Struct GateBlock 

pub struct GateBlock<'a> {
    pub design: ArrayView2<'a, f64>,
    pub y: ArrayView1<'a, f64>,
    pub b: ArrayView1<'a, f64>,
    pub offset: Option<ArrayView1<'a, f64>>,
    pub psi_hat: Option<ArrayView1<'a, f64>>,
    pub penalty: Option<ArrayView2<'a, f64>>,
    pub hess_rest: Option<ArrayView2<'a, f64>>,
    pub h_rest: Option<ArrayView1<'a, f64>>,
}

The data of one logit gate block to be evidence-integrated.

All matrices are in the gate coordinates g (dimension d_g). The h_rest / hess_rest blocks carry whatever the surrounding arrow Schur system contributes to the gate coordinates from the rest of the model (decoder/coordinate cross-terms already Schur-folded in); pass zeros when the gate block is isolated.

Fields

design: ArrayView2<'a, f64>

Gate design X_g, shape (n, d_g): row i is x_i with ψ_i = x_iᵀγ + o_i.

y: ArrayView1<'a, f64>

Binomial responses y_i (counts; 0..=b_i).

b: ArrayView1<'a, f64>

Binomial shapes b_i (1.0 for Bernoulli).

offset: Option<ArrayView1<'a, f64>>

Per-row offset o_i (the fixed part of the gate logit). Empty ⇒ zeros.

psi_hat: Option<ArrayView1<'a, f64>>

Current gate linear predictor ψ̂_i used to tilt the PG law (the inner optimum’s logits). Empty ⇒ the untilted PG(b, 0) rule.

penalty: Option<ArrayView2<'a, f64>>

Penalty S_g on the gate coordinates (d_g × d_g, SPD-or-PSD). Empty ⇒ zero.

hess_rest: Option<ArrayView2<'a, f64>>

Rest-of-model Hessian contribution to the gate coordinates H_rest,gg (d_g × d_g). Empty ⇒ zero.

h_rest: Option<ArrayView1<'a, f64>>

Rest-of-model linear contribution h_rest,g (length d_g). Empty ⇒ zero.