API: expected result for pow(1, pd.NA) or pow(pd.NA, 0)

[TomAugspurger]

Should pow(1, pd.NA) be 1 or NA?

In [1]: import pandas as pd

In [2]: 1 ** pd.NA
Out[2]: NA

In [3]: import numpy as np

In [5]: 1 ** np.nan
Out[5]: 1.0

cc @jorisvandenbossche

[TomAugspurger]

R gives 1

> 1 ^ NA
[1] 1

[jorisvandenbossche]

Julia seems to do missing:

julia> 1 ^ missing
missing

But I think the rationale of returning 1 (i.e. "whathever number is used, 1 to the power something will always be 1") also makes sense.

[jschendel]

Likewise there's an inconsistency with np.nan for pow(pd.NA, 0):

In [1]: import numpy as np; import pandas as pd; pd.__version__
Out[1]: '0.26.0.dev0+1155.ged20822a5'

In [2]: pd.NA ** 0
Out[2]: NA

In [3]: np.nan ** 0
Out[3]: 1.0

[jorisvandenbossche]

R also gives 1 then:

> NA ^ 0
[1] 1

[jorisvandenbossche]

I am +1 changing those two cases

[brandon-b-miller]

@TomAugspurger @jorisvandenbossche Sorry to raise an old issue, I've been looking at aligning cuDF with pandas for these special cases. I believe I follow what's been discussed so far here, as well as in the associated PR, however there's arguably some consistency that is broken by stopping nulls from propagating for special values. Would you mind sharing your thoughts on this tradeoff?

I don't think consistency in itself is necessarily a rock solid reason to change this, but also am struggling to be completely convinced that this is exactly right either.

[TomAugspurger]

I think our invariant is that NA represents unknown. So if you have an operation that is know with only one operand, like “True | NA”, then the output should be known and not NA.