BUG: Correct intermediate overflow in KS one asymptotic in SciPy.stats

[aherbert]

Overflow can occur for 6*n when integer n is large.

Change the use of exp to expm1 when appropriate.

See #17775

Reference issue

Closes gh-17775.

What does this implement/fix?

Corrects integer overflow in the c implementation when n is large by changing 6*n to 6.0*n.

Use exp or expm1 to compute the p-value choosing whichever will be smallest. The opposite p-value is computed by subtraction from 1.

[tupui]

Thank you @aherbert, this looks great. I only have some suggestions.

Also linter is failing. You can check locally doing python dev.py lint.

[aherbert]

assert_allclose uses a tolerance relative to the expected. This removed the requirement to use the ulp assertion. I switched to a parameterized test which makes it easier to add more test cases as required. Currently the 3 provided cases all failed with the old C code.

[tupui]

LGTM, let's see if all CI is happy now. Thanks for the quick updates.

[tupui]

CI failure is not related (and now fixed on main). Merging. Thanks again @aherbert, it was a great issue and PR 👍

[tupui]

Oh just for next time, try to avoid force pushing without telling first the reviewer. GitHub is not smart enough and links from email notifications are breaking after a force push. But even if, it's simpler for reviewer to know what commit they already reviewed and follow along.

[aherbert]

OK. Next time I will leave the commit history and you can squash them as appropriate.

[tupui]

Yes this is what we do 😃

[tylerjereddy]

Bug fix shortly after 1.10.0 release so tentatively labeling for backport to 1.10.1