How to implement custom gradient calculation procedure in TVirtualFitter?

[terawatt]

Hello,
I cannot find any tutorials with custom (provided by user) gradient calculation procedure for Minuit or Minuit2. When I try to make very simple test procedure like

void FCN(Int_t&npar, Double_t*gin, Double_t&f, Double_t*par, Int_t flag)
{// select case using flag
	switch(flag)
	{
		case 1: 
			// Initialization
			cout<<"case1\n";
		break;
		case 2:
			cout<<"case2\n";
			gin[0]=2*par[0]-8;
			// Compute derivatives   
			// store them in gin
		break;
		case 3:    
			cout<<"case3\n";
		break;
		default:
		cout<<"flag:"<<flag<<"\n";
		f=pow(par[0],2)-8*par[0]+4;
		cout<<f<<"\n";
		break;
	}
}
void test()
{
        TVirtualFitter::SetDefaultFitter("Minuit");
	TVirtualFitter * minuit = TVirtualFitter::Fitter(0,1);
	minuit->SetParameter(0,"x",0,0.01,0,0);
	minuit->SetFCN(FCN);
	
	double arglist[100];
	arglist[0] = 0;
	// set print level
	minuit->ExecuteCommand("SET PRINT",arglist,2);

	// minimize
	arglist[0] = 5000; // number of function calls
	arglist[1] = 0.01; // tolerance
	minuit->ExecuteCommand("MIGRAD",arglist,2);
}

Minuit always sets flag=4, so custom derivative procedure never used.

_ROOT Version:_6.20/04
Platform: x86-x64
_Compiler:_g++

---

[jblomer]

@moneta Do you have a suggestion?

[terawatt]

I think I has found a solution for Minuit, an additional command is needed:

double GrPar[]={1};
minuit->ExecuteCommand("SET GRA",GrPar,1);

where value in the GrPar is defined according to conditions:
If GrPar[0]==0, Minuit will check the user-defined gradient by numerical calculation before use it;
if GrPar[0]==1, Minuit will use the user-defined derivatives without any checks.

Anyway, it would be interesting how to do similar trick with other minimizers.

[moneta]

Hi,
For minimizers that supports it, you can provide custom gradient calculation when fitting. This is the case for example when Automatic differentiation is used when fitting an histogram with a function expressed by a formula.
The ROOT::Math::Minimizer interface supports it, by simply providing the right function interface to the minimizer. You need to call the right SetFunction method, see
ROOT: ROOT::Math::Minimizer Class Reference.

Lorenzo

[terawatt]

Thank you!