Struct FullRankGaussian 

pub struct FullRankGaussian {
    pub mean: Array1<f64>,
    pub chol_factor: Array2<f64>,
    pub dim: usize,
}

Full-rank Gaussian variational family

Approximates the posterior with a Gaussian with full covariance: q(z) = N(z; mu, Sigma) where Sigma = L L^T (Cholesky parameterization)

This can capture correlations but has O(d^2) parameters.

Fields

mean: Array1<f64>

Variational mean

chol_factor: Array2<f64>

Lower triangular Cholesky factor of the covariance Stored as a flattened lower-triangular matrix

dim: usize

Dimensionality

Implementations

impl FullRankGaussian

pub fn new(dim: usize) -> Result<Self>

Create a new full-rank Gaussian with given dimension

pub fn from_params(mean: Array1<f64>, chol_factor: Array2<f64>) -> Result<Self>

Create from specific parameters

pub fn covariance(&self) -> Array2<f64>

Get the covariance matrix: Sigma = L * L^T

pub fn precision(&self) -> Result<Array2<f64>>

Get the precision matrix (inverse covariance)

pub fn sample(&self, epsilon: &Array1<f64>) -> Result<Array1<f64>>

Sample from the variational distribution using reparameterization trick

z = mu + L * epsilon, where epsilon ~ N(0, I)

pub fn entropy(&self) -> f64

Compute the entropy of the full-rank Gaussian H[q] = 0.5 * d * (1 + log(2*pi)) + sum_i log(L_ii)

pub fn log_prob(&self, z: &Array1<f64>) -> Result<f64>

Compute log q(z) for a given z

pub fn n_params(&self) -> usize

Get total number of variational parameters

pub fn get_params(&self) -> Array1<f64>

Get all variational parameters as a flat vector Layout: [mean, lower-triangular elements of L]

pub fn set_params(&mut self, params: &Array1<f64>) -> Result<()>

Set variational parameters from a flat vector

Trait Implementations

impl Clone for FullRankGaussian

fn clone(&self) -> FullRankGaussian

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for FullRankGaussian

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

Formats the value using the given formatter.