MomentJacobianPreconditioner

phasic.svgd.MomentJacobianPreconditioner(
    model,
    observed_data,
    theta_dim,
    param_transform=None,
    rewards=None,
    epsilon=1e-08,
)

Moment Jacobian preconditioner for multi-scale SVGD.

Computes J[k,j] = d(moment_k)/d(theta_j) via finite differences at a reference point, then uses column norms as scaling factors. Scaling D_j = ||J[:,j]|| (column norm), normalized to mean 1.

This is simpler and more robust than Fisher preconditioning because it avoids dividing by PMF values (which can blow up when PMF is small).

Parameters

model : callable

Model function: model(theta, data, rewards=None) -> (pmf, moments)

observed_data : array

Observation data points

theta_dim : int

Number of parameters (learnable dimensions only if fixed params exist)

param_transform : callable or None = None

Transformation from unconstrained to constrained space (e.g., softplus)

rewards : array or None = None

Optional rewards for multivariate models

epsilon : float = 1e-8

Floor for scaling values to avoid division by zero

Methods

Name	Description
compute_scaling	Compute Jacobian column norms at reference point and derive scaling.

compute_scaling

phasic.svgd.MomentJacobianPreconditioner.compute_scaling(theta_ref)

Compute Jacobian column norms at reference point and derive scaling.

Uses moment-matching to find a better reference point before computing the Jacobian. The provided theta_ref is used as a starting point for the moment-matching search.

Parameters

theta_ref : array(theta_dim)

Initial reference point in unconstrained space (same space as particles). Used as starting point for moment-matching refinement.