/* normal.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. */ #include #include #include #include #include #include #include #include #include #include #include typedef struct { size_t p; /* number of columns of LS matrix */ gsl_matrix *ATA; /* A^T A, p-by-p */ gsl_vector *ATb; /* A^T b, p-by-1 */ double normb; /* || b || */ gsl_matrix *work_ATA; /* workspace for chol(ATA), p-by-p */ gsl_permutation *perm; /* permutation vector */ gsl_vector *workp; /* workspace size p */ gsl_vector *work3p; /* workspace size 3*p */ gsl_vector *D; /* scale factors for ATA, size p */ gsl_vector *c; /* solution vector for L-curve */ int eigen; /* 1 if eigenvalues computed */ double eval_min; /* minimum eigenvalue */ double eval_max; /* maximum eigenvalue */ gsl_eigen_symm_workspace *eigen_p; } normal_state_t; static void *normal_alloc(const size_t p); static void normal_free(void *vstate); static int normal_reset(void *vstate); static int normal_accumulate(gsl_matrix * A, gsl_vector * b, void * vstate); static int normal_solve(const double lambda, gsl_vector * x, double * rnorm, double * snorm, void * vstate); static int normal_rcond(double * rcond, void * vstate); static int normal_lcurve(gsl_vector * reg_param, gsl_vector * rho, gsl_vector * eta, void * vstate); static int normal_solve_system(const double lambda, gsl_vector * x, normal_state_t *state); static int normal_solve_cholesky(gsl_matrix * ATA, const gsl_vector * ATb, gsl_vector * x, normal_state_t *state); static int normal_calc_norms(const gsl_vector *x, double *rnorm, double *snorm, normal_state_t *state); static int normal_eigen(normal_state_t *state); /* normal_alloc() Allocate workspace for solving large linear least squares problems using the normal equations approach Inputs: p - number of columns of LS matrix Return: pointer to workspace */ static void * normal_alloc(const size_t p) { normal_state_t *state; if (p == 0) { GSL_ERROR_NULL("p must be a positive integer", GSL_EINVAL); } state = calloc(1, sizeof(normal_state_t)); if (!state) { GSL_ERROR_NULL("failed to allocate normal state", GSL_ENOMEM); } state->p = p; state->ATA = gsl_matrix_alloc(p, p); if (state->ATA == NULL) { normal_free(state); GSL_ERROR_NULL("failed to allocate ATA matrix", GSL_ENOMEM); } state->work_ATA = gsl_matrix_alloc(p, p); if (state->work_ATA == NULL) { normal_free(state); GSL_ERROR_NULL("failed to allocate temporary ATA matrix", GSL_ENOMEM); } state->ATb = gsl_vector_alloc(p); if (state->ATb == NULL) { normal_free(state); GSL_ERROR_NULL("failed to allocate ATb vector", GSL_ENOMEM); } state->perm = gsl_permutation_alloc(p); if (state->perm == NULL) { normal_free(state); GSL_ERROR_NULL("failed to allocate perm", GSL_ENOMEM); } state->D = gsl_vector_alloc(p); if (state->D == NULL) { normal_free(state); GSL_ERROR_NULL("failed to allocate D vector", GSL_ENOMEM); } state->workp = gsl_vector_alloc(p); if (state->workp == NULL) { normal_free(state); GSL_ERROR_NULL("failed to allocate temporary ATb vector", GSL_ENOMEM); } state->work3p = gsl_vector_alloc(3 * p); if (state->work3p == NULL) { normal_free(state); GSL_ERROR_NULL("failed to allocate work3p", GSL_ENOMEM); } state->c = gsl_vector_alloc(p); if (state->c == NULL) { normal_free(state); GSL_ERROR_NULL("failed to allocate c vector", GSL_ENOMEM); } state->eigen_p = gsl_eigen_symm_alloc(p); if (state->eigen_p == NULL) { normal_free(state); GSL_ERROR_NULL("failed to allocate eigen workspace", GSL_ENOMEM); } normal_reset(state); return state; } static void normal_free(void *vstate) { normal_state_t *state = (normal_state_t *) vstate; if (state->ATA) gsl_matrix_free(state->ATA); if (state->work_ATA) gsl_matrix_free(state->work_ATA); if (state->ATb) gsl_vector_free(state->ATb); if (state->perm) gsl_permutation_free(state->perm); if (state->D) gsl_vector_free(state->D); if (state->workp) gsl_vector_free(state->workp); if (state->work3p) gsl_vector_free(state->work3p); if (state->c) gsl_vector_free(state->c); if (state->eigen_p) gsl_eigen_symm_free(state->eigen_p); free(state); } static int normal_reset(void *vstate) { normal_state_t *state = (normal_state_t *) vstate; gsl_matrix_set_zero(state->ATA); gsl_vector_set_zero(state->ATb); state->normb = 0.0; state->eigen = 0; state->eval_min = 0.0; state->eval_max = 0.0; return GSL_SUCCESS; } /* normal_accumulate() Add a new block of rows to the normal equations system Inputs: A - new block of rows, n-by-p b - new rhs vector n-by-1 vstate - workspace Return: success/error */ static int normal_accumulate(gsl_matrix * A, gsl_vector * b, void * vstate) { normal_state_t *state = (normal_state_t *) vstate; const size_t n = A->size1; if (A->size2 != state->p) { GSL_ERROR("columns of A do not match workspace", GSL_EBADLEN); } else if (n != b->size) { GSL_ERROR("A and b have different numbers of rows", GSL_EBADLEN); } else { int s; /* ATA += A^T A, using only the lower half of the matrix */ s = gsl_blas_dsyrk(CblasLower, CblasTrans, 1.0, A, 1.0, state->ATA); if (s) return s; /* ATb += A^T b */ s = gsl_blas_dgemv(CblasTrans, 1.0, A, b, 1.0, state->ATb); if (s) return s; /* update || b || */ state->normb = gsl_hypot(state->normb, gsl_blas_dnrm2(b)); return GSL_SUCCESS; } } /* normal_solve() Solve normal equations system: (A^T A + \lambda^2 I) x = A^T b using Cholesky decomposition Inputs: lambda - regularization parameter x - (output) solution vector p-by-1 rnorm - (output) residual norm ||b - A x|| snorm - (output) solution norm ||x|| vstate - workspace Return: success/error */ static int normal_solve(const double lambda, gsl_vector * x, double * rnorm, double * snorm, void * vstate) { normal_state_t *state = (normal_state_t *) vstate; if (x->size != state->p) { GSL_ERROR("solution vector does not match workspace", GSL_EBADLEN); } else { int status; /* solve system (A^T A) x = A^T b */ status = normal_solve_system(lambda, x, state); if (status) { GSL_ERROR("failed to solve normal equations", status); } /* compute residual norm ||y - X c|| and solution norm ||x|| */ normal_calc_norms(x, rnorm, snorm, state); return GSL_SUCCESS; } } static int normal_rcond(double * rcond, void * vstate) { normal_state_t *state = (normal_state_t *) vstate; int status = GSL_SUCCESS; double rcond_ATA; status = gsl_linalg_pcholesky_rcond(state->work_ATA, state->perm, &rcond_ATA, state->work3p); if (status == GSL_SUCCESS) *rcond = sqrt(rcond_ATA); return status; } /* normal_lcurve() Compute L-curve of least squares system Inputs: reg_param - (output) vector of regularization parameters rho - (output) vector of residual norms eta - (output) vector of solution norms vstate - workspace Return: success/error */ static int normal_lcurve(gsl_vector * reg_param, gsl_vector * rho, gsl_vector * eta, void * vstate) { normal_state_t *state = (normal_state_t *) vstate; int status; double smin, smax; /* minimum/maximum singular values */ size_t i; if (state->eigen == 0) { status = normal_eigen(state); if (status) return status; } if (state->eval_max < 0.0) { GSL_ERROR("matrix is not positive definite", GSL_EDOM); } /* compute singular values which are sqrts of eigenvalues */ smax = sqrt(state->eval_max); if (state->eval_min > 0.0) smin = sqrt(state->eval_min); else smin = 0.0; /* compute vector of regularization parameters */ gsl_multifit_linear_lreg(smin, smax, reg_param); /* solve normal equations for each regularization parameter */ for (i = 0; i < reg_param->size; ++i) { double lambda = gsl_vector_get(reg_param, i); double rnorm, snorm; status = normal_solve_system(lambda, state->c, state); if (status) return status; /* compute ||y - X c|| and ||c|| */ normal_calc_norms(state->c, &rnorm, &snorm, state); gsl_vector_set(rho, i, rnorm); gsl_vector_set(eta, i, snorm); } return GSL_SUCCESS; } /* normal_solve_system() Compute solution to normal equations: (A^T A + lambda^2*I) x = A^T b using LDL decomposition. Inputs: x - (output) solution vector state - workspace Return: success/error */ static int normal_solve_system(const double lambda, gsl_vector * x, normal_state_t *state) { int status; const double lambda_sq = lambda * lambda; gsl_vector_view d = gsl_matrix_diagonal(state->work_ATA); /* copy ATA matrix to temporary workspace and regularize */ gsl_matrix_tricpy('L', 1, state->work_ATA, state->ATA); gsl_vector_add_constant(&d.vector, lambda_sq); /* solve with LDL decomposition */ status = normal_solve_cholesky(state->work_ATA, state->ATb, x, state); if (status) return status; return status; } static int normal_solve_cholesky(gsl_matrix * ATA, const gsl_vector * ATb, gsl_vector * x, normal_state_t *state) { int status; status = gsl_linalg_pcholesky_decomp2(ATA, state->perm, state->D); if (status) return status; status = gsl_linalg_pcholesky_solve2(ATA, state->perm, state->D, ATb, x); if (status) return status; return GSL_SUCCESS; } /* normal_calc_norms() Compute residual norm ||y - X c|| and solution norm ||c|| Inputs: x - solution vector rnorm - (output) residual norm ||y - X c|| snorm - (output) solution norm ||c|| state - workspace */ static int normal_calc_norms(const gsl_vector *x, double *rnorm, double *snorm, normal_state_t *state) { double r2; /* compute solution norm ||x|| */ *snorm = gsl_blas_dnrm2(x); /* compute residual norm ||b - Ax|| */ /* compute: A^T A x - 2 A^T b */ gsl_vector_memcpy(state->workp, state->ATb); gsl_blas_dsymv(CblasLower, 1.0, state->ATA, x, -2.0, state->workp); /* compute: x^T A^T A x - 2 x^T A^T b */ gsl_blas_ddot(x, state->workp, &r2); /* add b^T b */ r2 += state->normb * state->normb; *rnorm = sqrt(r2); return GSL_SUCCESS; } /* normal_eigen() Compute eigenvalues of A^T A matrix, which are stored in state->workp on output. Also, state->eval_min and state->eval_max are set to the minimum/maximum eigenvalues */ static int normal_eigen(normal_state_t *state) { int status; /* copy lower triangle of ATA to temporary workspace */ gsl_matrix_tricpy('L', 1, state->work_ATA, state->ATA); /* compute eigenvalues of ATA */ status = gsl_eigen_symm(state->work_ATA, state->workp, state->eigen_p); if (status) return status; gsl_vector_minmax(state->workp, &state->eval_min, &state->eval_max); state->eigen = 1; return GSL_SUCCESS; } static const gsl_multilarge_linear_type normal_type = { "normal", normal_alloc, normal_reset, normal_accumulate, normal_solve, normal_rcond, normal_lcurve, normal_free }; const gsl_multilarge_linear_type * gsl_multilarge_linear_normal = &normal_type;