Re: [boost] Boost Digest, Vol 1227, Issue 2

From: Paul A Bristow (pbristow_at_[hidden])
Date: 2005-09-27 04:25:26

| -----Original Message-----
| From: boost-bounces_at_[hidden] 
| [mailto:boost-bounces_at_[hidden]] On Behalf Of ajb_at_[hidden]
| Sent: 27 September 2005 03:20
| To: boost_at_[hidden]
| Subject: Re: [boost] Boost Digest, Vol 1227, Issue 2

|  AND the other missing, and in many cases MUCH more 
| important IMHO, math 'special' functions,
| > mainly requiring the Incomplete beta function,
| > as I proposed
| >
| > http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2004/n1668.pdf
| >
| > but which was rejected as 'too difficult' - for which I 
| suspect means 'too  expensive').
| 
| Certainly not "too expensive" in terms of computing resources; as I
| understand it, incomplete beta is no harder to compute than gamma.
| Perhaps they meant too difficult to get a useful TR out that people
| would actually be able to implement in reasonable time?

The expense is the brain power required to produce a high quality
implementation which proved more than expected for TR1 functions.

IMO part of the problem is the unrealistic expectation of the Standards WGs
for an accuracy similar to that
for most of the other functions like sin, cos for which it IS feasible,
albeit with care and skill, to get within a bit (ulp) or few.

For some of the other functions, getting within a few decimal places
is challenging.

But the results are still far more accurate than required to be
'fit for purpose' in real-life statistics applications,
like comparing two means to assess the probablity that they are different.

Paul

Paul A Bristow
Prizet Farmhouse, Kendal, Cumbria UK LA8 8AB
[phone]  [phone]
mailto: pbristow_at_[hidden]  www.hetp.u-net.com