FEA Make LedoitWolf estimator array API compatible

[betatim]

Reference Issues/PRs

Came up as part of pondering #33564

Towards #33584

What does this implement/fix? Explain your changes.

This makes LedoitWolf array API compatible. After this it might be easier to convert other covariance estimators as it also updates some of the basic infra shared by them

Todo

• it might be worth doing ledoit_wolf (the function) in this PR as well

AI usage disclosure

I used AI assistance for:

• Code generation (e.g., when writing an implementation or fixing a bug)
• Test/benchmark generation
• Documentation (including examples)
• Research and understanding

Any other comments?

[bruAristimunha]

Hey @betatim,

Sorry for dropping into the PR, but what you are working on is something super nice! At Pyriemann, with @agramfort and @qbarthelemy, we are using many of the covariance matrix implementations from scikit-learn and are also moving towards array API compatibility.

I just would like to offer some assistance with the covariance part api compatibility, if necessary.

[bruAristimunha]

I tried to find a macro issue to offer help, but I couldn't find a place that seemed appropriate.

Please feel free to delete my comment, given that it's off-topic for your PR.

[betatim]

I tried to find a macro issue to offer help, but I couldn't find a place that seemed appropriate.

Please feel free to delete my comment, given that it's off-topic for your PR.

No worries and welcome! It is fantastic to hear from someone using scikit-learn that says "this looks useful". I don't think there is a macro/mega issue related to this area of scikit-learn + array API support. I picked it more or less at random/as an exercise to see how much infrastructure would be missing.

I think if you are interested in contributing to converting estimators/tools from https://scikit-learn.org/stable/api/sklearn.covariance.html go for it! It could be that the first step towards that is helping review this PR or picking another estimator from the list and converting it. I'd have to think about it, it probably depends on how much interdependency there is.

Another thing that could be worth doing is creating an example for the gallery that shows off the array API support in LedoitWolf based on a real-world use-case.

I will create a mega-issue for sklearn.covariance so we can keep track of what is done, what needs doing, etc.

[bruAristimunha]

if you do this, should unlock too the empiral covariance

[betatim]

I undid the changes to EmpiricalCovariance again. By marking this class as supporting array API classes that inherit from it also get marked with support. So we need to tackle this at the end when everyone else has array API support.

[bruAristimunha]

I can create a separate PR for this, if you want, @betatim. I fixed in my draft PR, small tricks are necessary for empirical covariance.

[betatim]

I've addressed all the comments, let me know what y'all think. Maybe we could get this wrapped up in time for the release?

[ogrisel]

LGTM besides the following:

[ogrisel]

I ran the tests of this PR with DPNP on an Intel GPU with float32-only support and there are two failures barely above the rtol threshold for float64 inputs:

https://github.com/probabl-ai/scikit-learn-intel-workflow/actions/runs/26148891699/job/76911036653

I think we should bump it up a bit (e.g. 1e-6 instead of 1e-7).

[ogrisel]

@cakedev0 Actually there is something I don't understand in the above. Does your float32-only GPU:

How is the case the following does not fail if the GPU is float32 only xp.asarray(np.ones(3, dtype=np.float64, device="gpu").

Also: I would have expected _array_api_for_tests to have skipped that particular test case (dpnp-gpu-float64 on a float32-only GPU host) entirely.

[cakedev0]

Not sure to understand exactly your message but, on my laptop (float32-only GPU):

>>> dpnp.asarray(np.ones(3, dtype=np.float64), device="gpu")
array([1., 1., 1.], dtype=float32)

And this test should definitely have been skipped, let's not bump the rtol but make sure it's skipped instead (I'm looking into that)

[cakedev0]

Let's update the usage of _array_api_for_tests to skip tests for which the device doesn't support the dtype.

[ogrisel]

New Intel GPU workflow: https://github.com/probabl-ai/scikit-learn-intel-workflow/actions/runs/26577559267

EDIT: It's green!

Let me fix the linter.

[ogrisel]

@betatim this PR should be ready to merge once the following is addressed: #33573 (review)

[betatim]

I fixed the use of _convert_to_numpy. GitHub says that @cakedev0 requested changes, but I can't find what/where? Has that been resolved already or is there a change you'd like to see?

[cakedev0]

This should discard the "request changes" review. (I'll avoid that next time)

[betatim]

Yeah, my impression is that the "request changes" feature isn't super useful for a project like scikit-learn. It ends up making it more complicated to merge things and I don't think I've ever experienced it where someone's feedback was ignored/not worked on because they "only" left it as a comment/normal review feedback. The social fabric is strong here :D

[betatim]

@OmarManzoor and @ogrisel you both left 👍 , do you want to take a last look and then press merge? Or should we attract some more reviews?

[ogrisel]

Merged! Thanks all!