News for hackers. Stuff that matters.
Re:Intel Underestimates Error Bounds by 1.3 Quintillion by nhaehnle on Thursday October 09, 2014 @07:12PM (#8435475)
Actually, epsilon is guaranteed to be much larger than the minimum representable float: double(pi) is 1.57... * 2^1, so double(pi) - pi is going to be on the order of 2^-m, where m is the size of the mantissa. The only way that double(pi) - pi could be much smaller is if pi magically had a lot of 0s in its binary representation (it doesn't).

This caught me off-guard at first as well, but it's really quite logical.

Re:Intel Underestimates Error Bounds by 1.3 Quintillion by colanderman on Thursday October 09, 2014 @08:59PM (#8435748)
The only way that double(pi) - pi could be much smaller is if pi magically had a lot of 0s in its binary representation (it doesn't).

Yes, this is what I was getting at. Obviously the statement is true for pi and the majority of reals; I just don't understand why it should be a rule as the article implies.