\documentclass[xcolor=dvipsnames,table]{beamer}
\author[Paul Seyfert]{
Johannes Albrecht\inst{1}, Marta Calvi\inst{2}, Marcin Chrz\k{a}szcz\inst{3,4}, Laura Gavardi\inst{1}, Jon Harrison\inst{5}, Basem Khanji\inst{2}, George Lafferty\inst{5}, Tatiana Likhomanenko\inst{6}, Eduardo Rodrigues\inst{5}, Nicola Serra\inst{4}, \underline{Paul Seyfert\inst{7}}}
\institute[Uni Heidelberg]{
\inst{1}Dortmund,
\inst{2}Milano Bicocca,
\inst{3}Cracow, \inst{4}Zurich, \inst{5}Manchester, \inst{6}Yandex, \inst{7}Heidelberg University
}
\date{\today}
\subject{}
\AtBeginSection[]
{
\begin{frame}<beamer>{}
\tableofcontents[currentsection,currentsubsection]
\end{frame}
}
\begin{document}
\begin{frame}
\titlepage
\end{frame}
\begin{frame}
\begin{enumerate}
\item introduction\vspace{.5em}
\item multivariate technique\vspace{.5em}
\item choice of triggers\vspace{.5em}
\item normalisation\vspace{.5em}
\item backgrounds\vspace{.5em}
\item expected sensitivity\vspace{.5em}
\item model dependence\vspace{.5em}
\end{enumerate}
Major news wrt.\ the previous analysis rounds highlighted in \textcolor{darkgreen}{green}
\end{frame}
\section{introduction}
\begin{frame}
\frametitle{Status $\Ptau\to\Pmu\Pmu\Pmu$}
\begin{columns}
\begin{column}{.62\textwidth}
\begin{fmffile}{sm}
\begin{fmfgraph*}(200,150)
\fmfstraight
\fmfleft{is3,is2,is1,i1,i2,i3,i4,i5,i6}
\fmfright{os3,os2,os1,o1,o2,o3,o4,o5,o6}
\fmf{phantom}{o1,g1}
\fmf{phantom,tension=1.5}{g1,i1}
\fmffreeze
\fmf{photon, tension=2.,lab.side=left, lab=$\Pphoton$}{g1,g2}
\fmf{fermion,lab.side=left,label=$\APmuon$}{o3,g2}
\fmf{fermion,lab.side=left,label=$\Pmuon$}{g2,o5}
\fmf{phantom,tension=1.5}{i6,g2}
\fmffreeze
\fmf{fermion,lab.side=left,label=$\Pgtm$}{i1,w1}
\fmf{dashes,lab.side=right,label=$\PWminus$}{w1,w2}
\fmf{fermion,lab.side=left,label=$\Pmuon$}{w2,o1}
\fmfipath{p[]}
\fmf{plain,tension=0,right,label={$\Pgngt\to\Pgngm$},tag=1}{w1,w2}
\fmffreeze
\fmfiset{p1}{vpath1(__w1,__w2)}
\fmfiv{d.sh=cross,d.ang=0,d.size=5thick}{point length(p1)/2 of p1}
\end{fmfgraph*}
\end{fmffile}
{\footnotesize{
\begin{itemize}
\item charged Lepton Flavour Violation process
\item possible as penguin with neutrino oscillation
\item unmeasurable small
\end{itemize}
}}
\end{column}
\begin{column}{.37\textwidth}
\begin{block}{current limits ($ 90\,\%$ CL)}
\begin{description}
\item[BaBar] $3.3\times 10^{-8}$
\item[Belle] $2.1\times 10^{-8}$
\item[LHCb] $8.0\times 10^{-8}$
\end{description}
\end{block}
\begin{block}{BSM predictions}
\begin{description}
\item[var.\ SUSY] $10^{-10}$
\item[non universal $\PZprime$] $10^{-8}$
\item[mSUGRA+seesaw] $10^{-9}$
\end{description}
\end{block}
\end{column}
\end{columns}
\end{frame}
\begin{frame}
\frametitle{$\Ptau$ production}
\begin{itemize}
\item consider five production channels (fractions at $\unit{8}{\TeV}$):\begin{itemize}
\item prompt $\PDs\to\Ptau$ ($73.1\pm3.9\,\%$)
\item prompt $\PDplus\to\Ptau$ ($0.42\pm0.43\,\%$)
\item non-prompt $\PDs\to\Ptau$ ($9.7\pm2.1\,\%$)
\item non-prompt $\PDplus\to\Ptau$ ($0.02\pm0.02\,\%$)
\item $X_{\Pbottom}\to\Ptau$ (meson or baryon) ($16.8\pm3.0\,\%$)
\end{itemize}
\item \textcolor{darkgreen}{use $\sigma(\Pbottom\APbottom)$ at $\unit{8}{\TeV}$ from LHCb}
\item \textcolor{darkgreen}{use Pythia scaling for $\sigma(\Pcharm\APcharm)$ at $\unit{8}{\TeV}$}
\end{itemize}
\begin{exampleblock}{$\mathcal{B}(\PDplus\to\Ptau)$}
\begin{itemize}
\item old analysis: used upper limit
\item now: $\mathcal{B}(\PDplus\to\Pmu\Pnum)$ + helicity suppression + phase space
\item \texttt{hep-ex:0604043}
\item $\mathcal{B}(\PDplus\to\Ptau\Pnut)=1\pm1\times10^{-4}$
\end{itemize}
\end{exampleblock}
\end{frame}
\begin{frame}
\frametitle{Strategy}
\begin{itemize}
\item mostly as in many rare decay searches:
\item loose stripping selection
\item multivariate classification in: mass, PID, ``geometry/topology''
\item relative normalisation ($\PDs\to\Pphi(\Pmu\Pmu)\Ppi$)
\item invariant mass fit for expected background in each likelihood bin\newline fit in $m-m_{\Ptau}>\unit{30}{\MeV}$
\item ``middle sidebands'' for classifier evaluation ($\unit{20}{\MeV}<m-m_{\Ptau}<\unit{30}{\MeV}$)
\item CLs for limit calculation
\end{itemize}
\begin{block}{today}
\begin{itemize}
\item $\unit{3}{\reciprocal\femtobarn}$ analysis ``as is''
\item nice \& interesting things not included so far e.g.\ 2011 reanalysed
\end{itemize}
\end{block}
\end{frame}
\begin{frame}
\frametitle{Datasets}
\begin{itemize}
\item data from \textcolor{darkgreen}{Reco14Stripping20(r1)}
\item much MC\begin{itemize}
\item \textcolor{darkgreen}{24M} inclusive background events ($\Pbottom\APbottom$ and $\Pcharm\APcharm$)
\item \textcolor{darkgreen}{10M} exclusive background events ($\PDs\to\Peta(\Pmu\Pmu\Pphoton)\Pmu\Pnum$)
\item \textcolor{darkgreen}{2M} signal events (split over 5 production channels)
\end{itemize}
\item[$\Rightarrow$] \textcolor{darkgreen}{generator level cuts} for improved use of computing resources
\begin{itemize}
\item \textcolor{darkgreen}{$\sim 14$ times more} signal statistics after stripping
\item \textcolor{darkgreen}{$\sim 2$ times more} background statistics
\end{itemize}
\item \textcolor{darkgreen}{mix $\Ptau$ production on ntuple level} instead of reweighting.
\newline$\Rightarrow$ ease up ntuple usage (no forgotten weighting, no double weighting, \dots)
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{(Stripping) selection}
{\footnotesize{
\begin{tabular}{|c|cc|}
\hline
&$\Ptau\to\Pmu\Pmu\Pmu$&$\PDs\to\Pphi\Ppi$\\
\hline
$\mu^\pm$ , $ \pi^\pm$ &\multicolumn{2}{c|}{} \\
$p_T$ &\multicolumn{2}{c|}{$>300\MeV$} \\
Track $\chi^2$/ndf &\multicolumn{2}{c|}{$<3 $} \\
IP $\chi^2$/ndf &\multicolumn{2}{c|}{$>9 $} \\
\textcolor{darkgreen}{track ghost probability} &\multicolumn{2}{c|}{\textcolor{darkgreen}{$<0.3 $}} \\
\hline
$\mu$ pairs &\multicolumn{2}{c|}{} \\
$m_{\mu^+\mu^-} - m_{\phi}$ & $>20\MeV$ & $<20\MeV$\\
$m_{\mu^+\mu^-}$ & $> 450\MeV$ & - \\
$m_{\mu^+\mu^+}$ & $> 250\MeV$ & - \\
\hline
$\tau^\pm$ and \PDs &\multicolumn{2}{c|}{} \\
$\Delta m$ & $<400\MeV$ & $<50\MeV$\\
Vertex $\chi^2$ &\multicolumn{2}{c|}{$<15$} \\
\textcolor{darkgreen}{IP $\chi^2$} &\multicolumn{2}{c|}{$<225^1 $} \\
$\cos\alpha$ &\multicolumn{2}{c|}{$>0.99$} \\
$c\tau$ (stripping) &\multicolumn{2}{c|}{$>\unit{100}{\mu m}$} \\
&\multicolumn{2}{c|}{no PV refitting}\\
decay time (offline) &\multicolumn{2}{c|}{$> -0.01$ ns \& $< 0.025$ ns}\\
&\multicolumn{2}{c|}{PV refitting}\\
\hline
\end{tabular}
}}
{\footnotesize{$^1$ different LoKi functor}}
\end{frame}
\section{multivariate technique}
\begin{frame}
\frametitle{``geometric likelihood''}
\begin{itemize}
\item classify the displaced 3-body decay properties of a signal candidate
\item revisit variable choice
\item revisit classification technique
\item \textcolor{darkgreen}{more toolkits tried: MatrixNet, NeuroBayes}, TMVA
\item \textcolor{darkgreen}{retune input variables\newline($\PBs\to\Pmu\Pmu$ isolation $\rightarrow$ Laura's BDT isolation: CERN-THESIS-2013-259)}
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{setup}
\begin{itemize}
\item train $1/3$ signal MC against $1/2$ background MC
\item variables \begin{itemize}
\item $3\times$ DOCA
\item vertex $\chi^2$
\item $\tau$ decay time
\item $\tau$ IP$\chi^2$
\item min.\ $\mu$ IP$\chi^2$
\item $\Ptau$ pointing angle
\item $\tau$ $p_T$
\item max.\ track $\chi^2$
\item $\PBs\to\Pmu\Pmu$ track isolation
\item cone isolation
\item \textcolor{darkgreen}{BDT isolation}
\end{itemize}
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{futher tweaking}
\begin{itemize}
\item \textcolor{darkgreen}{remove fully reconstructed 3-body decays from background sample\newline (don't expect to be able to discriminate these)}
\item don't apply trigger prior to training
\end{itemize}
\begin{exampleblock}{``blending'' technique}
\begin{itemize}
\item for each signal channel we train: one BDT, three Fisher classifier, four MLPs, one FDA, and one LD classifier
\item[$\Rightarrow$] 50 classifiers
\item one final MatrixNet classifier using the 13 base variables and the 50 classifiers as input
\newline(trained on the second $1/3$ of signal MC and the second $1/2$ of background MC)
\end{itemize}
\end{exampleblock}
\end{frame}
\begin{frame}
\frametitle{performance}
\begin{itemize}
\item classifier prefers $\Ptau$ from prompt $\PDs$
\end{itemize}
\begin{columns}
\begin{column}{.48\textwidth}
\begin{block}{MC response for different\newline $\Ptau$ production channels}
\includegraphics[width=.95\textwidth]{./for_paul.png}
\end{block}
\end{column}
\begin{column}{.48\textwidth}
\begin{block}{response for $\PDs\to\Pphi\Ppi$\newline data and MC}
\includegraphics[width=.95\textwidth]{./MN_BLEND_FLAT.png}
\end{block}
\end{column}
\end{columns}
\end{frame}
\begin{frame}
\frametitle{calibration}
\begin{itemize}
\item assume all differences between $\Ptau\to\Pmu\Pmu\Pmu$ and $\PDs\to\Pphi\Ppi$ come from kinematics (mass, resonance, decay time)
\item get correction $\PDs\leadsto\Ptau$ from MC
\item apply corrections to $\PDs\to\Pphi\Ppi$ on data
\end{itemize}
\begin{block}{validation}
\begin{itemize}
\item done for 2011 analysis, treating smeared MC as data
\end{itemize}
\end{block}
\begin{columns}
\begin{column}{.45\textwidth}
\begin{itemize}
\item $\PDs\to\Pphi\Ppi$ well modelled in MC
\item until the very low likelihood end of the distribution
\item[$\rightarrow$] i.e.\ also badly pointing non-prompt $\PDs$
\end{itemize}
\end{column}
\begin{column}{.45\textwidth}
\includegraphics[width=.95\textwidth]{MN_BLEND_FLAT.png}
\end{column}
\end{columns}
\end{frame}
\begin{frame}
\frametitle{PID}
\begin{itemize}
\item we used ProbNNmu already in the previous round of the analysis
\item now use MC12TuneV2 (latest)
\item two-fold reason:\begin{itemize}
\item expect better performance than CombDLL variables
\item ``one variable for everything'':\newline with CombDLL we needed both CombDLL($\mu-\pi$) and CombDLL($\mu-K$)
\end{itemize}
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{PIDCalib}
\begin{itemize}
\item calibration strategy: use PIDCalib
\item confirm with $\PDs\to\Pphi\Ppi$ if everything is fine
\end{itemize}
\begin{columns}
\begin{column}{.45\textwidth}
cut\&fit:
\begin{itemize}
\item fit $\PDs\to\Pphi\Ppi$ with a TIS muon in data
\item cut on ProbNNmu of one muon
\item fit again
\item[$\rightarrow$] ratio is ``true'' cut efficiency
\end{itemize}
\begin{block}{ProbNNmu>0.4}
$\varepsilon=86.3\,\%$
\end{block}
\end{column}
\begin{column}{.45\textwidth}
PIDCalib
\begin{itemize}
\item apply full selection (incl. trigger) to $\PDs\to\Pphi\Ppi$ MC reference sample
\item avoid IsMuon bias
\end{itemize}
\begin{block}{ProbNNmu>0.4}
$\varepsilon=89.8\,\%$
\end{block}
\end{column}
\end{columns}
\end{frame}
\begin{frame}
\frametitle{?}
\begin{itemize}
\item first shown at \myhref{https://indico.cern.ch/event/291727/}{charming VRD meeting}
\item also mentionned at last \myhref{https://indico.cern.ch/event/298021/}{LHCb week}
\item many emails exchanged with Barbara Sciascia
\item mistakes on user side found
\item still no agreement
\end{itemize}
\begin{exampleblock}{phenomenologic treatment}
\begin{itemize}
\item correlations are small in $\PDs\to\Pphi\Ppi$ data and MC
\newline $\varepsilon(\text{cut on one muon})^2 = \varepsilon(\text{cut on two muons})$
\item[$\Rightarrow$] use $c^3=(\varepsilon(\text{cut and fit})/\varepsilon(\text{PIDCalib}))^3$ as correction to PIDCalib for $\Ptau\to\Pmu\Pmu\Pmu$
\item assign error of $0.02$ for $c$
\end{itemize}
\end{exampleblock}
\begin{itemize}
\item planned: investigate further (usage/bug/samples)
\item planned: use muons from $\PDs\to\Pphi\Ppi$ directly
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{binning optimisation}
\begin{itemize}
\item how to optimise the binning in two classifiers?
\item $\unit{1}{\reciprocal\femtobarn}$ CONF note: two one-dimensional optimisations as in $\PBs\to\Pmu\Pmu$
\item $\unit{1}{\reciprocal\femtobarn}$ PAPER: iterative loop of one-dimensional optimisations\newline optimising one classifier on the sensitive range of the other classifier
\item \textcolor{darkgreen}{now: optimise two-dimensions (optimise bin boundaries in both dimensions at the same time)}
\item unchanged: don't use lowest likelihood bins\newline(reflection backgrounds, no sensitivity gain)
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{impact of new binning optimisation}
\begin{itemize}
\item removal of tiny bins which contribute negligible sensitivity
\item colour: limit obtained, using only this particular bin
\item number: rank of that bin (1=best sensitivity bin)
\end{itemize}
~
\begin{columns}
\begin{column}{.5\textwidth}
old analysis
~
\includegraphics[width=.95\textwidth]{./90CLonebinlimit.eps}
\end{column}
\begin{column}{.5\textwidth}
new analysis
(2011 data, not final calibration)
\includegraphics[width=.95\textwidth]{./rank.eps}
\end{column}
\end{columns}
\end{frame}
\begin{frame}
\frametitle{mass shape}
\begin{itemize}
\item double-Gaussian with fixed fraction ($70\,\%$ inner Gaussian)
\item fix fraction to ease calibration:\newline
$\sigma_{data}^{\Ptau} = \frac{\sigma_{MC}^{\Ptau}}{\sigma_{MC}^{\PDs}}\times\sigma_{data}^{\PDs}$
\end{itemize}
\includegraphics[width=.44\textwidth]{./Ds_data_2011.pdf}
\includegraphics[width=.44\textwidth]{./Ds_data_2012.pdf}
{\footnotesize{
\begin{tabular}{|c|c|c|}
\hline
calibrated $\Ptau$ mass shape & 7~TeV & 8~TeV\\
\hline
Mean ($\MeV$) & $1779.1 \pm 0.1$ & $1779.0 \pm 0.1$\\
\hline
$\sigma_1$ ($\MeV$) & $7.7 \pm 0.1$ & $7.6 \pm 0.1$\\
\hline
$\sigma_2$ ($\MeV$) & $12.0 \pm 0.8$ & $11.5 \pm 0.5$\\
\hline
\end{tabular}
}
}
\end{frame}
\beamertemplateshadingbackground{darkgreen!30}{PineGreen!30}
\section{choice of triggers}
\begin{frame}
\frametitle{the story}
\begin{itemize}
\item first look at $\PDs\to\Pphi\Ppi$ revealed: signal/background much worse in 2012
\item[$\rightarrow$] charm got more bandwidth in 2012
\item taking all events quite unsatisfactory:\begin{itemize}
\item strictly speaking: we don't know why the events ended up on tape
\item trigger efficiencies unstable (TIS/TPS/TOS efficiencies for lots of lines and TCKs)
\item signal anyhow mostly TOS in a muon trigger
\end{itemize}
\item[$\Rightarrow$] if an event was background in the trigger, then don't consider it signal afterwards.
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{strategy}
\begin{itemize}
\item back of the envelope: limit scales with $1/\varepsilon_\text{sig}$
\item back of the envelope: limit scales with $\sqrt{\varepsilon_\text{bkg}}$
\item[$\Rightarrow$] minimise $FOM=\frac{\sqrt{\varepsilon_\text{bkg}}}{\varepsilon_\text{sig}}$
\item running TCKsh and some combinatorics: $\mathcal{O}(2^{256}\approx 10^{75})$ possibilities of using HLT2 line (ignoring that there are different TCKs, TIS/TOS/TPS/DEC, \dots)
\item[$\rightarrow$] be more pragmatic!
\end{itemize}
\begin{exampleblock}{trigger optimisation}
\begin{itemize}
\item sort trigger lines (Punzi FoM)
\item start with events from the best trigger line and compute $FOM$
\item add events from the next trigger and recompute $FOM$
\item iterate and converge to $FOM(\text{all triggers})$
\item use all triggers until $FOM$ starts rising.
\end{itemize}
\end{exampleblock}
\end{frame}
\begin{frame}
\frametitle{our triggers}
{\footnotesize{
\begin{tabular}{l|c|c}
& signal & normalisation \\\hline\hline
L0$^1$ & \multicolumn{2}{c}{L0Muon TOS}\\\hline
Hlt1$^1$ & \multicolumn{2}{c}{Hlt1TrackMuon TOS}\\\hline
Hlt2 2011 & Hlt2CharmSemilepD2HMuMu TOS & Hlt2DiMuonDetached$^2$ TOS \\
& || Hlt2TriMuonTau TOS & \\\hline
Hlt2 2012 & Hlt2TriMuonTau$^1$ TOS & Hlt2DiMuonDetached$^2$ TOS\\\hline
\end{tabular}
}
}
\only<1>{
\begin{block}{$^1$ triggers in 2012}
\begin{itemize}
\item cuts changed through 2012
\item[$\rightarrow$] \textcolor{darkgreen}{emulated two different TCKs for 2012}
\end{itemize}
\end{block}}
\only<2>{
\begin{block}{$^2$ word on Hlt2DiMuonDetached}
\begin{itemize}
\item keep it simple here
\item line unchanged in 2012
\item[$\rightarrow$] choice keeps Hlt2 trigger efficiency stable
\item $\PDs\to\Pphi\Ppi$ anyhow doesn't behave like $\Ptau\to\Pmu\Pmu\Pmu$ in the TriMuon trigger (requires misidentification)
\end{itemize}
\end{block}}
\end{frame}
\begin{frame}
\frametitle{cross check (not in the note)}
\begin{itemize}
\item shouldn't the multivariate classifier do better than any trigger?
\item back-of-the-envelope overestimates the improvement (expect $\sim 9\,\%$ improvement)
\item[$\rightarrow$] add events back to the ntuple, recalculate normalisation, redid fits
\item[$\Rightarrow$] \textcolor{darkgreen}{restricting the triggers gains $\sim 3\,\%$ sensitivity wrt.\ previous round}
\end{itemize}
\end{frame}
\beamertemplateshadingbackground{White}{White}
\section{normalisation}
\begin{frame}
\frametitle{relative normalisation}
$\mathcal{B}(\Ptau\to\Pmu\Pmu\Pmu) = \frac{\mathcal{B}(\PDs\to\Pphi\Ppi)}{\mathcal{B}(\PDs\to\Ptau\Pnut)} \times f_{\PDs}^{\Ptau} \times \frac{\varepsilon_\text{norm} }{\varepsilon_\text{sig} } \times \frac{N_\text{sig}}{N_\text{norm}} = \alpha\times N_\text{sig}$
\begin{itemize}
\item where $\varepsilon$ stands for trigger, reconstruction, selection
\item $\text{norm}$ = normalisation channel $\PDs\to\Pphi\Ppi$
\item $f_{\PDs}^{\Ptau}$ is the fraction of $\Ptau$ coming from $\PDs$
\newline i.e.\ $(83\pm3)\,\%$ for 2012
\end{itemize}
\includegraphics[width=.47\textwidth]{./Ds_data_2011.pdf}
\includegraphics[width=.47\textwidth]{./Ds_data_2012.pdf}
\end{frame}
\begin{frame}[allowframebreaks]
\frametitle{normalisation in numbers}
{\footnotesize{
$\begin{array}{c|c|c}
& \rm{7~TeV} & \rm{8~TeV}\\
\hline
\rm{\epsilon\mathstrut_{sig}}^{GEN} & 8.98 \pm 0.41 & 9.20 \pm 0.36\\
\hline
\rm{\epsilon_{cal}}^{GEN} & 11.19 \pm 0.34 & 11.53 \pm 0.32\\
\hline
\rm{\epsilon_{sig}}^{REC,isMuon,SEL} & 9.794 \pm 0.028 & 9.134 \pm 0.023 \\
\hline
\rm{\epsilon_{cal}}^{REC,isMuon,SEL} & 7.187 \pm 0.022 & 6.690 \pm 0.022 \\
\hline
\frac{\rm{c_{cal}}^{track}}{\rm{c_{sig}}^{track}} & 0.997 \pm 0.009 \pm 0.026 & 0.996 \pm 0.009 \pm 0.026 \\
\hline
\frac{\rm{c_{cal}}^{\mu ID}}{\rm{c_{sig}}^{\mu ID}} & 0.9731 \pm 0.0031 \pm 0.0264 & 1.0071 \pm 0.0022 \pm 0.0204 \\
\hline
\rm{c}^{\Pphi} & \multicolumn{2}{c}{0.98 \pm 0.01} \\
\hline
\rm{c}^{\Ptau} & 1.032 \pm 0.006 & 1.026 \pm 0.006\\
\hline
\rm{c}^{trash} & 1.95 \pm 0.12 & 2.05 \pm 0.12\\
\hline
\rm{\epsilon\mathstrut_{sig}}^{TRIG} & 35.45 \pm 0.11 \pm 0.14 & 39.1 \pm 1.7 \pm 2.0 \\
\hline
\rm{\epsilon\mathstrut_{cal}}^{TRIG} & 23.42 \pm 0.14 \pm 0.09 & 20.62 \pm 0.76 \pm 1.07 \\
\end{array}$
}}
\framebreak
{\footnotesize{
$\begin{array}{c|c|c}
& \rm{7~TeV} & \rm{8~TeV}\\
\hline
\mathcal{B}(\PDs\to\Pphi\Ppi) & \multicolumn{2}{c}{(1.317 \pm 0.099) \times 10^{-5}}\\
\hline
f^{\Ptau}_{\PDs} & 0.82 \pm 0.03 & 0.83 \pm 0.03 \\
\hline
\mathcal{B}(\PDs\to\Ptau\Pnut) & \multicolumn{2}{c}{0.0561 \pm 0.0024}\\
\hline
\rm{\epsilon\mathstrut_{norm}}^{REC\&SEL}/
\rm{\epsilon\mathstrut_{sig}}^{REC\&SEL}
& 0.897 \pm 0.061 & 0.926 \pm 0.056 \\
\hline
\rm{\epsilon\mathstrut_{norm}}^{TRIG}/
\rm{\epsilon\mathstrut_{sig}}^{TRIG}
& 0.6606 \pm 0.0059 & 0.527 \pm 0.041\\
\hline
N_{norm} & 28,162 \pm 434 & 51,998 \pm 684\\
\hline & \\[-1.5em]\hline
\alpha & (4.05 \pm 0.48) \times 10^{-9} & (1.83 \pm 0.25) \times 10^{-9}\\
\alpha^{trash} & (7.90 \pm 0.49) \times 10^{-9} & (3.75 \pm 0.27) \times 10^{-9}\\
\end{array}$
}}
\end{frame}
\section{backgrounds}
\begin{frame}
\frametitle{misidentification}
\begin{itemize}
\item most dominant: $\PDplus\to\PK\Ppi\Ppi$
\item experience from last round: cut away low ProbNNmu range
\item check remaining data under $\PK\Ppi\Ppi$ hypothesis for $\PDplus$ peak
\item[$\Rightarrow$] misid safely contained in ``trash'' bin
\item \textcolor{darkgreen}{$\PDplus\to\Ppi\Ppi\Ppi$ and $\PDs\to\Ppi\Ppi\Ppi$ start to become visible in 2012}
\end{itemize}
\includegraphics[width=.45\textwidth]{./Dp2Kpipi_all_2012_senseBins.pdf}
\includegraphics[width=.45\textwidth]{./FittoD23pi_2012.pdf}
\end{frame}
\begin{frame}
\frametitle{evil backgrounds}
\begin{itemize}
\item $\Pphi\to\Pmu\Pmu + X$: narrow veto on dimuon mass
\item $\PDs\to\Peta(\Pmu\Pmu\Pphoton)\Pmu\Pnum$: not so easy
\begin{itemize}
\item modelled in CONF note
\item optimised veto in PAPER
\item right now: both versions in the ANA note
\end{itemize}
\item baseline: veto $m_{\APmuon\Pmuon} < \unit{450}{\MeV}$
\begin{itemize}
\item fits better understood
\item sensitivity unchanged when removing veto
\item smaller uncertainty on expected background
\end{itemize}
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{remaining backgrounds}
\begin{itemize}
\item fit exponential to invariant mass spectrum in each likelihood bin
\item don't use $\pm \unit{30}{\MeV}$ in the fit
\item[$\rightarrow$] compatible results blinding only $\pm \unit{20}{\MeV}$\footnote{partially used in classifier developement}
\end{itemize}
{\begin{center}
most sensitive bins in 2011 and 2012
\includegraphics[width=.4\textwidth]{./fit2011.png}
\includegraphics[width=.4\textwidth]{./fit2012.png}
\end{center}}
\end{frame}
\section{results}
\begin{frame}
\frametitle{expected limit}
\begin{itemize}
\item still blinded
\item consider nuisance parameters from background fit, signal pdf calibration, normalisation
\item nuisance parameters due to $\Ptau$ production not included in signal pdf shape, yet
\item limit for combined 2011+2012 analysis
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{sensitivity}
$\mathcal{B}(\Ptau\to\Pmu\Pmu\Pmu)<5.6 \times 10^{-8}$ at 90\% CL
\includegraphics[width=.8\textwidth]{./banana.png}
\end{frame}
\beamertemplateshadingbackground{PineGreen!30}{White}
\section{model dependence}
\begin{frame}
\frametitle{model dependence}
\begin{itemize}
\item $\Peta$ veto $\Rightarrow$ our limit not applicable to New Physics with small $m_{\APmuon\Pmuon}$
\item model independent description in \texttt{arXiv:0707.0988}
\item 5 relevant Dalitz distributions: 2 four-point operators, 1 radiative operator, 2 interference terms
\end{itemize}
\only<2->{
\begin{itemize}
\item with radiative distribution limit gets worse by $51\,\%$ (dominantly from the $\Peta$ veto)
\item the other four Dalitz distributions behave nicely (within $7\,\%$)
\end{itemize}
}
\only<1>{
\includegraphics[width=.331\textwidth]{./gammallll.eps}
\includegraphics[width=.331\textwidth]{./gammallrr.eps}
\includegraphics[width=.331\textwidth]{./gammarad.eps}
\includegraphics[width=.331\textwidth]{./gammarad-llll.eps}
\includegraphics[width=.331\textwidth]{./gammarad-llrr.eps}
}
\end{frame}
\beamertemplateshadingbackground{White}{White}
\begin{frame}
\frametitle{Conclusion}
\begin{columns}
\begin{column}{.55\textwidth}
\begin{itemize}
\item finally all pieces put together
\item model (in)dependence of $\Peta$ veto investigated
\item expected sensitivity computed\newline $5.6\times 10^{-8}$
\end{itemize}
\end{column}
\begin{column}{.45\textwidth}
\includegraphics[width=\textwidth]{party-music-hd-wallpaper-1920x1200-3850.jpg}
\end{column}
\end{columns}
\end{frame}
\end{document}
mchrzasz