I am running into a problem when fitting in TGraphErrors.
My understanding from the TGraph:Fit document is that TGraph simply uses the un-weighted least-square fit, no uncertainty is considered at all. TGraphErrors uses the uncertainties as the weights, but if option “W” is provided, it will ignore the uncertainties and use the un-weighted least-square fit as well.
I tried both TGraph and TGraphErrors fitting in both ways, the FitResults are the same and the fitting looks great. However, if I include uncertainties in TGraphErrors, the fit results are different, and in my case, the fitting is much worse.
I am curious about:
- How TGraph and TGraphErrors determine the un-weighted least square fit uncertainty. (If it’s a histogram then it is the sqrt of y(i), but this is not the case in TGraph?)
- How to force the fitting to be better when uncertainties are considered? I tried setting bounds on the parameters but no luck so far.
A few plots are below.