Arxiv:0801.1531V4 [Nucl-Ex] 23 Sep 2008 ∗ Infra6-E Ipms S4 Is Mass WIMP 60-Gev Sec- a Best Cross for Current WIMP-Nucleon Tion the Spin-Independent the [1]

Home , ArDM, WIMP Argon Programme

arXiv:0801.1531v4 [nucl-ex] 23 Sep 2008 ∗ infra6-e IPms s4 is mass WIMP 60-GeV sec- a best cross for current WIMP-nucleon tion The spin-independent the [1]. for well candidate limit a matter (WIMPs), detecting dark Particles for motivated Massive materials as Interacting liquids Weakly noble using periments etlmtfo ol iuddtco s8 is detector liquid current noble The a from application. ease, limit this relative best for with promise masses great large hold to they scalable impurities, scintil- likely radioactive high are of have purified and liquids easily experi- are noble yields, (CDMS) many lation Search Because Matter [2]. Dark ment Cryogenic the by cnilto ih ih itnus lcrncadnuclear and electronic 7]. distinguish might [6, light target scintillation the as argon and dual- liquid employ (WARP) use that experiments Programme detectors (ArDM) phase Argon Matter WIMP Dark lower an Argon require- The the purification is simpler to argon its ments. and due Liquid argon xenon natural of liquid light. cost to to alternative produce charge recoils attractive of electronic ratios and event different recoils between nuclear discriminate as ion- to classes, dual- and collection light are scintillation charge These both ization 5]. use that [4, detectors xenon light phase liquid scintillation from both ionization collect and to ex- designed ZEPLIN are and XENON periments form The that events backgrounds. recoil primary electronic the and signal consti- WIMP that a events tute recoil nuclear between discriminating detec- liquid these Noble on improve likely [3]. will limits. masses XENON10 target by larger with set tors GeV, 100 at lcrncades [email protected] address: Electronic eetyashv ena nraei h ubro ex- of number the in increase an seen have years Recent lentvl,asnl-hs eetrcletn solely collecting detector single-phase a Alternatively, is detectors matter dark liquid noble the to key The cnilto iedpnec n us hp discriminat shape pulse and dependence time Scintillation lcrnceuvln eoleeg kVe,w esr t measure we (keVee), energy recoil electronic-equivalent ehd,w esr akrudadsaitc-iie le an statistics-limited keVe and 7 electronic background 5 be a between to measure down distinguish we argon methods, to liquid discrimination in shape recoils nuclear and electronic ehdo us hp iciiainbsdo h esrds measured 61.25.Bi,29.40.Mc,95.35.+d the numbers: PACS on Fina based keVr. 62 discrimination above shape events pulse recoil-like of nuclear method no with 50% of sn igepaelqi ro eetrwt inlyie signal a with detector argon liquid single-phase a Using .INTRODUCTION I. . 6 × 10 − .Kearns, E. 7 ewe 2ad10kVo ula eoleeg kV)franu a for (keVr) energy recoil nuclear of keV 110 and 52 between 2 ainlIsiueo tnad n ehooy ole,C Boulder, Technology, and Standards of Institute National 1 .H Lippincott, H. W. 3 eateto hsc,Yl nvriy e ae,C 06511 CT Haven, New University, Yale Physics, of Department eateto hsc,Bso nvriy otn A02215 MA Boston, University, Boston Physics, of Department 4 o lmsNtoa aoaoy o lms M87545 NM Alamos, Los Laboratory, National Alamos Los 3 .N McKinsey, N. D. . 6 × 10 . 8 × − 44 Dtd coe 4 2018) 24, October (Dated: 10 1 cm .J Coakley, J. K. − 44 2 set , cm 1, 2 ∗ .A Nikkel, A. J. ogrtiltlftm nlqi ensol lo supe- allow should much neon the liquid that in out lifetime pointed triplet [14] longer Coakley and McKinsey oeto h eto 1] S a lobe used been also has dipole PSD electric [11]. permanent suppress neutron the to the for of search moment the in events ilto ih nlqi eo n iudagni signifi- is as such argon ionizers liquid heavy and for different xenon cantly liquid in light 1977, tillation In (PSD). discrimination Kubota shape pulse using recoils AL :Lftmso h ige n rpe ttsfrneon 10]. for [9, states excimers triplet and xenon singlet and the argon, of Lifetimes I: TABLE iud.I iudhlu,PDhsbe tde nor- in studied from events been recoil has electronic PSD separate helium, noble to several liquid der with In use for studied liquids. been has light tillation odtriewihtp fectto curdfragiven a used for be occurred event. excitation can of components type slow which relative determine and the to fast Therefore, the causing of recoils, dependences. amplitudes nuclear time by observed produced den- the those excitation high as for believed such stronger sities is are it reactions exchange; by these ion- that spin intensity Penning electron-triplet in and primarily suppressed ization interactions, be emitted triplet light can sin- destructive scintillation molecules either slow triplet in The by different exist very I). have can (Table which that states, lifetimes molecular excimers produced triplet or is of glet gases decay noble the liquid by in scintillation because is sinfamnswe oprdt ih oiessuch ionizers light to compared as when fragments fission us hp iciiainbsdo h iigo scin- of timing the on based discrimination shape Pulse 2 β .Gastler, D. ea n opo-ctee lcrn 8.This [8]. electrons Compton-scattered and decay esitlaintm eedneo both of dependence time scintillation he Ne e4 Xe r7 Ar e feetoi eolcnaiainto contamination recoil electronic of vel ula eol.Uigoeo these of one Using recoils. nuclear d tal. et l,w eeo aiu likelihood maximum a develop we lly, .W eeo w ehd fpulse of methods two develop We e. do .5poolcrn e e of keV per photoelectrons 4.85 of ld 1 itlaintm dependence. time cintillation ige ieie(s rpe ieie(ns) Lifetime Triplet (ns) Lifetime Singlet n .C Stonehill C. L. and γ rybcgonsi iudxnn[2 13]. [12, xenon liquid in backgrounds -ray hwdta h iedpnec fscin- of dependence time the that showed < 3 18 . . 0 3 .Hime, A. . ± ± 2 ± 1 0 . . 80305 O 1600 0 22 6 0 la eolacceptance recoil clear . o nlqi argon liquid in ion 14900 2 4 4 . 0 ± ± ± α 100 2 300 . atce and particles 0 3 He( n,p ) 3 H , 2 rior PSD, which has recently been verified experimentally [10]. Following this observation, Boulay and Hime recognized that the similar properties of liquid argon could in principle achieve PSD with part per billion lev- els of electronic recoil contamination (ERC) [15]. ERC is defined to be the probability of incorrectly classifying an electronic recoil event as a nuclear recoil event given a particular level of nuclear recoil acceptance. An ERC 8 of 10− or better is required to perform a competitive WIMP search using liquid argon due to the presence of the radioactive isotope 39Ar, which produces about 1 Bq per kg of atmospheric argon [16, 17, 18]. The WARP collaboration has used scintillation timing in combina- tion with an ionization signal to reduce electronic recoil backgrounds in liquid argon [19]. The Dark Mat- ter Experiment using Argon Pulse Shape Discrimination (DEAP) has demonstrated a background limited ERC FIG. 1: (Color online) Schematic representation of the scin- 6 of 5 × 10− using the DEAP-0 single phase detector for tillation cell. nuclear recoil energies above 1 MeV [20]. At a given energy, ERC improves exponentially with scintillation light collection efficiency. For this reason, efficient scintillation the PMTs. Both windows are coated with (0.20 ± 0.01) light detection is the primary requirement for performing mg/cm2 of TPB, while the Teflon cylinder is coated with a sensitive WIMP search with negligible background at (0.30±0.01) mg/cm2. The Teflon cylinder, windows, and a suitably low energy threshold. PMTs are all immersed directly in liquid argon and con- In this paper we describe measurements of scintillation tained within a 25-cm-diameter by 91-cm-tall stainless in liquid argon due to low-energy nuclear and electronic steel vessel. recoils in the energy range relevant to a WIMP dark mat- The stainless steel vessel is in turn housed inside a vac- ter search. We measure the scintillation time dependence uum Dewar, and argon gas is introduced into the system of liquid argon for both event classes. We develop two though a tube on the top of the Dewar. The argon is liq- basic PSD methods and measure the level of discrimina- uefied in a copper cell mounted to the end of a pulse-tube tion in our apparatus. Finally, we use the measured time refrigerator [25] inside the Dewar before flowing through dependence to develop a maximum likelihood method of a tube to the stainless steel vessel. All components that PSD. come into contact with the gas or liquid are baked to at least 60◦C, and the ultra-high-purity argon gas (99.999%) is passed through a heated gas-purification getter [26] be- II. EXPERIMENTAL DETAILS fore entering the vessel. In addition, the argon is continually circulated through the getter and reliquefied at a A. Detector Design rate of at least 2.0 standard liters per minute (slpm) to ensure that high purity is maintained. The stability of The apparatus consists of a 3.14-liter active volume of the system is discussed further in Sec. II C. liquid argon viewed by two 200-mm-diameter photomul- The data acquisition system is custom-built around tiplier tubes (PMTs) [21], all contained within a stain- VME-bus waveform digitizers (WFDs); a sample WFD less steel vessel and vacuum Dewar. Figure 1 shows a trace from a scintillation event in argon can be seen in schematic of the central volume and PMTs. The active Fig. 2. The PMT signals from the detector are divided region is defined by a Teflon cylinder 200 mm in diameter three ways by a linear fan out with two copies of the and 100 mm in height with two 3-mm-thick fused-silica signal sent to the WFDs and one sent to a triggering windows enclosing the top and bottom. The PMTs are system. Each WFD has four channels that record eight- held in place by Teflon rings above and below the central bit samples at 500 MHz. These samples are stored in volume and view the active region through the windows. a separate programmable-length memory buffer for each They are powered by positive high voltage with a typical channel. For all data presented here, the record length gain of approximately 4 × 107. is set to 26 µs. The two copies of the PMT waveforms Because liquid argon scintillates in the ultraviolet are recorded separately at unity gain and at an atten- (≈ 128 nm) [22], all inner surfaces of the Teflon and uation of ten to increase the effective dynamic range of windows are coated with a thin film of tetraphenyl bu- the eight-bit digitization. The buffer is continually filled tadiene (TPB) [23] by use of a Tectra Mini-Coater evap- but saved to disk only when the triggering system regis- oration system [24][44]. The TPB shifts the wavelength ters a fraction of a photoelectron in both PMTs within of the ultraviolet light to approximately 440 nm so that a 100-ns coincidence window. Once a trigger has been it may pass through the windows and be detected by registered, the DAQ records the event for 22 µs, leaving 3 an additional 4 µs of baseline presamples in the data (see where ST and SB are the signal areas in the top and Fig. 2). The data are read to a computer via fiber-optic bottom PMTs. For most data, we require −0.275

measure the signal yield of the detector to be 4.85 photoelectrons per keV electron equivalent (keVee), where keVee refers to the amount of energy deposited by an electronic recoil. The response of the detector was stable to within 5% during the four months of data acquisition.

600 Counts 500 Cobalt data

400 Simulation

300

200

100 FIG. 3: (Color online) Schematic of the neutron scattering setup. 00 100 200 300 400 500 600 700 800 Number of photoelectrons

57 We also collect background data by looking at events FIG. 4: (Color online) Example spectrum of Co data, along in the liquid argon with no external source present. This with a simulation done using the Reactor Analysis Tool [30]. background data provides an estimate of the accidental 57 background rate that may be contaminating the γ-ray We use the 122-keV Co peak to provide a daily en- data sets, and it will be discussed further in Sec. III B ergy calibration. To check the quality of that calibration, we use the 511-keV γ rays produced by the 22Na source as a second point of reference. When calibrated using the 57Co source and the assumption that the signal scales lin- C. Detector Calibration early with deposited energy, the 511-keV absorption line appears in the 22Na spectrum as expected to within 1%. We use a 10-µCi sealed 57Co source for daily calibra- In addition, simulations of 511-keV γ rays are consistent tions. This source produces 122-, 137-, and 14.4-keV γ with the data. rays, with branching ratios of 86%, 11% and 9%, respec- Impurities in the detector can build up over time via tively. Any scintillation event in liquid argon produces outgassing. These impurities can quench argon excimers a significant triplet component; since this component is or absorb emitted UV photons, which would lead to a de- spread out over many µs, it appears in the signal as many crease in light yield. Additionally, one would observe a single photoelectrons well separated in time (for exam- decrease in the triplet molecule lifetime. Work by Himi et ple, there are a number of single photoelectrons that ap- al. suggests that an impurity level of 0.5 atoms of ni- pear after 7 µs in Fig. 2). Therefore, we measure the trogen per 106 atoms of argon in the liquid could de- gain of the PMTs using the 57Co source by selecting sin- crease the observed triplet lifetime by as much as 0.1 gle photoelectrons from the tail end of each pulse. The µs [31]. Further experiments quantifying the reduction of PMT traces are divided into 75-ns regions centered on the triplet lifetime due to nitrogen and oxygen impurites times at which the trace crossed an experimentally de- have recently been performed by the WARP collabora- termined threshold of roughly 1/3 of a photoelectron. tion [32, 33]. To avoid signal degradation, we continually These regions are then integrated to obtain the single circulate the argon through a getter before reliquefying photoelectron pulse area. The typical gain for the PMTs back into the detector. We use daily measurements of is approximately 4 × 107. both the light yield and the triplet lifetime to monitor Figure 4 shows an example 57Co spectrum along with the purity level. a simulation done with the Reactor Analysis Tool, a We measure the light yield in the manner described toolkit of Geant4 developed by the Braidwood collab- earlier in this section, and we use the same 57Co data to oration [29, 30]. From the simulation, we find that the measure the triplet lifetime. First, we select events in the position of the primary peak is dominated by the 122- 122-keV peak to make sure we use a similar data set for keV γ photoabsorption process. The simulation parame- each individual measurement. The top and bottom PMT ters describing absorption and reflection of the materials traces are normalized by the size of the single photoelec- in the detector are tuned to match the observed signal tron and summed together. We align each pulse based on yield, and the spectral shapes line up nicely. its estimated trigger time, defined as the time at which By comparing the integrated signal corresponding to the trace first crosses 20% of its maximum value. At this the 122-keV peak to that of a single photoelectron, we trigger time, the relative time for each pulse is t = 0. 5

Between 5000 and 10000 traces are averaged and the fol- III. EXPERIMENTAL RESULTS lowing model is ﬁt to the average trace between 1 and 7 µs from the trigger: A. Detection Time PDF Model

We measure the time dependence of scintillation light < V (t) >= A exp(−t/τl)+ B, (3) produced by electronic and nuclear recoil scattering events in our detector. For an event with energy deposited at time t0, we model the temporal probability where < V (t) > is the expected trace, A is a normaliza- density functions (PDFs) for the emission times of scin- tion factor, τl is the triplet lifetime and B is an additional tillation photons as the weighted sum or mixture of two baseline term that helps stabilize the fit over a range of exponential PDFs: fit windows. An example fit is shown in Fig. 5. f(t − t0) = q g(t − t0, τl)+(1 − q) g(t − t0, τs), (4)

where 1 −t g(t, τ)= exp , (5) τ τ

τl and τs are the long and short time constants for both nuclear and electronic recoil event classes at any given energy, and the probability parameter q takes diﬀerent values for the two classes of events. At each PMT, detected scintillation photons yield photoelectrons that produce observed voltage traces. For the small detector in this experiment, scintillation tran- sit times are negligible. We assume that the duration of energy deposition and excimer formation is instantaneous compared to the time scales relevant to scintillation light emission. Hence, we model the expected voltage trace FIG. 5: (Color online) An example ﬁt of a single exponential as a convolution of the impulse response function of the to an average trace to measure the long time constant in liquid PMT and the PDF model for the emission of scintillation argon. photons:

∞ < V (t) >∝ hV (t − s)f(s − t )ds, (6) ∗ We find two significant sources of systematic error in Zs=t∗ the fit parameters stemming from the voltage applied to the PMTs and the choice of fit window. Although we where hV (t) is the impulse response function of the PMT and t is an additional model parameter that relates the do not fully understand the PMT voltage effect, we esti- ∗ mate the systematic uncertainty to be 33 ns by changing energy deposit time t0 to the relative time scale we as- the PMT voltages by ±75 V (equivalent to dividing or sociate with our measurement of hV (t). We estimate the multiplying the gain by 2). We do not fit the data before impulse response function by averaging single photoelec- 1 µs or after 9 µs, as the fit becomes less reliable due tron events observed in our calibration data. Neglecting to contamination from the fast component and baseline additive noise and other instrumental systematic errors, noise, respectively. We estimate the systematic uncer- the integral of the voltage trace is proportional to the tainty associated with the fit window to be about 35 ns number of photoelectrons produced by the event. by both varying the end time by ±2 µs and by fitting the The data are divided into 15 non-uniform bins by pho- data within 5 µs windows ending at 6, 7 and 8 µs from toelectron number, with the smallest signal bin includ- the trigger. We combine the two sources of error into a ing events consisting of 20–24 photoelectrons and the single systematic uncertainty estimate of 50 ns. largest bin including events consisting of 240–279 photoelectrons. These bins define the region of interest. During the four months of operation, we mea- For each photoelectron bin, we generate template traces sure the long time constant and signal yield to be for both electronic and nuclear recoils by averaging all (1463 ± 5stat ± 50sys) ns and (4.85 ± 0.08) photo- “tagged” events of a given type. The traces are aver- electrons/keVee. The uncertainty on the signal yield is aged in the same manner as described in Sec. IIC and statistical, and varying the PMT voltage by ±75 V has normalized. no apparent effect. For a fixed fit window and PMT volt- We use these average voltage traces to determine the age, the long time constant is stable to within 1%, while model parameters: τl, τs, qnuclear and qelectronic. The ob- the signal yield is stable to within 5%. served and predicted fraction of a normalized trace in the 6 ith time bin are called pm(i) andp ˆ(i), respectively. We 10-1 obtain the model parameters by minimizing the squared Matusita distance [34, 35, 36] between pm and p: Nuclear recoil data 10-2 Nuclear recoil model prediction 2 |p − pm|M = ( pm(i) − pˆ(i)) , (7) Xi p p 10-3 where negative values of pm are set to 0. For each bin, we

fit our model to normalized mean voltage trace data in Normalized voltage trace -4 a time window that ends about 6800 ns after the trigger 10 time. We determine t to be 30 ns before the trigger time, ∗ chosen to minimize the value of Eq. 7. Table II shows the 10-5 estimated model parameters for each photoelectron bin. 102 103 Relative time (ns) Figure 6 shows the model prediction along with data for the 80–99 photoelectron bin. 10-1 Between 80 and 300 ns, there is a feature in both event classes that is not well predicted by the model. A scintil- Electronic recoil data 1 -2 Electronic recoil model prediction lation component that decays as approximately t− has 10 been observed in liquid helium, attributed to diffusion- dominated excimer-excimer destruction [11]. When we 10-3 fit a model including the full form of this component as described in [11], the parameter estimates are not well Normalized voltage trace determined. Previous observations of scintillation in ar- 10-4 gon have observed an intermediate exponential component with a decay time of 20–40 ns [9]. Unfortunately, a model that includes a third exponential component nei- 10-5 102 103 ther returns stable model parameters as a function of en- Relative time (ns) ergy nor accurately predicts the trace behavior between 80 and 300 ns, so we prefer the two-component model. FIG. 6: (Color online) Observed and predicted mean volt- We assume that the sharp bump localized at 150 ns is age traces for nuclear and electronic recoil events of 80 to 99 caused by the cabling and electronics. photoelectrons.

Bin (pe) τl (ns) τs (ns) qnuclear qelectronic 20–24 1634 ± 150 9 ± 3 0.378 ± 0.011 0.523 ± 0.015 voltage applied to the PMT, and variation in the re- 25–29 1535 ± 128 10 ± 3 0.382 ± 0.011 0.573 ± 0.015 sponse of each individual PMT is difficult. Nonetheless, 30–34 1478 ± 107 10 ± 3 0.357 ± 0.010 0.601 ± 0.014 we get approximate estimates of systematic uncertain- 35–39 1455 ± 102 11 ± 3 0.353 ± 0.010 0.627 ± 0.014 ties from these sources, and these systematics are the 40–49 1461 ± 96 12 ± 3 0.344 ± 0.010 0.658 ± 0.014 dominant source of error. We estimate 1-sigma systematic uncertainties in the parameters τ (for i = l,s) as 50–59 1459 ± 92 12 ± 3 0.327 ± 0.009 0.681 ± 0.015 i 0.5 × (τ − τ ), where τ and τ are the ± ± . ± . . ± . i,max i,min i,max i,min 60–69 1439 89 12 3 0 315 0 010 0 699 0 015 maximum and minimum values taken by τ for different ± ± ± ± i 70–79 1448 89 13 3 0.309 0.010 0.710 0.015 choices of time windows. We refit the model to data using 80–99 1447 ± 85 13 ± 3 0.298 ± 0.010 0.721 ± 0.015 time windows of approximately 5000 ns and 8600 ns from 100–119 1452 ± 84 13 ± 3 0.289 ± 0.010 0.733 ± 0.015 the trigger time, and vary the offset parameter t by ± ∗ 120–139 1447 ± 84 13 ± 3 0.284 ± 0.011 0.741 ± 0.016 2 ns. We also check for differences between the response 140–159 1446 ± 84 14 ± 3 0.278 ± 0.012 0.747 ± 0.016 of each PMT in five different photoelectron bins. We in- 160–199 1450 ± 84 14 ± 3 0.272 ± 0.013 0.752 ± 0.016 clude the voltage-dependent errors described in Sec. IIC 200–239 1460 ± 84 15 ± 3 0.265 ± 0.015 0.760 ± 0.016 in the systematic uncertainty of τl, and we estimate the voltage-dependent systematic variation of the q-values. 240–279 1467 ± 84 15 ± 3 0.258 ± 0.018 0.764 ± 0.016 We find that the time constants are most influenced by choice of fit window, while the q-values are most influ- TABLE II: Estimated model parameters and 1-sigma uncertainties for each photoelectron bin. The systematic errors enced by PMT effects. We combine all sources of error described in the text are the dominant source of error. to obtain the estimates of systematic uncertainty shown in Table II. We determine 1-sigma random uncertainties for the As a consistency check on the quality of fit, for each model parameters with a nonparametric bootstrap re- photoelectron bin we use our model to predict the prompt sampling scheme [37]. A rigorous quantification of sys- fraction, a discrimination statistic described in detail in tematic uncertainty arising from the fit window, the the next section. We compute the difference between 7 the predicted and observed prompt fractions, and the mean fp for both classes of events in the energy range root-mean-square values of this prediction error across of interest. We estimate the systematic uncertainty on all bins are respectively 0.007 and 0.003 for the nuclear the values in Table III to be 3%. This uncertainty esti- and electronic recoil event classes. The fractional root- mate comes from changing the PMT voltage by −75 V mean-square value of this prediction error is about 1% and from variations in the measured signals between the for both event classes. two PMTs. The mean values of the fp distributions for the two event classes are closer at low energies than at high energies, possibly because dE/dx for nuclear recoils B. Prompt Fraction Method decreases at low energies while increasing for electronic recoils. Therefore, the PSD improves at higher energies The prompt fraction method is a simple approach to both because of increased photoelectron statistics and be- pulse shape discrimination. For each trace, we define the cause of increased separation between the mean fp val- prompt fraction fp as ues. ξ V (t)dt ˆ ˆ Ti Energy (keVee) fp,electronic fp,nuclear fp = T , (8) R f V (t)dt 5–6 0.391 ± 0.012 0.566 ± 0.018 Ti R 6–7 0.376 ± 0.011 0.595 ± 0.018 where V (t) is the voltage trace from the PMT, ξ is 7–8 0.361 ± 0.011 0.607 ± 0.019 an integration time determined to optimize the ERC, 8–9 0.349 ± 0.011 0.625 ± 0.019 T = t − 50 ns, T = t +9 µs, and t is the trigger i 0 f 0 0 ± ± time as defined in Sec. II B. The measured discrimina- 9–10 0.339 0.010 0.638 0.020 . ± . . ± . tion does not significantly improve by extending Tf to 10–11 0 334 0 010 0 640 0 020 ± ± 20 µs. Fig. 7 shows a scatter plot of fp versus energy 11–12 0.328 0.010 0.649 0.020 for both electronic and nuclear recoils. The two popula- 12–13 0.322 ± 0.010 0.663 ± 0.020 tions of events represent the tagged data remaining in the 13–14 0.319 ± 0.010 0.658 ± 0.020 22 neutron generator data set and the Na data set after 14–15 0.314 ± 0.009 0.675 ± 0.020 the selection cuts described in Sec. II B have been made. 15–16 0.311 ± 0.009 0.683 ± 0.021 We choose ξ = 90 ns by estimating the ERC based on a 16–17 0.309 ± 0.009 0.678 ± 0.021 simple Gaussian model for values of ξ from 50 to 250 ns ± ± over a variety of different photoelectron bins. Although 17–18 0.304 0.009 0.685 0.021 ± ± the value of ξ has only a weak effect on the predicted 18–19 0.302 0.009 0.682 0.021 discrimination, a choice of ξ = 90 ns provides the best 19–20 0.299 ± 0.009 0.684 ± 0.021 results across the widest range of energies in the region of 20–21 0.297 ± 0.009 0.690 ± 0.021 interest, and that value is used for the analysis presented 21–22 0.295 ± 0.009 0.695 ± 0.021 in this paper. 22–23 0.292 ± 0.009 0.699 ± 0.021 23–24 0.290 ± 0.009 0.690 ± 0.021 24–25 0.288 ± 0.009 0.688 ± 0.021 25–26 0.289 ± 0.009 0.695 ± 0.021 26–27 0.288 ± 0.009 0.696 ± 0.021 27–28 0.285 ± 0.009 0.696 ± 0.021 28–29 0.284 ± 0.009 0.701 ± 0.021 29–30 0.283 ± 0.009 0.708 ± 0.021 30–31 0.281 ± 0.009 0.701 ± 0.021 31–32 0.282 ± 0.009 0.689 ± 0.021

TABLE III: This table presents estimated mean values of fp versus energy, where ξ = 90 ns. The main sources of uncertainty are systematic, stemming from voltage eﬀects and diﬀerences in the measured signals between the two PMTs.

FIG. 7: (Color online) A scatter plot of fp vs. energy for tagged electronic and nuclear recoils, where ξ = 90 ns. We estimate the ERC as the number of tagged electronic recoil events with fp > fˆp,nuclear divided by the For energy bins of width 1 keVee between 5 and 32 total number of electronic recoil events, where fˆp,nuclear is keVee, we form histograms of the electronic and nuclear the estimated mean fp for nuclear recoils of that energy. recoil fp statistics. To estimate the expected value of fp, This restriction sets a nuclear recoil acceptance level of fˆp, we ﬁt a Gaussian function to the empirical distribu- approximately 50%. Since the shielding, coincidence, and tions. In Table III and Fig. 8, we present the estimated timing cuts do not eliminate all neutron backgrounds in 8

1 0.8 Data analyzed with prompt fraction method 10-1 Data analyzed with multi-bin method Estimated neutron background 0.7 10-2 Statistical model (with fitted parameters) Statistical model (additional noise set to 0) 10-3 0.6 Nuclear recoils 10-4 Prompt fraction Electronic recoils -5 0.5 10 (errors are systematic) 10-6 0.4 10-7 10-8 0.3 Electronic Recoil Contamination 10-9 -10 0.2 10 5 10 15 20 25 30 20 40 60 80 100 Energy (keVee) Energy (keVr)

FIG. 8: (Color online) The estimated mean fp versus energy FIG. 9: (Color online) Measured electronic recoil contami- for both event classes where ξ = 90 ns. nation obtained by using the prompt fraction method and the multi-bin method. Also shown: the background estimation and model predictions described in the text. We include the detector, a background estimation Nbg is made by a model with the additional noise set to 0 for comparison. There were no contamination events above 69 keVr observed measuring the rate of background neutrons Rbgn and assuming that the background is dominated by these neu- by use of either method. The energy axis has been scaled from trons hitting in accidental coincidence with γ rays (R ) keVee to keVr by use of a constant nuclear recoil scintillation γ efficiency of 0.29 as discussed in the text. in the liquid scintillator during the time allowed by the time-of-flight cut: 2 dom variables with means µp and µl and variances σ Nbg = Rbgn × Rγ × TOF × Ta, (9) p 2 and σl . The estimated total number of photoelectrons, where Ta is the acquisition time of the data. To enable Ntot = Np + Nl, is also a random variable, with mean 2 2 2 comparison with the measured ERC, we divide Nbg for µtot = µp + µl and variance σtot = σp + σl . We express each energy bin by the total number of electronic recoil µp and µl in terms of µtot and fˆp and we decompose the events in that bin. variances into two components: Figure 9 shows the ERC observed using the prompt fraction method. We also plot the background estima- µp = fˆpµtot tion, two PSD projections based on the statistical model ˆ described below, and the ERC observed by applying a µl = (1 − fp)µtot 2 2 multibin method of PSD described in the next section. σp = µp + σp,add We convert the energy axis in Fig. 9 to keV of nuclear σ2 = µ + σ2 (10) recoil energy (keVr) from keVee by dividing all electron l l l,add equivalent energies by a constant nuclear recoil scintilla- where σp,add and σl,add represent additional sources of tion efficiency of 0.29. This value was obtained by mea- random variability beyond what we expect from Poisson surements using the same apparatus described in this pa- counting statistics (for example, integration noise). per and will be discussed in an upcoming publication [38]. Hinkley [41] has described in detail the probability den- We present the PSD results in keVr because that is the sity function of the ratio of two normally distributed, cor- unit of interest for a dark matter detector. Using the related random variables. For simplicity, we present here prompt fraction method, for a nuclear recoil acceptance an approximation [given by Eq. 9 of Ref. [41]] to the PDF level of approximately 50%, we measure a background- of fp = Np/Ntot: and statistics-limited level of ERC in our detector of 6 2 2 8.5 × 10− between 52 and 110 keVr (11 contamination σl µpx+σpµl(1 x) g (x)= 2 − × events). We observe no nuclear recoil-like events above fp √2π(σ x2+σ2(1 x)2)3/2 l p − 2 69 keVr. For comparison, there is an uncorrelated neu- (µlx µp(1 x) ) 2 − − exp[− 2(σ x2+σ2(1 x)2) ], (11) tron background rate of ∼ 6 mHz between 69 and 110 l p − keVr, corresponding to 0.25 expected background counts in that energy range. where we have used the fact that the correlation, ρ, be- Following work done by members of the DEAP collab- tween Np and Ntot is oration [39, 40], we model our estimates of the number σp of photoelectrons in the prompt and late time windows, ρ = . (12) 2 2 Np and Nl, as normally distributed, independent ran- σp + σl q 9

For the analysis below, we use the exact PDF [Eq. 1 ture studies will require more work to better calibrate the in Ref. [41]], although in practice the approximation is ERC predictions made by the ratio-of-Gaussians model. extremely good down to the lowest energy bin examined. We fit the electronic recoil data with the statistical 4 model in each energy bin by fixing Ntot according to our 10 measured light yield and treating fˆp, σp,add, and σl,add as Counts free parameters. We assume that the statistical distribu- 103 tion of fp does not strongly depend on the value of Ntot for events of the same energy. For any particular energy 102 deposit, Ntot is a random variable, and events due to many different energy deposits can contribute to prompt 10 ratio data in any one bin in Ntot space. The probability density function for prompt ratio data is a mixture of energy dependent PDFs, and we neglect this energy 1 blurring effect. Simple Monte Carlo studies suggest that 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 the ratio-of-Gaussians model breaks down for idealized Prompt fraction Gaussian data due to the constraint on Ntot resulting from the binning of data, and this effect has not been taken into account in our analysis. 104

We use the statistical model to estimate the expected Counts fraction of electronic recoils that are misclassiﬁed as nu- 103 clear recoils: 102 1

ERC = gfp (x)dx. (13) Zη 10

Here, we choose η = fˆp,nuclear to set the nuclear recoil 1 acceptance level to approximately 50%, and we choose the parameters of gf according to the fits to the elec- 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 p Prompt fraction tronic recoil fp distribution. Fig. 9 shows the predicted ERC versus energy according to this model. We also plot the idealized case where σl,add and σp,add are set to 0. In FIG. 10: (Color online) Projections of the electronic recoil general, we expect that the normal distribution model for data from Fig. 7 onto the y-axis for 14–15 keVee (top) and 30–31 keVee events (bottom), fitted by Eq. 11 as discussed in Np and Nl is an approximation for any energy of interest. the text. For the idealized case where σp,add = σl,add = 0, Np and Nl would be Poisson random variables rather than Gaus- sian random variables. Hence, the ERC predicted by the Figure 11 shows the fitted values of the variances, 2 2 2 2 ratio-of-Gaussians model for the idealized case should be σl = µl + σl,add and σp = µp + σp,add versus the es- interpreted with caution particularly at lower energies timated number of photoelectrons in µl and µp. The where the accuracy of a normal distribution model for a ideal case of zero additional noise (σ2 = µ) has been Poisson random variable can be very poor. shown for comparison. The relationship between the es- Figure 10 shows an example of the model fit for 14–15 timated variance parameters and the corresponding esti- keVee and 30–31 keVee electronic recoil events. There mates of the number of photoelectrons is approximately 2 2 is a deviation from the model at low fp values that we linear, suggesting that σl,add and σp,add are proportional attribute to pile-up and noise triggers. There is also an to µl and µp. Straight line fits to the late and prompt excess of events in the high fp region. This excess might variances have slopes of 2.2 and 1.3, respectively. Due to be caused by edge effects in our detector coming from γ the various uncertainties in the fit such as that associ- tracks going into the walls or the TPB layer, producing ated with data binning, these numbers do not represent extra prompt light. A larger detector with position re- a rigorous estimate of the amount of noise in the late and construction capability might be able to eliminate such prompt distributions; however, they might be used as a edge effects. There could be some unknown phenomena point of comparison with future detectors in trying to re- at work in the production and decay of argon molecular duce the overall noise in the system. Possible sources of states, yielding a small fraction of events with anoma- this additional noise include integration noise, the widths lously large fp values. A third possibility is approxima- of the single photoelectron spectra, and variability in the tion error stemming from incorrect model assumptions, prompt window size arising from uncertainty in the de- such as the Gaussianity of Np and Nl or the effect of data termined trigger position. To improve PSD, future ex- binning. These possible effects can be better investigated periments will need to reduce the size of this additional by detectors with improved neutron shielding, and fu- noise to approach the ideal case. 10

statistic based on analysis of an idealized experiment in 350 which we could observe the absolute detection time of Late component variance, σ2 300 l each photoelectron without error. In this ideal case, for Prompt component variance, σ2 p a ﬁxed number of detected photoelectrons, the observed σ2 µ 250 Ideal case, = number of photoelectrons in the time bins is a multino- 200 mial random variable.

Estimated variance Given that the fraction of detected photoelectrons in 150 the kth bin is pm(k,l), the multinomial log-likelihood statistics for the nuclear and electronic recoil classes are 100 K L 50 ln Yn = Ntot δll′ pm(k,l) ln pn(k,l′)+const (14) 0 Xk=1 Xl=1 0 20 40 60 80 100 120 140 Estimated number of photoelectrons and

2 2 K L FIG. 11: (Color online) Estimated σl and σp parameters from the statistical model plotted against the estimated number of ln Ye = Ntot δll′ pm(k,l) ln pe(k,l′) + const, (15) photoelectrons in the late and prompt components, respec- kX=1 Xl=1 tively. Each distribution is fitted by a straight line, and the 2 ideal case where σ = µ is also shown. respectively [42]. Here, Ntot is the total number of detected photoelectrons, and pn(k,l) and pe(k,l) are the expected values of the fraction of detected photoelectrons C. Multibin Method in the kth time bin for the nuclear and electronic recoil classes, respectively. For this idealized case, a natural choice for the discrimination statistic is the log-likelihood The prompt fraction method of PSD is based on bin- ratio statistic, ln R : ning the voltage trace into two time bins. We generalize m this approach by representing a normalized voltage trace as a K ×L dimensional matrix. In this representation, we ln Rm = ln Ye − ln Yn. (16) categorize the data by the number of photoelectrons in the event, and L refers to this division of data into photo- In our experiment, the observed data is not multino- electron bins. We also partition the voltage trace into K mial, primarily because we observe a noisy voltage wave- time bins. We choose K = 10 for ease of computation, form rather than discrete detection times. Nonetheless, but there may be a better choice of K. With the ex- we compute a discrimination statistic using Eqns. 14–16 ception of the four smallest signal bins, which have been where we estimate Ntot and pm(k,l) from any voltage combined to form bins of 20–29 and 30–39 photoelec- trace of interest, and determine pe(k,l) and pn(k,l) as trons, we use the same average voltage traces obtained described earlier. in Sec. III A as templates. Figure 12 shows a scatter plot of ln Rm versus energy, We partition each template into K = 10 time bins. analogous to Fig. 7. From this point the analysis parallels The upper and lower endpoints of each time bin are se- the prompt fraction method, as we form histograms of lected so that the fractions of the template trace for 80– ln Rm by energy bin and fit a Gaussian function to the 99 photoelectron electronic recoil events that falls in each observed ln Rm statistics to estimate the mean of ln Rm. time bin are approximately equal. The initial bin starts Fig. 13 shows an example of the fitted projections for 50 ns before the trigger, and the endpoints of each bin 14–15 keVee and 30–31 keVee events. are as follows, measured in nanoseconds from the trig- For each energy bin, we estimate the mean values of ger: 8, 18, 56, 200, 440, 750, 1180, 1800, 2950, and 8000. ln Rm for nuclear recoils based on the Gaussian fits to the We do not adjust these endpoints for different photoelec- observed distributions, and this estimated mean value de- tron number but use the 80–99 photoelectron based time termines an approximate 50% nuclear acceptance thresh- binning scheme for all cases. old. We then determine the fraction of events in the The kth component of the normalized template for an tagged electronic recoil data set that have discrimination event that falls in the lth photoelectron bin for the nu- statistics less than this mean value to determine the ob- clear and electronic recoil classes is denoted pn(k,l) or served level of ERC using the multi-bin method. If we pe(k,l), respectively. For instance, pn(1, 1) represents the determine the 50% nuclear acceptance threshold by find- fraction of the 20–29 photoelectron nuclear recoil tem- ing the median values of the nuclear recoil distributions, plate trace that falls in the first time bin (between 50 ns the observed ERC is not significantly affected. Figure 9 before the trigger and 8 ns after the trigger). in Sec. III B shows the ERC using the multibin method. To assign an event to either the nuclear recoil or elec- The multibin method outperforms the prompt fraction tronic recoil class, we first compute a discrimination method by as much as an order of magnitude. For a nu- statistic for the event. We motivate a discrimination clear recoil acceptance of approximately 50%, we have 11

IV. MAXIMUM LIKELIHOOD PSD

We predict the PSD achievable in a large detector with many PMTs by assuming we can measure discrete times for each photoelectron detected in an event. We simulate events in this detector using the PDFs measured in Sec. III A and assume that the detection times of photoelectrons are measured without error. We develop a maximum likelihood PSD method and apply it to the simulated data. For an event that generates N photoelectrons, we de- note the detection times as t = (t1,t2, ..., tN ). We deﬁne the log-likelihood function, ln L, of this data following Ref. [43] as

FIG. 12: (Color online) A scatter plot of ln Rm vs. energy for both electronic and nuclear recoils. N ln L = −m + ln(rbg + λf(t − t0)), (17) Xi=1 where

Tagged nuclear recoils Te 3 10 Tagged electronic recoils m = (rbg + λf(t − t0))dt, (18) Counts ZTb

102 t0 is the time at which the energy deposit occurs, f(t−t0) is the PDF for the observed photoelectrons [Eq. 4], λ is the expected number of detected photoelectrons gener- 10 ated by the event, Tb and Te are the start and end time of the observation, and rbg is the background rate. In our simulations, we set r = 0 and t > T . The 1 bg 0 b maximum likelihood estimates of t0 and λ are therefore ˆ -100 -50 0 50 100 t1, the first detection time, and λ, where Log-likelihood statistic, ln Rm N λˆ = (19) qF (T − t , τ )+(1 − q)F (T − t , τ ) 103 Tagged nuclear recoils e 1 1 e 1 2 Tagged electronic recoils Counts and F (T, τ)=1 − exp(−T/τ). Consequently, m = N. 102 Additionally, the time window in our simulations is 40 µs, Te so f(t − t0)dt ≈ 1. Thus, for our case the likelihood Tb functionR of the observed data is well approximated as 10 N N L(t) ≈ exp(−N) N f(ti − t1). (20) 1 iY=1 Following Ref. [13] and the discussion above, for each -100 -50 0 50 100 t Log-likelihood statistic, ln Rm simulated event we determine L( ) for both the electronic and nuclear recoil event classes as modeled in Sec. III A. t t FIG. 13: (Color online) Projections of Fig. 12 onto the y-axis These values are denoted Le( ) and Ln( ). We define the for 14–15 keVee (top) and 30–31 keVee (bottom) events, with log-likelihood ratio, ln R: Gaussian fits to both the electronic (left) and nuclear (right) recoil distributions. ln R = ln Le(t) − ln Ln(t). (21) Figure 14 shows this log-likelihood statistic for simulated electronic and nuclear recoil events that yield 50 photoelectrons. The event is assigned to the nuclear recoil measured a background- and statistics-limited ERC of class if ln R is less than an adjustable threshold that can 7 7.6 × 10− between 52 and 110 keVr (1 contamination be varied to increase the discrimination against electronic event). We observe no contamination events above 62 recoils at the cost of decreasing nuclear recoil acceptance. keVr using the multibin method. The Monte Carlo estimate of the distribution of ln R has 12 prominent ripples even though the Monte Carlo estimates dence of electronic and nuclear recoils in liquid argon of Ln and Le do not. In our Monte Carlo study, we simulate events that yield a fixed number of photoelectrons ML simulation 10-1 according to PDFs with the assumption of perfect knowl- ML gaussian extrapolation 10-2 edge. Furthermore, we neglect dark current noise. In an Idealized prompt fraction method 10-3 actual experiment, we expect that imperfect energy res- 10-4 olution due to variability in counting statistics and dark 10-5 current effects would attenuate the ripples in the ln R 10-6 distribution. 10-7 10-8 10-9 0.5 10-10

Electronic Recoil Contamination -11 Nuclear recoils Electronic recoils 10 0.4 -12 10 10 15 20 25 30 35 40 45 50 55 0.3 Energy (keVr)

0.2 FIG. 15: (Color online) Predicted performance of maximum likelihood and prompt ratio PSD methods for a detector yield- Relative counts (arb. units) 0.1 ing 6 photoelectrons/keVee (the energy axis has been scaled from keVee to keVr by use of a constant nuclear recoil scintillation eﬃciency of 0.29 as discussed in Sec. III B. -400 -30 -20 -10 0 10 20 30 40 Log-likelihood statistic, ln R

FIG. 14: (Color online) Monte Carlo estimates of the log- down to 5 keVee or 20 keVr. We developed a prompt frac- likelihood statistics for events that yield 50 photoelectrons. tion method of PSD in liquid argon, and for a nuclear recoil acceptance level of 50%, we measured a background- 6 To illustrate the maximum likelihood method, we sim- and statistics-limited level of ERC to be 8.5 × 10− be- ulate a detector with a signal yield of 6 photoelec- tween 52 and 110 keVr with no contamination events trons/keVee, as might be possible in a detector with full above 69 keVr. We also developed a multibin method PMT coverage. We neglect dark current and we set the of PSD, improving on the prompt fraction method by discrimination threshold to accept 50% of nuclear recoils. as much as an order of magnitude. With this method, we measured a background- and statistics-limited level For comparison, we simulate prompt fraction data by as- 7 suming the number of prompt photoelectrons is a bino- of ERC of 7.6 × 10− between 52 and 110 keVr for the mial random variable with an expected value determined same nuclear recoil acceptance of 50%. We modeled the by the measured PDFs, and the total number of photo- observed prompt fraction data as the ratio of two nor- electrons is assumed to be known without error. This ide- mally distributed, correlated random variables, where we alized binomial model predicts a much lower ERC than assumed Np and Nl were uncorrelated; we discussed dis- the statistical model discussed in Sec. III B. crepancies between observed and predicted prompt frac- The maximum likelihood discrimination method out- tion results. Finally, we developed a maximum likelihood performs the idealized prompt fraction method, as shown method of PSD for a detector capable of measuring a dis- in Fig. 15. We also expect that the maximum likelihood crete detection time for each observed photoelectron in method will be more robust than the prompt fraction an event. method to background noise. For the prompt fraction case, we select a threshold in prompt photoelectron space to yield a nuclear recoil acceptance probability as close to 0.5 as possible. Due to quantization eﬀects, the ac- Acknowledgments tual nuclear recoil acceptance probability varies about 0.5, and there are sawtooth-like artifacts in the prompt We thank M. Boulay, C. Jillings, B. Cai, and J. Lidgard fraction ERC curve shown in Fig. 15. for useful discussions and for developing the ratio-of- Gaussians model used in this analysis. We acknowledge S. Seibert and J. Klein for their contributions to the V. CONCLUSION Monte Carlo simulations software package. This work was supported by the David and Lucille Packard Foun- Using a detector with a signal yield of 4.85 photoelec- dation, the Los Alamos Directed Research and Develop- trons/keVee, we measured the scintillation time depen- ment Program, and the U.S. Department of Energy. 13

[1] For a review see G. Jungman, M. Kamionkowski, and and Experiments for SNOLAB (2006). K. Griest, Phys. Rep. 267, 195 (1996). [21] Hamamatsu R5912-MOD02, www.hamamatsu.com [2] Z. Ahmed, D. S. Akerib, S. Arrenberg, M. J. Attisha, [22] O. Cheshnovsky, B. Raz, and J. Jortner, J. Chem. Phys. C. N. Bailey, L. Baudis, D. A. Bauer, J. Beaty, P. L. 57, 4628 (1972). Brink, T. Bruch, et al., arXiv.org:0802.3530v2 [astro-ph] [23] D. N. McKinsey, C. R. Brome, J. S. Butterworth, (2008). R. Golub, K. Habicht, P. R. Huffman, S. K. Lamore- [3] J. Angle, E. Aprile, F. Arneodo, L. Baudis, A. Bern- aux, C. E. H. Mattoni, and J. M. Doyle, Nucl. Instrum. stein, A. Bolozdynya, P. Brusov, L. C. C. Coelho, C. E. Methods Phys. Res., Sect. B 132, 351 (1997). Dahl, L. DeViveiros, et al., Phys. Rev. Lett. 100, 021303 [24] Tectra Mini-Coater, www.tectra.de/mico.htm (2008). [25] Cryomech model PT805, www.cryomech.com [4] E. Aprile, K. L. Giboni, P. Majewski, K. Ni, M. Ya- [26] Omni Nupure III, www.nupure.com mashita, R. Gaitskell, P. Sorensen, L. DeViveiros, [27] R. Brun and F. Rademakers, Nucl. Instrum. Methods L. Baudis, A. Bernstein, et al., New Astron. Rev. 49, Phys. Res., Sect. A 389, 81 (1997). 289 (2005). [28] Thermo Electron Model MP320, www.thermo.com [5] D. B. Cline, Y. Seo, F. Sergiampietri, H. Wang, J. T. [29] S. Agostinelli, J. Allison, K. Amako, J. Apostolakis, White, J. Gao, P. Picchi, G. Mannocchi, L. Periale, H. Araujo, P. Arce, M. Asai, D. Axen, S. Banerjee, F. Pietropaolo, et al., Nucl. Phys. B, Proc. Suppl. 124, G. Barrand, et al., Nucl. Instrum. Methods Phys. Res., 229 (2003). Sect. A 506, 250 (2003). [6] A. Rubbia, Nucl. Phys. B, Proc. Suppl. 149, 197 (2005). [30] Reactor Analysis Tool, http://nu.ph.utexas.edu/rat/trac [7] R. Brunetti, E. Calligarich, M. Cambiaghi, F. Carbonara, [31] S. Himi, T. Takahashi, J. Ruan, and S. Kubota, Nucl. A. Cocco, C. De Vecchi, R. Dolfini, A. Ereditato, G. Fio- Instrum. Methods 203, 153 (1982). rillo, and L. Grandi, New Astron. Rev. 49, 265 (2005). [32] R. Acciarri, M. Antonello, B. Baibussinov, M. Baldo- [8] S. Kubota, M. Hishida, and A. Nohara, Nucl. Instrum. Ceolin, P. Benetti, F. Calaprice, E. Calligarich, Methods 150, 561 (1978). M. Cambiaghi, N. Canci, F. Carbonara, et al., [9] A. Hitachi, T. Takahashi, N. Funayama, K. Masuda, arXiv.org:0804.1222v1 [nucl-ex] (2008). J. Kikuchi, and T. Doke, Phys. Rev. B 27, 5279 (1983). [33] R. Acciarri, M. Antonello, B. Baibussinov, M. Baldo- [10] J. A. Nikkel, R. Hasty, W. H. Lippincott, and D. N. Ceolin, P. Benetti, F. Calaprice, E. Calligarich, McKinsey, Astropart. Phys. 29, 161 (2008). M. Cambiaghi, N. Canci, F. Carbonara, et al., [11] D. N. McKinsey, C. R. Brome, S. N. Dzhosyuk, R. Golub, arXiv.org:0804.1217v1 [nucl-ex] (2008). K. Habicht, P. R. Huffman, E. Korobkina, S. K. Lamore- [34] K. Matusita, Ann. I. Stat. Math. 5, 59 (1954). aux, C. E. H. Mattoni, A. K. Thompson, et al., Phys. [35] W. R. Dillon and M. Goldstein, J. Am. Stat. Assoc. 73, Rev. A 67, 062716 (2003). 305 (1978). [12] D. Akimov, A. Bewick, D. Davidge, J. Dawson, A. S. [36] M. Goldstein and W. Dillon, Discrete Discrimination Howard, I. Ivaniouchenkov, W. G. Jones, J. Joshi, V. A. Analysis (John Wiley and Sons, New York, 1978). Kudryavtsev, T. B. Lawson, et al., Phys. Lett. B 524, [37] B. Efron and R. J. Tibshirani, An Introduction to the 245 (2002). Bootstrap (John Wiley and Sons, New York, 1993). [13] G. J. Davies, J. D. Davies, J. D. Lewin, P. F. Smith, and [38] D. Gastler, A. Hime, E. Kearns, W. H. Lippincott, W. G. Jones, Phys. Lett. B 320, 395 (1994). D. N. McKinsey, D. Mei, J. A. Nikkel, and L. C. Stone- [14] D. N. McKinsey and K. J. Coakley, Astropart. Phys. 22, hill, Phys. Rev. C (2008), to be submitted. 355 (2005). [39] J. Lidgard, Masters Thesis, Queens University, CA [15] M. G. Boulay and A. Hime, Astropart. Phys. 25, 179 (2008), advised by Mark Boulay. (2006). [40] M. Boulay, et al., (2008), to be submitted. [16] P. Benetti, F. Calaprice, E. Calligarich, M. Cambiaghi, [41] D. V. Hinkley, Biometrika 56, 635 (1969). F. Carbonara, F. Cavanna, A. G. Cocco, F. Di Pompeo, [42] A. M. Mood, F. A. Graybill, and D. C. Boes, Introduction N. Ferrari, G. Fiorillo, et al., Nucl. Instrum. Methods to the Theory of Statistics (McGraw Hill, 1974), 3rd ed. Phys. Res., Sect. A 574, 83 (2007). [43] B. T. Cleveland, Nucl. Instrum. Methods Phys. Res., [17] H. H. Loosli and H. Oeschger, Earth Planet. Sci. Lett. 7, Sect. A 214, 451 (1983). 67 (1969). [44] Commercial equipment, instruments, or materials are [18] J. A. Formaggio and C. J. Martoff, Annu. Rev. Nucl. identified in this paper to adequately specify the experi- Part. Sci. 54, 361 (2004). mental procedure. Such identification implies no recom- [19] P. Benetti, R. Acciarri, F. Adamo, B. Baibussinov, mendation or endorsement by NIST, nor does it imply M. Baldo-Ceolin, M. Belluco, F. Calaprice, E. Calli- that the materials or equipment identified are necessar- garich, M. Cambiaghi, F. Carbonara, et al., Astropart. ily the best available for the purpose. Phys. 28, 495 (2008). [20] M. G. Boulay and A. Hime, in V’th Workshop on Science