covstream 0.1.2

Lean-backed fixed-dimension streaming covariance and Ledoit-Wolf shrinkage
Documentation
/// Controls how the shrinkage coefficient `alpha` is chosen.
#[derive(Debug, Clone, Copy, PartialEq)]
pub enum ShrinkageMode {
    /// Use the supplied coefficient as-is.
    FixedAlpha(f64),
    /// Clamp the supplied coefficient into the mathematically valid interval
    /// `[0, 1]`.
    ClippedAlpha(f64),
}

impl ShrinkageMode {
    /// Returns the effective shrinkage coefficient.
    pub fn alpha(self) -> f64 {
        match self {
            ShrinkageMode::FixedAlpha(alpha) => alpha,
            ShrinkageMode::ClippedAlpha(alpha) => alpha.clamp(0.0, 1.0),
        }
    }
}

/// Applies one shrinkage update toward the scaled-identity target `μI`.
#[inline]
pub fn shrink_entry(covariance_entry: f64, mu: f64, is_diagonal: bool, alpha: f64) -> f64 {
    let target_entry = if is_diagonal { mu } else { 0.0 };
    (1.0 - alpha) * covariance_entry + alpha * target_entry
}

/// Returns the arithmetic mean of the diagonal of a row-major square matrix.
pub fn diagonal_mean_row_major(matrix: &[f64], dimension: usize) -> f64 {
    let mut sum = 0.0;
    for i in 0..dimension {
        sum += matrix[i * dimension + i];
    }
    sum / dimension as f64
}

/// Builds the row-major matrix `μI`.
pub fn scaled_identity_row_major(mu: f64, dimension: usize) -> Vec<f64> {
    let mut out = vec![0.0; dimension * dimension];

    for i in 0..dimension {
        out[i * dimension + i] = mu;
    }

    out
}

/// Shrinks a row-major covariance matrix toward `μI` with coefficient `alpha`.
pub fn shrink_row_major(covariance: &[f64], dimension: usize, alpha: f64) -> Vec<f64> {
    let mu = diagonal_mean_row_major(covariance, dimension);
    let mut out = vec![0.0; covariance.len()];

    for row in 0..dimension {
        for col in 0..dimension {
            let idx = row * dimension + col;
            out[idx] = shrink_entry(covariance[idx], mu, row == col, alpha);
        }
    }

    out
}

/// Shrinks a row-major covariance matrix using a [`ShrinkageMode`].
pub fn shrink_with_mode_row_major(
    covariance: &[f64],
    dimension: usize,
    mode: ShrinkageMode,
) -> Vec<f64> {
    let alpha = mode.alpha();
    shrink_row_major(covariance, dimension, alpha)
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn diagonal_mean_row_major_returns_expected_value() {
        let matrix = vec![2.0, 1.0, 1.0, 4.0];

        let mean = diagonal_mean_row_major(&matrix, 2);

        assert_eq!(mean, 3.0);
    }

    #[test]
    fn scaled_identity_row_major_returns_expected_matrix() {
        let matrix = scaled_identity_row_major(3.0, 2);

        assert_eq!(matrix, vec![3.0, 0.0, 0.0, 3.0]);
    }

    #[test]
    fn shrink_row_major_with_zero_alpha_returns_original_matrix() {
        let covariance = vec![2.0, 1.0, 1.0, 4.0];

        let shrunk = shrink_row_major(&covariance, 2, 0.0);

        assert_eq!(shrunk, covariance);
    }

    #[test]
    fn shrink_row_major_with_one_alpha_returns_target_matrix() {
        let covariance = vec![2.0, 1.0, 1.0, 4.0];

        let shrunk = shrink_row_major(&covariance, 2, 1.0);

        assert_eq!(shrunk, vec![3.0, 0.0, 0.0, 3.0]);
    }

    #[test]
    fn shrink_row_major_with_half_alpha_blends_covariance_and_target() {
        let covariance = vec![2.0, 1.0, 1.0, 4.0];

        let shrunk = shrink_row_major(&covariance, 2, 0.5);

        assert_eq!(shrunk, vec![2.5, 0.5, 0.5, 3.5]);
    }

    #[test]
    fn shrink_with_mode_row_major_uses_fixed_alpha() {
        let covariance = vec![2.0, 1.0, 1.0, 4.0];

        let shrunk = shrink_with_mode_row_major(&covariance, 2, ShrinkageMode::FixedAlpha(0.5));

        assert_eq!(shrunk, vec![2.5, 0.5, 0.5, 3.5]);
    }

    #[test]
    fn shrink_with_mode_row_major_clips_large_alpha_to_one() {
        let covariance = vec![2.0, 1.0, 1.0, 4.0];

        let shrunk = shrink_with_mode_row_major(&covariance, 2, ShrinkageMode::ClippedAlpha(2.0));

        assert_eq!(shrunk, vec![3.0, 0.0, 0.0, 3.0]);
    }

    #[test]
    fn shrink_with_mode_row_major_clips_negative_alpha_to_zero() {
        let covariance = vec![2.0, 1.0, 1.0, 4.0];

        let shrunk = shrink_with_mode_row_major(&covariance, 2, ShrinkageMode::ClippedAlpha(-1.0));

        assert_eq!(shrunk, covariance);
    }
}