVCOL, GIVNUM, INDXP, INDX, INFO )
00004 *
00005 *  -- LAPACK routine (version 3.2.2) --
00006 *  -- LAPACK is a software package provided by Univ. of Tennessee,    --
00007 *  -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
00008 *     June 2010
00009 *
00010 *     .. Scalar Arguments ..
00011       INTEGER            CUTPNT, GIVPTR, ICOMPQ, INFO, K, LDQ, LDQ2, N,
00012      $                   QSIZ
00013       DOUBLE PRECISION   RHO
00014 *     ..
00015 *     .. Array Arguments ..
00016       INTEGER            GIVCOL( 2, * ), INDX( * ), INDXP( * ),
00017      $                   INDXQ( * ), PERM( * )
00018       DOUBLE PRECISION   D( * ), DLAMDA( * ), GIVNUM( 2, * ),
00019      $                   Q( LDQ, * ), Q2( LDQ2, * ), W( * ), Z( * )
00020 *     ..
00021 *
00022 *  Purpose
00023 *  =======
00024 *
00025 *  DLAED8 merges the two sets of eigenvalues together into a single
00026 *  sorted set.  Then it tries to deflate the size of the problem.
00027 *  There are two ways in which deflation can occur:  when two or more
00028 *  eigenvalues are close together or if there is a tiny element in the
00029 *  Z vector.  For each such occurrence the order of the related secular
00030 *  equation problem is reduced by one.
00031 *
00032 *  Arguments
00033 *  =========
00034 *
00035 *  ICOMPQ  (input) INTEGER
00036 *          = 0:  Compute eigenvalues only.
00037 *          = 1:  Compute eigenvectors of original dense symmetric matrix
00038 *                also.  On entry, Q contains the orthogonal matrix used
00039 *                to reduce the original matrix to tridiagonal form.
00040 *
00041 *  K      (output) INTEGER
00042 *         The number of non-deflated eigenvalues, and the order of the
00043 *         related secular equation.
00044 *
00045 *  N      (input) INTEGER
00046 *         The dimension of the symmetric tridiagonal matrix.  N >= 0.
00047 *
00048 *  QSIZ   (input) INTEGER
00049 *         The dimension of the orthogonal matrix used to reduce
00050 *         the full matrix to tridiagonal form.  QSIZ >= N if ICOMPQ = 1.
00051 *
00052 *  D      (input/output) DOUBLE PRECISION array, dimension (N)
00053 *         On entry, the eigenvalues of the two submatrices to be
00054 *         combined.  On exit, the trailing (N-K) updated eigenvalues
00055 *         (those which were deflated) sorted into increasing order.
00056 *
00057 *  Q      (input/output) DOUBLE PRECISION array, dimension (LDQ,N)
00058 *         If ICOMPQ = 0, Q is not referenced.  Otherwise,
00059 *         on entry, Q contains the eigenvectors of the partially solved
00060 *         system which has been previously updated in matrix
00061 *         multiplies with other partially solved eigensystems.
00062 *         On exit, Q contains the trailing (N-K) updated eigenvectors
00063 *         (those which were deflated) in its last N-K columns.
00064 *
00065 *  LDQ    (input) INTEGER
00066 *         The leading dimension of the array Q.  LDQ >= max(1,N).
00067 *
00068 *  INDXQ  (input) INTEGER array, dimension (N)
00069 *         The permutation which separately sorts the two sub-problems
00070 *         in D into ascending order.  Note that elements in the second
00071 *         half of this permutation must first have CUTPNT added to
00072 *         their values in order to be accurate.
00073 *
00074 *  RHO    (input/output) DOUBLE PRECISION
00075 *         On entry, the off-diagonal element associated with the rank-1
00076 *         cut which originally split the two submatrices which are now
00077 *         being recombined.
00078 *         On exit, RHO has been modified to the value required by
00079 *         DLAED3.
00080 *
00081 *  CUTPNT (input) INTEGER
00082 *         The location of the last eigenvalue in the leading
00083 *         sub-matrix.  min(1,N) <= CUTPNT <= N.
00084 *
00085 *  Z      (input) DOUBLE PRECISION array, dimension (N)
00086 *         On entry, Z contains the updating vector (the last row of
00087 *         the first sub-eigenvector matrix and the first row of the
00088 *         second sub-eigenvector matrix).
00089 *         On exit, the contents of Z are destroyed by the updating
00090 *         process.
00091 *
00092 *  DLAMDA (output) DOUBLE PRECISION array, dimension (N)
00093 *         A copy of the first K eigenvalues which will be used by
00094 *         DLAED3 to form the secular equation.
00095 *
00096 *  Q2     (output) DOUBLE PRECISION array, dimension (LDQ2,N)
00097 *         If ICOMPQ = 0, Q2 is not referenced.  Otherwise,
00098 *         a copy of the first K eigenvectors which will be used by
00099 *         DLAED7 in a matrix multiply (DGEMM) to update the new
00100 *         eigenvectors.
00101 *
00102 *  LDQ2   (input) INTEGER
00103 *         The leading dimension of the array Q2.  LDQ2 >= max(1,N).
00104 *
00105 *  W      (output) DOUBLE PRECISION array, dimension (N)
00106 *         The first k values of the final deflation-altered z-vector and
00107 *         will be passed to DLAED3.
00108 *
00109 *  PERM   (output) INTEGER array, dimension (N)
00110 *         The permutations (from deflation and sorting) to be applied
00111 *         to each eigenblock.
00112 *
00113 *  GIVPTR (output) INTEGER
00114 *         The number of Givens rotations which took place in this
00115 *         subproblem.
00116 *
00117 *  GIVCOL (output) INTEGER array, dimension (2, N)
00118 *         Each pair of numbers indicates a pair of columns to take place
00119 *         in a Givens rotation.
00120 *
00121 *  GIVNUM (output) DOUBLE PRECISION array, dimension (2, N)
00122 *         Each number indicates the S value to be used in the
00123 *         corresponding Givens rotation.
00124 *
00125 *  INDXP  (workspace) INTEGER array, dimension (N)
00126 *         The permutation used to place deflated values of D at the end
00127 *         of the array.  INDXP(1:K) points to the nondeflated D-values
00128 *         and INDXP(K+1:N) points to the deflated eigenvalues.
00129 *
00130 *  INDX   (workspace) INTEGER array, dimension (N)
00131 *         The permutation used to sort the contents of D into ascending
00132 *         order.
00133 *
00134 *  INFO   (output) INTEGER
00135 *          = 0:  successful exit.
00136 *          < 0:  if INFO = -i, the i-th argument had an illegal value.
00137 *
00138 *  Further Details
00139 *  ===============
00140 *
00141 *  Based on contributions by
00142 *     Jeff Rutter, Computer Science Division, University of California
00143 *     at Berkeley, USA
00144 *
00145 *  =====================================================================
00146 *
00147 *     .. Parameters ..
00148       DOUBLE PRECISION   MONE, ZERO, ONE, TWO, EIGHT
00149       PARAMETER          ( MONE = -1.0D0, ZERO = 0.0D0, ONE = 1.0D0,
00150      $                   TWO = 2.0D0, EIGHT = 8.0D0 )
00151 *     ..
00152 *     .. Local Scalars ..
00153 *
00154       INTEGER            I, IMAX, J, JLAM, JMAX, JP, K2, N1, N1P1, N2
00155       DOUBLE PRECISION   C, EPS, S, T, TAU, TOL
00156 *     ..
00157 *     .. External Functions ..
00158       INTEGER            IDAMAX
00159       DOUBLE PRECISION   DLAMCH, DLAPY2
00160       EXTERNAL           IDAMAX, DLAMCH, DLAPY2
00161 *     ..
00162 *     .. External Subroutines ..
00163       EXTERNAL           DCOPY, DLACPY, DLAMRG, DROT, DSCAL, XERBLA
00164 *     ..
00165 *     .. Intrinsic Functions ..
00166       INTRINSIC          ABS, MAX, MIN, SQRT
00167 *     ..
00168 *     .. Executable Statements ..
00169 *
00170 *     Test the input parameters.
00171 *
00172       INFO = 0
00173 *
00174       IF( ICOMPQ.LT.0 .OR. ICOMPQ.GT.1 ) THEN
00175          INFO = -1
00176       ELSE IF( N.LT.0 ) THEN
00177          INFO = -3
00178       ELSE IF( ICOMPQ.EQ.1 .AND. QSIZ.LT.N ) THEN
00179          INFO = -4
00180       ELSE IF( LDQ.LT.MAX( 1, N ) ) THEN
00181          INFO = -7
00182       ELSE IF( CUTPNT.LT.MIN( 1, N ) .OR. CUTPNT.GT.N ) THEN
00183          INFO = -10
00184       ELSE IF( LDQ2.LT.MAX( 1, N ) ) THEN
00185          INFO = -14
00186       END IF
00187       IF( INFO.NE.0 ) THEN
00188          CALL XERBLA( 'DLAED8', -INFO )
00189          RETURN
00190       END IF
00191 *
00192 *     Need to initialize GIVPTR to O here in case of quick exit
00193 *     to prevent an unspecified code behavior (usually sigfault) 
00194 *     when IWORK array on entry to *stedc is not zeroed 
00195 *     (or at least some IWORK entries which used in *laed7 for GIVPTR).
00196 *
00197       GIVPTR = 0
00198 *
00199 *     Quick return if possible
00200 *
00201       IF( N.EQ.0 )
00202      $   RETURN
00203 *
00204       N1 = CUTPNT
00205       N2 = N - N1
00206       N1P1 = N1 + 1
00207 *
00208       IF( RHO.LT.ZERO ) THEN
00209          CALL DSCAL( N2, MONE, Z( N1P1 ), 1 )
00210       END IF
00211 *
00212 *     Normalize z so that norm(z) = 1
00213 *
00214       T = ONE / SQRT( TWO )
00215       DO 10 J = 1, N
00216          INDX( J ) = J
00217    10 CONTINUE
00218       CALL DSCAL( N, T, Z, 1 )
00219       RHO = ABS( TWO*RHO )
00220 *
00221 *     Sort the eigenvalues into increasing order
00222 *
00223       DO 20 I = CUTPNT + 1, N
00224          INDXQ( I ) = INDXQ( I ) + CUTPNT
00225    20 CONTINUE
00226       DO 30 I = 1, N
00227          DLAMDA( I ) = D( INDXQ( I ) )
00228          W( I ) = Z( INDXQ( I ) )
00229    30 CONTINUE
00230       I = 1
00231       J = CUTPNT + 1
00232       CALL DLAMRG( N1, N2, DLAMDA, 1, 1, INDX )
00233       DO 40 I = 1, N
00234          D( I ) = DLAMDA( INDX( I ) )
00235          Z( I ) = W( INDX( I ) )
00236    40 CONTINUE
00237 *
00238 *     Calculate the allowable deflation tolerence
00239 *
00240       IMAX = IDAMAX( N, Z, 1 )
00241       JMAX = IDAMAX( N, D, 1 )
00242       EPS = DLAMCH( 'Epsilon' )
00243       TOL = EIGHT*EPS*ABS( D( JMAX ) )
00244 *
00245 *     If the rank-1 modifier is small enough, no more needs to be done
00246 *     except to reorganize Q so that its columns correspond with the
00247 *     elements in D.
00248 *
00249       IF( RHO*ABS( Z( IMAX ) ).LE.TOL ) THEN
00250          K = 0
00251          IF( ICOMPQ.EQ.0 ) THEN
00252             DO 50 J = 1, N
00253                PERM( J ) = INDXQ( INDX( J ) )
00254    50       CONTINUE
00255          ELSE
00256             DO 60 J = 1, N
00257                PERM( J ) = INDXQ( INDX( J ) )
00258                CALL DCOPY( QSIZ, Q( 1, PERM( J ) ), 1, Q2( 1, J ), 1 )
00259    60       CONTINUE
00260             CALL DLACPY( 'A', QSIZ, N, Q2( 1, 1 ), LDQ2, Q( 1, 1 ),
00261      $                   LDQ )
00262          END IF
00263          RETURN
00264       END IF
00265 *
00266 *     If there are multiple eigenvalues then the problem deflates.  Here
00267 *     the number of equal eigenvalues are found.  As each equal
00268 *     eigenvalue is found, an elementary reflector is computed to rotate
00269 *     the corresponding eigensubspace so that the corresponding
00270 *     components of Z are zero in this new basis.
00271 *
00272       K = 0
00273       K2 = N + 1
00274       DO 70 J = 1, N
00275          IF( RHO*ABS( Z( J ) ).LE.TOL ) THEN
00276 *
00277 *           Deflate due to small z component.
00278 *
00279             K2 = K2 - 1
00280             INDXP( K2 ) = J
00281             IF( J.EQ.N )
00282      $         GO TO 110
00283          ELSE
00284             JLAM = J
00285             GO TO 80
00286          END IF
00287    70 CONTINUE
00288    80 CONTINUE
00289       J = J + 1
00290       IF( J.GT.N )
00291      $   GO TO 100
00292       IF( RHO*ABS( Z( J ) ).LE.TOL ) THEN
00293 *
00294 *        Deflate due to small z component.
00295 *
00296          K2 = K2 - 1
00297          INDXP( K2 ) = J
00298       ELSE
00299 *
00300 *        Check if eigenvalues are close enough to allow deflation.
00301 *
00302          S = Z( JLAM )
00303          C = Z( J )
00304 *
00305 *        Find sqrt(a**2+b**2) without overflow or
00306 *        destructive underflow.
00307 *
00308          TAU = DLAPY2( C, S )
00309          T = D( J ) - D( JLAM )
00310          C = C / TAU
00311          S = -S / TAU
00312          IF( ABS( T*C*S ).LE.TOL ) THEN
00313 *
00314 *           Deflation is possible.
00315 *
00316             Z( J ) = TAU
00317             Z( JLAM ) = ZERO
00318 *
00319 *           Record the appropriate Givens rotation
00320 *
00321             GIVPTR = GIVPTR + 1
00322             GIVCOL( 1, GIVPTR ) = INDXQ( INDX( JLAM ) )
00323             GIVCOL( 2, GIVPTR ) = INDXQ( INDX( J ) )
00324             GIVNUM( 1, GIVPTR ) = C
00325             GIVNUM( 2, GIVPTR ) = S
00326             IF( ICOMPQ.EQ.1 ) THEN
00327                CALL DROT( QSIZ, Q( 1, INDXQ( INDX( JLAM ) ) ), 1,
00328      $                    Q( 1, INDXQ( INDX( J ) ) ), 1, C, S )
00329             END IF
00330             T = D( JLAM )*C*C + D( J )*S*S
00331             D( J ) = D( JLAM )*S*S + D( J )*C*C
00332             D( JLAM ) = T
00333             K2 = K2 - 1
00334             I = 1
00335    90       CONTINUE
00336             IF( K2+I.LE.N ) THEN
00337                IF( D( JLAM ).LT.D( INDXP( K2+I ) ) ) THEN
00338                   INDXP( K2+I-1 ) = INDXP( K2+I )
00339                   INDXP( K2+I ) = JLAM
00340                   I = I + 1
00341                   GO TO 90
00342                ELSE
00343                   INDXP( K2+I-1 ) = JLAM
00344                END IF
00345             ELSE
00346                INDXP( K2+I-1 ) = JLAM
00347             END IF
00348             JLAM = J
00349          ELSE
00350             K = K + 1
00351             W( K ) = Z( JLAM )
00352             DLAMDA( K ) = D( JLAM )
00353             INDXP( K ) = JLAM
00354             JLAM = J
00355          END IF
00356       END IF
00357       GO TO 80
00358   100 CONTINUE
00359 *
00360 *     Record the last eigenvalue.
00361 *
00362       K = K + 1
00363       W( K ) = Z( JLAM )
00364       DLAMDA( K ) = D( JLAM )
00365       INDXP( K ) = JLAM
00366 *
00367   110 CONTINUE
00368 *
00369 *     Sort the eigenvalues and corresponding eigenvectors into DLAMDA
00370 *     and Q2 respectively.  The eigenvalues/vectors which were not
00371 *     deflated go into the first K slots of DLAMDA and Q2 respectively,
00372 *     while those which were deflated go into the last N - K slots.
00373 *
00374       IF( ICOMPQ.EQ.0 ) THEN
00375          DO 120 J = 1, N
00376             JP = INDXP( J )
00377             DLAMDA( J ) = D( JP )
00378             PERM( J ) = INDXQ( INDX( JP ) )
00379   120    CONTINUE
00380       ELSE
00381          DO 130 J = 1, N
00382             JP = INDXP( J )
00383             DLAMDA( J ) = D( JP )
00384             PERM( J ) = INDXQ( INDX( JP ) )
00385             CALL DCOPY( QSIZ, Q( 1, PERM( J ) ), 1, Q2( 1, J ), 1 )
00386   130    CONTINUE
00387       END IF
00388 *
00389 *     The deflated eigenvalues and their corresponding vectors go back
00390 *     into the last N - K slots of D and Q respectively.
00391 *
00392       IF( K.LT.N ) THEN
00393          IF( ICOMPQ.EQ.0 ) THEN
00394             CALL DCOPY( N-K, DLAMDA( K+1 ), 1, D( K+1 ), 1 )
00395          ELSE
00396             CALL DCOPY( N-K, DLAMDA( K+1 ), 1, D( K+1 ), 1 )
00397             CALL DLACPY( 'A', QSIZ, N-K, Q2( 1, K+1 ), LDQ2,
00398      $                   Q( 1, K+1 ), LDQ )
00399          END IF
00400       END IF
00401 *
00402       RETURN
00403 *
00404 *     End of DLAED8
00405 *
00406       END
 All Files Functions