I understand the necessity for epsilon in floating-point comparisons, but is there some geometric interpretation to it? Something like "the minimum edge length to not cause errors is x * epsilon" for some factor x?