scikit-hep / iminuit / 34876504855

100%

push

github

fix: avoid NaN in binned chi2 gradients when mu is zero (#1204)

:robot: _AI text below_ :robot:

`_poisson_chi2_grad` and `_multinomial_chi2_grad` computed `n / mu`
without a guard. If the model prediction underflows to exactly zero in a
bin, this gives NaN or inf and the whole gradient becomes NaN with a
RuntimeWarning. For example, `ExtendedBinnedNLL` with an exponential CDF
at a small scale returns `[nan, nan]` from `grad`.

The value functions use `log_or_zero`, which is constant for `mu <= 0`,
so the consistent derivative of the log term in such bins is zero. Both
gradients now use a helper that evaluates `n / mu` only where `mu > 0`.
The previously commented-out gradient checks in
`test_ExtendedBinnedNLL_negative_weights` are restored, and a new test
covers the multinomial path.

Part of #1192

8 of 8 new or added lines in 1 file covered. (100.0%)

3967 of 3978 relevant lines covered (99.72%)

1.0 hits per line