Fix for #2374

[brownbaerchen]

See #2374 for the description of the problem reported by @maliesen.

Short recap: The function __moment_w_axis takes unbiased as input and forwards it to the respective torch function (var in this case). However, it only does so if unbiased is True. We need to forward unbiased=False as well, though.

The same happens with Fisher (see #2375) which may also be fixed by this PR. It remains to be checked, though.

Issue/s resolved: #2374

[Copilot]

Pull request overview

This PR fixes a bug in axis-based moment computations where falsy boolean options (notably unbiased=False) were not forwarded to the underlying torch implementation, causing ht.var(..., axis=..., ddof=0) to behave as if Bessel correction was enabled. It also adjusts the same forwarding logic for Fischer in kurtosis/skew paths and adds a regression test for the reported variance edge case.

Changes:

• Forward unbiased and Fischer through __moment_w_axis whenever they are provided (including False).
• Add a regression test for ht.var(..., axis=0, ddof=0) on a single-element input (issue #2374).
• Fix a minor docstring typo (“implicitely” → “implicitly”).

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
heat/core/statistics.py	Fixes boolean kwarg forwarding for axis-based moment helpers (unbiased, Fischer) and corrects a docstring typo.
tests/core/test_statistics.py	Adds a regression test covering the ddof=0 + axis edge case from #2374.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 91.74%. Comparing base (5d42619) to head (18639de).
⚠️ Report is 14 commits behind head on main.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2379   +/-   ##
=======================================
  Coverage   91.74%   91.74%           
=======================================
  Files          87       87           
  Lines       14120    14120           
=======================================
  Hits        12955    12955           
  Misses       1165     1165           

Flag	Coverage Δ
unit	91.74% <100.00%> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

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

[mtar]

unbiased is correction in torch.var since version 2.0 and works with integers. We will have to look what it is for the other functions.

[brownbaerchen]

It's also called correction in numpy. See the docs.
Scipy kurtosis and skew have bias, see docs.

I switched to correction internally, but left the interface of ht.skew and ht.kurtosis to keep unbiased. What do you think?

[mtar]

I think we should revert correction back for now. It is only in var and accepts ints. Maybe, change it to bias and flip the values.
Nonetheless, we should use the same parameters as in the other APIs for our public functions and update the calls to torch.var

[brownbaerchen]

correction is the closest to the torch and numpy APIs, so why would be do something yet completely different?

[mtar]

Actually, I don't see it close to numpys API.

Torch/NumPy:

• correction is only used in var
• correction takes ints and floats and is a array api compatible name for ddof.
• skew and kurtosis use bias in scipy and are bools

PR:

• correction is used everywhere
• correction takes bools in var and it has a different logic underneath
• skew and kurtosis also have the correction parameter name

Right now in the PR, it is nothing alike but the name in a single place.

[mtar]

We will keep it for now.

[maliesen]

I don't know if this is the right place for it, but upon merging the fix we can now omit the four atol=1e-1 argument values in the assertions in tests/naive_bayes/test_gaussiannb.py. This is to reflect the change in the numerical values of ht.naive_bayes.GaussianNB:

heat/tests/naive_bayes/test_gaussiannb.py

Lines 79 to 80 in 5d9851b

	self.assertTrue(ht.isclose(gnb_heat.sigma_, sklearn_sigma, atol=1e-1).all())
	self.assertTrue(ht.isclose(y_pred_proba_sklearn, y_pred_proba_local, atol=1e-1).all())

heat/tests/naive_bayes/test_gaussiannb.py

Line 91 in 5d9851b

self.assertTrue(ht.isclose(gnb_heat.sigma_, sklearn_sigma, atol=1e-1).all())

heat/tests/naive_bayes/test_gaussiannb.py

Line 105 in 5d9851b

# self.assertTrue(ht.isclose(gnb_heat.sigma_, sklearn_sigma, atol=1e-1).all())

[brownbaerchen]

@maliesen, thanks for pointing this out! You are solving all of our bugs! :)