Sebastian Nowozin <nowozin <at> gmail.com> writes:
The values I compare come from a Monte Carlo procedure, and are guaranteed to lie within a certain absolute tolerance. I compare the Monte Carlo results against exact values. Numerical stability is not an issue as these values are within the range zero to one.
This is quite a big range. You can get values around 0.1 or around 1e-15. Will you be using the same tolerance in these cases?
Therefore, you can safely assume that absolute deviation is the criteria I want to use, and is popularly used in the literature for this purpose (the so called total variation distance). I do not want to use relative comparisons, in particular because normalizing by a very small value such as 1.0e-10 can happen in that case.
So? This will only make the comparison fair. 1e-10 and 2e-10 will differ by 1e-10, but the second value is twice as big. 0.9000000001 and 0.9000000002 differs only insignificantly. How can you catch this difference with absolute comparison? BOOST_CHECK_CLOSE on the other hand will work correctly. Gennadiy