FIX Restore support for n_samples == n_features in MinCovDet.

[anntzer]

Fixes #30625

#29835 broke support for the (degenerate) n_samples == n_features (and support_fraction unset) case in MinCovDet because this led to n_support = n_samples + 1, which was implicitly clipped to n_support = n_samples previously (at np.argsort(dist)[:n_support]) but not anymore now (and crashes np.argpartition(dist, n_support - 1).

To fix this, explicitly clip n_support.

Noticed at pyRiemann/pyRiemann#335.

Reference Issues/PRs

What does this implement/fix? Explain your changes.

Any other comments?

[github-actions]

✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: 4aa5665. Link to the linter CI: here}

[virchan]

So, if n_samples == n_feautres (and assume support_fraction=None), then

$$\begin{align*} n_{\mathrm{support}} &= \left\lceil \frac{n_{\mathrm{samples}} + n_{\mathrm{features}} + 1}{2} \right\rceil \\ &= \left\lceil \frac{2 n_{\mathrm{samples}} + 1}{2} \right\rceil \\ &= \left\lceil n_{\mathrm{samples}} + \frac{1}{2} \right\rceil \\ &= n_{\mathrm{samples}} + 1 \end{align*}$$

This causes an out-of-range index because we must always have n_support <= n_samples.

@anntzer's solution---to choose the smaller value between n_samples and the originally implemented n_support---makes sense to me.

[virchan]

I think the CI test is failing because tol_support = 50 in the launch_mcd_on_dataset function. We should always have tol_support <= n_samples here.

[virchan]

I think a changelog is needed here @anntzer. You can refer to the following link for more details:

https://github.com/scikit-learn/scikit-learn/blob/main/doc/whats_new/upcoming_changes/README.md

[anntzer]

done

[virchan]

LGTM! Thanks @anntzer!

@agramfort, @ogrisel, just a friendly ping—would you like to take a look?

[jeremiedbb]

LGTM. Thanks @anntzer