Changes between Initial Version and Version 1 of Ticket #23061, comment 38


Ignore:
Timestamp:
Aug 29, 2017, 1:31:11 AM (2 years ago)
Author:
teor
Comment:

Legend:

Unmodified
Added
Removed
Modified
  • Ticket #23061, comment 38

    initial v1  
    44> Is there also no point in trying to fill the holes of missing representable values that are between the steps of size 2^-53^? (losing one bit of precision with probability 0.5; two bits with probability 0.25; etc.) Also intermediate results with greater precision may occur (e.g., on x87 -- this is probably how you got greater than 53 bits of precision reported in your tests with the "naive" approach) and maybe we should allow those?  Or does calling into `tor_mathlog()` round it back down to double precision anyway?
    55
    6 I don't believe there is any point in trying for anything more than N*2^-53^ (N integer) in a generic function. If we need more than that for a specific distribution, we should change the transform we use for the distribution, or feed it random bits directly.
     6I don't believe there is any point in trying for any more precision than N*2^-53^ (N integer) in a generic function. If we need more than that for a specific distribution, we should change the transform we use for the distribution, or feed it random bits directly.
     7
     8Edit: be more precise