[3.2.3] bayesian_info_criterion_lsq failure with ARM64 #9657
Description
On Debian, we are currently on the transition to Python 3.8. With this, almost all tests succeed on all platforms, except the following which happens on PowerPC 64 bit and Arm64 (both little endian):
171 >>> g_init = models.Gaussian1D(amplitude=1., mean=0, stddev=1.)
172 >>> fit_g = fitting.LevMarLSQFitter()
173 >>> g = fit_g(g_init, x, y)
174 >>> # Compute the mean squared errors
175 >>> ssr_t = np.sum((t(x) - y)*(t(x) - y))
176 >>> ssr_g = np.sum((g(x) - y)*(g(x) - y))
177 >>> # Compute the bics
178 >>> bic_t = bayesian_info_criterion_lsq(ssr_t, 4, x.shape[0])
179 >>> bic_g = bayesian_info_criterion_lsq(ssr_g, 3, x.shape[0])
180 >>> bic_t - bic_g # doctest: +FLOAT_CMP
Expected:
30.644474706065466
Got:
31.11155402550662
/<<PKGBUILDDIR>>/.pybuild/cpython3_3.8/build/astropy/stats/info_theory.py:180: DocTestFailure
Full build log here. With Python 3.7, these tests succeeded.
Python is 3.8.0.
BTW, Travis offers as Beta builds on arm64, powerpc64 and s390x. arm64 and s390x may be worthwile test environments for astropy: arm64 because of the usage in small devices (Raspberry), and s390x because it is big endian.