Struct GaussianLocationScaleFitResult 

pub struct GaussianLocationScaleFitResult {
    pub fit: BlockwiseTermFitResult,
    pub wiggle_knots: Option<Array1<f64>>,
    pub wiggle_degree: Option<usize>,
    pub beta_link_wiggle: Option<Vec<f64>>,
    pub response_scale: f64,
}

Fields

fit: BlockwiseTermFitResult

wiggle_knots: Option<Array1<f64>>

wiggle_degree: Option<usize>

beta_link_wiggle: Option<Vec<f64>>

response_scale: f64

Response standardization factor applied internally during fitting.

The Gaussian location-scale path fits on y / response_scale so the fixed log-σ soft floor LOGB_SIGMA_FLOOR = 0.01 is operationally scale-relative (1 % of the response spread) rather than absolute, keeping κ = dlogσ/dη ≈ 1 across the realistic σ range and informing the scale block like gamlss. The returned coefficient blocks, beta, and link-wiggle knots/coefficients are already mapped back to raw response units (the Location/Mean block scaled by response_scale, the Scale block intercept shifted by +ln(response_scale)), so downstream reconstruction μ = X_mean·β comes out in raw units with no further rescaling.

The σ reconstruction, however, must scale the floor too to stay response-scale-equivariant (#884):

σ = response_scale·LOGB_SIGMA_FLOOR + exp(X_scale·β)
  = response_scale·(LOGB_SIGMA_FLOOR + exp(η_internal)).

The intercept shift carries only the exp(η) term; reconstructing with a raw LOGB_SIGMA_FLOOR instead of response_scale·LOGB_SIGMA_FLOOR leaves the non-equivariant residual LOGB_SIGMA_FLOOR·(1 − response_scale).

This field records the factor that was applied for transparency, covariance bookkeeping, and the equivariant σ-floor reconstruction; it is 1.0 when no standardization was needed (degenerate constant response).