FAQ === Why not prioritize MVE? ----------------------- Minimum-volume ellipsoid estimators are historically important, but they are not the current differentiator for this package. The benchmark evidence is stronger for efficient FastMCD and regularized heavy-tail scatter estimators such as ``RegularizedCauchy`` and ``StudentTScatter``. When should I use FastMCD? -------------------------- Use ``FastMCD`` when the data are mostly clean, the outliers are separable, and ``n`` is comfortably larger than ``p``. It is the most interpretable choice for classical robust distances and support diagnostics. When should I use RegularizedCauchy? ------------------------------------ Use ``RegularizedCauchy`` for small-sample, high-dimensional, heavy-tailed covariance problems. It is currently the strongest method in the package's small-sample heavy-tail benchmark gallery. When should I use StudentTScatter? ---------------------------------- Use ``StudentTScatter`` when you want a smooth heavy-tail covariance-like estimator with a fixed degrees-of-freedom parameter. It is often competitive with Cauchy and can be easier to tune conceptually. When should I use AutoRobustScatter? ------------------------------------ Use ``AutoRobustScatter`` for exploratory workflows when you want the package to compare several robust scatter candidates. It is a diagnostic selector, not an oracle. For production, inspect the chosen estimator and the robust-distance diagnostics. Are OpenMP results deterministic? --------------------------------- For fixed random seeds, estimator choices are intended to be reproducible. Parallel floating-point reductions can still introduce tiny numerical differences because summation order changes. This is normal for parallel numerical code. Why are there both a Benchmark Gallery and a Use-case Gallery? ---------------------------------------------------------------- The benchmark gallery answers evidence questions: speed, ranking, scaling, and failure modes. The use-case gallery answers application questions: what to do for finance, fraud, sensors, embeddings, and ML preprocessing.