MeasurementSelector.h

Go to the documentation of this file.

/*
  Copyright (C) 2002-2026 CERN for the benefit of the ATLAS collaboration
  */
#pragma once
// Alternative measurement selector
//
// This measurement selector is assuming the following
// - the number of selected measurements is small (~<10)
// - the number of measurement candidates is typically >> the
//   number of selected measuremebts
// - the total number of candidate measurements can be large
// - there is a simple one-to-one relation between bound state
//   parameters and measurement coordinates.
 
#include "Acts/Utilities/Result.hpp"
#include "Acts/Utilities/Delegate.hpp"
#include "Acts/EventData/SourceLink.hpp"
#include "Acts/TrackFinding/CombinatorialKalmanFilterError.hpp"
#include "Acts/Surfaces/Surface.hpp"
#include "Acts/Geometry/GeometryHierarchyMap.hpp"
#include "Acts/EventData/Types.hpp"
#include "Acts/EventData/TrackStatePropMask.hpp"
#include "boost/container/small_vector.hpp"
 
// for BaseTypes
#include "Acts/EventData/TrackStateProxy.hpp"
#include "Acts/Utilities/CalibrationContext.hpp"
#include "Acts/EventData/BoundTrackParameters.hpp"
 
// for MeasurementSizeMax
#include "Acts/EventData/MultiTrajectory.hpp"
 
#include "ActsCalibBase/MeasurementCalibratorBase.h"
#include "TrackStateFlagHelper.h"
 
#include <utility>
#include <type_traits>
#include <optional>
 
// Types to be used during measurement selection for the prediction and the
// measurement for calibrated measurements after selection if the actual calibration is
// executed after the calibration, and the trajectory, track state, bound parameter
// types.
template <typename derived_t>
struct MeasurementSelectorTraits
{
   // the measurement type after the selection e.g. a Matrix<N,1>
   template <std::size_t N>
   using CalibratedMeasurement = typename Acts::detail_tsp::FixedSizeTypes<N>::Coefficients;
 
   // the  measurement covariance type after the selection e.g. a Matrix<N,N>
   template <std::size_t N>
   using CalibratedMeasurementCovariance = typename Acts::detail_tsp::FixedSizeTypes<N>::Covariance;
 
   // the measurement type before the selection e.g. an Eigen::Map< Matrix<N,1> > if
   // the calibration is performed after the selection
   template <std::size_t N>
   using PreSelectionMeasurement = typename Acts::detail_tsp::FixedSizeTypes<N>::Coefficients;
 
   // the measurement covariance type before the selection e.g. an Eigen::Map<Matrix<N,N> > if
   // the calibration is performed after the selection
   template <std::size_t N>
   using PreSelectionMeasurementCovariance = typename Acts::detail_tsp::FixedSizeTypes<N>::Covariance;
 
   // e.g. the same as CalibratedMeasurement
   template <std::size_t N>
   using Predicted = typename Acts::detail_tsp::FixedSizeTypes<N>::Coefficients;
 
   // e.g. the same as CalibratedMeasurementCovariance
   template <std::size_t N>
   using PredictedCovariance = typename Acts::detail_tsp::FixedSizeTypes<N>::Covariance;
 
   // e.g. helper template to get the value_type from the container type
   // e.g. helper template to get the value_type from the measurement range iterator type
 
   template <typename T_MeasurementRangeIterator>
   struct MeasurementContainerTraits {
      using value_type = typename T_MeasurementRangeIterator::value_type;
   };
 
   // abstract measurement range
   // implements empty()
   // and can be converted by the derived_t into an iterable range
   // over measurements of concrete measurement container_type.
   using abstract_measurement_range_t = std::ranges::iota_view<unsigned int, unsigned int>;
 
   // the trajectory type to which states for selected measurements are to be added
   using trajectory_t = typename derived_t::traj_t;
   // the track state type for new track states
   using TrackStateProxy = trajectory_t::TrackStateProxy;
 
   // the value type usd for matrices
   using MatrixFloatType = double;
   using BoundTrackParameters = Acts::BoundTrackParameters;
   using BoundMatrix = Acts::BoundMatrix;
 
   using BoundState = std::tuple<BoundTrackParameters, BoundMatrix, double>;
 
   // maximum dimension of measurements
   static const std::size_t s_dimMax = 3;
 
   // must be the same as what is used for the CKF
   // @TODO how to get rid of this ?
   static constexpr std::size_t s_maxBranchesPerSurface = 10;
};
 
//
struct MeasurementSelectorMatrixTraits {
   template <class T_Matrix>
   static constexpr std::size_t matrixColumns() { return T_Matrix::ColsAtCompileTime;}
   template <class T_Matrix>
   static constexpr std::size_t matrixRows() { return T_Matrix::RowsAtCompileTime;}
 
   template <typename T_Float, class T_Matrix>
   static auto matrixTypeCast(const T_Matrix &matrix) { return matrix.template cast<T_Float>(); }
 
   template <class T_Matrix>
   static auto transpose(const T_Matrix &matrix) { return matrix.transpose(); }
 
   template <class T_Matrix>
   static auto invert(const T_Matrix &matrix) { return matrix.inverse(); }
};
 
 
// Map from the measurement to bound state domain
// it is assumed that there is a simple unambiguous association
// between coordinates of the measurement domain and the bound state domain
// e.g. measurement coordinate 1 maps to loc0 of the bound state.
// @TODO use FixedSizeSubspace ? currently does not provide methods to directly
//       create the a sub-space covariance matrix from a full covariance matrix,
//       and to create the projector bitset, which is stored in the TrackState.
struct ParameterMapping {
 
   template <std::size_t N>
   using type = std::array<unsigned char, N>;
 
   template <std::size_t N>
   static constexpr type<N> identity() {
      type<N> ret;
      for(int i=0; i<N; ++i) {
         ret[i]=i;
      }
      return ret;
   }
};
 
// utility to "project" a bound state parameter vector or covariance matrix onto the 1,2, ... N dimensional measurement domain
// @TODO allow to influence resulting matrix type ?
template <std::size_t N,class T_ResultType,class T_Matrix>
T_ResultType project(ParameterMapping::type<N> parameter_map, const T_Matrix &matrix)
{
   using MatrixIndexMapType = unsigned char; // "char" to reduce the size of the map, and if not wide enough this entire
                                             //        concept is likely inefficient.
   using MatrixIndexType = unsigned int;     // @TODO or std::size_t ? does not matter
 
   // ensure that index types are wide enough
   static_assert( MeasurementSelectorMatrixTraits::matrixRows<T_Matrix>() < std::numeric_limits<MatrixIndexMapType>::max());
   static_assert( N*MeasurementSelectorMatrixTraits::matrixRows<T_Matrix>() < std::numeric_limits<MatrixIndexType>::max());
 
   T_ResultType ret;
   if constexpr(MeasurementSelectorMatrixTraits::matrixColumns<T_Matrix>() == 1) {
      // handle projection of paramteter vector
      for (MatrixIndexType meas_i=0; meas_i<N; ++meas_i) {
         assert( meas_i < parameter_map.size() );
         ret(meas_i,0) = matrix( parameter_map[meas_i], 0);
      }
   }
   else {
      // handle projection of covariance matrix
      // "project" matrix
      for (MatrixIndexType meas_i=0; meas_i<N; ++meas_i) {
         assert( meas_i < parameter_map.size());
         MatrixIndexType param_i = parameter_map[meas_i];
         for (MatrixIndexType meas_j=0; meas_j<N; ++meas_j) {
            assert( meas_j < parameter_map.size());
            ret(meas_i,meas_j) = matrix(param_i, parameter_map[meas_j]);
         }
      }
   }
   return ret;
}
 
 
 
 
// helper method to compute a chi2 for the difference of two "measurement" and covariance pairs
template <typename measurement_vector_t, typename measurement_cov_matrix_t,
          typename predicted_vector_t, typename predicted_cov_matrix_t>
double computeChi2(const measurement_vector_t &a,
                   const measurement_cov_matrix_t &a_cov,
                   const predicted_vector_t &b,
                   const predicted_cov_matrix_t &b_cov) {
 
   // just sum sanity checks that a and b have the correct dimensions and that
   // the chi2 can actually be computed.
   static_assert( MeasurementSelectorMatrixTraits::matrixColumns<measurement_vector_t>() == 1); // a is vector
   static_assert( MeasurementSelectorMatrixTraits::matrixColumns<predicted_vector_t>() == 1); // b is vector
   static_assert( MeasurementSelectorMatrixTraits::matrixRows<measurement_cov_matrix_t>()
                  == MeasurementSelectorMatrixTraits::matrixColumns<measurement_cov_matrix_t>() ); // a is square matrix
   static_assert( MeasurementSelectorMatrixTraits::matrixRows<predicted_cov_matrix_t>()
                  == MeasurementSelectorMatrixTraits::matrixColumns<predicted_cov_matrix_t>() ); // b is square matrix
   static_assert( MeasurementSelectorMatrixTraits::matrixRows<measurement_cov_matrix_t>()
                  == MeasurementSelectorMatrixTraits::matrixRows<measurement_vector_t>() );        // a vector matches matrix
   static_assert( MeasurementSelectorMatrixTraits::matrixRows<predicted_cov_matrix_t>()
                  == MeasurementSelectorMatrixTraits::matrixRows<predicted_vector_t>() );        // b vector matches matrix
   static_assert( MeasurementSelectorMatrixTraits::matrixRows<measurement_cov_matrix_t>()
                  == MeasurementSelectorMatrixTraits::matrixRows<predicted_cov_matrix_t>() );    // a and b match
 
   // @TODO remove abstraction i.e. assume matrix has the interface of an eigen matrix
   auto inv_ab_cov( MeasurementSelectorMatrixTraits::invert(a_cov+b_cov) );
   auto  diff( a-b);
   return (MeasurementSelectorMatrixTraits::transpose(diff) * inv_ab_cov * diff)(0,0);
}
 
// Collection to hold the n-"best" candidates
// The objects of type PayloadType must support assignment operation, and must
// be default constructible. Moreover it must be possible to provide a
// "comparison" operator to order the payload objects.
template <std::size_t N, class PayloadType >
struct TopCollection {
   using IndexType = unsigned short; // @TODO or char ? If N>>10 this concept is likely
                                     //       inefficient
   //   using PayloadType = Payload<DIM>;
   TopCollection(std::size_t max_n) {
      init(max_n);
   }
 
   // @param max_n the maximum number of top-candidates is fixed by the template parameter
   //    N but can be reduced further to this number
   void init(std::size_t max_n) {
      assert( max_n < N);
      m_nextSlot=0;
      m_maxSlots=max_n;
      m_order[0]=0;
   }
   // @param get a slot to hold a new candidate which is not necessarily accepted in the list
   //     of the n-top candidates
   PayloadType &slot() {
      return m_slots[m_order[m_nextSlot] ];
   }
   // @param idx get the specified filled slot (read only) indicated by the index, where the index does not
   //    indicate the order in the top-candidate list
   const PayloadType &getSlot(IndexType idx) const {
      return m_slots[idx];
   }
   // @param idx get the specified filled slot indicated by the index, where the index does not
   //    indicate the order in the top-candidate list
   PayloadType &getSlot(IndexType idx) {
      return m_slots[idx];
   }
   // @param test whether the given index points to one of the accepted top candidates
   bool isValid(IndexType idx) const {
      return idx < m_nextSlot;
   }
   // Accept the element of the latest slot provided there is still a free slot or it is better than
   // the worst element.
   // @param comparison operator to compute e.g. a<b where "smaller" means "better"
   void acceptAndSort(std::function<bool(const PayloadType &a, const PayloadType &b)> comparison) {
      // bubble best element to top
      for (unsigned int slot_i = m_nextSlot;
           slot_i-- > 0
           && !comparison(m_slots[ m_order[slot_i] ],m_slots[ m_order[slot_i+1] ]);) {
         std::swap(m_order[slot_i],m_order[slot_i+1]);
      }
      // if there are still free slot increase the number of used slots
      if (m_nextSlot < m_maxSlots) {
         ++m_nextSlot;
         m_order[m_nextSlot]=m_nextSlot;
      }
   }
   // Accept the element of the latest slot provided there is still a free slot.
   bool acceptNoSort() {
      // if there are still free slot increase the number of used slots
      if (m_nextSlot < m_maxSlots) {
         ++m_nextSlot;
         m_order[m_nextSlot]=m_nextSlot;
      }
      return (m_nextSlot < m_maxSlots);
   }
 
   bool empty() const {
      return m_nextSlot==0;
   }
 
   IndexType size() const { return m_nextSlot; }
 
   // helper to iterate over the slot indices in sorting order from best to worst
   typename std::array<IndexType, N+1>::const_iterator begin() { return m_order.begin(); }
   typename std::array<IndexType, N+1>::const_iterator end()   { return m_order.begin()+m_nextSlot; }
 
   std::array<PayloadType, N+1> m_slots;         // storage for the slots
   std::array<IndexType,   N+1> m_order;         // order of the filled slots
   IndexType                    m_nextSlot = 0;  // the index of the next free slot
   IndexType                    m_maxSlots = 0;  // maximum number of top-slots
};
 
 
 
struct AtlasMeasurementSelectorCuts {
  std::vector<float> etaBins{};
  std::vector<std::pair<float, float> > chi2CutOff{ {15,25} };
  std::vector<std::size_t> numMeasurementsCutOff{1};
  std::optional<Acts::BoundaryTolerance> edgeTolerance{};
};
 
// Measurement type specific measirement selector
// Assumptions:
//  - begin and end source_link_iterator point to a contiguous range of measurements in a single container
//  - all measurements in this range have the same dimensionality.
//  - the mapping of the bound parameters to the measurement domain for each measurement in this range
//    is identical
template <std::size_t NMeasMax,
          std::size_t DIMMAX,
          typename derived_t >
struct MeasurementSelectorBase {
 
   using Config = Acts::GeometryHierarchyMap<AtlasMeasurementSelectorCuts>;
   using traits = MeasurementSelectorTraits<derived_t>;
 
   Config m_config;
 
protected:
   using T_BoundState = typename MeasurementSelectorTraits<derived_t>::BoundState;
   using trajectory_t = typename MeasurementSelectorTraits<derived_t>::trajectory_t;
   using TrackStateProxy = typename MeasurementSelectorTraits<derived_t>::TrackStateProxy;
   static constexpr std::size_t s_maxBranchesPerSurface = MeasurementSelectorTraits<derived_t>::s_maxBranchesPerSurface;
 
   const derived_t &derived() const { return *static_cast<const derived_t *>(this); }
 
   // helper to create a projector bitset from a map from bound parameters to coordinates
   struct ProjectorBitSetMaker {
      template <std::size_t N>
      static
      Acts::ProjectorBitset create(const ParameterMapping::type<N> &parameter_map) {
         constexpr std::size_t nrows = Acts::kMeasurementSizeMax;
         constexpr std::size_t ncols = Acts::eBoundSize;
 
         std::bitset<nrows * ncols> proj_bitset {};
 
         for (unsigned int col_i=0; col_i<N; ++col_i) {
            unsigned int row_i = parameter_map[col_i];
            unsigned int idx = col_i *nrows + row_i;      // @TODO handle row major and column major correctly
            proj_bitset[ (nrows * ncols - 1) - idx ] = 1;
         }
         return proj_bitset.to_ullong();
      }
   };
 
   // helper to create states without the calibrated measurement, covariance, and sourcelink
   // if more than one state is created the states after the first will share
   // the jacobi and the prediction, if outlier_states is true no storage is created
   // for filtered.
   // @TODO should there be the possibility to have multiple states which mix outlier and non-outlier states ?
   static void createStates(std::size_t n_new_states,
                            const T_BoundState& boundState,
                            std::size_t prevTip,
                            trajectory_t& trajectory,
                            const Acts::BoundSubspaceIndices& subspaceIndices,
                            boost::container::small_vector< typename TrackStateProxy::IndexType, s_maxBranchesPerSurface> &track_states,
                            const Acts::Logger& logger,
                            bool outlier_states) {
      // create track states for all compatible measurement candidates
      const auto &boundParams = derived_t::boundParams(boundState);
      const auto &pathLength = derived_t::pathLength(boundState);
 
      using PM = Acts::TrackStatePropMask;
 
      track_states.reserve( n_new_states );
      std::optional<TrackStateProxy> firstTrackState{
         std::nullopt};
      for (unsigned int state_i=0; state_i<n_new_states; ++state_i) {
         PM mask = PM::Predicted | PM::Filtered | PM::Jacobian | PM::Calibrated;
 
         if (firstTrackState.has_value()) {
            // subsequent track states don't need storage for these as they will
            // be shared
            mask &= ~PM::Predicted & ~PM::Jacobian;
         }
 
         if (outlier_states) {
            // outlier won't have separate filtered parameters
            mask &= ~PM::Filtered;
         }
 
         TrackStateProxy trackState = trajectory.makeTrackState(mask, prevTip);
         ACTS_VERBOSE("Create SourceLink output track state #"
                      << trackState.index() << " with mask: " << mask);
 
         if (firstTrackState.has_value()) {
            trackState.shareFrom(*firstTrackState, PM::Predicted);
            trackState.shareFrom(*firstTrackState, PM::Jacobian);
         }
         else {
            // only set these for first
            trackState.predicted() = boundParams.parameters();
            if (boundParams.covariance()) {
               trackState.predictedCovariance() = *boundParams.covariance();
            }
            trackState.jacobian() = derived_t::boundJacobiMatrix(boundState);
            firstTrackState = trackState;
         }
         trackState.pathLength() = pathLength;
 
         trackState.setReferenceSurface(boundParams.referenceSurface().getSharedPtr());
 
         trackState.setProjectorSubspaceIndices(subspaceIndices);
 
         auto typeFlags = trackState.typeFlags();
         if (trackState.referenceSurface().surfaceMaterial() != nullptr) {
            typeFlags.setHasMaterial();
         }
         typeFlags.setHasParameters();
 
         // @TODO these track states still need some additional processing. Should there be a special
         //       flag for this ?
         track_states.push_back( trackState.index());
      }
   }
 
 
   struct Flags {
      bool isSplitHit()   const { return ActsTrk::detail::testTrackStateFlag(Acts::TrackStateFlag::IsSplitHit, flags);}
      bool isOutlier()   const { return ActsTrk::detail::testTrackStateFlag(Acts::TrackStateFlag::IsOutlier, flags);}
      operator unsigned int() { return flags; }
      unsigned int flags;
   };
 
   template <typename T1, typename T2>
   struct PairWithFlags : std::tuple<T1,T2,unsigned int> {
      PairWithFlags<T1,T2> &operator=(std::tuple<T1,T2,unsigned int> &&a) {
         std::tuple<T1,T2,unsigned int>::operator=(std::move(a));
         return *this;
      }
 
      const T1 &position() const { return std::get<0>(*this); }
      const T2 &covariance() const { return std::get<1>(*this); }
      Flags flags() const { return Flags{ std::get<2>(*this) }; }
   };
 
   unsigned int makeOutlierFlag(bool value) { return static_cast<unsigned int>(value)<<2u; }
   // utility struct to temporarily store data about the best measurements
   template <std::size_t DIM, typename T_SourceLink>
   struct MatchingMeasurement {
      using MeasCovPair = PairWithFlags < typename MeasurementSelectorTraits<derived_t>::template PreSelectionMeasurement<DIM>,
                                          typename MeasurementSelectorTraits<derived_t>::template PreSelectionMeasurementCovariance<DIM> >;
      MeasCovPair      m_measurement;
      float            m_chi2{};
      std::optional<T_SourceLink> m_sourceLink;
   };
 
   // type and dimension specific function to select measurements from the range defined by the source link iterators.
   // will iterate over the contiguous measurement range defined by the source link iterators where the measurements
   // are contained in the given container. The selection lopp will get the measurement and covariance with the
   // help of a preCalibrator. select measurements based on smallest chi2 wrt. the prediction, then optionally
   // apply a full calibrator after the selection and finally create track states.
   template <std::size_t DIM, typename T_MeasurementRange>
   Acts::Result<boost::container::small_vector< typename TrackStateProxy::IndexType, s_maxBranchesPerSurface> >
   selectMeasurementsCreateTrackStates(const Acts::GeometryContext& geometryContext,
                                       const Acts::CalibrationContext& calibrationContext,
                                       const Acts::Surface& surface,
                                       const T_BoundState& boundState,
                                       T_MeasurementRange &&measurement_range,
                                       std::size_t prevTip,
                                       trajectory_t& trajectory,
                                       const Acts::Logger& logger,
                                       const std::size_t numMeasurementsCut,
                                       const std::pair<float,float>& maxChi2Cut,
                                       bool forced) const {
      Acts::Result<boost::container::small_vector< typename TrackStateProxy::IndexType, s_maxBranchesPerSurface> >
         result = boost::container::small_vector< typename TrackStateProxy::IndexType, s_maxBranchesPerSurface>{};
 
      using iterator_t = decltype(measurement_range.begin());
      using container_value_t  = typename MeasurementSelectorTraits<derived_t>::template MeasurementContainerTraits<iterator_t>::value_type;
      using BaseElementType = std::remove_cv_t<std::remove_pointer_t< container_value_t > >;
      // get calibrator
      using TheMatchingMeasurement = MatchingMeasurement<DIM, container_value_t >;
      using MeasCovPair = TheMatchingMeasurement::MeasCovPair;
 
      using Predicted = typename MeasurementSelectorTraits<derived_t>::template Predicted<DIM>;
      using PredictedCovariance = typename MeasurementSelectorTraits<derived_t>::template PredictedCovariance<DIM>;
 
 
 
      // get prediction in the measurement domain
      Acts::SubspaceIndices<DIM> parameter_map = derived().template parameterMap<DIM>(geometryContext,
                                                                                       calibrationContext,
                                                                                       surface);
      auto predicted
         = std::make_pair( project<DIM, Predicted>(parameter_map,
                                                   derived().boundParams(boundState).parameters()),
                           project<DIM, PredictedCovariance>(parameter_map,
                                                             derived().boundParams(boundState).covariance().value()));
 
      // select n measurents with the smallest chi2.
      auto preCalibrator = derived().template preCalibrator<DIM, BaseElementType>();
      auto postCalibrator = derived().template postCalibrator<DIM, BaseElementType>();
      TopCollection<NMeasMax, TheMatchingMeasurement > selected_measurements(numMeasurementsCut);
      for ( const auto &measurement : measurement_range ) {
         TheMatchingMeasurement &matching_measurement=selected_measurements.slot();
         if (forced && postCalibrator) {
            // skip preCalibrator and computeChi2, by doing everything else. We start with what's done if no preCalibrator.
            const auto &m = derived().forwardToCalibrator(measurement);
            matching_measurement.m_measurement = std::make_tuple( m.template localPosition<DIM>(), m.template localCovariance<DIM>(), 0u );
            matching_measurement.m_chi2 = 0.0f;
            matching_measurement.m_sourceLink=measurement;
            if (!selected_measurements.acceptNoSort()) break;
         } else {
            matching_measurement.m_measurement = preCalibrator(geometryContext,
                                                               calibrationContext,
                                                               surface,
                                                               derived().forwardToCalibrator(measurement),
                                                               derived().boundParams(boundState));
            matching_measurement.m_chi2 = computeChi2(matching_measurement.m_measurement.position(),
                                                      matching_measurement.m_measurement.covariance(),
                                                      predicted.first,
                                                      predicted.second);
            // only consider measurements which pass the outlier chi2 cut
            if (matching_measurement.m_chi2<maxChi2Cut.second) {
                matching_measurement.m_sourceLink=measurement;
                selected_measurements.acceptAndSort([](const TheMatchingMeasurement &a,
                                                       const TheMatchingMeasurement &b) {
                   return a.m_chi2 < b.m_chi2;
                });
            }
         }
      }
 
      if (selected_measurements.empty()) {
         auto config_for_surface = m_config.find(surface.geometryId());
         if (config_for_surface != m_config.end()) {
            if (config_for_surface->edgeTolerance.has_value()) {
               // if the  prediction failes the bound test, where the edgeTolerance is expected to be negative
               // then the "hole" is considered to be an edge hole which is not treated as a hole.
               auto local_coords  = derived().boundParams(boundState).parameters().template block<2,1>(0,0);
               if (!surface.insideBounds(local_coords,config_for_surface->edgeTolerance.value())) {
                  result = result.failure(Acts::CombinatorialKalmanFilterError::NoMeasurementExpected);
               }
            }
         }
      }
      else {
 
         // apply final calibration to n-best measurements
         // First Create states without setting information about the calibrated measurement for the selected measurements
         // @TODO first create state then copy measurements, or crete state by state and set measurements ?
         //       the lastter has the "advantage" that the outlier flag can be set individually
         //       the former has the advantage that part of the state creation code is independent of the
         //       the measurement.
 
         Acts::BoundSubspaceIndices boundSubspaceIndices;
         std::copy(parameter_map.begin(), parameter_map.end(), boundSubspaceIndices.begin());
         createStates( selected_measurements.size(),
                       boundState,
                       prevTip,
                       trajectory,
                       boundSubspaceIndices,
                       *result,
                       logger,
                       (!selected_measurements.empty()
                        ? selected_measurements.getSlot( *(selected_measurements.begin())).m_measurement.flags().isOutlier() || postCalibrator
                        : false) );
         assert( result->size() == selected_measurements.size() );
 
         if (postCalibrator) {
            // either run post selection calibrator which will set the calibrated position and covariance
            // directly to the state.
            auto predicted_pos_dbl = predicted.first.template cast<double>();
            auto predicted_cov_dbl = predicted.second.template cast<double>();
            unsigned int state_i=0u;
            for (typename TopCollection<NMeasMax, MeasCovPair >::IndexType
                    idx: selected_measurements) {
               assert( state_i < result->size());
               TrackStateProxy trackState( trajectory.getTrackState( (*result)[state_i] ) );
               TheMatchingMeasurement &a_selected_measurement = selected_measurements.getSlot(idx);
               trackState.setUncalibratedSourceLink(ActsTrk::detail::MeasurementCalibratorBase::pack(std::move(a_selected_measurement.m_sourceLink.value())));
 
               // apply the calibration
               postCalibrator(geometryContext,
                              calibrationContext,
                              derived().forwardToCalibrator(a_selected_measurement.m_sourceLink.value()),
                              trackState
                              );
               // update chi2 using calibrated measurement
               trackState.chi2() = computeChi2(trackState.template calibrated<DIM>(),
                                               trackState.template calibratedCovariance<DIM>(),
                                               predicted_pos_dbl,
                                               predicted_cov_dbl);
               // ... and set outlier flag
               // and share or create storage for filtered.
               bool is_outlier = !forced && trackState.chi2() >= maxChi2Cut.first;
               if (is_outlier) {
                  trackState.typeFlags().setIsOutlier();
                  trackState.shareFrom(trackState, Acts::TrackStatePropMask::Predicted,
                                       Acts::TrackStatePropMask::Filtered);
               } else {
                  trackState.typeFlags().setIsMeasurement();
                  trackState.addComponents(Acts::TrackStatePropMask::Filtered);
               }
               ++state_i;
            }
         }
         else {
            // or copy the pre-selection calibrated position and covariance to the
            // track states and adjust the flags.
            unsigned int state_i=0u;
            for (typename TopCollection<NMeasMax, MeasCovPair >::IndexType
                    idx: selected_measurements) {
               assert( state_i < result->size());
               TrackStateProxy trackState( trajectory.getTrackState( (*result)[state_i] ) );
               TheMatchingMeasurement &a_selected_measurement = selected_measurements.getSlot(idx);
 
               bool is_outlier = !forced && a_selected_measurement.m_chi2 >= maxChi2Cut.first;
               auto type_flags = trackState.typeFlags();
               if (is_outlier) {
                  // the filtered parameters will not be computed for outliers, so has to be shared from the prediction.
                  type_flags.setIsOutlier();
                  trackState.shareFrom(trackState, Acts::TrackStatePropMask::Predicted,
                                       Acts::TrackStatePropMask::Filtered);
               } else {
                  // for measurements the filtered parameters will be computed, so space needs to be allocated.
                  type_flags.setIsMeasurement();
                  trackState.addComponents(Acts::TrackStatePropMask::Filtered);
               }
               if (a_selected_measurement.m_measurement.flags().isSplitHit()) {
                  type_flags.setIsSplitHit();
               }
 
               trackState.setUncalibratedSourceLink(ActsTrk::detail::MeasurementCalibratorBase
                                                        ::pack(std::move(a_selected_measurement.m_sourceLink.value())));
 
               trackState.allocateCalibrated(DIM);
               trackState.template calibrated<DIM>()
                  = MeasurementSelectorMatrixTraits::matrixTypeCast<typename MeasurementSelectorTraits<derived_t>
                                                                    ::MatrixFloatType>(a_selected_measurement.m_measurement.position());
               trackState.template calibratedCovariance<DIM>()
                  = MeasurementSelectorMatrixTraits::matrixTypeCast<typename MeasurementSelectorTraits<derived_t>
                                                                    ::MatrixFloatType>(a_selected_measurement.m_measurement.covariance());
               trackState.chi2() = a_selected_measurement.m_chi2;
               ++state_i;
            }
         }
      }
 
      return result;
   }
 
   template <typename parameters_t>
   static std::size_t getEtaBin(const parameters_t& boundParameters,
                                const std::vector<float> &etaBins) {
      if (etaBins.empty()) {
         return 0u;  // shortcut if no etaBins
      }
      const float eta = std::abs(std::atanh(std::cos(boundParameters.parameters()[Acts::eBoundTheta])));
      std::size_t bin = 0;
      for (auto etaBin : etaBins) {
         if (etaBin >= eta) {
            break;
         }
         bin++;
      }
      return bin;
   }
 
public:
   // get numMeasurement and maxChi2 cuts for the given surface
   // @TODO should the cuts just depend on the surface or really on the bound parameters (i.e. bound eta)
   std::tuple< std::size_t, std::pair<float,float> > getCuts(const Acts::Surface& surface,
                                                             const T_BoundState& boundState,
                                                             const Acts::Logger& logger) const {
      std::tuple< std::size_t, std::pair<float,float> > result;
      std::size_t &numMeasurementsCut = std::get<0>(result);
      std::pair<float,float> &maxChi2Cut = std::get<1>(result);
      // Get geoID of this surface
      auto geoID = surface.geometryId();
      // Find the appropriate cuts
      auto cuts = m_config.find(geoID);
      if (cuts == m_config.end()) {
         // indicates failure
         numMeasurementsCut = 0;
      }
      else {
         // num measurement Cut
         // getchi2 cut
         std::size_t eta_bin = getEtaBin(derived().boundParams(boundState), cuts->etaBins);
         numMeasurementsCut = (!cuts->numMeasurementsCutOff.empty()
                               ? cuts->numMeasurementsCutOff[ std::min(cuts->numMeasurementsCutOff.size()-1, eta_bin) ]
                               : NMeasMax);
         maxChi2Cut = ! cuts->chi2CutOff.empty()
            ? cuts->chi2CutOff[ std::min(cuts->chi2CutOff.size()-1, eta_bin) ]
            : std::make_pair<float,float>(std::numeric_limits<float>::max(),
                                          std::numeric_limits<float>::max());
         ACTS_VERBOSE("Get cut for eta-bin="
                      << (eta_bin < cuts->etaBins.size() ? cuts->etaBins[eta_bin] : std::numeric_limits<float>::max())
                      << ": chi2 (max,max-outlier) " << maxChi2Cut.first << ", " << maxChi2Cut.second
                      << " max.meas." << numMeasurementsCut);
      }
      return result;
   }
};
 
// Measurement selector which calls a type specific selection method
// the measurement selector expects a list of measurement containers to be provided
// by the source link iterators as well as the index of the container to be used for the
// given measurement range, which is defined by the source link iterators.
// the measurement range must refer to a contiguous range of measurements contained
// within a single measurement container.
// The "container" itself is a variant of particular measurement containers with associated
// dimension i.e. number of coordinates per measurement. A member funcion specific to one of
// the alternatives of the variant will be called to perform the selection and track state
// creation.
template <std::size_t NMeasMax,
          typename derived_t,
          typename measurement_container_variant_t >
struct MeasurementSelectorWithDispatch : public MeasurementSelectorBase< NMeasMax, MeasurementSelectorTraits<derived_t>::s_dimMax, derived_t> {
 
   using Base = MeasurementSelectorBase< NMeasMax, MeasurementSelectorTraits<derived_t>::s_dimMax, derived_t>;
   using T_BoundState = typename Base::T_BoundState;
   using trajectory_t = typename Base::trajectory_t;
   using TrackStateProxy = typename Base::TrackStateProxy;
   static constexpr std::size_t s_maxBranchesPerSurface = Base::s_maxBranchesPerSurface;
 
   // helper to get the maximum number of measurement diemsions.
   static constexpr std::size_t dimMax() {
      return MeasurementSelectorTraits<derived_t>::s_dimMax;
   }
 
  Acts::Result<boost::container::small_vector< typename TrackStateProxy::IndexType, s_maxBranchesPerSurface> >
  createTrackStates(const Acts::GeometryContext& geometryContext,
                    const Acts::CalibrationContext& calibrationContext,
                    const Acts::Surface& surface,
                    const T_BoundState& boundState,
                    typename TrackStateProxy::IndexType prevTip,
                    [[maybe_unused]] std::vector<TrackStateProxy>& trackStateCandidates,
                    trajectory_t& trajectory,
                    const Acts::Logger& logger) const {
 
      Acts::Result<boost::container::small_vector< typename TrackStateProxy::IndexType, s_maxBranchesPerSurface> >
         result = Acts::CombinatorialKalmanFilterError::MeasurementSelectionFailed;
      // get associated measurement container and the relevant measurement range for the given surface.
      auto [a_measurement_container_variant_ptr, range, forced] = this->derived().containerAndRange(surface);
      if (!forced && !this->derived().expectMeasurements( surface, a_measurement_container_variant_ptr, range)) {
         result = result.failure(Acts::CombinatorialKalmanFilterError::NoMeasurementExpected);
         return result;
      }
      if (!range.empty()) {
         auto [numMeasurementsCut, maxChi2Cut] = this->getCuts(surface,boundState, logger);
         // numMeasurementsCut is == 0 in case getCuts failed
         // anyway cannot select anything if numMeasurementsCut==0;
         if (numMeasurementsCut>0) {
 
            result = std::visit( [this,
                                  &geometryContext,
                                  &calibrationContext,
                                  &surface,
                                  &boundState,
                                  &range,
                                  prevTip,
                                  &trajectory,
                                  &logger,
                                  numMeasurementsCut,
                                  &maxChi2Cut,
                                  forced] (const auto &measurement_container_with_dimension) {
               using ArgType = std::remove_cv_t<std::remove_reference_t< decltype(measurement_container_with_dimension) > >;
               constexpr std::size_t DIM = ArgType::dimension();
               auto measurement_range = this->derived().rangeForContainer(measurement_container_with_dimension,range);
               return
                  this->template
                  selectMeasurementsCreateTrackStates<DIM, decltype(measurement_range)>(geometryContext,
                                                           calibrationContext,
                                                           surface,
                                                           boundState,
                                                           std::move(measurement_range),
                                                           prevTip,
                                                           trajectory,
                                                           logger,
                                                           numMeasurementsCut,
                                                           maxChi2Cut,
                                                           forced);
            },
            *a_measurement_container_variant_ptr);
         }
      }
      else {
         result = Acts::Result<boost::container::small_vector< typename TrackStateProxy::IndexType, s_maxBranchesPerSurface> >::success({});
      }
      return result;
   }
};
 
template <std::size_t NMeasMax,
          typename derived_t,
          typename measurement_container_variant_t>
struct MeasurementSelectorBaseImpl : public MeasurementSelectorWithDispatch<NMeasMax, derived_t, measurement_container_variant_t> {
   using T_BoundState = MeasurementSelectorTraits<derived_t>::BoundState;
   using TrackStateProxy = MeasurementSelectorTraits<derived_t>::TrackStateProxy;
   using abstract_measurement_range_t = MeasurementSelectorTraits<derived_t>::abstract_measurement_range_t;
 
   // get the bound parameters and covariance of the bound state
   static const MeasurementSelectorTraits<derived_t>::BoundTrackParameters &boundParams(const T_BoundState &boundState)  {
      return std::get<0>(boundState);
   }
 
   // get the jacobi matrix of the bound state
   static const MeasurementSelectorTraits<derived_t>::BoundMatrix &boundJacobiMatrix(const T_BoundState &boundState)  {
      return std::get<1>(boundState);
   }
   // get the accumulated path length of the bound state
   static double pathLength(const T_BoundState &boundState)  {
      return std::get<2>(boundState);
   }
 
   // perform simple transformation to cerate the type
   // that is passed to the calibrator.
   // by default pass the measurement by reference not pointer
   template <typename T>
   static const auto &forwardToCalibrator(const T &a) {
      if constexpr( std::is_same<T, std::remove_pointer_t<T> >::value ) {
         return a;
      }
      else {
         return *a;
      }
   }
 
 
   // get mapping between bound state parameters and coordinates in the measurement domain
   // in most cases this is just the identity operation e.g. loc0 -> coord0 and loc1 -> coord1
   // i.e. map[0]=0; map[1]=1;
   // @TODO should this be measurement type specific, or is dimension and the surface good enough ?
   template <std::size_t DIM>
   ParameterMapping::type<DIM>
   parameterMap(const Acts::GeometryContext&,
                const Acts::CalibrationContext&,
                const Acts::Surface&) {
      return ParameterMapping::identity<DIM>();
   }
 
   struct CalibratedMeasurementTraits {
      // the types used for the calibrated measurement and covariance
      template <std::size_t DIM>
      using Measurement = typename MeasurementSelectorTraits<derived_t>::template CalibratedMeasurement<DIM>;
      template <std::size_t DIM>
      using MeasurementCovariance = typename MeasurementSelectorTraits<derived_t>::template CalibratedMeasurementCovariance<DIM>;
   };
 
   // By default the methods postCalibrator and preCalibrator return delegates:
   template <std::size_t DIM, typename measurement_t>
   using PreCalibrator = Acts::Delegate<
      std::tuple < typename MeasurementSelectorTraits<derived_t>::template PreSelectionMeasurement<DIM>,
                   typename MeasurementSelectorTraits<derived_t>::template PreSelectionMeasurementCovariance<DIM>,
                   unsigned int>
                (const Acts::GeometryContext&,
                 const Acts::CalibrationContext&,
                 const Acts::Surface&,
                 const measurement_t &,
                 const typename MeasurementSelectorTraits<derived_t>::BoundTrackParameters &)>;
 
   // Since the measurement types used during measurement selection and after measurement selection
   // might be different so are the types of the calibrator delegates
   template <std::size_t DIM, typename measurement_t>
   using PostCalibrator = Acts::Delegate<
                void
                (const Acts::GeometryContext&,
                 const Acts::CalibrationContext&,
                 const measurement_t &,
                 TrackStateProxy &)>;

   template <std::size_t DIM, typename measurement_t>
   const PostCalibrator<DIM, measurement_t> &
   postCalibrator() const; // not implemented
 
   // the "calibrator" which is used during the measuremnt selection
   //  this delegate must be connected or be a valid functor. If not this likely will lead to an
   //  exception or seg fault.
   template <std::size_t DIM, typename measurement_t>
   const PreCalibrator<DIM, measurement_t> &
   preCalibrator() const; // not implemented
 

   std::tuple<const measurement_container_variant_t &, abstract_measurement_range_t>
   containerAndRange(const Acts::Surface &surface) const; // not implemented

   bool expectMeasurements([[maybe_unused]] const Acts::Surface &surface,
                           [[maybe_unused]] const measurement_container_variant_t *container_variant_ptr,
                           const abstract_measurement_range_t &abstract_range) const;

   template <typename measurement_container_t>
   static auto
   rangeForContainer(const measurement_container_t &concrete_container, const abstract_measurement_range_t &abstract_range);
};