BUG: numpy logarithm behavior does not match CPython 3.14 cmath.log behavior #30470
Description
Describe the issue:
In CPython 3.14, the behavior of logarithms of complex numbers changed for some edge case inputs in python/cpython@2cb84b1, causing behavior to differ beyond floating point precision differences between
import cmath
cmath.log(x, base)and
import numpy as np
np.log(x) / np.log(base)where
x = 1j
base = -cmath.inf + cmath.nanjReproduce the code example:
import cmath
import numpy as np
import warnings
def python(x, base):
return cmath.log(x, base)
def numpy(x, base):
return np.log(x) / np.log(base)
def numpy_scimath(x, base):
return np.emath.logn(base, x)
def main():
x = 1j
base = -cmath.inf + cmath.nanj
funcs = python, numpy, numpy_scimath
names = list(map(lambda x: x.__name__, funcs))
padding = max(map(len, names))
for name, func in zip(names, funcs):
print(f"{name}".ljust(padding) + ":", func(x, base))
if __name__ == "__main__":
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=RuntimeWarning)
main()In 3.13, this produces:
python : (nan+nanj)
numpy : (nan+nanj)
numpy_scimath: (nan+nanj)
In 3.14, this produces:
python : 0j
numpy : (nan+nanj)
numpy_scimath: (nan+nanj)
Error message:
N/A, there's no error here.Python and NumPy Versions:
Python versions are 3.13 and 3.14
numpy version is 2.3.4 in both cases
Runtime Environment:
[{'numpy_version': '2.3.4',
'python': '3.13.9 (main, Oct 14 2025, 13:52:31) [GCC 14.3.0]',
'uname': uname_result(system='Linux', machine='x86_64')},
{'simd_extensions': {'baseline': ['SSE', 'SSE2', 'SSE3'],
'found': ['SSSE3',
'SSE41',
'POPCNT',
'SSE42',
'AVX',
'F16C',
'FMA3',
'AVX2'],
'not_found': ['AVX512F',
'AVX512CD',
'AVX512_KNL',
'AVX512_KNM',
'AVX512_SKX',
'AVX512_CLX',
'AVX512_CNL',
'AVX512_ICL',
'AVX512_SPR']}},
{'filepath': '/nix/store/xm08aqdd7pxcdhm0ak6aqb1v7hw5q6ri-gcc-14.3.0-lib/lib/libgomp.so.1.0.0',
'internal_api': 'openmp',
'num_threads': 32,
'prefix': 'libgomp',
'user_api': 'openmp',
'version': None}]
Context for the issue:
I'm not sure about how to prioritize this, but the change affects numba and numba-cuda, since they depend on numpy's behavior for these functions -- apparently even in the stdlib case, since neither of them calls the stdlib function directly when a user uses these functions.