Struct VarianceJet 

pub struct VarianceJet {
    pub v: f64,
    pub v1: f64,
    pub v2: f64,
    pub v3: f64,
    pub v4: f64,
}

Fields

v: f64

v1: f64

v2: f64

v3: f64

v4: f64

Implementations

impl VarianceJet

pub fn bernoulli(mu: f64) -> Self

Bernoulli / binomial variance V(μ) = μ(1−μ).

pub fn bernoulli_with_complement(mu: f64, one_minus_mu: f64) -> Self

Bernoulli variance jet from the (μ, 1−μ) PAIR (#2273).

1.0 - mu is a hard zero once mu rounds to exactly 1.0, which a bounded inverse link reaches far inside its tail — cloglog at η ≈ 3.62, probit at η ≈ 8.29 — while the true complement is still a perfectly representable ~1e-18. V = μ(1−μ) then collapses to 0 and every observed-information quantity divides by it, so a row whose Fisher weight is a healthy 1e-14 is refused as unrepresentable instead of evaluated. crate::mixture_link::inverse_link_complement_for_inverse_link recovers the complement from η without cancellation; this constructor is how it reaches the variance.

Only v changes: v1 = 1 − 2μ needs no complement (it is −1 to full precision wherever μ has saturated, from either expression), and rewriting it would move existing well-conditioned fits by an ulp for no accuracy gained.

pub fn poisson(mu: f64) -> Self

Poisson variance V(μ) = μ.

pub fn gamma(mu: f64) -> Self

Gamma variance V(μ) = μ².

pub fn tweedie(mu: f64, p: f64) -> Self

Tweedie variance V(μ) = μ^p.

pub fn negative_binomial(mu: f64, theta: f64) -> Self

Negative-binomial variance V(μ) = μ + μ² / theta.

pub fn gaussian() -> Self

Gaussian (identity) variance V(μ) = 1.

pub fn binomial_n(mu: f64, one_minus_mu: f64) -> Self

Binomial(n, p) variance V(p) = p(1−p), identical to Bernoulli.

The trial count n enters as a prior-weight multiplier, not through the variance function itself.

pub fn beta(mu: f64, one_minus_mu: f64, phi: f64) -> Self

Beta-regression variance V(μ) = μ(1−μ)/(1+φ).