When I run through my head examples of how that would work in
algorithms I have worked on, it seems to me that it (a) it starts out
with less precision than floating point (because of the extra bits
being used for representing an estimate of accuracy), and (b) that it
would tend to also lose precision faster (because it started out with
less, so the fractional bits being lost are more significant).

Put differently: as long as (a) this representation stays close to
original data, and (b) the people using it understand in detail how it
works, it will probably be ok. But run this through a lengthy sequence
of calculations and it'll mess up faster than floating point.

Put differently: I prefer J's approach of providing ":!.precision (or
9!:11) over these things.

That said, running through the calculation once using unums (to get a
precision estimate) and then running through it again using floating
point (to get a more precise result) might be a useful approach (for
applications where the factor of 2 time and space cost is acceptable).

That said, it is fun reading about how other people think about these things.

Then again, there presumably must be applications where unums are a
better fit than floating point? (I just don't know much about what
those might be, off the top of my head.)

-- 
Raul



On Wed, Apr 27, 2016 at 11:36 PM, 'Jon Hough' via Chat
<[email protected]> wrote:
> This interview is pretty interesting, about a new number format that will 
> solve floating point related errors:
> http://ubiquity.acm.org/article.cfm?id=2913029
>
> see also: 
> http://motherboard.vice.com/read/a-new-number-format-for-computers-could-nuke-approximation-errors-for-good
>
> I wonder about a J implementation of unums...  ...it seems Julia (among 
> others) has one
> https://github.com/REX-Computing/unumjl
>
> ----------------------------------------------------------------------
> For information about J forums see http://www.jsoftware.com/forums.htm
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm

Reply via email to