b-tagging for the high-pT
physics program of ATLAS
Laurent Vacavant
FZU/Academy of Sciences & FJFI/Czech Technical University – Prague – Dec 19 2008
Introduction
• b-jet tagging: identification of jets stemming from the
hadronization of a b-quark, in high-pT processes
  B-physics
• A key ingredient for:
– top quark physics
– Higgs searches and properties
– SUSY and BSM models
• This talk:
–
–
–
–
–
a few physics cases (top, Higgs, SUSY)
b-tagging basics
tagging algorithms
performance
how to measure the performance in data
All material to be available in: Expected Performance of the ATLAS Experiment,
Detector, Trigger and Physics, CERN-OPEN-2008-20
Laurent Vacavant
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
2
Top-quark physics
Muon+jets channel
Top physics @ LHC:
3 jets pT> 40 GeV +
• production (BSM?)
1 jet pT> 30 GeV
• precise top mass
• properties (charge, spin, etc)
Signal extraction:
•σ = 830 (350) pb @ 14 (10) TeV
•favorable situation: S/B 10x
PT(lep) > 20 GeV
higher than Tevatron
No b-tagging
• b-tagging still useful to kill
light jet background
Missing ET > 20 GeV
(ttbar) 23.6%
• At least 1 btag: Ru ~200, b = 63%
Lepton+jets, 100 pb-1:
S/B 2x (4x) better
requiring 1 (2) b-jets
 modest efficiency OK
Laurent Vacavant
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
3
Higgs sector
SM Higgsbb (searches & properties):
• natural interest for low mass region (Γff ∼ mf2)
• revived interest for WH(bb) (w/ high-pT jets)
(Butterworth et al., PRL 100,242001,2008)
• direct production not visible
associated prod.
 ttH(bb) channel, WH channel
≳70% needed !
• H properties: Yukawa coupling with ttH
bbH, H
SUSY Higgses:
• coupling to b enhanced by sinα/cosβ for 2HDM
models like MSSM  @ high tanβ
• production of neutral H in association with
b quarks:
Laurent Vacavant
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
soft b-jets
4
Low mass Higgs: ttH
• ttH(bb) @ LHC (14 TeV), mH=120 GeV:
• σ = 0.519 pb (LO), K-factor=1.2
• complex final state: for l+jets:
1 lepton, 6 jets (4 b-jets), Etmiss
• b-tagging crucial to reject huge
tt+light jets background
cut
ttH
ttbb(EW)
ttbb(QCD)
tt+jets
1 Lepton (fb)
56.9
141
1356
63710
+ 6 jets (fb)
36.2
76.7
665
26214
+ 4 b-jets loose (fb)
16.2
23.4
198
2589
+ 4 b-jets tight (fb)
3.76
4.2
29.6
50.7
500
• excellent efficiency required (4 b-jets)
• systematics critical  performance under
control: b  20% change in selection
eff., Ru  3-10% change
Significance of lepton+jets channel, 30 fb-1: 2.2
but requires precise background extraction
Laurent Vacavant
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
5
Exotics
• Plethora of models
• Often heavy particles decay primarily to heavy quarks  t/b/c
RS Extra-dim: KK-gluon
σ ~15 pb for m=1 TeV (@ 10 TeV)
LR Twin Higgs model
Large branching to top: > 90%
Common pattern for exotics: most often high-mass particles giving
rise to (very) high-pT b-jets ( O(TeV) )  challenging experimentally
Laurent Vacavant
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
6
b-tagging basics
b-tagging basics
Hard fragmentation of b quarks xB ~ 70%
High mass mB ~ 5 GeV
Soft lepton
Lifetime of B hadrons:
c ~ 470 m (mixture B+/B0/Bs) , ~ 390 m (b)
for E(B) ~ 50 GeV, flight length ~ 5 mm, d0 ~ 500 m
 Spatial tagging:
• Signed impact parameter
of tracks (or significance)
• Secondary vertex
Secondary Vertex
B
Jet axis
a0>0
a0<0
Primary vertex
y
x
Laurent Vacavant
 Soft-lepton tagging:
• Low pT electron from B (D)
• Low pT muon from B (D)
(limited by Br: around 20% each)
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
8
b-tagging ingredients
• Jets (direction):
•to assign tracks: R      0.4
•to sign impact parameters
2
2
•Tracks:
• impact parameter resolution
(35m for a central 5 GeV track)
• tracking in (dense) jets
• specific quality requirements
•
•
σ(d0) ~ 10 ⊕ 150/pT√sinθ m
N(hit B-layer) > 0
shared hits, etc
• Primary vertex:
• mostly needed along Z
• @ high-luminosity
(σ~50m)
• Lepton ID (soft leptons)
Laurent Vacavant
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
9
Pattern-recognition inside dense jets
Example of very different treatment (shared hits, errors) :
• Specific to tracking in jets
• NewTracking can be tuned differently (e.g. r12, r14)
• Important for strategy with real data !
 20% diff. in b-tagging
Laurent Vacavant
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
10
Tracks with shared hits
0.95
1
Laurent Vacavant
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
11
Tagging algorithms
Impact parameter-based b-tagging
Impact parameter:
• signed w.r.t. jet direction
• use significance (d0/σ)
Simplest algorithm:
• counting tracks with
large d0 or d0/σ
Laurent Vacavant
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
13
JetProb tagging algorithm
JetProb algorithm:
(à la ALEPH)
• compatibility of tracks with primary vertex
•resolution function extracted from data
Laurent Vacavant
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
14
Likelihood ratio IP taggers
• smoothed & normalized significances for
the two hypotheses: b(S), u(S)
 requires PDFs (from MC or data)
• for each track: ratio b(S)/u(S)
N tr
b( S i )
W jet   ln
u ( Si )
i 1
In 3D:
• use 2D PDF for (d0,z0)
 correlations
Z vs R
Z impact parameter significance
IP likelihood ratios:
R impact parameter significance
N tr
b( Si R , Si Z )
i 1
u ( Si R , Si Z )
W jet   ln
Laurent Vacavant
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
15
Secondary vertex-based b-tagging
Inclusive vertex:
1. Find all two-track vertices (tracks away from primary vertex: L3D/σ > 2)
2. Remove vertices compatibles w/ γ conv., Ks, Λ, material inter.
Ks
Λ
3.
4.
5.
the
Fit all remaining tracks into single inclusive vertex
Remove iteratively worst tracks
Use distance to PV or properties of
fitted vertex for discrimination:
• number of two-track vertices
• mass of the vertex
• energy fraction
N
(NB: Lxy not used, correlated with IP)
Laurent Vacavant
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
16
Secondary vertex-based b-tagging
Basic algorithm:
• SV0 tagger: returns L3D/σ
LR approach: templates (PDFs) for b and
light hypotheses:
• SV1 tagger: 1D for N’, 2D for (M’, F’)
• SV2 tagger: 3D for (N’, M’, F’)
• Fold in also probability  to find vertex
W jet  Wtracks  Wvertex
N track
b( d 0 i , z 0 i )
b( M , F , N )
  log
 log
u ( d 0 i , z0 i )
u(M , F , N )
i 1
F
SVX mass M
M
light jets
b-jets
F ratio 
Laurent Vacavant
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
17
Topological reconstruction of B/D decay
Main idea: try to constrain all tracks from the
B/D hadron decays to intersect the same flight
axis (instead of the common vertex approach)
JetFitter: a new dedicated Kalman filter fitting:
•
•
•
•
categories for topology
vertex mass
energy fraction
significances of {disti}
Laurent Vacavant
~20% improvement
in light jet rej.
promising for b/c
separation
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
light jet rejection
Based on likelihood:
charm jet rejection
Treatment of incomplete/1-prong topologies:
• on average 1.9 (1.7) OK tracks from D (B) decays
18
Tagging with Soft Muons
• Br(bX)+Br(bcX) = 11% + 10%
• Muon reconstruction: high-efficiency down to 4 GeV/c pT (90% above 5 GeV), low
fakes (5p1000)
•
2 steps: association muon-jet (cone ∆R < 0.5), likelihood ratios (1D, 2D) for
b/light hypothesis using pT, pTrel.
 Rejection of 300 for 10% b-tagging efficiency (incl. Br)
 Fake muons vs pile-up/cavern bckgd: 15% impact on rej. (@ 2.1033)
Laurent Vacavant
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
19
Their Performance
Performance of b-tagging algorithms
5% dead pixels
tt
IP3D+SV1, w>4, tt
•Track counting: R~30 @ 60%
•Soft muon: R~300 @ 10% (i.e. 80% w/ BR)
•Soft electron: R~100 @ 8%
•HLT: R~20 @ 60%
•Charm rejection: 5 to 7 @ 60%, up to
20 with JetFitter
Laurent Vacavant
Factorization (≠ channels):
dependency on jet pT,  and env. (∆R)
Degraded performance:
•low pT: MS, secondaries
•high η: lever-arm (z0),secondaries
•high pT: next slide
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
21
Tagging high and very high-pT jets
Challenges:
• fixed ∆R for track/jet assoc 
dilution for jets of high pT
Example: Z’(2 TeV)  bb, uu
after retuning of IP3D+SV1:
(x3 worse otherwise)
• gains (x2,x3) possible for non-isolated jets
• pattern-recognition issues
• ‘late’ B-decays
Fraction of jets (WH400):
 Require dedicated treatment
for clustering, pattern-recognition
Laurent Vacavant
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
22
Impact of residual misalignments
Very important studies:
Impact on b-tagging:
MC geometry includes misplacements
as expected in detector (incl. surveys)
10-100 m for modules, mm for elements,
some systematic deformations (global shifts,
clocking effects,…)
 next slide
1. all displacements known in reco
2. 1 + additional pixel random
misalignments introduced in reco only
3. actual ATLAS alignment procedure !
Both 1 and 2 require a specific procedure
to increase the hit errors  fully recovers
tracking efficiency and fake levels
Alignment procedure suggests residual
misalignments of ~ 3 m in rφ (pixels)
Laurent Vacavant
 after alignment, relative loss
in rejection between 11-18%
encouraging, even though
probably not all systematic
deformations were accounted for
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
23
Impact of misalignments: CSC challenge
Laurent Vacavant
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
24
Improving performance: Track Categories
Idea:
• use dedicated p.d.f. for various
classes of tracks.
• already used for Shared
10%(20%) improvement for b=60%(50%)
Laurent Vacavant
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
25
Optimizing the tracks-jet association
Distance track-jet vs pT:
All jets in ttbar events:
Tracks from B/D
Other tracks
• choose f(pT) such that 95%
of decay products of B/D are
included for all pT
Laurent Vacavant
• For non-isolated jets: x3 (x2)
the rejection power for
(b)= 50% (60%)
• Sample-dependent
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
26
Measuring b-tagging performance in data
Calibrating b-tagging efficiency with di-jet events
• Key-ingredient: soft muons
• Dedicated -jet trigger:
(pTjet<30 GeV)
– staged jet ET thresholds
– 1 Hz  100k in 30h (1 pb-1@1031 cm-2s-1)
• Method 1: pTrel templates
– b & c templates from MC
– fit in pT,  bins
• Method 2: non-linear system (à la D0)
–
–
2 samples
2 different b-jet/light jet fractions
Muon
jet
–
2 non-correlated taggers
 system can be solved analytically
• Both methods work well for pTjet<80 GeV
• Systematic uncertainties dominate for >
50 pb-1
 Current precision on b-tag efficiency: 6%
Laurent Vacavant
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008


Muon
jet
SVT
tag jet
28
System8 on di-jet events
Measurements
 system solvable analytically for , r and sample composition !
k: correlation between soft muon & track-based taggers
,: sample dependency
Laurent Vacavant
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
29
Calibrating b-tagging efficiency with top events
Tag counting method:
•
•
count 1,2,3 b-tags  likelihood
best accuracy on b: for 100 pb-1
•
integrated efficiency only
Several selection methods:
•topological
•likelihood
•kinematic fit
 signal purity 60-80%
±2.7%(stat)±3.4%(syst)
Selecting unbiased b-jets
W jets:
ET>40 GeV,
20
b-veto
(weight < 3)
(leptonic side)
Hadronic side
b-jet: ET>40
GeV
b-tag
(weight>3)
b-tagging efficiency determination:
Lepton:
ET>20 GeV
ETmiss>
20 GeV
Laurent Vacavant
Leptonic
side b-jet:
ET>20 GeV
No tag
requirement
Topo.
Like.
Kine.
Stat. (200 pb-1)
6.4%
4.4%
5.5%
Syst. error
3.4%
14.2%
6.2%
 can also be used to extract b’s p.d.fs
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
30
Tag counting in top events
• selection ttbar l+jets: 4j > 20 GeV
• counting events with 1,2,3 tags
• HF contents:
• 2 b-jets
• possibly 1 c-jet from W
• possibly cc/bb from gluon splitting
• first order:
• actual: (complex!) likelihood
with rough estimate of Ru
4% total error with 100 pb-1
slightly better for di-lepton
Laurent Vacavant
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
31
Conclusion
•
•
•
•
Expected b-tagging performance of ATLAS studied in details
For simple taggers, light jet rejection of 30 (for b=60%)
Secondary vertex taggers can achieve a rejection of 150
Most advanced taggers can reach a rejection of 50 for b=70%, very
important for multi b-jet signatures (more is needed: NN, BDT)
• Many critical aspects studied: misalignment, material, state of detector,
Monte Carlo modelling, etc – More to do: misalignment, pile-up
• Methods to measure the b-tagging efficiency in data have been
established, allowing a ~5% accuracy with 100 pb-1 in top events
• Work is on-going for the measurement of mistag rates
• New developments, e.g. b/c separation, can broaden the physics
reach of ATLAS
• Commissioning of the detector progressing well, our work will now
focus on the commissioning of the b-tagging to transform our
expectations into reality !
Laurent Vacavant
FZU/ASCR & FJFI/CVUT – Prague – Dec 19 2008
32