nans and object identity

[sjmillius]

Description

nans (like float("nan"), np.nan) follow the IEEE 754 standard for floating-point arithmetic, which specifies that NaN is not equal to anything, including itself, so:

np.nan == np.nan  # False
float("nan") == float("nan") # False

But

np.nan is np.nan  # True

So these are objects that are not equal but identical to themselves. This leads to unexpected results in cases where object identity is taken as a short cut before equality comparisons, e.g. in tuples.

Snippet preview

>>> import numpy as np
>>> np.nan == np.nan
False
>>> (np.nan, np.nan) == (np.nan, np.nan)
True
>>> nan1 = float('nan'); nan2 = float('nan')
>>> nan1 == nan2
False
>>> (nan1, nan2) == (nan1, nan2)
True
>>> (nan1, nan2) == (nan2, nan1)
False

Checklist before calling for maintainers

• Have you checked to ensure there aren't other open Issues for the same update/change?
• Have you checked that this snippet is not similar to any of the existing snippets?

• Have you added an Explanation section? It shall include the reasons for changes and why you'd like us to include them

[Henrik1248]

This behaviour is intriguing, another interesting result:

>>> import numpy as np
>>> [np.nan] == [np.nan]
True