Struct PenaltyConfig 

pub struct PenaltyConfig<'a> {
    pub canonical_penalties: &'a [CanonicalPenalty],
    pub balanced_penalty_root: Option<&'a Array2<f64>>,
    pub reparam_invariant: Option<&'a ReparamInvariant>,
    pub p: usize,
    pub coefficient_lower_bounds: Option<&'a Array1<f64>>,
    pub linear_constraints_original: Option<&'a LinearInequalityConstraints>,
    pub penalty_shrinkage_floor: Option<f64>,
    pub kronecker_factored: Option<&'a KroneckerFactoredBasis>,
}

Fields

canonical_penalties: &'a [CanonicalPenalty]

Block-local canonical penalties with precomputed roots and spectral data. This is the single canonical penalty representation — no full-width rank × p roots are stored. When the reparameterization engine needs full-width roots, they are derived on-the-fly from these block-local roots.

balanced_penalty_root: Option<&'a Array2<f64>>

reparam_invariant: Option<&'a ReparamInvariant>

p: usize

coefficient_lower_bounds: Option<&'a Array1<f64>>

linear_constraints_original: Option<&'a LinearInequalityConstraints>

penalty_shrinkage_floor: Option<f64>

Relative shrinkage floor for eigenvalues of the penalized block. If Some(epsilon), a rho-independent ridge of epsilon * max_balanced_eigenvalue is added to prevent barely-penalized directions from causing pathological non-Gaussianity in the posterior. Typical value: 1e-6. None disables.

kronecker_factored: Option<&'a KroneckerFactoredBasis>

When set, the penalties have Kronecker (tensor-product) structure. The reparameterization engine will use factored Qs = U_1 ⊗ … ⊗ U_d instead of eigendecomposing the full p×p balanced penalty.