Could `verify` use either `float_rounding` or `numeric_precision` and not both?

[gsganden]

float_rounding and numeric_precision do almost the same thing in verify, and having both is potentially confusing. For instance, verify(parse("0.0000001"), parse("0.0000002"), numeric_precision=99999) comes out True because float_rounding is 6 by default.

Could we get rid of float_rounding and just relying on numeric_precision? Would you like me to submit a PR that makes this change?

[hynky1999]

Hi, they are almost the same but have different purpose.
float_rounding is to ensure that in some cases we would like restrict direct float precision to specific precission. Many eval datasets use 5/6 for float approximation e.g 1/3 = 0.33333. Float rounding ensures that in case of 0.33333 and 1/3 with proper rounding the verify will return True.

numeric_precision is something I would prefer ditching completely, but because computers are still running in binary system with finite precision, we have to assume that there is a difference in number not because of incorrect answer but because of inherent precision issues.

Therefore I wouldn't remove the args

[gsganden]

Thanks for the response!

I'm afraid I'm not following. Rounding to say 5 decimal places before checking agreement and checking agreement to 5 decimal places is effectively the same thing except in the occasional case where one number gets rounded up and the other gets rounded down. I don't see why you would use one mechanism for approximating rational numbers with decimals and another for coping with approximating decimals on binary systems. It still seems to me that providing both as parameters is unnecessary and confusing.

That said, I could be missing something.

[hank0316]

+1. I’m also confused. Can anyone explain the design?