Scaling of fit errors for weighted events

Hello.

I’m doing an unbinned fit of weighted events with rooFit using ROOT 5.34/01 build for mac os X 10.6.8.

The events are fitted with a Crystalball function convoluted with a Breit-wigner. My problem is about the scaling of the uncertainties of the fitted parameters as a function of the weights applied to the events.
In (*) are the results obtained using SumW2Error(true) and SumW2Error(false): left and right columns respectively. I’m applying uniform weights on all the events like 1/1000, 1/100, 1/10, 1, 10 and 100 going
from the first row to the last one.

As expected, when using SumW2Error(true) the uncertainties don’t depend on the weights… for weights > 1/10. However, they seem to be affected by the weights when the weights are smaller than 1/10.

Is this because of some numerical precision limit that I’m hitting? Or it could be due to a bug? or better to a mistake on my side? In case the results that I’m describing are not expected, I can provide more information to reproduce the problem.

Thanks,
Boris

(*)
mangano.web.cern.ch/mangano/dro … caling.pdf