Re: [cc65] Floating point support?

On 28 Sep 2004, at 13:54, Maciej Witkowiak wrote:

>> IMHO, the problem with FP is the following: The people mostly using 
>> them
>> are the people not knowledgeable enough to decide if this is
>> appropriate. People who are knowledgeable try to avoid them, if this 
>> is
>> possible.
>
> This is so true. Floats should be used only by people who know what 
> they
> are doing and can design both algorithm and implementation correctly.

With all due respect, that is taking it quite a bit too far. Sure, if 
you're implementing a linear algebra library... But if you're 
calculating the average cost of N shopping items?

For my work, I see a lot of science being done using numerical models. 
These guys use higly non-linear models, a complete formal error 
analysis is completely out of the question. What you see is that people 
tend to do a sensitivity test: if I change my input parameters by a 
small amount, do I still get (essentially) the same result?

Only in very regular problems (e.g., the aforementioned linear algebra 
problem) or the infamous example of solving 2nd degree polynomials it 
is possible to do a formal analysis. They told me the very thing you 
say back in school but I don't see this happen in practice.

> Otherwise possible accumulation of errors leads to gibberish results 
> and
> increasing the size of FP numbers can't help that.

Yes it can. E.g. moving up from single to double precision (6 to 15 
decimal digits!) gets many realistic problems out of a zone where 
numerical roundoff introduces so much noise that all information is 
lost.

Regards, Sidney

----------------------------------------------------------------------
To unsubscribe from the list send mail to majordomo@musoftware.de with
the string "unsubscribe cc65" in the body(!) of the mail.