Both f(x) and g(x) could be calc'ed to proper machine precision (say, GSL double prec ~ 1E-15). Would this imply, that beyond the machine precision the values are effectively fixed to resp. zero or infinity, instead of carrying around the extreme orders of magnitude?
I'm not exactly sure what you're asking here, but the point is that "to machine precision" is relative: if f(x) and g(x) are O(1e200), then the absolute error of each is still O(1e185). However f(x)/g(x) will still be very accurate (with absolute error O(1e-15)).
