IMP_Eigen::MatrixXd
GaussianProcessInterpolation::get_posterior_covariance_matrix(FloatsList x)
const {
unsigned N(x.size());
IMP_Eigen::MatrixXd Wpri(M_, N);
for (unsigned i = 0; i < N; i++) Wpri.col(i) = get_wx_vector(x[i]);
IMP_Eigen::LDLT<IMP_Eigen::MatrixXd, IMP_Eigen::Upper> ldlt(get_ldlt());
// we can now use covariance_function_ because it is up to date
IMP_Eigen::MatrixXd Wpost((*covariance_function_)(x));
return Wpost - Wpri.transpose() * ldlt.solve(Wpri);
}
IMPISD_BEGIN_NAMESPACE
GaussianProcessInterpolation::GaussianProcessInterpolation(
FloatsList x, Floats sample_mean, Floats sample_std, unsigned n_obs,
UnivariateFunction *mean_function, BivariateFunction *covariance_function,
Particle *sigma, double sparse_cutoff)
: Object("GaussianProcessInterpolation%1%"),
x_(x),
n_obs_(n_obs),
mean_function_(mean_function),
covariance_function_(covariance_function),
sigma_(sigma),
cutoff_(sparse_cutoff) {
// O(M)
// store dimensions
M_ = x.size();
N_ = x[0].size();
sigma_val_ = Scale(sigma_).get_nuisance();
// basic checks
IMP_USAGE_CHECK(sample_mean.size() == M_,
"sample_mean should have the same size as x");
IMP_USAGE_CHECK(sample_std.size() == M_,
"sample_std should have the same size as x");
IMP_USAGE_CHECK(mean_function->get_ndims_x() == N_,
"mean_function should have " << N_ << " input dimensions");
IMP_USAGE_CHECK(mean_function->get_ndims_y() == 1,
"mean_function should have 1 output dimension");
IMP_USAGE_CHECK(covariance_function->get_ndims_x1() == N_,
"covariance_function should have "
<< N_ << " input dimensions for first vector");
IMP_USAGE_CHECK(covariance_function->get_ndims_x2() == N_,
"covariance_function should have "
<< N_ << " input dimensions for second vector");
IMP_USAGE_CHECK(covariance_function->get_ndims_y() == 1,
"covariance_function should have 1 output dimension");
// set all flags to false = need update.
force_mean_update();
force_covariance_update();
// compute needed matrices
compute_I(sample_mean);
compute_S(sample_std);
}
