TFumiliFCN.cxx

Go to the documentation of this file.

// @(#)root/minuit2:$Id: TFumiliFCN.cxx 29198 2009-06-24 13:05:43Z brun $
// Author: L. Moneta    10/2005  

/**********************************************************************
 *                                                                    *
 * Copyright (c) 2005 ROOT Foundation,  CERN/PH-SFT                   *
 *                                                                    *
 **********************************************************************/

#include "TFumiliFCN.h"
#include "TChi2FCN.h"
#include "TBinLikelihoodFCN.h"
#include "TChi2FitData.h"
#include "FitterUtil.h"

#include "TF1.h"
#include "TVirtualFitter.h"


//#define DEBUG

#ifdef DEBUG
#include <iostream>
#endif

#include <cmath>

static const double kPrecision = 1.E-16; 
static const double kEpsilon = 1.E-300; 




TFumiliFCN::TFumiliFCN( const TVirtualFitter & fitter, double up, int strategy, bool skipEmptyBins) : 
FumiliFCNBase()
{
   // constructor for a class with a FumiliFCNBase interface.
   // Need to use default constructor for FumiliFCNBase (don't know number of parameters at this stage)
   // Create also FitData class. 

   fUp = up;
   fFunc = dynamic_cast<TF1 *> ( fitter.GetUserFunc() );
   assert(fFunc != 0);
   // default skip empty bins
   fData = new TChi2FitData(fitter, skipEmptyBins); 
   //std::cout << "Created FitData with size = " << fData->Size() << std::endl;
   
   // need to set the size so ROOT can calculate ndf.
   fFunc->SetNumberFitPoints(fData->Size());
   
   fStrategy = strategy; 
   
   //std::cout << "created FumiliFCN  with dimension " << dimension() << std::endl;
}


TFumiliFCN::~TFumiliFCN() {  
   //  this class manages the fit data class. Delete it at the end

   if (fData) { 
      //std::cout << "deleting the data - size is " << fData->Size() << std::endl; 
      delete fData; 
   }
}


void TFumiliFCN::Initialize(unsigned int nPar) { 
   // need an initialize method with number of fit parameters to make space for cached 
   // quantities (gradient , etc..)

   fParamCache = std::vector<double>(nPar);
   fFunctionGradient = std::vector<double>( nPar );
   // call init function on FumiliFCN
   InitAndReset(nPar);
}


void  TFumiliFCN::EvaluateAll( const std::vector<double> & p) { 
   // evaluate at same time function, gradient and hessian for parameter p. FumiliFCN interface
   Calculate_gradient_and_hessian(p);
}


void TFumiliFCN::Calculate_gradient_and_hessian(const std::vector<double> & p)  {
   // evaluate at same time function, gradient and hessian for parameter p
      
   unsigned int npar = p.size(); 
   if (npar != Dimension() ) { 
      // re-initialize the store gradient
      //std::cout << "initialize FumiliFCN and cache" << std::endl;
      Initialize(npar);
   }
   
   const FumiliFitData & points = *fData; 
   
   
   // set parameters
   fFunc->SetParameters( &p.front() );
   fParamCache = p;
   
   std::vector<double> & grad = Gradient(); 
   std::vector<double> & hess = Hessian(); 
   
   // dimension of hessian symmetric matrix  
   unsigned int nhdim = static_cast<int>( 0.5*npar*(npar + 1) ); 
   assert( npar == fFunctionGradient.size() ); 
   assert( npar == grad.size() ); 
   assert( nhdim == hess.size() );
   // reset  
   grad.assign( npar, 0.0); 
   hess.assign( nhdim, 0.0);
   
   double sum = 0; 
   
   
   // loop on measurements
   unsigned int nMeasurements = points.Size();
   unsigned int nRejected = 0; 
   for (unsigned int i = 0; i < nMeasurements; ++i) {
      
      fFunc->RejectPoint(false); 
      
      const std::vector<double> & x =  points.Coords(i); 
      fFunc->InitArgs( &x.front(), &fParamCache.front() ); 
      
      // one should implement integral option (in TFumili is not correct)
      double fval; 
      if ( fData->UseIntegral()) {
         const std::vector<double> & x2 = fData->Coords(i+1); 
         // need to implement derivatives of integral
         fval = FitterUtil::EvalIntegral(fFunc,x,x2,fParamCache); 
         if (fFunc->RejectedPoint() ) { 
            nRejected++; 
            continue;
         }
         Calculate_numerical_gradient_of_integral( x, x2, fval);        
         
      }
      else { 
         
         fval = fFunc->EvalPar(&x.front(), &fParamCache.front() ); 
         if (fFunc->RejectedPoint() ) { 
            nRejected++; 
            continue;
         }
         Calculate_numerical_gradient( x, fval); 
      }
      
      // calculate gradient 
      // eventually use function if provides it 
      
      Calculate_element(i, points, fval, sum, grad, hess); 
      
      // calculate i -element contribution to the chi2
      // add contributions to previous one 
      
      
   }
   
#ifdef DEBUG
   std::cout << "Calculated Gradient and hessian " << std::endl; 
   for (unsigned int i = 0; i < npar; ++i) 
      std::cout << " par " << i << " = " << fParamCache[i] << " grad = " << grad[i] << std::endl;
   for (unsigned int i = 0; i < npar; ++i) {
      for (unsigned int j = 0; j < npar; ++j) 
         std::cout << hess[i*npar+j]; 
      
      std::cout << std::endl; 
   }
#endif
   
   // set value of Obj function to be used by Minuit
   SetFCNValue(sum);
   
   // reset the number of fitting data points
   if (nRejected != 0)  fFunc->SetNumberFitPoints(nMeasurements-nRejected);
   
        
}


void TFumiliFCN::Calculate_numerical_gradient( const std::vector<double> & x, double f0) {   
   // calculate the numerical gradient for f(f). 
   // Different method (5 point r 2 points) according to the strategy set
   
   // use a cache for parameters to avoid to copy parameters each time this function is called
   
   int n = fParamCache.size();
   //std::cout << "Model function Gradient " << std::endl;
   for (int ipar = 0; ipar < n ; ++ipar) { 
      double p0 = fParamCache[ipar]; 
      // use 0.001 of par 
      double h = std::max( 0.001* std::fabs(p0), 8.0*kPrecision*(std::fabs(p0) + kPrecision) );  
      fParamCache[ipar] = p0 + h;  
      double f2 =  fFunc->EvalPar( &x.front(), &fParamCache.front() ); 
      
      
      if (fStrategy == 2) { 
         //  USE 5 POINT_RULE
         fParamCache[ipar] = p0 - h;  
         double f1 =  fFunc->EvalPar( &x.front(), &fParamCache.front() ); 
         fParamCache[ipar] = p0 + h/2; 
         double g1 =  fFunc->EvalPar( &x.front(), &fParamCache.front() ); 
         fParamCache[ipar] = p0 - h/2; 
         double g2 =  fFunc->EvalPar( &x.front(), &fParamCache.front() ); 
         
         double h2    = 1/(2.*h);
         double d0    = f1 - f2;
         double d2    = 2*(g1 - g2);
         
         //fFunctionGradient[ipar] =  0.5*( f2 - f1)/h;
         fFunctionGradient[ipar] =  h2*(4*d2 - d0)/3.;
         
      }
      else { 
         //default 2  point rule
         fFunctionGradient[ipar] =  (f2 - f0)/h;
      }
      
      // reset to old value
      fParamCache[ipar] = p0; 
      //std::cout << " i " << ipar << par[ipar] << "  " << fFunctionGradient[ipar] << " xi = " << x[0] << " fval " << f0 << std::endl; 
   }
   
}


void TFumiliFCN::Calculate_numerical_gradient_of_integral( const std::vector<double> & x1, const std::vector<double> & x2, double f0) {   
   // calculate the numerical gradient when the integral of the model function is used in the fit
    // Different method (5 point r 2 points) according to the strategy set
   // use a cache for parameters to avoid to copy parameters each time this function is called
   
   int n = fParamCache.size();
   //std::cout << "Model function Gradient " << std::endl;
   for (int ipar = 0; ipar < n ; ++ipar) { 
      double p0 = fParamCache[ipar]; 
      // use 0.001 of par 
      double h = std::max( 0.001* std::fabs(p0), 8.0*kPrecision*(std::fabs(p0) + kPrecision) );  
      fParamCache[ipar] = p0 + h;  
      double f2 =   FitterUtil::EvalIntegral(fFunc,x1,x2,fParamCache);
      
      
      if (fStrategy == 2) { 
         //  USE 5 POINT_RULE
         fParamCache[ipar] = p0 - h;  
         double f1 =  FitterUtil::EvalIntegral(fFunc,x1,x2,fParamCache);
         fParamCache[ipar] = p0 + h/2; 
         double g1 =  FitterUtil::EvalIntegral(fFunc,x1,x2,fParamCache);
         fParamCache[ipar] = p0 - h/2; 
         double g2 =  FitterUtil::EvalIntegral(fFunc,x1,x2,fParamCache); 
         
         double h2    = 1/(2.*h);
         double d0    = f1 - f2;
         double d2    = 2*(g1 - g2);
         
         //fFunctionGradient[ipar] =  0.5*( f2 - f1)/h;
         fFunctionGradient[ipar] =  h2*(4*d2 - d0)/3.;
         
      }
      else { 
         //default 2  point rule
         fFunctionGradient[ipar] =  (f2 - f0)/h;
      }
      
      // reset to old value
      fParamCache[ipar] = p0; 
      //std::cout << " i " << ipar << par[ipar] << "  " << fFunctionGradient[ipar] << " xi = " << x[0] << " fval " << f0 << std::endl; 
   }
   
}



void TFumiliChi2FCN::Calculate_element(int i, const FumiliFitData & points, double fval, double & chi2, std::vector<double> & grad,   std::vector<double> & hess ) {
   // calculate the element i : grad(i) and hessian(i) for a chi2 FCN
   
   double invError =  points.InvError(i); 
   double value = points.Value(i);
   double element = invError*( fval - value );
   unsigned int npar = grad.size();
   
   chi2 += element*element;
   
   for (unsigned int j = 0; j < npar; ++j) { 
      
      double fj =  invError * fFunctionGradient[j]; 
      grad[j] += 2.0 * element * fj; 
      
      //std::cout << " ---------j " << " j  " <<  fFunctionGradient[j] << "  " << grad[j] << std::endl; 
      
      for (unsigned int k = j; k < npar; ++ k) { 
         int idx =  j + k*(k+1)/2; 
         hess[idx] += 2.0 * fj * invError * fFunctionGradient[k]; 
      }
   }
   //      std::cout << "element " << i << " x " << x[0] << " val = " << value << " sig = " << 1/invError << " f(x) " << fval << " element " << element << " gradients:  " << grad[0] << "  " << grad[1] << "  " << grad[2] << std::endl; 
   
}


void TFumiliBinLikelihoodFCN::Calculate_element(int i, const FumiliFitData & points, double fval, double & logLike, std::vector<double> & grad,   std::vector<double> & hess ) {
   // calculate the element i : grad(i) and hessian(i) for a binned log-likelihood FCN
   
   unsigned int npar = grad.size();
   
   // kEpsilon is smalles number ( 10-300) 
   double logtmp, invFval; 
   if(fval<=kEpsilon) { 
      logtmp = fval/kEpsilon + std::log(kEpsilon) - 1; 
      invFval = 1.0/kEpsilon; 
   } else {        
      logtmp = std::log(fval);
      invFval = 1.0/fval;
   }
   
   double value =  points.Value(i); 
   logLike +=  2.*( fval - value*logtmp );
   
   
   for (unsigned int j = 0; j < npar; ++j) {
      
      double fj; 
      if ( fval < kPrecision &&  std::fabs(fFunctionGradient[j]) < kPrecision ) 
         fj = 2.0; 
      else 
         fj =  2.* fFunctionGradient[j] * ( 1.0 - value*invFval); 
      
      
      //            if ( ( ! (fj <= 0) )  && ( ! ( fj > 0) ) ) { 
      //              std::cout << "fj is nan -- " << fj << "  " << j << " x " << x[0] << " f(x) = " << fval << "  inv =  " << invFval << "gradient = " 
      //                        << fFunctionGradient[j] << "  " << fFunctionGradient[j]/fval << std::endl;
      //              fj = 0; 
      
      //            }
      
      grad[j] += fj;
      
      for (unsigned int k = j; k < npar; ++ k) { 
         int idx =  j + k*(k+1)/2; 
         double fk; 
         if ( fval < kPrecision &&  std::fabs(fFunctionGradient[k]) < kPrecision ) 
            fk = 1.0; 
         else 
            fk =  fFunctionGradient[k]* ( 1.0 - value*invFval); 
         
         
         hess[idx] += fj * fk;
      }
   }
   
}


void TFumiliUnbinLikelihoodFCN::Calculate_element(int , const FumiliFitData &, double fval, double & logLike, std::vector<double> & grad,   std::vector<double> & hess ) {
   // calculate the element i : grad(i) and hessian(i) for an unbinned log-likelihood FCN
   
   unsigned int npar = grad.size();
   // likelihood
   //if (fval < 1.0E-16) fval = 1.0E-16; // truncate for precision
   
   // kEpsilon is smalles number ( 10-300) 
   double logtmp, invFval; 
   if(fval<=kEpsilon) { 
      logtmp = fval/kEpsilon + std::log(kEpsilon) - 1; 
      invFval = 1.0/kEpsilon; 
   } else {        
      logtmp = std::log(fval);
      invFval = 1.0/fval;
   }
   
   logLike += logtmp;
   for (unsigned int j = 0; j < npar; ++j) {
      
      double fj; 
      if ( fval < kPrecision &&  std::fabs(fFunctionGradient[j]) < kPrecision ) 
         fj = 2.0; 
      else 
         fj =  2.* invFval * fFunctionGradient[j]; 
        
      grad[j] -= fj;
      
      for (unsigned int k = j; k < npar; ++ k) { 
         int idx =  j + k*(k+1)/2; 
         double fk; 
         if ( fval < kPrecision &&  std::fabs(fFunctionGradient[k]) < kPrecision ) 
            fk = 1.0; 
         else 
            fk =  invFval * fFunctionGradient[k]; 
         
         
         hess[idx] += fj * fk;
      }
   }
}


//  
double TFumiliChi2FCN::operator()(const std::vector<double>& par) const {
   // implement chi2 function 
   assert(fData != 0); 
   assert(fFunc != 0); 
   
   TChi2FCN  fcn(fData,fFunc); 
   return fcn(par); 
}


double TFumiliBinLikelihoodFCN::operator()(const std::vector<double>& par) const {
    // implement binned-likelihood function 
   assert(fData != 0); 
   assert(fFunc != 0); 
   
   TBinLikelihoodFCN  fcn(fData,fFunc); 
   return fcn(par); 
}

double TFumiliBinLikelihoodFCN::Chi2(const std::vector<double>& par) const {
   // implement function to evaluate chi2 equivalent
   TChi2FCN chi2Fcn(fData,fFunc);
   return chi2Fcn(par);
}


double TFumiliUnbinLikelihoodFCN::operator()(const std::vector<double>& /*par */) const {
   // not yet implemented (to be done)
   assert(fData != 0); 
   assert(fFunc != 0); 
   
   //TUnbinLikelihoodFCN  fcn(*fData,*fFunc); 
   //return fcn(par);
   // to be implemented
   return 0; 
}

---