Release scikit-dimension 0.4.0 · scikit-learn-contrib/scikit-dimension

New estimators and options

id.GRIDE (Denti et al. 2022), reimplemented from the paper with DADApy's conventions.
TwoNN(method="mle"): closed-form unbiased maximum-likelihood estimate.
id.PH: persistent-homology (PH0) dimension, ported from the kmyim/scikit-dimension fork (Ka Man Yim).
lPCA signal-rank versions:
- random matrix theory: MP (Marchenko-Pastur edge), GD (Gavish-Donoho threshold),
  KN (Kritchman-Nadler sequential Tracy-Widom test);
- parallel: Horn's parallel analysis with a permutation or sign-flip null;
- Minka: PPCA Laplace evidence, ported from the fork (Ka Man Yim, John Harvey);
- denoised: scran's denoisePCANumber rule (technical variance from noise_var).
DANCo(ver="DANCoFit"): calibration from a precomputed table (k=10, D <= 100,
20 <= n_samples <= 5000), 10-60 times faster than simulated calibration.
Datasets product, hyperTwinPeaksUniform, toroidal_spiral, torus, lorenz_attractor,
dumbbell, SurfaceOfRevolution, hyperSphere(radius, center), and skdim.embeddings
(delay_embedding, monomial_embedding), ported from the fork (John Harvey, Ka Man Yim).

Every pre-existing estimator and option was run on master (0.3.x) and on this version with the
same inputs; only the following differ.

lPCA(ver="broken_stick"): off-by-one fixed; estimates decrease by 1 (and can be 0 when no
component exceeds the broken-stick expectation).
FisherS: inseparability probabilities are divided by n - 1 (the other points), as in the
original MATLAB code, instead of n; estimates change by about 0.1-0.2% at n = 300, more for small n.
TLE: degenerate neighbourhoods now follow the original MATLAB code: NaN measurements from
duplicate points are dropped, distances below epsilon are excluded correctly (they previously
discarded all distances), and the estimate is 0 when (almost) all kNN distances are equal.
Results on data without duplicate or tied points are unchanged.

scikit-learn >= 1.6 (__sklearn_tags__, n_features_in_), Python 3.13+ and numpy 2.x.
Fit-time attributes n_neighbors and comb of local estimators renamed n_neighbors_ and comb_.
CorrInt and MLE no longer overwrite k1/k2/K during fit; KNN gains random_state
(it previously used numpy's global random generator).
lPCA(ver="Fan") no longer modifies a user-supplied explained-variance array.
transform(X) raises when X has a different number of features than the fitted data.
pandas is imported only by datasets.BenchmarkManifolds.
datasets.hyperSphere gains radius and center after random_state, so existing positional
calls are unchanged; all pre-existing dataset generators return identical samples for a given seed.
Requires Python >= 3.10.

Packaging metadata moves from setup.py/setup.cfg to pyproject.toml; the README is the PyPI description.
GitHub Actions replace Travis, AppVeyor and CircleCI: tests on Python 3.10-3.13 (Linux) and 3.13
(macOS, Windows), a documentation build and a package build check on every pull request; releases
are published to PyPI from GitHub releases with trusted publishing.
Documentation uses the pydata-sphinx-theme, builds with Sphinx >= 7, and gains an estimator overview
and this changelog; outdated pinned documentation dependencies are removed.
pixi.toml provides development, documentation and reference-implementation environments.

skdim/tests/test_reference.py checks agreement with reference implementations on identical
inputs: R intrinsicDimension (lPCA FO/Fan/maxgap, MLE and its options, ESS, MiND_ML, DANCo given
shared calibration data), R ider (CorrInt, MADA), PCAtools (MP, GD, parallel), scran (denoised),
DADApy (TwoNN, GRIDE), and the original MATLAB code run in Octave (TLE, MOM, FisherS).
Fixtures are regenerated with pixi run -e ref make-fixtures.
No runnable reference: KNN (random resampling), KN, PH, lPCA ratio/participation_ratio/Kaiser/broken_stick.

Per-file licence headers shortened to SPDX identifiers; third-party notices in LICENSES_bundled.txt;
original implementations and their authors listed in AUTHORS.md; CITATION.cff added.