Time construction with format='mjd' inaccurate by >600 ns #9327
Description
If I construct an astropy.time.Time object with format='mjd' and scale='tai', for certain (future) values I lose >600 ns of accuracy. The same does not occur if I construct directly with format='jd' that I constructed using day_frac. Example in the form of a test case:
def test_mjd_initialization_precise():
i, f = 65536, 3.637978807091714e-12 # Found using hypothesis
t = Time(val=i, val2=f, format="mjd", scale="tai")
jd1, jd2 = day_frac(i + erfa.DJM0, f)
jd1_t, jd2_t = day_frac(t.jd1, t.jd2)
assert (abs((jd1-jd1_t) + (jd2-jd2_t))*u.day).to(u.ns) < 1*u.nsand the output:
___________________________________________ test_mjd_initialization_precise ___________________________________________
def test_mjd_initialization_precise():
i, f = 65536, 3.637978807091714e-12 # Found using hypothesis
t = Time(val=i, val2=f, format="mjd", scale="tai")
jd1, jd2 = day_frac(i + erfa.DJM0, f)
jd1_t, jd2_t = day_frac(t.jd1, t.jd2)
> assert (abs((jd1-jd1_t) + (jd2-jd2_t))*u.day).to(u.ns) < 1*u.ns
E assert <Quantity 628.64273787 ns> < (1 * Unit("ns"))
E + where <Quantity 628.64273787 ns> = <bound method Quantity.to of <Quantity 7.27595761e-12 d>>(Unit("ns"))
E + where <bound method Quantity.to of <Quantity 7.27595761e-12 d>> = (7.275957614183426e-12 * Unit("d")).to
E + where 7.275957614183426e-12 = abs(((2465537.0 - 2465537.0) + (-0.499999999996362 - -0.49999999998908606)))
E + and Unit("d") = u.day
E + and Unit("ns") = u.ns
E + and Unit("ns") = u.ns
astropy/time/tests/test_precision.py:146: AssertionError