/* linalg/qrpt.c * * Copyright (C) 1996, 1997, 1998, 1999, 2000, 2007 Gerard Jungman, Brian Gough * * This program is free software; you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by * the Free Software Foundation; either version 3 of the License, or (at * your option) any later version. * * This program is distributed in the hope that it will be useful, but * WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU * General Public License for more details. * * You should have received a copy of the GNU General Public License * along with this program; if not, write to the Free Software * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA. */ #include #include #include #include #include #include #include #include #include #include "apply_givens.c" /* Factorise a general M x N matrix A into * * A P = Q R * * where Q is orthogonal (M x M) and R is upper triangular (M x N). * When A is rank deficient, r = rank(A) < n, then the permutation is * used to ensure that the lower n - r rows of R are zero and the first * r columns of Q form an orthonormal basis for A. * * Q is stored as a packed set of Householder transformations in the * strict lower triangular part of the input matrix. * * R is stored in the diagonal and upper triangle of the input matrix. * * P: column j of P is column k of the identity matrix, where k = * permutation->data[j] * * The full matrix for Q can be obtained as the product * * Q = Q_k .. Q_2 Q_1 * * where k = MIN(M,N) and * * Q_i = (I - tau_i * v_i * v_i') * * and where v_i is a Householder vector * * v_i = [1, m(i+1,i), m(i+2,i), ... , m(M,i)] * * This storage scheme is the same as in LAPACK. See LAPACK's * dgeqpf.f for details. * */ int gsl_linalg_QRPT_decomp (gsl_matrix * A, gsl_vector * tau, gsl_permutation * p, int *signum, gsl_vector * norm) { const size_t M = A->size1; const size_t N = A->size2; if (tau->size != GSL_MIN (M, N)) { GSL_ERROR ("size of tau must be MIN(M,N)", GSL_EBADLEN); } else if (p->size != N) { GSL_ERROR ("permutation size must be N", GSL_EBADLEN); } else if (norm->size != N) { GSL_ERROR ("norm size must be N", GSL_EBADLEN); } else { size_t i; *signum = 1; gsl_permutation_init (p); /* set to identity */ /* Compute column norms and store in workspace */ for (i = 0; i < N; i++) { gsl_vector_view c = gsl_matrix_column (A, i); double x = gsl_blas_dnrm2 (&c.vector); gsl_vector_set (norm, i, x); } for (i = 0; i < GSL_MIN (M, N); i++) { /* Bring the column of largest norm into the pivot position */ double max_norm = gsl_vector_get(norm, i); size_t j, kmax = i; for (j = i + 1; j < N; j++) { double x = gsl_vector_get (norm, j); if (x > max_norm) { max_norm = x; kmax = j; } } if (kmax != i) { gsl_matrix_swap_columns (A, i, kmax); gsl_permutation_swap (p, i, kmax); gsl_vector_swap_elements(norm,i,kmax); (*signum) = -(*signum); } /* Compute the Householder transformation to reduce the j-th column of the matrix to a multiple of the j-th unit vector */ { gsl_vector_view c_full = gsl_matrix_column (A, i); gsl_vector_view c = gsl_vector_subvector (&c_full.vector, i, M - i); double tau_i = gsl_linalg_householder_transform (&c.vector); gsl_vector_set (tau, i, tau_i); /* Apply the transformation to the remaining columns */ if (i + 1 < N) { gsl_matrix_view m = gsl_matrix_submatrix (A, i, i + 1, M - i, N - (i+1)); gsl_linalg_householder_hm (tau_i, &c.vector, &m.matrix); } } /* Update the norms of the remaining columns too */ if (i + 1 < M) { for (j = i + 1; j < N; j++) { double x = gsl_vector_get (norm, j); if (x > 0.0) { double y = 0; double temp= gsl_matrix_get (A, i, j) / x; if (fabs (temp) >= 1) y = 0.0; else y = x * sqrt (1 - temp * temp); /* recompute norm to prevent loss of accuracy */ if (fabs (y / x) < sqrt (20.0) * GSL_SQRT_DBL_EPSILON) { gsl_vector_view c_full = gsl_matrix_column (A, j); gsl_vector_view c = gsl_vector_subvector(&c_full.vector, i+1, M - (i+1)); y = gsl_blas_dnrm2 (&c.vector); } gsl_vector_set (norm, j, y); } } } } return GSL_SUCCESS; } } int gsl_linalg_QRPT_decomp2 (const gsl_matrix * A, gsl_matrix * q, gsl_matrix * r, gsl_vector * tau, gsl_permutation * p, int *signum, gsl_vector * norm) { const size_t M = A->size1; const size_t N = A->size2; if (q->size1 != M || q->size2 !=M) { GSL_ERROR ("q must be M x M", GSL_EBADLEN); } else if (r->size1 != M || r->size2 !=N) { GSL_ERROR ("r must be M x N", GSL_EBADLEN); } else if (tau->size != GSL_MIN (M, N)) { GSL_ERROR ("size of tau must be MIN(M,N)", GSL_EBADLEN); } else if (p->size != N) { GSL_ERROR ("permutation size must be N", GSL_EBADLEN); } else if (norm->size != N) { GSL_ERROR ("norm size must be N", GSL_EBADLEN); } gsl_matrix_memcpy (r, A); gsl_linalg_QRPT_decomp (r, tau, p, signum, norm); /* FIXME: aliased arguments depends on behavior of unpack routine! */ gsl_linalg_QR_unpack (r, tau, q, r); return GSL_SUCCESS; } /* Solves the system A x = b using the Q R P^T factorisation, R z = Q^T b x = P z; to obtain x. Based on SLATEC code. */ int gsl_linalg_QRPT_solve (const gsl_matrix * QR, const gsl_vector * tau, const gsl_permutation * p, const gsl_vector * b, gsl_vector * x) { if (QR->size1 != QR->size2) { GSL_ERROR ("QR matrix must be square", GSL_ENOTSQR); } else if (QR->size1 != p->size) { GSL_ERROR ("matrix size must match permutation size", GSL_EBADLEN); } else if (QR->size1 != b->size) { GSL_ERROR ("matrix size must match b size", GSL_EBADLEN); } else if (QR->size2 != x->size) { GSL_ERROR ("matrix size must match solution size", GSL_EBADLEN); } else { gsl_vector_memcpy (x, b); gsl_linalg_QRPT_svx (QR, tau, p, x); return GSL_SUCCESS; } } int gsl_linalg_QRPT_svx (const gsl_matrix * QR, const gsl_vector * tau, const gsl_permutation * p, gsl_vector * x) { if (QR->size1 != QR->size2) { GSL_ERROR ("QR matrix must be square", GSL_ENOTSQR); } else if (QR->size1 != p->size) { GSL_ERROR ("matrix size must match permutation size", GSL_EBADLEN); } else if (QR->size2 != x->size) { GSL_ERROR ("matrix size must match solution size", GSL_EBADLEN); } else { /* compute sol = Q^T b */ gsl_linalg_QR_QTvec (QR, tau, x); /* Solve R x = sol, storing x inplace in sol */ gsl_blas_dtrsv (CblasUpper, CblasNoTrans, CblasNonUnit, QR, x); gsl_permute_vector_inverse (p, x); return GSL_SUCCESS; } } /* Find the least squares solution to the overdetermined system * * A x = b * * for M >= N using the QRPT factorization A P = Q R. Assumes * A has full column rank. */ int gsl_linalg_QRPT_lssolve (const gsl_matrix * QR, const gsl_vector * tau, const gsl_permutation * p, const gsl_vector * b, gsl_vector * x, gsl_vector * residual) { const size_t N = QR->size2; int status = gsl_linalg_QRPT_lssolve2(QR, tau, p, b, N, x, residual); return status; } /* Find the least squares solution to the overdetermined system * * A x = b * * for M >= N using the QRPT factorization A P = Q R, where A * is assumed rank deficient with a given rank. */ int gsl_linalg_QRPT_lssolve2 (const gsl_matrix * QR, const gsl_vector * tau, const gsl_permutation * p, const gsl_vector * b, const size_t rank, gsl_vector * x, gsl_vector * residual) { const size_t M = QR->size1; const size_t N = QR->size2; if (M < N) { GSL_ERROR ("QR matrix must have M>=N", GSL_EBADLEN); } else if (M != b->size) { GSL_ERROR ("matrix size must match b size", GSL_EBADLEN); } else if (rank == 0 || rank > N) { GSL_ERROR ("rank must have 0 < rank <= N", GSL_EBADLEN); } else if (N != x->size) { GSL_ERROR ("matrix size must match solution size", GSL_EBADLEN); } else if (M != residual->size) { GSL_ERROR ("matrix size must match residual size", GSL_EBADLEN); } else { gsl_matrix_const_view R11 = gsl_matrix_const_submatrix (QR, 0, 0, rank, rank); gsl_vector_view QTb1 = gsl_vector_subvector(residual, 0, rank); gsl_vector_view x1 = gsl_vector_subvector(x, 0, rank); size_t i; /* compute work = Q^T b */ gsl_vector_memcpy(residual, b); gsl_linalg_QR_QTvec (QR, tau, residual); /* solve R_{11} x(1:r) = [Q^T b](1:r) */ gsl_vector_memcpy(&(x1.vector), &(QTb1.vector)); gsl_blas_dtrsv (CblasUpper, CblasNoTrans, CblasNonUnit, &(R11.matrix), &(x1.vector)); /* x(r+1:N) = 0 */ for (i = rank; i < N; ++i) gsl_vector_set(x, i, 0.0); /* compute x = P y */ gsl_permute_vector_inverse (p, x); /* compute residual = b - A x = Q (Q^T b - R x) */ gsl_vector_set_zero(&(QTb1.vector)); gsl_linalg_QR_Qvec(QR, tau, residual); return GSL_SUCCESS; } } int gsl_linalg_QRPT_QRsolve (const gsl_matrix * Q, const gsl_matrix * R, const gsl_permutation * p, const gsl_vector * b, gsl_vector * x) { if (Q->size1 != Q->size2 || R->size1 != R->size2) { return GSL_ENOTSQR; } else if (Q->size1 != p->size || Q->size1 != R->size1 || Q->size1 != b->size) { return GSL_EBADLEN; } else { /* compute b' = Q^T b */ gsl_blas_dgemv (CblasTrans, 1.0, Q, b, 0.0, x); /* Solve R x = b', storing x inplace */ gsl_blas_dtrsv (CblasUpper, CblasNoTrans, CblasNonUnit, R, x); /* Apply permutation to solution in place */ gsl_permute_vector_inverse (p, x); return GSL_SUCCESS; } } int gsl_linalg_QRPT_Rsolve (const gsl_matrix * QR, const gsl_permutation * p, const gsl_vector * b, gsl_vector * x) { if (QR->size1 != QR->size2) { GSL_ERROR ("QR matrix must be square", GSL_ENOTSQR); } else if (QR->size1 != b->size) { GSL_ERROR ("matrix size must match b size", GSL_EBADLEN); } else if (QR->size2 != x->size) { GSL_ERROR ("matrix size must match x size", GSL_EBADLEN); } else if (p->size != x->size) { GSL_ERROR ("permutation size must match x size", GSL_EBADLEN); } else { /* Copy x <- b */ gsl_vector_memcpy (x, b); /* Solve R x = b, storing x inplace */ gsl_blas_dtrsv (CblasUpper, CblasNoTrans, CblasNonUnit, QR, x); gsl_permute_vector_inverse (p, x); return GSL_SUCCESS; } } int gsl_linalg_QRPT_Rsvx (const gsl_matrix * QR, const gsl_permutation * p, gsl_vector * x) { if (QR->size1 != QR->size2) { GSL_ERROR ("QR matrix must be square", GSL_ENOTSQR); } else if (QR->size2 != x->size) { GSL_ERROR ("matrix size must match x size", GSL_EBADLEN); } else if (p->size != x->size) { GSL_ERROR ("permutation size must match x size", GSL_EBADLEN); } else { /* Solve R x = b, storing x inplace */ gsl_blas_dtrsv (CblasUpper, CblasNoTrans, CblasNonUnit, QR, x); gsl_permute_vector_inverse (p, x); return GSL_SUCCESS; } } /* Update a Q R P^T factorisation for A P= Q R , A' = A + u v^T, Q' R' P^-1 = QR P^-1 + u v^T = Q (R + Q^T u v^T P ) P^-1 = Q (R + w v^T P) P^-1 where w = Q^T u. Algorithm from Golub and Van Loan, "Matrix Computations", Section 12.5 (Updating Matrix Factorizations, Rank-One Changes) */ int gsl_linalg_QRPT_update (gsl_matrix * Q, gsl_matrix * R, const gsl_permutation * p, gsl_vector * w, const gsl_vector * v) { const size_t M = R->size1; const size_t N = R->size2; if (Q->size1 != M || Q->size2 != M) { GSL_ERROR ("Q matrix must be M x M if R is M x N", GSL_ENOTSQR); } else if (w->size != M) { GSL_ERROR ("w must be length M if R is M x N", GSL_EBADLEN); } else if (v->size != N) { GSL_ERROR ("v must be length N if R is M x N", GSL_EBADLEN); } else { size_t j, k; double w0; /* Apply Given's rotations to reduce w to (|w|, 0, 0, ... , 0) J_1^T .... J_(n-1)^T w = +/- |w| e_1 simultaneously applied to R, H = J_1^T ... J^T_(n-1) R so that H is upper Hessenberg. (12.5.2) */ for (k = M - 1; k > 0; k--) { double c, s; double wk = gsl_vector_get (w, k); double wkm1 = gsl_vector_get (w, k - 1); gsl_linalg_givens (wkm1, wk, &c, &s); gsl_linalg_givens_gv (w, k - 1, k, c, s); apply_givens_qr (M, N, Q, R, k - 1, k, c, s); } w0 = gsl_vector_get (w, 0); /* Add in w v^T (Equation 12.5.3) */ for (j = 0; j < N; j++) { double r0j = gsl_matrix_get (R, 0, j); size_t p_j = gsl_permutation_get (p, j); double vj = gsl_vector_get (v, p_j); gsl_matrix_set (R, 0, j, r0j + w0 * vj); } /* Apply Givens transformations R' = G_(n-1)^T ... G_1^T H Equation 12.5.4 */ for (k = 1; k < GSL_MIN(M,N+1); k++) { double c, s; double diag = gsl_matrix_get (R, k - 1, k - 1); double offdiag = gsl_matrix_get (R, k, k - 1); gsl_linalg_givens (diag, offdiag, &c, &s); apply_givens_qr (M, N, Q, R, k - 1, k, c, s); gsl_matrix_set (R, k, k - 1, 0.0); /* exact zero of G^T */ } return GSL_SUCCESS; } } /* gsl_linalg_QRPT_rank() Estimate rank of triangular matrix from QRPT decomposition Inputs: QR - QRPT decomposed matrix tol - tolerance for rank determination; if < 0, a default value is used: 20 * (M + N) * eps(max(|diag(R)|)) Return: rank estimate */ size_t gsl_linalg_QRPT_rank (const gsl_matrix * QR, const double tol) { const size_t M = QR->size1; const size_t N = QR->size2; gsl_vector_const_view diag = gsl_matrix_const_diagonal(QR); double eps; size_t i; size_t r = 0; if (tol < 0.0) { double min, max, absmax; int ee; gsl_vector_minmax(&diag.vector, &min, &max); absmax = GSL_MAX(fabs(min), fabs(max)); ee = (int) (log(absmax) / M_LN2); /* can't use log2 since its not ANSI */ eps = 20.0 * (M + N) * pow(2.0, (double) ee) * GSL_DBL_EPSILON; } else eps = tol; /* count number of diagonal elements with |di| > eps */ for (i = 0; i < GSL_MIN(M, N); ++i) { double di = gsl_vector_get(&diag.vector, i); if (fabs(di) > eps) ++r; } return r; } int gsl_linalg_QRPT_rcond(const gsl_matrix * QR, double * rcond, gsl_vector * work) { const size_t M = QR->size1; const size_t N = QR->size2; if (M < N) { GSL_ERROR ("M must be >= N", GSL_EBADLEN); } else if (work->size != 3 * N) { GSL_ERROR ("work vector must have length 3*N", GSL_EBADLEN); } else { gsl_matrix_const_view R = gsl_matrix_const_submatrix (QR, 0, 0, N, N); int status; status = gsl_linalg_tri_rcond(CblasUpper, &R.matrix, rcond, work); return status; } }