Re: [boost] [rfc] Inexact quantities arithmetics

From: Marat Khalili (0x8.0p15_at_[hidden])
Date: 2007-07-15 02:35:45

Hello,

Paul A Bristow wrote:
> The word error gives off some very wrong vibes, and I would strongly urge you to change to using uncertainty instead of error,
> despite the historial use of the word error.  It *really, really* isn't an error - it is a interval of (un-)certainty.

Not an interval, but it's half-width, right? What do you think about 
'uncertainty_interval_half_width'? Seriously I wanted to call it 
'deviation' first, but opted for shorter identifier; same number of 
letters as in 'value' is additional bonus, makes code more symmetrical. 
BTW shouldn't 'value' be called average, mean or expectancy?

I thought it is still colloquially being called 'error', but I'll think 
about it. Some help from English-speaking folks will be appreciated.

> You should also see
> 
> http://trs-new.jpl.nasa.gov/dspace/bitstream/2014/29189/1/95-0368.pdf

Yes, this one looks very much like one I'm thinking about. As a work of 
a government employee his code must be free, shouldn't it?

> You will see from the full code that he templates two versions, one (unusual) where the uncertainties are correlated (must add up to
> 100%) - as well as the normal uncorrelated situation.  Both are needed I think.

I'm thinking about providing a class of uncorrelated values, but return 
value from any operation on them will be correlated. Even this allows to 
underestimate the er... uncertainty if someone uses a value twice. I 
don't know how to prohibit this (compile time).

> and FWIW some of my comments some years ago
> 
> http://lists.boost.org/Archives/boost/2003/04/46828.php

Is it open-source?

> My uncertain class has a value, standard deviation, degrees of freedom, and 16
> uncertainty type flags - square, triangular, gaussian, exact etc

> I also felt it was vital to store the degrees of freedom with the value and uncertainty, and I also added some other bits of
> information like quantization (from A/D conversion), exact values, where the uncertainty estimate came from ...

This looks too heavy for me, it was tuned for some specific task, right? 
Generally, adding or substracting two lognormal values produces 
god-knows-what. Either you keep the whole (joint!) distribution or nothing.

> I built my attempt without using Boost interval library, but you might be able to use it.

I wanted to keep the main value intact. Probably value and interval is a 
way to go... Also, uBLAS vectors have problems even with wrapped double, 
I should test intervals first.

With Best Regards,
Marat