[Python] Numpy handling small imaginary parts differently

There is a difference in behaviour on macOS with arm64 CPU vs other Linux with amd64.

For numpy v1 on arm64, it seems there is an internal logic guaranteeing the imaginary part to be exactly 0:

>>> import numpy as np
>>> np.__version__
'1.26.4'
>>> x = 0.1 + 0.2j
>>> x * np.conj(x)
(0.05000000000000001+0j)


For v2 on arm64, this check is disabled and the imaginary part fluctuated at the border of double precision as expected:

>>> import numpy as np
>>> np.__version__
'2.3.5'
>>> x = 0.1 + 0.2j
>>> x * np.conj(x)
np.complex128(0.05000000000000001+1.6653345369377347e-18j)


For v1 & v2 on amd64, the check seems to be enabled and the imaginary part is 0 exactly.

Is there a native way to re-enable this check on macosx-arm64 and obtain results with the imaginary part equal to 0?

Continue reading...