/* multifit_nlinear/cholesky.c * * Copyright (C) 2015, 2016 Patrick Alken * * 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. */ /* * This module calculates the solution of the normal equations least squares * system: * * [ J~^T J~ + mu D~^T D~ ] p~ = -J~^T f * * using the modified Cholesky decomposition. Quantities are scaled * according to: * * J~ = J S * D~ = D S * p~ = S^{-1} p * * where S is a diagonal matrix and S_jj = || J_j || and J_j is column * j of the Jacobian. This balancing transformation seems to be more * numerically stable for some Jacobians. */ #include #include #include #include #include #include #include #include #include #include "common.c" typedef struct { gsl_matrix *JTJ; /* J^T J */ gsl_matrix *work_JTJ; /* copy of J^T J */ gsl_vector *rhs; /* -J^T f, size p */ gsl_permutation *perm; /* permutation matrix for modified Cholesky */ gsl_vector *work3p; /* workspace, size 3*p */ double mu; /* current regularization parameter */ } cholesky_state_t; static void *cholesky_alloc (const size_t n, const size_t p); static int cholesky_init(const void * vtrust_state, void * vstate); static int cholesky_presolve(const double mu, const void * vtrust_state, void * vstate); static int cholesky_solve(const gsl_vector * f, gsl_vector *x, const void * vtrust_state, void *vstate); static int cholesky_solve_rhs(const gsl_vector * b, gsl_vector *x, cholesky_state_t *state); static int cholesky_regularize(const double mu, const gsl_vector * diag, gsl_matrix * A, cholesky_state_t * state); static void * cholesky_alloc (const size_t n, const size_t p) { cholesky_state_t *state; (void)n; state = calloc(1, sizeof(cholesky_state_t)); if (state == NULL) { GSL_ERROR_NULL ("failed to allocate cholesky state", GSL_ENOMEM); } state->JTJ = gsl_matrix_alloc(p, p); if (state->JTJ == NULL) { GSL_ERROR_NULL ("failed to allocate space for JTJ", GSL_ENOMEM); } state->work_JTJ = gsl_matrix_alloc(p, p); if (state->work_JTJ == NULL) { GSL_ERROR_NULL ("failed to allocate space for JTJ workspace", GSL_ENOMEM); } state->rhs = gsl_vector_alloc(p); if (state->rhs == NULL) { GSL_ERROR_NULL ("failed to allocate space for rhs", GSL_ENOMEM); } state->perm = gsl_permutation_alloc(p); if (state->perm == NULL) { GSL_ERROR_NULL ("failed to allocate space for perm", GSL_ENOMEM); } state->work3p = gsl_vector_alloc(3 * p); if (state->work3p == NULL) { GSL_ERROR_NULL ("failed to allocate space for work3p", GSL_ENOMEM); } state->mu = -1.0; return state; } static void cholesky_free(void *vstate) { cholesky_state_t *state = (cholesky_state_t *) vstate; if (state->JTJ) gsl_matrix_free(state->JTJ); if (state->work_JTJ) gsl_matrix_free(state->work_JTJ); if (state->rhs) gsl_vector_free(state->rhs); if (state->perm) gsl_permutation_free(state->perm); if (state->work3p) gsl_vector_free(state->work3p); free(state); } static int cholesky_init(const void * vtrust_state, void * vstate) { const gsl_multifit_nlinear_trust_state *trust_state = (const gsl_multifit_nlinear_trust_state *) vtrust_state; cholesky_state_t *state = (cholesky_state_t *) vstate; /* compute J^T J */ gsl_blas_dsyrk(CblasLower, CblasTrans, 1.0, trust_state->J, 0.0, state->JTJ); return GSL_SUCCESS; } /* cholesky_presolve() Compute the modified Cholesky decomposition of J^T J + mu D^T D. Modified Cholesky is used in case mu = 0 and there are rounding errors in forming J^T J which could lead to an indefinite matrix. Inputs: mu - LM parameter vstate - workspace Notes: 1) On output, state->work_JTJ contains the Cholesky decomposition of J^T J + mu D^T D */ static int cholesky_presolve(const double mu, const void * vtrust_state, void * vstate) { const gsl_multifit_nlinear_trust_state *trust_state = (const gsl_multifit_nlinear_trust_state *) vtrust_state; cholesky_state_t *state = (cholesky_state_t *) vstate; gsl_matrix *JTJ = state->work_JTJ; const gsl_vector *diag = trust_state->diag; int status; /* copy lower triangle of A to workspace */ gsl_matrix_tricpy('L', 1, JTJ, state->JTJ); /* augment normal equations: A -> A + mu D^T D */ status = cholesky_regularize(mu, diag, JTJ, state); if (status) return status; /* compute modified Cholesky decomposition */ status = gsl_linalg_mcholesky_decomp(JTJ, state->perm, NULL); if (status) return status; state->mu = mu; return GSL_SUCCESS; } /* cholesky_solve() Compute (J^T J + mu D^T D) x = -J^T f Inputs: f - right hand side vector f x - (output) solution vector vstate - cholesky workspace */ static int cholesky_solve(const gsl_vector * f, gsl_vector *x, const void * vtrust_state, void *vstate) { const gsl_multifit_nlinear_trust_state *trust_state = (const gsl_multifit_nlinear_trust_state *) vtrust_state; cholesky_state_t *state = (cholesky_state_t *) vstate; int status; /* compute rhs = -J^T f */ gsl_blas_dgemv(CblasTrans, -1.0, trust_state->J, f, 0.0, state->rhs); status = cholesky_solve_rhs(state->rhs, x, state); if (status) return status; return GSL_SUCCESS; } static int cholesky_rcond(double * rcond, void * vstate) { int status; cholesky_state_t *state = (cholesky_state_t *) vstate; double rcond_JTJ; if (state->mu != 0) { /* * Cholesky decomposition hasn't been computed yet, or was computed * with mu > 0 - recompute Cholesky decomposition of J^T J */ /* copy lower triangle of JTJ to workspace */ gsl_matrix_tricpy('L', 1, state->work_JTJ, state->JTJ); /* compute modified Cholesky decomposition */ status = gsl_linalg_mcholesky_decomp(state->work_JTJ, state->perm, NULL); if (status) return status; } status = gsl_linalg_mcholesky_rcond(state->work_JTJ, state->perm, &rcond_JTJ, state->work3p); if (status == GSL_SUCCESS) *rcond = sqrt(rcond_JTJ); return status; } /* solve: (J^T J + mu D^T D) x = b */ static int cholesky_solve_rhs(const gsl_vector * b, gsl_vector *x, cholesky_state_t *state) { int status; gsl_matrix *JTJ = state->work_JTJ; status = gsl_linalg_mcholesky_solve(JTJ, state->perm, b, x); if (status) return status; return GSL_SUCCESS; } /* A <- A + mu D^T D */ static int cholesky_regularize(const double mu, const gsl_vector * diag, gsl_matrix * A, cholesky_state_t * state) { (void) state; if (mu != 0.0) { size_t i; for (i = 0; i < diag->size; ++i) { double di = gsl_vector_get(diag, i); double *Aii = gsl_matrix_ptr(A, i, i); *Aii += mu * di * di; } } return GSL_SUCCESS; } static const gsl_multifit_nlinear_solver cholesky_type = { "cholesky", cholesky_alloc, cholesky_init, cholesky_presolve, cholesky_solve, cholesky_rcond, cholesky_free }; const gsl_multifit_nlinear_solver *gsl_multifit_nlinear_solver_cholesky = &cholesky_type;