Agreed, a benchmark would be cool. From what I've read before it ends up taking longer than the integer methods anyway because you still need to keep log n bit

Agreed, a benchmark would be cool. From what I've read before it ends up taking longer than the integer methods anyway because you still need to keep log n bits of precision in the floating point values around to get the correct answer. Perhaps if you had a use case for approximate values of very large fibonacci numbers it would be useful? I'd love to hear that use case.