To clarify, it took me/us quite a while to understand why MPFR has what it has; see https://github.com/sagemath/sage/pull/41288#issuecomment-3671350964 Some takeaways from the explorations there: - MPFR is right in having its different "rint_*" routines, as these all have subtly different behaviour and their behaviour cannot be achieved simply by reducing the number of digits and using a rounding mode. That means that borrowing the rounding mode on the relevant "real field" does not solve the problem completely. "trunc, ceil, floor" are still specific - IEEE does seem to prefer round-to-even in various settings and so do python floats. - It doesn't look like MPFR made a *choice*. They just provide the different possibilities; not necessarily under a completely regular naming scheme. - We should make sure that existing code continues to run; hopefully with the same results, but if really necessary some different rounding would be OK (when you rely on floats there is always the possibility of changes due to versions not quite preserving same behaviour on the bit level). I'd be against any environment variables or global state for this. The rounding issue here is really an edge case and would not matter for code that genuinely depends on floating point, so we should do something sensible but foremost simple. Having to deal with deprecation messages popping up is almost certainly going to be much more disruptive than even a silent change to different sensible behaviour.
Note that the mpfr rounding routines express the resulting integer value still as an mpfr float with a limited number of bits, contrary to our rounding routine that is supposed to spit out an actual integer. On Tuesday, 7 April 2026 at 06:59:55 UTC-7 cxzhong wrote: > I have opened a new PR about this. If we do not want to change the > rounding method. we can follow this > Add rnd parameter to RealNumber.round() by cxzhong · Pull Request #41945 · > sagemath/sage <https://github.com/sagemath/sage/pull/41945> > > 在2026年4月7日星期二 UTC+8 16:24:25<cxzhong> 写道： > >> I think this fixes the strange behavior like >> ``` >> sage: (5/2).round() >> 2 >> sage: (-5/2).round() >> -2 >> sage: (2.5).round() >> 3 >> sage: (-2.5).round() >> -3 >> ``` >> >> 在2026年4月7日星期二 UTC+8 16:20:41<cxzhong> 写道： >> >>> It mainly about this >>> Modify rounding method to use banker’s rounding by cxzhong · Pull >>> Request #41268 · sagemath/sage >>> <https://github.com/sagemath/sage/pull/41268> >>> I would like to change the default rounding method of MPFR number from >>> rounding to infinity to rounding to even. to satisfy the python's general >>> implement >>> and you can also see the discussion in Fix incorrect tests and >>> override canonical associate by cxzhong · Pull Request #41288 · >>> sagemath/sage <https://github.com/sagemath/sage/pull/41288> >>> Thanks >>> Chenxin >> >> -- You received this message because you are subscribed to the Google Groups "sage-devel" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion visit https://groups.google.com/d/msgid/sage-devel/f5241791-7c12-4c6c-8193-590ae044b940n%40googlegroups.com.
