Hi Felix,
> what I mean by relatively large is this 1 x 10-3 compared
> with something like 4.9406564584124654 x 10−324
> (Double floating-point).
Yeah, but you may have gotten a bit of a wrong idea about floating point numbers though - even though the smallest representable number will be something similar to what you're saying there that doesn't mean that you can do much with that number other than just look at it by itself. Once you start doing arithmetic between different numbers you're looking at accuracy of more like 1e-16 or something more in that range.
The smallest possible number that can be represented is not also the accuracy that arithmetic works with...
Each floating point number has a separate mantissa and exponent part to it, and when you do arithmetic between 2 numbers if their exponents are way different the one that's way smaller will sort of evaporate away.
The "floating" part is kind of like a floating range - you get precision with arithmetic between numbers that are within the same exponent range of each other, so like 3.0e-100 + 4.0e-100 will generate 7.0e-100 ok but if you try to do 1 + 3.0e-100 it will just give a result of 1 because there's a 100 digit difference between those numbers.
> Maybe, just maybe it would be possible to have something
> like an on-demand tolerance control.
I may have this eventually but it tends to open up a pandora's box of various problems when it is set wrongly and I've often seen it set badly in practice before.
See this thread for some previous discussion on it:
http://moi3d.com/forum/index.php?webtag=MOI&msg=4215.4
- Michael
|