Rounding error in (**)/2

[ridgeworks]

@jan

I’m seriously wondering how many people care about the precision of floating point operations (assuming the error is not more than a couple of ulps).

I strongly suspect very few people care.Otherwise, there would be correct, precise C math libraries in common usage.

What I care about is that the float rounding modes correctly bound any precise real values. The underlying assumption is that compiled arithmetic is IEEE compliant for all rounding modes and the functions in math libraries are correct within 1 ulp in default “to_nearest” rounding mode. Largely that seems to be true, but rigorous testing is hard and there are no guarantees.

@j4n_bur53

But I am not publicitly testing intervals, I rather leave that to the interval owner.

Most likely buggy intervals will affect clpBNR, but again, also not my roadworks.

Hence my concern. But interval arithmetic is not built into SWIP, so it’s not Jan’s “roadworks” either. It does require IEEE compliance as noted above, unless you’re building a mixed Prolog and C implementation.