| 1 |
friis |
1.1 |
\documentclass{article}
|
| 2 |
|
|
|
| 3 |
veelken |
1.3 |
%\title{New techniques for decay mode reconstruction and identification of
|
| 4 |
|
|
%hadronic tau lepton decays [outline]}
|
| 5 |
|
|
\title{The Tau Neural Classifier algorithm: tau identification and decay mode reconstruction using neural networks}
|
| 6 |
friis |
1.1 |
\author{Evan K. Friis}
|
| 7 |
|
|
|
| 8 |
|
|
\begin{document}
|
| 9 |
|
|
|
| 10 |
|
|
\maketitle
|
| 11 |
|
|
\tableofcontents
|
| 12 |
|
|
|
| 13 |
friis |
1.2 |
\abstract{
|
| 14 |
veelken |
1.3 |
%Description of a new method for identifying hadronically decaying taus that
|
| 15 |
|
|
%improves the tau identification efficiency on hadroncially decaying taus from
|
| 16 |
|
|
%Z->tautau events while lowering the number of quark and gluon jets from QCD
|
| 17 |
|
|
%di--jet events that are mis-tagged as taus.
|
| 18 |
|
|
%jets.
|
| 19 |
|
|
%\begin{itemize}
|
| 20 |
|
|
% \item Reconstructs the decay mode of the tau
|
| 21 |
|
|
% \item Novel neural networks corresponding to different decay modes of the tau
|
| 22 |
|
|
%\end{itemize}
|
| 23 |
|
|
The Tau Neural Clssifier (TaNC) is a novel algorithm for identification of hadronic tau decays.
|
| 24 |
|
|
The algorithm includes two ocmponents, the reconstruction of tau lepton hadronic decay modes
|
| 25 |
|
|
and discrimination of tau lepton hadronic decays from quark and gluon jets.
|
| 26 |
|
|
The reconstruction of decay modes is based on the reconstruction of individual charged hadrons and photons
|
| 27 |
|
|
by the particle--flow algorithm
|
| 28 |
|
|
and is utilized in the discrimination to train a set of neural networks using input variables
|
| 29 |
|
|
that are sensitive to particular decay modes.
|
| 30 |
|
|
We observe a significant improvement in identification performance in comparisson to previous algorithms.
|
| 31 |
friis |
1.2 |
}
|
| 32 |
|
|
|
| 33 |
friis |
1.1 |
\section{Introduction}
|
| 34 |
veelken |
1.3 |
%Taus are an important part of the physics program at CMS.
|
| 35 |
|
|
%\begin{itemize}
|
| 36 |
|
|
% \item Higgs Boson have an enhanced coupling to taus due to their high mass.
|
| 37 |
|
|
% \item In MSSM, this coupling is enhanced by tanBeta
|
| 38 |
|
|
% \item For certain Higgs mass ranges, the tau decay channel offers best
|
| 39 |
|
|
% discovery potential.
|
| 40 |
|
|
% \item Tau leptons can decay to electrons or muons.
|
| 41 |
|
|
% \item But Tau leptons are unique in that their are the only lepton that can decay
|
| 42 |
|
|
% to hadrons. (1 or 3 pions)
|
| 43 |
|
|
% \item In this paper we describe a novel method for identifying hadronic
|
| 44 |
|
|
% decays of taus.
|
| 45 |
|
|
% \item Methods for discriminating against electron and muons are described in
|
| 46 |
|
|
% PFT-08-001
|
| 47 |
|
|
%\end{itemize}
|
| 48 |
|
|
%
|
| 49 |
|
|
%Identifying taus is difficult at hadron colliders.
|
| 50 |
|
|
%\begin{itemize}
|
| 51 |
|
|
% \item Taus production in channels of interest is a relatively rare
|
| 52 |
|
|
% phenomenon.
|
| 53 |
|
|
% \item The decay signature of the tau lepton is very similar to electron,
|
| 54 |
|
|
% muon, quark and gluon jets which are produced in abundance.
|
| 55 |
|
|
%\end{itemize}
|
| 56 |
|
|
%
|
| 57 |
|
|
%\subsection{Tau Identification}
|
| 58 |
|
|
%A description of the tau identification algorithms used in past CMS physics analysis.
|
| 59 |
|
|
%We propose an extension to these methods.
|
| 60 |
|
|
%\begin{itemize}
|
| 61 |
|
|
% \item CaloTaus versus PFTaus
|
| 62 |
|
|
% \item ParticleFlow blurb
|
| 63 |
|
|
% \item PFTau have better ET and angular resolution and can resolve individual
|
| 64 |
|
|
% photons
|
| 65 |
|
|
% \item To remove QCD, and isolation requirement is applied, described in
|
| 66 |
|
|
% PFT-08-001
|
| 67 |
|
|
% \item A Et dependent signal cone has been developed to separate
|
| 68 |
|
|
% signal and isolation regions.
|
| 69 |
|
|
% \item Performance is on the order of O(0.01)
|
| 70 |
|
|
% \item Plot: Shrinking Cone performance from PFT-08-001
|
| 71 |
|
|
%\end{itemize}
|
| 72 |
|
|
|
| 73 |
|
|
A good tau identification performance is important for the discovery potential of many possible new physics signals at the LHC.
|
| 74 |
|
|
\begin{itemize}
|
| 75 |
|
|
\item typically are signal processes
|
| 76 |
|
|
\item quark and gluon jets produced with significantly larger cross--sections
|
| 77 |
|
|
\item efficient identification of hadronic tau decays and low misidentification rate for quarks and gluons
|
| 78 |
|
|
thus essential for many searches for new physics
|
| 79 |
|
|
\end{itemize}
|
| 80 |
|
|
|
| 81 |
|
|
New physics signals may be discovered via tau lepton hadronic decays in early CMS data.
|
| 82 |
|
|
\begin{itemize}
|
| 83 |
|
|
\item for example, MSSM Higgs to production cross--section of which is enhanced by tan(beta)
|
| 84 |
|
|
\item but also for discovery of Standard Model Higgs, a good tau identification performance is important,
|
| 85 |
|
|
as Higgs $\rightarrow$ tau decays have the second largest branching fraction
|
| 86 |
|
|
\end{itemize}
|
| 87 |
|
|
|
| 88 |
|
|
Tau leptons are unique in that they are the only type of leptons which are heavy enough to decay to hadrons.
|
| 89 |
|
|
\begin{itemize}
|
| 90 |
|
|
\item lifetime $c \cdot \tau = 87 \mu$~m
|
| 91 |
|
|
\item BR(e) ~ BR(mu) ~ 17%
|
| 92 |
|
|
\item BR(hadrons) ~ 65%;
|
| 93 |
|
|
mostly either one or three charged pions plus zero to two neutral pions,
|
| 94 |
|
|
which almost instanteneously decay to photons
|
| 95 |
|
|
\end{itemize}
|
| 96 |
|
|
|
| 97 |
|
|
In this note, we will concentrate on the identification of hadronic tau decays.
|
| 98 |
|
|
\begin{itemize}
|
| 99 |
|
|
\item tau decays to electrons and muons are difficult to distinguish from electrons and muons produced in $pp$ collision
|
| 100 |
|
|
(strategy depends on analysis, tau decays to electrons and muons typically identified by requiring
|
| 101 |
|
|
two leptons of differenct flavor)
|
| 102 |
|
|
\item discrimination of hadronic tau decays from electrons and muons is described in PFT--08--001
|
| 103 |
|
|
\item ``signal'' signature the identification of which we aim to improve with the Tau Neural Classifier (TaNC)
|
| 104 |
|
|
is collimated jet containing either one or three tracks reconstructed in Pixel and silicon Strip tracker,
|
| 105 |
|
|
plus low number of neutral electromagnetic showers reconstructed in the ECAL
|
| 106 |
friis |
1.2 |
\end{itemize}
|
| 107 |
friis |
1.1 |
|
| 108 |
|
|
\subsection{TaNC motivation}
|
| 109 |
friis |
1.2 |
The different hadronic decay modes of the tau come from different resonance. Provides
|
| 110 |
|
|
additional information. Can re-frame the search into search for rhos, a1s, etc.
|
| 111 |
|
|
\begin{itemize}
|
| 112 |
|
|
\item Each decay mode has a different topology and different possibilities
|
| 113 |
|
|
for discrimination.
|
| 114 |
|
|
\item The tau decay can have 1 || 3 pions and a number of pi0s.
|
| 115 |
|
|
\item Each decay mode multiplicity maps directly to a resonance (@ 95\%
|
| 116 |
|
|
level)
|
| 117 |
|
|
\item This note presents two complimentary techniques: a method to
|
| 118 |
|
|
reconstruct the decay mode and an ensemble of neural network discriminants
|
| 119 |
|
|
used to classify tau--candidates.
|
| 120 |
|
|
\item Plot: True visible invariant mass for different decay modes
|
| 121 |
|
|
\end{itemize}
|
| 122 |
|
|
|
| 123 |
|
|
\section{Decay Mode Reconstruction}
|
| 124 |
veelken |
1.3 |
The signal
|
| 125 |
|
|
CV: add reference to shrinking cone note CMS AN--2008/026
|
| 126 |
|
|
cone photons are merged into candidate pi0s and the candidates are
|
| 127 |
friis |
1.2 |
subject to a minimum pT quality requirement to remove contamination from various
|
| 128 |
|
|
sources.
|
| 129 |
|
|
\begin{itemize}
|
| 130 |
|
|
\item pi0s undergo prompt decay to photons.
|
| 131 |
|
|
\item The number of photons present in the signal cone has a long tail due to
|
| 132 |
|
|
UE, PU, showers, photon conversions.
|
| 133 |
|
|
\item Plot: number of photons versus number of pi-zeros
|
| 134 |
|
|
\end{itemize}
|
| 135 |
friis |
1.1 |
|
| 136 |
|
|
\subsection{Photon Merging}
|
| 137 |
friis |
1.2 |
Photons are merged into composite pi0s by looking at the invariant mass of each
|
| 138 |
|
|
combination of photons.
|
| 139 |
|
|
\begin{itemize}
|
| 140 |
|
|
\item Only photon pairs that have mass less than 0.2GeV are considered.
|
| 141 |
|
|
\item CMS Ecal granularity and particle flow clustering provide excellent
|
| 142 |
|
|
resolution.
|
| 143 |
|
|
\item Plot: di photon mass for decay mode 1.
|
| 144 |
|
|
\end{itemize}
|
| 145 |
|
|
|
| 146 |
|
|
\subsection{Quality requirements}
|
| 147 |
|
|
To remove contamination from pile-up and underlying event, a minimum pt quality
|
| 148 |
|
|
requirement is applied to the remaining photon candidates.
|
| 149 |
|
|
\begin{itemize}
|
| 150 |
|
|
\item The lowest pt photon is required to carry 10\% of the composite visible
|
| 151 |
|
|
pt
|
| 152 |
|
|
\item This removes contaminant photons while preserving single photons that
|
| 153 |
|
|
correspond to pi0s
|
| 154 |
|
|
\item Plots: photon pt fraction for DM0 and DM1
|
| 155 |
|
|
\end{itemize}
|
| 156 |
|
|
|
| 157 |
|
|
\subsection{Results}
|
| 158 |
|
|
The decay mode reconstruction algorithm dramatically improves the determination
|
| 159 |
|
|
of the decay mode.
|
| 160 |
|
|
\begin{itemize}
|
| 161 |
|
|
\item Tails removed
|
| 162 |
|
|
\item Mean improved
|
| 163 |
|
|
\item Plot: correlation plot
|
| 164 |
|
|
\end{itemize}
|
| 165 |
|
|
|
| 166 |
|
|
The distribution of the decay modes is different for signal and background. The
|
| 167 |
|
|
decay mode determination is slightly dependent on pt and eta.
|
| 168 |
|
|
|
| 169 |
|
|
\begin{itemize}
|
| 170 |
|
|
\item pt turn on curve is due to pt quality thresholds and cone size
|
| 171 |
|
|
\item Blowup of 1prong1pi0 fraction at eta = 2.5 due to loss of tracker + no
|
| 172 |
|
|
loss of ECAL?
|
| 173 |
|
|
\item NB that the distribution of the decay modes is another handle that the
|
| 174 |
|
|
TaNC has.
|
| 175 |
|
|
\item Plot: Decay mode for sig/bkg vs. pt and eta
|
| 176 |
|
|
\end{itemize}
|
| 177 |
friis |
1.1 |
|
| 178 |
|
|
\section{Neural network classification}
|
| 179 |
friis |
1.2 |
For each decay mode, a different neural network is used.
|
| 180 |
|
|
\begin{itemize}
|
| 181 |
|
|
\item The five decay modes we use constitute 95\% of hadronic decays.
|
| 182 |
|
|
\item Table of the five decays
|
| 183 |
|
|
\item Other decay modes are discarded.
|
| 184 |
|
|
\item Each neural net has inputs that are specific to that decay mode.
|
| 185 |
|
|
\item Each neural net is trained on a tau--candidates reconstructed with the
|
| 186 |
|
|
associated decay mode.
|
| 187 |
|
|
\item During final discrimination, the neural network associated with the
|
| 188 |
|
|
reconstructed decay mode of the tau candidate is used to do the
|
| 189 |
|
|
classification.
|
| 190 |
|
|
\item Since five neural networks are used a strategy must be used to select
|
| 191 |
|
|
the cut used on each neural network output.
|
| 192 |
|
|
\end{itemize}
|
| 193 |
|
|
|
| 194 |
|
|
\subsection{Neural network discriminants}
|
| 195 |
veelken |
1.3 |
The neural networks use
|
| 196 |
|
|
%discriminants
|
| 197 |
|
|
as input variables observables
|
| 198 |
|
|
specific to each decay mode.
|
| 199 |
|
|
%Discriminants
|
| 200 |
|
|
The observables are listed in the appendix.
|
| 201 |
|
|
Common
|
| 202 |
|
|
%discriminants
|
| 203 |
|
|
observables include:
|
| 204 |
friis |
1.2 |
\begin{itemize}
|
| 205 |
|
|
\item Pt/Eta
|
| 206 |
|
|
\item Invariant mass
|
| 207 |
|
|
\item Pt and DR from axis of signal objects
|
| 208 |
|
|
\item Pt and DR from axis of isolation objects
|
| 209 |
|
|
\item Number of charged isolation objects
|
| 210 |
|
|
\item Sum charged pt in isolation
|
| 211 |
|
|
\item For three body decays, the two dalitz variables
|
| 212 |
|
|
\item Include separation and correlation plots for all variables?
|
| 213 |
veelken |
1.3 |
CV: yes, please (in appendix)
|
| 214 |
friis |
1.2 |
\end{itemize}
|
| 215 |
|
|
|
| 216 |
|
|
\subsection{Neural network training}
|
| 217 |
|
|
|
| 218 |
|
|
The signal and background samples are split into five subsamples corresponding
|
| 219 |
|
|
to each decay mode.
|
| 220 |
|
|
\begin{itemize}
|
| 221 |
|
|
\item Ztautau matched to hadronic taus for signal, QCD Dijet for bkg
|
| 222 |
|
|
\item The leading pion pt requirement is applied.
|
| 223 |
|
|
\item Table of signal/background training events for each mode.
|
| 224 |
|
|
\end{itemize}
|
| 225 |
|
|
|
| 226 |
|
|
The decay mode is dependent on pt and eta and this dependence must be invisible
|
| 227 |
|
|
to the neural network.
|
| 228 |
|
|
\begin{itemize}
|
| 229 |
|
|
\item The kinematics are very different for signal/background
|
| 230 |
|
|
\item We want to prevent the NN from training on these differences
|
| 231 |
|
|
\item Weighting is applied so the weighted pt/eta distributions are identical
|
| 232 |
|
|
\item Since the probability for a given decay mode to occur is kinematically
|
| 233 |
|
|
dependent, the weighting is applied to the subset of the sample that
|
| 234 |
|
|
corresponds to ensemble of allowed decay modes.
|
| 235 |
|
|
\end{itemize}
|
| 236 |
|
|
|
| 237 |
|
|
The neural networks are implemented as TMVA back-propagating neural networks.
|
| 238 |
|
|
\begin{itemize}
|
| 239 |
|
|
\item Number of hidden nodes = Kolmogorov function N + 1 (2*N + 1)
|
| 240 |
|
|
\item 500 training epochs, testing for over-training every ten
|
| 241 |
|
|
\item No over-training is detected. (need plots?)
|
| 242 |
veelken |
1.3 |
CV: yes, please show NN output error on training and on validation dataset
|
| 243 |
|
|
(two curves overlayed on same plot which has training epoch on the x--axis and NN output error on the y--axis)
|
| 244 |
|
|
for at least one of the decay modes/neural networks (as example)
|
| 245 |
friis |
1.2 |
\end{itemize}
|
| 246 |
|
|
|
| 247 |
|
|
\subsection{Individual neural network performance}
|
| 248 |
|
|
The separation power of the individual neural net is different. The ultimate separation
|
| 249 |
|
|
power of the algorithm depends on both the individual neural net separation
|
| 250 |
|
|
performance and decay mode distribution differences between signal and
|
| 251 |
|
|
background.
|
| 252 |
|
|
\begin{itemize}
|
| 253 |
|
|
\item Plots of each decay mode separation
|
| 254 |
|
|
\item Example: 1prong1pi0 has no discrimination power for isolated OneProng
|
| 255 |
|
|
QCD
|
| 256 |
|
|
\end{itemize}
|
| 257 |
|
|
|
| 258 |
|
|
\subsection{Neural network output selections}
|
| 259 |
|
|
Since there are five neural networks, a discrimination working point requires
|
| 260 |
|
|
selection of a point in five-D space.
|
| 261 |
|
|
\begin{itemize}
|
| 262 |
|
|
\item Monte Carlo cut point selection
|
| 263 |
|
|
\item A 5D point is added to the performance curve if it has a higher
|
| 264 |
|
|
signal efficiency than the current point with the same background mis-tag
|
| 265 |
|
|
rate.
|
| 266 |
|
|
\item Separate samples are used for selecting the 5D curve, and evaluating
|
| 267 |
|
|
its performance.
|
| 268 |
|
|
\end{itemize}
|
| 269 |
|
|
|
| 270 |
|
|
The 5D performance curve can also be parameterized by using the probability for a
|
| 271 |
|
|
tau--candidate to be identified for a given decay mode.
|
| 272 |
|
|
\begin{itemize}
|
| 273 |
|
|
\item The method transforms the output of each neural net according to the
|
| 274 |
|
|
decay mode probability
|
| 275 |
|
|
\item The decay mode probability is dependent on pt/eta
|
| 276 |
|
|
\item Derivation of transform
|
| 277 |
|
|
\item Net discriminant output is now a single continuous variable
|
| 278 |
|
|
\item Recommended method of using the TaNC
|
| 279 |
|
|
\item Plot: comparison of transform to MC-determined optimal curve
|
| 280 |
|
|
\end{itemize}
|
| 281 |
|
|
|
| 282 |
|
|
\subsection{Algorithm Performance}
|
| 283 |
|
|
The TaNC algorithm identifies true hadronic tau decays with a much higher purity
|
| 284 |
|
|
than algorithms previously used in CMS analyses.
|
| 285 |
|
|
\begin{itemize}
|
| 286 |
|
|
\item Plot: performance curve
|
| 287 |
|
|
\item With transform, cut is a continuous variable
|
| 288 |
|
|
\item Comparison with shrinking/fixed cone
|
| 289 |
|
|
\end{itemize}
|
| 290 |
|
|
|
| 291 |
|
|
\section{Future work}
|
| 292 |
|
|
The TaNC algorithm has been optimized for the initial stages of LHC operation.
|
| 293 |
|
|
\begin{itemize}
|
| 294 |
|
|
\item Will need to be retrained when luminosity changes
|
| 295 |
|
|
\item Once enough data comes, backgrounds will be trained with data events
|
| 296 |
|
|
\end{itemize}
|
| 297 |
friis |
1.1 |
|
| 298 |
|
|
\end{document}
|
