50 lines
2.1 KiB
C++
50 lines
2.1 KiB
C++
|
#pragma once
|
||
|
|
|
||
|
|
#include "spectra_reml/types.hpp"
|
||
|
|
|
||
|
|
#include <limits>
|
||
|
|
#include <string>
|
||
|
|
#include <vector>
|
||
|
|
|
||
|
|
namespace spectra::reml {
|
||
|
|
|
||
|
|
// Fits y = X beta + g + e in the GRM eigenspace, where
|
||
|
|
// Var(y) = sigma_g^2 diag(lambda) + sigma_e^2 I.
|
||
|
|
//
|
||
|
|
// The parameters are the signed standard deviations (sigma_e, sigma_g), as in
|
||
|
|
// thesis Equations 4.24-4.25. The covariance depends on their squares, so the
|
||
|
|
// optimization is unconstrained. It uses AI-REML and a strong-Wolfe line
|
||
|
|
// search; no EM update is performed and P is never materialized.
|
||
|
|
[[nodiscard]] RemlResult fit_ai_reml_spectral(
|
||
|
|
const std::vector<double>& y_star, const ColMajorMatrix& x_star,
|
||
|
|
const std::vector<double>& eigenvalues,
|
||
|
|
const RemlOptions& options = {});
|
||
|
|
|
||
|
|
// Exposed for finite-difference and independent-oracle tests. The ordering is
|
||
|
|
// theta=(sigma_e,sigma_g); ai is [ee,eg;eg,gg] in column-major order.
|
||
|
|
struct RemlEvaluation {
|
||
|
|
bool valid = false;
|
||
|
|
double log_likelihood = -std::numeric_limits<double>::infinity();
|
||
|
|
double gradient_e = std::numeric_limits<double>::quiet_NaN();
|
||
|
|
double gradient_g = std::numeric_limits<double>::quiet_NaN();
|
||
|
|
// Scores with respect to the variance parameters v_e=sigma_e^2 and
|
||
|
|
// v_g=sigma_g^2. These remain informative when sigma_g=0, unlike the
|
||
|
|
// signed-standard-deviation gradient gradient_g=2*sigma_g*score_v_g.
|
||
|
|
double variance_score_e = std::numeric_limits<double>::quiet_NaN();
|
||
|
|
double variance_score_g = std::numeric_limits<double>::quiet_NaN();
|
||
|
|
double ai_ee = std::numeric_limits<double>::quiet_NaN();
|
||
|
|
double ai_eg = std::numeric_limits<double>::quiet_NaN();
|
||
|
|
double ai_gg = std::numeric_limits<double>::quiet_NaN();
|
||
|
|
std::vector<double> beta;
|
||
|
|
std::vector<double> beta_covariance;
|
||
|
|
std::string error;
|
||
|
|
};
|
||
|
|
|
||
|
|
[[nodiscard]] RemlEvaluation evaluate_reml_spectral(
|
||
|
|
const std::vector<double>& y_star, const ColMajorMatrix& x_star,
|
||
|
|
const std::vector<double>& eigenvalues, double sigma_e, double sigma_g,
|
||
|
|
bool compute_ai = true, bool compute_beta_covariance = false,
|
||
|
|
double covariance_floor_relative = 1e-12);
|
||
|
|
|
||
|
|
} // namespace spectra::reml
|