Thanks for the PR @awinml. I directly pushed some changes because this is now the last function to deal with, so we can clean-up the filtering in the test at the same time. I also simplified a bit some parts, especially the example: we usually keep the example simple and don't include special cases.

That's good @jeremiedbb.
Do you think this function should support broadcasting like numpy.linalg.slogdet? Should I open an issue for review?
I have highlighted the limitation in the PR. #24605 (comment)

A limitation of this function as compared to numpy.linalg.slogdet is that it cannot broadcast for multiple matrices in a single function call. When we pass a stack of matrices as input it throws a ValueError. Ex:

>>> a = np.array([ [[1, 2], [3, 4]], [[1, 2], [2, 1]], [[1, 3], [3, 1]] ])
>>> fast_logdet(a)
ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()

Do you think this function should support broadcasting like numpy.linalg.slogdet? Should I open an issue for review?

We don't have the need for that in scikit-learn so far so I don't see the benefit.