Struct InformationProjection 

pub struct InformationProjection {
    pub target_family: Vec<fn(&[f64]) -> f64>,
}

Information projection onto an exponential family.

Finds the distribution in the target family closest (in KL divergence) to a given distribution.

Fields

target_family: Vec<fn(&[f64]) -> f64>

Sufficient statistics defining the target exponential family.

Implementations

impl InformationProjection

pub fn new(target_family: Vec<fn(&[f64]) -> f64>) -> Self

Create a new InformationProjection.

pub fn project(&self, p: &[f64]) -> Vec<f64>

Project distribution p (given as a probability vector) onto the exponential family via moment matching (forward KL projection).

Returns the natural parameters θ of the projected distribution.

pub fn reverse_kl_projection(&self, p: &[f64], init_theta: &[f64]) -> Vec<f64>

Reverse KL projection: minimize KL(q ‖ p) over the exponential family.

Returns the natural parameters of the reverse projection. Uses a simple gradient descent on the KL divergence.