Statistics for Resonance Search

Similar documents
arxiv: v1 [hep-ex] 21 Aug 2011

Two Early Exotic searches with dijet events at ATLAS

How to find a Higgs boson. Jonathan Hays QMUL 12 th October 2012

On behalf of the ATLAS and CMS Collaborations

FYST17 Lecture 8 Statistics and hypothesis testing. Thanks to T. Petersen, S. Maschiocci, G. Cowan, L. Lyons

Statistical Data Analysis Stat 3: p-values, parameter estimation

Primer on statistics:

Discovery and Significance. M. Witherell 5/10/12

Statistical methods in CMS searches

Statistics for the LHC Lecture 1: Introduction

Hypothesis testing (cont d)

P Values and Nuisance Parameters

Extra dimensions and black holes at the LHC

ATLAS and CMS diphoton resonance searches at 13 TeV

Physics 736. Experimental Methods in Nuclear-, Particle-, and Astrophysics. - Statistics and Error Analysis -

Di-photon at 750 GeV! (A first read)

Boosted Top Resonance Searches at CMS

W. Fedorko for the ATLAS collaboration

Evidence for D 0 - D 0 mixing. Marko Starič. J. Stefan Institute, Ljubljana, Slovenia. March XLII Rencontres de Moriond, La Thuile, Italy

arxiv: v1 [hep-ex] 8 Nov 2010

Search for high mass diphoton resonances at CMS

Hà γγ in the VBF production mode and trigger photon studies using the ATLAS detector at the LHC

Journeys of an Accidental Statistician

Higgs and Z τ + τ in CMS

arxiv: v6 [physics.data-an] 20 Feb 2018

Harrison B. Prosper. CMS Statistics Committee

Searches using displaced leptons or jets

Probing Dark Matter at the LHC

Feasibility of a cross-section measurement for J/ψ->ee with the ATLAS detector

Inclusive top pair production at Tevatron and LHC in electron/muon final states

Tutorial 8: Discovery of the Higgs boson

Statistical Tools in Collider Experiments. Multivariate analysis in high energy physics

Why I never believed the Tevatron Higgs sensitivity claims for Run 2ab Michael Dittmar

Seminario di fine III anno

PoS(ICHEP2012)238. Search for B 0 s µ + µ and other exclusive B decays with the ATLAS detector. Paolo Iengo

Statistical Methods in Particle Physics Lecture 2: Limits and Discovery

First two sided limit on BR(B s μ + μ - ) Matthew Herndon, University of Wisconsin Madison SUSY M. Herndon, SUSY

Recall the Basics of Hypothesis Testing

Physics 509: Error Propagation, and the Meaning of Error Bars. Scott Oser Lecture #10

Statistical Tools in Collider Experiments. Multivariate analysis in high energy physics

E. Santovetti lesson 4 Maximum likelihood Interval estimation

Statistics for Particle Physics. Kyle Cranmer. New York University. Kyle Cranmer (NYU) CERN Academic Training, Feb 2-5, 2009

Bayes at the Frontier: The Promise and Challenges

B-quark discovery. e+ e- PETRA GeV TRISTAN 61.4 GeV LEP Mz Mt > 45.8 GeV

Search for b Ø bz. CDF note Adam Scott, David Stuart UCSB. 1 Exotics Meeting. Blessing

The X(3872) at the Tevatron

The Tevatron Higgs Search

Search for Quark Substructure in 7 TeV pp Collisions with the ATLAS Detector

Statistical Methods for Discovery and Limits in HEP Experiments Day 3: Exclusion Limits

Early physics with Atlas at LHC

Hypothesis testing. Chapter Formulating a hypothesis. 7.2 Testing if the hypothesis agrees with data

Recent Heavy Flavors results from Tevatron. Aleksei Popov (Institute for High Energy Physics, Protvino) on behalf of the CDF and DØ Collaborations

Hadronic Exotica Searches at CMS

Search for W' tb in the hadronic final state at ATLAS

Searches for Exotic Physics with the ATLAS Detector

First two sided limit on BR(B s µ + µ - ) Matthew Herndon, University of Wisconsin Madison SUSY M. Herndon, SUSY

b hadron properties and decays (ATLAS)

Study on the Two-Photon Transition from ψ(2s) to J/ψ at BESIII

Recent ATLAS measurements in Higgs to diboson channels

Statistical Methods for Particle Physics (I)

QCD Studies at LHC with the Atlas detector

Searches for BSM Physics in Rare B-Decays in ATLAS

The Higgs boson discovery. Kern-und Teilchenphysik II Prof. Nicola Serra Dr. Annapaola de Cosa Dr. Marcin Chrzaszcz

Higgs Production at LHC

Higgs Boson at the CMS experiment

First V+jets results with CMS. Vitaliano Ciulli (Univ. & INFN Firenze) V+jets workshop, 8-10 Sep 2010, Durham

Statistical Methods for Particle Physics Lecture 4: discovery, exclusion limits

STATS 200: Introduction to Statistical Inference. Lecture 29: Course review

Colin Jessop. University of Notre Dame

Use of the likelihood principle in physics. Statistics II

Higgs-related SM Measurements at ATLAS

B-physics with ATLAS and CMS

Searches for Dark Matter with in Events with Hadronic Activity

Median Statistics Analysis of Non- Gaussian Astrophysical and Cosmological Data Compilations

Measurement of EW production! of Z+2j at the LHC

ATLAS: LHC 2016: 240 GeV Higgs Mass State at 3.6 sigma

Lecture 2. G. Cowan Lectures on Statistical Data Analysis Lecture 2 page 1

arxiv: v1 [hep-ex] 8 Jan 2018

Physics 509: Bootstrap and Robust Parameter Estimation

Is there evidence for a peak in this data?

Search for di-higgs to 4b with the ATLAS detector. Tony(Baojia)Tong, Harvard University CLHCP, PKU, Dec. 18, 2016

Hypotheses and Errors

PoS(EPS-HEP2011)250. Search for Higgs to WW (lνlν, lνqq) with the ATLAS Detector. Jonas Strandberg

Combination of top quark physics results at the LHC

Dimitris Iliadis AUTh, LAPP. On behalf of the ATLAS collaboration. XIIth Quark Confinement and the Hadron Spectrum 29 August 2016, Thessaloniki

Text. Decays of heavy flavour hadrons measured by CMS and ATLAS. M.Smizanska, Lancaster University

Contact the LHC

Detection of Z Gauge Bosons in the Di-muon Decay Mode in CMS

Physics 403. Segev BenZvi. Classical Hypothesis Testing: The Likelihood Ratio Test. Department of Physics and Astronomy University of Rochester

b and c production in CMS

Top Quark Production at the LHC. Masato Aoki (Nagoya University, Japan) For the ATLAS, CMS Collaborations

Dark Matter in ATLAS

arxiv: v1 [hep-ph] 24 Oct 2011

Recent Results from the Tevatron

Multivariate Analysis for Setting the Limit with the ATLAS Detector

Hypotheses Testing. Chapter Hypotheses and tests statistics

Search for High Mass SM Higgs at the Tevatron

A search for t t resonances using 3.2 fb 1 of proton-proton collisions at 13 TeV: Boosted analysis

Statistics notes. A clear statistical framework formulates the logic of what we are doing and why. It allows us to make precise statements.

Transcription:

Statistics for Resonance Search Georgios Choudalakis University of Chicago ATLAS exotic diphoton resonances meeting Nov. 5, 0

Intro Introduction I was asked to explain how we treated (statistically) the dijet resonance search. The BumpHunter is already being used in boosted object resonance searches (Elin, Muge et al), and in dilepton searches (Dan Hayden). The same tool could be used in diphoton resonance search, for consistency. Georgios Choudalakis (U.Chicago) Statistics for resonance search page

Introduction Background by fitting We fit a smooth function, through the whole spectrum, to obtain background. We justify this by showing it can fit SM QCD. Fitting the data is not a good argument. If we couldn t fit SM QCD, we wouldn t use a fit; we would use SM QCD as background. Georgios Choudalakis (U.Chicago) Statistics for resonance search page 3

Introduction Two statistical questions Is there a discrepancy between data and background? What is the probability density function (p.d.f.) of the number of RS G γγ signal events? Georgios Choudalakis (U.Chicago) Statistics for resonance search page 4

Introduction The Kolmogorov-Smirnov test, the χ test, etc. KS test statistic: D = max CDF CDF K&S computed the p.d.f. of D, when the two distributions are consistent (i.e. in the null hypothesis). a Knowing the p.d.f. of D, one can reject the null hypothesis if D obs is uncomfortably big: P(D > D obs ) = i= ( )i e i C (N,N )D obs. a This is directly analogous to Pearson s theorem, where he computed the p.d.f. of χ in the null hypothesis, and showed it to follow a χ - distribution. Georgios Choudalakis (U.Chicago) Statistics for resonance search page 5

Genius vs Computer : 0- Introduction Georgios Choudalakis (U.Chicago) Statistics for resonance search page 6

Introduction We can define any test statistic we want, and find experimentally its p.d.f. in the null hypothesis. It turns out there are better tests than KS and χ. In the recent dijet resonance search, we used 6 tests (to be as model-independent as possible, and just because it didn t hurt): KS χ lnl Jeffreys divergence TailHunter BumpHunter Georgios Choudalakis (U.Chicago) Statistics for resonance search page 7

The code Introduction You are welcome to use the code! Writing proper documentation, but in the meanwhile you can cite: Phys.Rev.D79:0,009 & Phys.Rev.D78:000,008 Download package (with example) from SVN: https://svnweb.cern.ch/trac/atlasgrp/browser/institutes/uchicago/dijetmassandchi/trunk/macros/package I ll be happy to provide technical support. Georgios Choudalakis (U.Chicago) Statistics for resonance search page 8

The BumpHunter BumpHunter This version operates on binned histograms. [Generalization to unbinned data is easy.] Decide width, in number of bins, of central window (W C ). Width of each sideband = max(, W C ) 3 Decide position of central window. Start from a position that gives enough room for the left sideband. 4 Cound data (D C ) and background (B C ) in the central window 5 Cound data (D L, D R ) and background (B L, B R ) in the Left and Right sidebands 6 Compute P L,P C,P R, where P X = BX d d=d X d! e B X if D X B X, or P X = D X BX d d=0 d! e B X if D X < B X. [Big excesses or big deficits are both treated as discrepancies.] 7 If D C < B C, or P L < P C, or P R < P C, then P C. [Look for an excess, surrounded by agreeing sidebands. This of course we can tweak to our liking.] 8 Translate the central window and its sidebands by bin to the right, and find P L,P C,P R again. 9 Start a scan with bigger W C, until W C is of the whole spectrum. [This too can be tweaked, if you have a certitude that the new physics will be narrow or wide or have a specific mass-width relationship.] Return lnp min C from all positions and all widths tried. Georgios Choudalakis (U.Chicago) Statistics for resonance search page 9

The BumpHunter BumpHunter Toy example, injecting signal Gaus(00 ± 50 GeV) Events 4 3 Interval range 000 00-0 events 500 00 500 000 500 jj Reconstructed m [GeV] 0 0 500 00 Interval number Georgios Choudalakis (U.Chicago) Statistics for resonance search page

The BumpHunter BumpHunter Toy example, injecting signal Gaus(00 ± 50 GeV) Events 4 3 Interval range 000 00 - events 500 00 500 000 500 jj Reconstructed m [GeV] 0 0 500 00 Interval number Georgios Choudalakis (U.Chicago) Statistics for resonance search page

The BumpHunter BumpHunter Toy example, injecting signal Gaus(00 ± 50 GeV) Events 4 3 Interval range 000 00-30 events 500 00 500 000 500 jj Reconstructed m [GeV] 0 0 500 00 Interval number Georgios Choudalakis (U.Chicago) Statistics for resonance search page

BumpHunter The BumpHunter Toy example, injecting signal Gaus(00 ± 50 GeV) Events 4 3 Interval range 500 000 500 00 500-40 events 500 00 500 000 500 jj Reconstructed m [GeV] 0 500 00 Interval number Georgios Choudalakis (U.Chicago) Statistics for resonance search page 3

BumpHunter The BumpHunter Toy example, injecting signal Gaus(00 ± 50 GeV) Events 4 3 Interval range 500 000 500 00 500-50 events 500 00 500 000 500 jj Reconstructed m [GeV] 0 500 00 Interval number Georgios Choudalakis (U.Chicago) Statistics for resonance search page 4

BumpHunter The BumpHunter Toy example, injecting signal Gaus(00 ± 50 GeV) Events 4 3 Interval range 500 000 500 00 500-60 events 500 00 500 000 500 jj Reconstructed m [GeV] 0 500 00 Interval number Georgios Choudalakis (U.Chicago) Statistics for resonance search page 5

The BumpHunter BumpHunter Toy example, injecting signal Gaus(00 ± 50 GeV) Events 4 3 observed statistic.9668 PoissonPval of interval - - -3-4 500 00 500 000 500 jj Reconstructed m [GeV] 00 500 000 dijet mass [GeV] Georgios Choudalakis (U.Chicago) Statistics for resonance search page 6

The BumpHunter BumpHunter Toy example, injecting signal Gaus(00 ± 50 GeV) Events 4 3 observed statistic.563 PoissonPval of interval - -3-5 -7 500 00 500 000 500-9 00 500 jj Reconstructed m [GeV] dijet mass [GeV] Georgios Choudalakis (U.Chicago) Statistics for resonance search page 7

The BumpHunter BumpHunter Toy example, injecting signal Gaus(00 ± 50 GeV) Pseudo-experiments Pseudo-experiments bump hunter statistic 30 σ band 0 0 5 signal = 0 bump hunter statistic 0 bump hunter statistic signal = 50 events 0 50 0 Number of Signal Events So, if I observe a BumpHunter statistic =, I know the p-value in the Null Hypothesis; it s about 00. If 50 signal events are expected, it will be likely to observe such a statistic, so, it will be likely to reject the Null Hypothesis at the corresponding confidence level. Later we ll talk about sensitivity: What s the chance of making a.6σ discovery ( 5% false-positive probability) if s signal events are expected? How much does s need to be, to have probability β of declaring a discovery with false-positive probability α? Georgios Choudalakis (U.Chicago) Statistics for resonance search page 8

Summary of tests Other statistical tests BumpHunter Pseudo-experiments 3 bump hunter statistic 30 σ band 0 0 5 bump hunter statistic 0 50 0 Number of Signal Events Georgios Choudalakis (U.Chicago) Statistics for resonance search page 9

Summary of tests Other statistical tests TailHunter : Like the BumpHunter, but instead of windows look for tails. No sideband criteria. Pseudo-experiments 3 tail hunter statistic 5 0 5 σ band 5 0 5 tail hunter statistic 0 50 0 Number of Signal Events Georgios Choudalakis (U.Chicago) Statistics for resonance search page 0

Summary of tests Other statistical tests logl = log i Poisson(d i;b i ) Pseudo-experiments 3 -ln(l) 0 σ band 90 80 90 0 -ln(l) 80 0 50 0 Number of Signal Events Georgios Choudalakis (U.Chicago) Statistics for resonance search page

Summary of tests Other statistical tests Pearson s famous χ Pseudo-experiments 3 ) ln(χ 4.5 4.0 σ band 3.5 3.0 3 4 ln(χ ) 0 50 0 Number of Signal Events Georgios Choudalakis (U.Chicago) Statistics for resonance search page

Summary of tests Other statistical tests Jeffreys divergence: A symmetric version of the Kullback-Leibler divergence. It expresses how surprised, on average, I ll be if I expected one distribution and observed the other. See literature on information theory. Pseudo-experiments 3 Jeffreys divergence 0.4 0.3-3 σ band 0. 0. 0. Jeffreys divergence 0. 0 50 0 Number of Signal Events Georgios Choudalakis (U.Chicago) Statistics for resonance search page 3

Summary of tests Other statistical tests The famous KS test. Pseudo-experiments 3 Kolmogorov-Smirnov 0.004 0.003 σ band 0.00 0.000 0.00 0.004 0.006 0.008 Kolmogorov-Smirnov 0.00 0 50 0 Number of Signal Events Georgios Choudalakis (U.Chicago) Statistics for resonance search page 4

Sensitivity Power curves of different tests, for signal = Gaus(00,50) Prob p-val <.6σ effect.5.0 -ln(l(s=0)/l(s)) BumpHunter BumpHunter, no sidebands -ln(l(s=0)) χ TailHunter KS Jeffreys 0.5 0.0 0 50 0 Number of Signal Events Georgios Choudalakis (U.Chicago) Statistics for resonance search page 5

Sensitivity Power curves of different tests, for signal = Gaus(00,400) Prob p-val <.6σ effect.5.0 -ln(l(s=0)/l(s)) BumpHunter BumpHunter, no sidebands -ln(l(s=0)) χ TailHunter KS Jeffreys 0.5 0.0 0 00 400 Number of Signal Events Georgios Choudalakis (U.Chicago) Statistics for resonance search page 6

Sensitivity Power curves of different tests, for signal = Gaus(000,50) Prob p-val <.6σ effect.5.0 -ln(l(s=0)/l(s)) BumpHunter BumpHunter, no sidebands -ln(l(s=0)) χ TailHunter KS Jeffreys 0.5 0.0 0 5 Number of Signal Events Georgios Choudalakis (U.Chicago) Statistics for resonance search page 7

Sensitivity Power curves of different tests, for signal = Gaus(000,400) Prob p-val <.6σ effect.5.0 -ln(l(s=0)/l(s)) BumpHunter BumpHunter, no sidebands -ln(l(s=0)) χ TailHunter KS Jeffreys 0.5 0.0 0 0 30 Number of Signal Events Georgios Choudalakis (U.Chicago) Statistics for resonance search page 8

Bayesian Limit-setting Limt-setting L(s) L(data s) = i bins Poisson(d i;b i +s i ) P(s data) = L(s) Prior(s)/Norm Georgios Choudalakis (U.Chicago) Statistics for resonance search page 9

Convolution Convolution of systematics L(s;λ JES,λ lumi,λ fit uncertainty ) = i bins Poisson(d i;b i (λ fit )+s i (λ JES,λ lumi )) 3-dimensional Gaussian (Normal) p.d.f. in the 3-dimensional space of λs: π( λ). Integration is not made by throwing pseudo-experiments. Instead, we use a grid from -3 to +3 in each dimension. L(s) = d λl(s; λ)π( λ) Georgios Choudalakis (U.Chicago) Statistics for resonance search page 30

Convolution of systematics Frequentist coverage of Bayesian limit Coverage of 95% Bayesian limit.00 ATLAS Preliminary 0.98 0.96 The coverage probability, or the fraction of times that the confidence interval defined by the Bayesian limit contains the true number of signal events, as a function of signal yield for a hypothetical q mass of 900 GeV. In this study, the coverage probabilities were found to lie in the vicinity of 95%, indicating compatibility between Bayesian and frequentist approaches. 0 50 0 Number of Signal Events Georgios Choudalakis (U.Chicago) Statistics for resonance search page 3

Convolution of systematics Systematics Convolution P(signal data) for flat prior 0.04 0.0 ATLAS Preliminary - s=7 TeV Data ( Ldt=96 nb ) Posterior p.d.f. for q*(400) No systematic uncertainty Ldt uncertainty JES and Ldt uncertainty JES and Ldt and fit uncertainty 0.00 0 0 00 Number of Signal Events Georgios Choudalakis (U.Chicago) Statistics for resonance search page 3

Convolution of systematics Conclusion The BumpHunter package offers a variety of model-independent tests. (The BumpHunter and TailHunter are just two of them.) The sensitivity of each method depends on the kind of signal, of course. Convolution of systematics was done more accurately and carefully, using the grid convolution (integration) method. You can use the software already. Documentation is being written. Georgios Choudalakis (U.Chicago) Statistics for resonance search page 33