TST: linalg: fix complex sort in test_bad_geneig #20197
Conversation
I realized that I never finished the cross-BLAS-flavour analysis for 1.12. I've restarted that now, and will try to write a summary later. If you want to check yourself, you should see1 it in the Footnotes
|
|
Curious about the cross-flavor blas analysis. Is the intention to proceed in the 1.13rc phase, or is it realistically for 1.14 and beyond? |
|
@h-vetinari Any chance you'd be able to check that this PR fixes the failure seen on conda-forge? An alternative is of course to just land it and see what falls out in your next round of cross-blas analysis. |
Gave it a shot in conda-forge/scipy-feedstock@c0a0f48. CI will run for 3-4 hours before we have results. |
|
So it appear that this fixes the errors - I've started a second run to be sure. In any case thanks a lot for coming up with a fix! In this case I'd like to have this in 1.13, so I'm adding the backport-candidate label. |
|
Always happy to remove the old crust anyways. Thanks @ev-br, @h-vetinari ! |
Reference issue
closes #17125
What does this implement/fix?
The failure is relatively benign if annoying: the computation is most likely always correct, what fails is ordering of eigenvalues consistently on the complex plane, as Warren's analysis in #17125 (comment) shows.
So try a bit harder to trim numerical noise and account for complex conjugate pairs.
Have to admit I cannot repro locally. @h-vetinari you mentioned it's reproducible on conda-forge? #17125 (comment)
Additional information