Fix compatibility with numpy 2.5 by IAlibay · Pull Request #5404 · MDAnalysis/mdanalysis

IAlibay · 2026-06-22T10:32:11Z

Fixes #5403

Changes made in this Pull Request:

With numpy 2.5, compatibility in casting between inputs and outputs is more strongly verified.
This PR fixes an issue where np.degrees could be fed a complex number from principal_axes due to floating point noise.

LLM / AI generated code disclosure

LLMs or other AI-powered tools (beyond simple IDE use cases) were used in this contribution: no

PR Checklist

Issue raised/referenced?
Tests updated/added?
Documentation updated/added?
package/CHANGELOG file updated?
Is your name in package/AUTHORS? (If it is not, add it!)
LLM/AI disclosure was updated.

Developers Certificate of Origin

I certify that I can submit this code contribution as described in the Developer Certificate of Origin, under the MDAnalysis LICENSE.

IAlibay · 2026-06-22T10:36:13Z

        """
        atomgroup = group.atoms
-        e_val, e_vec = np.linalg.eig(atomgroup.moment_of_inertia(wrap=wrap))
+        e_val, e_vec = np.linalg.eigh(atomgroup.moment_of_inertia(wrap=wrap))


I think this should fix it?

Someone with better maths knowledge should double check this - my understanding is that the moment of intertia tensor is always a real symmetric matrix, so it should satisfy the requirements of np.linalg.eigh.

My understanding is that eigh doesn't have the stability issues of eig, so I tried swapping it out and it seems to be working fine.

Indeed, eigh is preferred for symmetric/Hermitian matrices.

IAlibay · 2026-06-22T10:37:19Z

            [
-                [0.78787867, 0.26771575, -0.55459488],
-                [-0.40611024, -0.45112859, -0.7947059],
+                [-0.78787867, -0.26771575, 0.55459488],


I believe these are fine since the eigenvector sign is arbitrary?

Yes, the sign is algorithm-dependent and therefore can differ between eig and eigh.

Looks like some tests are still failing because of sign issues. Maybe to make them more robust we could check for all the eigenvector tests that they are numerically equivalent up to a sign change?

Yeah I'm thinking we should do an abs check, will update the tests accordingly.

Done - just to sanity check me, can I get a re-review please?

I thought about it again. abs is not the same as sign-check. Will comment on review.

RMeli · 2026-06-22T12:20:30Z

        """
        atomgroup = group.atoms
-        e_val, e_vec = np.linalg.eig(atomgroup.moment_of_inertia(wrap=wrap))
+        e_val, e_vec = np.linalg.eigh(atomgroup.moment_of_inertia(wrap=wrap))


Indeed, eigh is preferred for symmetric/Hermitian matrices.

RMeli · 2026-06-22T12:20:52Z

            [
-                [0.78787867, 0.26771575, -0.55459488],
-                [-0.40611024, -0.45112859, -0.7947059],
+                [-0.78787867, -0.26771575, 0.55459488],


Yes, the sign is algorithm-dependent and therefore can differ between eig and eigh.

codecov · 2026-06-22T18:37:19Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 93.85%. Comparing base (d25f074) to head (2f46b14).
⚠️ Report is 3 commits behind head on develop.

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #5404      +/-   ##
===========================================
+ Coverage    93.63%   93.85%   +0.21%     
===========================================
  Files          182      182              
  Lines        22509    22509              
  Branches      3202     3202              
===========================================
+ Hits         21076    21125      +49     
+ Misses         965      922      -43     
+ Partials       468      462       -6

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

orbeckst

Although testing for abs() equality works, it's not equivalent to testing that the two vectors are equal up to a sign change. The abs check can, in principle, lead to false positives.

I suggest to do this properly and use a helper function such as

def assert_vectors_no_sign_close(a, b):
   if not (np.allclose(a, b) or np.allclose(a, -b)):
      # raise assertion with the sign choice that looks closest
      x = b if np.linalg.norm(a - b) < np.linalg.norm(a + b) else -b
      assert_allclose(a, x)

allclose() may need proper rtol and atol if the defaults aren't appropriate.

orbeckst · 2026-06-22T18:35:12Z

+ * The `principal_axes` method in :class:`Masses` now uses
+   `np.linalg.eigh` instead of `np.linalg.eig`, improving numerical
+   stability. This may lead to slightly different results from previous
+   version of MDAnalysis.


Should go under changes, not really a fix. Reference PR #5404

Interesting, I see this as a a fix in the sense of "before we were using something that could yield a complex number because it was prone to floating point precision issues, now it's fixed" - what do you think?

My view was that this was a robustness change that was surfaced by a newer version of numpy. However, your interpretation also makes sense so I am happy to leave it as a fix. I'd still add the issue/PR number, though.

orbeckst · 2026-06-22T18:36:58Z

            assert_almost_equal(result[0], ref[method_name][0], self.prec)
            assert_almost_equal(result[1], ref[method_name][1], self.prec)
+        elif method_name == "principal_axes":
+            assert_almost_equal(


Adding comment regarding sign-check would be useful + ref GH-5404

comment along the lines of

The direction (sign) of the principal axes is dependent on the specific algorithm
used in numpy.linear.eigh() but the direction itself is not physically relevant
so we have to test both directions.

orbeckst · 2026-06-22T18:41:14Z

            [
-                [0.78787867, 0.26771575, -0.55459488],
-                [-0.40611024, -0.45112859, -0.7947059],
+                [-0.78787867, -0.26771575, 0.55459488],


I thought about it again. abs is not the same as sign-check. Will comment on review.

IAlibay · 2026-06-22T22:09:24Z

Thanks @orbeckst - here's a second attempt with something a little bit more fancy (einsum -> sign then use that to flip antiparallel vectors before the assert).

orbeckst

lgtm, thanks for the fix

orbeckst · 2026-06-22T23:25:35Z

+            # specific algorithm used, but the direction itself is not physically
+            # relevant, so we get the signs to flip any anti-parallel vectors before
+            # comparing the two results arrays
+            signs = np.sign(np.einsum("ij,ij->i", result, ref[method_name]))


fancy, I convinced myself that it does the right thing for the output of principal_axes().

RMeli

Agree with @orbeckst, abs() is not what we wanted here. The ensum implementation is a nice touch (and well commented for posterity). Thanks @IAlibay!

Fix compatibility with numpy 2.5

0df2fd7

IAlibay commented Jun 22, 2026

View reviewed changes

IAlibay requested review from orbeckst and tylerjereddy June 22, 2026 10:37

RMeli approved these changes Jun 22, 2026

View reviewed changes

IAlibay added 2 commits June 22, 2026 19:12

Merge branch 'develop' into fix-ci

a0dce8c

fix CI

3bcf762

IAlibay requested a review from RMeli June 22, 2026 18:22

format two files

fd6fa86

orbeckst requested changes Jun 22, 2026

View reviewed changes

orbeckst mentioned this pull request Jun 22, 2026

ENH: Add PDB reader benchmark to track parsing performance #5394

Merged

6 tasks

IAlibay added 2 commits June 22, 2026 23:05

Try something a little bit more clever to compare the arrays

7cecd60

add more comments

0aebde9

IAlibay requested a review from orbeckst June 22, 2026 22:08

orbeckst approved these changes Jun 22, 2026

View reviewed changes

Update package/CHANGELOG

2f46b14

orbeckst mentioned this pull request Jun 22, 2026

MDAnalysis testsuite fails with numpy 2.5 #5403

Closed

RMeli approved these changes Jun 23, 2026

View reviewed changes

RMeli merged commit f765ecd into develop Jun 23, 2026
23 of 25 checks passed

RMeli deleted the fix-ci branch June 23, 2026 07:25

Uh oh!

Conversation

IAlibay commented Jun 22, 2026

LLM / AI generated code disclosure

PR Checklist

Developers Certificate of Origin

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

codecov Bot commented Jun 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

orbeckst left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

IAlibay commented Jun 22, 2026

Uh oh!

orbeckst left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

RMeli left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

codecov Bot commented Jun 22, 2026 •

edited

Loading

orbeckst left a comment •

edited

Loading