tracking at lhcb introduction: tracking performance at lhcb kalman filter technique speed...
TRANSCRIPT
Tracking at LHCb
• Introduction: Tracking Performance at LHCb
• Kalman Filter Technique
• Speed Optimization
• Status & Plans
June 4, 2004 NIKHEF B-physics meeting 2Jeroen van Hunen
Detector response: test-beam data (resolution, efficiency, noise, cross-talk) Spill-over effects included (25 ns bunch spacing)
PYTHIA and GEANT simulation
Trigger simulation: thresholds tuned to get maximal signal efficiencies at limited output rates of 1 MHz (L0) and 40 kHz (L1)
Offline reconstruction: Full pattern recognition (track finding, RICH reco.)
June 4, 2004 NIKHEF B-physics meeting 3Jeroen van Hunen
Track finding strategy
VELO seeds
Long track (forward)
Long track (matched)
T seeds
Upstream track
Downstream track
T track
VELO track
T tracks useful for RICH2 pattern recognition
Long tracks highest quality for physics (good IP & p resolution)Downstream tracks needed for efficient KS finding (good p resolution)Upstream tracks lower p, worse p resolution, but useful for RICH1 pattern recognition
VELO tracks useful for primary vertex reconstruction (good IP resolution)
June 4, 2004 NIKHEF B-physics meeting 4Jeroen van Hunen
Tracking performance
Red: Geant hits
Blue: reconstructed tracks
- 2050 hits assigned to a long track - 98.7% correctly assigned- Efficiency: 94%
On average:26 long tracks11 upstream tracks4 downstream tracks5 T tracks26 VELO tracks
June 4, 2004 NIKHEF B-physics meeting 5Jeroen van Hunen
Tracking performance
Eff = 94% (p > 10 GeV)
Ghost rate = 3%(for pT > 0.5 GeV)
June 4, 2004 NIKHEF B-physics meeting 6Jeroen van Hunen
Tracking performance
IP resolutionp/p
0.37 %p
p
IP 20μm Typical:
June 4, 2004 NIKHEF B-physics meeting 7Jeroen van Hunen
Kalman Filter Technique
June 4, 2004 NIKHEF B-physics meeting 8Jeroen van Hunen
Prediction and filter step
Kalman technique : Efficient iterative solution for a least-squares fit
Detector planes Measurement and error
State vector (x,y,tx,ty,q/p) is updated at each measurement by performing a prediction and filter step.
• prediction Estimate of state at a given z-position based on previous planes
• filter (update) Weighted mean of the prediction and measurement at given z-pos.
Direction of Kalman filter
June 4, 2004 NIKHEF B-physics meeting 9Jeroen van Hunen
Detector Material
Detector material is taken into account by adjusting the state vector and it’s covariance matrix:
• Multiple scattering For small deflection angles Gaussian, increases the covariance matrix elements of tx, ty. Additional factor tuned to give track parameters with pull=1.
• Energy loss:
Charged hadrons + muons ( ) Affects the momentum, but not the covariance matrix. The factor cion is tuned to give a momentum pull that is centered around zero.
Electrons (bremsstrahlung) Affects the momentum and the covariance matrix. (Increase of covariance allows to correct a sudden momentum change LHCb classic only)
0X
c
dx
dE ion
ion
Note : Covariance matrix V:
22 2/
2
1)(
xexG
uxVux
n
T
eV
xG
1
21
||2
1)( 2/ ))(( jjiiij uxuxEV
June 4, 2004 NIKHEF B-physics meeting 10Jeroen van Hunen
Magnetic field
Prediction and with Fk = Propagation matrix.
Fk : determined using Runge-Kutta (The LHCb B-field is not homogeneous and has Bx,By and Bz components). A 5th order adaptive Runge-Kutta is used (the step size is automatically adjusted to obtain a given precision, by comparing with a fourth order Runge-Kutta).
11 ~~
kkkk xFx
Filter (update) Weighted mean of the prediction and measurement.
Smoother Information of last state is propagated backwards to earlier states.
kTkkk
kk QFCFC
11
June 4, 2004 NIKHEF B-physics meeting 11Jeroen van Hunen
Kalman Filter needs a state to start with ( ) taken from pattern recognition
Seeding (example):
Seed for Kalman filter
1~
kx
magnet
1kCCovariance matrix
should be ‘infinite’ in order to use the measurements only one time and create artificially small errors
June 4, 2004 NIKHEF B-physics meeting 12Jeroen van Hunen
Speed Issues
June 4, 2004 NIKHEF B-physics meeting 13Jeroen van Hunen
Long Tracks Filter and Smooth : 17ms per track on (1GHz Pentium III)
• Filtering (prediction and update)
• Smoothing
• Outlier removal
Reference : HLT processing time 60 ms (1GHz Pentium III). HLT should do pattern recognition, track fitting and a selection for specific final states.
States Including the determination of the states at 5 specified z-positions (beam line, RICH1, RICH2) it is 24ms.
From this 24 ms, 21ms is used for the Transport Service (88%).
How long takes the Kalman filtering
( FilterandSmooth = 240s, for 14000 tracks from which 13000 have 3 hits in the Velo and 3 hits in the OT )
Tool to obtain the radiation length of detector material that is encountered by the particleVELO (50%), OT/IT (19%), RICH2
(13%), TT (9%), RICH1 (4%)
June 4, 2004 NIKHEF B-physics meeting 14Jeroen van Hunen
Gaining time
Look for possible improvements in:
-TransportSvc / XML
-Extrapolation (Runge-Kutta)
TransportSvc / XML
• TransportSvc implementation ( C++ and Gaudi)
• XML
• Check if we can simplify
• Volume distribution
• Used logical volumes with solid
• Use assemblies (but reimplemented)
June 4, 2004 NIKHEF B-physics meeting 15Jeroen van Hunen
Gaining time
• Use Transport service differently:
Request the transport service less often, but for longer distances, thus:
• Call transportSvc for long distance
• Store information from transportSvc
• Take small steps in extrapolation using stored information until the deviation of the track with respect to the original estimate is too large.
Needs testing
Extrapolation (Runge-Kutta)
• C++ and Gaudi (only use msgSvc in case of error, etc.)
• Runge-kutta (4th , 5th, other ?)
• Use of extrapolator : Runge-kutta parabolic extrapolator (small steps)
In Velo region: It is 3.8 times faster to call the TransportSvc one time for 80 cm than 10 times for 8 cm
June 4, 2004 NIKHEF B-physics meeting 16Jeroen van Hunen
Other applications of Kalman Technique
Kalman technique is generally implemented because it is fast
Higher level trigger applications
At LHCb:
- HLT processing time 60 ms
-Current track fit 17 ms/track , thus about 0.5 sec/event.
Would like to do the track fit in a few ms/event, thus need to speed up by a factor 100
Seems possible :
- Ignore detector material (gives factor 10)
- Replace/Improve Runge-Kutta (G. Raven : Pt-kick to get through magnet)
- Only prediction and update (no smoother step)
June 4, 2004 NIKHEF B-physics meeting 17Jeroen van Hunen
Status & Plans
• Implement fast Kalman Fit (for HLT, etc.) without material, Runge-kutta, smoother, etc.
(a first version exist, but still has some strange features)
• Try improve use of transport svc, by calling it less often (at the moment 700.000 times for 500 events) and store information.