Float Precision

[mzuenni]

The spec states:

[...] floats of arbitrary size and precision are supported.

But the following code fails:

ASSERT(0.1+0.1+0.1==0.3)

I would have expected 0.1 and 0.3 are represented perfectly and therefore causing no issues.

[eldering]

That's because even if the precision can be arbitrarily high (using libGMP floats), it is not infinite or using a rational representation, so they still suffer the same issue that some decimal floating points numbers cannot be exactly represented, hence equality will fail, no matter what.

[mzuenni]

hm that sounds like a "won't fix" to me ^^
(my reimplementation uses fractions, since all expressions in a ctd checker can always be represented that way, i.e., nothing like sqrt, log, pi, ... can occur)

[eldering]

hm that sounds like a "womt fix" to me ^^ (my reimplementation uses fractions, since all expressions in a ctd checker can always be represented that way, i.e., nothing like sqrt, log, pi, ... can occure)

Yeah, I was thinking about fractions indeed when I initially replied. It's a good point that currently all operations are rational, but using that fact would preclude adding non-rational functions later...

Do you actually want to use this anywhere? Because I wonder where this would be useful, and actually whether using GMP floats are even that useful.

[mzuenni]

Do you actually want to use this anywhere?

what exactly do you mean here?

Because I wonder where this would be useful

I wouldn't be surprised if there is a problem with floats where you want to ensure that the sum of all floats is exactly X (e.g. probabilities with sum 1.0).