Increase in uncertainties with version 3.0.1 wrt 2.0.1
Dear experts,
when changing version of RooUnfold from 2.0.1 to 3.01, there is suddenly see significant increase in statistical uncertainty when using kCovToys with bayesian unfolding:
where left it 2.0.1 and right is 3.0.1. In this example it is just single iteration where the input and matrix are the same MC.
The only difference code-wise is due to change in interface between the releases:
unf = unfold.Hreco(ROOT.RooUnfold.kCovToy) # 2.0.1
unf = unfold.Hunfold(ROOT.RooUnfolding.ErrorTreatment.kCovToys) # 3.0.1
This problem was confirmed by @adye in the support mailing list:
"I don't know for sure which is the better measurement (did we break something, or fix it?), but in my tests, I noticed that the 2.0.1 CovToys errors agree with the default (propagation of errors) calculation, while the 3.0.1 errors are much larger. Also 2.0.1 gives a pull RMS of ~1, but ~0.4 in 3.0.1 (is this a good check?)."
It would be good to know what caused this change and which version is correct.