New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
ENH: update to arpack-ng 3.8.0 release #14786
base: main
Are you sure you want to change the base?
Conversation
Important changes are more standard-conforming code. We ignore their Fortran 90 files for the time being, although they might allow a simpler interface.
One tests for correctness eigenpairs, in a loop, which is meant to allow for a slack in Krylov subspaces method. One should 1st check whether one got correct eigenvalues, and try again if they are not the wanted ones. The current implementation does it backwards: it checks whether the eigenpair works, without 1st checking for eigenvalues. This commit corrects this programming error
The new error is in a different test (coming from the same, changed by the last commit, loop):
this does look like a numerical noise - I don't understand why it was not seen earlier (in fact, the same can be seen on the master branch with only the last commit applied, so that was somehow hidden. |
Thanks @dimpase. It may be useful to submit the test bugfix commit as a separate PR, so we can see its effect on master and merge it quickly (with a tolerance bump if needed). I checked the SciPy-specific patch instructions in We're also pulling in ~5 years worth of bug fixes, the last upgrade was in 2016. Hopefully that addresses some of the flaky test failures that we occasionally see. Let me Cc @pv on this one to see if he has any concerns or things that need doing/checking. |
This should be rebased over #14798 once it's merged. |
Also, there's a big merge conflict since |
@rgommers I no longer have |
Hey @dimpase, I think we both lost track of this. Yes, I think this is still of interest, lots of useful changes get pulled in by upgrading to a more recent Arpack-ng (and 3.9.0 appeared 6 months ago). You don't need |
Important changes are more standard-conforming code.
We ignore their Fortran 90 files for the time being, although they
might allow a simpler interface.
helps with gh-13222
More conformance to Fortran standards
I took the *.f files in https://github.com/opencollab/arpack-ng/tree/3.8.0/SRC and
copied them over the corresponding files in ARPACK module of scipy.
Additional information
one test fails, but it might be merely numerical noise (different scaling of eigenvectors):
Edit: the above failure was due to a bug the scipy test code, which I corrected in the 2nd commit.