point is ... I only roughly remember my CS lessons on error propagation and I not very keen to open old books again. ... ;)

If I had the time I would try to reproduce your results with Math::BigFloat cause I don't trust much into those C libraires.

Please keep in mind a method can have a better error margin (ie worst case scenario) while an alternative approach is statistically far better (in the middle).

it's even more complicated if you start discussing which input set of calculations should be considered representative for such a statistical investigation.

for instance x**6e7 ... seriously? :P