TGraph fit errors: 5.20 vs. 5.22

Hi all,

I noticed in 5.22 the reported errors on the fit parameters have changed with respect to 5.20. Root used to build a Chi2 function assuming 1 as the errors, pass it to MINUIT, and then scale the errors so that one gets Chi2/ndof=1.

In 5.22 it does scale the errors but not using the formula it used to be, namely sigma = sigma_fit * sqrt(chi2/(ndof-1)).

Before I submit a bug report, does anybody have a good explanation of why this was changed?

  • Moritz

Hi,

this is a bug in 5.22 which has been fixed in 5.23.01 (rev. >= 27613).
Bug report is savannah.cern.ch/bugs/?46006

However, the scale of the error is done by using sqrt(chi2/(ndf)).
and not ndf-1. I don’t see why one should use ndf-1 and excluding the case
when ndf=1.
This is the formula reported also at
en.wikipedia.org/wiki/Least_squa … statistics

Best Regards

Lorenzo