You are viewing a single comment's thread.

view the rest of the comments →

[–] 26560437? ago  (edited ago)

The issue here is that floating point arithmetic is not perfectly reversible. The exponent and mantissa are both variable and there are numbers without perfectly round representations which lead to an increase in Epsilon as the number of operations increases. Yeah, we can create a "probable" reversal but with the number of adjustments I've seen it's likely raised Epsilon really high.

The question is how much tolerance for error does an election need? If we cannot accept a single vote error then Epsilon is a serious factor. If not, we'd have to make the estimation so that Epsilon was minimized.