Newer
Older
TB_Chris / TbAlignment / src / TbMillepede.cpp
#include <iomanip>
#include <algorithm>
#include <cmath>

// Local
#include "TbMillepede.h"

// TbKernel
#include "TbKernel/TbGeomFunctions.h"

DECLARE_TOOL_FACTORY(TbMillepede)

//=============================================================================
// Standard constructor, initializes variables
//=============================================================================
TbMillepede::TbMillepede(const std::string& type, const std::string& name,
                         const IInterface* parent)
    : TbAlignmentBase(type, name, parent),
      m_nalc(4),
      m_debug(false),
      m_iterate(true) {

  declareProperty("NIterations", m_nIterations = 5);
  declareProperty("ResCut", m_rescut = 0.05);
  declareProperty("ResCutInit", m_rescut_init = 0.6);
  declareProperty("NStdDev", m_nstdev = 0);
  declareProperty("Sigmas", m_sigmas = {});

  m_dofsDefault = {true, true, false, true, true, true};
}

//=============================================================================
// Destructor
//=============================================================================
TbMillepede::~TbMillepede() {}

//=============================================================================
// Initialization
//=============================================================================
StatusCode TbMillepede::initialize() {

  // Initialise the base class.
  StatusCode sc = TbAlignmentBase::initialize();
  if (sc.isFailure()) return sc;

  // Get the verbosity level.
  m_debug = msgLevel(MSG::DEBUG);

  // Renumber the planes in Millepede, ignoring masked planes.
  m_millePlanes.assign(m_nPlanes, 999);
  unsigned int index = 0;
  for (unsigned int i = 0; i < m_nPlanes; ++i) {
    if (masked(i)) continue;
    m_millePlanes[i] = index;
    ++index;
  }
  // Use default values for the sigmas, unless specified explicitly.
  if (m_sigmas.size() != 6) {
    m_sigmas = {0.05, 0.05, 0.5, 0.005, 0.005, 0.005};
  }
  return StatusCode::SUCCESS;
}

//=============================================================================
// Main alignment function
//=============================================================================
void TbMillepede::align(std::vector<TbAlignmentTrack*>& alignmentTracks) {

  info() << "Millepede alignment" << endmsg;
  const unsigned int nPlanes = m_nPlanes - m_maskedPlanes.size();
  const unsigned int nParameters = 6 * nPlanes;
  for (unsigned int iteration = 0; iteration < m_nIterations; ++iteration) {
    const unsigned int nTracks = alignmentTracks.size();
    // Define the constraint equations.
    setConstraints(nPlanes);
    const double startfact = 100.;
    // Initialise all matrices and vectors.
    reset(nPlanes, startfact);
    info() << "Feeding Millepede with " << nTracks << " tracks..." << endmsg;
    // Feed Millepede with tracks.
    unsigned int nSkipped = 0;
    unsigned int nOutliers = 0;
    for (unsigned int i = 0; i < nTracks; ++i) {
      LHCb::TbTrack* track = alignmentTracks[i]->track();
      if (track->size() != nPlanes) {
        ++nSkipped;
        continue;
      }
      if (!putTrack(track, nPlanes)) {
        ++nOutliers;
      }
    }
    if (nSkipped > 0) {
      info() << "Skipped " << nSkipped << " tracks with less than " << nPlanes
             << " clusters." << endmsg;
    }
    if (nOutliers > 0) {
      info() << "Rejected " << nOutliers << " outlier tracks." << endmsg;
    }
    // Do the global fit.
    info() << "Determining global parameters..." << endmsg;
    fitGlobal();
    // Calculate the convergence metric (sum of misalignments).
    double converg = 0.;
    for (unsigned int i = 0; i < nParameters; ++i) {
      converg += fabs(m_dparm[i]);
    }
    converg /= nParameters;
    info() << "Convergence: " << converg << endmsg;
    // Update the module positions and orientations.
    info() << "Updating geometry..." << endmsg;
    updateGeometry();
    // Update the cluster coordinates based on the new geometry.
    for (unsigned int i = 0; i < nTracks; ++i) {
      LHCb::TbTrack* track = alignmentTracks[i]->track();
      SmartRefVector<LHCb::TbCluster> clusters = track->clusters();
      for (auto ic = clusters.begin(), end = clusters.end(); ic != end; ++ic) {
        Gaudi::XYZPoint pLocal((*ic)->xloc(), (*ic)->yloc(), 0.);
        const auto pGlobal = geomSvc()->localToGlobal(pLocal, (*ic)->plane());
        (*ic)->setX(pGlobal.x());
        (*ic)->setY(pGlobal.y());
        (*ic)->setZ(pGlobal.z());
      }
    }
    if (converg < 0.00001) break;
  }
}

//=============================================================================
// Setup the constraint equations.
//=============================================================================
void TbMillepede::setConstraints(const unsigned int nPlanes) {

  // Calculate the mean z-position.
  double avgz = 0.;
  for (auto im = m_modules.cbegin(), end = m_modules.cend(); im != end; ++im) {
    if (masked(im - m_modules.cbegin())) continue;
    avgz += (*im)->z();
  }
  avgz /= nPlanes;
  // Calculate the variance.
  double varz = 0.0;
  for (auto im = m_modules.cbegin(), end = m_modules.cend(); im != end; ++im) {
    if (masked(im - m_modules.cbegin())) continue;
    const double dz = (*im)->z() - avgz;
    varz += dz * dz;
  }
  varz /= nPlanes;

  // Define the 9 constraints equations according to the requested geometry.
  const unsigned int nParameters = 6 * nPlanes;
  std::vector<double> ftx(nParameters, 0.);
  std::vector<double> fty(nParameters, 0.);
  std::vector<double> ftz(nParameters, 0.);
  std::vector<double> frx(nParameters, 0.);
  std::vector<double> fry(nParameters, 0.);
  std::vector<double> frz(nParameters, 0.);
  std::vector<double> fscaz(nParameters, 0.);
  std::vector<double> shearx(nParameters, 0.);
  std::vector<double> sheary(nParameters, 0.);

  m_constraints.clear();
  for (auto im = m_modules.cbegin(), end = m_modules.cend(); im != end; ++im) {
    if (masked(im - m_modules.begin())) continue;
    const unsigned int i = m_millePlanes[im - m_modules.begin()];
    const double sz = ((*im)->z() - avgz) / varz;
    ftx[i] = 1.0;
    fty[i + nPlanes] = 1.0;
    ftz[i + 2 * nPlanes] = 1.0;
    frx[i + 3 * nPlanes] = 1.0 ; // i > 4 ? 1.0 : -1.0;
    fry[i + 4 * nPlanes] = 1.0 ; //  i > 4 ? 1.0 : -1.0;
    frz[i + 5 * nPlanes] = 1.0;
    shearx[i] = sz;
    sheary[i + nPlanes] = sz;
    fscaz[i + 2 * nPlanes] = sz;
  }

  const std::vector<bool> constraints = {true,  true, true, true, false,
                                         false, true, true, true};
  //  Put the constraints information in the basket.
  if (constraints[0] && m_dofs[0]) addConstraint(ftx, 0.0);
  if (constraints[1] && m_dofs[0]) addConstraint(shearx, 0.);
  if (constraints[2] && m_dofs[1]) addConstraint(fty, 0.0);
  if (constraints[3] && m_dofs[1]) addConstraint(sheary, 0.);
  if (constraints[4] && m_dofs[2]) addConstraint(ftz, 0.0);
  // if (constraints[5] && m_dofs[2]) addConstraint(fscaz, 0.0);
  if (constraints[6] && m_dofs[3]) addConstraint(frx, 0.0);
  if (constraints[7] && m_dofs[4]) addConstraint(fry, 0.0);
  if (constraints[8] && m_dofs[5]) addConstraint(frz, 0.0);
}

//=============================================================================
// Define a single constraint equation.
//=============================================================================
void TbMillepede::addConstraint(const std::vector<double>& dercs,
                                const double rhs) {

  Constraint constraint;
  // Set the right-hand side (Lagrange multiplier value, sum of equation).
  constraint.rhs = rhs;
  // Set the constraint equation coefficients.
  constraint.coefficients = dercs;
  m_constraints.push_back(constraint);
}

//=============================================================================
// Add the equations for one track to the matrix
//=============================================================================
bool TbMillepede::putTrack(LHCb::TbTrack* track, const unsigned int nPlanes) {

  std::vector<Equation> equations;
  const unsigned int nParameters = 6 * nPlanes;
  // Global derivatives
  std::vector<double> dergb(nParameters, 0.);
  // Global derivatives non linearly related to residual
  std::vector<double> dernl(nParameters, 0.);
  // Local parameter indices associated with non-linear derivatives
  std::vector<int> dernli(nParameters, 0);
  // Track slopes
  std::vector<double> dernls(nParameters, 0.);

  /// Refit the track for the reference states.
  m_trackFit->fit(track);
  auto state = track->firstState();
  const double tx = state.tx();
  const double ty = state.ty();

  // Iterate over each cluster on the track.
  const SmartRefVector<LHCb::TbCluster>& clusters = track->clusters();
  for (auto itc = clusters.cbegin(), end = clusters.cend(); itc != end; ++itc) {
    if (masked((*itc)->plane())) continue;
    const auto normal = geomSvc()->module((*itc)->plane())->normal();
    double nx = normal.x() / normal.z();
    double ny = normal.y() / normal.z();
    const double xg = (*itc)->x();
    const double yg = (*itc)->y();
    const double zg = (*itc)->z();
    // Calculate quasi-local coordinates.
    const double zl = zg - m_modules[(*itc)->plane()]->z();
    const double xl = xg - m_modules[(*itc)->plane()]->x();
    const double yl = yg - m_modules[(*itc)->plane()]->y();

    std::vector<double> nonlinear = {
        nx, ny, 1., -yl + zl * ny, -xl + zl * nx, xl * ny - yl * nx};
    const double den = 1. + tx * nx + ty * ny;
    for (auto& a : nonlinear) a /= den;
    // Get the errors on the measured x and y coordinates.
    const double errx = (*itc)->xErr();
    const double erry = (*itc)->yErr();
    // Get the internal plane index in Millepede.
    const unsigned int plane = m_millePlanes[(*itc)->plane()];
    // Set the local derivatives for the X equation.
    std::vector<double> derlc = {1., zg, 0., 0.};
    // Set the global derivatives (see LHCb-2005-101) for the X equation.
    std::fill(dergb.begin(), dergb.end(), 0.);
    std::fill(dernl.begin(), dernl.end(), 0.);
    std::fill(dernli.begin(), dernli.end(), 0);
    std::fill(dernls.begin(), dernls.end(), 0.);
    if (m_dofs[0]) dergb[plane] = -1.;
    if (m_dofs[4]) dergb[4 * nPlanes + plane] = -zl;
    if (m_dofs[5]) dergb[5 * nPlanes + plane] = yl;
    for (unsigned int i = 0; i < 6; ++i) {
      if (!m_dofs[i]) continue;
      dergb[i * nPlanes + plane] += tx * nonlinear[i];
      dernl[i * nPlanes + plane] = nonlinear[i];
      dernli[i * nPlanes + plane] = 1;
      dernls[i * nPlanes + plane] = tx;
    }
    // Store the X equation.
    addEquation(equations, derlc, dergb, dernl, dernli, dernls, xg, errx);
    // Set the local derivatives for the Y equation.
    derlc = {0., 0., 1., zg};
    // Set the global derivatives (see LHCb-2005-101) for the Y equation.
    std::fill(dergb.begin(), dergb.end(), 0.);
    std::fill(dernl.begin(), dernl.end(), 0.);
    std::fill(dernli.begin(), dernli.end(), 0);
    std::fill(dernls.begin(), dernls.end(), 0.);
    if (m_dofs[1]) dergb[nPlanes + plane] = -1.;
    if (m_dofs[3]) dergb[3 * nPlanes + plane] = -zl;
    if (m_dofs[5]) dergb[5 * nPlanes + plane] = -xl;
    for (unsigned int i = 0; i < 6; ++i) {
      if (!m_dofs[i]) continue;
      dergb[i * nPlanes + plane] += ty * nonlinear[i];
      dernl[i * nPlanes + plane] = nonlinear[i];
      dernli[i * nPlanes + plane] = 3;
      dernls[i * nPlanes + plane] = ty;
    }
    // Store the Y equation.
    addEquation(equations, derlc, dergb, dernl, dernli, dernls, yg, erry);
  }
  // Vector containing the track parameters
  std::vector<double> trackParams(2 * m_nalc + 2, 0.);
  // Fit the track.
  const unsigned int iteration = 1;
  const bool ok = fitTrack(equations, trackParams, false, iteration);
  if (ok) m_equations.push_back(equations);
  return ok;
}

//=============================================================================
// Store the parameters for one measurement
//=============================================================================
void TbMillepede::addEquation(std::vector<Equation>& equations,
                              const std::vector<double>& derlc,
                              const std::vector<double>& dergb,
                              const std::vector<double>& dernl,
                              const std::vector<int>& dernli,
                              const std::vector<double>& slopes,
                              const double rmeas, const double sigma) {

  if (sigma <= 0.) {
    error() << "Invalid cluster error (" << sigma << ")" << endmsg;
    return;
  }
  Equation equation;
  equation.rmeas = rmeas;
  equation.weight = 1. / (sigma * sigma);
  // Add non-zero local derivatives and corresponding indices.
  equation.derL.clear();
  equation.indL.clear();
  for (unsigned int i = 0; i < m_nalc; ++i) {
    if (derlc[i] == 0.) continue;
    equation.indL.push_back(i);
    equation.derL.push_back(derlc[i]);
  }
  // Add non-zero global derivatives and corresponding indices.
  equation.derG.clear();
  equation.indG.clear();
  equation.derNL.clear();
  equation.indNL.clear();
  equation.slopes.clear();
  const unsigned int nG = dergb.size();
  for (unsigned int i = 0; i < nG; ++i) {
    if (dergb[i] == 0.) continue;
    equation.indG.push_back(i);
    equation.derG.push_back(dergb[i]);
    equation.derNL.push_back(dernl[i]);
    equation.indNL.push_back(dernli[i]);
    equation.slopes.push_back(slopes[i]);
  }
  // Add the equation to the list.
  equations.push_back(std::move(equation));
}

//=============================================================================
// Track fit (local fit)
//=============================================================================
bool TbMillepede::fitTrack(const std::vector<Equation>& equations,
                           std::vector<double>& trackParams,
                           const bool singlefit, const unsigned int iteration) {

  std::vector<double> blvec(m_nalc, 0.);
  std::vector<std::vector<double> > clmat(m_nalc,
                                          std::vector<double>(m_nalc, 0.));

  // First loop: local track fit
  for (const auto& equation : equations) {
    double rmeas = equation.rmeas;
    // Suppress the global part (only relevant with iterations).
    const unsigned int nG = equation.derG.size();
    for (unsigned int i = 0; i < nG; ++i) {
      const unsigned int j = equation.indG[i];
      rmeas -= equation.derG[i] * m_dparm[j];
    }
    const double w = equation.weight;
    // Fill local matrix and vector.
    const unsigned int nL = equation.derL.size();
    for (unsigned int i = 0; i < nL; ++i) {
      const unsigned int j = equation.indL[i];
      blvec[j] += w * rmeas * equation.derL[i];
      if (m_debug) debug() << "blvec[" << j << "] = " << blvec[j] << endmsg;
      // Symmetric matrix, don't bother k > j coefficients.
      for (unsigned int k = 0; k <= i; ++k) {
        const unsigned int ik = equation.indL[k];
        clmat[j][ik] += w * equation.derL[i] * equation.derL[k];
        if (m_debug) {
          debug() << "clmat[" << j << "][" << ik << "] = " << clmat[j][ik]
                  << endmsg;
        }
      }
    }
  }

  // Local parameter matrix is completed, now invert to solve.
  // Rank: number of non-zero diagonal elements.
  int rank = invertMatrixLocal(clmat, blvec, m_nalc);
  // Store the track parameters and errors.
  for (unsigned int i = 0; i < m_nalc; ++i) {
    trackParams[2 * i] = blvec[i];
    trackParams[2 * i + 1] = sqrt(fabs(clmat[i][i]));
  }

  // Second loop: residual calculation.
  double chi2 = 0.0;
  int ndf = 0;
  for (const auto& equation : equations) {
    double rmeas = equation.rmeas;
    // Suppress global and local terms.
    const unsigned int nL = equation.derL.size();
    for (unsigned int i = 0; i < nL; ++i) {
      const unsigned int j = equation.indL[i];
      rmeas -= equation.derL[i] * blvec[j];
    }
    const unsigned int nG = equation.derG.size();
    for (unsigned int i = 0; i < nG; ++i) {
      const unsigned int j = equation.indG[i];
      rmeas -= equation.derG[i] * m_dparm[j];
    }
    // Now rmeas contains the residual value.
    if (m_debug) debug() << "Residual value: " << rmeas << endmsg;
    // Reject the track if the residual is too large (outlier).
    const double rescut = iteration <= 1 ? m_rescut_init : m_rescut;
    if (fabs(rmeas) >= rescut) {
      if (m_debug) {
        debug() << "Reject track due to residual cut in iteration " << iteration
                << endmsg;
      }
      return false;
    }
    chi2 += equation.weight * rmeas * rmeas;
    ++ndf;
  }
  ndf -= rank;
  const bool printDetails = false;
  if (printDetails) {
    info() << endmsg;
    info() << "Local track fit (rank " << rank << ")" << endmsg;
    info() << " Result of local fit :      (index/parameter/error)" << endmsg;
    for (unsigned int i = 0; i < m_nalc; ++i) {
      info() << std::setprecision(6) << std::fixed << std::setw(20) << i
             << "   /   " << std::setw(10) << blvec[i] << "   /   "
             << sqrt(clmat[i][i]) << endmsg;
    }
    info() << "Final chi square / degrees of freedom: " << chi2 << " / " << ndf
           << endmsg;
  }

  // Stop here if just updating the track parameters.
  if (singlefit) return true;

  if (m_nstdev != 0 && ndf > 0) {
    const double chi2PerNdf = chi2 / ndf;
    const double cut = chi2Limit(m_nstdev, ndf) * m_cfactr;
    if (chi2 > cut) {
      // Reject the track.
      if (m_debug) {
        debug() << "Rejected track because chi2 / ndof (" << chi2PerNdf
                << ") is larger than " << cut << endmsg;
      }
      return false;
    }
  }
  // Store the chi2 and number of degrees of freedom.
  trackParams[2 * m_nalc] = ndf;
  trackParams[2 * m_nalc + 1] = chi2;

  // Local operations are finished. Track is accepted.
  // Third loop: update the global parameters (other matrices).
  m_clcmat.assign(m_nagb, std::vector<double>(m_nalc, 0.));
  unsigned int nagbn = 0;
  std::vector<int> indnz(m_nagb, -1);
  std::vector<int> indbk(m_nagb, 0);
  for (const auto& equation : equations) {
    double rmeas = equation.rmeas;
    // Suppress the global part.
    const unsigned int nG = equation.derG.size();
    for (unsigned int i = 0; i < nG; ++i) {
      const unsigned int j = equation.indG[i];
      rmeas -= equation.derG[i] * m_dparm[j];
    }
    const double w = equation.weight;
    // First of all, the global/global terms.
    for (unsigned int i = 0; i < nG; ++i) {
      const unsigned int j = equation.indG[i];
      m_bgvec[j] += w * rmeas * equation.derG[i];
      if (m_debug) debug() << "bgvec[" << j << "] = " << m_bgvec[j] << endmsg;
      for (unsigned int k = 0; k < nG; ++k) {
        const unsigned int n = equation.indG[k];
        m_cgmat[j][n] += w * equation.derG[i] * equation.derG[k];
        if (m_debug) {
          debug() << "cgmat[" << j << "][" << n << "] = " << m_cgmat[j][n]
                  << endmsg;
        }
      }
    }
    // Now we have also rectangular matrices containing global/local terms.
    const unsigned int nL = equation.derL.size();
    for (unsigned int i = 0; i < nG; ++i) {
      const unsigned int j = equation.indG[i];
      // Index of index.
      int ik = indnz[j];
      if (ik == -1) {
        // New global variable.
        indnz[j] = nagbn;
        indbk[nagbn] = j;
        ik = nagbn;
        ++nagbn;
      }
      // Now fill the rectangular matrix.
      for (unsigned int k = 0; k < nL; ++k) {
        const unsigned int ij = equation.indL[k];
        m_clcmat[ik][ij] += w * equation.derG[i] * equation.derL[k];
        if (m_debug)
          debug() << "clcmat[" << ik << "][" << ij << "] = " << m_clcmat[ik][ij]
                  << endmsg;
      }
    }
  }

  // Third loop is finished, now we update the correction matrices.
  multiplyAVAt(clmat, m_clcmat, m_corrm, m_nalc, nagbn);
  multiplyAX(m_clcmat, blvec, m_corrv, m_nalc, nagbn);
  for (unsigned int i = 0; i < nagbn; ++i) {
    const unsigned int j = indbk[i];
    m_bgvec[j] -= m_corrv[i];
    for (unsigned int k = 0; k < nagbn; ++k) {
      const unsigned int ik = indbk[k];
      m_cgmat[j][ik] -= m_corrm[i][k];
    }
  }
  return true;
}

//=============================================================================
// Update the module positions and orientations.
//=============================================================================
void TbMillepede::updateGeometry() {

  const unsigned int nPlanes = m_nPlanes - m_maskedPlanes.size();
  for (auto im = m_modules.cbegin(), end = m_modules.cend(); im != end; ++im) {
    if (masked(im - m_modules.begin())) continue;
    const unsigned int plane = m_millePlanes[im - m_modules.begin()];
    const double tx = (*im)->x() + m_dparm[plane + 0 * nPlanes];
    const double ty = (*im)->y() + m_dparm[plane + 1 * nPlanes];
    const double tz = (*im)->z() + m_dparm[plane + 2 * nPlanes];
    double rx = 0.;
    double ry = 0.;
    double rz = 0.;
    Tb::SmallRotation((*im)->rotX(), m_dparm[plane + 3 * nPlanes],
                      (*im)->rotY(), m_dparm[plane + 4 * nPlanes],
                      m_dparm[plane + 5 * nPlanes], rx, ry, rz);
    (*im)->setAlignment(tx, ty, tz, (*im)->rotX() - rx, (*im)->rotY() + ry,
                        (*im)->rotZ() - rz, 0., 0., 0., 0., 0., 0.);
  }
}

//=============================================================================
// Initialise the vectors and arrays.
//=============================================================================
bool TbMillepede::reset(const unsigned int nPlanes, const double startfact) {

  info() << "                                           " << endmsg;
  info() << "            * o o                   o      " << endmsg;
  info() << "              o o                   o      " << endmsg;
  info() << "   o ooooo  o o o  oo  ooo   oo   ooo  oo  " << endmsg;
  info() << "    o  o  o o o o o  o o  o o  o o  o o  o " << endmsg;
  info() << "    o  o  o o o o oooo o  o oooo o  o oooo " << endmsg;
  info() << "    o  o  o o o o o    ooo  o    o  o o    " << endmsg;
  info() << "    o  o  o o o o  oo  o     oo   ooo  oo  ++ starts" << endmsg;
  info() << "                                           " << endmsg;

  // Reset the list of track equations.
  m_equations.clear();
  // Set the number of global derivatives.
  m_nagb = 6 * nPlanes;
  // Reset matrices and vectors.
  const unsigned int nRows = m_nagb + m_constraints.size();
  m_bgvec.resize(nRows, 0.);
  m_cgmat.assign(nRows, std::vector<double>(nRows, 0.));
  m_corrv.assign(m_nagb, 0.);
  m_corrm.assign(m_nagb, std::vector<double>(m_nagb, 0.));
  m_dparm.assign(m_nagb, 0.);

  // Define the sigmas for each parameter.
  m_fixed.assign(m_nagb, true);
  m_psigm.assign(m_nagb, 0.);
  for (unsigned int i = 0; i < 6; ++i) {
    if (!m_dofs[i]) continue;
    for (unsigned int j = i * nPlanes; j < (i + 1) * nPlanes; ++j) {
      m_fixed[j] = false;
      m_psigm[j] = m_sigmas[i];
    }
  }
  // Fix modules if requested.
  for (auto it = m_fixedPlanes.cbegin(); it != m_fixedPlanes.cend(); ++it) {
    info() << "You are fixing module " << (*it) << endmsg;
    // TODO: check if this is the "millePlanes" index or the regular one.
    const unsigned ndofs = m_fix_all ? 6 : 3;
    for (unsigned int i = 0; i < ndofs; ++i) {
      m_fixed[(*it) + i * nPlanes] = true;
      m_psigm[(*it) + i * nPlanes] = 0.;
    }
  }

  // Set the chi2 / ndof cut used in the local track fit.
  // Iterations are stopped when the cut factor reaches the ref. value (1).
  m_cfactref = 1.;
  m_cfactr = std::max(1., startfact);

  return true;
}

//=============================================================================
//
//=============================================================================
bool TbMillepede::fitGlobal() {

  m_diag.assign(m_nagb, 0.);
  std::vector<double> bgvecPrev(m_nagb, 0.);
  std::vector<double> trackParams(2 * m_nalc + 2, 0.);
  const unsigned int nTracks = m_equations.size();
  std::vector<std::vector<double> > localParams(
      nTracks, std::vector<double>(m_nalc, 0.));

  const unsigned int nMaxIterations = 10;
  unsigned int iteration = 1;
  unsigned int nGoodTracks = nTracks;
  while (iteration <= nMaxIterations) {
    info() << "Iteration " << iteration << " (using " << nGoodTracks
           << " tracks)" << endmsg;

    // Save the diagonal elements.
    for (unsigned int i = 0; i < m_nagb; ++i) {
      m_diag[i] = m_cgmat[i][i];
    }

    unsigned int nFixed = 0;
    for (unsigned int i = 0; i < m_nagb; ++i) {
      if (m_fixed[i]) {
        // Fixed global parameter.
        ++nFixed;
        for (unsigned int j = 0; j < m_nagb; ++j) {
          // Reset row and column.
          m_cgmat[i][j] = 0.0;
          m_cgmat[j][i] = 0.0;
        }
      } else {
        m_cgmat[i][i] += 1. / (m_psigm[i] * m_psigm[i]);
      }
    }
    // Add the constraints equations.
    unsigned int nRows = m_nagb;
    for (const auto& constraint : m_constraints) {
      double sum = constraint.rhs;
      for (unsigned int j = 0; j < m_nagb; ++j) {
        if (m_psigm[j] == 0.) {
          m_cgmat[nRows][j] = 0.0;
          m_cgmat[j][nRows] = 0.0;
        } else {
          m_cgmat[nRows][j] = nGoodTracks * constraint.coefficients[j];
          m_cgmat[j][nRows] = m_cgmat[nRows][j];
        }
        sum -= constraint.coefficients[j] * m_dparm[j];
      }
      m_cgmat[nRows][nRows] = 0.0;
      m_bgvec[nRows] = nGoodTracks * sum;
      ++nRows;
    }

    double cor = 0.0;
    if (iteration > 1) {
      for (unsigned int j = 0; j < m_nagb; ++j) {
        for (unsigned int i = 0; i < m_nagb; ++i) {
          if (m_fixed[i]) continue;
          cor += bgvecPrev[j] * m_cgmat[j][i] * bgvecPrev[i];
          if (i == j) {
            cor -= bgvecPrev[i] * bgvecPrev[i] / (m_psigm[i] * m_psigm[i]);
          }
        }
      }
    }
    if (m_debug) debug() << " Final corr. is " << cor << endmsg;

    // Do the matrix inversion.
    const int rank = invertMatrix(m_cgmat, m_bgvec, nRows);
    // Update the global parameters values.
    for (unsigned int i = 0; i < m_nagb; ++i) {
      m_dparm[i] += m_bgvec[i];
      bgvecPrev[i] = m_bgvec[i];
      if (m_debug) {
        debug() << "bgvec[" << i << "] = " << m_bgvec[i] << endmsg;
        debug() << "dparm[" << i << "] = " << m_dparm[i] << endmsg;
        debug() << "cgmat[" << i << "][" << i << "] = " << m_cgmat[i][i]
                << endmsg;
        debug() << "err = " << sqrt(fabs(m_cgmat[i][i])) << endmsg;
        debug() << "cgmat * diag = " << std::setprecision(5)
                << m_cgmat[i][i] * m_diag[i] << endmsg;
      }
    }
    if (nRows - nFixed - rank != 0) {
      warning() << "The rank defect of the symmetric " << nRows << " by "
                << nRows << " matrix is " << nRows - nFixed - rank << endmsg;
    }
    if (!m_iterate) break;
    ++iteration;
    if (iteration == nMaxIterations) break;
    // Update the factor in the track chi2 cut.
    const double newcfactr = sqrt(m_cfactr);
    if (newcfactr > 1.2 * m_cfactref) {
      m_cfactr = newcfactr;
    } else {
      m_cfactr = m_cfactref;
    }
    if (m_debug) {
      debug() << "Refitting tracks with cut factor " << m_cfactr << endmsg;
    }

    // Reset global variables.
    for (unsigned int i = 0; i < nRows; ++i) {
      m_bgvec[i] = 0.0;
      for (unsigned int j = 0; j < nRows; ++j) m_cgmat[i][j] = 0.0;
    }
    // Refit the tracks.
    double chi2 = 0.;
    double ndof = 0.;
    nGoodTracks = 0;
    for (unsigned int i = 0; i < nTracks; ++i) {
      // Skip invalidated tracks.
      if (m_equations[i].empty()) continue;
      std::vector<Equation> equations(m_equations[i].begin(),
                                      m_equations[i].end());
      for (auto equation : equations) {
        const unsigned int nG = equation.derG.size();
        for (unsigned int j = 0; j < nG; ++j) {
          const double t = localParams[i][equation.indNL[j]];
          if (t == 0) continue;
          equation.derG[j] += equation.derNL[j] * (t - equation.slopes[j]);
        }
      }
      std::fill(trackParams.begin(), trackParams.end(), 0.);
      // Refit the track.
      bool ok = fitTrack(equations, trackParams, false, iteration);
      // Cache the track state.
      for (unsigned int j = 0; j < m_nalc; ++j) {
        localParams[i][j] = trackParams[2 * j];
      }
      if (ok) {
        // Update the total chi2.
        chi2 += trackParams[2 * m_nalc + 1];
        ndof += trackParams[2 * m_nalc];
        ++nGoodTracks;
      } else {
        // Disable the track.
        m_equations[i].clear();
      }
    }
    info() << "Chi2 / DOF after re-fit: " << chi2 / (ndof - nRows) << endmsg;
  }

  // Print the final results.
  printResults();

  info() << endmsg;
  info() << "            * o o                   o      " << endmsg;
  info() << "              o o                   o      " << endmsg;
  info() << "   o ooooo  o o o  oo  ooo   oo   ooo  oo  " << endmsg;
  info() << "    o  o  o o o o o  o o  o o  o o  o o  o " << endmsg;
  info() << "    o  o  o o o o oooo o  o oooo o  o oooo " << endmsg;
  info() << "    o  o  o o o o o    ooo  o    o  o o    " << endmsg;
  info() << "    o  o  o o o o  oo  o     oo   ooo  oo ++ ends." << endmsg;
  info() << "                       o                   " << endmsg;
  info() << endmsg;
  return true;
}

//=============================================================================
// Obtain solution of a system of linear equations with symmetric matrix
// and the inverse (using 'singular-value friendly' GAUSS pivot).
// Solve the equation V * X = B.
// V is replaced by its inverse matrix and B by X, the solution vector
//=============================================================================
int TbMillepede::invertMatrix(std::vector<std::vector<double> >& v,
                              std::vector<double>& b, const int n) {
  int rank = 0;
  const double eps = 0.0000000000001;

  std::vector<double> diag(n, 0.);
  std::vector<bool> used_param(n, true);
  std::vector<bool> flag(n, true);
  for (int i = 0; i < n; i++) {
    for (int j = 0; j <= i; j++) {
      v[j][i] = v[i][j];
    }
  }

  // Find max. elements of each row and column.
  std::vector<double> r(n, 0.);
  std::vector<double> c(n, 0.);
  for (int i = 0; i < n; i++) {
    for (int j = 0; j < n; j++) {
      if (fabs(v[i][j]) >= r[i]) r[i] = fabs(v[i][j]);
      if (fabs(v[j][i]) >= c[i]) c[i] = fabs(v[j][i]);
    }
  }
  for (int i = 0; i < n; i++) {
    if (0.0 != r[i]) r[i] = 1. / r[i];
    if (0.0 != c[i]) c[i] = 1. / c[i];
    // Check if max elements are within requested precision.
    if (eps >= r[i]) r[i] = 0.0;
    if (eps >= c[i]) c[i] = 0.0;
  }

  // Equilibrate the V matrix
  for (int i = 0; i < n; i++) {
    for (int j = 0; j < n; j++) {
      v[i][j] = sqrt(r[i]) * v[i][j] * sqrt(c[j]);
    }
  }

  for (int i = 0; i < n; i++) {
    // Save the absolute values of the diagonal elements.
    diag[i] = fabs(v[i][i]);
    if (r[i] == 0. && c[i] == 0.) {
      // This part is empty (non-linear treatment with non constraints)
      flag[i] = false;
      used_param[i] = false;
    }
  }

  for (int i = 0; i < n; i++) {
    double vkk = 0.0;
    int k = -1;
    // First look for the pivot, i. e. the max unused diagonal element.
    for (int j = 0; j < n; j++) {
      if (flag[j] && (fabs(v[j][j]) > std::max(fabs(vkk), eps))) {
        vkk = v[j][j];
        k = j;
      }
    }

    if (k >= 0) {
      // Pivot found.
      rank++;
      // Use this value.
      flag[k] = false;
      // Replace pivot by its inverse.
      vkk = 1.0 / vkk;
      v[k][k] = -vkk;
      for (int j = 0; j < n; j++) {
        for (int jj = 0; jj < n; jj++) {
          if (j != k && jj != k && used_param[j] && used_param[jj]) {
            // Other elements (do them first as you use old v[k][j])
            v[j][jj] = v[j][jj] - vkk * v[j][k] * v[k][jj];
          }
        }
      }
      for (int j = 0; j < n; j++) {
        if (j != k && used_param[j]) {
          v[j][k] = v[j][k] * vkk;
          v[k][j] = v[k][j] * vkk;
        }
      }
    } else {
      // No more pivot value (clear those elements)
      for (int j = 0; j < n; j++) {
        if (flag[j]) {
          b[j] = 0.0;
          for (int k = 0; k < n; k++) {
            v[j][k] = 0.0;
            v[k][j] = 0.0;
          }
        }
      }
      break;
    }
  }
  // Correct matrix V
  for (int i = 0; i < n; i++) {
    for (int j = 0; j < n; j++) {
      v[i][j] = sqrt(c[i]) * v[i][j] * sqrt(r[j]);
    }
  }

  std::vector<double> temp(n, 0.);
  for (int j = 0; j < n; j++) {
    // Reverse matrix elements
    for (int jj = 0; jj < n; jj++) {
      v[j][jj] = -v[j][jj];
      temp[j] += v[j][jj] * b[jj];
    }
  }

  for (int j = 0; j < n; j++) {
    b[j] = temp[j];
  }
  return rank;
}

//=============================================================================
// Simplified version.
//=============================================================================
int TbMillepede::invertMatrixLocal(std::vector<std::vector<double> >& v,
                                   std::vector<double>& b, const int n) {

  int rank = 0;
  const double eps = 0.0000000000001;

  std::vector<bool> flag(n, true);
  std::vector<double> diag(n, 0.);
  for (int i = 0; i < n; i++) {
    // Save the absolute values of the diagonal elements.
    diag[i] = fabs(v[i][i]);
    for (int j = 0; j <= i; j++) {
      v[j][i] = v[i][j];
    }
  }

  for (int i = 0; i < n; i++) {
    double vkk = 0.0;
    int k = -1;
    // First look for the pivot, i. e. the max. unused diagonal element.
    for (int j = 0; j < n; j++) {
      if (flag[j] && (fabs(v[j][j]) > std::max(fabs(vkk), eps * diag[j]))) {
        vkk = v[j][j];
        k = j;
      }
    }

    if (k >= 0) {
      // Pivot found
      rank++;
      flag[k] = false;
      // Replace pivot by its inverse.
      vkk = 1.0 / vkk;
      v[k][k] = -vkk;
      for (int j = 0; j < n; j++) {
        for (int jj = 0; jj < n; jj++) {
          if (j != k && jj != k) {
            // Other elements (do them first as you use old v[k][j])
            v[j][jj] = v[j][jj] - vkk * v[j][k] * v[k][jj];
          }
        }
      }

      for (int j = 0; j < n; j++) {
        if (j != k) {
          v[j][k] = v[j][k] * vkk;
          v[k][j] = v[k][j] * vkk;
        }
      }
    } else {
      // No more pivot values (clear those elements).
      for (int j = 0; j < n; j++) {
        if (flag[j]) {
          b[j] = 0.0;
          for (k = 0; k < n; k++) v[j][k] = 0.0;
        }
      }
      break;
    }
  }

  std::vector<double> temp(n, 0.);
  for (int j = 0; j < n; j++) {
    // Reverse matrix elements
    for (int jj = 0; jj < n; jj++) {
      v[j][jj] = -v[j][jj];
      temp[j] += v[j][jj] * b[jj];
    }
  }
  for (int j = 0; j < n; j++) {
    b[j] = temp[j];
  }
  return rank;
}

//=============================================================================
// Return the limit in chi^2 / nd for n sigmas stdev authorized.
// Only n=1, 2, and 3 are expected in input.
//=============================================================================
double TbMillepede::chi2Limit(const int n, const int nd) const {
  constexpr double sn[3] = {0.47523, 1.690140, 2.782170};
  constexpr double table[3][30] = {
      {1.0000, 1.1479, 1.1753, 1.1798, 1.1775, 1.1730, 1.1680, 1.1630,
       1.1581, 1.1536, 1.1493, 1.1454, 1.1417, 1.1383, 1.1351, 1.1321,
       1.1293, 1.1266, 1.1242, 1.1218, 1.1196, 1.1175, 1.1155, 1.1136,
       1.1119, 1.1101, 1.1085, 1.1070, 1.1055, 1.1040},
      {4.0000, 3.0900, 2.6750, 2.4290, 2.2628, 2.1415, 2.0481, 1.9736,
       1.9124, 1.8610, 1.8171, 1.7791, 1.7457, 1.7161, 1.6897, 1.6658,
       1.6442, 1.6246, 1.6065, 1.5899, 1.5745, 1.5603, 1.5470, 1.5346,
       1.5230, 1.5120, 1.5017, 1.4920, 1.4829, 1.4742},
      {9.0000, 5.9146, 4.7184, 4.0628, 3.6410, 3.3436, 3.1209, 2.9468,
       2.8063, 2.6902, 2.5922, 2.5082, 2.4352, 2.3711, 2.3143, 2.2635,
       2.2178, 2.1764, 2.1386, 2.1040, 2.0722, 2.0428, 2.0155, 1.9901,
       1.9665, 1.9443, 1.9235, 1.9040, 1.8855, 1.8681}};

  if (nd < 1) return 0.;
  const int m = std::max(1, std::min(n, 3));
  if (nd <= 30) return table[m - 1][nd - 1];
  return ((sn[m - 1] + sqrt(float(2 * nd - 3))) *
          (sn[m - 1] + sqrt(float(2 * nd - 3)))) /
         float(2 * nd - 2);
}

//=============================================================================
// Multiply general M-by-N matrix A and N-vector X
//=============================================================================
bool TbMillepede::multiplyAX(const std::vector<std::vector<double> >& a,
                             const std::vector<double>& x,
                             std::vector<double>& y, const unsigned int n,
                             const unsigned int m) {

  // Y = A * X, where
  //   A = general M-by-N matrix
  //   X = N vector
  //   Y = M vector
  for (unsigned int i = 0; i < m; ++i) {
    y[i] = 0.0;
    for (unsigned int j = 0; j < n; ++j) {
      y[i] += a[i][j] * x[j];
    }
  }
  return true;
}

//=============================================================================
// Multiply symmetric N-by-N matrix V from the left with general M-by-N
// matrix A and from the right with the transposed of the same general
// matrix to form a symmetric M-by-M matrix W.
//=============================================================================
bool TbMillepede::multiplyAVAt(const std::vector<std::vector<double> >& v,
                               const std::vector<std::vector<double> >& a,
                               std::vector<std::vector<double> >& w,
                               const unsigned int n, const unsigned int m) {

  // W = A * V * AT, where
  //   V = symmetric N-by-N matrix
  //   A = general N-by-M matrix
  //   W = symmetric M-by-M matrix
  for (unsigned int i = 0; i < m; ++i) {
    for (unsigned int j = 0; j <= i; ++j) {
      // Reset the final matrix.
      w[i][j] = 0.0;
      // Fill the upper left triangle.
      for (unsigned int k = 0; k < n; ++k) {
        for (unsigned int l = 0; l < n; ++l) {
          w[i][j] += a[i][k] * v[k][l] * a[j][l];
        }
      }
      // Fill the rest.
      w[j][i] = w[i][j];
    }
  }

  return true;
}

//=============================================================================
// Print results
//=============================================================================
bool TbMillepede::printResults() {
  const std::string line(85, '-');
  info() << line << endmsg;
  info() << " Result of fit for global parameters" << endmsg;
  info() << line << endmsg;
  info() << "   I    Difference    Last step      "
         << "Error        Pull     Global corr." << endmsg;
  info() << line << endmsg;
  for (unsigned int i = 0; i < m_nagb; ++i) {
    double err = sqrt(fabs(m_cgmat[i][i]));
    if (m_cgmat[i][i] < 0.0) err = -err;
    if (fabs(m_cgmat[i][i] * m_diag[i]) > 0) {
      info() << std::setprecision(6) << std::fixed;
      // Calculate the pull.
      const double pull =
          m_dparm[i] / sqrt(m_psigm[i] * m_psigm[i] - m_cgmat[i][i]);
      // Calculate the global correlation coefficient
      // (correlation between the parameter and all the other variables).
      const double gcor = sqrt(fabs(1.0 - 1.0 / (m_cgmat[i][i] * m_diag[i])));
      info() << std::setw(4) << i << "   " << std::setw(10) << m_dparm[i]
             << "   " << std::setw(10) << m_bgvec[i] << "   " << std::setw(10)
             << std::setprecision(5) << err << "   " << std::setw(10)
             << std::setprecision(5) << pull << "   " << std::setw(10) << gcor
             << endmsg;
    } else {
      info() << std::setw(4) << i << "   " << std::setw(10) << "OFF"
             << "   " << std::setw(10) << "OFF"
             << "   " << std::setw(10) << "OFF"
             << "   " << std::setw(10) << "OFF"
             << "   " << std::setw(10) << "OFF" << endmsg;
    }
    if ((i + 1) % (m_nagb / 6) == 0) info() << line << endmsg;
  }
  if (m_debug) {
    for (unsigned int i = 0; i < m_nagb; ++i) {
      debug() << " i=" << i
              << "  sqrt(fabs(cgmat[i][i]))=" << sqrt(fabs(m_cgmat[i][i]))
              << " diag = " << m_diag[i] << endmsg;
    }
  }

  return true;
}