Struct PACBayesGeneralization 

pub struct PACBayesGeneralization {
    pub kl_qp: f64,
    pub n: usize,
    pub delta: f64,
}

PAC-Bayes generalization bound computation.

McAllester’s bound: L_D(Q) ≤ L_S(Q) + √((KL(Q‖P) + ln(n/δ)) / (2n)).

Fields

kl_qp: f64

KL divergence KL(Q‖P) in nats.

n: usize

Number of training samples n.

delta: f64

Confidence parameter δ.

Implementations

impl PACBayesGeneralization

pub fn new(kl_qp: f64, n: usize, delta: f64) -> Self

Create a new PAC-Bayes bound instance.

pub fn mcallester_bound(&self, empirical_loss: f64) -> f64

McAllester bound: empirical_loss + √((KL + ln(n/δ)) / (2n)).

pub fn catoni_bound(&self, empirical_loss: f64, lambda: f64) -> f64

Catoni’s tighter bound (using solve for λ-parameterized form). Approximation: L_D(Q) ≤ (1/(1-λ/2)) * (L_S(Q) + KL/(2λn)).

pub fn optimal_lambda(&self, empirical_loss: f64) -> f64

Optimal λ for Catoni bound (minimizing RHS).