Struct Gaussian 

pub struct Gaussian {
    pub mean: Vec<f64>,
    pub cov: Vec<f64>,
    /* private fields */
}

Fields

mean: Vec<f64>

Posterior mean — length D.

cov: Vec<f64>

Posterior covariance, row-major D×D.

Implementations

impl Gaussian

pub fn new(mean: Vec<f64>, cov: Vec<f64>) -> Result<Self, BLRError>

Create a new Gaussian, validating dimensions.

pub fn dim(&self) -> usize

Dimension D of the distribution.

pub fn std(&self) -> Vec<f64>

Per-dimension standard deviations: sqrt(diag(cov)).

pub fn marginal(&self, i: usize) -> (f64, f64)

Marginal distribution at index i: returns (mean[i], std[i]).

pub fn log_pdf(&self, x: &[f64]) -> f64

Log probability density log N(x; mean, cov).

Uses Cholesky of cov for numerical stability.

pub fn condition( self, a: &[f64], n_obs: usize, d_feat: usize, y: &[f64], noise_variance: f64, ) -> Result<Self, BLRError>

Bayesian update: computes p(self | y) where y = A·self + ε, ε ~ N(0, σ²·I_N) (homoscedastic noise).

a is the measurement matrix A (n_obs × d_feat), row-major flat slice. noise_variance is the scalar observation noise variance σ² > 0.

Adaptive dispatch

Two algebraically equivalent forms are available; this method selects whichever minimises the size of the required Cholesky factorisation:

Condition	Form chosen	Cholesky size
n_obs < d_feat	Gram / Kalman-gain (observation-space)	N×N
n_obs >= d_feat	Precision / Woodbury (parameter-space)	D×D

The two forms are related by the Woodbury matrix identity; see dev/blog/blr-and-ard.md Appendix A for the derivation. The precision form derives Σ_prior⁻¹ directly from self.cov — no isotropic approximation is made, and the forms agree within floating-point rounding error.

Returns the updated Gaussian representing the posterior distribution.