Struct FitInference 

pub struct FitInference {

    pub edf_by_block: Vec<f64>,
    pub penalty_block_trace: Vec<f64>,
    pub edf_total: f64,
    pub smoothing_correction: Option<Array2<f64>>,
    pub smoothing_correction_method: Option<SmoothingCorrectionMethod>,
    pub smoothing_correction_first_order: Option<Array2<f64>>,
    pub smoothing_correction_method_first_order: Option<SmoothingCorrectionMethod>,
    pub penalized_hessian: UnscaledPrecision,
    pub reparam_qs: Option<Array2<f64>>,
    pub dispersion: Dispersion,
    pub beta_covariance: Option<PhiScaledCovariance>,
    pub beta_standard_errors: Option<Array1<f64>>,
    pub beta_covariance_corrected: Option<Array2<f64>>,
    pub beta_standard_errors_corrected: Option<Array1<f64>>,
    pub beta_covariance_frequentist: Option<Array2<f64>>,
    pub coefficient_influence: Option<Array2<f64>>,
    pub weighted_gram: Option<Array2<f64>>,
    pub bias_correction_beta: Option<Array1<f64>>,
    pub bias_correction_jacobian: Option<Array2<f64>>,
}

Fields

edf_by_block: Vec<f64>

penalty_block_trace: Vec<f64>

Raw per-penalty-block trace tr_kk = λ_kk·tr(H⁻¹ S_kk), one entry per smoothing parameter (aligned 1:1 with lambdas, like edf_by_block). Unclamped, in either coefficient basis (the trace of a matrix product is basis-invariant). This is the quantity both the dense and survival EDF paths already form internally; storing it lets per-term EDF be assembled as edf_term = |coeff_range| − Σ_{kk∈term} tr_kk, which equals the trace of the influence matrix F = H⁻¹X'WX over the term’s coefficient block, is additive across terms, and sums exactly to edf_total. The legacy Σ_kk(rank(S_kk) − tr_kk) block-sum over-counts whenever several penalties share one coefficient range (te/ti, anisotropic, adaptive), reporting a per-term EDF that can exceed the model total (issue #1219). May be empty for fits produced before this field existed or by paths that do not record traces; consumers fall back to coefficient_influence.

edf_total: f64

smoothing_correction: Option<Array2<f64>>

smoothing_correction_method: Option<SmoothingCorrectionMethod>

Method that produced smoothing_correction. Required whenever a matrix is present; None means no correction was retained.

smoothing_correction_first_order: Option<Array2<f64>>

The exact first-order IFT smoothing-parameter-uncertainty correction, RETAINED even when smoothing_correction/smoothing_correction_method above hold a cubature upgrade instead. compute_smoothing_correction_auto always computes the first-order correction before deciding whether to escalate to sigma-point cubature; discarding it once cubature is chosen made the #946 WPS-corrected-EDF/AIC channel go dark precisely when smoothing-parameter uncertainty is large enough to matter — the regime the correction exists for (see model_comparison_from_unified’s method_certified_exact gate, which is exact-provenance-only by design). Some exactly when smoothing_correction_method_first_order is Some(FirstOrderIdentifiedSubspace{..}); None when the first-order geometry itself was unavailable (mirrors smoothing_correction in that case — there is nothing to retain either way).

smoothing_correction_method_first_order: Option<SmoothingCorrectionMethod>

Provenance for smoothing_correction_first_order. Always either None or Some(FirstOrderIdentifiedSubspace{..}) — this field never holds SigmaPointCubature, unlike smoothing_correction_method above.

penalized_hessian: UnscaledPrecision

Penalised Hessian H = X'W_HX + S(λ) with NO dispersion scaling. When UnifiedFitResult::geometry is present, this matrix shares its exact active coefficient frame and therefore has dimension geometry.coefficient_gauge.reduced_total(). Without saved geometry it is in the saved/raw coefficient frame. Stored as [UnscaledPrecision] so callers that need the φ-scaled covariance Vb know they must pair this with Self::dispersion. #[serde(transparent)] on the newtype keeps the on-disk encoding identical to the pre-newtype Array2<f64> storage.

reparam_qs: Option<Array2<f64>>

dispersion: Dispersion

Dispersion/scale used to scale all coefficient covariance matrices. Dispersion is a validated newtype with no meaningful default (its source tag and φ are always established by the fit), so this field is required on the wire — no #[serde(default)], which would both demand a nonexistent Default impl and silently fabricate an unvalidated scale.

beta_covariance: Option<PhiScaledCovariance>

Conditional Bayesian covariance under fixed smoothing parameters (mgcv Vb). In an unreduced coefficient frame, Vb = H^{-1} * phi. With an active geometry gauge β = Tθ + a, the saved/raw covariance is Vb = T H_θ^{-1} Tᵀ * phi. Do not use an unscaled H^{-1} for standard errors when scale is estimated.

beta_standard_errors: Option<Array1<f64>>

Marginal SEs from beta_covariance.

beta_covariance_corrected: Option<Array2<f64>>

Optional smoothing-parameter-corrected Bayesian covariance (mgcv Vp): Vp = Vb + V_lambda, on the same dispersion scale as Vb. Usually this is first-order: Var*(β) ≈ Var(β|λ) + J Var(ρ) J^T; high-risk regimes may use adaptive cubature for higher-order terms.

beta_standard_errors_corrected: Option<Array1<f64>>

Marginal SEs from beta_covariance_corrected (Vp).

beta_covariance_frequentist: Option<Array2<f64>>

Frequentist covariance Ve = H⁻¹ X’WX H⁻¹ * φ̂.

coefficient_influence: Option<Array2<f64>>

Coefficient-space influence matrix F = H⁻¹ X’WX. Its trace is the total EDF.

weighted_gram: Option<Array2<f64>>

Weighted Gram X'WX = H − S(λ) in the original coefficient basis — symmetric PSD by construction. Stored directly (issue #1027) so the Wood–Pya–Säfken corrected-EDF correction tr(X'WX·Σ_ρ) pairs the true PSD Gram with Σ_ρ, rather than reconstructing it as H·F from a Hessian surface that need not satisfy H·F = X'WX (which made the correction indefinite and the corrected EDF drop below the conditional).

bias_correction_beta: Option<Array1<f64>>

O(n⁻¹) frequentist bias-correction vector b̂ = H⁻¹ S(λ̂) β̂ in the original (untransformed) coefficient basis. Predictions apply η̂_BC(x) = η̂(x) + s_*(x)^T b̂ to remove first-order shrinkage bias.

bias_correction_jacobian: Option<Array2<f64>>

O(n⁻¹) frequentist bias-correction Jacobian A = I + H⁻¹ S(λ̂) — the fixed-ρ linearization dβ_BC/dβ̂ of the bias-corrected coefficient β_BC = β̂ + b̂. A credible band centred at β_BC must report the covariance of that estimator, A·V·Aᵀ. The smoothing-corrected covariance already folds A in (see the optimizer), but the conditional covariance is stored raw, so prediction applies A to the conditional band through this Jacobian to avoid the over-narrow band #1870 documents. None when the full inverse (hence A) was unavailable.

Implementations

impl FitInference

pub fn validate_numeric_finiteness(&self) -> Result<(), EstimationError>

Trait Implementations

impl Clone for FitInference

fn clone(&self) -> FitInference

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for FitInference

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for FitInference

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for FitInference

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.