CMS Tracking POG Performance Plots For 2017 with PhaseI pixel detector

Track Reconstruction Performance

The reconstructed tracks are selected with the "high purity" flag (see arXiv:1405.6569).
The samples used are ttbar events simulated at sqrt(s)=13 TeV with different pile-up conditions (=0, 35, 50, 70; bunch crossing=25ns).
The releases used are either the one used during Run2 or the current development candidate for 2017 data-taking.
The matching of reconstructed track with simulated ones requires that >75% of the reconstructed track hits come from the simulated track.

General description

The PU scenarios considered are 0, 35, 50, 70.

The plots that follow come always in three:

1. Track reconstruction efficiency vs eta
2. Fake and Duplicate rate vs eta
3. Track reconstruction efficiency vs track production radius

The definition of efficiency, fake rate and duplicate rate follow.

simulated track

track of a simulated particle in the Geant4 detector simulation ("MC truth")

reconstucted track

result of the track reconstruction run on the simulated data

efficiency

it is defined as the number of matched reconstructed tracks divided by number of simulated tracks. The association between reconstructed and simulated tracks has been introduced and explained in the previous paragraph

fake rate

it is defined as the number of non-matched reconstructed tracks divided by number of reconstructed tracks.

duplicate rate

it is defined as the number of reconstructed tracks associated to the very same simulated track divided by number of reconstructed tracks.

2017 tracking

Plot	Description
pdf png	Track reconstruction efficiency per tracking iteration as a function of simulated track pT for the Phase I tracker.
pdf png	Track reconstruction efficiency per tracking iteration as a function of simulated track production vertex radius for the Phase I tracker.

Tracking performance at different PU condition

Plot	Description
pdf png	Track reconstruction efficiency as a function of simulated track for the Phase I tracker at different pileup conditions.
pdf png	Track reconstruction fake rate as a function of reconstructed track for the Phase I tracker at different pileup conditions.
pdf png	Track reconstruction efficiency as a function of simulated track pT for the Phase I tracker at different pileup conditions.
pdf png	Track reconstruction fake rate as a function of reconstructed track pT for the Phase I tracker at different pileup conditions.
pdf png	Track reconstruction efficiency as a function of simulated track production vertex radius for the Phase I tracker at different pileup conditions.
pdf png	Track reconstruction fake rate as a function of radius of the point of closest approach to beam line (PCA) of reconstructed track for the Phase I tracker at different pileup conditions.

Vertex Reconstruction Performance

Reconstructed vertices are selected requiring: ndof>4, |z|< 24 cm, R<2 cm.
Samples used are ttbar simulated at sqrt(s)=13 TeV, 25ns bunch crossing and different average pile-up conditions (=0, 35, 50 and 70).
Release used is the current development candidate for Run2 data-taking.

Plot	Description
pdf png	Number of reconstructed vertices vs number of simulated interactions.

Comparison against 2016 performances

Tracking performance

Plot	Description
pdf png	Track reconstruction efficiency as a function of simulated track η for 2016 and 2017 detectors. The 2017 detector shows better performance than 2016 over all the η spectrum.
pdf png	Track reconstruction fake rate as a function of reconstructed track η for 2016 and 2017 detectors. The 2017 detector shows better performance than 2016 over all the η spectrum.
pdf png	Track reconstruction efficiency as a function of simulated track pT for 2016 and 2017 detectors. The 2017 detector shows better performance than 2016 over all the pT spectrum.
pdf png	Track reconstruction fake rate as a function of reconstructed track pT for 2016 and 2017 detectors. The 2017 detector shows better performance than 2016 over all the pT spectrum.
pdf png	Track reconstruction efficiency as a function of simulated track production vertex radius for 2016 and 2017 detectors. The 2017 detector shows better performance than 2016 especially at small radii, where the new pixel detector plays a vital role at seeding and pattern recognition level.
pdf png	Track reconstruction fake rate as a function of radius of the point of closest approach to beam line (PCA) of reconstructed track for 2016 and 2017 detectors. The 2017 detector shows better performance than 2016 especially at small radii, where the new pixel detector plays a vital role at seeding and pattern recognition level.

Expected resolutions on track parameters

Plot	Description
pdf png	Track d0 (transverse impact point) resolution as a function of the simulated track η for 2016 and 2017 detectors The 2017 detector shows better performance than 2016 over all the η spectrum.
pdf png	Track dz (longitudinal impact point) resolution as a function of the simulated track η for 2016 and 2017 detectors The 2017 detector shows better performance than 2016 over all the η spectrum.
pdf png	Track pT resolution as a function of the simulated track pT for 2016 and 2017 detectors. The performances between 2016 and 2017 are comparable.
pdf png	Track pT resolution as a function of the simulated track η for 2016 and 2017 detectors. The performances between 2016 and 2017 are comparable. The pT resolution improves in	η	1.2-1.6, because the 4th pixel layer yields better precision on the track extrapolation to the strip tracker in the pixel barrel-forward transition region.

General Vertex Reconstruction Performance

Reconstructed vertices are selected requiring: ndof>4, |z|< 24 cm, R<2 cm.
Samples used are ttbar simulated at sqrt(s)=13 TeV, 25ns bunch crossing and average pile-up of 35.
Release used is the current development candidate for 2017 data-taking.
Matching with simulated vertices requiring: |deltaZ|<1 mm AND |deltaZ|<3sigmaZ (where sigmaZ is the reconstructed vertex Z uncertainty).

Plot	Description
pdf png	Number of reconstructed vertices vs number of simulated interactions.
pdf png	Vertex merge rate as a function of the distance to closest other vertex in z (right). Vertex merge rate is defined as the number of reconstructed vertices matched to at least two simulated vertices divided by the number of all reconstructed vertices.

Vertex Resolutions

Plot	Description
pdf png	Vertex transverse resolution as a function of the number of tracks used in the vertex fit. The 2017 detector shows better performance than 2016 detector.
pdf png	Vertex longitudinal resolution as a function of the number of tracks used in the vertex fit. The 2017 detector shows better performance than 2016 detector.

Comparison using CA seeding.

We already noticed in the previous sections that the increase in the combinatoric problem shows up mainly at seeding level, while, if seeds are of good-enough quality, the overall time spent in the patter recognition step does not suffer too much. For this reasons a huge effort has been spent during the past years to improve the seeding level algorithms. A novel approach based on Cellular Automaton(CA) techniques has been developed to created seeds. The algorithm has been developed in a GPU friendly way. The current plan is to switch the full tracking seeding to use the CA seeding code, that, as we will see, guarantees much better seeding time, while preserving if not improving the overall outcome in terms of good tracks reconstructed.

Generic Tracking Performances vs. standard seeding

Plot	Description
pdf png	Track reconstruction efficiency as a function of simulated track eta for standard seeding and Cellular Automaton (CA) seeding.
pdf png	Track reconstruction fake rate as a function of reconstructed track eta for standard seeding and Cellular Automaton (CA) seeding.

2017 tracking with CA seeeding

Plot	Description
pdf png	Track reconstruction efficiency per tracking iteration as a function of simulated track pT for the Phase I tracker.
pdf png	Track reconstruction efficiency per tracking iteration as a function of simulated track production vertex radius for the Phase I tracker.

Tracking performance at different PU condition

Plot	Description
pdf png	Track reconstruction efficiency as a function of simulated track for the Phase I tracker at different pileup conditions.
pdf png	Track reconstruction fake rate as a function of reconstructed track for the Phase I tracker at different pileup conditions.
pdf png	Track reconstruction efficiency as a function of simulated track pT for the Phase I tracker at different pileup conditions.
pdf png	Track reconstruction fake rate as a function of reconstructed track pT for the Phase I tracker at different pileup conditions.
pdf png	Track reconstruction efficiency as a function of simulated track production vertex radius for the Phase I tracker at different pileup conditions.
pdf png	Track reconstruction fake rate as a function of radius of the point of closest approach to beam line (PCA) of reconstructed track for the Phase I tracker at different pileup conditions.

Comparison of tracking performance against 2016 performances

Plot	Description
pdf png	Track reconstruction efficiency as a function of simulated track η for 2016 and 2017 detectors, latter with Cellular Automaton (CA) seeding.. The 2017 detector shows better performance than 2016 over all the η spectrum.
pdf png	Track reconstruction fake rate as a function of reconstructed track η for 2016 and 2017 detectors, latter with Cellular Automaton (CA) seeding.. The 2017 detector shows better performance than 2016 over all the η spectrum.
pdf png	Track reconstruction efficiency as a function of simulated track pT for 2016 and 2017 detectors, latter with Cellular Automaton (CA) seeding.. The 2017 detector shows better performance than 2016 over all the pT spectrum.
pdf png	Track reconstruction fake rate as a function of reconstructed track pT for 2016 and 2017 detectors, latter with Cellular Automaton (CA) seeding.. The 2017 detector shows better performance than 2016 over all the pT spectrum.
pdf png	Track reconstruction efficiency as a function of simulated track production vertex radius for 2016 and 2017 detectors, latter with Cellular Automaton (CA) seeding. The 2017 detector shows better performance than 2016 especially at small radii, where the new pixel detector plays a vital role at seeding and pattern recognition level.
pdf png	Track reconstruction fake rate as a function of radius of the point of closest approach to beam line (PCA) of reconstructed track for 2016 and 2017 detectors, latter with Cellular Automaton (CA) seeding.. The 2017 detector shows better performance than 2016 especially at small radii, where the new pixel detector plays a vital role at seeding and pattern recognition level.

Timing

Plot	Description
pdf png	Time spent in seeding as a function of average pileup for 2016 tracking, 2017 tracking with conventional seeding, and 2017 tracking with Cellular Automaton (CA) seeding. All time measurements are normalized such that 1 = tracking time of 2016 no pileup. The timing of 2017 CA seeding is back at the level of 2016, a remarkable achievement given the increased number of layer combinations involved in the seeding phase with respect to the 2016 case. CA is expected to be faster than conventional seeding because of CA needing fewer and simpler calculations that are localized in memory.
pdf png	Time spent in pattern recognition as a function of average pileup for 2016 tracking, 2017 tracking with conventional seeding, and 2017 tracking with Cellular Automaton (CA) seeding. All time measurements are normalized such that 1 = tracking time of 2016 no pileup. In pattern recognition, CA seeding brings no additional gain, as expected.
pdf png	Time spent in the entire tracking sequence (right) as a function of average pileup for 2016 tracking, 2017 tracking with conventional seeding, and 2017 tracking with Cellular Automaton (CA) seeding. All time measurements are normalized such that 1 = tracking time of 2016 no pileup. The overall gain by the CA comes completely from the gain in the seeding stage. CA is expected to be faster than conventional seeding because of CA needing fewer and simpler calculations that are localized in memory.