Provided Matérn container classes

gpmpcontrib.models provides configured ModelContainer subclasses. Each container stores one independent gpmp.core.Model per output and builds a gpmp.parameter.Param object from the selected raw vectors. The common select_params, predict, run_diagnosis, and state-access contracts are documented in Model containers.

Naming and covariance parameters

Maternp fixes the Matérn regularity at \(\nu=p+1/2\) and uses

[log(sigma2), -log(rho_0), ..., -log(rho_{d-1})]

Matern selects \(\nu\) and uses

[log(sigma2), log(nu), -log(rho_0), ..., -log(rho_{d-1})]

The corresponding Param paths are covparam/variance, covparam/regularity, and covparam/lengthscale. Variance and regularity use logarithmic normalization. Lengthscales use inverse-log normalization.

Constructor contracts

class Model_ConstantMean_Maternp_ML(name, output_dim, covariance_specification=None)

Build output_dim fixed-regularity Matérn models. Each output has one explicit constant-mean parameter. Mean and covariance parameters are selected by maximum likelihood. covariance_specification must contain {"p": int} unless a custom covariance function is supplied.

class Model_ConstantMean_Maternp_REML(name, output_dim, mean_specification, covariance_specification)

Build fixed-regularity Matérn models with covariance parameters selected by restricted likelihood. mean_specification accepts {"type": "constant"} or {"type": "linear"}. covariance_specification must contain {"p": int} unless a custom covariance function is supplied.

class Model_ConstantMean_Matern_REML(name, output_dim, mean_specification, covariance_specification=None)

Build Matérn models whose variance, regularity, and componentwise lengthscales are selected by restricted likelihood. A d-dimensional input gives d + 2 covariance parameters. covariance_specification is normally omitted; the class uses gp.kernel.matern_covariance and its default numerical policy.

The initial guess is computed by gp.kernel.anisotropic_parameters_initial_guess_matern. After select_params, use model[k].get_param() and path ["covparam", "regularity"] to read log(nu). Exponentiation gives the selected positive regularity.

class Model_ConstantMean_Maternp_REMAP_logsigma2(name, output_dim, mean_specification, covariance_specification, gamma=None, sigma2_coverage=None)

Build fixed-regularity Matérn models selected by REMAP with a Gaussian prior on log(sigma2). Missing hyperparameters are resolved from gpmp.kernel.prior_defaults when select_params builds the criterion.

class Model_ConstantMean_Maternp_REMAP_logsigma2_and_logrho_prior(name, output_dim, mean_specification, covariance_specification, gamma=None, sigma2_coverage=None, alpha=None, rho_min_range_factor=None, logrho_min=None, covparam0_prior=None, logsigma2_0_prior=None, logrho_0_prior=None)

Build fixed-regularity Matérn models selected by REMAP with priors on log(sigma2) and logrho. Direct anchors logsigma2_0_prior and logrho_0_prior take priority over covparam0_prior. Missing anchors are obtained from the anisotropic initial guess on the current observations.

class Model_ConstantMean_Maternp_REMAP

Alias of Model_ConstantMean_Maternp_REMAP_logsigma2_and_logrho_prior.

class Model_Noisy_ConstantMean_Maternp_REML(name, output_dim, mean_specification, covariance_specification)

Build fixed-regularity Matérn models with known, point-dependent observation noise. The input matrix contains physical coordinates followed by one noise variance column per output. Prediction points use zero in these columns when prediction concerns the latent function.

Parameter adapters

Model_ConstantMean_Matern_REML.build_param_procedures accepts the optional arguments logsigma2_bounds, lognu_bounds, loginvrho_bounds, and name_prefix. It returns (param_from_vectors, vectors_from_param). Bounds use the normalized coordinates stored in Param:

import gpmp.num as gnp

param_from_vectors, vectors_from_param = model.build_param_procedures(
    0,
    logsigma2_bounds=(-10.0, 10.0),
    lognu_bounds=(-2.0, 4.0),
    loginvrho_bounds=(-8.0, 8.0),
)
param = param_from_vectors(gnp.asarray([]), covparam)

The first callable accepts (meanparam, covparam) and returns a Param object. The second accepts a Param object and returns (meanparam, covparam).

Prior access

Model_ConstantMean_Maternp_REMAP_logsigma2.set_prior(*, gamma=..., sigma2_coverage=..., covparam0_prior=..., logsigma2_0_prior=..., output_idx=None)

Update the supplied log-variance prior fields. Omitted fields remain unchanged. output_idx=None applies each update to every output.

Model_ConstantMean_Maternp_REMAP_logsigma2_and_logrho_prior.set_prior(*, gamma=..., sigma2_coverage=..., alpha=..., rho_min_range_factor=..., logrho_min=..., covparam0_prior=..., logsigma2_0_prior=..., logrho_0_prior=..., output_idx=None)

Update the supplied log-variance and log-lengthscale prior fields. Omitted fields remain unchanged. output_idx=None applies each update to every output. covparam0_prior uses covariance coordinates [log(sigma2), -log(rho_0), ...]. logrho_0_prior uses logrho and therefore has the opposite sign on lengthscale coordinates.

get_prior(output_idx=None, resolved=True)

Return one resolved prior object or a list with one object per output. resolved=True raises ValueError before data-dependent defaults have been resolved by select_params.