robustat-py 0.1.0

Robust statistics for Python: M/S/MM regression, robust scale and MCD/OGK/Tyler covariance; Rust-powered bindings to robust-rs.

robustat-py

Robust statistics for Python: Rust-powered bindings to robust-rs. M/S/MM regression, robust scale, robust multivariate location–scatter (MCD, OGK, Tyler) and the influence-function sampling theory, all with a numpy interface.

The numerics run in compiled Rust (via PyO3 and rust-numpy); randomized estimators (S/MM/LTS/MCD) are reproducible by default and expose a seed.

Install

pip install robustat-py

From source (this repo)

Requires a Rust toolchain and maturin:

pip install maturin
maturin develop -m py/Cargo.toml --features extension-module   # into the active venv
# or build a wheel:
maturin build  -m py/Cargo.toml --release

Quick start

import numpy as np
import robustat_py as rp

# --- Robust regression -------------------------------------------------------
# starsCYG: four giant stars flip the OLS slope negative; MM recovers the
# physical positive slope and rejects them (R's lmrob default).
x_raw, y = rp.datasets.stars_cyg()
X = np.column_stack([np.ones(len(y)), x_raw[:, 0]])   # prepend an intercept

mm = rp.MMEstimator(seed=1).fit(X, y)     # 50% breakdown AND ~95% efficiency
mm.coefficients                           # -> np.ndarray; slope > 0
mm.gaussian_efficiency()                  # ~0.95
mm.breakdown_point                        # 0.5

# Outliers only in y? A plain M-estimator is cheaper (but 0 breakdown vs leverage):
fit = rp.MEstimator(rp.Huber(), rp.Mad()).fit(X, y)
fit.weights                               # near 0 for down-weighted rows

# --- Robust covariance / multivariate outliers -------------------------------
Xs, _ = rp.datasets.stackloss()
mcd = rp.Mcd(seed=1).fit(Xs)              # FAST-MCD, affine equivariant
mcd.outliers(0.975)                       # bool per row, χ² cutoff
mcd.location, mcd.scatter

# --- Robust location & scale -------------------------------------------------
data = [2.1, 2.3, 1.9, 2.0, 2.2, 47.0]    # one gross outlier
rp.m_location(data).estimate              # ~2.16, not the mean (~9.58)
rp.Mad().scale(data)
rp.hodges_lehmann(data).estimate

API at a glance

Estimators (build, then .fit)

Class	.fit returns	Notes
MEstimator(loss, scale)	RegressionFit	IRLS; fast/convex; 0 breakdown vs leverage
SEstimator(loss, seed=…)	RegressionFit	FAST-S; 50% breakdown
MMEstimator(seed=…)	RegressionFit	start here; 50% breakdown + ~95% efficiency
Lts(seed=…)	LtsFit	FAST-LTS; coverage knob
Mcd(seed=…)	McdFit	start here for covariance; RMCD by default
Ogk(scale)	ScatterFit	deterministic, positive-definite
MScatter(loss)	ScatterFit	monotone M-scatter (low breakdown)
Tyler()	TylerFit	distribution-free shape (unit det)

Functions: m_location, trimmed_mean, winsorized_mean, hodges_lehmann, theil_sen, mahalanobis_distances, classical_covariance, distance_cutoff, outlier_flags, gaussian_efficiency, asymptotic_variance, breakdown_point.

Losses (redescending ones marked *): Huber(k=1.345), Tukey(c=4.685) *, Cauchy *, Welsch *, Andrews *, Hampel(a,b,c) *, LeastSquares, L1. Each exposes rho/psi/weight/psi_prime, tuning, is_redescending, rho_sup and gaussian_efficiency()/asymptotic_variance()/influence().

Scales: Mad(), Qn(), Sn(), HuberProposal2(). Every loss/scale argument also accepts a name string ("huber", "tukey", "mad", "qn", …).

Datasets: robustat_py.datasets.stackloss(), robustat_py.datasets.stars_cyg() (each returns (X, y) with predictors only; prepend your own intercept column).

Errors surface as robustat_py.RobustError (a ValueError subclass).

License

MIT OR Apache-2.0, matching robust-rs.