Marcel Vreeswijk (NIKHEF) B tagging, performance vertexing Neural Net studies tt event selection mass reconstruction in tt events conclusions B tagging.

Marcel Vreeswijk (NIKHEF)

• B tagging, performance vertexing• Neural Net studies• tt event selection• mass reconstruction in tt events• conclusions

B tagging in the tt all jets channel

By: Graziano Massaro Michiel Vogelvang (university student) Marcel Vreeswijk


• Performance secprobsecprob algorithm in tt events (p05 & p08)

• Reminder:

B tagging & vertexing

Start with (non PV) selected tracksSignificance>3

Make all possible 2-track vertices(vertex fits)

Keep/Kill vertices with shared tracksAdd tracks

(based on probability: opening angle, Pt)

Vertex fit based onimpact parameters


• Signal Events:

Performance vertexing

scaled to 100evts #Bjets efficiency Tagged (>1 sig. tracks) Tagged % #Non Bjets Purity%

SECPROB ttbar (p5) 137 59 14 82SECPROB ttbar+2mb (p5) 128 57 10 86SECPROB ttbar+2mb (p8) 126 58 13 82KALMAN ttbar+2mb (p8) 126 62 22 72

Background events#tagged #tagged #tagged Event-rate (ttbar/QCD)jets non-B jets PV jets S/B x 1000 S:B

SECPROB QCD+2mb (p8) 2.2 1.6 1.1 0.040 24839KALMAN QCD+2mb (p8) 7.8 6.5 5.7 0.014 72999

•KALMAN selects significantly more QCD jets (used without any additional cuts: what are they?)

•Efficiency SECPROB and KALMAN compatible.•No large effect from min. bias.

MC samples, thanks MC samples, thanks to Suyong!!!to Suyong!!!


Performance vertexingSECPROB

0

0.1

0.2

0.3

0.4

0.5

0.6

0 20 30 40

Et cuts

Eff

Bjet eff

S(ttbar)/B(QCD)x1000

KALMAN

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0 20 30 40Et cuts

Bjet eff


Performance SECPROB as func of Et

Performance KALMAN as func of Et

S(ttbar)/B(QCD)

Bjet eff.

KALMAN: higher QCD background


• Signal Events (cuts):


KALMAN

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0 0.2 0.5

Decay Length cuts

Bjet eff


SECPROB

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0 0.2 0.5

Decay Length cuts (cm)

Bjet eff


S(ttbar)/B(QCD)

Bjet eff.

S/B ratio not dependent on Decay Length


• Reminder: vertex constructed based on probability (Opening angle, Et)

• Now: try to find variables to discriminate between B vertices and QCD fakes, using a Probalistic Neural Network

vertexing and beyond

var description1 eneglo The full event energy2 ptglo The full event pt3 etglo The full event et4 njet The full event njets5 enejetx The jet energy measured from calorimeter6 etjetx The et measured from calorimeter7 massjx The mass measured from calorimeter8 sphebjet The jet boosted sphericity9 apbjet The jet boosted aplanarity10 enejet The jet energy11 etjet The jet et12 massj The jet mass13 sprijet The jet chi^2 jet tracks wrt primary vertex14 ntrajet The jet number of tracks15 E/ M Energy over mass= boost16 ddlife The decay length (x,y,z)17 dlife The decay length (x,y)18 sdlife chi^2 decay length (x,y)19 ctptvjet opening angle weigthed with pt20 evtxjet energy21 etvtxjet et22 massjv mass23 sprivjet chi^2 wrt PV24 sassvjet chi^2 wrt SV25 ntvtxjet number of tracks in vertex

Event

CAL

Jet-Tracks

Vertex in Jet

Preliminary!!!!!!!


• Strategy NN: take Et_vtx and Opening_Angle_vtx as base variables and see the effect of a third variable.

• the Et_jet and based on jet-track impact parameters appear promising


Jets-QCD

Bjets-ttbar

Ratio

Probability from NN

-jet-track impact parameters

Preliminary!!!!!!!

2bcontinued


Conclusions• The performance of the SECPROB

and KALMAN algorithm are studied using ttbar and QCD events.

• KALMAN has a slightly higher efficiency for B-vtxs, but finds significantly more QCD fake vtxs

• To find discriminating variables between good/fake vtxs a NN is used as tool.

• Many variables are tried: Et_jet and based on jet-track impact parameters appear promising


• For the ‘All jet’ channel

tt event (pre)selection

At least 5 jets with ||<2

Et of jetstt

qcd

Simple, effective, but: QCD has to be multiplied by 107


• From D0-RunI pubs: ET3= Et of jets, skipping 2 highest Et jets.


Cut appears less effective than in RunI. Why?

•In RunI: Initial jets in QCD events have large Et. The additional jets originate from QCD splittings and have low Et. Skipping 2 highest Et jets has large effect. For ttbar event effect is average: ET3(QCD) < ET3 (ttbar)•In RunII: QCD background has significant contribution from min. bias, which dillutes this effect.

Note: multiply QCD by 107


• Alternative:


tt

qcd

Et(5-jets)/Et(jets) vs <Et(jets)>

QCD: low Et per jet, many jetsttbar: high Et per jet contained in not so many jets.

Need many more QCD events!!!!


Mass reconstructionin tt--> all jets

• A very preliminary study

bt

W

W

b

jj

t

j

jDifficult final state: 4+2 jets

But, many constraints:

• W mass (2x)

•Both branches should yield similar top mass

Selection (no preselection):•At least 6 jets. Keep 6 highest Et jets•2 jets have vertex--> B candidates.Reconstruction:•2x2 W jets lead to 3 mass combinations•These mass combinations are then assigned to B candidates: 6 mass combinations.•Take combination with best 2 based on Mw (2x) and Mt1-Mt2



100 evts True full hadronicTop mass study 4+2jets 4+2jets 2 B ok 2B+2W ok All okttbar p5 14.3 10.7 7.9 5.7 3.9ttbar+2mb p5 25.1 15.7 14.4 7.0 5.3ttbar+2mb p8 22.9 15.6 11.0 5.9 3.6

Background: 5*5000000 QCD events <--> need more MC!!!!!!

tt

QCD

True mass

ALL



• Mass peak looks fine, but….

Good mass combs.

Bad mass combs.

The mass peak seems independend on bad/good combinations of the jets?!?!

Side remark: particle info in IN_PRT is corrupted as reported. In this study we attempted to take this into account properly.



•

W-mass (recoed)

tt

qcd

qcd

tt Note: multiply QCD by 107


Conclusions• The performance of the SECPROB and KALMAN

algorithm are studied using ttbar and QCD events. KALMAN has a slightly higher efficiency for B-vtxs, but finds significantly more QCD fake vtxs

• To find discriminating variables between good/fake vtxs a NN is used as tool. Many variables are tried: Et_jet and based on jet-track impact parameters appear promising

• The (pre)selection of ttbar events was studied. Cuts used in RunI apeared to have less effects due to min. bias overlay. New cuts are suggested.

• Can we measure the top mass in ttbar->All jet channel? A preliminary study, using all mass constraints yield a mass peak. However, this peak also show up for wrong jet-combinations(?).


• Background Events (cuts):


#tagged #tagged #tagged Event-rate (ttbar/QCD)No Cuts jets non-B jets PV jets S/B x 1000 S:B SECPROB QCD+2mb (p8) 2.2 1.6 1.1 0.040 24839KALMAN QCD+2mb (p8) 7.8 6.5 5.7 0.014 72999

Et jet>20 GeVSECPROB QCD+2mb (p8) 1.7 1.2 0.8 0.052 19221KALMAN QCD+2mb (p8) 5.0 4.1 3.7 0.021 48569

Etvtx>10 GeVSECPROB QCD+2mb (p8) 1.1 0.8 0.6 0.061 16319KALMAN QCD+2mb (p8) 3.2 2.8 2.7 0.017 57662

L>0.5cmSECPROB QCD+2mb (p8) 1.2 0.9 0.8 0.042 24096KALMAN QCD+2mb (p8) 2.5 2.2 2.0 0.020 48866

Et jet>40 GeVSECPROB QCD+2mb (p8) 0.1 0.1 0.0 0.544 1838KALMAN QCD+2mb (p8) 0.7 0.5 0.4 0.118 8441


• Signal Events (cuts):


No cuts 100evts #B jets efficiency Tagged(>1 sig. tracks) Tagged % #Non B jets Purity%

SECPROB ttbar+2mb 126 58 13 82KALMAN ttbar+2mb 126 62 22 72

Et jet>20 GeVSECPROB ttbar+2mb 122 59 13 82KALMAN ttbar+2mb 122 62 20 73

Et vtx>10 GeVSECPROB ttbar+2mb 127 45 9 84KALMAN ttbar+2mb 127 34 11 75

L>0.5cmSECPROB ttbar+2mb 127 31 7 82KALMAN ttbar+2mb 127 29 11 71

Et jet>40 GeVSECPROB ttbar+2mb 95 63 10 83KALMAN ttbar+2mb 95 63 15 76


• Strategy NN: take Et_vtx and Opening_Angle_vtx as base variables and see the effect of a third variable.

• the Et_jet and based on jet-track impact parameters appear promising


QCD jetsB jets ttbar

Probability from NN

Ratio

For Et-jet


• Validate ‘P9’ WH and QCD events versus ‘p8’ events (All samples from Suyong)

• First check distributions. Plots added of the tracking in jets related quantities:

Check p8 vs p9

jetintracks

errorPVparameterimpactSpri__

/)(_

SVintracks

errorSVparameterimpactSass__

/)(_

Sum of significances of tracks in jet wrt PV

Sum of significances of tracks in SV wrt SV

•See Plots, distributions look ok. Differences probably due to different cuts (Et in QCD generation), #min bias events and code changes


QCD p8


QCD p9


ttbar p8


WH p9


• Signal events

Efficiencies

No cuts 100evts #B jets efficiency Tagged(>1 sig. tracks) Tagged % #Non B jets Purity%

SECPROB ttbar+2mb 126 58 13 82SECPROB WH 121 52 1 98

• Backgroundfor 100 evts #tagged #tagged #tagged Event-rate (ttbar/QCD)No Cuts jets non-B jets PV jets S/B x 1000 S:B SECPROB QCD+2mb (p8) 2.2 1.6 1.1 0.040 24839SECPROB QCD p9 12.0 9.0 8.0

Number of ‘taggable’ Bjets + efficiency look fine

Number of tagged PV jets in QCD is significantly higher in P9. Probably explained by Et generator cut and/or #min. bias events.

Marcel Vreeswijk (NIKHEF) B tagging, performance vertexing Neural Net studies tt event selection mass reconstruction in tt events conclusions B tagging.

Documents