1 | | The potential problem is less about omitting rare small values and more about missing lots of representable intermediate values. I think with this kind of approach we're losing half of the representable values between 0.25 and 0.5-2^-52^; 3/4 of the possible values between 0.125 and 0.25-2^-51^, etc., which was kind of the motivation for the Downey paper. We might decide we don't care, but I think that's highly dependent on what the caller ends up doing with the results. (We should also document this choice; the comment currently focuses on small rare values instead of "holes" in the range.) |
| 1 | The potential problem is less about omitting rare small values and more about missing lots of representable intermediate values. I think with this kind of approach we're losing half of the representable values between 0.25 and 0.5-2^-54^; 3/4 of the possible values between 0.125 and 0.25-2^-55^, etc., which was kind of the motivation for the Downey paper. We might decide we don't care, but I think that's highly dependent on what the caller ends up doing with the results. (We should also document this choice; the comment currently focuses on small rare values instead of "holes" in the range.) |
| 2 | |
| 3 | edit: fix some exponents I miscomputed due to a sign error |