np.log1p(inf) returns NaN while it should return inf #4225

Activity

1. WarrenWeckesser commented on Jan 23, 2014

   

   WarrenWeckesser

   on Jan 23, 2014

   Member

   More actions

   I get the following on Windows 7 (64 bit), using Anaconda 1.7.0, which includes numpy 1.7.1 (that's all I have installed in my Windows VM at the moment):

   >>> import numpy as np
   >>> np.__version__
   '1.7.1'
   >>> np.log1p(np.inf)
   __main__:1: RuntimeWarning: invalid value encountered in log1p
   nan
   

   np.log works as expected:

   >>> np.log(np.inf)
   inf
   

   Also, np.expm1 gives nan when it should give inf:

   >>> np.expm1(np.inf)
   __main__:1: RuntimeWarning: invalid value encountered in expm1
   nan
   >>> np.exp(np.inf)
   inf
   
2. juliantaylor commented on Jan 23, 2014

   

   juliantaylor

   on Jan 23, 2014

   Contributor

   More actions

   looks like our no-C99 fallback is not C99 compliant again, it divides inf/inf = nan

   does windows really not have C99 math or are our binary builds just not correct?
3. charris

   added

   Defect

   on Feb 24, 2014
4. juliantaylor commented on Mar 2, 2014

   

   juliantaylor

   on Mar 2, 2014

   Contributor

   More actions

   the mingw binaries which we have on sourceforge seem ok, so its just the ms compiler builds that are broken.

   while i know ms does not care about C99 as the language, do they really not care about correct math? C++ likely uses the same library.
   Fixes for the non mingw buildsystem are appreciated.
5. juliantaylor

   added 3 commits that reference this issue on Mar 2, 2014

   BUG: fix non-c99 fallback for np.inf input to log1p/expm1

   ...

   This issue will close once commit 18cda3b is merged into the 'master' branch.

   18cda3b

   BUG: fix non-c99 fallback for np.inf input to log1p/expm1

   ...

   This issue will close once commit f53ab41 is merged into the 'master' branch.

   f53ab41
6. charris

   closed this as completedin #4415on Mar 2, 2014