| 1 |
< |
\fixme{need to summarize decay mode} |
| 1 |
> |
The Tau Neural classifier introduces two complimentary new techniques for tau |
| 2 |
> |
lepton physics at CMS: reconstruction of the hadronic tau decay mode and |
| 3 |
> |
discrimination from quark and gluon jets using neural networks. The decay mode |
| 4 |
> |
reconstruction strategy presented in section~\ref{sec:decay_mode_reco} |
| 5 |
> |
significantly improves the determination of the decay mode. This information has |
| 6 |
> |
the potential to be useful in studies of tau polarization and background |
| 7 |
> |
estimation. |
| 8 |
> |
|
| 9 |
> |
The Tau Neural classifier tau identification algorithm significantly improves |
| 10 |
> |
tau discrimination performance compared to isolation--based |
| 11 |
> |
approaches~\cite{PFT08001} used in previous CMS analyses. |
| 12 |
> |
Figure~\ref{fig:finalPerfCurve} compares the performance of the ``shrinking |
| 13 |
> |
cone'' isolation tau--identification algorithm~\cite{PFT08001} to the |
| 14 |
> |
performance of the TaNC for a scan of requirements on the transformed neural |
| 15 |
> |
network output. The signal efficiency and QCD di--jet fake rate versus |
| 16 |
> |
tau--candidate transverse momentum and pseudo--rapidity for the four benchmark |
| 17 |
> |
points and the isolation based tau identification are show in |
| 18 |
> |
figure~\ref{fig:kinematicPerformance}. For tau--candidates with transverse |
| 19 |
> |
momentum between 20 and 50 GeV/c, the TaNC operating cut can be chosen such that |
| 20 |
> |
the two methods have identical signal efficiency; at this point the TaNC |
| 21 |
> |
algorithm reduces the background fake rate by an additional factor of 3.9. This |
| 22 |
> |
reduction in background will directly improve the significance of searches for |
| 23 |
> |
new physics using tau leptons at CMS. |
| 24 |
|
|
| 25 |
< |
The Tau Neural classifier algorithm significantly improves tau identification |
| 4 |
< |
performance compared to isolation--based strategies~\cite{PFT08001} used in |
| 5 |
< |
previous CMS strategies. Figure~\ref{fig:finalPerfCurve} compares the |
| 6 |
< |
performance of the ``shrinking cone'' isolation tau--identification |
| 7 |
< |
algorithm~\cite{PFT08001} to the performance of the TaNC for a scan of |
| 8 |
< |
requirements on the transformed neural network output. For tau--candidates with |
| 9 |
< |
transverse momentum between 20 and 50 GeV/c, the TaNC operating cut can be |
| 10 |
< |
chosen such that the two methods have identical signal efficiency; at this point the TaNC |
| 11 |
< |
algorithm reduces the background fake rate by an additional factor of |
| 12 |
< |
\fixme{get exact}. |
| 13 |
< |
|
| 14 |
< |
The discriminator output of the TaNC algorithm is a continuous quantity, |
| 15 |
< |
enabling analysis specific optimization of the selection to maximize |
| 16 |
< |
sensitivity. For the convenience of the user, four operating point benchmark |
| 17 |
< |
selections are provided in addition to the continuous output. The four operating |
| 18 |
< |
points are chosen such that for tau--candidates with transverse momentum between |
| 19 |
< |
20 and 50 GeV/c, the expected QCD di--jet fake rate will be 0.1\%, 0.25\%, |
| 20 |
< |
0.50\% and 1.0\%, respectively. The signal efficiency and QCD di--jet fake rate |
| 21 |
< |
versus tau--candidate transverse momentum and pseudo--rapidity for the four |
| 22 |
< |
benchmark points and the isolation based tau identification are show in |
| 23 |
< |
figure~\ref{fig:kinematicPerformance}. |
| 24 |
< |
|
| 25 |
< |
\fixme{needs a zippy end} |
| 26 |
< |
|
| 27 |
< |
\begin{figure}[t] |
| 25 |
> |
\begin{figure}[thbp] |
| 26 |
|
\setlength{\unitlength}{1mm} |
| 27 |
|
\begin{center} |
| 28 |
|
\begin{picture}(150, 150)(0,0) |
| 38 |
|
di--jets with generator--level transverse momentum between 20 and 50 GeV/c are incorrectly |
| 39 |
|
identified as taus. The performance point for the same tau--candidates using |
| 40 |
|
the isolation based tau--identification~\cite{PFT08001} used in many previous |
| 41 |
< |
CMS analyses is indicated by the black star in the figure. |
| 41 |
> |
CMS analyses is indicated by the black star in the figure. An additional requirement |
| 42 |
> |
that the signal cone of the tau contain one or three charged hadrons (typical |
| 43 |
> |
in a final physics analysis) has been |
| 44 |
> |
applied to the isolation based tau--identification to ensure a conservative |
| 45 |
> |
comparison. |
| 46 |
|
} |
| 47 |
|
\label{fig:finalPerfCurve} |
| 48 |
|
\end{center} |
| 49 |
|
\end{figure} |
| 50 |
|
|
| 51 |
|
|
| 52 |
< |
\begin{figure}[t] |
| 52 |
> |
\begin{figure}[thbp] |
| 53 |
|
\setlength{\unitlength}{1mm} |
| 54 |
|
\begin{center} |
| 55 |
|
\begin{picture}(150, 150)(0,0) |
| 60 |
|
\put(2.5, 0) |
| 61 |
|
{\mbox{\includegraphics*[height=70mm]{figures/eff_signal_pt.pdf}}} |
| 62 |
|
\put(75, 0) |
| 63 |
< |
{\mbox{\includegraphics*[height=70mm]{figures/eff_signal_pt.pdf}}} |
| 63 |
> |
{\mbox{\includegraphics*[height=70mm]{figures/eff_signal_eta.pdf}}} |
| 64 |
|
\end{picture} |
| 65 |
< |
\caption{ |
| 66 |
< |
\fixme{add caption and legend, isolation+1 or 3 prong is the black trend, |
| 67 |
< |
pt greater than 20 cut applied on eta plots} |
| 65 |
> |
\caption{ Comparison of the identification efficiency for hadronic tau decays from |
| 66 |
> |
$Z \rightarrow \tau^{+} \tau^{-}$ decays (bottom row) and the misidentification |
| 67 |
> |
rate for QCD di--jets (top row) versus tau--candidate transverse momentum |
| 68 |
> |
(left) and pseudo-rapidity (right) for different tau identification |
| 69 |
> |
algorithms. The efficiency (fake--rate) in a given bin is defined as the |
| 70 |
> |
quotient of the number of true tau hadronic decays (generator level jets) in |
| 71 |
> |
that bin that are matched to a reconstructed tau--candidate that passes the |
| 72 |
> |
identification algorithm divided by the number of true tau hadronic decays |
| 73 |
> |
(generator level jets) in that bin. In the low transverse momentum region |
| 74 |
> |
both the number of tau--candidates in the denominator and the algorithm |
| 75 |
> |
acceptance vary rapidly with respect to $P_T$ for both signal and background; |
| 76 |
> |
a minimum transverse momentum requirement of 20 GeV/c is applied to the |
| 77 |
> |
psuedorapidity plots to facilitate interpretation of the plots. |
| 78 |
|
} |
| 79 |
|
\label{fig:kinematicPerformance} |
| 80 |
|
\end{center} |
