Re: How do I bisect unbounded intervals?
Dan Zuras wrote:
> Someone suggested an information theory midpoint. In
> floating-point this is quite close to a geometric mean
> in that it finds that floating-point number for which
> the number of representable numbers between lo & mid
> is nearly equal to the number between mid & hi.
>
> There is merit in this but I would not give it the name
> "midpoint". Perhaps "split".
Neat! Convert both endpoints to sortable, take the average,
and convert back to floating-point! ("Sortable" is a crude but
invertible transformation that permits basic string comparison
to be applied to the bits. On big-endian machines for binary
formats this involves inverting the sign bit and, if the input
was negative, inverting all other bits as well. For DFP it's a
bit more complicated but still doable.) (NaNs need to be dealt
with first, or they would be bigger than infinity.)
(This is unfortunately a Level 4 concept. Well, we are supposed
to avoid Level 3 issues, but nobody said anything about Level 4!)
Michel.
---Sent: 2012-01-13 05:46:59 UTC