Re: Generalised bisection

[Steven Tweed <orthochronous@gmail.com>]

On Thu, Mar 12, 2009 at 10:55 AM, Johannes Schindelin
<Johannes.Schindelin@gmx.de> wrote:
> On Thu, 12 Mar 2009, John Tapsell wrote:
> > 2009/3/11 Ealdwulf Wuffinga <ealdwulf@googlemail.com>:
> > > What I use is the multiprecision floating point number class. doubles
> > > don't seem to be long enough.
> >
> > Hmm, really really?  Sometimes this sort of thing can be fixed by just
> > readjusting the formulas.  What formulas are you using that require more
> > precision than doubles?
>
> Maybe you could post the formulae instead of forcing people to deduct them
> from the source code?

I haven't even looked at the source code so a description of the
mathematical algorithm would help, but I'll just point out that
underflow (in the case of working with probabilities) and overflow
(when working with their negated logarithms) is inherent in most
multi-step Bayesian algorithms. The only solution is to rescale things
as you go so that things stay in a "computable" range. (You're almost
never interested in absolute probabilities anyway but rather relative
probabilities or, in extreme cases, just the biggest probability, so
rescaling isn't losing any useful information.)

cheers,
dave tweed