CDF Single-Top-Quark Searches

Preview:

Citation preview

C. Ciobanu, page 1

CDF Single-Top-Quark Searches

Catalin CiobanuUniversity of Illinois at Urbana-Champaign

University of Michigan, visiting

HEP Seminar, Michigan State UniversityJanuary 16, 2006

C. Ciobanu, page 2

CDF at the Tevatron

Tevatron and CDF performance:CDF recorded over 1.5fb-1Current analyses use up to ~1 fb-1

C. Ciobanu, page 3

Single Top Production

s-channel production (W*)

t-channel production (Wg fusion)

At the Tevatron, top quarks are:

Mostly produced in pairs (6pb):

qq annihilation (85%)gg fusion (15%)

Also electroweak (single-top):S-channelT-channelWt associated production

Motivation:Cross-section ~|Vtb|2Direct measurement of ~|Vtb|2Source of polarized top quarksBackground to HiggsNew phenomena:

W’ bosons, charged HiggsFCNCAnomalous W-t-b couplings

s-channel production (W*)

s1/2 =1.96TeV NLO Cross-sectionst-channel 1.98±0.25 pbs-channel 0.88±0.11 pb

B.W. Harris et al.: Phys. Rev. D 66, 054024Z. Sullivan hep-ph/0408049

C. Ciobanu, page 4

Run II Analyses

CDF: Phys.Rev.D 71, 012005, 2005.D0: Phys. Lett. B 622, 265, 2005. D0: hep-ex/0612052 (evidence for single-top production)

Current CDF:1 lepton with ET>20 GeV, |η|<2.0missing transverse energy: MET>25 GeV2 jets : ET> 15 GeV, |η| < 2.8at least one b-tag (displaced sec. vertex)Veto Z, dilepton, conversion events

Backgrounds: non-top and tt

C. Ciobanu, page 5

Why is it Hard?

Small cross section 3pb – not the main problemHuge backgrounds. In 955 pb-1:

W→ℓν + 2 jets S/B=1/200, S/sqrtB = 0.6W→ℓν + 2 jets +≥1btag: S/B=1/15, S/sqrtB = 1.5W→ℓν + 2 jets +≥1btag + discrim: S/B=1/3, S/sqrtB = 2.5

Backgrounds:W+heavy flavor (Wbb, Wcc, Wc)W+light flavor (mistags)Diboson, Z-decays, non-WTop pair production

C. Ciobanu, page 6

AnalysesThree most recent analyses (0.95 fb-1)

Multivariate: likelihood function, matrix element, neural networksLikelihood function: search for s- and t-channel signalMatrix Element: search for s+t channel signalNeural Networks: search for both s+t and individual s- and t-channel production

All 3 analyses use common selection criteria (same dataset and MC samples)

All analyses use a neural networks b-tag extensionNN b-tagger applies to secondary vertex tags; uses information such as:

Vertex mass, decay length, number of tracks, etc.

C. Ciobanu, page 7

Event Yield

Predicted event yield with 955 pbPredicted event yield with 955 pb--11

t-channel 22.36 ± 3.64

s-channel 15.44 ± 2.23

Single top 37.80 ± 5.87

Total background 549.3 ± 95.2

tt 58.35 ± 13.46Diboson 13.72 ± 1.85Z + jets 11.92 ± 4.42Wbb 170.9 ± 50.7Wcc 63.5 ± 19.9Wc 68.6 ± 19.0Non-W 26.2 ± 15.9Mistags 136.1 ± 19.7

Total prediction 587.1 ± 96.6

Observed 644

C. Ciobanu, page 8

Systematic UncertaintiesSystematic (-1σ/+1σ) s-channel t-channel All single top Shape variations

Jet energy scale +1.3%−1.4%

+1.8%−2.4%

+1.6%−2.0%+2.0%+0.3%+2.6%+1.9%+1.4%−0.4%±1.6%

±7.4%

Luminosity ±6% ±6% ±6%

Mistag model N/A N/A N/A

Non-W model N/A N/A N/A

Neural net b-tagger N/A N/A N/A

Q 2 scale in Alpgen MC N/A N/A N/A

±10.5%

Initial state radiation +1.1%−2.0%

+2.6%+2.0%

Final state radiation +1.3%+1.4%

+3.4%+2.2%

Total rate uncertainty ±9.1% ±11.3%

Monte Carlo generator ±1% ±2%

Parton distribution function +1.0%−0.6%

+1.7%−0.3%

±7.8%Event detection efficiency ±6.1%

C. Ciobanu, page 9

LF analysis

Single Top signature: W(e,µ+MissET) and 2 tight jets, ≥1b-tag

The Problem: Jet Energies not Well-Measured ET imprecisely measuredAmbiguities in:

choosing the Pz(ν) solutionchoosing b quark from top decay (s-channel)

Use a χ2 in which we float Pb, ET ν, Φνcentral values = measured valuesuncertainties derived from HEPG comparisons with reconstructed values

Without looking at the b-tag, minimize χ2 under four scenarios2 choices of which jet is labeled ‘b from top decay’2 neutrino solutions

C. Ciobanu, page 10

LF variablesForm a combined probability:

i: variable index, k: sample index (s or t) ji: histogram bin

Four background classes used: Wbb, tt, Wcc/Wc and mistags

t-channel LF Variables:• HT• hybrid Mlνb• cosθlepton,other-jet in top decay frame• Q*η• mjj• log(MEt-chan) from MADGRAPH• NN(b); Neural Net b-tag output• If 2 SECVTX tags, set LT=0.01

s-channel LF Variables:• Mlνb(hybrid,s-chan)• log(HT*Mlνb)• ET(jet 1)• log(MEt-chan) • HT • NN(b)

C. Ciobanu, page 11

t-chan LF inputs

init state q = u,dq’(if t) q’(if t)

init state q = u,dq’(if t) q’(if t)

z

z

C. Ciobanu, page 12

t-chan LF inputs (cont)

C. Ciobanu, page 13

t-chan LF inputs (cont)

C. Ciobanu, page 14

t-chan LF inputs (cont)

C. Ciobanu, page 15

Data. t-chan LF

C. Ciobanu, page 16

s-chan LF inputs

C. Ciobanu, page 17

s-chan Likelihood Function

ET (leading jet) log of s-channel Likelihood function

C. Ciobanu, page 18

(LFt vs LFs)

C. Ciobanu, page 19

Test Statistic

CLs method used in Higgs searches at LEPTest between two hypotheses:

H1: Data is described by signal and backgroundsH2: Data is described by backgrounds only.

Poisson probabilities:

Test statistics Q = - 2log[P(data|H1)/P(data|H2)]

∏∏==

⋅==

bins iHibins N

i i

dHi

nN

ii

i

dnPHdataP

1

1

!)()1|(

1

e

2

1

1ln2 H

i

Hi

N

ii n

ndconstQbins

∑=

⋅−=

C. Ciobanu, page 20

Statistical MethodExclusion (yellow) and Discovery (blue)

no signalSM signal

P-value = 50% = 0 σ

P-value = 5% = 1.6σ

P-value = 1% = 2.3σ

P-value = 0.1% = 3.1 σ

p-value: smaller is better

C. Ciobanu, page 21

(LFt vs LFs) result

Best fit point: σfit,tchan= 0.2+0.9

-0.2 pbσfit.schan= 0.1+0.7

-0.1 pb

S+B (test) and B only (null) hypothesesp-value: 51% (2.3% expected)95% CL limit = 2.7 pb (2.9pb expected)

C. Ciobanu, page 22

Matrix Element Analysis

C. Ciobanu, page 23

Transfer Functions

C. Ciobanu, page 24

ME Inputs

C. Ciobanu, page 25

Probability Discriminant

C. Ciobanu, page 26

Data Result

C. Ciobanu, page 27

Data Result

( | 1)( | 0)

P data HQP data H

=

Fit result: Fit result: 2.7 2.7 +1.5+1.5--1.31.3 pbpb

Expected Expected pp--value: value: 0.6% (2.50.6% (2.5σσ))

Observed Observed pp--value: value: 1.0% (2.31.0% (2.3σσ))

C. Ciobanu, page 28

Q x η in the high-region discrim

C. Ciobanu, page 29

Single top like candidate

C. Ciobanu, page 30

Neural Networks Analysis

Similar to the Likelihood Function:More input variablesCorrelation among input variables accounted for

C. Ciobanu, page 31

Comparison to Data

S- vs T-channelCombined s+t channel fit: 0.0 pb

Individual Search:

s-channel: 0.7+2.2-0.7 pb

t-channel: 0.2+1.7-0.2 pb

C. Ciobanu, page 32

Compatibility (ME-LF)

Is the ME result of 2.7+1.5-1.3 pb compatible with the LF result (0.3 pb)?Answer this question by using simulated experiments (based on MC).

Throw events according to 0.95 fb-1

predictions:For each pseudo-experiment, form the ME and LFtchan distrib.Fit the two (separately) as weighted sums of signal/background reference histograms

Correlation is 59%Which leads to 4-5% probability of getting ME and LF results which are at least as far apart as they are in the data

C. Ciobanu, page 33

Compatibility (2)

Why do results differ?

While the analyses are correlated, there are several conceptual differences :ME analysis uses Transfer Functions, LF and NN analyses do not. ME does not use measured Missing Et, LF and NN do. ME integrates over all neutrino Pz values, the NN chooses the solution with the smaller |Pz| (remember the W=l+nu constraint gives two Pz(nu) solutions).There are two jets, one of which corresponds to the b-quark from top decay. NN analysis chooses the jet with a displaced vertex (b-tag) as the b from top, the ME analysis sums over both possibilities.NN uses events with soft jets (15GeV>Et(jet)>8 GeV), ME does not.

At the and of the day, ME and NN have very similar sensitivity 2.5-2.6 sigma expected.

C. Ciobanu, page 34

Why are the results different?

The overlap between the 5% highest-ME events and 5% highest-NN(s+t) events is 30% or 43% for the s-chan, or t-chan MC events, respectively. See the left plot below (x-axis = 5%). Red = t-channel MC. Blue = s-channel MC.

Middle plot shows the effect of the transfer functions – the NN needs better measured-jets (lower Delta E = E(measured jet)-E(hepg b-quark) ) in the high signal region (close to 0). For ME the well-measurement requirement is not so stringent. The effect is only significant for t-channel (upper black -NN- and blue -ME- graphs).

Right plot shows the effect of Missing ET (MET) measurement. The mean of the Delta E = MET (measured) – Et(hepg neutrino) distribution is 0 for all cases but the RMS gives an indication of how well we measure it (the smaller the better). Again, NN needs a better (lower RMS) determination of the neutrino energy than the ME in the high signal region (x-axis close to 0). That is, ME graphs are higher (less well-measured) than the NN graphs.

MadEvent signal MC studies (s-chan and t-chan events)

Signal-likeness Signal-likenessW

ell m

easu

red

MET

Signal-likeness

Wel

l-mea

sure

d je

ts

C. Ciobanu, page 35

NN(s+t) vs ME discriminants

Data compared with null hypothesis (B) –left plotData compared with test hypothesis (S+B) –right plot

ME

NN

Bin 1: NN < 0.8 && EPD < 0.9Bin 2: NN > 0.8 && EPD < 0.9Bin 3: NN < 0.8 && EPD > 0.9Bin 4: NN > 0.8 && EPD > 0.9

Chisquare null hyp: 5.22 / 4 bins prob = 26.6% Chisquare test hyp: 5.38 / 4 bins prob = 25.0%

C. Ciobanu, page 36

NN(s+t) vs LF(t-chan) discriminants

Data compared with null hypothesis (B) –left plotData compared with test hypothesis (S+B) –right plot

NN

TLF

Bin 1: TLF < 0.9 && NN < 0.8 Bin 2: TLF > 0.9 && NN < 0.8Bin 3: TLF < 0.9 && NN > 0.8Bin 4: TLF > 0.9 && NN > 0.8

Chisquare null hyp: 2.53 / 4 bins prob = 63.9%Chisquare test hyp: 5.35 / 4 bins prob = 25.3%

C. Ciobanu, page 37

Combination

The same datasetCombine analyses, not resultsOut of the 6 discriminants:

LFt, LFsMENNs, NNt, NNst

Form linear combinations (superdiscriminants):SDst = a*LFs + b*ME + c*NNst (combined)SDs = a*LFs + b*ME + c*NNs (s-channel)SDt = a*LFt + b*ME + c*NNt (t-channel)

Expected sensitivity: ~3σ

C. Ciobanu, page 38

Summary

CDF single-top results with 1/fb Expected sensitivities:

2.0, 2.4, 2.6σ (combined 3σ)Observe fluctuation low:

LF: 0 σ , ME: 2.3 σ, NN : 0 σAbout O(1%) probability to get this result (number in preparation)

Look forward for more dataA taste of LHC physics:

Challenging backgroundsPushing the limits on MC modeling (multiple analyses on the samedataset).

C. Ciobanu, page 39

Summary 2For completeness, D0 latest

Expected sensitivities:1.3, 1.8, 2.1σ

Observe fluctuation high:2.4, 2.9, 3.4 σ

CDF Run II

0.95 fb-1

CDFExpected sensitivities:

2.0, 2.5, 2.6σObserve fluctuation low:

0, 2.3, 0 σ

Recommended