Gradient calculation in GSLMultiFit

ROOT
1

I think I mentioned it a long time ago (although I am not sure), but perhaps now it could be fixed.

It seems that GSLMultiFit during the gradient calculation procedure before starting first iteration of the fit, calculates gradient for all values - even those fixed. No such thing happens for Minuit. I think this could mean a large efficiency loss when performing many small, quick separate fits for functions with lot’s of fixed parameters.

2

Hi,

are you sure this is happening with the current trunk revision or the latest release, 5.27.06 ?
Can you reproduce it with a small example ?

Lorenzo

3

My ROOT displays: 5.27/07 :slight_smile:

I’ll try to make an example.

4

Seems that the problem exists only with fitting via ROOT::Math etc. interfaces, not via TH::Fit(). I modified the script for the simulatnous fitting of two functions. It displays the changes of the fixed parameter before the first fit iteration.
combinedFit2.C (5.69 KB)

5

HI,
the problem was in the macro that I have attached in the post

The global chi2 function needs to be modified to not force the gradient calculation of all parameters. See the changes in the new macro which I attach to this post

Best Regards

Lorenzo
combinedFit2.C (5.99 KB)

6

Is this example in general is easily changeable to a 2D case? I thought yes, but it still counts gradient for me in the 2D case after modification. Perhaps a bug in my copy&paste or perhaps something more general.

To change it into a 2D case, I simply changed the wrapper functions to 2D:

		ROOT::Math::WrappedMultiTF1 wfRed(*psf_fin_red,2);
		ROOT::Math::WrappedMultiTF1 wfBlue(*psf_fin_blue,2);

where psf_fin_* are TF2

and data ranges to twodimensional:

		rangeRed.SetRange(-5.5, 5.5, -5.5, 5.5);

Ofcourse I fit to TH2Ds. Should any additional changes be made to the GlobalChi2 function or anything else? Sorry, my own code is two long and complicated to attach it here…

7

Yes , these changes should be enough !

Lorenzo

8

Found a bug in my copy& paste. Seems that this change not only skips unnecessary gradient calculations at the beginning, but also some calculations during normal iterations - normal iterations are now, I guess, a few dozen times faster :slight_smile: