BUG: bug in nanops.var with ddof=1 and 1 elements would sometimes return inf rather than nan (GH6136)

[jreback]

closes #6136

[jreback]

@y-p I could not reproduce, but this seems likely the problem

[ghost]

Thanks. Could this be a numpy issue? should we ping someone there?

[jreback]

It comes down to precision I think;

The following 2 cases end up slightly different if the numerator is exactly 0 or not (which is why it was 'randomly' failing)

I changed it so it doesn't do the computation if the denomincator is 0, problem solved.

In [9]: Series(0.,index=np.arange(5)).values/0.
Out[9]: array([ nan,  nan,  nan,  nan,  nan])

In [10]: (Series(0.,index=np.arange(5)).values+0.1)/0.
Out[10]: array([ inf,  inf,  inf,  inf,  inf])

[ghost]

That makes sense. Still, the same data behaves differently on the same platform
and numpy. You have to cut fp some slack, but is this really an instance of that?

I'll open a numpy issue just in case. My numerical-fu isn't good enough to call it.

[jreback]

not sure I never could reproduce
even with the same exact specs

[ghost]

are you on conda? they use MKL for numerics.

[jreback]

this was a grohkle 1.8 MKL build for windows

[ghost]

Same here (didn't realize those were compiled against MKL). strange.