robust-rs-core

Core abstractions for robust-rs: the RhoFunction loss trait (Huber, Tukey biweight, Cauchy, Welsch andrews, Hampel, L1, least squares), robust scale estimators (MAD, Huber Proposal 2, S-scale, Qn, Sn) and the influence-function / asymptotic-variance / Gaussian-efficiency / breakdown-point theory built on Gauss–Hermite quadrature.

Dependency-light (only libm, num-traits, thiserror) and wasm32-friendly. it carries no linear-algebra dependency, so it can be depended on for just the losses and theory.

Most users want the robust-rs crate instead, which layers the location, regression (M/S/MM/LTS/Theil–Sen) and multivariate (MCD/OGK/M-scatter/Tyler) estimators on top of this core.

use robust_rs_core::rho::{Huber, RhoFunction};
use robust_rs_core::theory::gaussian_efficiency;

let huber = Huber::default();          // k = 1.345
assert_eq!(huber.psi(10.0), 1.345);    // clipped score: bounded influence
assert!((gaussian_efficiency(&huber, 128) - 0.95).abs() < 0.01);

License

Dual-licensed under either MIT or Apache-2.0, at your option.