Question

Answering one question Why does scipy.stats.nanmean give different result from numpy.nansum?, I realized, multiplying numpy.int32 by a float results in different float result compared to a Python POD int with a float.

Is there a reason to cause a float approximation when using numpy.int32

>>> numpy.int32(1) * 0.2
0.20000000000000001
>>> 1 * 0.2
0.2

Answer 1

The two expressions give results that are identical in value but have different types:

In [17]: numpy.int32(1) * 0.2 == 1 * 0.2
Out[17]: True

In [18]: type(numpy.int32(1) * 0.2)
Out[18]: numpy.float64

In [19]: type(1 * 0.2)
Out[19]: float

The different output is purely due to the difference in default formatting between numpy.float64 and float.

If we reverse the types, the output also reverses:

In [12]: float(numpy.int32(1) * 0.2)
Out[12]: 0.2

In [13]: numpy.float64(1 * 0.2)
Out[13]: 0.20000000000000001

It’s purely a display issue. There is no numerical difference here.