Struct StatisticalDiffusionBackend 

pub struct StatisticalDiffusionBackend { /* private fields */ }

A diffusion backend that generates samples matching target statistical properties.

The forward process adds Gaussian noise according to the noise schedule. The reverse process uses Langevin-inspired updates to guide samples toward the target distribution (means, standard deviations, correlations).

Implementations

impl StatisticalDiffusionBackend

pub fn new(means: Vec<f64>, stds: Vec<f64>, config: DiffusionConfig) -> Self

Create a new statistical diffusion backend.

Arguments

• means - Target means for each feature dimension
• stds - Target standard deviations for each feature dimension
• config - Diffusion configuration (steps, schedule type, seed)

pub fn with_correlations(self, corr_matrix: Vec<Vec<f64>>) -> Self

Set the correlation matrix for multi-dimensional generation.

The matrix should be symmetric positive-definite with ones on the diagonal. After denoising, Cholesky decomposition is used to impose this correlation structure on the generated samples.

Trait Implementations

impl Clone for StatisticalDiffusionBackend

fn clone(&self) -> StatisticalDiffusionBackend

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for StatisticalDiffusionBackend

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

Formats the value using the given formatter.

impl DiffusionBackend for StatisticalDiffusionBackend

fn forward(&self, x: &[Vec<f64>], t: usize) -> Vec<Vec<f64>>

Forward process: add noise at timestep t.

x_t = sqrt(alpha_bar_t) * x_0 + sqrt(1 - alpha_bar_t) * noise

fn reverse(&self, x_t: &[Vec<f64>], t: usize) -> Vec<Vec<f64>>

Reverse process: denoise at timestep t using Langevin-inspired updates.

x_{t-1} = x_t - step_size * (x_t - mu) / sigma^2 + noise_scale * noise

fn generate( &self, n_samples: usize, n_features: usize, seed: u64, ) -> Vec<Vec<f64>>

Generate n_samples with n_features by starting from pure noise and iteratively denoising using the reverse process.

At each reverse step t, the sample is updated via: x_{t-1} = (1 - blend) * x_t + blend * (mu + sigma * z) + noise where blend is derived from the schedule’s signal-to-noise progression, z is standard normal for stochastic variation, and noise decreases to zero at t=0.

fn name(&self) -> &str

Backend name.