FWIW in pandas we post-process floordiv (and divmod) ops to get the "Expected Result" behavior from the OP.

On Fri, May 8, 2020 at 11:56 AM Anirudh Subramanian <anirudh2290@gmail.com> wrote:
Hi all,

There has been a discussion about divmod (1.0, 0.0) bug here : https://github.com/numpy/numpy/issues/14900 and https://github.com/numpy/numpy/pull/16161 . 


Currently divmod(1.0, 0.0) sets the "Invalid error" and returns (nan, nan). This is not consistent with IEEE 754 standard which says that 1.0/0.0 divisions should return inf and raise dividebyzero error. Although this may not apply to divmod, it should apply to floor_divide and mod. 
I have summarized in the table below, summarizing current state and expected state.

OperatorWarning messageExpected warningResultExpected Result
np.divmodInvalid errorinvalid and dividebyzero ??nan, naninf, nan
np.fmod(1.0, 0.0)Invalid errorInvalidnannan
np.floor_divide(1.0, 0.0)Invalid errorDividebyzeronaninf
np.remainder(1.0, 0.0)Invalid error Invalidnannan

For remainder and fmod above, according to the standard, it is supposed to raise invalid error. We need to change the code to also raise dividebyzero error for floor_divide.

The question is what to do for np.divmod (since this is not defined by standard). My opinion is in this case we need to set both dividebyzero and invalid error flags since its a combination of these two operations.


This is going to cause a breaking change/silent incorrect results to atleast some users who are either doing one or two of the following:
1. expecting nans from their output and check isnan but not isinf in their code and accordingly do further computations. 
2. who currently call raise only on invalid and not dividebyzero errors for any of  the above listed operations.

Considering this we can try one of the two things:
1. Create a futurewarning for every divmod(1.0, 0.0) path. This may be very noisy and I cannot think of an action for a user to take to suppress this.
2. Since bug fixes are exempt from backwards compatibility policy just notify in the release notes, maybe after a couple of releases. (More Impactful)


Depending on the compiler, and if it implements annex F of the C standard, it may not support 1.0/0.0 operation and may crash. Also this is the case for true_divide also, so we wont be breaking more users than we currently are.

Would like to hear your thoughts about this!


NumPy-Discussion mailing list