Fitting problem with TGraph and TF1 in 5.27

Roofit and RooStats
1

Hi,

I posted this originally in the PyRoot section http://root.cern.ch/phpBB3//viewtopic.php?f=14&t=12701, but have since verified that I see the same thing in Cint as well.

{
   const Double_t yVal[3] = {0.95, 1, 1.1};
   const Double_t xVal[3] = {-1, 0, 1};
   TCanvas *canvas = new TCanvas;
   TGraph *graph = new TGraph ( 3, xVal, yVal);
   graph->SetMarkerStyle (20);
   graph->SetMarkerColor (kRed);
   graph->SetMarkerSize  (1.4);
   graph->SetMaximum (1.3);
   graph->SetMinimum (0.7);
   TF1 *fun = new TF1("fun", "pol2");
   fun->FixParameter(0, 1);
   graph->Fit (fun);
   graph->Draw ("AP");
   canvas->Print ("fit.png");
}

Produces a good fit in 5.22:

[code]root [0]
Processing fitSTC.C…

Minimizer is Linear
Chi2 = 6.16298e-32
NDf = 0
Par_0 = 1 (fixed)
Par_1 = 0.075 +/- 0.707107
Par_2 = 0.025 +/- 0.707107
Info in TCanvas::Print: file fit.png has been created
root [1] .q
[/code]

But messes the fit up in 5.27.06b

Processing fitSTC.C...

****************************************
Minimizer is Linear
Chi2                      =            2
NDf                       =            0
p0                        =            1                         (fixed)
p1                        =        0.075   +/-   0.707107    
p2                        =        1.025   +/-   0.707107    
Info in <TCanvas::Print>: file fit.png has been created

One thing that is peculiar is that the fit parameter p1 (or par_1) has the same value, but p2/par_2 value differ by exactly 1.

Cheers,
Charles

2

One addition point (for the poor souls who are debugging this):

If I change “pol2” to “pol1”, the fit now works as expected in both 5.27 as well as 5.22.

3

Hi Charles,

Maybe this is something quite obscure, but when fixing or setting bounds on predefined functions, like
"polN", “gaus” , “landau”, etc…, one needs to use the fit option “B”

Cheers, Lorenzo

4

Hi Lorenzo,

[quote=“moneta”]Maybe this is something quite obscure, but when fixing or setting bounds on predefined functions, like
"polN", “gaus” , “landau”, etc…, one needs to use the fit option “B”
[/quote]

Thanks! It is, well, obscure. :slight_smile: Why did it work before? Is this really only necessary for built-in functions? (Even if I coded up the same function by hand?)

Thanks,
Charles

5

Hi Charles,

I was too fast before. Looking more deeply in your example, there is actually a bug (versions >= 5.26) when doing linear fits with fixed parameters.
As a workaround, you can use option F (or option B) which switch automatically to use Minuit instead of the linear fitter.
The bug is now fixed in the current trunk. Thanks a lot for reporting this problem

Best Regards

Lorenzo