romtools.workflows.inverse.vi.newton#
Natural-coordinate Newton support for full-covariance Gaussian VI/MF-VI.
This module augments the full-covariance score-function implementation with a
second-order curvature model in (mu, svec(Sigma)) coordinates. Newton
steps can be formed in ordinary coordinates or in locally Fisher-whitened
coordinates; the latter is selected with newton_metric='natural'.
Functions
|
|
|
Estimate full-covariance ELBO curvature using log-joint score terms. |
Classes
|
Newton state owned by one run; only accepted states commit curvature. |
- class romtools.workflows.inverse.vi.newton.FullCovarianceNewton(config, *, multifidelity=False)[source]#
Bases:
objectNewton state owned by one run; only accepted states commit curvature.
- romtools.workflows.inverse.vi.newton.estimate_ordinary_hessian(optimizer_samples, mean, cholesky, scaled_log_joint_terms, baseline_method, elbo_scaling_factor)[source]#
Estimate full-covariance ELBO curvature using log-joint score terms.
The optional joint entropy strategy affects the gradient estimator. Newton curvature follows the existing romtools split estimator: log-joint score curvature plus analytic Gaussian-entropy curvature.
- Parameters:
optimizer_samples (ndarray)
mean (ndarray)
cholesky (ndarray)
scaled_log_joint_terms (ndarray)
baseline_method (str)
elbo_scaling_factor (float)
- Return type:
ndarray