MuScleFitUtils.cc

Go to the documentation of this file.

// Some notes:
// - M(Z) after simulation needs to be extracted as a function of |y_Z| in order to be
//   a better reference point for calibration. In fact, the variation of PDF with y_Z
//   in production is sizable <---- need to check though.
// - ResHalfWidth needs to be optimized - this depends on the level of background.
// - Background parametrization still to be worked on, so far only a constant (type=1, and
//   parameter 2 fixed to 0) works.
// - weights have to be assigned to dimuon mass values in regions where different resonances
//   overlap, and one has to decide which resonance mass to assign the event to - this until
//   we implement in the fitter a sum of probabilities of an event to belong to different
//   resonances. The weight has to depend on the mass and has relative cross sections of
//   Y(1S), 2S, 3S as parameters. Some overlap is also expected in the J/psi-Psi(2S) region
//   when reconstructing masses with Standalone muons.
//
//   MODS 7/7/08 TD:
//   - changed parametrization of resolution in Pt: from sigma_pt = a*Pt + b*|eta| to
//                                                       sigma_pt = (a*Pt + b*|eta|)*Pt
//                                                  which is more correct (I hope)
//   - changed parametrization of resolution in cotgth: from sigma_cotgth = f(eta) to f(cotgth)
// --------------------------------------------------------------------------------------------

#include "MuScleFitUtils.h"
#include "DataFormats/HepMCCandidate/interface/GenParticle.h"
#include "SimDataFormats/Track/interface/SimTrack.h"
#include "DataFormats/Candidate/interface/LeafCandidate.h"
#include "DataFormats/Math/interface/LorentzVector.h"
#include "TString.h"
#include "TFile.h"
#include "TTree.h"
#include "TCanvas.h"
#include "TH2F.h"
#include "TF1.h"
#include "TF2.h"
#include <iostream>
#include <fstream>
#include <memory> // to use the auto_ptr

// Includes the definitions of all the bias and scale functions
// These functions are selected in the constructor according
// to the input parameters.
#include "MuonAnalysis/MomentumScaleCalibration/interface/Functions.h"

// To use callgrind for code profiling uncomment also the following define.
//#define USE_CALLGRIND
#ifdef USE_CALLGRIND
#include "valgrind/callgrind.h"
#endif

// Lorenzian Peak function
// -----------------------
Double_t lorentzianPeak (Double_t *x, Double_t *par) {
  return (0.5*par[0]*par[1]/TMath::Pi()) /
  TMath::Max(1.e-10,(x[0]-par[2])*(x[0]-par[2]) + .25*par[1]*par[1]);
}

// Gaussian function
// -----------------
Double_t Gaussian (Double_t *x, Double_t *par) {
  return par[0]*exp(-0.5*((x[0]-par[1])/par[2])*((x[0]-par[1])/par[2]));
}

// Array with number of parameters in the fitting functions
// (not currently in use)
// --------------------------------------------------------
//const int nparsResol[2] = {6, 4};
//const int nparsScale[13] = {2, 2, 2, 3, 3, 3, 4, 4, 2, 3, 4, 6, 8};
//const int nparsBgr[3] = {1, 2, 3};

// Quantities used for h-value computation
// ---------------------------------------
double mzsum;
double isum;
double f[11][100];
double g[11][100];

// Lorentzian convoluted with a gaussian:
// --------------------------------------
TF1 * GL = new TF1 ("GL",
            "0.5/3.1415926*[0]/(pow(x-[1],2)+pow(0.5*[0],2))*exp(-0.5*pow((x-[2])/[3],2))/([3]*sqrt(6.283185))",
            0, 1000);

TF2 * GL2= new TF2 ("GL2",
  "0.5/3.1415926*[0]/(pow(x-[1],2)+pow(0.5*[0],2))*exp(-0.5*pow((x-y)/[2],2))/([2]*sqrt(6.283185))",
  0, 200, 0, 200);

// // Lorentzian convoluted with a gaussian over a linear background:
// // ---------------------------------------------------------------
// TF1 * GLBL = new TF1 ("GLBL",
//   "0.5/3.1415926*[0]/(pow(x-[1],2)+pow(0.5*[0],2))*exp(-0.5*pow((x-[2])/[3],2))/([3]*sqrt(6.283185))+[4]+[5]*x",
//   0, 1000);

// // Lorentzian convoluted with a gaussian over an exponential background:
// // ---------------------------------------------------------------
// TF1 * GLBE = new TF1 ("GLBE",
//   "0.5/3.1415926*[0]/(pow(x-[1],2)+pow(0.5*[0],2))*exp(-0.5*pow((x-[2])/[3],2))/([3]*sqrt(6.283185))+exp([4]+[5]*x)",
//   0, 1000);

std::vector<int> MuScleFitUtils::doResolFit;
std::vector<int> MuScleFitUtils::doScaleFit;
std::vector<int> MuScleFitUtils::doCrossSectionFit;
std::vector<int> MuScleFitUtils::doBackgroundFit;

int MuScleFitUtils::minuitLoop_ = 0;
TH1D* MuScleFitUtils::likelihoodInLoop_ = 0;
TH1D* MuScleFitUtils::signalProb_ = 0;
TH1D* MuScleFitUtils::backgroundProb_ = 0;

bool MuScleFitUtils::duringMinos_ = false;

const int MuScleFitUtils::totalResNum = 6;

int MuScleFitUtils::SmearType = 0;
smearFunctionBase * MuScleFitUtils::smearFunction = 0;
int MuScleFitUtils::BiasType  = 0;
// No error, we take functions from the same group for bias and scale.
scaleFunctionBase<std::vector<double> > * MuScleFitUtils::biasFunction = 0;
int MuScleFitUtils::ResolFitType = 0;
resolutionFunctionBase<double *> * MuScleFitUtils::resolutionFunction = 0;
resolutionFunctionBase<std::vector<double> > * MuScleFitUtils::resolutionFunctionForVec = 0;
int MuScleFitUtils::ScaleFitType = 0;
scaleFunctionBase<double*> * MuScleFitUtils::scaleFunction = 0;
scaleFunctionBase<std::vector<double> > * MuScleFitUtils::scaleFunctionForVec = 0;
int MuScleFitUtils::BgrFitType   = 0;

CrossSectionHandler * MuScleFitUtils::crossSectionHandler;
// const int MuScleFitUtils::backgroundFunctionsRegions = 3;
// backgroundFunctionBase * MuScleFitUtils::backgroundFunctionForRegion[MuScleFitUtils::backgroundFunctionsRegions];
// backgroundFunctionBase * MuScleFitUtils::backgroundFunction[MuScleFitUtils::totalResNum];
BackgroundHandler * MuScleFitUtils::backgroundHandler;
std::vector<double> MuScleFitUtils::parBias;
std::vector<double> MuScleFitUtils::parSmear;
std::vector<double> MuScleFitUtils::parResol;
std::vector<double> MuScleFitUtils::parResolStep;
std::vector<double> MuScleFitUtils::parResolMin;
std::vector<double> MuScleFitUtils::parResolMax;
std::vector<double> MuScleFitUtils::parScale;
std::vector<double> MuScleFitUtils::parScaleStep;
std::vector<double> MuScleFitUtils::parScaleMin;
std::vector<double> MuScleFitUtils::parScaleMax;
std::vector<double> MuScleFitUtils::parCrossSection;
std::vector<double> MuScleFitUtils::parBgr;
std::vector<int> MuScleFitUtils::parResolFix;
std::vector<int> MuScleFitUtils::parScaleFix;
std::vector<int> MuScleFitUtils::parCrossSectionFix;
std::vector<int> MuScleFitUtils::parBgrFix;
std::vector<int> MuScleFitUtils::parResolOrder;
std::vector<int> MuScleFitUtils::parScaleOrder;
std::vector<int> MuScleFitUtils::parCrossSectionOrder;
std::vector<int> MuScleFitUtils::parBgrOrder;

std::vector<int> MuScleFitUtils::resfind;
int MuScleFitUtils::debug = 0;

bool MuScleFitUtils::ResFound = false;
int MuScleFitUtils::goodmuon = 0;
int MuScleFitUtils::counter_resprob = 0;

std::vector<std::vector<double> > MuScleFitUtils::parvalue;

int MuScleFitUtils::FitStrategy = 1; // Strategy in likelihood fit (1 or 2)
bool MuScleFitUtils::speedup = false; // Whether to cut corners (no sim study, fewer histos)

std::vector<std::pair<lorentzVector,lorentzVector> > MuScleFitUtils::SavedPair; // Pairs of reconstructed muons making resonances
std::vector<std::pair<MuScleFitMuon,MuScleFitMuon> > MuScleFitUtils::SavedPairMuScleFitMuons; // Pairs of reconstructed muons making resonances
std::vector<std::pair<lorentzVector,lorentzVector> > MuScleFitUtils::ReducedSavedPair; // Pairs of reconstructed muons making resonances inside smaller windows
std::vector<std::pair<lorentzVector,lorentzVector> > MuScleFitUtils::genPair; // Pairs of generated muons making resonances
//std::vector<GenMuonPair> MuScleFitUtils::genPairMuScleMuons; // Pairs of generated muons making resonances
std::vector<std::pair<lorentzVector,lorentzVector> > MuScleFitUtils::simPair; // Pairs of simulated muons making resonances

// Smearing parameters
// -------------------
double MuScleFitUtils::x[][10000];

// Probability matrices and normalization values
// ---------------------------------------------
int MuScleFitUtils::nbins = 1000;
double MuScleFitUtils::GLZValue[][1001][1001];
double MuScleFitUtils::GLZNorm[][1001];
double MuScleFitUtils::GLValue[][1001][1001];
double MuScleFitUtils::GLNorm[][1001];
double MuScleFitUtils::ResMaxSigma[];

// Masses and widths from PDG 2006, half widths to be revised
// NB in particular, halfwidths have to be made a function of muonType
// -------------------------------------------------------------------
const double MuScleFitUtils::mMu2 = 0.011163612;
const double MuScleFitUtils::muMass = 0.105658;
double MuScleFitUtils::ResHalfWidth[];
double MuScleFitUtils::massWindowHalfWidth[][6];
int MuScleFitUtils::MuonType;
int MuScleFitUtils::MuonTypeForCheckMassWindow;

double MuScleFitUtils::ResGamma[] = {2.4952, 0.000020, 0.000032, 0.000054, 0.000317, 0.0000932 };
// ATTENTION:
// This is left because the values are used by the BackgroundHandler to define the center of the regions windows,
// but the values used in the code are read computed using the probability histograms ranges.
// The histograms are read after the initialization of the BackgroundHandler (this can be improved so that
// the background handler too could use the new values).
// At this time the values are consistent.
double MuScleFitUtils::ResMinMass[] = {-99, -99, -99, -99, -99, -99};
double MuScleFitUtils::ResMass[] = {91.1876, 10.3552, 10.0233, 9.4603, 3.68609, 3.0969};
// From Summer08 generator production TWiki: https://twiki.cern.ch/twiki/bin/view/CMS/ProductionSummer2008
// - Z->mumu              1.233 nb
// - Upsilon3S->mumu      0.82  nb
// - Upsilon2S->mumu      6.33  nb
// - Upsilon1S->mumu     13.9   nb
// - Prompt Psi2S->mumu   2.169 nb
// - Prompt J/Psi->mumu 127.2   nb
// double MuScleFitUtils::crossSection[] = {1.233, 0.82, 6.33, 13.9, 2.169, 127.2};
// double MuScleFitUtils::crossSection[] = {1.233, 2.07, 6.33, 13.9, 2.169, 127.2};

unsigned int MuScleFitUtils::loopCounter = 5;

// According to the pythia manual, there is only a code for the Upsilon and Upsilon'. It does not distinguish
// between Upsilon(2S) and Upsilon(3S)
const unsigned int MuScleFitUtils::motherPdgIdArray[] = {23, 100553, 100553, 553, 100443, 443};

// double MuScleFitUtils::leftWindowFactor = 1.;
// double MuScleFitUtils::rightWindowFactor = 1.;

// double MuScleFitUtils::internalLeftWindowFactor = 1.;
// double MuScleFitUtils::internalRightWindowFactor = 1.;

// int MuScleFitUtils::backgroundWindowEvents_ = 0;
// int MuScleFitUtils::resonanceWindowEvents_ = 0;

// double MuScleFitUtils::oldEventsOutInRatio_ = 0.;

bool MuScleFitUtils::scaleFitNotDone_ = true;

bool MuScleFitUtils::sherpa_ = false;

bool MuScleFitUtils::useProbsFile_ = true;

bool MuScleFitUtils::rapidityBinsForZ_ = true;

bool MuScleFitUtils::separateRanges_ = true;
double MuScleFitUtils::minMuonPt_ = 0.;
double MuScleFitUtils::maxMuonPt_ = 100000000.;
double MuScleFitUtils::minMuonEtaFirstRange_ = -6.;
double MuScleFitUtils::maxMuonEtaFirstRange_ = 6.;
double MuScleFitUtils::minMuonEtaSecondRange_ = -100.;
double MuScleFitUtils::maxMuonEtaSecondRange_ = 100.;
double MuScleFitUtils::deltaPhiMinCut_ = -100.;
double MuScleFitUtils::deltaPhiMaxCut_ = 100.;

bool MuScleFitUtils::debugMassResol_;
MuScleFitUtils::massResolComponentsStruct MuScleFitUtils::massResolComponents;

bool MuScleFitUtils::normalizeLikelihoodByEventNumber_ = true;
TMinuit * MuScleFitUtils::rminPtr_ = 0;
double MuScleFitUtils::oldNormalization_ = 0.;
unsigned int MuScleFitUtils::normalizationChanged_ = 0;

bool MuScleFitUtils::startWithSimplex_;
bool MuScleFitUtils::computeMinosErrors_;
bool MuScleFitUtils::minimumShapePlots_;

int MuScleFitUtils::iev_ = 0;

// Find the best simulated resonance from a vector of simulated muons (SimTracks)
// and return its decay muons
// ------------------------------------------------------------------------------
std::pair<SimTrack,SimTrack> MuScleFitUtils::findBestSimuRes (const std::vector<SimTrack>& simMuons) {

  std::pair<SimTrack, SimTrack> simMuFromBestRes;
  double maxprob = -0.1;

  // Double loop on muons
  // --------------------
  for (std::vector<SimTrack>::const_iterator simMu1=simMuons.begin(); simMu1!=simMuons.end(); simMu1++) {
    for (std::vector<SimTrack>::const_iterator simMu2=simMu1+1; simMu2!=simMuons.end(); simMu2++) {
      if (((*simMu1).charge()*(*simMu2).charge())>0) {
    continue; // this also gets rid of simMu1==simMu2...
      }
      // Choose the best resonance using its mass. Check Z, Y(3S,2S,1S), Psi(2S), J/Psi in order
      // ---------------------------------------------------------------------------------------
      double mcomb = ((*simMu1).momentum()+(*simMu2).momentum()).mass();
      double Y = ((*simMu1).momentum()+(*simMu2).momentum()).Rapidity();
      for (int ires=0; ires<6; ires++) {
    if (resfind[ires]>0) {
      double prob = massProb( mcomb, Y, ires, 0. );
      if (prob>maxprob) {
        simMuFromBestRes.first = (*simMu1);
        simMuFromBestRes.second = (*simMu2);
        maxprob = prob;
      }
    }
      }
    }
  }

  // Return most likely combination of muons making a resonance
  // ----------------------------------------------------------
  return simMuFromBestRes;
}

// Find the best reconstructed resonance from a collection of reconstructed muons
// (MuonCollection) and return its decay muons
// ------------------------------------------------------------------------------
std::pair<MuScleFitMuon,MuScleFitMuon> MuScleFitUtils::findBestRecoRes( const std::vector<MuScleFitMuon>& muons ){
  // NB this routine returns the resonance, but it also sets the ResFound flag, which
  // is used in MuScleFit to decide whether to use the event or not.
  // --------------------------------------------------------------------------------
  if (debug>0) std::cout << "In findBestRecoRes" << std::endl;
  ResFound = false;
  std::pair<MuScleFitMuon, MuScleFitMuon> recMuFromBestRes;

  // Choose the best resonance using its mass probability
  // ----------------------------------------------------
  double maxprob = -0.1;
  double minDeltaMass = 999999;
  std::pair<MuScleFitMuon,MuScleFitMuon> bestMassMuons;
  for (std::vector<MuScleFitMuon>::const_iterator Muon1=muons.begin(); Muon1!=muons.end(); ++Muon1) {
    //rc2010
    if (debug>0) std::cout << "muon_1_charge:"<<(*Muon1).charge() << std::endl;
    for (std::vector<MuScleFitMuon>::const_iterator Muon2=Muon1+1; Muon2!=muons.end(); ++Muon2) {
   //rc2010
      if (debug>0) std::cout << "after_2" << std::endl;
      if ((((*Muon1).charge())*((*Muon2).charge()))>0) {
    continue; // This also gets rid of Muon1==Muon2...
      }
      // To allow the selection of ranges at negative and positive eta independently we define two
      // ranges of eta: (minMuonEtaFirstRange_, maxMuonEtaFirstRange_) and (minMuonEtaSecondRange_, maxMuonEtaSecondRange_).
      // If the interval selected is simmetric, one only needs to specify the first range. The second has
      // default values that accept all muons (minMuonEtaSecondRange_ = -100., maxMuonEtaSecondRange_ = 100.).
      double pt1 = (*Muon1).Pt();
      double pt2 = (*Muon2).Pt();
      double eta1 = (*Muon1).Eta();
      double eta2 = (*Muon2).Eta();
      if( pt1 >= minMuonPt_ && pt1 < maxMuonPt_ &&
      pt2 >= minMuonPt_ && pt2 < maxMuonPt_ &&
      ( (eta1 >= minMuonEtaFirstRange_ && eta1 < maxMuonEtaFirstRange_ &&
         eta2 >= minMuonEtaFirstRange_ && eta2 < maxMuonEtaFirstRange_) ||
        (eta1 >= minMuonEtaSecondRange_ && eta1 < maxMuonEtaSecondRange_ &&
         eta2 >= minMuonEtaSecondRange_ && eta2 < maxMuonEtaSecondRange_) ) ) {
        double mcomb = ((*Muon1).p4()+(*Muon2).p4()).mass();
    double Y = ((*Muon1).p4()+(*Muon2).p4()).Rapidity();
    if (debug>1) {
      std::cout<<"muon1 "<<(*Muon1).p4().Px()<<", "<<(*Muon1).p4().Py()<<", "<<(*Muon1).p4().Pz()<<", "<<(*Muon1).p4().E()<<", "<<(*Muon1).charge()<<std::endl;
      std::cout<<"muon2 "<<(*Muon2).p4().Px()<<", "<<(*Muon2).p4().Py()<<", "<<(*Muon2).p4().Pz()<<", "<<(*Muon2).p4().E()<<", "<<(*Muon2).charge()<<std::endl;
      std::cout<<"mcomb "<<mcomb<<std::endl;}
    double massResol = 0.;
    if( useProbsFile_ ) {
      massResol = massResolution ((*Muon1).p4(), (*Muon2).p4(), parResol);
    }
    double prob = 0;
    for( int ires=0; ires<6; ires++ ) {
      if( resfind[ires]>0 ) {
        if( useProbsFile_ ) {
          prob = massProb( mcomb, Y, ires, massResol );
        }
        if( prob>maxprob ) {
          if( (*Muon1).charge()<0 ) { // store first the mu minus and then the mu plus
        recMuFromBestRes.first = (*Muon1);
        recMuFromBestRes.second = (*Muon2);
          } else {
        recMuFromBestRes.first = (*Muon2);
        recMuFromBestRes.second = (*Muon1);
          }
          if (debug>0) std::cout << "muon_1_charge (after swapping):"<<recMuFromBestRes.first.charge() << std::endl;
          ResFound = true; // NNBB we accept "resonances" even outside mass bounds
          maxprob = prob;
        }
        // if( ResMass[ires] == 0 ) {
        //   std::cout << "Error: ResMass["<<ires<<"] = " << ResMass[ires] << std::endl;
        //   exit(1);
        // }
        double deltaMass = fabs(mcomb-ResMass[ires])/ResMass[ires];
        if( deltaMass<minDeltaMass ){
          bestMassMuons = std::make_pair((*Muon1),(*Muon2));
          minDeltaMass = deltaMass;
        }
      }
    }
      }
    }
  }
  //If outside mass window (maxprob==0) then take the two muons with best invariant mass
  //(anyway they will not be used in the likelihood calculation, only to fill plots)
  if(!maxprob){
    if(bestMassMuons.first.charge()<0){
      recMuFromBestRes.first = bestMassMuons.first;
      recMuFromBestRes.second = bestMassMuons.second;
    }
    else{
      recMuFromBestRes.second = bestMassMuons.first;
      recMuFromBestRes.first = bestMassMuons.second;
    }
  }
  return recMuFromBestRes;
}

// Resolution smearing function called to worsen muon Pt resolution at start
// -------------------------------------------------------------------------
lorentzVector MuScleFitUtils::applySmearing (const lorentzVector& muon)
{
  double pt = muon.Pt();
  double eta = muon.Eta();
  double phi = muon.Phi();
  double E = muon.E();

  double y[7] = {};
  for (int i=0; i<SmearType+3; i++) {
    y[i] = x[i][goodmuon%10000];
  }

  // Use the smear function selected in the constructor
  smearFunction->smear( pt, eta, phi, y, parSmear );

  if (debug>9) {
    std::cout << "Smearing Pt,eta,phi = " << pt << " " <<  eta << " "
     << phi << "; x = ";
    for (int i=0; i<SmearType+3; i++) {
      std::cout << y[i];
    }
    std::cout << std::endl;
  }

  double ptEtaPhiE[4] = {pt, eta, phi, E};
  return( fromPtEtaPhiToPxPyPz(ptEtaPhiE) );
}

// Biasing function called to modify muon Pt scale at the start.
// -------------------------------------------------------------
lorentzVector MuScleFitUtils::applyBias( const lorentzVector& muon, const int chg )
{
  double ptEtaPhiE[4] = {muon.Pt(),muon.Eta(),muon.Phi(),muon.E()};

  if (MuScleFitUtils::debug>1) std::cout << "pt before bias = " << ptEtaPhiE[0] << std::endl;

  // Use functors (although not with the () operator)
  // Note that we always pass pt, eta and phi, but internally only the needed
  // values are used.
  // The functors used are takend from the same group used for the scaling
  // thus the name of the method used is "scale".
  ptEtaPhiE[0] = biasFunction->scale(ptEtaPhiE[0], ptEtaPhiE[1], ptEtaPhiE[2], chg, MuScleFitUtils::parBias);

  if (MuScleFitUtils::debug>1) std::cout << "pt after bias = " << ptEtaPhiE[0] << std::endl;

  return( fromPtEtaPhiToPxPyPz(ptEtaPhiE) );
}

// Version of applyScale accepting a std::vector<double> of parameters
// --------------------------------------------------------------
lorentzVector MuScleFitUtils::applyScale (const lorentzVector& muon,
                                          const std::vector<double> & parval, const int chg)
{
  double * p = new double[(int)(parval.size())];
  // Replaced by auto_ptr, which handles delete at the end
  // std::auto_ptr<double> p(new double[(int)(parval.size())]);
  // Removed auto_ptr, check massResolution for an explanation.
  int id = 0;
  for (std::vector<double>::const_iterator it=parval.begin(); it!=parval.end(); ++it, ++id) {
    //(&*p)[id] = *it;
    // Also ok would be (p.get())[id] = *it;
    p[id] = *it;
  }
  lorentzVector tempScaleVec( applyScale (muon, p, chg) );
  delete[] p;
  return tempScaleVec;
}

// This is called by the likelihood to "taste" different values for additional corrections
// ---------------------------------------------------------------------------------------
lorentzVector MuScleFitUtils::applyScale (const lorentzVector& muon,
                                          double* parval, const int chg)
{
  double ptEtaPhiE[4] = {muon.Pt(),muon.Eta(),muon.Phi(),muon.E()};
  int shift = parResol.size();

  if (MuScleFitUtils::debug>1) std::cout << "pt before scale = " << ptEtaPhiE[0] << std::endl;

  // the address of parval[shift] is passed as pointer to double. Internally it is used as a normal array, thus:
  // array[0] = parval[shift], array[1] = parval[shift+1], ...
  ptEtaPhiE[0] = scaleFunction->scale(ptEtaPhiE[0], ptEtaPhiE[1], ptEtaPhiE[2], chg, &(parval[shift]));

  if (MuScleFitUtils::debug>1) std::cout << "pt after scale = " << ptEtaPhiE[0] << std::endl;

  return( fromPtEtaPhiToPxPyPz(ptEtaPhiE) );
}

// Useful function to convert 4-vector coordinates
// -----------------------------------------------
lorentzVector MuScleFitUtils::fromPtEtaPhiToPxPyPz( const double* ptEtaPhiE )
{
  double px = ptEtaPhiE[0]*cos(ptEtaPhiE[2]);
  double py = ptEtaPhiE[0]*sin(ptEtaPhiE[2]);
  double tmp = 2*atan(exp(-ptEtaPhiE[1]));
  double pz = ptEtaPhiE[0]*cos(tmp)/sin(tmp);
  double E  = sqrt(px*px+py*py+pz*pz+muMass*muMass);

  // lorentzVector corrMu(px,py,pz,E);
  // To fix memory leaks, this is to be substituted with
  // std::auto_ptr<lorentzVector> corrMu(new lorentzVector(px, py, pz, E));

  return lorentzVector(px,py,pz,E);
}

// Dimuon mass
// -----------
inline double MuScleFitUtils::invDimuonMass( const lorentzVector& mu1,
                                             const lorentzVector& mu2 )
{
  return (mu1+mu2).mass();
}

// Mass resolution - version accepting a std::vector<double> parval
// -----------------------------------------------------------
double MuScleFitUtils::massResolution( const lorentzVector& mu1,
                                       const lorentzVector& mu2,
                                       const std::vector<double> & parval )
{
  // double * p = new double[(int)(parval.size())];
  // Replaced by auto_ptr, which handles delete at the end
  // --------- //
  // ATTENTION //
  // --------- //
  // auto_ptr calls delete, not delete[] and thus it must
  // not be used with arrays. There are alternatives see
  // e.g.: http://www.gotw.ca/gotw/042.htm. The best
  // alternative seems to be to switch to vector though.
  // std::auto_ptr<double> p(new double[(int)(parval.size())]);

  double * p = new double[(int)(parval.size())];
  std::vector<double>::const_iterator it = parval.begin();
  int id = 0;
  for ( ; it!=parval.end(); ++it, ++id) {
    // (&*p)[id] = *it;
    p[id] = *it;
  }
  double massRes = massResolution (mu1, mu2, p);
  delete[] p;
  return massRes;
}

double MuScleFitUtils::massResolution( const lorentzVector& mu1,
                                       const lorentzVector& mu2,
                                       double* parval )
{
  double mass   = (mu1+mu2).mass();
  double pt1    = mu1.Pt();
  double phi1   = mu1.Phi();
  double eta1   = mu1.Eta();
  double theta1 = 2*atan(exp(-eta1));
  double pt2    = mu2.Pt();
  double phi2   = mu2.Phi();
  double eta2   = mu2.Eta();
  double theta2 = 2*atan(exp(-eta2));

  double dmdpt1  = (pt1/std::pow(sin(theta1),2)*sqrt((std::pow(pt2/sin(theta2),2)+mMu2)/(std::pow(pt1/sin(theta1),2)+mMu2))-
            pt2*(cos(phi1-phi2)+cos(theta1)*cos(theta2)/(sin(theta1)*sin(theta2))))/mass;
  double dmdpt2  = (pt2/std::pow(sin(theta2),2)*sqrt((std::pow(pt1/sin(theta1),2)+mMu2)/(std::pow(pt2/sin(theta2),2)+mMu2))-
            pt1*(cos(phi2-phi1)+cos(theta2)*cos(theta1)/(sin(theta2)*sin(theta1))))/mass;
  double dmdphi1 = pt1*pt2/mass*sin(phi1-phi2);
  double dmdphi2 = pt2*pt1/mass*sin(phi2-phi1);
  double dmdcotgth1 = (pt1*pt1*cos(theta1)/sin(theta1)*
                       sqrt((std::pow(pt2/sin(theta2),2)+mMu2)/(std::pow(pt1/sin(theta1),2)+mMu2)) -
               pt1*pt2*cos(theta2)/sin(theta2))/mass;
  double dmdcotgth2 = (pt2*pt2*cos(theta2)/sin(theta2)*
                       sqrt((std::pow(pt1/sin(theta1),2)+mMu2)/(std::pow(pt2/sin(theta2),2)+mMu2)) -
               pt2*pt1*cos(theta1)/sin(theta1))/mass;

  if( debugMassResol_ ) {
    massResolComponents.dmdpt1 = dmdpt1;
    massResolComponents.dmdpt2 = dmdpt2;
    massResolComponents.dmdphi1 = dmdphi1;
    massResolComponents.dmdphi2 = dmdphi2;
    massResolComponents.dmdcotgth1 = dmdcotgth1;
    massResolComponents.dmdcotgth2 = dmdcotgth2;
  }

  // Resolution parameters:
  // ----------------------
  double sigma_pt1 = resolutionFunction->sigmaPt( pt1,eta1,parval );
  double sigma_pt2 = resolutionFunction->sigmaPt( pt2,eta2,parval );
  double sigma_phi1 = resolutionFunction->sigmaPhi( pt1,eta1,parval );
  double sigma_phi2 = resolutionFunction->sigmaPhi( pt2,eta2,parval );
  double sigma_cotgth1 = resolutionFunction->sigmaCotgTh( pt1,eta1,parval );
  double sigma_cotgth2 = resolutionFunction->sigmaCotgTh( pt2,eta2,parval );
  double cov_pt1pt2 = resolutionFunction->covPt1Pt2( pt1, eta1, pt2, eta2, parval );

  // Sigma_Pt is defined as a relative sigmaPt/Pt for this reason we need to
  // multiply it by pt.
  double mass_res = sqrt(std::pow(dmdpt1*sigma_pt1*pt1,2)+std::pow(dmdpt2*sigma_pt2*pt2,2)+
             std::pow(dmdphi1*sigma_phi1,2)+std::pow(dmdphi2*sigma_phi2,2)+
             std::pow(dmdcotgth1*sigma_cotgth1,2)+std::pow(dmdcotgth2*sigma_cotgth2,2)+
             2*dmdpt1*dmdpt2*cov_pt1pt2*sigma_pt1*sigma_pt2);

  if (debug>19) {
    std::cout << "  Pt1=" << pt1 << " phi1=" << phi1 << " cotgth1=" << cos(theta1)/sin(theta1) << " - Pt2=" << pt2
     << " phi2=" << phi2 << " cotgth2=" << cos(theta2)/sin(theta2) << std::endl;
    std::cout << " P[0]="
     << parval[0] << " P[1]=" << parval[1] << "P[2]=" << parval[2] << " P[3]=" << parval[3] << std::endl;
    std::cout << "  Dmdpt1= " << dmdpt1 << " dmdpt2= " << dmdpt2 << " sigma_pt1="
     << sigma_pt1 << " sigma_pt2=" << sigma_pt2 << std::endl;
    std::cout << "  Dmdphi1= " << dmdphi1 << " dmdphi2= " << dmdphi2 << " sigma_phi1="
     << sigma_phi1 << " sigma_phi2=" << sigma_phi2 << std::endl;
    std::cout << "  Dmdcotgth1= " << dmdcotgth1 << " dmdcotgth2= " << dmdcotgth2
     << " sigma_cotgth1="
     << sigma_cotgth1 << " sigma_cotgth2=" << sigma_cotgth2 << std::endl;
    std::cout << "  Mass resolution (pval) for muons of Pt = " << pt1 << " " << pt2
     << " : " << mass << " +- " << mass_res << std::endl;
  }

  // Debug std::cout
  // ----------
  bool didit = false;
  for (int ires=0; ires<6; ires++) {
    if (!didit && resfind[ires]>0 && fabs(mass-ResMass[ires])<ResHalfWidth[ires]) {
      if (mass_res>ResMaxSigma[ires] && counter_resprob<100) {
    counter_resprob++;
    LogDebug("MuScleFitUtils") << "RESOLUTION PROBLEM: ires=" << ires << std::endl;
    didit = true;
      }
    }
  }

  return mass_res;
}

double MuScleFitUtils::massResolution( const lorentzVector& mu1,
                                       const lorentzVector& mu2,
                       const ResolutionFunction & resolFunc )
{
  double mass   = (mu1+mu2).mass();
  double pt1    = mu1.Pt();
  double phi1   = mu1.Phi();
  double eta1   = mu1.Eta();
  double theta1 = 2*atan(exp(-eta1));
  double pt2    = mu2.Pt();
  double phi2   = mu2.Phi();
  double eta2   = mu2.Eta();
  double theta2 = 2*atan(exp(-eta2));

  double dmdpt1  = (pt1/std::pow(sin(theta1),2)*sqrt((std::pow(pt2/sin(theta2),2)+mMu2)/(std::pow(pt1/sin(theta1),2)+mMu2))-
            pt2*(cos(phi1-phi2)+cos(theta1)*cos(theta2)/(sin(theta1)*sin(theta2))))/mass;
  double dmdpt2  = (pt2/std::pow(sin(theta2),2)*sqrt((std::pow(pt1/sin(theta1),2)+mMu2)/(std::pow(pt2/sin(theta2),2)+mMu2))-
            pt1*(cos(phi2-phi1)+cos(theta2)*cos(theta1)/(sin(theta2)*sin(theta1))))/mass;
  double dmdphi1 = pt1*pt2/mass*sin(phi1-phi2);
  double dmdphi2 = pt2*pt1/mass*sin(phi2-phi1);
  double dmdcotgth1 = (pt1*pt1*cos(theta1)/sin(theta1)*
                       sqrt((std::pow(pt2/sin(theta2),2)+mMu2)/(std::pow(pt1/sin(theta1),2)+mMu2)) -
               pt1*pt2*cos(theta2)/sin(theta2))/mass;
  double dmdcotgth2 = (pt2*pt2*cos(theta2)/sin(theta2)*
                       sqrt((std::pow(pt1/sin(theta1),2)+mMu2)/(std::pow(pt2/sin(theta2),2)+mMu2)) -
               pt2*pt1*cos(theta1)/sin(theta1))/mass;

  // Resolution parameters:
  // ----------------------
  double sigma_pt1 = resolFunc.sigmaPt( mu1 );
  double sigma_pt2 = resolFunc.sigmaPt( mu2 );
  double sigma_phi1 = resolFunc.sigmaPhi( mu1 );
  double sigma_phi2 = resolFunc.sigmaPhi( mu2 );
  double sigma_cotgth1 = resolFunc.sigmaCotgTh( mu1 );
  double sigma_cotgth2 = resolFunc.sigmaCotgTh( mu2 );

  // Sigma_Pt is defined as a relative sigmaPt/Pt for this reason we need to
  // multiply it by pt.
  double mass_res = sqrt(std::pow(dmdpt1*sigma_pt1*pt1,2)+std::pow(dmdpt2*sigma_pt2*pt2,2)+
             std::pow(dmdphi1*sigma_phi1,2)+std::pow(dmdphi2*sigma_phi2,2)+
             std::pow(dmdcotgth1*sigma_cotgth1,2)+std::pow(dmdcotgth2*sigma_cotgth2,2));

  return mass_res;
}


// Mass probability - version with linear background included, accepts std::vector<double> parval
// -----------------------------------------------------------------------------------------
double MuScleFitUtils::massProb( const double & mass, const double & resEta, const double & rapidity, const double & massResol, const std::vector<double> & parval, const bool doUseBkgrWindow, const double & eta1, const double & eta2 )
{
#ifdef USE_CALLGRIND
  CALLGRIND_START_INSTRUMENTATION;
#endif

  double * p = new double[(int)(parval.size())];
  // Replaced by auto_ptr, which handles delete at the end
  // Removed auto_ptr, check massResolution for an explanation.
  // std::auto_ptr<double> p(new double[(int)(parval.size())]);

  std::vector<double>::const_iterator it = parval.begin();
  int id = 0;
  for ( ; it!=parval.end(); ++it, ++id) {
    // (&*p)[id] = *it;
    p[id] = *it;
  }
  // p must be passed by value as below:
  double massProbability = massProb( mass, resEta, rapidity, massResol, p, doUseBkgrWindow, eta1, eta2 );
  delete[] p;

#ifdef USE_CALLGRIND
  CALLGRIND_STOP_INSTRUMENTATION;
  CALLGRIND_DUMP_STATS;
#endif

  return massProbability;
}

double MuScleFitUtils::probability( const double & mass, const double & massResol,
                                    const double GLvalue[][1001][1001], const double GLnorm[][1001],
                                    const int iRes, const int iY )
{
  if( iRes == 0 && iY > 23 ) {
    std::cout << "WARNING: rapidity bin selected = " << iY << " but there are only histograms for the first 24 bins" << std::endl;
  }

  double PS = 0.;
  bool insideProbMassWindow = true;
  // Interpolate the four values of GLZValue[] in the
  // grid square within which the (mass,sigma) values lay
  // ----------------------------------------------------
  // This must be done with respect to the width used in the computation of the probability distribution,
  // so that the bin 0 really matches the bin 0 of that distribution.
  //  double fracMass = (mass-(ResMass[iRes]-ResHalfWidth[iRes]))/(2*ResHalfWidth[iRes]);
  double fracMass = (mass - ResMinMass[iRes])/(2*ResHalfWidth[iRes]);
  if (debug>1) std::cout << std::setprecision(9)<<"mass ResMinMass[iRes] ResHalfWidth[iRes] ResHalfWidth[iRes]"
                    << mass << " "<<ResMinMass[iRes]<<" "<<ResHalfWidth[iRes]<<" "<<ResHalfWidth[iRes]<<std::endl;
  int iMassLeft  = (int)(fracMass*(double)nbins);
  int iMassRight = iMassLeft+1;
  double fracMassStep = (double)nbins*(fracMass - (double)iMassLeft/(double)nbins);
  if (debug>1) std::cout<<"nbins iMassLeft fracMass "<<nbins<<" "<<iMassLeft<<" "<<fracMass<<std::endl;

  // Simple protections for the time being: the region where we fit should not include
  // values outside the boundaries set by ResMass-ResHalfWidth : ResMass+ResHalfWidth
  // ---------------------------------------------------------------------------------
  if (iMassLeft<0) {
    edm::LogInfo("probability") << "WARNING: fracMass=" << fracMass << ", iMassLeft="
                           << iMassLeft << "; mass = " << mass << " and bounds are " << ResMinMass[iRes]
                           << ":" << ResMinMass[iRes]+2*ResHalfWidth[iRes] << " - iMassLeft set to 0" << std::endl;
    iMassLeft  = 0;
    iMassRight = 1;
    insideProbMassWindow = false;
  }
  if (iMassRight>nbins) {
    edm::LogInfo("probability") << "WARNING: fracMass=" << fracMass << ", iMassRight="
                           << iMassRight << "; mass = " << mass << " and bounds are " << ResMinMass[iRes]
                           << ":" << ResMass[iRes]+2*ResHalfWidth[iRes] << " - iMassRight set to " << nbins-1 << std::endl;
    iMassLeft  = nbins-1;
    iMassRight = nbins;
    insideProbMassWindow = false;
  }
  double fracSigma = (massResol/ResMaxSigma[iRes]);
  int iSigmaLeft = (int)(fracSigma*(double)nbins);
  int iSigmaRight = iSigmaLeft+1;
  double fracSigmaStep = (double)nbins * (fracSigma - (double)iSigmaLeft/(double)nbins);
  // std::cout << "massResol = " << massResol << std::endl;
  // std::cout << "ResMaxSigma["<<iRes<<"] = " << ResMaxSigma[iRes] << std::endl;
  // std::cout << "fracSigma = " << fracSigma << std::endl;
  // std::cout << "nbins = " << nbins << std::endl;
  // std::cout << "ISIGMALEFT = " << iSigmaLeft << std::endl;
  // std::cout << "ISIGMARIGHT = " << iSigmaRight << std::endl;
  // std::cout << "fracSigmaStep = " << fracSigmaStep << std::endl;

  // Simple protections for the time being: they should not affect convergence, since
  // ResMaxSigma is set to very large values, and if massResol exceeds them the fit
  // should not get any prize for that (for large sigma, the prob. distr. becomes flat)
  // ----------------------------------------------------------------------------------
  if (iSigmaLeft<0) {
    edm::LogInfo("probability") << "WARNING: fracSigma = " << fracSigma << ", iSigmaLeft="
                           << iSigmaLeft << ", with massResol = " << massResol << " and ResMaxSigma[iRes] = "
                           << ResMaxSigma[iRes] << " -  iSigmaLeft set to 0" << std::endl;
    iSigmaLeft  = 0;
    iSigmaRight = 1;
  }
  if (iSigmaRight>nbins ) {
    if (counter_resprob<100)
      edm::LogInfo("probability") << "WARNING: fracSigma = " << fracSigma << ", iSigmaRight="
                             << iSigmaRight << ", with massResol = " << massResol << " and ResMaxSigma[iRes] = "
                             << ResMaxSigma[iRes] << " -  iSigmaRight set to " << nbins-1 << std::endl;
    iSigmaLeft  = nbins-1;
    iSigmaRight = nbins;
  }

  // If f11,f12,f21,f22 are the values at the four corners, one finds by linear interpolation the
  // formula below for PS
  // --------------------------------------------------------------------------------------------
  if( insideProbMassWindow ) {
    double f11 = 0.;
    if (GLnorm[iY][iSigmaLeft]>0)
      f11 = GLvalue[iY][iMassLeft][iSigmaLeft] / GLnorm[iY][iSigmaLeft];
    double f12 = 0.;
    if (GLnorm[iY][iSigmaRight]>0)
      f12 = GLvalue[iY][iMassLeft][iSigmaRight] / GLnorm[iY][iSigmaRight];
    double f21 = 0.;
    if (GLnorm[iY][iSigmaLeft]>0)
      f21 = GLvalue[iY][iMassRight][iSigmaLeft] / GLnorm[iY][iSigmaLeft];
    double f22 = 0.;
    if (GLnorm[iY][iSigmaRight]>0)
      f22 = GLvalue[iY][iMassRight][iSigmaRight] / GLnorm[iY][iSigmaRight];
    PS = f11 + (f12-f11)*fracSigmaStep + (f21-f11)*fracMassStep +
      (f22-f21-f12+f11)*fracMassStep*fracSigmaStep;
    if (PS>0.1 || debug>1) LogDebug("MuScleFitUtils") << "iRes = " << iRes << " PS=" << PS << " f11,f12,f21,f22="
                                                      << f11 << " " << f12 << " " << f21 << " " << f22 << " "
                                                      << " fSS=" << fracSigmaStep << " fMS=" << fracMassStep << " iSL, iSR="
                                                      << iSigmaLeft << " " << iSigmaRight
                                                      << " GLvalue["<<iY<<"]["<<iMassLeft<<"] = " << GLvalue[iY][iMassLeft][iSigmaLeft]
                                                      << " GLnorm["<<iY<<"]["<<iSigmaLeft<<"] = " << GLnorm[iY][iSigmaLeft] << std::endl;

//     if (PS>0.1) std::cout << "iRes = " << iRes << " PS=" << PS << " f11,f12,f21,f22="
//                      << f11 << " " << f12 << " " << f21 << " " << f22 << " "
//                      << " fSS=" << fracSigmaStep << " fMS=" << fracMassStep << " iSL, iSR="
//                      << iSigmaLeft << " " << iSigmaRight << " GLV,GLN="
//                      << GLvalue[iY][iMassLeft][iSigmaLeft]
//                      << " " << GLnorm[iY][iSigmaLeft] << std::endl;

  }
  else {
    edm::LogInfo("probability") << "outside mass probability window. Setting PS["<<iRes<<"] = 0" << std::endl;
  }

//   if( PS != PS ) {
//     std::cout << "ERROR: PS = " << PS << " for iRes = " << iRes << std::endl;

//     std::cout << "mass = " << mass << ", massResol = " << massResol << std::endl;
//     std::cout << "fracMass = " << fracMass << ", iMassLeft = " << iMassLeft
//          << ", iMassRight = " << iMassRight << ", fracMassStep = " << fracMassStep << std::endl;
//     std::cout << "fracSigma = " << fracSigma << ", iSigmaLeft = " << iSigmaLeft
//          << ", iSigmaRight = " << iSigmaRight << ", fracSigmaStep = " << fracSigmaStep << std::endl;
//     std::cout << "ResMaxSigma["<<iRes<<"] = " << ResMaxSigma[iRes] << std::endl;

//     std::cout << "GLvalue["<<iY<<"]["<<iMassLeft<<"] = " << GLvalue[iY][iMassLeft][iSigmaLeft]
//          << " GLnorm["<<iY<<"]["<<iSigmaLeft<<"] = " << GLnorm[iY][iSigmaLeft] << std::endl;
//   }

  return PS;
}

// Mass probability - version with linear background included
// ----------------------------------------------------------
double MuScleFitUtils::massProb( const double & mass, const double & resEta, const double & rapidity, const double & massResol, double * parval, const bool doUseBkgrWindow, const double & eta1, const double & eta2 ) {

  // This routine computes the likelihood that a given measured mass "measMass" is
  // the result of a reference mass ResMass[] if the resolution
  // expected for the two muons is massResol.
  // This version includes two parameters (the last two in parval, by default)
  // to size up the background fraction and its relative normalization with respect
  // to the signal shape.
  //
  // We model the signal probability with a Lorentz L(M,H) of resonance mass M and natural width H
  // convoluted with a gaussian G(m,s) of measured mass m and expected mass resolution s,
  // by integrating over the intersection of the supports of L and G (which can be made to coincide with
  // the region where L is non-zero, given that H<<s in most cases) the product L(M,H)*G(m,s) dx as follows:
  //
  //   GL(m,s) = Int(M-10H,M+10H) [ L(x-M,H) * G(x-m,s) ] dx
  //
  // The above convolution is computed numerically by an independent root macro, Probs.C, which outputs
  // the values in six 1001x1001 grids, one per resonance.
  //
  // NB THe following block of explanations for background models is outdated, see detailed
  // explanations where the code computes PB.
  // +++++++++++++++++++++++
  //   For the background, instead, we have two choices: a linear and an exponential model.
  //     * For the linear model, we choose a one-parameter form whereby the line is automatically normalized
  //       in the support [x1,x2] where we defined our "signal region", as follows:
  //
  //         B(x;b) = 1/(x2-x1) + {x - (x2+x1)/2} * b
  //
  //       Defined as above, B(x) is a line passing for the point of coordinates (x1+x2)/2, 1/(x2-x1),
  //       whose slope b has as support the interval ( -2/[(x1-x2)*(x1+x2)], 2/[(x1-x2)*(x1+x2)] )
  //       so that B(x) is always positive-definite in [x1,x2].
  //
  //     * For the exponential model, we define B(x;b) as
  //
  //         B(x;b) = b * { exp(-b*x) / [exp(-b*x1)-exp(-b*x2)] }
  //
  //       This one-parameter definition is automatically normalized to unity in [x1,x2], with a parameter
  //       b which has to be positive in order for the slope to be negative.
  //       Please note that this model is not useful in most circumstances; a more useful form would be one
  //       which included a linear component.
  // ++++++++++++++++++++++
  //
  // Once GL(m,s) and B(x;b) are defined, we introduce a further parameter a, such that we can have the
  // likelihood control the relative fraction of signal and background. We first normalize GL(m,s) for
  // any given s by taking the integral
  //
  //   Int(x1,x2) GL(m,s) dm = K_s
  //
  // We then define the probability as
  //
  //   P(m,s,a,b) = GL(m,s)/K_s * a  +  B(x,b) * (1-a)
  //
  // with a taking on values in the interval [0,1].
  // Defined as above, the probability is well-behaved, in the sense that it has a value between 0 and 1,
  // and the four parameters m,s,a,b fully control its shape.
  //
  // It is to be noted that the formulation above requires the computation of two rather time-consuming
  // integrals. The one defining GL(m,s) can be stored in a TH2D and loaded by the constructor from a
  // file suitably prepared, and this will save loads of computing time.
  // ----------------------------------------------------------------------------------------------------

  double P = 0.;
  int crossSectionParShift = parResol.size() + parScale.size();
  // Take the relative cross sections
  std::vector<double> relativeCrossSections = crossSectionHandler->relativeCrossSections(&(parval[crossSectionParShift]), resfind);
  //   for( unsigned int i=0; i<relativeCrossSections.size(); ++i ) {
  //     std::cout << "relativeCrossSections["<<i<<"] = " << relativeCrossSections[i] << std::endl;
  //     std::cout << "parval["<<crossSectionParShift+i<<"] = " << parval[crossSectionParShift+i] << std::endl;
  //   }

  // int bgrParShift = crossSectionParShift + parCrossSection.size();
  int bgrParShift = crossSectionParShift + crossSectionHandler->parNum();
  double Bgrp1 = 0.;
  //   double Bgrp2 = 0.;
  //   double Bgrp3 = 0.;

  // NB defined as below, P is a non-rigorous "probability" that we observe
  // a dimuon mass "mass", given ResMass[], gamma, and massResol. It is what we need for the
  // fit which finds the best resolution parameters, though. A definition which is
  // more properly a probability is given below (in massProb2()), where the result
  // cannot be used to fit resolution parameters because the fit would always prefer
  // to set the res parameters to the minimum possible value (best resolution),
  // to have a probability close to one of observing any mass.
  // -------------------------------------------------------------------------------

  // Determine what resonance(s) we have to deal with
  // NB for now we assume equal xs for each resonance
  // so we do not assign them different weights
  // ------------------------------------------------
  double PS[6] = {0.};
  double PB = 0.;
  double PStot[6] = {0.};

  // Should be removed because it is not used
  bool resConsidered[6] = {false};

  bool useBackgroundWindow = (doBackgroundFit[loopCounter] || doUseBkgrWindow);
  // bool useBackgroundWindow = (doBackgroundFit[loopCounter]);

  // First check the Z, which is divided in 24 rapidity bins
  // NB max value of Z rapidity to be considered is 2.4 here
  // -------------------------------------------------------

  // Do this only if we want to use the rapidity bins for the Z
  if( MuScleFitUtils::rapidityBinsForZ_ ) {
    // ATTENTION: cut on Z rapidity at 2.4 since we only have histograms up to that value
    // std::pair<double, double> windowFactors = backgroundHandler->windowFactors( useBackgroundWindow, 0 );
    std::pair<double, double> windowBorders = backgroundHandler->windowBorders( useBackgroundWindow, 0 );
    if( resfind[0]>0
        // && checkMassWindow( mass, 0,
        //                     backgroundHandler->resMass( useBackgroundWindow, 0 ),
        //                     windowFactors.first, windowFactors.second )
        && checkMassWindow( mass, windowBorders.first, windowBorders.second )
        // && fabs(rapidity)<2.4
        ) {

      int iY = (int)(fabs(rapidity)*10.);
      if( iY > 23 ) iY = 23;

      if (MuScleFitUtils::debug>1) std::cout << "massProb:resFound = 0, rapidity bin =" << iY << std::endl;

      // In this case the last value is the rapidity bin
      PS[0] = probability(mass, massResol, GLZValue, GLZNorm, 0, iY);

      if( PS[0] != PS[0] ) {
        std::cout << "ERROR: PS[0] = nan, setting it to 0" << std::endl;
        PS[0] = 0;
      }

      // std::pair<double, double> bgrResult = backgroundHandler->backgroundFunction( doBackgroundFit[loopCounter],
      //                                           &(parval[bgrParShift]), MuScleFitUtils::totalResNum, 0,
      //                                           resConsidered, ResMass, ResHalfWidth, MuonType, mass, resEta );

      std::pair<double, double> bgrResult = backgroundHandler->backgroundFunction( doBackgroundFit[loopCounter],
                                           &(parval[bgrParShift]), MuScleFitUtils::totalResNum, 0,
                                           resConsidered, ResMass, ResHalfWidth, MuonType, mass, eta1, eta2 );

      Bgrp1 = bgrResult.first;
      // When fitting the background we have only one Bgrp1
      // When not fitting the background we have many only in a superposition region and this case is treated
      // separately after this loop
      PB = bgrResult.second;
      if( PB != PB ) PB = 0;
      PStot[0] = (1-Bgrp1)*PS[0] + Bgrp1*PB;


      // PStot[0] *= crossSectionHandler->crossSection(0);
      // PStot[0] *= parval[crossSectionParShift];
      PStot[0] *= relativeCrossSections[0];
      if ( MuScleFitUtils::debug > 0 ) std::cout << "PStot["<<0<<"] = " << "(1-"<<Bgrp1<<")*"<<PS[0]<<" + "<<Bgrp1<<"*"<<PB<<" = " << PStot[0] << " (reletiveXS = )" << relativeCrossSections[0] << std::endl;
    }
    else {
      if( debug > 0 ) {
        std::cout << "Mass = " << mass << " outside range with rapidity = " << rapidity << std::endl;
        std::cout << "with resMass = " << backgroundHandler->resMass( useBackgroundWindow, 0 ) << " and left border = " << windowBorders.first << " right border = " << windowBorders.second << std::endl;
      }
    }
  }
  // Next check the other resonances
  // -------------------------------
  int firstRes = 1;
  if( !MuScleFitUtils::rapidityBinsForZ_ ) firstRes = 0;
  for( int ires=firstRes; ires<6; ++ires ) {
    if( resfind[ires] > 0 ) {
      // First is left, second is right (returns (1,1) in the case of resonances, it could be improved avoiding the call in this case)
      // std::pair<double, double> windowFactor = backgroundHandler->windowFactors( useBackgroundWindow, ires );
      std::pair<double, double> windowBorder = backgroundHandler->windowBorders( useBackgroundWindow, ires );
      if( checkMassWindow(mass, windowBorder.first, windowBorder.second) ) {
        if (MuScleFitUtils::debug>1) std::cout << "massProb:resFound = " << ires << std::endl;

        // In this case the rapidity value is instead the resonance index again.
        PS[ires] = probability(mass, massResol, GLValue, GLNorm, ires, ires);

        std::pair<double, double> bgrResult = backgroundHandler->backgroundFunction( doBackgroundFit[loopCounter],
                                             &(parval[bgrParShift]), MuScleFitUtils::totalResNum, ires,
                                             // resConsidered, ResMass, ResHalfWidth, MuonType, mass, resEta );
                                             resConsidered, ResMass, ResHalfWidth, MuonType, mass, eta1, eta2 );
        Bgrp1 = bgrResult.first;
        PB = bgrResult.second;

        if( PB != PB ) PB = 0;
        PStot[ires] = (1-Bgrp1)*PS[ires] + Bgrp1*PB;
        if( MuScleFitUtils::debug>0 ) std::cout << "PStot["<<ires<<"] = " << "(1-"<<Bgrp1<<")*"<<PS[ires]<<" + "<<Bgrp1<<"*"<<PB<<" = " << PStot[ires] << std::endl;

        PStot[ires] *= relativeCrossSections[ires];
      }
    }
  }

  for( int i=0; i<6; ++i ) {
    P += PStot[i];
  }

  if( MuScleFitUtils::signalProb_ != 0 && MuScleFitUtils::backgroundProb_ != 0 ) {
    double PStotTemp = 0.;
    for( int i=0; i<6; ++i ) {
      PStotTemp += PS[i]*relativeCrossSections[i];
    }
    if( PStotTemp != PStotTemp ) {
      std::cout << "ERROR: PStotTemp = nan!!!!!!!!!" << std::endl;
      int parnumber = (int)(parResol.size()+parScale.size()+crossSectionHandler->parNum()+parBgr.size());
      for( int i=0; i<6; ++i ) {
        std::cout << "PS[i] = " << PS[i] << std::endl;
        if( PS[i] != PS[i] ) {
          std::cout << "mass = " << mass << std::endl;
          std::cout << "massResol = " << massResol << std::endl;
      for( int j=0; j<parnumber; ++j ) {
        std::cout << "parval["<<j<<"] = " << parval[j] << std::endl;
      }
        }
      }
    }
    if( PStotTemp == PStotTemp ) {
      MuScleFitUtils::signalProb_->SetBinContent(MuScleFitUtils::minuitLoop_, MuScleFitUtils::signalProb_->GetBinContent(MuScleFitUtils::minuitLoop_) + PStotTemp);
    }
    if (debug>0) std::cout << "mass = " << mass << ", P = " << P << ", PStot = " << PStotTemp << ", PB = " << PB << ", bgrp1 = " << Bgrp1 << std::endl;

    MuScleFitUtils::backgroundProb_->SetBinContent(MuScleFitUtils::minuitLoop_, MuScleFitUtils::backgroundProb_->GetBinContent(MuScleFitUtils::minuitLoop_) + PB);
  }
  return P;
}

// Method to check if the mass value is within the mass window of the i-th resonance.
// inline bool MuScleFitUtils::checkMassWindow( const double & mass, const int ires, const double & resMass, const double & leftFactor, const double & rightFactor )
// {
//   return( mass-resMass > -leftFactor*massWindowHalfWidth[MuonTypeForCheckMassWindow][ires]
//           && mass-resMass < rightFactor*massWindowHalfWidth[MuonTypeForCheckMassWindow][ires] );
// }
inline bool MuScleFitUtils::checkMassWindow( const double & mass, const double & leftBorder, const double & rightBorder )
{
  return( (mass > leftBorder) && (mass < rightBorder) );
}

// Function that returns the weight for a muon pair
// ------------------------------------------------
double MuScleFitUtils::computeWeight( const double & mass, const int iev, const bool doUseBkgrWindow )
{
  // Compute weight for this event
  // -----------------------------
  double weight = 0.;

  // Take the highest-mass resonance within bounds
  // NB this must be revised once credible estimates of the relative xs of Y(1S), (2S), and (3S)
  // are made. Those are priors in the decision of which resonance to assign to an in-between event.
  // -----------------------------------------------------------------------------------------------

  if( doUseBkgrWindow && (debug > 0) ) std::cout << "using backgrond window for mass = " << mass << std::endl;
  // bool useBackgroundWindow = (doBackgroundFit[loopCounter] || doUseBkgrWindow);
  bool useBackgroundWindow = (doBackgroundFit[loopCounter]);

  for (int ires=0; ires<6; ires++) {
    if (resfind[ires]>0 && weight==0.) {
      // std::pair<double, double> windowFactor = backgroundHandler->windowFactors( useBackgroundWindow, ires );
      std::pair<double, double> windowBorder = backgroundHandler->windowBorders( useBackgroundWindow, ires );
      // if( checkMassWindow(mass, ires, backgroundHandler->resMass( useBackgroundWindow, ires ),
      //                     windowFactor.first, windowFactor.second) ) {
      if( checkMassWindow(mass, windowBorder.first, windowBorder.second) ) {
        weight = 1.0;
        if( doUseBkgrWindow && (debug > 0) ) std::cout << "setting weight to = " << weight << std::endl;
      }
    }
  }

  return weight;
}

// Likelihood minimization routine
// -------------------------------
void MuScleFitUtils::minimizeLikelihood()
{
  // Output file with fit parameters resulting from minimization
  // -----------------------------------------------------------
  ofstream FitParametersFile;
  FitParametersFile.open ("FitParameters.txt", std::ios::app);
  FitParametersFile << "Fitting with resolution, scale, bgr function # "
            << ResolFitType << " " << ScaleFitType << " " << BgrFitType
            << " - Iteration " << loopCounter << std::endl;

  // Fill parvalue and other vectors needed for the fitting
  // ------------------------------------------------------





  // ----- //
  // FIXME //
  // ----- //
  // this was changed to verify the possibility that fixed parameters influence the errors.
  // It must be 0 otherwise the parameters for resonances will not be passed by minuit (will be always 0).
  // Should be removed.
  int parForResonanceWindows = 0;
  // int parnumber = (int)(parResol.size()+parScale.size()+parCrossSection.size()+parBgr.size() - parForResonanceWindows);
  int parnumber = (int)(parResol.size()+parScale.size()+crossSectionHandler->parNum()+parBgr.size() - parForResonanceWindows);






  int parnumberAll = (int)(parResol.size()+parScale.size()+crossSectionHandler->parNum()+parBgr.size());

  // parvalue is a std::vector<std::vector<double> > storing all the parameters from all the loops
  parvalue.push_back(parResol);
  std::vector<double> *tmpVec = &(parvalue.back());

  // If this is not the first loop we want to start from neutral values
  // Otherwise the scale will start with values correcting again a bias
  // that is already corrected.
  if( scaleFitNotDone_ ) {
    tmpVec->insert( tmpVec->end(), parScale.begin(), parScale.end() );
    std::cout << "scaleFitNotDone: tmpVec->size() = " << tmpVec->size() << std::endl;
  }
  else {
    scaleFunction->resetParameters(tmpVec);
    std::cout << "scaleFitDone: tmpVec->size() = " << tmpVec->size() << std::endl;
  }
  tmpVec->insert( tmpVec->end(), parCrossSection.begin(), parCrossSection.end() );
  tmpVec->insert( tmpVec->end(), parBgr.begin(), parBgr.end() );
  int i = 0;
  std::vector<double>::const_iterator it = tmpVec->begin();
  for( ; it != tmpVec->end(); ++it, ++i ) {
    std::cout << "tmpVec["<<i<<"] = " << *it << std::endl;
  }

  // Empty vector of size = number of cross section fitted parameters. Note that the cross section
  // fit works in a different way than the others and it uses ratios of the paramters passed via cfg.
  // We use this empty vector for compatibility with the rest of the structure.
  std::vector<int> crossSectionParNumSizeVec( MuScleFitUtils::crossSectionHandler->parNum(), 0 );

  std::vector<int> parfix(parResolFix);
  parfix.insert( parfix.end(), parScaleFix.begin(), parScaleFix.end() );
  parfix.insert( parfix.end(), crossSectionParNumSizeVec.begin(), crossSectionParNumSizeVec.end() );
  parfix.insert( parfix.end(), parBgrFix.begin(), parBgrFix.end() );

  std::vector<int> parorder(parResolOrder);
  parorder.insert( parorder.end(), parScaleOrder.begin(), parScaleOrder.end() );
  parorder.insert( parorder.end(), crossSectionParNumSizeVec.begin(), crossSectionParNumSizeVec.end() );
  parorder.insert( parorder.end(), parBgrOrder.begin(), parBgrOrder.end() );

  // This is filled later
  std::vector<double> parerr(3*parnumberAll,0.);

  if (debug>19) {
    std::cout << "[MuScleFitUtils-minimizeLikelihood]: Parameters before likelihood " << std::endl;
    for (unsigned int i=0; i<(unsigned int)parnumberAll; i++) {
      std::cout << "  Par # " << i << " = " << parvalue[loopCounter][i] << " : free = " << parfix[i] << "; order = "
       << parorder[i] << std::endl;
    }
  }

//   // Background rescaling from regions to resonances
//   // -----------------------------------------------
//   // If we are in a loop > 0 and we are not fitting the background, but we have fitted it in the previous iteration
//   if( loopCounter > 0 && !(doBackgroundFit[loopCounter]) && doBackgroundFit[loopCounter-1] ) {
//     // This rescales from regions to resonances
//     int localMuonType = MuonType;
//     if( MuonType > 2 ) localMuonType = 2;
//     backgroundHandler->rescale( parBgr, ResMass, massWindowHalfWidth[localMuonType],
//                                 MuScleFitUtils::SavedPair);
//   }

  // Init Minuit
  // -----------
  TMinuit rmin (parnumber);
  rminPtr_ = &rmin;
  rmin.SetFCN (likelihood);     // Unbinned likelihood
  // Standard initialization of minuit parameters:
  // sets input to be $stdin, output to be $stdout
  // and saving to a file.
  rmin.mninit (5, 6, 7);
  int ierror = 0;
  int istat;
  double arglis[4];
  arglis[0] = FitStrategy;      // Strategy 1 or 2
  // 1 standard
  // 2 try to improve minimum (slower)
  rmin.mnexcm ("SET STR", arglis, 1, ierror);

  arglis[0] = 10001;
  // Set the random seed for the generator used in SEEk to a fixed value for reproducibility
  rmin.mnexcm("SET RAN", arglis, 1, ierror);

  // Set fit parameters
  // ------------------
  double * Start = new double[parnumberAll];
  double * Step  = new double[parnumberAll];
  double * Mini  = new double[parnumberAll];
  double * Maxi  = new double[parnumberAll];
  int * ind = new int[parnumberAll]; // Order of release of parameters
  TString * parname = new TString[parnumberAll];

  if( !parResolStep.empty() && !parResolMin.empty() && !parResolMax.empty() ) {
    MuScleFitUtils::resolutionFunctionForVec->setParameters( Start, Step, Mini, Maxi, ind, parname, parResol, parResolOrder, parResolStep, parResolMin, parResolMax, MuonType );
  }
  else {
    MuScleFitUtils::resolutionFunctionForVec->setParameters( Start, Step, Mini, Maxi, ind, parname, parResol, parResolOrder, MuonType );
  }

  // Take the number of parameters in the resolutionFunction and displace the arrays passed to the scaleFunction
  int resParNum = MuScleFitUtils::resolutionFunctionForVec->parNum();

  if( !parScaleStep.empty() && !parScaleMin.empty() && !parScaleMax.empty() ) {
    MuScleFitUtils::scaleFunctionForVec->setParameters( &(Start[resParNum]), &(Step[resParNum]),
                            &(Mini[resParNum]), &(Maxi[resParNum]),
                            &(ind[resParNum]), &(parname[resParNum]),
                            parScale, parScaleOrder, parScaleStep,
                            parScaleMin, parScaleMax, MuonType );
  }
  else {
    MuScleFitUtils::scaleFunctionForVec->setParameters( &(Start[resParNum]), &(Step[resParNum]),
                            &(Mini[resParNum]), &(Maxi[resParNum]),
                            &(ind[resParNum]), &(parname[resParNum]),
                            parScale, parScaleOrder, MuonType );
  }

  // Initialize cross section parameters
  int crossSectionParShift = resParNum + MuScleFitUtils::scaleFunctionForVec->parNum();
  MuScleFitUtils::crossSectionHandler->setParameters( &(Start[crossSectionParShift]), &(Step[crossSectionParShift]), &(Mini[crossSectionParShift]),
                                                      &(Maxi[crossSectionParShift]), &(ind[crossSectionParShift]), &(parname[crossSectionParShift]),
                                                      parCrossSection, parCrossSectionOrder, resfind );

  // Initialize background parameters
  int bgrParShift = crossSectionParShift + crossSectionHandler->parNum();
  MuScleFitUtils::backgroundHandler->setParameters( &(Start[bgrParShift]), &(Step[bgrParShift]), &(Mini[bgrParShift]), &(Maxi[bgrParShift]),
                                                    &(ind[bgrParShift]), &(parname[bgrParShift]), parBgr, parBgrOrder, MuonType );

  for( int ipar=0; ipar<parnumber; ++ipar ) {
    std::cout << "parname["<<ipar<<"] = " << parname[ipar] << std::endl;
    std::cout << "Start["<<ipar<<"] = " << Start[ipar] << std::endl;
    std::cout << "Step["<<ipar<<"] = " << Step[ipar] << std::endl;
    std::cout << "Mini["<<ipar<<"] = " << Mini[ipar] << std::endl;
    std::cout << "Maxi["<<ipar<<"] = " << Maxi[ipar] << std::endl;


    rmin.mnparm( ipar, parname[ipar], Start[ipar], Step[ipar], Mini[ipar], Maxi[ipar], ierror );


    // Testing without limits
    // rmin.mnparm( ipar, parname[ipar], Start[ipar], Step[ipar], 0, 0, ierror );


  }

  // Do minimization
  // ---------------
  if (debug>19)
    std::cout << "[MuScleFitUtils-minimizeLikelihood]: Starting minimization" << std::endl;
  double fmin;
  double fdem;
  double errdef;
  int npari;
  int nparx;
  rmin.mnexcm ("CALL FCN", arglis, 1, ierror);

  // First, fix all parameters
  // -------------------------
  if (debug>19)
    std::cout << "[MuScleFitUtils-minimizeLikelihood]: First fix all parameters ...";
  for (int ipar=0; ipar<parnumber; ipar++) {
    rmin.FixParameter (ipar);
  }

  // Then release them in the specified order and refit
  // --------------------------------------------------
  if (debug>19) std::cout << " Then release them in order..." << std::endl;

  TString name;
  double pval;
  double pmin;
  double pmax;
  double errp;
  double errl;
  double errh;
  int ivar;
  double erro;
  double cglo;
  int n_times = 0;
  // n_times = number of loops required to unlock all parameters.

  if (debug>19) std::cout << "Before scale parNum" << std::endl;
  int scaleParNum = scaleFunction->parNum();
  if (debug>19) std::cout << "After scale parNum" << std::endl;
  // edm::LogInfo("minimizeLikelihood") << "number of parameters for scaleFunction = " << scaleParNum << std::endl;
  // edm::LogInfo("minimizeLikelihood") << "number of parameters for resolutionFunction = " << resParNum << std::endl;
  // edm::LogInfo("minimizeLikelihood") << "number of parameters for cross section = " << crossSectionHandler->parNum() << std::endl;
  // edm::LogInfo("minimizeLikelihood") << "number of parameters for backgroundFunction = " << parBgr.size() << std::endl;
  std::cout << "number of parameters for scaleFunction = " << scaleParNum << std::endl;
  std::cout << "number of parameters for resolutionFunction = " << resParNum << std::endl;
  std::cout << "number of parameters for cross section = " << crossSectionHandler->parNum() << std::endl;
  std::cout << "number of parameters for backgroundFunction = " << parBgr.size() << std::endl;
  // edm::LogInfo("minimizeLikelihood") << "number of parameters for backgroundFunction = " << backgroundFunction->parNum() << std::endl;

  for (int i=0; i<parnumber; i++) {
    // NB ind[] has been set as parorder[] previously
    if (n_times<ind[i]) {
      edm::LogInfo("minimizeLikelihood") << "n_times = " << n_times << ", ind["<<i<<"] = " << ind[i] << ", scaleParNum = " << scaleParNum << ", doScaleFit["<<loopCounter<<"] = " << doScaleFit[loopCounter] << std::endl;
      // Set the n_times only if we will do the fit
      if ( i<resParNum ) {
        if( doResolFit[loopCounter] ) n_times = ind[i];
      }
      else if( i<resParNum+scaleParNum ) {
        if( doScaleFit[loopCounter] ) n_times = ind[i];
      }
      else if( doBackgroundFit[loopCounter] ) n_times = ind[i];
    }
  }
  for (int iorder=0; iorder<n_times+1; iorder++) { // Repeat fit n_times times
    std::cout << "Starting minimization " << iorder << " of " << n_times << std::endl;

    bool somethingtodo = false;

    // Use parameters from cfg to select which fit to do
    // -------------------------------------------------
    if( doResolFit[loopCounter] ) {
      // Release resolution parameters and fit them
      // ------------------------------------------
      for( unsigned int ipar=0; ipar<parResol.size(); ++ipar ) {
    if( parfix[ipar]==0 && ind[ipar]==iorder ) {
      rmin.Release( ipar );
      somethingtodo = true;
    }
      }
    }
    if( doScaleFit[loopCounter] ) {
      // Release scale parameters and fit them
      // -------------------------------------
      for( unsigned int ipar=parResol.size(); ipar<parResol.size()+parScale.size(); ++ipar ) {
    if( parfix[ipar]==0 && ind[ipar]==iorder ) { // parfix=0 means parameter is free
      rmin.Release( ipar );
      somethingtodo = true;
    }
      }
      scaleFitNotDone_ = false;
    }
    unsigned int crossSectionParShift = parResol.size()+parScale.size();
    if( doCrossSectionFit[loopCounter] ) {
      // Release cross section parameters and fit them
      // ---------------------------------------------
      // Note that only cross sections of resonances that are being fitted are released
      bool doCrossSection = crossSectionHandler->releaseParameters( rmin, resfind, parfix, ind, iorder, crossSectionParShift );
      if( doCrossSection ) somethingtodo = true;
    }
    if( doBackgroundFit[loopCounter] ) {
      // Release background parameters and fit them
      // ------------------------------------------
      // for( int ipar=parResol.size()+parScale.size(); ipar<parnumber; ++ipar ) {
      // Free only the parameters for the regions, as the resonances intervals are never used to fit the background
      unsigned int bgrParShift = crossSectionParShift+crossSectionHandler->parNum();
      for( unsigned int ipar = bgrParShift; ipar < bgrParShift+backgroundHandler->regionsParNum(); ++ipar ) {
        // Release only those parameters for the resonances we are fitting
    if( parfix[ipar]==0 && ind[ipar]==iorder && backgroundHandler->unlockParameter(resfind, ipar - bgrParShift) ) {
      rmin.Release( ipar );
      somethingtodo = true;
    }
      }
    }

    // OK, now do minimization if some parameter has been released
    // -----------------------------------------------------------
    if( somethingtodo ) {
// #ifdef DEBUG

      std::stringstream fileNum;
      fileNum << loopCounter;

      minuitLoop_ = 0;
      char name[50];
      sprintf(name, "likelihoodInLoop_%d_%d", loopCounter, iorder);
      TH1D * tempLikelihoodInLoop = new TH1D(name, "likelihood value in minuit loop", 10000, 0, 10000);
      likelihoodInLoop_ = tempLikelihoodInLoop;
      char signalProbName[50];
      sprintf(signalProbName, "signalProb_%d_%d", loopCounter, iorder);
      TH1D * tempSignalProb = new TH1D(signalProbName, "signal probability", 10000, 0, 10000);
      signalProb_ = tempSignalProb;
      char backgroundProbName[50];
      sprintf(backgroundProbName, "backgroundProb_%d_%d", loopCounter, iorder);
      TH1D * tempBackgroundProb = new TH1D(backgroundProbName, "background probability", 10000, 0, 10000);
      backgroundProb_ = tempBackgroundProb;
// #endif



      // Before we start the minimization we create a list of events with only the events inside a smaller
      // window than the one in which the probability is != 0. We will compute the probability for all those
      // events and hopefully the margin will avoid them to get a probability = 0 (which causes a discontinuity
      // in the likelihood function). The width of this smaller window must be optimized, but we can start using
      // an 90% of the normalization window.
      double protectionFactor = 0.9;

      MuScleFitUtils::ReducedSavedPair.clear();
      for( unsigned int nev=0; nev<MuScleFitUtils::SavedPair.size(); ++nev ) {
        const lorentzVector * recMu1 = &(MuScleFitUtils::SavedPair[nev].first);
        const lorentzVector * recMu2 = &(MuScleFitUtils::SavedPair[nev].second);
        double mass = MuScleFitUtils::invDimuonMass( *recMu1, *recMu2 );
        // Test all resonances to see if the mass is inside at least one of the windows
        bool check = false;
        for( int ires = 0; ires < 6; ++ires ) {
      // std::pair<double, double> windowFactor = backgroundHandler->windowFactors( doBackgroundFit[loopCounter], ires );
      std::pair<double, double> windowBorder = backgroundHandler->windowBorders( doBackgroundFit[loopCounter], ires );
          // if( resfind[ires] && checkMassWindow( mass, ires, backgroundHandler->resMass( doBackgroundFit[loopCounter], ires ),
          //                                       0.9*windowFactor.first, 0.9*windowFactor.second ) ) {
      // double resMassValue = backgroundHandler->resMass( doBackgroundFit[loopCounter], ires );
      // double windowBorderLeft = resMassValue - protectionFactor*(resMassValue - windowBorder.first);
      // double windowBorderRight = resMassValue + protectionFactor*(windowBorder.second - resMassValue);
      double windowBorderShift = (windowBorder.second - windowBorder.first)*(1-protectionFactor)/2.;
      double windowBorderLeft = windowBorder.first + windowBorderShift;
      double windowBorderRight = windowBorder.second - windowBorderShift;
          if( resfind[ires] && checkMassWindow( mass, windowBorderLeft, windowBorderRight ) ) {
            check = true;
          }
        }
        if( check ) {
          MuScleFitUtils::ReducedSavedPair.push_back(std::make_pair(*recMu1, *recMu2));
        }
      }
      std::cout << "Fitting with " << MuScleFitUtils::ReducedSavedPair.size() << " events" << std::endl;


      // rmin.SetMaxIterations(500*parnumber);

      //Print some informations
      std::cout<<"MINUIT is starting the minimization for the iteration number "<<loopCounter<<std::endl;

      //Try to set iterations
      //      rmin.SetMaxIterations(100000);

      std::cout<<"maxNumberOfIterations (just set) = "<<rmin.GetMaxIterations()<<std::endl;

      MuScleFitUtils::normalizationChanged_ = 0;

      // Maximum number of iterations
      arglis[0] = 100000;
      // tolerance
      arglis[1] = 0.1;

      // Run simplex first to get an initial estimate of the minimum
      if( startWithSimplex_ ) {
    rmin.mnexcm( "SIMPLEX", arglis, 0, ierror );
      }

      rmin.mnexcm( "MIGRAD", arglis, 2, ierror );




// #ifdef DEBUG
      likelihoodInLoop_->Write();
      signalProb_->Write();
      backgroundProb_->Write();
      delete tempLikelihoodInLoop;
      delete tempSignalProb;
      delete tempBackgroundProb;
      likelihoodInLoop_ = 0;
      signalProb_ = 0;
      backgroundProb_ = 0;
// #endif


      // Compute again the error matrix
      rmin.mnexcm( "HESSE", arglis, 0, ierror );

      // Peform minos error analysis.
      if( computeMinosErrors_ ) {
    duringMinos_ = true;
    rmin.mnexcm( "MINOS", arglis, 0, ierror );
    duringMinos_ = false;
      }

      if( normalizationChanged_ > 1 ) {
        std::cout << "WARNING: normalization changed during fit meaning that events exited from the mass window. This causes a discontinuity in the likelihood function. Please check the scan of the likelihood as a function of the parameters to see if there are discontinuities around the minimum." << std::endl;
      }
    }

    // bool notWritten = true;
    for (int ipar=0; ipar<parnumber; ipar++) {

      rmin.mnpout (ipar, name, pval, erro, pmin, pmax, ivar);
      // Save parameters in parvalue[] vector
      // ------------------------------------
      if (ierror!=0 && debug>0) {
        std::cout << "[MuScleFitUtils-minimizeLikelihood]: ierror!=0, bogus pars" << std::endl;
      }
      //     for (int ipar=0; ipar<parnumber; ipar++) {
      //       rmin.mnpout (ipar, name, pval, erro, pmin, pmax, ivar);
      parvalue[loopCounter][ipar] = pval;
      //     }


      // int ilax2 = 0;
      // Double_t val2pl, val2mi;
      // rmin.mnmnot (ipar+1, ilax2, val2pl, val2mi);
      rmin.mnerrs (ipar, errh, errl, errp, cglo);


      // Set error on params
      // -------------------
      if (errp!=0) {
    parerr[3*ipar] = errp;
      } else {
    parerr[3*ipar] = (((errh)>(fabs(errl)))?(errh):(fabs(errl)));
      }
      parerr[3*ipar+1] = errl;
      parerr[3*ipar+2] = errh;

      if( ipar == 0 ) {
        FitParametersFile << " Resolution fit parameters:" << std::endl;
      }
      if( ipar == int(parResol.size()) ) {
        FitParametersFile << " Scale fit parameters:" << std::endl;
      }
      if( ipar == int(parResol.size()+parScale.size()) ) {
        FitParametersFile << " Cross section fit parameters:" << std::endl;
      }
      if( ipar == int(parResol.size()+parScale.size()+crossSectionHandler->parNum()) ) {
        FitParametersFile << " Background fit parameters:" << std::endl;
      }
//       if( ipar >= int(parResol.size()+parScale.size()) && ipar < int(parResol.size()+parScale.size()+crossSectionHandler->parNum()) && notWritted ) {

//         std::vector<double> relativeCrossSections = crossSectionHandler->relativeCrossSections(&(parvalue[loopCounter][parResol.size()+parScale.size()]));
//         std::vector<double>::const_iterator it = relativeCrossSections.begin();
//         for( ; it != relativeCrossSections.end(); ++it ) {
//           FitParametersFile << "  Results of the fit: parameter " << ipar << " has value "
//                             << *it << "+-" << 0
//                             << " + " << 0 << " - " << 0
//                             << " /t/t (" << 0 << ")" << std::endl;
//         }

//         notWritten = false;
//       }
//       else {
        FitParametersFile << "  Results of the fit: parameter " << ipar << " has value "
                          << pval << "+-" << parerr[3*ipar]
                          << " + " << parerr[3*ipar+1] << " - " << parerr[3*ipar+2]
                          // << " \t\t (" << parname[ipar] << ")"
                          << std::endl;



    }
    rmin.mnstat (fmin, fdem, errdef, npari, nparx, istat); // NNBB Commented for a check!
    FitParametersFile << std::endl;

    if( minimumShapePlots_ ) {
      // Create plots of the minimum vs parameters
      // -----------------------------------------
      // Keep this after the parameters filling because it recomputes the values and it can compromise the fit results.
      if( somethingtodo ) {
    std::stringstream iorderString;
    iorderString << iorder;
    std::stringstream iLoopString;
    iLoopString << loopCounter;

    for (int ipar=0; ipar<parnumber; ipar++) {
      if( parfix[ipar] == 1 ) continue;
      std::cout << "plotting parameter = " << ipar+1 << std::endl;
      std::stringstream iparString;
      iparString << ipar+1;
      std::stringstream iparStringName;
      iparStringName << ipar;
      rmin.mncomd( ("scan "+iparString.str()).c_str(), ierror );
      if( ierror == 0 ) {
        TCanvas * canvas = new TCanvas(("likelihoodCanvas_loop_"+iLoopString.str()+"_oder_"+iorderString.str()+"_par_"+iparStringName.str()).c_str(), ("likelihood_"+iparStringName.str()).c_str(), 1000, 800);
        canvas->cd();
        // arglis[0] = ipar;
        // rmin.mnexcm( "SCA", arglis, 0, ierror );
        TGraph * graph = (TGraph*)rmin.GetPlot();
        graph->Draw("AP");
        // graph->SetTitle(("parvalue["+iparStringName.str()+"]").c_str());
        graph->SetTitle(parname[ipar]);
        // graph->Write();

        canvas->Write();
      }
    }

    //       // Draw contours of the fit
    //       TCanvas * canvas = new TCanvas(("contourCanvas_oder_"+iorderString.str()).c_str(), "contour", 1000, 800);
    //       canvas->cd();
    //       TGraph * contourGraph = (TGraph*)rmin.Contour(4, 2, 4);
    //       if( (rmin.GetStatus() == 0) || (rmin.GetStatus() >= 3) ) {
    //         contourGraph->Draw("AP");
    //       }
    //       else {
    //         std::cout << "Contour graph error: status = " << rmin.GetStatus() << std::endl;
    //       }
    //       canvas->Write();
      }
    }

  } // end loop on iorder
  FitParametersFile.close();

  std::cout << "[MuScleFitUtils-minimizeLikelihood]: Parameters after likelihood " << std::endl;
  for (unsigned int ipar=0; ipar<(unsigned int)parnumber; ipar++) {
    std::cout << ipar << " " << parvalue[loopCounter][ipar] << " : free = "
     << parfix[ipar] << "; order = " << parorder[ipar] << std::endl;
  }

  // Put back parvalue into parResol, parScale, parCrossSection, parBgr
  // ------------------------------------------------------------------
  for( int i=0; i<(int)(parResol.size()); ++i ) {
    parResol[i] = parvalue[loopCounter][i];
  }
  for( int i=0; i<(int)(parScale.size()); ++i ) {
    parScale[i] = parvalue[loopCounter][i+parResol.size()];
  }
  parCrossSection = crossSectionHandler->relativeCrossSections(&(parvalue[loopCounter][parResol.size()+parScale.size()]), resfind);
  for( unsigned int i=0; i<parCrossSection.size(); ++i ) {
  //   parCrossSection[i] = parvalue[loopCounter][i+parResol.size()+parScale.size()];
    std::cout << "relative cross section["<<i<<"] = " << parCrossSection[i] << std::endl;
  }
  // Save only the fitted background parameters
  for( unsigned int i = 0; i<(parBgr.size() - parForResonanceWindows); ++i ) {
    parBgr[i] = parvalue[loopCounter][i+parResol.size()+parScale.size()+crossSectionHandler->parNum()];
  }

  // Background rescaling from regions to resonances
  // -----------------------------------------------
  // Only if we fitted the background
  if( doBackgroundFit[loopCounter] ) {
    // This rescales from regions to resonances
    int localMuonType = MuonType;
    if( MuonType > 2 ) localMuonType = 2;
    backgroundHandler->rescale( parBgr, ResMass, massWindowHalfWidth[localMuonType],
                                MuScleFitUtils::ReducedSavedPair );
  }

  // Delete the arrays used to set some parameters
  delete[] Start;
  delete[] Step;
  delete[] Mini;
  delete[] Maxi;
  delete[] ind;
  delete[] parname;
}

// Likelihood function
// -------------------
extern "C" void likelihood( int& npar, double* grad, double& fval, double* xval, int flag ) {

  if (MuScleFitUtils::debug>19) std::cout << "[MuScleFitUtils-likelihood]: In likelihood function" << std::endl;

  const lorentzVector * recMu1;
  const lorentzVector * recMu2;
  lorentzVector corrMu1;
  lorentzVector corrMu2;

  //   if (MuScleFitUtils::debug>19) {
  //     int parnumber = (int)(MuScleFitUtils::parResol.size()+MuScleFitUtils::parScale.size()+
  //                           MuScleFitUtils::parCrossSection.size()+MuScleFitUtils::parBgr.size());
  //     std::cout << "[MuScleFitUtils-likelihood]: Looping on tree with ";
  //     for (int ipar=0; ipar<parnumber; ipar++) {
  //       std::cout << "Parameter #" << ipar << " with value " << xval[ipar] << " ";
  //     }
  //     std::cout << std::endl;
  //   }

  // Loop on the tree
  // ----------------
  double flike = 0;
  int evtsinlik = 0;
  int evtsoutlik = 0;
  // std::cout << "SavedPair.size() = " << MuScleFitUtils::SavedPair.size() << std::endl;
  if( MuScleFitUtils::debug>0 ) {
    std::cout << "SavedPair.size() = " << MuScleFitUtils::SavedPair.size() << std::endl;
    std::cout << "ReducedSavedPair.size() = " << MuScleFitUtils::ReducedSavedPair.size() << std::endl;
  }
  // for( unsigned int nev=0; nev<MuScleFitUtils::SavedPair.size(); ++nev ) {
  for( unsigned int nev=0; nev<MuScleFitUtils::ReducedSavedPair.size(); ++nev ) {

    //     recMu1 = &(MuScleFitUtils::SavedPair[nev].first);
    //     recMu2 = &(MuScleFitUtils::SavedPair[nev].second);
    recMu1 = &(MuScleFitUtils::ReducedSavedPair[nev].first);
    recMu2 = &(MuScleFitUtils::ReducedSavedPair[nev].second);

    // Compute original mass
    // ---------------------
    double mass = MuScleFitUtils::invDimuonMass( *recMu1, *recMu2 );

    // Compute weight and reference mass (from original mass)
    // ------------------------------------------------------
    double weight = MuScleFitUtils::computeWeight(mass, MuScleFitUtils::iev_);
    if( weight!=0. ) {
      // Compute corrected mass (from previous biases) only if we are currently fitting the scale
      // ----------------------------------------------------------------------------------------
      if( MuScleFitUtils::doScaleFit[MuScleFitUtils::loopCounter] ) {
//  std::cout << "Original pt1 = " << corrMu1.Pt() << std::endl;
//  std::cout << "Original pt2 = " << corrMu2.Pt() << std::endl;
        corrMu1 = MuScleFitUtils::applyScale(*recMu1, xval, -1);
        corrMu2 = MuScleFitUtils::applyScale(*recMu2, xval,  1);

//         if( (corrMu1.Pt() != corrMu1.Pt()) || (corrMu2.Pt() != corrMu2.Pt()) ) {
//           std::cout << "Rescaled pt1 = " << corrMu1.Pt() << std::endl;
//           std::cout << "Rescaled pt2 = " << corrMu2.Pt() << std::endl;
//         }
//  std::cout << "Rescaled pt1 = " << corrMu1.Pt() << std::endl;
//  std::cout << "Rescaled pt2 = " << corrMu2.Pt() << std::endl;
      }
      else {
        corrMu1 = *recMu1;
        corrMu2 = *recMu2;

//         if( (corrMu1.Pt() != corrMu1.Pt()) || (corrMu2.Pt() != corrMu2.Pt()) ) {
//           std::cout << "Not rescaled pt1 = " << corrMu1.Pt() << std::endl;
//           std::cout << "Not rescaled pt2 = " << corrMu2.Pt() << std::endl;
//         }
      }
      double corrMass = MuScleFitUtils::invDimuonMass(corrMu1, corrMu2);
      double Y = (corrMu1+corrMu2).Rapidity();
      double resEta = (corrMu1+corrMu2).Eta();
      if( MuScleFitUtils::debug>19 ) {
    std::cout << "[MuScleFitUtils-likelihood]: Original/Corrected resonance mass = " << mass
         << " / " << corrMass << std::endl;
      }

      // Compute mass resolution
      // -----------------------
      double massResol = MuScleFitUtils::massResolution(corrMu1, corrMu2, xval);
      if (MuScleFitUtils::debug>19)
    std::cout << "[MuScleFitUtils-likelihood]: Resolution is " << massResol << std::endl;

      // Compute probability of this mass value including background modeling
      // --------------------------------------------------------------------
      if (MuScleFitUtils::debug>1) std::cout << "calling massProb inside likelihood function" << std::endl;

      // double prob = MuScleFitUtils::massProb( corrMass, resEta, Y, massResol, xval );
      double prob = MuScleFitUtils::massProb( corrMass, resEta, Y, massResol, xval, false, corrMu1.eta(), corrMu2.eta() );
      if (MuScleFitUtils::debug>1) std::cout << "likelihood:massProb = " << prob << std::endl;

      // Compute likelihood
      // ------------------
      if( prob>0 ) {
    // flike += log(prob*10000)*weight; // NNBB! x10000 to see if we can recover the problem of boundary
    flike += log(prob)*weight;
    evtsinlik += 1;  // NNBB test: see if likelihood per event is smarter (boundary problem)
      } else {
        if( MuScleFitUtils::debug > 0 ) {
          std::cout << "WARNING: corrMass = " << corrMass << " outside window, this will cause a discontinuity in the likelihood. Consider increasing the safety bands which are now set to 90% of the normalization window to avoid this problem" << std::endl;
          std::cout << "Original mass was = " << mass << std::endl;
      std::cout << "WARNING: massResol = " << massResol << " outside window" << std::endl;
        }
    evtsoutlik += 1;
      }
      if (MuScleFitUtils::debug>19)
    std::cout << "[MuScleFitUtils-likelihood]: Mass probability = " << prob << std::endl;
    } // weight!=0

  } // End of loop on tree events

//   // Protection for low statistic. If the likelihood manages to throw out all the signal
//   // events and stays with ~ 10 events in the resonance window it could have a better likelihood
//   // because of ~ uniformly distributed events (a random combination could be good and spoil the fit).
//   // We require that the number of events included in the fit does not change more than 5% in each minuit loop.
//   bool lowStatPenalty = false;
//   if( MuScleFitUtils::minuitLoop_ > 0 ) {
//     double newEventsOutInRatio = double(evtsinlik);
//     // double newEventsOutInRatio = double(evtsoutlik)/double(evtsinlik);
//     double ratio = newEventsOutInRatio/MuScleFitUtils::oldEventsOutInRatio_;
//     MuScleFitUtils::oldEventsOutInRatio_ = newEventsOutInRatio;
//     if( ratio < 0.8 || ratio > 1.2 ) {
//       std::cout << "Warning: too much change from oldEventsInLikelihood to newEventsInLikelihood, ratio is = " << ratio << std::endl;
//       std::cout << "oldEventsInLikelihood = " << MuScleFitUtils::oldEventsOutInRatio_ << ", newEventsInLikelihood = " << newEventsOutInRatio << std::endl;
//       lowStatPenalty = true;
//     }
//   }

  // It is a product of probabilities, we compare the sqrt_N of them. Thus N becomes a denominator of the logarithm.
  if( evtsinlik != 0 ) {

    if( MuScleFitUtils::normalizeLikelihoodByEventNumber_ ) {
      // && !(MuScleFitUtils::duringMinos_) ) {
      if( MuScleFitUtils::rminPtr_ == 0 ) {
        std::cout << "ERROR: rminPtr_ = " << MuScleFitUtils::rminPtr_ << ", code will crash" << std::endl;
      }
      double normalizationArg[] = {1/double(evtsinlik)};
      // Reset the normalizationArg only if it changed
      if( MuScleFitUtils::oldNormalization_ != normalizationArg[0] ) {
        int ierror = 0;
//         if( MuScleFitUtils::likelihoodInLoop_ != 0 ) {
//           // This condition is set only when minimizing. Later calls of hesse and minos will not change the value
//           // This is done to avoid minos being confused by changing the UP parameter during its computation.
//           MuScleFitUtils::rminPtr_->mnexcm("SET ERR", normalizationArg, 1, ierror);
//         }
        MuScleFitUtils::rminPtr_->mnexcm("SET ERR", normalizationArg, 1, ierror);
    std::cout << "oldNormalization = " << MuScleFitUtils::oldNormalization_ << " new = " << normalizationArg[0] << std::endl;
        MuScleFitUtils::oldNormalization_ = normalizationArg[0];
        MuScleFitUtils::normalizationChanged_ += 1;
      }
      fval = -2.*flike/double(evtsinlik);
      // fval = -2.*flike;
      //     if( lowStatPenalty ) {
      //       fval *= 100;
      //     }
    }
    else {
      fval = -2.*flike;
    }
  }
  else {
    std::cout << "Problem: Events in likelihood = " << evtsinlik << std::endl;
    fval = 999999999.;
  }
  // fval = -2.*flike;
  if (MuScleFitUtils::debug>19)
    std::cout << "[MuScleFitUtils-likelihood]: End tree loop with likelihood value = " << fval << std::endl;

//  #ifdef DEBUG

//  if( MuScleFitUtils::minuitLoop_ < 10000 ) {
  if( MuScleFitUtils::likelihoodInLoop_ != 0 ) {
    ++MuScleFitUtils::minuitLoop_;
    MuScleFitUtils::likelihoodInLoop_->SetBinContent(MuScleFitUtils::minuitLoop_, fval);
  }
  //  }
  // else std::cout << "minuitLoop over 10000. Not filling histogram" << std::endl;

  std::cout<<"MINUIT loop number "<<MuScleFitUtils::minuitLoop_<<", likelihood = "<<fval<<std::endl;

  if( MuScleFitUtils::debug > 0 ) {
    //     if( MuScleFitUtils::duringMinos_ ) {
    //       int parnumber = (int)(MuScleFitUtils::parResol.size()+MuScleFitUtils::parScale.size()+
    //                             MuScleFitUtils::parCrossSection.size()+MuScleFitUtils::parBgr.size());
    //       std::cout << "[MuScleFitUtils-likelihood]: Looping on tree with ";
    //       for (int ipar=0; ipar<parnumber; ipar++) {
    //         std::cout << "Parameter #" << ipar << " with value " << xval[ipar] << " ";
    //       }
    //       std::cout << std::endl;
    //       std::cout << "[MuScleFitUtils-likelihood]: likelihood value = " << fval << std::endl;
    //     }
    std::cout << "Events in likelihood = " << evtsinlik << std::endl;
    std::cout << "Events out likelihood = " << evtsoutlik << std::endl;
  }

//  #endif
}

// Mass fitting routine
// --------------------
std::vector<TGraphErrors*> MuScleFitUtils::fitMass (TH2F* histo) {

  if (MuScleFitUtils::debug>0) std::cout << "Fitting " << histo->GetName() << std::endl;

  std::vector<TGraphErrors *> results;

  // Results of the fit
  // ------------------
  std::vector<double> Ftop;
  std::vector<double> Fwidth;
  std::vector<double> Fmass;
  std::vector<double> Etop;
  std::vector<double> Ewidth;
  std::vector<double> Emass;
  std::vector<double> Fchi2;
  // X bin center and width
  // ----------------------
  std::vector<double> Xcenter;
  std::vector<double> Ex;

  // Fit with lorentzian peak
  // ------------------------
  TF1 *fitFcn = new TF1 ("fitFcn", lorentzianPeak, 70, 110, 3);
  fitFcn->SetParameters (100, 3, 91);
  fitFcn->SetParNames ("Ftop", "Fwidth", "Fmass");
  fitFcn->SetLineWidth (2);

  // Fit slices projected along Y from bins in X
  // -------------------------------------------
  double cont_min = 20;    // Minimum number of entries
  Int_t binx = histo->GetXaxis()->GetNbins();
  // TFile *f= new TFile("prova.root", "recreate");
  // histo->Write();
  for (int i=1; i<=binx; i++) {
    TH1 * histoY = histo->ProjectionY ("", i, i);
    // histoY->Write();
    double cont = histoY->GetEntries();
    if (cont>cont_min) {
      histoY->Fit ("fitFcn", "0", "", 70, 110);
      double *par = fitFcn->GetParameters();
      double *err = fitFcn->GetParErrors();

      Ftop.push_back(par[0]);
      Fwidth.push_back(par[1]);
      Fmass.push_back(par[2]);
      Etop.push_back(err[0]);
      Ewidth.push_back(err[1]);
      Emass.push_back(err[2]);

      double chi2 = fitFcn->GetChisquare();
      Fchi2.push_back(chi2);

      double xx = histo->GetXaxis()->GetBinCenter(i);
      Xcenter.push_back(xx);
      double ex = 0; // FIXME: you can use the bin width
      Ex.push_back(ex);
    }
  }
  // f->Close();

  // Put the fit results in arrays for TGraphErrors
  // ----------------------------------------------
  const int nn = Fmass.size();
  double *x    = new double[nn];
  double *ym   = new double[nn];
  double *e    = new double[nn];
  double *eym  = new double[nn];
  double *yw   = new double[nn];
  double *eyw  = new double[nn];
  double *yc   = new double[nn];

  for (int j=0; j<nn; j++) {
    x[j]   = Xcenter[j];
    ym[j]  = Fmass[j];
    eym[j] = Emass[j];
    yw[j]  = Fwidth[j];
    eyw[j] = Ewidth[j];
    yc[j]  = Fchi2[j];
    e[j]   = Ex[j];
  }

  // Create TGraphErrors
  // -------------------
  TString name = histo->GetName();
  TGraphErrors *grM = new TGraphErrors (nn, x, ym, e, eym);
  grM->SetTitle (name+"_M");
  grM->SetName (name+"_M");
  TGraphErrors *grW = new TGraphErrors (nn, x, yw, e, eyw);
  grW->SetTitle (name+"_W");
  grW->SetName (name+"_W");
  TGraphErrors *grC = new TGraphErrors (nn, x, yc, e, e);
  grC->SetTitle (name+"_chi2");
  grC->SetName (name+"_chi2");

  // Cleanup
  // -------
  delete[] x;
  delete[] ym;
  delete[] eym;
  delete[] yw;
  delete[] eyw;
  delete[] yc;
  delete[] e;
  delete fitFcn;

  results.push_back(grM);
  results.push_back(grW);
  results.push_back(grC);
  return results;
}

// Resolution fitting routine
// --------------------------
std::vector<TGraphErrors*> MuScleFitUtils::fitReso (TH2F* histo) {
  std::cout << "Fitting " << histo->GetName() << std::endl;
  std::vector<TGraphErrors *> results;

  // Results from fit
  // ----------------
  std::vector<double> maxs;
  std::vector<double> means;
  std::vector<double> sigmas;
  std::vector<double> chi2s;
  std::vector<double> Emaxs;
  std::vector<double> Emeans;
  std::vector<double> Esigmas;

  // X bin center and width
  // ----------------------
  std::vector<double> Xcenter;
  std::vector<double> Ex;

  // Fit with a gaussian
  // -------------------
  TF1 *fitFcn = new TF1 ("fitFunc", Gaussian, -0.2, 0.2, 3);
  fitFcn->SetParameters (100, 0, 0.02);
  fitFcn->SetParNames ("max", "mean", "sigma");
  fitFcn->SetLineWidth (2);

  // Fit slices projected along Y from bins in X
  // -------------------------------------------
  double cont_min = 20;    // Minimum number of entries
  Int_t binx = histo->GetXaxis()->GetNbins();
  for (int i=1; i<=binx; i++) {
    TH1 * histoY = histo->ProjectionY ("", i, i);
    double cont = histoY->GetEntries();
    if (cont>cont_min) {
      histoY->Fit ("fitFunc", "0", "", -0.2, 0.2);
      double *par = fitFcn->GetParameters();
      double *err = fitFcn->GetParErrors();

      maxs.push_back (par[0]);
      means.push_back (par[1]);
      sigmas.push_back (par[2]);
      Emaxs.push_back (err[0]);
      Emeans.push_back (err[1]);
      Esigmas.push_back (err[2]);

      double chi2 = fitFcn->GetChisquare();
      chi2s.push_back (chi2);

      double xx = histo->GetXaxis()->GetBinCenter(i);
      Xcenter.push_back (xx);
      double ex = 0; // FIXME: you can use the bin width
      Ex.push_back (ex);
    }
  }

  // Put the fit results in arrays for TGraphErrors
  // ----------------------------------------------
  const int nn = means.size();
  double *x    = new double[nn];
  double *ym   = new double[nn];
  double *e    = new double[nn];
  double *eym  = new double[nn];
  double *yw   = new double[nn];
  double *eyw  = new double[nn];
  double *yc   = new double[nn];

  for (int j=0; j<nn; j++) {
    x[j]   = Xcenter[j];
    ym[j]  = means[j];
    eym[j] = Emeans[j];
    // yw[j]  = maxs[j];
    // eyw[j] = Emaxs[j];
    yw[j]  = sigmas[j];
    eyw[j] = Esigmas[j];
    yc[j]  = chi2s[j];
    e[j]   = Ex[j];
  }

  // Create TGraphErrors
  // -------------------
  TString name = histo->GetName();
  TGraphErrors *grM = new TGraphErrors (nn, x, ym, e, eym);
  grM->SetTitle (name+"_mean");
  grM->SetName (name+"_mean");
  TGraphErrors *grW = new TGraphErrors (nn, x, yw, e, eyw);
  grW->SetTitle (name+"_sigma");
  grW->SetName (name+"_sigma");
  TGraphErrors *grC = new TGraphErrors (nn, x, yc, e, e);
  grC->SetTitle (name+"_chi2");
  grC->SetName (name+"_chi2");

  // Cleanup
  // -------
  delete[] x;
  delete[] ym;
  delete[] eym;
  delete[] yw;
  delete[] eyw;
  delete[] yc;
  delete[] e;
  delete fitFcn;

  results.push_back (grM);
  results.push_back (grW);
  results.push_back (grC);
  return results;
}

// Mass probability for likelihood computation - no-background version (not used anymore)
// --------------------------------------------------------------------------------------
double MuScleFitUtils::massProb( const double & mass, const double & rapidity, const int ires, const double & massResol )
{
  // This routine computes the likelihood that a given measured mass "measMass" is
  // the result of resonance #ires if the resolution expected for the two muons is massResol
  // ---------------------------------------------------------------------------------------

  double P = 0.;

  // Return Lorentz value for zero resolution cases (like MC)
  // --------------------------------------------------------
  if (massResol==0.) {
    if (debug>9) std::cout <<  "Mass, gamma , mref, width, P: " << mass
              << " " << ResGamma[ires] << " " << ResMass[ires]<< " " << massResol
              << " : used Lorentz P-value" << std::endl;
    return (0.5*ResGamma[ires]/TMath::Pi())/((mass-ResMass[ires])*(mass-ResMass[ires])+
                         .25*ResGamma[ires]*ResGamma[ires]);
  }

  // NB defined as below, P is not a "probability" but a likelihood that we observe
  // a dimuon mass "mass", given mRef, gamma, and massResol. It is what we need for the
  // fit which finds the best resolution parameters, though. A definition which is
  // more properly a probability is given below (in massProb2()), where the result
  // cannot be used to fit resolution parameters because the fit would always prefer
  // to set the res parameters to the minimum possible value (best resolution),
  // to have a probability close to one of observing any mass.
  // -------------------------------------------------------------------------------
  // NNBB the following two lines have been replaced with the six following them,
  // which provide an improvement of a factor 9 in speed of execution at a
  // negligible price in precision.
  // ----------------------------------------------------------------------------
  // GL->SetParameters(gamma,mRef,mass,massResol);
  // P = GL->Integral(mass-5*massResol, mass+5*massResol);

  Int_t np = 100;
  double * x = new double[np];
  double * w = new double[np];
  GL->SetParameters (ResGamma[ires], ResMass[ires], mass, massResol);
  GL->CalcGaussLegendreSamplingPoints (np, x, w, 0.1e-15);
  P = GL->IntegralFast (np, x, w, ResMass[ires]-10*ResGamma[ires], ResMass[ires]+10*ResGamma[ires]);
  delete[] x;
  delete[] w;

  // If we are too far away we set P to epsilon and forget about this event
  // ----------------------------------------------------------------------
  if (P<1.0e-12) {
    P = 1.0e-12;
    if (debug>9) std::cout << "Mass, gamma , mref, width, P: " << mass
              << " " << ResGamma[ires] << " " << ResMass[ires] << " " << massResol
              << ": used epsilon" << std::endl;
    return P;
  }

  if (debug>9) std::cout << "Mass, gamma , mref, width, P: " << mass
            << " " << ResGamma[ires] << " " << ResMass[ires] << " " << massResol
            << " " << P << std::endl;
  return P;
}

std::pair<lorentzVector, lorentzVector> MuScleFitUtils::findSimMuFromRes( const edm::Handle<edm::HepMCProduct> & evtMC,
                                      const edm::Handle<edm::SimTrackContainer> & simTracks )
{
  //Loop on simulated tracks
  std::pair<lorentzVector, lorentzVector> simMuFromRes;
  for( edm::SimTrackContainer::const_iterator simTrack=simTracks->begin(); simTrack!=simTracks->end(); ++simTrack ) {
    //Chose muons
    if (fabs((*simTrack).type())==13) {
      //If tracks from IP than find mother
      if ((*simTrack).genpartIndex()>0) {
    HepMC::GenParticle* gp = evtMC->GetEvent()->barcode_to_particle ((*simTrack).genpartIndex());
        if( gp != 0 ) {

          for (HepMC::GenVertex::particle_iterator mother = gp->production_vertex()->particles_begin(HepMC::ancestors);
               mother!=gp->production_vertex()->particles_end(HepMC::ancestors); ++mother) {

            bool fromRes = false;
            unsigned int motherPdgId = (*mother)->pdg_id();
            for( int ires = 0; ires < 6; ++ires ) {
              if( motherPdgId == motherPdgIdArray[ires] && resfind[ires] ) fromRes = true;
            }
            if( fromRes ) {
              if(gp->pdg_id() == 13)
                simMuFromRes.first = lorentzVector(simTrack->momentum().px(),simTrack->momentum().py(),
                                                   simTrack->momentum().pz(),simTrack->momentum().e());
              else
                simMuFromRes.second = lorentzVector(simTrack->momentum().px(),simTrack->momentum().py(),
                                                    simTrack->momentum().pz(),simTrack->momentum().e());
            }
          }
        }
        // else LogDebug("MuScleFitUtils") << "WARNING: no matching genParticle found for simTrack" << std::endl;
      }
    }
  }
  return simMuFromRes;
}

std::pair<lorentzVector, lorentzVector> MuScleFitUtils::findGenMuFromRes( const edm::HepMCProduct* evtMC )
{
  const HepMC::GenEvent* Evt = evtMC->GetEvent();
  std::pair<lorentzVector,lorentzVector> muFromRes;
  //Loop on generated particles
  for (HepMC::GenEvent::particle_const_iterator part=Evt->particles_begin();
       part!=Evt->particles_end(); part++) {
    if (fabs((*part)->pdg_id())==13 && (*part)->status()==1) {
      bool fromRes = false;
      for (HepMC::GenVertex::particle_iterator mother = (*part)->production_vertex()->particles_begin(HepMC::ancestors);
       mother != (*part)->production_vertex()->particles_end(HepMC::ancestors); ++mother) {
        unsigned int motherPdgId = (*mother)->pdg_id();

        // For sherpa the resonance is not saved. The muons from the resonance can be identified
        // by having as mother a muon of status 3.
        if( sherpa_ ) {
          if( motherPdgId == 13 && (*mother)->status() == 3 ) fromRes = true;
        }
        else {
          for( int ires = 0; ires < 6; ++ires ) {
        if( motherPdgId == motherPdgIdArray[ires] && resfind[ires] ) fromRes = true;
          }
        }
      }
      if(fromRes){
    if((*part)->pdg_id()==13)
      //   muFromRes.first = (*part)->momentum();
      muFromRes.first = (lorentzVector((*part)->momentum().px(),(*part)->momentum().py(),
                       (*part)->momentum().pz(),(*part)->momentum().e()));
    else
      muFromRes.second = (lorentzVector((*part)->momentum().px(),(*part)->momentum().py(),
                        (*part)->momentum().pz(),(*part)->momentum().e()));
      }
    }
  }
  return muFromRes;
}

std::pair<lorentzVector, lorentzVector> MuScleFitUtils::findGenMuFromRes( const reco::GenParticleCollection* genParticles)
{
  std::pair<lorentzVector,lorentzVector> muFromRes;

  //Loop on generated particles
  if( debug>0 ) std::cout << "Starting loop on " << genParticles->size() << " genParticles" << std::endl;
  for( reco::GenParticleCollection::const_iterator part=genParticles->begin(); part!=genParticles->end(); ++part ) {
    if (fabs(part->pdgId())==13 && part->status()==1) {
      bool fromRes = false;
      unsigned int motherPdgId = part->mother()->pdgId();
      if( debug>0 ) {
    std::cout << "Found a muon with mother: " << motherPdgId << std::endl;
      }
      for( int ires = 0; ires < 6; ++ires ) {
    if( motherPdgId == motherPdgIdArray[ires] && resfind[ires] ) fromRes = true;
      }
      if(fromRes){
    if(part->pdgId()==13) {
      muFromRes.first = part->p4();
      if( debug>0 ) std::cout << "Found a genMuon + : " << muFromRes.first << std::endl;
      //      muFromRes.first = (lorentzVector(part->p4().px(),part->p4().py(),
      //                       part->p4().pz(),part->p4().e()));
    }
    else {
      muFromRes.second = part->p4();
      if( debug>0 ) std::cout << "Found a genMuon - : " << muFromRes.second << std::endl;
      //      muFromRes.second = (lorentzVector(part->p4().px(),part->p4().py(),
      //                        part->p4().pz(),part->p4().e()));
    }
      }
    }
  }
  return muFromRes;
}