76X - no pt reweighting

Sample	A (0.2<R<0.4, <2 tag)	B (R>0.4, <2 tag)	C (0.2<R<0.4, 2 tag)	D = B*C/A pred.	D (R>0.4, 2 tag) obs.	Ratio pred./obs.	Pull (pred-obs)/error	KS test
T5tttt	1682.95 +- 38.09	2345.72 +- 45.22	1587.65 +- 36.79	2212.89 +- 83.41	2098.93 +- 42.08	1.05 +- 0.05	1.22	0.10
Sum Bkg.	102.06 +- 6.07	4.93 +- 0.24	10.77 +- 1.58	0.76 +- 0.00	0.52 +- 0.06	1.46 +- 0.16	4.15	-
All Bkg.	102.06 +- 6.07	4.93 +- 0.24	10.77 +- 1.58	0.52 +- 0.09	0.52 +- 0.06	1.00 +- 0.20	0.00	1.00
TTJetsMGHT	12.71 +- 0.26	1.14 +- 0.03	3.20 +- 0.06	0.29 +- 0.01	0.21 +- 0.01	1.39 +- 0.11	4.54	0.00
WJets	6.34 +- 0.28	1.82 +- 0.15	0.72 +- 0.09	0.21 +- 0.03	0.07 +- 0.03	2.98 +- 1.31	3.13	0.63
ZJets	1.66 +- 0.14	1.25 +- 0.12	0.13 +- 0.04	0.10 +- 0.03	0.13 +- 0.04	0.76 +- 0.33	-0.62	0.72
TTX	0.92 +- 0.10	0.25 +- 0.04	0.38 +- 0.06	0.10 +- 0.03	0.09 +- 0.03	1.13 +- 0.43	0.33	0.81
Top	1.67 +- 0.35	0.16 +- 0.10	0.44 +- 0.18	0.04 +- 0.03	0.00 +- 0.00	inf +- -nan	1.32	0.49
QCD	78.48 +- 6.05	0.17 +- 0.09	5.88 +- 1.56	0.01 +- 0.01	0.00 +- 0.00	inf +- -nan	1.59	1.00
Diboson	0.28 +- 0.11	0.14 +- 0.04	0.03 +- 0.06	0.01 +- 0.03	0.03 +- 0.02	0.53 +- 1.29	-0.34	0.99

1) Increase to 4 fb^-1 AK8 - R

2) Correctly weight NLO Events -> Has to do, OK!

3) Only use two Highest pt jets for Delta-phi and pre-tags -> Much better! Reduced R-DPhi correlation

AK8 - R

4) Compare AK8-R with AK4 and TT-R TT-R: Less correlation, but more background AK4-R: Somewhere between AK8 and TT -> Should probably stay with AK8-R

TT-R

5) Test TT-bar behavior AK8

TT

1) Use AK4 jets for R, a'la Razor recipe

Used Old JEC MC samples

This doesn't seem to work. The smaller cone size jets collect better all the hadronic activity, but the signal to background discrimination becomes worse. I think the reason might be that the larger cone size jet algorithm finds better the merged top jets, while AK4 fragments them. I tried tuning the cuts fo AK4 jet R, but for the same amount of background I got less singal events.

(I will also do a tt-pair R vs AK8 R comparison which might be useful)

2) Try shrinking R control region from [0,0.4] to [~0.3,0.4]

R sideband: [0,0.4]

R sideband: [0.3,0.4]

R sideband: [0.2,0.4] - DPhi<2.7

This works! The TTbar prediction is closer to the actual values. I used [0.2, 0.4] bin so the statistical precision doesn't decrease, we need to have a little bit of compromise here to keep statistical error low but also having a better TTbar estimate. Signal contamination is also less in this bin.

3) How would we treat signal pollution ? Show yields for all 8 boxes defined by 3-axis: R, Ntop, DPhi

R sideband: [0.2,0.4] - DPhi>2.7

In |Dphi|<2.7, signal pollution is: A: ~1% B: 120% C: ~6%

In |DPhi|>2.7, the signal pollution is one order of magnitude smaller! A': 0.1% B': 9% C~: 0.6%

A'/C' ratio is very similar to the A/C ratio, we can substitute that instead. The B/B' background ratio could be taken from MC and use it to scale the number of events in B' to have an order of magnitude less polluted estimate for B and then D).

D is then: D = (B/B')_MC * B' * C' / A' = (9.45/12.24) * 13.36 * 1273.35 / 5801.51 = 2.26

c.f: 2.10 with no signal contamination. The error then is around the signal pollution level of B' which can also be substracted maybe, but it's definitaley lower than ABCD only prediction. i.e: Polluted A/B/C would cause: D_poll = B_poll * C_poll / A_poll = 4.67 --> This would be much larger and therefore it is a legitimate question

This method I think looks good! Is this the logic you described, Petar?

4) Try other TTbar samples as a comparison

Warning: NLO event weights are not correct (negative gen-weights aren't subtracted), but the obs/predicted ratio is still useful - Overall number of ttbar events are around ~1+-0.4, prediction to observed ratio: ~1.4+-0.3 - In all cases, the overall event number prediction stays very close to the observed number of events: ratio: 0.97 +-0.09, pulls are low too

5) Why TT+W/Z not similar to ttbar? Plot DeltaPhi shape for TTV

Plots attached. For some reason the Dphi distribution looks very similar regardless the Ntop/Dphi bin. These are NLO samples, so I will need to correct for the negative weights.

6) When backgrounds are combined scaling might not work: --> Try mixing together different ratios of bacgrounds to check how robust the method is

If I scale ttbar by 2 times the cross section, the pull gets larger, the ratio is sligthly lower, but still within statistical error. - Should we apply a ~3% correction to account for this? - If I blow up ttbar to x5 xsec, the ratio is still 0.85 --> Should we simply add a systematic error?

The method seems to be very robust.

7) Try to find a way to describe why the method works from first principles Plot correlation of Razor variables for all regions of the ABCD methods in TTbar (and other) samples, eg: a) DeltaPhi vs MR/MTR for Ntop==2 and Ntop<2

The MTR vs DPhi plots show no correlation, that's great. MR vs DPhi shows some correlation, but I have an idea: MR for the top-pair is uncorrelated to DPhi, this is potentially useful. But I have to update my top-pair definition to the latest recipe and rerun the code.

I am planning to do a full AK4/AK8/TT - R comparison to settle this issue (I already did for Phys14, but I only checked signal efficiency, which was very similar).

Hopefully TT-R is uncorrelated to DPhi and then the explanation why the method works might be finally there.

b) MR/MTR vs Ntop for DeltaPhi <2.7 and DeltaPhi >2.7

I missed these plots, sorry. They are on my to-do list.

Other 4 regions in 3D box:

R sideband 0,2,0.4 - slightly more statistics is better in A and C for overall estimate. For TTbar is is a bit less optimal, needs to correct for it

R sideband 0,25,0.4

Add filter + trigger (QCD back to normal)

R sideband 0,25,0.4

Add some samples, New JEC - no reweighting, no trigger, no filter (QCD somehow higher than previously!)

dphi < 2.7 pt400 old tau32 old - 2.5fb^-1

dphi < 2.7 pt350 old tau32 old - 2.5fb^-1

dphi < 2.7 pt400 old tau32 new - 2.5fb^-1

dphi < 2.8 pt400 old tau32 old - 2.5fb^-1

Manual cuts are ( R > 0.4 && DPhi < 2.7 && HTall > 0 )

New

Old

5 fb^-1

TTJest (10 fb^-1)

TT

bad (prev, manual)

Cuts:

• NTop,had == 2

Cutflow (latest):

• NTop_hadronic == 2
• Nlepton_veto == 0

Background estimate:

Preliminary:

Tight lepton veto (R>0.4, DPhi<2.8)

Loose lepton veto (R>0.4, DPhi<2.8)

Loose lepton veto (R>0.35, DPhi<2.8)

Loose lepton veto (R>0.30, DPhi<2.8)

Cuts:

• NTop_hadronic == 2
• NTop_leptonic == 0

Best cuts were ( AK8JetRFine >0.32 && TT_AbsDeltaPhiFine<2.8 && H_TallFine>1450 ) Smin = 0.18291

--> Switched to Eff = 1 (for No cut, instead of full efficiency, taking into account 2 hadronic tops)

Best cuts are ( AK8JetRFine >0.32 && TT_AbsDeltaPhiFine<2.8 && H_TallFine>1450 ) Smin = 0.180292

--> Added New Samples to table: DYJets_HT, QCD_PT_bcToE, TToLep _s_t, WJets, ZJets_HT, DYJets_HT, GJets_HT

1) m Gluino = 1300 GeV

Best cuts are ( RFine>0.32 && TTHadDPhiFine <2.8 && HtAllFine >1450 ) Smin = 0.170138

2) m Gluino = 1000 GeV

Best cuts are ( RFine>0.35 && TTHadDPhiFine <2.8 && HtAllFine >1150 ) Smin = 0.131402

--> Below removed 0 leptonic top cut (which was not intentional for the moment)

• NTop_hadronic == 2

1) m Gluino = 1300 GeV

Best cuts are ( RFine>0.41 && TTHadDPhiFine <2.8 && HtAllFine >1100 ) Smin = 0.134513

2) m Gluino = 1300 GeV

Best cuts are ( RFine>0.41 && TTHadDPhiFine <2.8 && HtAllFine >1150 ) Smin = 0.0940552

-- JanosKarancsi - 2015-03-03