| 1 |
SUBROUTINE SGGSVD( JOBU, JOBV, JOBQ, M, N, P, K, L, A, LDA, B, |
| 2 |
$ LDB, ALPHA, BETA, U, LDU, V, LDV, Q, LDQ, WORK, |
| 3 |
$ IWORK, INFO ) |
| 4 |
* |
| 5 |
* -- LAPACK driver routine (version 3.0) -- |
| 6 |
* Univ. of Tennessee, Univ. of California Berkeley, NAG Ltd., |
| 7 |
* Courant Institute, Argonne National Lab, and Rice University |
| 8 |
* June 30, 1999 |
| 9 |
* |
| 10 |
* .. Scalar Arguments .. |
| 11 |
CHARACTER JOBQ, JOBU, JOBV |
| 12 |
INTEGER INFO, K, L, LDA, LDB, LDQ, LDU, LDV, M, N, P |
| 13 |
* .. |
| 14 |
* .. Array Arguments .. |
| 15 |
INTEGER IWORK( * ) |
| 16 |
REAL A( LDA, * ), ALPHA( * ), B( LDB, * ), |
| 17 |
$ BETA( * ), Q( LDQ, * ), U( LDU, * ), |
| 18 |
$ V( LDV, * ), WORK( * ) |
| 19 |
* .. |
| 20 |
* |
| 21 |
* Purpose |
| 22 |
* ======= |
| 23 |
* |
| 24 |
* SGGSVD computes the generalized singular value decomposition (GSVD) |
| 25 |
* of an M-by-N real matrix A and P-by-N real matrix B: |
| 26 |
* |
| 27 |
* U'*A*Q = D1*( 0 R ), V'*B*Q = D2*( 0 R ) |
| 28 |
* |
| 29 |
* where U, V and Q are orthogonal matrices, and Z' is the transpose |
| 30 |
* of Z. Let K+L = the effective numerical rank of the matrix (A',B')', |
| 31 |
* then R is a K+L-by-K+L nonsingular upper triangular matrix, D1 and |
| 32 |
* D2 are M-by-(K+L) and P-by-(K+L) "diagonal" matrices and of the |
| 33 |
* following structures, respectively: |
| 34 |
* |
| 35 |
* If M-K-L >= 0, |
| 36 |
* |
| 37 |
* K L |
| 38 |
* D1 = K ( I 0 ) |
| 39 |
* L ( 0 C ) |
| 40 |
* M-K-L ( 0 0 ) |
| 41 |
* |
| 42 |
* K L |
| 43 |
* D2 = L ( 0 S ) |
| 44 |
* P-L ( 0 0 ) |
| 45 |
* |
| 46 |
* N-K-L K L |
| 47 |
* ( 0 R ) = K ( 0 R11 R12 ) |
| 48 |
* L ( 0 0 R22 ) |
| 49 |
* |
| 50 |
* where |
| 51 |
* |
| 52 |
* C = diag( ALPHA(K+1), ... , ALPHA(K+L) ), |
| 53 |
* S = diag( BETA(K+1), ... , BETA(K+L) ), |
| 54 |
* C**2 + S**2 = I. |
| 55 |
* |
| 56 |
* R is stored in A(1:K+L,N-K-L+1:N) on exit. |
| 57 |
* |
| 58 |
* If M-K-L < 0, |
| 59 |
* |
| 60 |
* K M-K K+L-M |
| 61 |
* D1 = K ( I 0 0 ) |
| 62 |
* M-K ( 0 C 0 ) |
| 63 |
* |
| 64 |
* K M-K K+L-M |
| 65 |
* D2 = M-K ( 0 S 0 ) |
| 66 |
* K+L-M ( 0 0 I ) |
| 67 |
* P-L ( 0 0 0 ) |
| 68 |
* |
| 69 |
* N-K-L K M-K K+L-M |
| 70 |
* ( 0 R ) = K ( 0 R11 R12 R13 ) |
| 71 |
* M-K ( 0 0 R22 R23 ) |
| 72 |
* K+L-M ( 0 0 0 R33 ) |
| 73 |
* |
| 74 |
* where |
| 75 |
* |
| 76 |
* C = diag( ALPHA(K+1), ... , ALPHA(M) ), |
| 77 |
* S = diag( BETA(K+1), ... , BETA(M) ), |
| 78 |
* C**2 + S**2 = I. |
| 79 |
* |
| 80 |
* (R11 R12 R13 ) is stored in A(1:M, N-K-L+1:N), and R33 is stored |
| 81 |
* ( 0 R22 R23 ) |
| 82 |
* in B(M-K+1:L,N+M-K-L+1:N) on exit. |
| 83 |
* |
| 84 |
* The routine computes C, S, R, and optionally the orthogonal |
| 85 |
* transformation matrices U, V and Q. |
| 86 |
* |
| 87 |
* In particular, if B is an N-by-N nonsingular matrix, then the GSVD of |
| 88 |
* A and B implicitly gives the SVD of A*inv(B): |
| 89 |
* A*inv(B) = U*(D1*inv(D2))*V'. |
| 90 |
* If ( A',B')' has orthonormal columns, then the GSVD of A and B is |
| 91 |
* also equal to the CS decomposition of A and B. Furthermore, the GSVD |
| 92 |
* can be used to derive the solution of the eigenvalue problem: |
| 93 |
* A'*A x = lambda* B'*B x. |
| 94 |
* In some literature, the GSVD of A and B is presented in the form |
| 95 |
* U'*A*X = ( 0 D1 ), V'*B*X = ( 0 D2 ) |
| 96 |
* where U and V are orthogonal and X is nonsingular, D1 and D2 are |
| 97 |
* ``diagonal''. The former GSVD form can be converted to the latter |
| 98 |
* form by taking the nonsingular matrix X as |
| 99 |
* |
| 100 |
* X = Q*( I 0 ) |
| 101 |
* ( 0 inv(R) ). |
| 102 |
* |
| 103 |
* Arguments |
| 104 |
* ========= |
| 105 |
* |
| 106 |
* JOBU (input) CHARACTER*1 |
| 107 |
* = 'U': Orthogonal matrix U is computed; |
| 108 |
* = 'N': U is not computed. |
| 109 |
* |
| 110 |
* JOBV (input) CHARACTER*1 |
| 111 |
* = 'V': Orthogonal matrix V is computed; |
| 112 |
* = 'N': V is not computed. |
| 113 |
* |
| 114 |
* JOBQ (input) CHARACTER*1 |
| 115 |
* = 'Q': Orthogonal matrix Q is computed; |
| 116 |
* = 'N': Q is not computed. |
| 117 |
* |
| 118 |
* M (input) INTEGER |
| 119 |
* The number of rows of the matrix A. M >= 0. |
| 120 |
* |
| 121 |
* N (input) INTEGER |
| 122 |
* The number of columns of the matrices A and B. N >= 0. |
| 123 |
* |
| 124 |
* P (input) INTEGER |
| 125 |
* The number of rows of the matrix B. P >= 0. |
| 126 |
* |
| 127 |
* K (output) INTEGER |
| 128 |
* L (output) INTEGER |
| 129 |
* On exit, K and L specify the dimension of the subblocks |
| 130 |
* described in the Purpose section. |
| 131 |
* K + L = effective numerical rank of (A',B')'. |
| 132 |
* |
| 133 |
* A (input/output) REAL array, dimension (LDA,N) |
| 134 |
* On entry, the M-by-N matrix A. |
| 135 |
* On exit, A contains the triangular matrix R, or part of R. |
| 136 |
* See Purpose for details. |
| 137 |
* |
| 138 |
* LDA (input) INTEGER |
| 139 |
* The leading dimension of the array A. LDA >= max(1,M). |
| 140 |
* |
| 141 |
* B (input/output) REAL array, dimension (LDB,N) |
| 142 |
* On entry, the P-by-N matrix B. |
| 143 |
* On exit, B contains the triangular matrix R if M-K-L < 0. |
| 144 |
* See Purpose for details. |
| 145 |
* |
| 146 |
* LDB (input) INTEGER |
| 147 |
* The leading dimension of the array B. LDA >= max(1,P). |
| 148 |
* |
| 149 |
* ALPHA (output) REAL array, dimension (N) |
| 150 |
* BETA (output) REAL array, dimension (N) |
| 151 |
* On exit, ALPHA and BETA contain the generalized singular |
| 152 |
* value pairs of A and B; |
| 153 |
* ALPHA(1:K) = 1, |
| 154 |
* BETA(1:K) = 0, |
| 155 |
* and if M-K-L >= 0, |
| 156 |
* ALPHA(K+1:K+L) = C, |
| 157 |
* BETA(K+1:K+L) = S, |
| 158 |
* or if M-K-L < 0, |
| 159 |
* ALPHA(K+1:M)=C, ALPHA(M+1:K+L)=0 |
| 160 |
* BETA(K+1:M) =S, BETA(M+1:K+L) =1 |
| 161 |
* and |
| 162 |
* ALPHA(K+L+1:N) = 0 |
| 163 |
* BETA(K+L+1:N) = 0 |
| 164 |
* |
| 165 |
* U (output) REAL array, dimension (LDU,M) |
| 166 |
* If JOBU = 'U', U contains the M-by-M orthogonal matrix U. |
| 167 |
* If JOBU = 'N', U is not referenced. |
| 168 |
* |
| 169 |
* LDU (input) INTEGER |
| 170 |
* The leading dimension of the array U. LDU >= max(1,M) if |
| 171 |
* JOBU = 'U'; LDU >= 1 otherwise. |
| 172 |
* |
| 173 |
* V (output) REAL array, dimension (LDV,P) |
| 174 |
* If JOBV = 'V', V contains the P-by-P orthogonal matrix V. |
| 175 |
* If JOBV = 'N', V is not referenced. |
| 176 |
* |
| 177 |
* LDV (input) INTEGER |
| 178 |
* The leading dimension of the array V. LDV >= max(1,P) if |
| 179 |
* JOBV = 'V'; LDV >= 1 otherwise. |
| 180 |
* |
| 181 |
* Q (output) REAL array, dimension (LDQ,N) |
| 182 |
* If JOBQ = 'Q', Q contains the N-by-N orthogonal matrix Q. |
| 183 |
* If JOBQ = 'N', Q is not referenced. |
| 184 |
* |
| 185 |
* LDQ (input) INTEGER |
| 186 |
* The leading dimension of the array Q. LDQ >= max(1,N) if |
| 187 |
* JOBQ = 'Q'; LDQ >= 1 otherwise. |
| 188 |
* |
| 189 |
* WORK (workspace) REAL array, |
| 190 |
* dimension (max(3*N,M,P)+N) |
| 191 |
* |
| 192 |
* IWORK (workspace/output) INTEGER array, dimension (N) |
| 193 |
* On exit, IWORK stores the sorting information. More |
| 194 |
* precisely, the following loop will sort ALPHA |
| 195 |
* for I = K+1, min(M,K+L) |
| 196 |
* swap ALPHA(I) and ALPHA(IWORK(I)) |
| 197 |
* endfor |
| 198 |
* such that ALPHA(1) >= ALPHA(2) >= ... >= ALPHA(N). |
| 199 |
* |
| 200 |
* INFO (output)INTEGER |
| 201 |
* = 0: successful exit |
| 202 |
* < 0: if INFO = -i, the i-th argument had an illegal value. |
| 203 |
* > 0: if INFO = 1, the Jacobi-type procedure failed to |
| 204 |
* converge. For further details, see subroutine STGSJA. |
| 205 |
* |
| 206 |
* Internal Parameters |
| 207 |
* =================== |
| 208 |
* |
| 209 |
* TOLA REAL |
| 210 |
* TOLB REAL |
| 211 |
* TOLA and TOLB are the thresholds to determine the effective |
| 212 |
* rank of (A',B')'. Generally, they are set to |
| 213 |
* TOLA = MAX(M,N)*norm(A)*MACHEPS, |
| 214 |
* TOLB = MAX(P,N)*norm(B)*MACHEPS. |
| 215 |
* The size of TOLA and TOLB may affect the size of backward |
| 216 |
* errors of the decomposition. |
| 217 |
* |
| 218 |
* Further Details |
| 219 |
* =============== |
| 220 |
* |
| 221 |
* 2-96 Based on modifications by |
| 222 |
* Ming Gu and Huan Ren, Computer Science Division, University of |
| 223 |
* California at Berkeley, USA |
| 224 |
* |
| 225 |
* ===================================================================== |
| 226 |
* |
| 227 |
* .. Local Scalars .. |
| 228 |
LOGICAL WANTQ, WANTU, WANTV |
| 229 |
INTEGER I, IBND, ISUB, J, NCYCLE |
| 230 |
REAL ANORM, BNORM, SMAX, TEMP, TOLA, TOLB, ULP, UNFL |
| 231 |
* .. |
| 232 |
* .. External Functions .. |
| 233 |
LOGICAL LSAME |
| 234 |
REAL SLAMCH, SLANGE |
| 235 |
EXTERNAL LSAME, SLAMCH, SLANGE |
| 236 |
* .. |
| 237 |
* .. External Subroutines .. |
| 238 |
EXTERNAL SCOPY, SGGSVP, STGSJA, XERBLA |
| 239 |
* .. |
| 240 |
* .. Intrinsic Functions .. |
| 241 |
INTRINSIC MAX, MIN |
| 242 |
* .. |
| 243 |
* .. Executable Statements .. |
| 244 |
* |
| 245 |
* Test the input parameters |
| 246 |
* |
| 247 |
WANTU = LSAME( JOBU, 'U' ) |
| 248 |
WANTV = LSAME( JOBV, 'V' ) |
| 249 |
WANTQ = LSAME( JOBQ, 'Q' ) |
| 250 |
* |
| 251 |
INFO = 0 |
| 252 |
IF( .NOT.( WANTU .OR. LSAME( JOBU, 'N' ) ) ) THEN |
| 253 |
INFO = -1 |
| 254 |
ELSE IF( .NOT.( WANTV .OR. LSAME( JOBV, 'N' ) ) ) THEN |
| 255 |
INFO = -2 |
| 256 |
ELSE IF( .NOT.( WANTQ .OR. LSAME( JOBQ, 'N' ) ) ) THEN |
| 257 |
INFO = -3 |
| 258 |
ELSE IF( M.LT.0 ) THEN |
| 259 |
INFO = -4 |
| 260 |
ELSE IF( N.LT.0 ) THEN |
| 261 |
INFO = -5 |
| 262 |
ELSE IF( P.LT.0 ) THEN |
| 263 |
INFO = -6 |
| 264 |
ELSE IF( LDA.LT.MAX( 1, M ) ) THEN |
| 265 |
INFO = -10 |
| 266 |
ELSE IF( LDB.LT.MAX( 1, P ) ) THEN |
| 267 |
INFO = -12 |
| 268 |
ELSE IF( LDU.LT.1 .OR. ( WANTU .AND. LDU.LT.M ) ) THEN |
| 269 |
INFO = -16 |
| 270 |
ELSE IF( LDV.LT.1 .OR. ( WANTV .AND. LDV.LT.P ) ) THEN |
| 271 |
INFO = -18 |
| 272 |
ELSE IF( LDQ.LT.1 .OR. ( WANTQ .AND. LDQ.LT.N ) ) THEN |
| 273 |
INFO = -20 |
| 274 |
END IF |
| 275 |
IF( INFO.NE.0 ) THEN |
| 276 |
CALL XERBLA( 'SGGSVD', -INFO ) |
| 277 |
RETURN |
| 278 |
END IF |
| 279 |
* |
| 280 |
* Compute the Frobenius norm of matrices A and B |
| 281 |
* |
| 282 |
ANORM = SLANGE( '1', M, N, A, LDA, WORK ) |
| 283 |
BNORM = SLANGE( '1', P, N, B, LDB, WORK ) |
| 284 |
* |
| 285 |
* Get machine precision and set up threshold for determining |
| 286 |
* the effective numerical rank of the matrices A and B. |
| 287 |
* |
| 288 |
ULP = SLAMCH( 'Precision' ) |
| 289 |
UNFL = SLAMCH( 'Safe Minimum' ) |
| 290 |
TOLA = MAX( M, N )*MAX( ANORM, UNFL )*ULP |
| 291 |
TOLB = MAX( P, N )*MAX( BNORM, UNFL )*ULP |
| 292 |
* |
| 293 |
* Preprocessing |
| 294 |
* |
| 295 |
CALL SGGSVP( JOBU, JOBV, JOBQ, M, P, N, A, LDA, B, LDB, TOLA, |
| 296 |
$ TOLB, K, L, U, LDU, V, LDV, Q, LDQ, IWORK, WORK, |
| 297 |
$ WORK( N+1 ), INFO ) |
| 298 |
* |
| 299 |
* Compute the GSVD of two upper "triangular" matrices |
| 300 |
* |
| 301 |
CALL STGSJA( JOBU, JOBV, JOBQ, M, P, N, K, L, A, LDA, B, LDB, |
| 302 |
$ TOLA, TOLB, ALPHA, BETA, U, LDU, V, LDV, Q, LDQ, |
| 303 |
$ WORK, NCYCLE, INFO ) |
| 304 |
* |
| 305 |
* Sort the singular values and store the pivot indices in IWORK |
| 306 |
* Copy ALPHA to WORK, then sort ALPHA in WORK |
| 307 |
* |
| 308 |
CALL SCOPY( N, ALPHA, 1, WORK, 1 ) |
| 309 |
IBND = MIN( L, M-K ) |
| 310 |
DO 20 I = 1, IBND |
| 311 |
* |
| 312 |
* Scan for largest ALPHA(K+I) |
| 313 |
* |
| 314 |
ISUB = I |
| 315 |
SMAX = WORK( K+I ) |
| 316 |
DO 10 J = I + 1, IBND |
| 317 |
TEMP = WORK( K+J ) |
| 318 |
IF( TEMP.GT.SMAX ) THEN |
| 319 |
ISUB = J |
| 320 |
SMAX = TEMP |
| 321 |
END IF |
| 322 |
10 CONTINUE |
| 323 |
IF( ISUB.NE.I ) THEN |
| 324 |
WORK( K+ISUB ) = WORK( K+I ) |
| 325 |
WORK( K+I ) = SMAX |
| 326 |
IWORK( K+I ) = K + ISUB |
| 327 |
ELSE |
| 328 |
IWORK( K+I ) = K + I |
| 329 |
END IF |
| 330 |
20 CONTINUE |
| 331 |
* |
| 332 |
RETURN |
| 333 |
* |
| 334 |
* End of SGGSVD |
| 335 |
* |
| 336 |
END |