From: Brook Milligan (brook_at_[hidden])
Date: 2007-05-07 20:19:31
Several people have commented, both publicly and privately, on the
Probability library I mentioned last week. There is now a new version
that addresses most of the concerns. In what follows I will address
the salient points as I see them.
- The main documentation page now begins with a brief definition of
probability and likelihood.
- Runtime costs have now been quantified in a fairly simple manner.
The results are summarized on the main page, but indicate that there
is less than a 0.5% effect in a test involving a large fraction of
operations on these quantities. Is suspect this is well within the
noise, but input from those with greater benchmarking experience is
- Additive operators are now provided within the log domain. This
completes the full set of arithmetic operators.
- A suggestion was made to combine this with the math toolkit (and
possibly the units) library. I hesitate to do this immediately
until it is clear that the Probability library is indeed acceptable.
It seems that the process would occur in stages: handle this one on
its own, then work on integration if that is generally a desirable
direction. This should stand on its own merits, at least initially.
- Another suggestion focused on the potential for a numerical value
type for the log domain, independent of probabilities. Clearly,
that is contained within this and such a type could be extracted out
for independent use. Had such a type existed, a portion of this
library would have been simpler. However, such a type will not
address the interconversions between probabilities and likelihoods
that form a natural part of much statistical modeling. Thus, the
higher level types incorporated here remain important, with or
without a general log domain type. For now it seems that this is an
implementation detail from the perspective of the Probability
library. If there is a strong interest in such a type, perhaps this
library could be refactored into two.. Again, I would opt for
waiting to assess the acceptability of this library and the general
level of interest in these different facets.
I appreciate the comments and welcome other ideas. I hope that more
people will look over the new version of the library and provide
Thanks for your interest.
-- Brook Milligan Internet: brook_at_[hidden] Department of Biology New Mexico State University Telephone: (505) 646-7980 Las Cruces, New Mexico 88003 U.S.A. FAX: (505) 646-5665
Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk