closing in on the higgs boson at cdf zurich, 28 may 2008 aidan robson glasgow university
TRANSCRIPT
Closing in on the Higgs Boson at CDF
Zurich, 28 May 2008
Aidan RobsonGlasgow University
CDF
2/54
Aidan Robson Glasgow University
Limit-setting Analysis techniques Analysis stability Improving sensitivity
Motivation
Higgs WW
Low mass channels
CDF and Tevatron Combination
Outlook
3/54
Aidan Robson Glasgow University
Higgs?
Simplest way of breaking electroweak symmetry
A complex doublet of scalar Higgs fields
that couple to the SU(2)U(1) vector bosons maintaining gauge invariance
and have associated potential V() = ()2–2()with minimum away from =0
Gauge bosons acquire massand can write lepton mass terms
We are left with one dynamical field single neutral scalar particle
4/54
Aidan Robson Glasgow University
Higgs?WW scattering
W+
W–
Z/ W+
W–
W+
W–
W+
W–
Z/W+
W–
W+
W–
W+
W–
W+
W–
W+
W–
W+
W–
HH
required to cancel high-energy behaviour
5/54
Aidan Robson Glasgow University
Direct searches / Indirect limits
e+
e–
Z
H
Z b
bmH>114GeV mH<160GeV
6/54
Aidan Robson Glasgow University
Tevatron
proton–antiproton
√s = 1.96TeV
7/54
Aidan Robson Glasgow University
CDF
proton–antiproton
√s = 1.96TeV
= 1.0= 1.0
= 0.6= 0.6
= 2.0= 2.0
8/54
Aidan Robson Glasgow University
Decay
Br.
Br
Production
mH/GeV
/ fb
mH/GeV
ggH
qqWH
qqqqH
qqZHbbH
gg,qqttH
Br
q
q’
W,Z
H ’
9/54
Aidan Robson Glasgow University
A physics programme
In last year
(ppZZ) 3 x(ppH)(mH=160)
10/54
Aidan Robson Glasgow University
Higgs WW
H0W+
W–
11/54
Aidan Robson Glasgow University
Hadronic Ws?
WW ljj
Signal + background
Background only!
W
W
W
W
l
q
q’
l
q
q’
Leptonic final statesfor our Higgs search
12/54
Aidan Robson Glasgow University
HWW
H0W+
l+
W–l–
ee, e, ; ET
Isolation
mll > 16
Z and topsuppression
Dilepton sample composition
Dre
ll-Y
ando
min
ated
WW
dom
inat
ed
Hig
gs e
nhan
ced
B
Z Z
tt ttWW WW WW WW
H H HH
signalseparation
H0W+
l+
W–l–
W+
W–
q
q’
q
q’
90%
10%
13/54
Aidan Robson Glasgow University
Limit settingIsolation
mll > 16 or 25Z and topsuppression
signalseparation
Background
Higgs signal x 10
eve
nts
X
X = some observable
H1=SM+Higgs (of mass mH)H0=SM only
Construct test statistic Q = P(data|H1)/P(data|H0) –2lnQ = 2(data|H1) – 2(data|H0) , marginalized over nuisance params except H
Find 95th percentile of resulting H distribution – this is 95% CL upper limit.
When computed with collider data this is the “observed limit”
Repeat for pseudoexperiments drawn from expected distributions to build up expected outcomes
Median of expected outcomes is “expected limit”E
xpec
ted
ou
tcom
es
95% CL Limit/SM
Median = expected limit
H (pb)
95%
H/SM
95%
0 20 1 2
rescalePD
F
14/54
Aidan Robson Glasgow University
Limit setting (2)
95%
CL
Lim
it /
SM
mH / GeVRepeat for different values of mH build up exclusion plot
median 12
illustrative
mH=160
15/54
Aidan Robson Glasgow University
What have we found?95
% C
L L
imit
/ S
M
mH / GeV mH / GeV
expected limitobserved limit
illustrative
expected limitobserved limit
illustrative
Deficit Excess
16/54
Aidan Robson Glasgow University
Analysis overview
Spin structure WW vs H->WW lepton
NNscore
0 1
var1
var2var n
Cut-based analysis
Neural net approach Matrix element likelihood approach
d ∫ ƒa(x,Q2) ƒb(x,Q2) |Mab(4;)|2 …
1. sequential combination2. understanding relationship at a deeper level
HW–
W+l –
l+
ex
tend
sen
stiv
ityextend senstivity
Looking for some final distribution to which to fit templates and from which to extract limits
17/54
Aidan Robson Glasgow University
Matrix element method Use LO matrix element (MCFM) to compute event probability
HWWllWWllZZllW+partonl+jetWl+
ET modellepton energy resn
px
py
pz
lep1
LO |M|2 :px
py
pz
lep2
Ex , Eyparton lepton fake rate conversion rate
xobs:
(with true values y)
Compute likelihood ratio discriminator
R =Ps
Ps + kbiPb
i
i
kb is relative fraction of expected background contrib.Ps computed for each mH
Fit templates (separately for high S/B and low S/B dilepton types)
18/54
Aidan Robson Glasgow University
Neural network method
NNscore
0 1
var1
var2var n
Background Higgs
Various versions. Current: Apply preselection (eg ET to remove Drell-Yan) Train on {all backgrounds / WW} against Higgs mH=110,120…160…200 { possibly separate ee,e,
x10
Pass signal/all backgrounds through net Form templates
NN
0 1
Pass templates and data to fitter
ET
ET
mll
Elep1
Elep2
ETsigData
HWWWWDYWgWZZZt t
fakes
ETjet1
Rleptons
leptons
ET lep or jet
ETjet2
Njets
Most recent CDF“combined ME/NN” analysis also uses ME LRs as NN input variables
19/54
Aidan Robson Glasgow University
What have we found???
Last summer, CDF had 3 analyses: Matrix Element, Neurobayes Neural Net, TMVA Neural Net expected sensitivities all similar input distributions: well modeled observed limits...
20/54
Aidan Robson Glasgow University
What have we found???
Last summer, CDF had 3 analyses: Matrix Element, Neurobayes Neural Net, TMVA Neural Net expected sensitivities all similar input distributions: well modeled observed limits...
mH / GeV
expected limitobserved limit
illustrative
Excess in one of the 3 analyses!
21/54
Aidan Robson Glasgow University
Grounding in SM measurements
TCE
TC
ETC
E P
HX
PH
X P
HX
TCE
CM
UP
TCE
CM
X
TCE
CM
IOC
ES
TCE
CM
IOP
ES
PH
X C
MU
PP
HX
CM
X
PH
X C
MIO
PE
S
PH
X C
MIO
CE
S
Neural net: Matrix element:
Different subdetector combinations
22/54
Aidan Robson Glasgow University
ME Input Variables
leading lepton px subleading lepton px
Ex Ey
ET sin( ET ,nearest lep or jet)
23/54
Aidan Robson Glasgow University
Stability
convergence
Assessing NN stability
rogue variables – had checked data-simulation agreement in as many regions as possible
Drell Yan-rich WW-rich
MetMet– applicability?
training epoch training epoch
successful training unsuccessful training
tra
inin
g e
sti
ma
tor
tra
inin
g e
sti
ma
tor
control regions…
24/54
Aidan Robson Glasgow University
Complementarity
Redefine discriminant for WW hypothesis:
R’ =PWW
PWW + kbiPb
i
i
exploit different sensitivities of matrix element / neural net
verify matrix element method: cycle signal
– ME is leading order - remove variables that use jet information from neural net for comparison
25/54
Aidan Robson Glasgow University
NN same-sign
After DYnet cut
WW NN score WW NN scoreWW NN score
W
26/54
Aidan Robson Glasgow University
NN correlations
Data
WWHiggs
ET :mll
ETsubleading lepton
:mll
ET :mll
ET :ET
27/54
Aidan Robson Glasgow University
Increased acceptance by adding plug calorimeter (no tracking) and tracks pointing to cracks
Lepton coverage
electrons muons
28/54
Aidan Robson Glasgow University
Luminosity
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Inst luminosity (1030)
(n
b)
recent hardware upgrades clever prescale strategies – complicate analysis!
a trigger consisting of hits in the “CMX” muon system, matched to a track
Higher statistics…… but affects trigger rates
Year2002 2003 2004 2005 2006 2007 2008
4fb–1
3fb–1
2fb–1
To
tal
lum
ino
sity
(p
b–1
)
4fb–1 delivered
CDF: 3.3fb–1 to tape
Store number
29/54
Aidan Robson Glasgow University
Systematics
JESLepton Energy ScaleISRPDFFakes
Shape uncertainties(neural net):
40 weights stored per event(Higgs and all backgrounds!)
(neural net)
– and similar for matrix element analysis
30/54
Aidan Robson Glasgow University
CDF H WW
expected signalevents
Current result is NN trained on kinematic + ME likelihood variables
31/54
Aidan Robson Glasgow University
CDF limit
1.6
2.5
32/54
Aidan Robson Glasgow University
CDF limit development
Oct 05
Jan 07Mar 07
Aug 07
Expected limits
Feb 08
2.8
33/54
Aidan Robson Glasgow University
Achieving a deeper understanding
Matrix element likelihood / Neural net
Much work into demonstrating relationship between ME and NN approaches
First attempt: removed variables that use jet information from neural net for comparison(ME is leading order plus an approximation for pT) – saw ~equivalence
More advanced attempt:swapping between kinematic 4-vector inputs, and ME likelihood inputs– excellent equivalence
using all Rleptons
ET variableLRHWW
using kinematic only Rleptons
ET variableHT
E2lep
E2lep
NjetET
lep
ETjet
ET
Don’t know exactly what in ME is distinguishing backgrounds
Similarly we “guess” at the “right” variables for the NN
Closed form calculations could help to bridge the gap
34/54
Aidan Robson Glasgow University
Ongoing improvements
W+
W–
VBF
W+
W–
VH VWWl
l / Other channels:
2-jet bin Hadronic taus Inclusion of more triggers Improvement of S/B through lepton selection
ee mH=130 mH=160
W
+ Increase WW analysissensitivity at lower masses
factor 1.5in sensitivity?
Sensitivity to low mH from low mll
35/54
Aidan Robson Glasgow University
Low mass channels
q
q’
W,Z
H
36/54
Aidan Robson Glasgow University
b-tagging
Secondary vertex-finding algorithm
Attempt to fit tracks to decay vertex
Jet probability
Compares track impact parameters to measured resolution functions
Neural network filters
ntracks in secondary vertexpT fraction carried by those tracksgoodness of vertex fitvertex masstransverse decay length & significance…
37/54
Aidan Robson Glasgow University
ZHllbb
Double tag
One tight or two loose b-tagsZ from ee or
ANN2 is the equivalent of 2.5 x more data compared to a dijet mass peak!
Missing transverse energy Projection Dijet Fitter corrects jet energies for the missing energy in the event – improves ZH dijet resolution from 16% to 10%.
Single tagq
q’
Z
H
Z
l
l
b
b
38/54
Aidan Robson Glasgow University
WHlbb
25% improvement
10% gain
q
q’
W
H
W
l
b
b
double tight tag
one tight tagwith NN tag
17% gain
Three categories: double tight tag one tight tag + one jet probability tag one tight tag with NN tagANN to improve signal discriminationAdd isolated tracks
mHNN output (120)
39/54
Aidan Robson Glasgow University
WHlbbq
q’
W
H
W
l
b
b
ME technique from single top search
40/54
Aidan Robson Glasgow University
VHETbb
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
ET > 70GeVOne tight or two loose tagsveto leptons
q
q’
W/Z
H
W/Z
(l)
b
b
mjj (GeV/c2) mjj (GeV/c2)
1 tight tag 2 loose tags
41/54
Aidan Robson Glasgow University
VHETbb
q
q’
W/Z
H
W/Z
(l)
b
b
Track-based discriminant for removing QCD background
double tight tagone tight tag + one jet probability tag
42/54
Aidan Robson Glasgow University
VHqqbb4 jets2 tags
ET trigger
SMH=120
/B
5 / 20000
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
data-driven bg
Tag rate fractionmeasured as fnET, ntracks, mbb,Rjj
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
mbb
mqq
43/54
Aidan Robson Glasgow University
lephad: Br 45%
leptonhadronic tau (1&3prong)2 jets
opposite charge
3 ANN: sig vs: Z, tt, QCDcombined for fit
Alpgen Z+jetsMT(lep,ET) / GeV/c2
0 100
44/54
Aidan Robson Glasgow University
CDF and Tevatron Combinations
45/54
Aidan Robson Glasgow University
1.62.6
46/54
Aidan Robson Glasgow University
Current limit
47/54
Aidan Robson Glasgow University
Current limit
48/54
Aidan Robson Glasgow University
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
1 Oct 07
1 Oct 09
3.3 fb-1
4.5 fb-1
LuminosityYear2002 2003 2004 2005 2006 2007 2008
4fb–1
3fb–1
2fb–1
To
tal
lum
ino
sity
(p
b–1
)
4fb–1 delivered
CDF: 3.3fb–1 to tape
Store number
49/54
Aidan Robson Glasgow University
Achievable Sensitivity
Sensitivity factors Minimum = x1.5Further = x2.25
CDF+D0 combined- curves are sqrt(L)
95%
CL
50/54
Aidan Robson Glasgow University
Improving sensitivity: low mH
Low Mass Higgs (~ 115 GeV)
Minimum Achievable Improvements– 25% b tagging (improved usage of existing
taggers)• Into ZH=>llbb and VH=>MET+bb
– Implemented in WH=>lnubb Summer’07
– 25% trigger acceptance (pre-existing triggers)• Into ZH/WH => llbb, lnubb
– Completed S vs B studies
– 20% from advanced analysis techniques studies & better usage of MET
• Into MET+bb and lnubb– Implemented in ZH=>llbb Summer’07
x1.5 avg. sensitivity improvement for all analysis
All improvements validated on analysis/studies with real data/tools
%’s are in sensitivity
51/54
Aidan Robson Glasgow University
Low mass Higgs( ~115 GeV)
Further Achievable Improvements
– 25% b tagging (NN-based) • All channels
– Tagger in advanced stage– Efficiency studied
– 25% trigger acceptance• More pre-existing triggers
– Based on HTTF studies
– 10% Tau channels (hadronic)
Id well understood from H analysis
%’s are in sensitivity
All improvements validated on analysis/studies with real data/tools
Additional x1.5 avg. sensitivity improvement for all analysis
Mis
tag
rat
e Tagging Efficiency
Standard secondary vertex
b-tagging
52/54
Aidan Robson Glasgow University
Achievable Sensitivity
Sensitivity factors Minimum = x1.5Further = x2.25
CDF+D0 combined- curves are sqrt(L)
95%
CL
115 GeV
53/54
Aidan Robson Glasgow University
Exclusion region grows
With 7 fb-1 • exclude all masses (except real mass)• 3 150:170
7.0
With 5.5 fb-1 tougher:• Exclude 140:180 range• 3 in one point: 160
5.5
54/54
Aidan Robson Glasgow University
Scenario from Feb 2006
LHC 2007: Pilot Run, Z,W calib? 200pb-1
LHC 2008: Physics, 1fb-1
Tev 2007: 4fb-1 : HWW 4x3: at SM limit in the 140-170 range. TOP and W Mass improved as well, so SM fit limits narrower.
• Deviations building from expected limit: we focus on this range for ATLAS 2009. Perhaps SM fit narrowing on this range.
• Higgs is 130-150 OR 170-185. Perhaps SM Fit excludes upper range.
Tev 2009: 3at 116: ATLAS 2011? for discovery. CDF keeps running!?LHC 2010: 10fb-1 : Discover it for > 130.
2008
2009
2008
March 2007May 2008
2010
55/54
Aidan Robson Glasgow University
Jets
W/Z
HiggsSusy
bottom-quark
top-quark
56/54
Aidan Robson Glasgow University
Backup...
57/54
Aidan Robson Glasgow University
58/54
Aidan Robson Glasgow University
59/54
Aidan Robson Glasgow University
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
60/54
Aidan Robson Glasgow University
Do we have to get lucky?
“further” @ 115 GeV
7 fb-1 => 70% experiments w/230% experiments w/3
“further” @ 160 GeV
7 fb-1 => 95% experiments w/275% experiments w/ 3
Solid lines = 2.25 improvementDash lines = 1.50 improvement
Analyzed Lum. Analyzed Lum.
61/54
Aidan Robson Glasgow University
CDF H WW
expected signalevents
62/54
Aidan Robson Glasgow University
Systematics: CDF HWW
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
%
63/54
Aidan Robson Glasgow University
Systematics: D0 HWW
64/54
Aidan Robson Glasgow University
Systematics: WHWWW
65/54
Aidan Robson Glasgow University
Neural network method TMVA
DYNN
score
0 1
var1
var2var n
WWNN
score
0 1
var1
var2var nDY Higgs WW Higgs
Use TMVA neural nets twice Train on Drell-Yan and Higgs mH=160 ee,e,
Train on WW and Higgs mH=110,120…160…200; ee,e,
DYNN
x3 x30
0 1
Pass signal/data/background through DY–H net Cut Pass remaining events through WW–H net
WWNN
0 1 Fit templates
ET
ET
mll
Elep1
Elep2
ETsig
DataHWW
WWDYWgWZZZt t
fakes
ETjet1
Rleptons
leptons
ET lep or jet
ETjet2
Njets
66/54
Aidan Robson Glasgow University
Achievable sensitivity (CDF only)
67/54
Aidan Robson Glasgow University
NN: low-mass, low mll
Sensitivity to low mH from low mll
68/54
Aidan Robson Glasgow University
Sensitivity at low mass
ee
e
mH=130 160 200
WW NN score
Low mass: W understanding becomes key
W
69/54
Aidan Robson Glasgow University
Improving sensitivity: high mH
High mass Higgs (~160 GeV)
CDF range of achievable improvements– 10-20% from hadronic taus in W decay (including better id)
• Ongoing studies
– 25-40% VHVWW and VBF (jj in final state)• Expect good S/B• Ongoing studies
– 10-15% more triggers (existing triggers)+ more leptons
%’s are in sensitivity
Improvements from x1.5 to x2 in sensitivityAll improvements validated on analysis/studies with real data/tools
ee
mH=130 mH=160
W+ Increase WW analysissensitivity at lower masses
70/54
Aidan Robson Glasgow University
Selection
B
Z Z
t t
WW WW WW WWH H H H
Matrix Element
mll>25 GeV
Metspecial
N jets<=1
Matrix element discriminator
Neural Net
mll>16 GeV
N jets(ET>15)=0 || N jets(ET<55)=1 || N jets(ET<40)=2
DY–Higgs neural net
WW–Higgs neural net
Isolation20GeV/10GeV
>25(ee,)>15(e)
Cosmic rejectionOpposite sign
71/54
Aidan Robson Glasgow University
Selection
Matrix element Neural net
Using extended lepton coveragefrom WZ observation
Using standard leptons
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
Zee
(Met,nearest l or j)
Me
t
(Met,nearest l or j)
Me
t
(Met,nearest l or j)
HWW 160
Me
t
WW
Metspecial = Met x sin(Met) Met (if Met > 900)
Leptons:
“Metspecial”:(Matrix element)
Avoid Met pointing along lepton or jet direction
72/54
Aidan Robson Glasgow University
Tevatron
proton–antiproton
√s = 1.96TeV
QuickTime™ and aTIFF (LZW) decompressor
are needed to see this picture.
73/54
Aidan Robson Glasgow University
CDF
= 1.0= 1.0
= 0.6= 0.6
= 2.0= 2.0muonchambers=
2= 3
0 1 2 3 m
2
1
0
tracker had cal
hadronic calEM cal
had cal
solenoid
pre-radiator shower max
silicon
EM cal
= 1
Drift chamber to ||<1Further tracking from SiCalorimeter to ||<3Muon system to ||<1.5
74/54
Aidan Robson Glasgow University
CDF muon patchwork
75/54
Aidan Robson Glasgow University
Production Decay
mH/GeV
Br
Br( W l) ≈ 0.32 Br( W jj ) ≈ 0.68
H0W+
W–
/ fb
mH/GeV
ggH
qqWH
qqqqH
qqZHbbH
gg,qqttH
76/54
Aidan Robson Glasgow University
Precision EWK fits
1-sided 95%CL upper limit 144 GeV
increases to 182 GeV including LEP direct exclusion to 114GeV
LEPEWWG, July 2007
77/54
Aidan Robson Glasgow University
ME Input Variables
leading lepton px subleading lepton px
Ex Ey
ET sin( ET ,nearest lep or jet)
78/54
Aidan Robson Glasgow University
Higgs yields
Matrix element analysis1/fb
Neural net analysis1/fb
Matrix element analysis2/fb
79/54
Aidan Robson Glasgow University
ME same sign
80/54
Aidan Robson Glasgow University
ME Likelihood Ratio
81/54
Aidan Robson Glasgow University
Systematics
JESLepton Energy ScaleISRPDFFakes
Shape uncertainties(neural net):
40 weights stored per event(Higgs and all backgrounds!)
(neural net)
– and similar for matrix element analysis
82/54
Aidan Robson Glasgow University
Systematics
Common uncertainties treated as nuisance parameters:
LuminosityTrack isolationHiggs: s, NNLOWW: Jet veto, PDF/Q2, generatorDY: Met modeling, low mass modeling
Shape uncertainties:
83/54
Aidan Robson Glasgow University
NN: low-mass, low mll
Sensitivity to low mH from low mll
84/54
Aidan Robson Glasgow University
D0 HWW (e/)
e
ee, e channels 1.1fb–1 ; channel 1.7fb–1
W+
W–
VBF
includes VBF
W+
W–ggH
85/54
Aidan Robson Glasgow University
D0: Other channels HWW had channel 1fb–1
– select using neural net – event likelihoods to separate signal (not currently contributing to overall limit)
VHVWW 1.1fb–1 search for ll’+X (like-sign dileptons) 2d likelihood: physics/instrumental backgrounds
W+
W–
VH VWWl
l /
86/54
Aidan Robson Glasgow University
mH (GeV) 120 140 160 180 200
Median /SM 22.2 6.7 2.8 4.4 9.7Observed /SM 47.3 12.0 2.4 4.7 11.1
D0 RunIIa+b Preliminary