Enum NoiseModel 

pub enum NoiseModel {
    Gaussian {
        sigma: Array1<f64>,
    },
    Poisson,
    Tweedie {
        p: f64,
        phi: Array1<f64>,
    },
    NegativeBinomial {
        theta: Array1<f64>,
    },
    Beta {
        phi: Array1<f64>,
    },
    Gamma {
        shape: Array1<f64>,
    },
    Bernoulli,
    TransformationNormalQuantile {
        grid_y: Array1<f64>,
        h_grid: Array2<f64>,
    },
}

Observation-noise model used for generative sampling.

Variants

Gaussian

Fields

sigma: Array1<f64>

Per-observation standard deviation.

Poisson

Tweedie

Fields

p: f64

phi: Array1<f64>

Per-observation dispersion φ (> 0). A scalar-dispersion fit broadcasts one value to every row; a dispersion location-scale fit (#913/#1125) supplies the fitted per-row φ = 1/exp(eta_d(x)).

NegativeBinomial

Fields

theta: Array1<f64>

Per-observation overdispersion θ (> 0); see Tweedie::phi.

Beta

Fields

phi: Array1<f64>

Per-observation precision φ (> 0); see Tweedie::phi.

Gamma

Fields

shape: Array1<f64>

Per-observation Gamma shape k (> 0), with mean-driven scale; see Tweedie::phi.

Bernoulli

TransformationNormalQuantile

Inverse-transform sampling for a conditional transformation-normal (CTM) model (issue #1613). The fitted latent transform h(·|x_i) is strictly increasing in y and h(Y|x) ~ N(0, 1), so a response-scale draw is Y = h⁻¹(Z | x_i) with Z ~ N(0, 1). The earlier generate path drew Gaussian noise around the mean, which produced latent-scale draws whose per-row mean moved the wrong way with the covariate; this variant instead samples from the genuine conditional law F(·|x).

Both this sampler and the response-scale conditional mean E[Y|x] used by predict (#1612) invert the SAME per-row monotone curve, materialized on a shared response grid, so the two paths cannot disagree on the underlying transform.

Fields

grid_y: Array1<f64>

Shared, strictly increasing response grid (length g ≥ 2).

h_grid: Array2<f64>

h_grid[[i, k]] = h(grid_y[k] | x_i), strictly increasing in k for every row i (one row per observation).

Implementations

impl NoiseModel

pub fn from_likelihood( likelihood: &LikelihoodSpec, nobs: usize, gaussian_scale: Option<f64>, ) -> Result<NoiseModel, EstimationError>

Single canonical mapping from a fitted LikelihoodSpec (response distribution + dispersion gaussian_scale) to the observation NoiseModel used for generative sampling. Both simulation (FamilyStrategy::simulate_noise) and generative inference (generativespec_from_predict) route through this one helper so the set of supported likelihoods and the interpretation of dispersion parameters can never diverge between the two paths.

nobs is the number of observations the resulting per-observation Gaussian sigma vector should span; it is ignored for families whose noise carries no per-observation state.

pub fn from_likelihood_with_per_row_dispersion( likelihood: &LikelihoodSpec, dispersion: Array1<f64>, ) -> Result<NoiseModel, EstimationError>

Build the observation NoiseModel for a dispersion location-scale fit (#1125) from a fitted PER-ROW dispersion surface dispersion[i] (the predictor’s exp(eta_d(x_i)) mapped into NoiseModel units — NB θ, Gamma shape, Beta φ directly, Tweedie φ as the reciprocal). Unlike from_likelihood, which broadcasts a single scalar dispersion to every row, this threads the genuine per-observation precision channel so generated data reproduces the fitted non-constant dispersion instead of coming out homoscedastic at the seed.

Trait Implementations

impl Clone for NoiseModel

fn clone(&self) -> NoiseModel

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for NoiseModel

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

Formats the value using the given formatter.