Top pick: Marchenko-Pastur + RIE (pyRMT (BSD-style)). Fills the gap: Matrix-level orthogonality — which the production baseline (Pearson + Spearman + h_norm + Tier-2 MI-vs-target) cannot detect.
Lane = how usable the metric is today. Top lane = production-baseline duplicates (Pearson, Spearman). Below = permissive FOSS / GPL / unmaintained / code-released-but-not-packaged / no-FOSS. Per-circle text = first 3 letters of language (Pyt / R / Jav).
| Metric | Library | License | Lang | Maint | Complexity | Param-free? | URL |
|---|---|---|---|---|---|---|---|
| Bai-Ng / Onatski factor selection | dfms | GPL | R | ✅ | O(N^3) | yes | ↗ |
| ICA (FastICA/JADE/Infomax) — R | ica | GPL | R | ✅ | varies | ? | ↗ |
| Ledoit-Wolf quadratic / analytical NLS | nlshrink | GPL | R | ✅ | O(N^3) | no | ↗ |
| Robust PCA (Candes 2011) | rsvd / pyrpca | GPL / MIT | R / Python | ✅ | O(NT min(N,T)) | yes | ↗ |
| Tracy-Widom edge test | RMTstat | GPL | R | ✅ | O(N^3) | no | ↗ |
| FastICA | sklearn.decomposition.FastICA | Apache-2 | Python | ✅ | O(NTK iter) | yes | ↗ |
| Graphical Lasso | sklearn.covariance.GraphicalLasso | Apache-2 | Python | ✅ | O(N^3) per iter | yes | ↗ |
| ICA portfolio optimization | FastICA (sklearn) + custom opt | Apache-2 | Python | ✅ | ? | ↗ | |
| Ledoit-Wolf linear shrinkage | sklearn.covariance.LedoitWolf | Apache-2 | Python | ✅ | O(N^2) | no | ↗ |
| Marchenko-Pastur eigenvalue clipping | pyRMT | BSD-style | Python | ✅ | O(N^3) | no | ↗ |
| NMF | sklearn.decomposition.NMF | Apache-2 | Python | ✅ | O(NTK iter) | yes | ↗ |
| PCA | sklearn.decomposition.PCA | Apache-2 | Python | ✅ | O(min(N^2 T, N T^2)) | yes | ↗ |
| Rotationally Invariant Estimator (RIE) | pyRMT | BSD-style | Python | ✅ | O(N^3) | no | ↗ |
| Sparse PCA | sklearn.decomposition.SparsePCA | Apache-2 | Python | ✅ | O(N^2 T) L1 | yes | ↗ |
| VIF | statsmodels.stats.outliers_influence.variance_inflation_factor | BSD-3 | Python | ✅ | O(N^2 T) per col | yes | ↗ |