You can feed NaNs to almost any numpy function. The result will simply be nan. Except for sign: It's the only function I've found that additionally emits a warning.

MWE:

import numpy

numpy.sqrt(numpy.nan)  # okay, nan
numpy.sin(numpy.nan)  # okay, nan
numpy.abs(numpy.nan)  # okay, nan

numpy.sign(numpy.nan)  # warning! nan