BUG: stats: distribution methods emit unnecessary warnings from Boost

[tupui]

Describe your issue

Several statistical distribution methods emit warnings stemming from Boost. This issue tracks the status of all of them.

• beta.ppf as reported here; should be fixed by short-circuit overflow protection for M1 Mac clang boostorg/math#827
• ncf as reported in BUG: stats: spurious warning from ncf.cdf #17101, should be fixed by Fix for issue 845 boostorg/math#846
• nct - I'm not sure that this has been reported separately yet. Is it caused by the same sort of thing?
• ncx2 - Ditto.

gh-17272 will silence the failures in SciPy's tests temporarily, but as many of these marks as possible should be removed when gh-17207 merges.

---

Original Post

Describe your issue.

macOS CI is failing due to an overflow. I am still trying to reproduce locally.

NumPy was updated to 1.21.3, could be this or another dependency.

https://github.com/scipy/scipy/runs/3960228039

Reproducing Code Example

# from scipy/stats/tests/test_distributions.py:2878: in test_issue_12796
import numpy as np
from scipy import stats
q = 0.999995
a = np.array([ 2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20])
b = np.array([99999, 99998, 99997, 99996, 99995, 99994, 99993, 99992, 99991, 99990, 99989, 99988, 99987, 99986, 99985, 99984, 99983, 99982, 99981])
stats.beta.ppf(q, a, b)

Error message

FAILED scipy/stats/tests/test_distributions.py::TestBeta::test_issue_12635 - ...
927
FAILED scipy/stats/tests/test_distributions.py::TestBeta::test_issue_12794 - ...
928
FAILED scipy/stats/tests/test_distributions.py::TestBeta::test_issue_12796 - ...
...
scipy/stats/_continuous_distns.py:626: in _ppf
916
    return _boost._beta_ppf(q, a, b)
917
E   RuntimeWarning: overflow encountered in _beta_ppf

SciPy/NumPy/Python version information

master on macOS/CI

[tupui]

Not sure anymore it's NumPy since I cannot reproduce locally. Will have a look at the other deps. I am pretty sure it's not on our side since the issue appeared after 14341ea. In the PR CI passed while after the merge it did not anymore.

[tupui]

suite-sparse dependency changed. Now it's Installing dependencies for suite-sparse: metis and openblas and before Installing dependencies for suite-sparse: metis, openblas and tbb. Homebrew/homebrew-core@083d997#diff-2ac98c43f5e9e7b776c4d0883ac3d65d61db22b46d98918ce2ce26433f2ddfbb

[andyfaff]

I'm not sure it's to do with the suite-sparse dependencies. Another thing is that the default macOS version (runs-on: macos-latest) was updated from 10.15 to 11. I think this happened when macOS tests started failing.

The most obvious thing is that a previously valid calculation is now not happening correctly. However, I'm not sure that a RuntimeWarning should be raised, these can be easy to ignore, should it be a RuntimeError instead?

[lordgrenville]

I can reproduce this on my machine, with this code (from the docs):

a, b = 2.31, 0.627
x = np.linspace(beta.ppf(0.01, a, b), beta.ppf(0.99, a, b), 100)

Happy to share details about my system info. (Running the latest macOS on M1.)

[andyfaff]

I think CI is on Macos Intel.

[tupui]

It's on intel and apparently not failing on Travis for the wheels.

Might be the macOS version as you said.

[rgommers]

This is getting annoying, I will try to silence these warnings before the PyData Global sprint tomorrow.

[rgommers]

• I cannot reproduce this on an arm64 machine with macOS 11.6 and XCode 12.5.1.
• GitHub Actions uses macOS 11.6. Don't see an XCode version, but I assume it's up-to-date.
• Azure has macOS 10.15.7, that's not failing.

The relevant tests and code seems to not have been touched after being introduced in the big Boost PR (gh-13328).

[rgommers]

@mckib2 if you have any memory of that code and some idea of what this could be, that would be great to hear (not very high-prio, we can live with just silencing the failures for a while).

[mckib2]

@mckib2 if you have any memory of that code and some idea of what this could be, that would be great to hear (not very high-prio, we can live with just silencing the failures for a while).

That warning was introduced by gh-14618, we've seen some other instances of strange behavior when building on different platforms, e.g. gh-14833. The only levers I know to pull in cases like this are the Boost.Math policy defines in the setup.py. Hard to experiment without local hardware though

[rgommers]

Thanks for the details @mckib2!

That warning was introduced by gh-14901

That's this issue, want to edit your comment to point at the PR you meant?

[mckib2]

Thanks for the details @mckib2!

That warning was introduced by gh-14901

That's this issue, want to edit your comment to point at the PR you meant?

Good catch -- edited

[rgommers]

So this happened on Intel macOS in CI; I'm seeing them locally with a conda build on arm64 macOS, and gh-14722 sees them on arm64 macOS too (miniforge-from-Homebrew install). Looking at a fix or some skips for 1.7.3.

[mckib2]

Looking this over again and am a little confused as to why a RuntimeWarning is being produced by an overflow. gh-14833 set the Boost.Math policy to throw an OverflowError, so it's currently unclear to me why it's being treated as a runtime exception. I'm just getting back from a trip and getting settled in, but I will find time this weekend to review this more closely. beta seems to be the problem child in the Boost distributions...

[mdhaber]

@mckib2 some of these have been fixed in Boost, do you know if the fixes have made their way into SciPy yet? It would be good to close this before 1.10 branches, if possible.

[mborland]

nct should be fixed by boostorg/math#892. I'll look at ncx2 this week.

[mborland]

@mdhaber please make sure I am reading this right. I used this CI run's values where df = 21, nc = 1.0560466, and x = 6.77671634 or 21.36270226 or 49.09451392 to generate overflow. I was not able to re-produce overflow (only FE_INEXACT) with the following platform:

Apple clang version 14.0.0 (clang-1400.0.29.202)
Target: arm64-apple-darwin22.1.0
Thread model: posix
InstalledDir: /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin

Are there other test values that are know to cause overflow with ncx2? The other distribution's test cases have all be integers that caused denominators to be 0.

[mdhaber]

Thanks for your patience @mborland. I don't know the details because I can't reproduce locally, but I think they will become more clear once we get all these fixes you've made into SciPy. We'll let you know if it's not already resolved.

---

Just another note relevant to this issue: we're now seeing these warnings in CI (e.g. gh-17853) again after gh-13699 was merged:

FAILED scipy/stats/tests/test_continuous_fit_censored.py::test_nct - RuntimeWarning: overflow encountered in _nct_sf
FAILED scipy/stats/tests/test_continuous_fit_censored.py::test_ncx2 - RuntimeWarning: overflow encountered in _ncx2_sf

We may xfail or skip these tests to silence the CI failure. Part of closing this issue will be removing the xfail/skip mark.

[mborland]

Do you have the test cases that cause those two to overflow? If they are like the others it is floating-point environment oddities.

[mdhaber]

I think you already looked into these (#14901 (comment)), so I don't think there's anything new to be done yet.
It difficult for me to give you a test case associated with those because I can't reproduce locally, but based on the log, it looks like sf of ncx2 raise the warning with df = 1.0, nc = 1.0, loc=0, scale=1, and x as one of the following:

[2.7, 0.2, 6.5, 0.4, 0.1, 8., 0.6, 1.]

For nct.sf, that's df=0.4125, nc=1.571875, loc=0, scale=1. and x as one of the following:

[1., 2., 3., 5., 8., 10., 25.,]

[mborland]

I think for those I would disable failure. They were setting the flag for FE_DIVBYZERO, but now the only one I can reproduce is FE_INEXACT which is defined as: "inexact result: rounding was necessary to store the result of an earlier floating-point operation". I don't think I would worry about that.

[h-vetinari]

Let's hope that this got ameliorated by #17865, but unless that gets backported, this issue is not for 1.10 anymore - I adapted the milestone.

[mdhaber]

No, gh-17865 just treated one of the symptoms, silencing warnings in two tests. The np.errstate context I added should be removed when this issue is resolved.

[mdhaber]

@mckib2 was there anything else you wanted to investigate here (e.g. I'm not sure if you saw #17432 (comment)), or do we just wait for upstream changes?

@tupui Do you think #17432 (comment) and the np.errstates we added in gh-17432 close this issue?

[tupui]

@mdhaber looks like you covered all the warnings. +1 to close from me then.

[mckib2]

@mckib2 was there anything else you wanted to investigate here (e.g. I'm not sure if you saw #17432 (comment)), or do we just wait for upstream changes?

+1 to wait for upstream resolution, I think this issue can be closed

[mdhaber]

Thanks everyone!

[saprykin]

I still have problems with Boost warnings on macOS running on M2 using SciPy 1.11.3. The following code throws a warning:

import scipy.stats as st
st.ncf.cdf(1.0, 9.3068452315593326, 2.5033839617676019, 2.2008593797340321, loc=-0.00428410694979, scale=2.96164351766)

The warnings looks like:

~/miniconda3/envs/scipy-test/lib/python3.11/site-packages/scipy/stats/_continuous_distns.py:7087: RuntimeWarning: overflow encountered in _ncf_cdf
  return _boost._ncf_cdf(x, dfn, dfd, nc)

Tested the same code on a Linux machine and no warnings.

[toolslive]

This might be a similar/identical issue

from scipy.stats import ncf

In [2]: ncf.cdf(1, 1.5, 2.5, 5)
Out[2]: 0.12644056682002736

In [3]: ncf.sf([2.5],[1],[1],[2.])
../3.11.6/envs/.../python3.11/site-packages/scipy/stats/_continuous_distns.py:7094: RuntimeWarning: divide by zero encountered in _ncf_sf
  return _boost._ncf_sf(x, dfn, dfd, nc)
Out[3]: array([0.58431814])

scipy version: 1.11.3 (scipy-1.11.3-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl)
python version: 3.11.6
OS: linux
Arch: 13th Gen Intel(R) Core(TM) i7-1365U

[mdhaber]

@mborland looks like we have another one of these warnings.

from scipy.stats import ncf
ncf.sf(2., 1., 1., 1.)

Can you help us? Thank you!

[mborland]

@mborland looks like we have another one of these warnings.

from scipy.stats import ncf
ncf.sf(2., 1., 1., 1.)

Can you help us? Thank you!

I will take a look

[mborland]

@mdhaber Fix for has been merged on my end.

[mdhaber]

Thanks @mborland!