hep-ex0408099/AnalysisMethod.tex
1: \section{ANALYSIS METHOD}
2: \label{sec:Analysis}
3: 
4: The $U$ and $I$ coefficients and the $\Btopipipi$ event yield are
5: determined by a maximum-likelihood fit of the signal model to the 
6: selected candidate events. Kinematic and event shape variables 
7: exploiting the characteristic properties of the events are used 
8: in the fit to discriminate signal from background. We limit the 
9: size of the data sample that enters the fit by tightening the acceptance
10: requirements for the discriminant variables compared to similar analyses, 
11: because the fit model with at least 17 physical parameters is rather involved.
12: With the same goal and because the modeling of the distribution of 
13: the continuum events in the Dalitz plot is delicate, we remove the center 
14: of the Dalitz plot from the analysis. This requirement does not affect 
15: the signal.
16: 
17: \subsection{EVENT SELECTION AND BACKGROUND SUPPRESSION}
18: \label{subsec:selection}
19: 
20: We reconstruct $\Btopipipi$ candidates from pairs of 
21: oppositely-charged tracks, forming a good quality vertex, and a 
22: $\pi^0$ candidate. 
23: %
24: We use information from the tracking system, EMC, and DIRC to 
25: remove tracks for which the PID is consistent with the electron, kaon, 
26: or proton hypotheses. In addition, we require that at least one track 
27: has a signature in the IFR that is inconsistent with the muon 
28: hypothesis.
29: %
30: The $\pi^0$ candidate mass must satisfy $0.11<m(\gamma\gamma)<0.16\gevcc$, 
31: where each photon is required to have an energy greater than $50\mev$
32: in the laboratory frame (LAB) and to exhibit a lateral profile of energy 
33: deposition in the EMC consistent with an electromagnetic shower.
34: 
35: Two requirements are applied on the Dalitz plot. Firstly, 
36: the invariant mass of the two tracks $m_0$ 
37: must be larger than $0.52\gevcc$. This rejects
38: about 80\% of the $B^0 \to \KS(\to\pipi) \pi^0$ background events, which due
39: to the long lifetime of the $\KS$ would require a dedicated $\dt$
40: treatment. This cut retains 98\% (100\%) of signal $B^0 \to \rho^0\pi^0$ 
41: ($B^0 \to \rho^\pm \pi^\mp$) events. Secondly, we remove the center 
42: of the Dalitz plot by requiring that at least one of the three invariant 
43: masses, $m_0$, $m_+$ or $m_-$, is lower than $1.5~\mathrm{GeV/c^2}$. 
44: % The advantages of this last 
45: % cut are twofold: first, it decreases strongly the correlation existing 
46: % between the Dalitz plot and the Neural-Network (NN) used to discriminate 
47: % between signal and continuum $e^+e^- \to q\bar{q} (q=u,d,s,c)$ 
48: % events -- see below for details on the NN. Then, it removes 12\% 
49: % of continuum background in the sample selection without affecting 
50: % the signal events, dominated by the $\rho(770)$ resonance and which 
51: % heavily acculumate along the Dalitz plot borders.
52: 
53: A $B$-meson candidate is characterized kinematically by the energy-substituted 
54: mass $\mes=\lbrack{(\half s+\pvec_0\cdot\pvec_B)^2/E_0^2-\pvec_B^2}\rbrack^\half$
55: and energy difference $\de = E_B^*-\half\sqrt{s}$, 
56: where $(E_B,\pvec_B)$ and $(E_0,\pvec_0)$ are the four-vectors
57: of the $B$-candidate and the initial electron-positron system,
58: respectively. The asterisk denotes the \FourS\  frame,
59: and $s$ is the square of the invariant mass of the electron-positron system.  
60: We require $5.272 < \mes <5.288\gevcc$, which retains $81\%$
61: of the signal and $8\%$ of the continuum background events. 
62: The $\de$ resolution 
63: exhibits a dependence on the $\pi^0$ energy and therefore varies 
64: across the Dalitz plot. We account for this effect by introducing
65: the transformed quantity $\deprime=(2\de - \demax - \demin)/(\demax - \demin)$,
66: with $\deminmax(\mpm)=c_{\pm}-\left(c_{\pm}\mp\bar c\right)(\mpm/\mpmMax)^2$,
67: where $\mpm$ monitors the $\piz$-energy dependence. We use the values
68: $\bar c = 0.045\gev$, $c_{-} = -0.140\gev$, $c_{+} = 0.080\gev$,
69: $\mpmMax = 5.0\gev$, and require $-1<\deprime<1$. 
70: These settings have been obtained from Monte Carlo (MC) simulation and
71: are tuned to maximize the selection of correctly reconstructed over 
72: misreconstructed signal events. The cut retains $75\%$ ($25\%$)
73: of the signal (continuum).
74: 
75: Backgrounds arise primarily from random combinations in continuum events.
76: To enhance discrimination between signal and continuum, we 
77: use a neural network (NN) to combine four discriminating variables: 
78: the angles with respect to the beam axis of the $B$ momentum and $B$ thrust 
79: axis in the \FourS\ frame, and the zeroth and second order monomials
80: $L_{0,2}$ of the energy flow about the $B$ thrust axis.  The monomials
81: are defined by $ L_j = \sum_i p_i\times\left|\cos\theta_i\right|^j$,
82: where $\theta_i$ is the angle with respect to the $B$ thrust axis of
83: track or neutral cluster $i$, $p_i$ is its momentum, and the sum
84: excludes the $B$ candidate.  
85: The NN is trained in the signal region with off-resonance data and
86: simulated signal events. The final sample of signal candidates 
87: is selected with a cut on the NN output that retains $77\%$ ($8\%$) 
88: of the signal (continuum).
89: 
90: The time difference $\deltat$ is obtained from the measured distance between 
91: the $z$ positions (along the beam direction) of the $\Bz_{\tpi}$ and 
92: $\Bz_{\rm tag}$ decay vertices, and the boost $\beta\gamma=0.56$ of 
93: the \epem\ system\footnote
94: {
95:   $\deltat$ is defined as: $\deltat = \Delta z/\beta\gamma c$ 
96: }. 
97: To determine the flavor of the $\Bz_{\rm tag}$ 
98: we use the tagging algorithm of Ref.~\cite{BabarS2b}.
99: This produces four mutually exclusive tagging categories. We also 
100: retain untagged events in a fifth category to improve the efficiency 
101: of the signal selection and because these events contribute to the 
102: measurement of direct \CP violation. Events with multiple \B 
103: candidates passing the full selection occur 
104: in $16\%$ $(\rho^\pm\pi^\mp)$ and $9\%$ $(\rho^0\pi^0)$ 
105: of the cases. If the multiple candidates have different $\pi^0$'s, 
106: we choose the candidate with the reconstructed $\pi^0$ mass closest 
107: to the nominal one; if not, one candidate is selected at random. 
108: 
109: The signal efficiency determined from MC simulation is $24\%$ for 
110: $B^0 \to \rho^\pm\pi^\mp$ and $B^0 \to \rho^0\pi^0$ events, and 
111: $11\%$ for non-resonant $\Btopipipi$ events. 
112: 
113: Of the selected signal events, $22\%$ ($B^0 \to \rho^\pm\pi^\mp$), 
114: $13\% $ ($B^0 \to \rho^0\pi^0$), and $6\%$ (non-resonant) are 
115: misreconstructed, mostly due to combinatorial background from 
116: low-momentum tracks and photons. They concentrate in the corners
117: of the Dalitz plot. The fraction of misreconstructed events strongly
118: varies across the tagging categories.
119: 
120: \subsection{BACKGROUND FROM OTHER {\em B} DECAYS}
121: 
122: We use MC-simulated events to study the background from other $B$ 
123: decays. The exclusive \B-background modes are grouped into eighteen 
124: classes with similar kinematic and topological properties. More
125: than hundred decay channels have been considered of which thirty-six are 
126: retained in the likelihood model. The most significant ones are
127: $B^+ \to \rho^+\rho^0$ with longitudinal polarization ($27 \pm 18$ events 
128: expected), $B^+ \to \pi^+\rho^0$ 
129: ($48 \pm 6$), $B^+ \to \pi^0\rho^+$ ($43 \pm 7$), $B^0 \to \rho^+\rho^-$ 
130: with longitudinal polarization ($50 \pm 10$), $B^0 \to (a_1\pi)^0$ 
131: ($29 \pm 11$), $B^0 \to \rho^-K^+$ ($61 \pm 11$), and $B^0 \to$~higher 
132: kaon resonances ($6 \pm 1$). The charmed modes 
133: $B^0 \to D^- (\to \pi^-\pi^0) \pi^+$ and $B^0 \to \Dzb (\to \pi^+\pi^-) \pi^0$ 
134: contribute to the selected data sample and are considered 
135: in individual classes.  They do not interfere with the signal 
136: due to the long $D$ lifetime. We also assign classes to the modes
137: $B^0 \rar \Dzb(\rar K^+\pi^-)\pi^0$ and $B^0 \rar J/\psi(\rar\ell^+\ell^-)\pi^0$.
138: In total we expect $49\pm15$ exclusive $b\to c$ events.
139: Two additional classes account for inclusive neutral and charged 
140: $b\to c$ decays, where we expect $82 \pm 6$ and $181 \pm 9$ events, 
141: respectively.
142: 
143: %
144: % ============================ MAXIMUM-LIKELIHOOD FIT =============================
145: %
146: 
147: \subsection{THE MAXIMUM-LIKELIHOOD FIT}
148: \label{subsec:ML}
149: 
150: We perform an unbinned extended maximum-likelihood fit to extract
151: the inclusive $\Btopipipi$ event yield and the $U$ and $I$ coefficients
152: defined in Eqs.~(\ref{eq:firstObs})--(\ref{eq:lastObs}). 
153: The fit uses the variables $\mes$, $\deprime$, the NN output, and the 
154: Dalitz plot to discriminate signal from background. The 
155: $\dt$ measurement allows to determine mixing-induced \CP violation
156: and provides additional continuum-background rejection. 
157: 
158: The selected on-resonance data sample is assumed to consist of signal, 
159: continuum-background and \B-background components, separated by the 
160: flavor and tagging category of the tag side \B decay. 
161: The signal likelihood consists of the sum of a correctly 
162: reconstructed (``truth-matched'', TM) component and a misreconstructed 
163: (``self-cross-feed'', SCF) component.
164: 
165: The probability density function (PDF) ${\cal P}_i^\cat$ for an
166: event $i$ in tagging category $\cat$ is the sum of the probability densities 
167: of all components, namely
168: %
169: \beqn
170: \label{eq:theLikelihood}
171: 	{\cal P}_i^\cat
172: 	&\equiv& 
173: 		N_{\tpi} f^\cat_{\tpi}
174: 		\left[ 	(1-\fscfave^\cat){\cal P}_{\tpi-\TM,i}^\cat +
175: 			\fscfave^\cat{\cal P}_{\tpi-\SCF,i}^\cat 
176: 		\right] 
177: 		\nonumber\\[0.3cm]
178: 	&&
179: 		+\; N^\cat_{q\bar q}\frac{1}{2}
180: 		\left(1 + \Qtagi\Atagqq\right){\cal P}_{q\bar q,i}^\cat
181: 		\nonumber \\[0.3cm]
182: 	&&
183: 		+\; \sum_{j=1}^{N^{B^+}_{\rm class}}
184: 		N_{B^+j} f^\cat_{B^+j}
185: 		\frac{1}{2}\left(1 + \Qtagi \Atagj\right){\cal P}_{B^+,ij}^\cat
186: 		\nonumber \\[0.3cm]
187: 	&&
188: 		+\; \sum_{j=1}^{N^{B^0}_{\rm class}}
189: 		N_{B^0j} f^\cat_{B^0j}
190: 		{\cal P}_{B^0,ij}^\cat~,
191: \eeqn
192: where: 
193: 	$N_{\tpi}$ is the total number of $\pi^+\pi^-\pi^0$ signal events 
194: 	the data sample;
195: %
196:  	$f^\cat_{\tpi}$ is the fraction of signal events that are 
197:        	tagged in category $\cat$;
198: %
199: 	$\fscfave^\cat$ is the fraction of SCF events in tagging category $\cat$, 
200: 	averaged over the Dalitz plot;
201: %
202: 	${\cal P}_{\tpi-\TM,i}^\cat$ and ${\cal P}_{\tpi-\SCF,i}^\cat$
203: 	are the products of PDFs of the discriminating variables used
204: 	in tagging category $\cat$ for TM and SCF
205: 	events, respectively; 
206: %
207:  	$N^\cat_{q\bar q}$ is the number of continuum events that are 
208: 	tagged in category $\cat$;
209: %
210: 	$\Qtagi$ is the tag flavor of the event, defined to be 
211: 	$+1$ for a $\Bz_{\rm tag}$ and $-1$ for a $\Bzb_{\rm tag}$; 
212: %	
213: 	$\Atagqq$ parameterizes possible tag asymmetry in continuum events; 
214: %
215: 	${\cal P}_{q\bar q,i}^\cat$ is the continuum PDF for tagging 
216: 	category $\cat$;
217: %
218: 	$N^{B^+}_{\rm class}$ ($N^{B^0}_{\rm class}$) is the number of 
219: 	charged (neutral) $B$-related background classes considered in the fit;
220: %
221: 	$N_{B^+j}$ ($N_{B^0j}$) is the number of expected events in
222: 	the charged (neutral) $B$-background class $j$;
223: %
224: 	$f^\cat_{B^+j}$ ($f^\cat_{B^0j}$) is the fraction of 
225: 	charged (neutral) $B$-background events of class $j$
226: 	that are tagged in category $\cat$;
227: %
228: 	$\Atagj$ describes a possible tag asymmetry in the charged-$B$ background
229: 	class $j$; 
230: 	correlations between the tag and the position in the Dalitz plot 
231: 	(the ``charge'') are absorbed in tag-flavor-dependent 
232: 	Dalitz plot PDFs that are used for charged-\B and continuum
233: 	background;
234: %
235: 	${\cal P}_{B^+,ij}^\cat$ is the $B^+$-background PDF for tagging 
236: 	category $\cat$ and class $j$;
237: %
238: 	finally, ${\cal P}_{B^0,ij}^\cat$ is the neutral-$B$-background 
239: 	PDF for tagging category $\cat$ and class $j$.
240: %
241: 
242: The PDFs ${\cal P}_{X}^{\cat}$ are the product of the four PDFs of the 
243: discriminating variables,
244: $x_1 = m_{ES}$, $x_2 = \deprime$, $x_3 = {\rm NN output}$, and the triplet
245: $x_4 = \{\mprime, \thetaprime, \deltat\}$:
246: % ($m_{ES}$, $\Delta E$, the NN output, 
247: %\{the Dalitz plot and $\Delta t$\}, $x_k$, $k=1,\dots,4$:
248: \beq
249: \label{eq:likVars}
250: 	{\cal P}_{X,i(j)}^{\cat} \;\equiv\; 
251: 	\prod_{k=1}^4 P_{X,i(j)}^\cat(x_k)~.
252: \eeq
253: The extended likelihood over all tagging categories is given by
254: %
255: \beq
256: 	{\cal L} \;\equiv\;  
257: 	\prod_{\cat=1}^{5} e^{-\overline N^\cat}\,
258: 	\prod_{i}^{N^\cat} {\cal P}_{i}^\cat~,
259: \eeq
260: %
261: where $\overline N^\cat$ is the total number of events expected in category 
262: $\cat$. 
263: 
264: A total of 39 parameters, including the inclusive signal yield and the
265: parameters from Eq.~(\ref{eq:dt}), are varied in the fit.
266: 
267: 
268: \subsubsection{\boldmath THE $\dt$ AND DALITZ PLOT PDFS}
269: 
270: \begin{figure}[t]
271:   \centerline{\epsfxsize8.2cm\epsffile{figures/effsquareall.eps}
272:    	      \epsfxsize8.2cm\epsffile{figures/effsquarescf.eps}}
273:   \vspace{-0.4cm}
274:   \caption[.]{\label{fig:effscfDP}
275: 	Selection efficiency of $\Btopipipi$ events (left) 
276: 	and fraction of misreconstructed events (right)
277: 	in the (symmetrized) square Dalitz plot for
278: 	MC-simulated events.}
279: \end{figure}
280: 
281: {\bf Signal Parameterization}.
282: 	The Dalitz plot PDFs require as input the Dalitz plot-dependent 
283: 	relative selection efficiency, $\e=\e(\mprime,\thetaprime)$, 
284: 	and SCF fraction, $\fscf=\fscf(\mprime,\thetaprime)$.
285: 	Both quantities are taken from MC simulation. They are 
286: 	given in Fig.~\ref{fig:effscfDP} (left plot for $\e$ and right 
287: 	plot for $\fscf$), where the symmetry of the Dalitz plot has been 
288: 	used to fold the upper $\thetaprime$ half into the lower one.
289: 	Away from the Dalitz plot corners the efficiency is uniform, while it 
290: 	decreases when approaching the corners, where one out of the 
291: 	three bodies in the final state is close to rest so that the 
292: 	acceptance requirements on the particle reconstruction become incisive.
293: 	Combinatorial backgrounds and hence SCF fractions are large in the corners
294: 	of the Dalitz plot due to the presence of soft neutral clusters 
295: 	and tracks. 
296: 
297: 	For an event~$i$, we define the time-dependent Dalitz plot PDFs
298: 	\beqn
299: 		P_{\tpi-\TM,i} &=&
300: 		\varepsilon_i\,(1 - \fscfi)\,\detJi\,\AmpAll~,
301: 		\\[0.3cm]
302: 		P_{\tpi-\SCF,\,i} &=&
303: 		\varepsilon_i\,\fscfi\,\detJi\,\AmpAll~,
304: 	\eeqn	
305: 	where $P_{\tpi-\TM,i}$ and $P_{\tpi-\SCF,\,i}$ are normalized. The 
306: 	corresponding phase space integration involves the expectation values 	
307: 	$\langle \varepsilon\,(1-\fscf)\,\detJ \,f^\kappa f^{\sigma*}\rangle$
308: 	and 
309: 	$\langle \varepsilon\,\fscf\,\detJ\, f^\kappa f^{\sigma*}\rangle$
310: 	for TM and SCF events, where the indices $\kappa$, $\sigma$ 
311: 	run over all resonances belonging to the signal model.
312: 	The expectation values are model-dependent and are 
313: 	computed with the use of MC integration over the square Dalitz plot:
314: 	\beq
315: 	\label{eq:normAverage}
316: 		\langle \varepsilon\,(1-\fscf)\,\detJ\, f^\kappa f^{\sigma*}\rangle
317: 		\;=\; \frac{\int_0^1\int_0^1 
318: 			    \varepsilon\,(1-\fscf)\,\detJ\, f^\kappa f^{\sigma*}
319: 			\,d\mprime d\thetaprime}
320: 		       {\int_0^1\int_0^1 \varepsilon\,\detJ\, f^\kappa f^{\sigma*}
321: 			\,d\mprime d\thetaprime}~,
322: 	\eeq
323: 	and similarly for 
324: 	$\langle \varepsilon\,\,\detJ\, f^\kappa f^{\sigma*}\rangle$,
325: 	where all quantities in the integrands are Dalitz plot-dependent.
326: 
327: 	Equation~(\ref{eq:theLikelihood}) invokes the phase 
328: 	space-averaged SCF fraction 
329: 	$\fscfave\equiv\langle\fscf\,\detJ\, f^\kappa f^{\sigma*}\rangle$. 
330: 	As for the PDF normalization, it is decay-dynamics-dependent.
331: 	It has to be computed iteratively, though the remaining systematic
332: 	uncertainty after one iteration step is small. We 
333: 	determine the average SCF fractions separately for each tagging category 
334: 	from MC simulation. 
335: 	
336: 	The width of the dominant $\rho(770)$ resonance is large compared 
337: 	to the mass resolution for TM events (about $8\mevcc$ core Gaussian
338: 	resolution). We can therefore neglect resolution effects in the TM 
339: 	model.	
340: 	Misreconstructed events	have a poor mass resolution that strongly 
341: 	varies across the Dalitz plot. It is described in the fit by a 
342: 	$2\times 2$-dimensional resolution function
343: \begin{figure}[t]
344:   \centerline{\epsfxsize8.2cm\epsffile{figures/dpressquaretm.eps}
345:    	      \epsfxsize8.2cm\epsffile{figures/dpressquarescf.eps}}
346:   \vspace{-0.4cm}
347:   \caption[.]{\label{fig:dpres}
348: 	Resolution for TM (left) and SCF events (right hand plot) 
349: 	in the square Dalitz plot for two coordinates indicated by 
350: 	the open stars. }
351: \end{figure}
352: 	\beq
353: 	\label{eq:rscf}
354: 		\Rscf(\mprime_r,\thetaprime_r,\mprime_t,\thetaprime_t)~,
355: 	\eeq
356: 	which represents the probability to reconstruct at the coordinate
357: 	$(\mprime_r,\thetaprime_r)$ an event that has the true coordinate 
358: 	$(\mprime_t,\thetaprime_t)$. It obeys the unitarity condition
359: 	\beq
360: 		\intl_0^1\intl_0^1 
361: 		\Rscf(\mprime_r,\thetaprime_r,\mprime_t,\thetaprime_t)
362: 		\,d\mprime_r d\thetaprime_r \;=\; 1~,
363: 		\hspace{0.5cm} \forall\; (\mprime_t,\thetaprime_t)\in {\rm SDP}~,
364: 	\eeq
365: 	and is convolved with the signal model. 
366: 	The $\Rscf$ function is obtained from MC simulation.
367: 	Figure~\ref{fig:dpres} shows the resolution function of TM (left) 
368: 	and SCF events (right) for two coordinates depicted by the 
369: 	open stars.
370: 
371: 	We use the signal model described in Section~\ref{sec:kinmeatics}. 
372: 	It contains the dynamical information and is connected with $\dt$ via 
373: 	the matrix element~(\ref{eq:dt}), which serves as PDF. It is diluted 
374: 	by the effects of mistagging and the limited vertex 
375: 	resolution~\cite{rhopipaper}. 
376: 	The $\deltat$ resolution function for signal and \B-background 
377: 	events is a sum of three Gaussian distributions, with parameters 
378: 	determined by a fit to fully reconstructed $\Bz$ 
379: 	decays~\cite{BabarS2b}. 
380: \\[0.3cm]\noindent
381: {\bf Background Parameterization}.
382: 	The Dalitz plot- and $\dt$-dependent PDFs factorize for the 
383: 	charged-$B$-background modes, but not (necessarily) 
384: 	for the neutral-$B$ background due to $\BzBzb$ mixing.
385: 
386:  	The charged \B-background
387: 		contribution to the likelihood~(\ref{eq:theLikelihood}) invokes 
388: 		the parameter $\Atag$, multiplied by the tag flavor $\Qtag$ of 
389: 		the event. In the presence of significant tag-``charge'' 
390: 		correlation (represented by an effective 
391: 		flavor-tag-versus-Dalitz-coordinate correlation),
392: 		it parameterizes possible direct \CP violation in these events.
393: 		We also use distinct square Dalitz plot PDFs for each 
394: 		reconstructed $B$ flavor tag, and a flavor-tag-averaged PDF for 
395: 		untagged events. The PDFs are obtained from MC simulation and are 
396: 		described with the use of non-parametric functions.
397: 		The $\dt$ resolution parameters are determined by a fit to fully 
398: 		reconstructed $\Bp$ decays. For each $\Bp$-background class we adjust 
399: 		effective lifetimes to account for the misreconstruction of the 
400: 		event that modifies the nominal $\dt$ resolution function.
401: 
402: 	The neutral-$B$ background is parameterized with PDFs that
403: 		depend on the flavor tag of the event. In the case of \CP
404: 		eigenstates, correlations between the flavor tag and the Dalitz 
405: 		coordinate are expected to be small. However, non-\CP  eigenstates,
406: 		such as $a_1^\pm\pi^\mp$, may exhibit such correlation. Both types 
407: 		of decays can have direct
408: 		and mixing-induced \CP  violation. A third type of decays
409: 		involves charged kaons and does not exhibit mixing-induced
410: 		\CP  violation, but usually has a strong correlation between the
411: 		flavor tag and the Dalitz plot coordinate (the kaon charge), because 
412: 		it consists of $B$-flavor eigenstates.
413: 		The Dalitz plot PDFs are obtained from MC simulation and are 
414: 		described with the use of non-parametric functions.
415: 		For neutral $B$ background, the signal $\dt$ resolution model 
416: 		is assumed.
417: 
418: 	The Dalitz plot
419: 		treatment of the continuum events is similar to the one used
420: 		for charged-$B$ background. 
421: 		The square Dalitz plot PDF for continuum background is 
422: 		obtained from on-resonance events selected in the
423: 		$\mes$ sidebands and corrected for feed-through
424: 		from \B decays. A large number of cross checks has been 
425: 		performed to ensure the high fidelity of the empirical shape 
426: 		parameterization. Analytical models have been found insufficient.
427: 		The continuum $\deltat$ distribution is parameterized as the sum of 
428: 		three Gaussian distributions with common mean and 
429: 		three distinct widths that scale the $\dt$ per-event error. 
430: 		This yields six shape parameters that are determined by 
431: 		the fit.
432:  		The model is motivated by the observation that 
433: 		the $\dt$ average is independent of its error, and that the 
434: 		$\dt$ RMS depends linearly on the $\dt$ error.
435: 
436: \subsubsection{PARAMETERIZATION OF THE OTHER VARIABLES}
437: \label{sec:likmESanddE}
438: 
439: 	The $\mes$ distribution of TM signal events is
440: 		parameterized by a bifurcated Crystal Ball function~\cite{PDFsCB},
441: 		which is a combination of a one-sided Gaussian and 
442: 		a Crystal Ball function. The mean of this function
443: 		is determined by the fit. A non-parametric
444: 		function is used to describe the SCF signal component.
445: 
446: 	The $\deprime$ distribution of TM events is
447: 		parameterized by a double Gaussian function, where
448: 		all five parameters depend linearly on $\mpm^2$.
449: 		Misreconstructed events are parameterized by a broad
450: 		single Gaussian function.
451: 		
452: 	Both $\mes$ and $\deprime$ PDFs are parameterized by non-parametric
453: 		functions for all $B$-background classes.
454: 
455: 	The $\mes$ and $\deprime$ PDFs for continuum events are
456: 		parameterized with an Argus shape function~\cite{PDFsArgus} and 
457: 		a second order polynomial, respectively, with parameters 
458: 		determined by the fit.
459: 
460: 	We use non-parametric functions to empirically describe the distributions 
461: 		of the NN outputs
462: 		found in the MC simulation for TM and SCF signal events, 
463: 		and for \B-background events. We distinguish tagging categories 
464: 		for TM signal events to account for differences observed in the 
465: 		shapes.
466: 	
467: 	The continuum NN distribution is parameterized by a 
468: 		third order polynomial that is defined to be positive. 
469: 		The coefficients of the polynomial are determined by the fit.
470: 		Continuum events exhibit a correlation between the Dalitz plot 
471: 		coordinate
472: 		and the shape of the event that is exploited in the NN. The tight 
473: 		requirement that eliminates the center of the Dalitz plot has the 
474: 		purpose to reduce such correlation. To correct for residual effects,
475: 		we introduce a linear dependence of the polynomial coefficients
476: 		on the distance of the Dalitz plot coordinate to the kinematic 
477: 		boundaries of the Dalitz plot. The parameters describing this
478: 		dependence are determined by the fit.
479: 
480: