pileup per particle id

PileUp Per Particle Id Daniele Bertolini MIT Philip Harris CERN Matthew Low University of Chicago Nhan Tran Fermilab/LPC

work soon to appear !

!Mitigation of pileup effects at the LHC

May 16, 2014

PUPPI Daniele Bertolini MIT Philip Harris CERN Matthew Low University of Chicago Nhan Tran Fermilab/LPC

work soon to appear !

!Mitigation of pileup effects at the LHC

May 16, 2014

May 16, 2014 3

• what is PUPPI? !

• defining the algorithm !

• setup and results !

• discussion and extensions

outline

May 16, 2014 4handles on pileup

•global event metric •local shape •tracking/vertexing •precision timing •depth segmentation

• (apologies, not a complete list!) !

• ρ correction/subtraction (area, 4-vector, shape, particle)

groomingtopoclustering

charged hadron subtractionjet cleansing pileup jet ID

• …







• …







• …{•stand on the shoulders of giants!

May 16, 2014 8PUPPI

grooming/local shape

tracking/vertexing information

global event information

depth segmentation

precision timing

following the “jets without jets” paradigm…

May 16, 2014 9PUPPI




depth segmentation

precision timing

following the “jets without jets” paradigm… Define on a per particle basis, before jet clustering, a weight for how likely a particle (or jet constituent) is to be from pileup or the leading vertex, then rescale each particle four momentum by that likelihood

May 16, 2014 10PUPPI




depth segmentation

precision timing

in one algorithm: correct the jet pT

correct the jet mass/shapes

perform pileup jet ID classify particles for MET determination

following the “jets without jets” paradigm… Define on a per particle basis, before jet clustering, a weight for how likely a particle (or jet constituent) is to be from pileup or the leading vertex, then rescale each particle four momentum by that likelihood

May 16, 2014 11





outline

May 16, 2014 12PUPPI overview

[1] define a local metric, α, that differs between pileup (PU) and

leading vertex (LV)

for a particle i with nearby particles jLet’s assume something similar to Particle

Flow inputs (not a requirement): neutral hadrons

charged hadrons from LVcharged hadrons from PU



leading vertex (LV)

for a particle i with nearby particles j

[2] using tracking information (e.g. charged particles) “sample”

the event, define unique distributions of α for PU and LV

Let’s assume something similar to Particle Flow inputs (not a requirement):

neutral hadronscharged hadrons from LVcharged hadrons from PU



leading vertex (LV)




[3] for the neutrals, ask “how PU-like is α for this particle?”, compute a weight

for how un-PU-like (or LV-like) it is





leading vertex (LV)




[3] for the neutrals, ask “how PU-like is α for this particle?”, compute a weight

for how un-PU-like (or LV-like) it is

[4] reweight the four-vector of the particle by this weight, then proceed to cluster the

event as usual



May 16, 2014 16[1] define a metricthere are many possibilities, we have tested many and find these to be near-optimal

example: 2-body system, for a particle i, what does particle j tell us?

⬆ for harder, collinear particles!⬇ for softer, wide angle particles

May 16, 2014 17[1] define a metricthere are many possibilities, we have tested many and find these to be near-optimal

example: 2-body system, for a particle i, what does particle j tell us?

⬆ for harder, collinear particles!⬇ for softer, wide angle particles

define 2 metrics, with and without using tracking information

R0 is the radius around particle i which is considered

!the log regularizes the scale

for event sampling (next slides)

“C” = charged or central, “F” = full or forward

May 16, 2014

Cα -5 0 5 10 15

frac

tion

of p

artic

les

0

0.02

0.04

0.06

LV particles

PU particles

18[2] sample the event

Fα -5 0 5 10 15

frac

tion

of p

artic

les

0

0.05

0.1

0.15

LV particles

PU particlesαF

charged particles with pT > 1 GeV, populated over 200 events

** dĳet events, more details on simulation later

sample the charged particles in the event, get separate distributions of α for LV and PU particles

!Determine the median and (left-side) RMS values for the charged PU

(good) assumption: α for charged PU and neutral PU are similar need to extrapolate αF to the forward region (η-dependence)

αC

does not include particles with

αC = 0.0

May 16, 2014 19[3] compute weight

Use χ2 CDF to determine probability, pi:

weight, wi = 1 - pi

χ2 construction allows for additional information, use χ2 with N DOF)Cα weight (

0 0.2 0.4 0.6 0.8 1

frac

tion

of p

artic

les

-310

-210

-110

LV particles

PU particles

)Fα weight (0 0.2 0.4 0.6 0.8 1

frac

tion

of p

artic

les

-310

-210

-110

LV particles

PU particles

Use a χ2 approximation to derive a probability of pileup

grooming and local shapeglobal event metrics


signed information because we know LV particles lie to the right of

the PU distribution

May 16, 2014 20[4] recompute 4-vectors

•step 1: assuming perfect tracking for charged particles, set wChLV = 1 and wChPU = 0.

•step 2: for neutrals, re-weight 4-vectors, wi × pi •step 3: cut on neutrals with wi < wcut and wi*pTi < βcut

!•wcut and βcut are tunable parameters

•this study: wcut = 0.001•βcut = f(nPV) ≈ 0.7e-2 × npu + 0.1 GeV (central)•βcut = f(nPV) ≈ 1.1e-2 × npu + 0.2 GeV (forward)

•R0, PUPPI cone size, is a tunable parameter•this study: R0 = 0.3

May 16, 2014 21





outline

May 16, 2014 22‘experimental’ setup

•dĳet events, Pythia 8.176 •flat pT spectrum for populating higher pT bins

!•pileup scenarios, nPU = 20-140, every 15 nPU

•not Poisson-distributed !

•tracker exists to |η| < 2.5, all particles out to |η| < 5 •assume perfect tracking to determine if particles are LV or PU

!•neutrals are reconstructed in cells (Δη×Δφ = 0.1×0.1)

•cut on neutrals, pT > 0.1 GeV

May 16, 2014 23input collections

•LV •only the particles from the leading vertex of interest

•PF: like particle flow •all charged and neutral particles

•PFCHS: like particle flow with charged hadron subtraction •all charged and neutral particles except for charged particles from PU in the tracking volume

•PUPPI: like particle flow •particles after the PUPPI algorithm

!•PF and PFCHS jet collections are corrected using fastjet 4-vector grid ρ subtraction

•PFCHS uses its own ρ inside tracker volume and PF ρ outside

May 16, 2014 24event displays

η-4 -2 0 2 4

φ

0

2

4

6

pT (G

eV)

-110

1

10

210

η-4 -2 0 2 4

φ

0

2

4

6

pT (G

eV)

-110

1

10

210

η-4 -2 0 2 4

φ

0

2

4

6

pT (G

eV)

-110

1

10

210

η-4 -2 0 2 4

φ

0

2

4

6pT

(GeV

)

-110

1

10

210

colored cells = LV process, black cells = pileup

LV PF

PFCHS PUPPI

nPU = 80 scenario

May 16, 2014 25jet kinematicsnPU = 80 scenarioNumber of jets with

pT > 25 GeV per event

eta-4 -2 0 2 4

jets

/eve

nt0

0.1

0.2

0.3

0.4LVPFlowPFlowCHSPUPPI

= 80PUAnti-kT (R=0.7), n

n jets 0 5 10 15 20

jets

/eve

nt

0

0.1

0.2

0.3

0.4 LVPFlowPFlowCHSPUPPI


PUPPI performs pileup jet identification reducing the amount of spurious jets formed from overlapping pileup events

May 16, 2014 26jet pT resolutionnPU = 80 scenario

LV)/pT

LV (pT - pT

-0.2 0 0.2

frac

tion

of je

ts

0

0.1

0.2PFlow

= 0.03x = 0.071σ

PFlowCHS = -0.01x = 0.043σ

PUPPI = 0.001x = 0.026σ


| < 2.5ηpT = [100-200] GeV, |

LV)/pT

LV (pT - pT

-0.4 -0.2 0 0.2 0.4

frac

tion

of je

ts

0

0.05

0.1 PFlow = 0.108x = 0.206σ

PFlowCHS = 0.003x = 0.154σ

PUPPI = -0.025x = 0.098σ


| < 2.5ηpT = [25-50] GeV, |

lower pT jets (25-50 GeV) higher pT jets (100-200 GeV)

match PF, PFCHS, PUPPI jets to LV jets and compare the jet pT resolution

May 16, 2014 27jet pT resolution vs. pT

nPU = 80 scenario

pT (GeV)0 100 200 300 400 500

LV)/p

TLV

, (pT

- pT

σ fi

tted

0

0.1

0.2

0.3PFlowPFlowCHSPUPPI


pT (GeV)0 100 200 300 400 500

LV)/p

TLV

fitte

d m

ean,

(pT

- pT

-0.2

-0.1

0

0.1

0.2PFlowPFlowCHSPUPPI


jet pT scale as a function of LV jet pT

jet pT resolution as a function of LV jet pT

May 16, 2014 28jet massnPU = 80 scenario

mass (GeV) 0 50 100

fract

ion

of je

ts

0

0.1

0.2



| < 2.5ηpT = [100-200] GeV, |

trimmed mass (GeV) 0 50 100

fract

ion

of je

ts0

0.1

0.2



| < 2.5ηpT = [100-200] GeV, |

(GeV)LV m - m-50 0 50

frac

tion

of je

ts

0

0.1

0.2PFlow

= 17.335x = 10.149σ

PFlowCHS = 5.926x = 7.191σ

PUPPI = 2.115x = 3.913σ


| < 2.5ηpT = [100-200] GeV, |

(GeV)LV m - m-50 0 50

frac

tion

of je

ts

0

0.1

0.2 PFlow = 27.503x = 11.0σ

PFlowCHS = 12.716x = 9.145σ

PUPPI = 0.408x = 4.069σ


| < 2.5ηpT = [100-200] GeV, |

jet mass trimmed jet mass

PF and PFCHS jets corrected doing trimming, then ρ

4-vector subtraction

May 16, 2014 29jet mass resolutionnPU = 80 scenario

pT (GeV)100 200 300 400 500

(GeV

)LV

, m -

mσ

fitte

d

0

10

20

30PFlowPFlowCHSPUPPI


pT (GeV)100 200 300 400 500

(GeV

)LV

fitte

d m

ean,

m -

m

0

20

40

60PFlowPFlowCHSPUPPI


jet mass scale as a function of LV jet pT

jet mass resolution as a function of LV jet pT

May 16, 2014 30results vs. PU

nPU50 100

(GeV

)G

ENm

- m

5

10

15

20

PUPPIPFlowPFlowCHS

Anti-kT (R=0.7)

| < 2.5ηpT = [100-200] GeV, |

nPU50 100

GEN

/pT

GEN

pT-pT

0.05

0.1

0.15

0.2

PUPPIPFlowPFlowCHS

Anti-kT (R=0.7)

| < 2.5ηpT = [100-200] GeV, |

jet pT resolution as a function of nPU

jet mass resolution as a function of nPU

May 16, 2014 31





outline

May 16, 2014

!•iPUPPI •isolation of leptons and photons using PUPPI-weight particles should have improved stability in high pileup environments!

!•PUPPET •missing transverse energy (MET) determination using input particles from PUPPI algorithm should improve MET resolution and tails

32discussion and extensions

MissX E∆

-100 -50 0 50 100

frac

tion

of e

vent

s

0

0.05

0.1

0.15PFlow

PFlowCHS

PUPPI

= 80PU

Z+jet, n

T EΣ ∆ 0 500 1000 1500 2000

frac

tion

of e

vent

s

0

0.2

0.4

0.6

0.8PFlow

PFlowCHS

PUPPI

= 80PU

Z+jet, n

May 16, 2014 33discussion and extensions•not an algorithm set-in-stone, but rather a framework

•takes some ideas from existing algorithms, uses as much information as possible

•classify particles (or smallest detector units) before clustering •can be used for lepton isolation and MET determination

•can combine experimental and theoretical considerations •include probabilistic information about tracking, depth segmentation and timing !

•needs more rigorous comparisons against other methods •looking forward to working sessions!

•

May 16, 2014 34discussion and extensions

•For the future… !

•would like to understand the definition of the metrics more rigorously from the QCD point-of-view

•while α is IRC-unsafe under collinear splitting of particle i, the weight is IRC-safe !

•consider algorithm outside of the PF framework •i.e. in topoclustering, consider nearby tracks

!•extend jet shape studies beyond jet mass !

•the “M” word: MVA particle classification using several metrics shows modest improvement

May 16, 2014 35summary

•a new algorithm is presented for pileup mitigation at the LHC !

•aims to classify particles before jet clustering •using local information, global event metrics and charged particle tracking •can be extended to use further experimental information !

•showing good performance for jet pT and mass resolution, comparing to ρ-subtracted PF and PFCHS

•need to exercise in detector environment! !

•not the final story, algorithm can be tuned for different experiments or improved further

•

May 16, 2014 36thanks!

•thank you to many people for helpful discussions! !

•We’d like to thank in particular Jeff Berryhill, Dinko Ferencek, Andrew Larkoski, David Krohn, Jesse Thaler, Lian Tao Wang + many CMS colleagues

pileup per particle id

Documents