B-Tagging at CDF and D

1
B-Tagging at CDF and DØLessons for LHC

• Thomas Wright
• University of Michigan
• For the CDF and DØ Collaborations
• Hadron Collider Physics Symposium
• May 26, 2006

2
Introduction

• Lots of interesting physics involves high-pT
  b-quarks
• Top production - BR(t?Wb) 100
• Higgs searches
• For mH lt 135 GeV/c2, H?bb is the most common
  decay
• SUSY bbg ? bbA ? bbbb at high tan?
• Sparticle searches sbottom, stop decays into
  bX
• Fortunately, there are ways to tell if a jet is a
  b-jet
• Outline
• Unique characteristics of b-jets
• CDF/DØ detector subsystems relevant to b-tagging
• B-tagging algorithms
• Efficiency and fake rate measurements
• Recent algorithmic advancements
• High luminosity and thoughts on the LHC
• Conclusions

3
Signatures of B-Jets

• High b-hadron mass
• 5.3 GeV/c2
• Relatively long lifetime
• 1.5 ps ? c? 450 ?m
• Hard fragmentation
• b-hadron retains 70 of b-quark momentum
• Put em all together and you get
• High-pT tracks
• Large impact parameters
• Secondary vertices (few mm)
• Lepton production
• 10 b ? l?
• Also 10 c/c ? l?
• High-pT tracks
• High-pT relative to b-jet

4
The CDF Detector

• COT 96-layer wire drift chamber
• 4 axial, 4 stereo superlayers
• 1.4 T magnetic field
• Central lepton ID out to ?lt1.1
• Silicon tracker
• 95 cm long Layer 00
• Radiation hard
• 3 SVX-II barrels, each 29 cm long

• L00 singled-sided, r 1.2 cm
• SVX-II 5 double-sided layers
• L0 r? rz, r2.5 cm
• L1 r? rz, r4.5 cm
• L2 r? 1.2 stereo, r6.5 cm
• L3 r? rz, r8.5 cm
• L4 r? 1.2 stereo, r10.5 cm
• Total sensor area 6 m2
• 720k electronics channels

5
The DØ Detector

• Central Fiber Tracker
• 8 layers of scintillating fiber (axial and
  stereo)
• 20 lt r lt 51 cm in 2 T B-field
• Unified muon system covers ?lt2
• Silicon tracker
• 6 barrels, 4 layers each, 1 m
• Disks extend coverage to ?lt3
• 790k electronics channels

• Original silicon starts at r 2.7 cm
• Layers 13 r? rz
• Layers 24 r? 2 stereo
• Central disks ?15 stereo
• Forward disks ?7.5 stereo
• New radiation-hard Layer 0
• Impact parameter resolution improves going to r
  1.6 cm
• Mitigates possible Layer 1 efficiency loss as
  dose increases

6
Common Issues

• Reference for impact parameter calculations
• Tevatron beam profile is 30 ?m in xy
• Can improve by computing event-by-event primary
  vertex
• Fit all tracks to a common point discard
  outliers iterate
• Final resolution is 10-30 ?m depending on event
  topology
• Poorly-reconstructed tracks are a killer
  quality cuts are critical
• Both experiments use silicon hit multiplicity and
  fit ?2
• Remove KS/? decay products
• Raising pT cuts helps control fake tag rate
• Inactive detector regions and data/simulation
  mismatch
• CDF model inactive silicon ladders in the
  simulation run-by-run
• DØ everything relative to taggable jets
  (couple of silicon tracks) get taggability
  rate from data (80)
• Both experiments apply a scale factor to
  simulation efficiency to match whats seen in the
  data (0.7 0.9, depending on tagger)

7
Soft Lepton Tagging

• Tag b-tags by identifying a lepton from b or c
  decay
• CDF DØ both do it for muons
• ID requirements somewhat different than high-pT
  case
• Cant use calorimeter energy
• Track-stub angular matching is more effective
• Typical ID efficiency 80-90
• Per-jet efficiency 10
• Mis-ID rates 0.5 per track
• Electrons are more difficult
• Had it in Run I
• Under development at CDF

CDF

• Soft lepton taggers use different information
  than lifetime taggers
• Can use overlap rate for calibration

8
Impact Parameter Taggers

• Count displaced tracks (DØ)
• three 2? or two 3?
• sign IPs against jet direction
• Jet probability (CDF DØ)
• Joint probability for all tracks to come from
  primary vertex
• Track resolution derived from negative IP tracks
• Use only positive IP tracks in probability
  calculation
• Displaced vertex (CDF DØ)
• Fit displaced tracks (above pT cuts) to a common
  vertex
• Prune tracks and cut on fit ?2
• Cut on Lxy significance
• All algorithms can be tuned by adjusting cuts
  2-6 operating points

DØ
DØ
CDF
9
Efficiency Measurement (CDF)

• Based on 8 GeV/c electron and muon data samples
• Tag away jet to enhance b-fraction
• Generate matching MC samples
• Method A muon pTrel fits
• Fit muon pTrel against jet in tagged and untagged
  jet
• Extract numbers of b-jets in each and compute
  efficiency
• Systematics 3, mostly from modeling of pTrel
  templates
• Method B electron double-tags
• Infer non-b component from electron jet
  single-tag rate and conversion sample
• Systematics 5, mostly from mistag subtraction

10
Efficiency Measurement (CDF)

• Systematics quoted earlier for each method were
  only the internal ones
• Additional systematics related to extrapolation
  to physics samples (in particular, b-jets from
  top)
• Jet ET dependence
• Convolute binned scale factor with ET spectrum
  from top
• 7 uncertainty
• Similar procedure for jet ?
• 1 uncertainty
• 2 uncertainty assigned for differences between
  semileptonic/generic B-decays
• Total uncertainty on data/MC scaled factor 7.3

11
Efficiency Results (CDF)

• CDF displaced vertex tagger
• Top pair simulation, scaled by data/MC tag
  efficiency ratio
• Width of the bands shows the systematic errors

12
Efficiency Measurement (DØ)

• Based on low-pT muon sample
• Require away jet similar to CDF
• Define a muon tag pTrel gt 0.7
• Count events with
• Muon tag
• B-tag (any of the IP taggers)
• Both muon tag and b-tag
• Same three, when away jet is tagged
• From these rates, can solve for
• Muon tag efficiency
• B-tag efficiency
• Sample b-fractions
• Systematic uncertainties 3
• Data statistical errors dominate at high jet ET
  (similar to CDFs 7 from convoluting with top
  b-jet spectrum)

JLIP
total uncertainty
13
Efficiency Results (DØ)

• TRFb MC efficiency, scaled by data/MC
  efficiency ratio
• Jet probability tagger for various cut values
• Combine ET and ? parametrizations into a 2D
  function
• Used to predict tag rates in Monte Carlo samples

14
Fake Tag Rates

• Fake tags (mistags) mostly from misreconstructed
  tracks
• /- symmetry
• Estimate fake rates by
• Using tracks with negative signed impact
  parameters
• Using displaced vertices behind the primary
  w.r.t. the jet direction
• Both experiments use parametrizations of
  negative tag rates (in ET, ?, etc) to predict
  mistag rate in data samples
• Complications
• Negative tags from heavy flavor (10-15)
• Mistag rate not exactly symmetric stray KS/?
  and interactions with detector material
• Fit POS-NEG excess in pseudo-c? to estimate
  (40)
• Net effect 30 enhancement over negative tage
  rate

CDF
15
Fake Tag Results (CDF)

• CDF displaced vertex tagger
• Negative tag rates, scaled by the asymmetry
  factor 1.3
• Bands show the systematic errors assigned (15)

16
Fake Tag Results (DØ)

• Jet probability tagger for various cut values
• Includes asymmetry factor (close to unity due to
  cancellation of the two effects)
• Uncertainties are 5-15 depending on operating
  point (looser tag smaller uncertainty)

17
Multivariate Algorithms

• Taggers are correlated, but not 100
• Can gain by combining them
• CDF has a simple combined tagger logical OR
  of displaced vertex and jet probability taggers
• Gain 15-25 efficiency, at cost of factor two
  mistag rate
• Use more information than just the tagger outputs
• Displaced vertex tagger gives you more than just
  yes/no
• Many vertex properties discriminate
  signal/background
• Both experiments have new multivariate taggers
• CDF start with displaced vertex tag, try to
  reject charm/light while preserving b-tags
• DØ no displaced vertex prerequisite can
  optimize for better purity or enchanced efficiency

18
Multivariate Tagger (CDF)

• Two 16-variable neural networks
• Vertex mass, Lxy, ?2, pT
• High-IP track pT, multiplicity
• Jet probability
• and so much more!
• Train one to separate b-vs-light, the other
  b-vs-c
• Choose cuts to preserve 90 of b
• Reject 45 of c, 65 of light
• Already in use
• Top cross section
• WH search
• Similar, dedicatedalgorithm forsingle-top search

19
Multivariate Tagger (DØ)

• 7-variable neural network
• Vertex mass, Lxy/?, ?2
• Number of vertices/jet
• Jet total and displaced track multiplicities
• Jet probability
• Vertex tagger uses a super-loose tune to get
  info on more jets
• Train to separate b-vs-light
• Can improve efficiency or mistag by 25, up to a
  factor of two in the efficiency/purity extremes
• Efficiency and fake rates have been measured in
  data
• Ready for analysis use

20
B-Tagging at High Luminosity

• Starting to study effects of high luminosity on
  b-tagging (results are extremely preliminary)
• Nvtx luminosity
• At 3E32 cm-2s-1, ltNvtxgt 4
• MC study of top events with extra minimum-bias
  overlaid indicates that efficiency decreases
• High tracker occupancy
• Hits merge together
• Pattern recognition harder
• Negative tag rates in data rise
• Silicon hit merging?
• Another interaction nearby?
• Disabling dE/dx pulse shaping on innermost COT
  layers should help
• Revisit track quality cuts

CDF
CDF
21
Some LHC Thoughts (from a non-LHCer)

• B-tagging at a hadron collider has many years
  history to draw upon
• CDF/DØ learning how to deal with multiple
  interactions
• May be useful for LHC
• Pixelstrip inner detectors will handle occupancy
  better
• Get your calibration triggers in
• Low-pT leptons
• Jets at various thresholds
• Likely have to prescale a lot
• Dont let them get turned off at high
  luminosity!
• Measure everything in data
• But, a well-tuned simulation comes in pretty handy

• Yesterdays discovery is todays calibration
• Measure b-tag efficiency (or data/MC scale
  factor) in top events?
• Parametrize acceptance and backgrounds in SF,
  require single- and double-tagged top cross
  sections equal get ?tt and SF
• Just about feasible at Tevatron - _at_LHC ?

22
Conclusion

• B-tagging is a critical component of the Tevatron
  physics program
• Both experiments have robust and well-understood
  tools
• We are still building upon them and making
  improvements
• Entering a new luminosity regime
• Some work ahead to understand/minimize the
  effects
• Tevatron experience should be useful at LHC
• B-tagging is a fun and challenging topic
• Design/calibration of taggers has a lot of
  physics in it
• Not a solved problem still room for good ideas