it seems that using TFractionFitter the errors on the weights of the MC
template histograms are not taken into account in computing the chi2.
Infact I tried to set by hand the weights errors to zero and the chi2 of the fit
does not change. Am I missing something?
Thanks for your help,
the uncertainty in the weights are not taken into account in the fit procedure. More information is provided in the original paper of R. Barlow,
Computer Physics Communication 77 (1993), 219.
See the paragraph 6 (Weighted events)
I remind you that this method provides also a biased error, which can be relevant in some case (.See Assumptions in the TFractionFitter documentation and arXiv:0803.2711).
A full extended maximum likelihood fit is then maybe the right choice in this case. You should be able to do that using RooFit