Evgeni Burovski

> removing old fortran bits We have some of this in the detailed roadmap: https://docs.scipy.org/doc/scipy-1.11.4/dev/roadmap-detailed.html#use-of-venerable-fortran-libraries EDIT: tangentially related, detailed roadmap might benefit from an update, too, for things which are...

Confirmed locally: `$ git revert 4751485b29` (which is the commit from gh-17918) fixes the MWE from https://github.com/akug/mwe_scipy_sqrtm_bug. The "error" goes down to ~1e-28 and the difference `out@out - data` become...

WDYT about committing the revert? IOW, what's worse: zero imaginary parts or plain wrong answers in at least some cases? That said, if you see a quick simple fix to...

A great loss indeed :-(.

Now that I actually looked at the code, indeed I doubt it is going to work, at all. That is, without exposing a tolerance to a user. So yeah, rewriting...

OK, https://github.com/scipy/scipy/pull/20212 suggests a minimal fix. It actually does look at subdiagonals only: what happens I believe is that while the lower triangular part is all ~1e-16 or less, some...

Reopening: the https://github.com/scipy/scipy/pull/20218 xfailed the test which is broken by gh-20212 on MacOS. Somehow the test passes on our regular MacOS CI , but fails on the wheel builds, https://github.com/scipy/scipy/actions/runs/8201161926/job/22429344365...

x-ref https://github.com/scipy/scipy/pull/20218#issuecomment-1985671107: with enough luck it's old(-er) OpenBLAS on MacOS specific.

https://github.com/scipy/scipy/pull/19816#issuecomment-1986693242 thank you @thalassemia for the details! Let's move the sqrtm discussion here. So what happens on main in the now-xfailed test, https://github.com/scipy/scipy/blob/v1.12.0/scipy/linalg/tests/test_matfuncs.py#L413, after gh-20212: on MacOS: ``` (Pdb) d0...

FWIW, I fail to see how a lapack routine helps here. Eigenvalues we get easily already. A tricky bit (as in, an arbitrary cutoff) is to decide it's real or...