Small Discrepancy between Real and Fourier Space Shift about galsim HOT 4 CLOSED

I couldn't reproduce the exact result you had. On my laptop, this results in the following:

A little bit smaller residuals, and a different pattern. Apparently the rounding behavior on my laptop is somewhat different than on collab. If you use double precision though, the differences are ~1.e-15.

So I think this isn't a problem with the method per se. Just single-precision floating point errors accumulating. But they are a fair bit higher than the single-precision epsilon value, so I switched the code to use double precision for the phases, which basically fixed it. cf. #1234

from galsim.

Hum interesting, I wouldn't have guessed platform-specific discrepancies at that level. But ok, if you recover 1e-15 using double precision then you are right it's not related to an approximation in this formula used to compute the phase.

from galsim.

The first equation is nominally exact, but can have rounding errors. The line you pointed to was intended to get back to |z|=1, which it does to the accuracy of the data type. So the errors are just that numerical rounding errors can add up, maybe semi-coherently, rather than purely randomly. And apparently differently on different systems. That surprised me too, FWIW.

from galsim.

Done. #1234

from galsim.