fix eigh to accept complex-valued DNDarrays

[GioPede]

Description

Small relaxation of the checks in linalg.eigh to allow complex-valued matrices

Issue/s resolved: #
#2327

Changes proposed:

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)

Does this change modify the behaviour of other functions? If so, which?

no

[brownbaerchen]

@GioPede, thanks for getting this going! A function that implies Hermitian input in its name but doesn't allow it is indeed a bug! :D I started a PR with some tests on your fork. Since you probably don't watch your own fork, I write a note about this here.

Also, just to make sure we have this discussion also here: This fix only works for non-distributed input, since the _eigh function for distributed input requires percentile, which requires sort, which doesn't allow complex input.

[ClaudiaComito]

@GioPede, thanks for getting this going! A function that implies Hermitian input in its name but doesn't allow it is indeed a bug! :D I started a PR with some tests on your fork. Since you probably don't watch your own fork, I write a note about this here.

Also, just to make sure we have this discussion also here: This fix only works for non-distributed input, since the _eigh function for distributed input requires percentile, which requires sort, which doesn't allow complex input.

Addendum after discussing this in our meeting on Jun 8th.

ht.sort does not support complex dtypes because torch.sort doesn't. We could of course provide at least the CPU functionality via numpy.