Re: [boost] Accumulators review

From: Eric Niebler (eric_at_[hidden])
Date: 2007-01-31 13:42:00

John Maddock wrote:
> Paul A Bristow wrote:
>> PS and about your torture test - I don't think this should be the
>> focus of the review.
>>
>> There are horses for courses, and the not-always-accurate version
>> should be fine for most applications. (If the data are going to be as
>> unreasonable as the torture data are, one could argue that they are
>> wrong!
> 
> In case it wasn't clear from my post, I actually agree with this.  The 
> library can be extended to more/better algorithms at a later date anyway, 
> the important thing at this stage is the framework.

Right! :-)

> What is important though is documentation: most people would miss the suble 
> difference between the "immediate" and "naive" variance calculations, and 
> potentially (although very rairly) fall into all kinds of hard-to-spot 
> traps.

Agree 100% -- the docs should be clearer on this point. I may be 
following up with you and Matthias offline for ideas about how best to 
document the complexity and error of the various statistical 
accumulators. But a big WARNING up front couldn't hurt, either.

-- 
Eric Niebler
Boost Consulting
www.boost-consulting.com