Timing for pileup mitigation in forward jets (& more) F. Rubbo, A. Schwartzman, 9/4/2015 1
Status Recap: Developed and characterized timing algorithms for pileup mitigation exploiting features of LHC pileup-leveling with crab-cavities. INT Note (work in progress) and last talk at Upgrade Physics meeting. Today s result: Investigating timing algorithms independent of collision configuration. Applications: pileup/hard-scatter jet tagging selection of VBF topologies. 2
Geometry and event simulation Simulated as disk at z=3.5 m with η coverage [2.5,4.3]. Timing detector: disk at z=3.5 m with η coverage [2.5,4.3]. Pythia simulation of signal overlapping with 200 pileup vertices, keeping the full particle record. η=0 absolute time η=2.5 t1 = ths+d1/v1 d1 = (z0-zhs)/tanh(η1) zhs η1 Run jet finding (fastjet) on the HS event only (HS truth jets) and on the full (signal+pileup) event record ( reco jets). zpu t2 = tpu+d2/v2 Apply jet area subtraction. z0=3.5 m 3
Space-time degeneracy of collisions Run 1-2 bunch configuration: ct vertex [m] z0 z vertex [m] Space-time PU interaction probability density is degenerate for head-on collisions. 4
Timing as discriminant The inclusive particle time distribution for head-on collisions provides little discrimination power between HS and PU. Two approaches: Remove time-space degeneracy of the interaction region with LHC crab cavities. Exploit jet and topology features to extract discriminating information (this talk). 5
Crab cavities for pileup leveling The crab-kissing configuration (ψ=5) squeezes the time component of the interaction region while keeping the same spatial spread. Pseudo-rectangular bunches flatten the spatial distributions, reducing the pileup density. head-on collisions crab-kissing ct vertex [m] ct vertex [m] z vertex [m] z vertex [m] 6
Timing as discriminant (w/ CK) The CK configuration allows using directly timing for jet pileup mitigation (HS jets: time <threshold). Relies on LHC upgrade and bunch crossing configuration. 1.0 0.8 No Discrimination Pseudo-Rectangular Gaussian 0.6 Fake Rate 0.4 0.2 0.0 0.0 0.2 0.4 0.6 0.8 1.0 Efficiency head-on collision - ψ=0 crab-kissing - ψ=5 7
Jet timing substructure 8
Time density The hard-scatter timing (corrected for zpv and η) is the same for all hard-scatter particles, within resolution. Look for timing clusters! R=0.1 9
Time density HS jets have a large number of HS time measurements while PU time measurements are evenly distributed. PU time measurements in (stochastic) PU jets are evenly distributed. HS jet PU jet 10
Density difference Build a HS vs PU discriminant: Find timing clusters with gaussian kernel density estimation. Order cluster by decreasing density. Δd = density[0] - density[1] Δd Δd 11
Density difference - performance rejection ~5 @80% efficiency for σt=0 ps performance quickly degrades with resolution. encouraging first results indicating timing-based PU. suppression w/o crab-kissing is possible. Δd > threshold Efficiency for pile-up jets 1.0 0.8 0.6 0.4 0.2 p s = 14 TeV, µ = 200 Pythia8 dijets p T > 20 GeV t=0 ps t=10 ps t=20 ps t=30 ps 0.0 0.0 0.2 0.4 0.6 0.8 1.0 Efficiency for hard-scatter jets 12
Possible improvements Use pt measurement (from multiple layers?) to enhance density of HS time measurement. Improve discriminant definition: e.g. use median density instead of second peak to define baseline. Combine with CK. New ideas..? 13
VBF tagging 14
VBF-like selection For VBF-like topologies, want to select forward and backward HS jets. Double tag Require both forward and backward HS jets Event tag Identify forwardbackward jet events compatible with VBF topology 15
VBF-like selection Tagging individual jet: HS-HS efficiency drops as εhs 2 High HS-PU rate as εhs*εpu PU-PU rate strongly suppressed as εpu*εpu QCD pileup jets have higher εpu because more similar to HS jets. Subdominant at high µ and low pt. VBF signal VBF background HS-HS HS-PU (stochastic) PU-PU PU-PU (same vertex) zpv z 16
Double tag Efficiency for pile-up jets The efficiencies of identifying each jets are uncorrelated VBF signal efficiency is quadratic > Requires very high single tag efficiency > High PU rate. N.B. same applies for a track-based tagger. 1.0 0.8 0.6 0.4 0.2 p s = 14 TeV, µ = 200 Pythia8 dijets p T > 20 GeV t=0 ps t=10 ps t=20 ps t=30 ps single jet tagging 2X Background efficiency 1.0 0.8 0.6 0.4 0.2 t=0 ps t=10 ps t=20 ps t=30 ps p s = 14 TeV, µ = 200 Pythia8 VBF inv. H jet p T > 20 GeV double jet tagging 0.0 0.0 0.2 0.4 0.6 0.8 1.0 Efficiency for hard-scatter jets 0.0 0.0 0.2 0.4 0.6 0.8 1.0 17 Signal efficiency
Event tag Use forward-backward timing measurements to identify jets from the same interaction. η=-2.5 η=0 η=2.5 t1 = tpv+d1/v1 t2 = tpv+d2/v2 d2 = (z0+zpv)/tanh(η2) d1 = (z0-zpv)/tanh(η1) η2 η1 zpv z0 tpv = t1-(z0-zpv)*cotanh(η1)/v2 = t2-(z0+zpv)*cotanh(η2)/v2 18
Forward-backward time difference Select events with the two leading jets at opposite η ( η >2.5). Use kernel density estimation to extract absolute timing of each jet. Given zpv, compute Δt = tpv 1 -tpv 2. Tails in HS-HS due to relativistic approximation: not all time measurements correspond to particles with v=c. pt info could help. Using jet η as particle η. Room for improvement by tweaking density-based time algorithm. 19
Performance Excellent performance at low efficiency with mild degradation from time resolution. Kink at εs~0.6 due to double gaussian shape of signal Δt distribution (room for improvement). Background efficiency 1.0 0.8 0.6 0.4 0.2 t=0 ps t=10 ps t=20 ps t=30 ps p s = 14 TeV, µ = 200 Pythia8 VBF inv. H jet p T > 20 GeV Event tagging potential improvement from removing Δt tails for HS-HS. 0.0 0.0 0.2 0.4 0.6 0.8 1.0 VBF Signal efficiency 20
Performance Significant improvement wrt double-tag algorithm. Background efficiency 1.0 0.8 0.6 0.4 0.2 t=0 ps t=10 ps Event tag Double tag p s = 14 TeV, µ = 200 Pythia8 VBF inv. H jet p T > 20 GeV improvement 0.0 0.0 0.2 0.4 0.6 0.8 1.0 VBF Signal Signal efficiency 21
Conclusions Use of timing for pileup mitigation is challenging. Studied two approaches: LHC crab-kissing to minimize time spread of collisions. Advanced timing algorithms exploiting jet and event time structure. Crab-kissing enables ~10% PU rate @ 80% HS efficiency. New time-density-based algorithms give ~20% PU @ 80% HS efficiency. Independent of bunch crossing configuration! Many handles for further optimization. New technique for VBF tagging based on timing of forwardbackward jet pairs. To be implemented and tested in full VBF analysis. 22
and next steps Continue algorithm R&D: new (or old) ideas to be (re-)implemented to improve further performance. e.g. vertex identification in VBF events: same algorithm as VBF-tagging in reverse (HS-HS jets >vertex z-position). Dedicated GEANT simulation for more accurate estimation of performance and R&D of detector geometry (see Ariel s slides). The tools are ~ready! Many interesting studies are on the way and help is welcome! 23