Re: Accuracy of Statistical Functions

[Leonard Mada <discoleo gmx net>]

Hello,

seems gnumeric is OK. :-)

If you receive this e-mail 2 times, take my apologies. I noticed too 
late that I sent it first to the wrong e-mail address.

Well, it was my fault as I opened the OOo ods document, and saw the 
results displayed as in OOo Calc, because gnumeric did *NOT recalculate* 
the formulas. Maybe there should be a method to recalculate the formulas 
in a document, especially IF the document was last *written* by a 
different application (I do NOT know if this can be read in the .ods 
document, though).

I noticed only today, that gnumeric is indeed more accurate. Good to 
know that! :-)

However, there are still some interesting points I discussed in the 
respective thread on the OOo mailing list (although the OOo developers 
were more or less reluctant for any fresh thoughts).

I have gnumeric 1.7.1, Windows edition (W2k SP4). I will make additional 
tests over the weekend, but until now it looks good for gnumeric.


Morten Welinder wrote:
> Note, that sorting is not the greatest thing to do.  For example if 
> you have,
> just three numbers
>
>     small, +huge, and -huge
>
> then you will actually want to add ±huge and get zero before you look at
> the small number

*Sorting* is a more general mechanism and should work good for many 
situation, not just for a small number of situations like the one 
mentioned above. But if actually the mathematical handling is more 
robust, its influence on accuracy won't be as big. I now believe that 
gnumeric is *quite robust* (and I saw the misleading values because 
gnumeric did NOT recalculate the formulas; sometimes shit happens).

The reason why sorting works is that, as you add the smaller numbers 
first, more less significant decimal places will be added, too, and may 
influence more significant digits. When you add larger numbers first, 
there is a higher chance to drop/round less significant digits.

Kind regards,

Leonard Mada