First Results from MiniBooNE

1
First Results from MiniBooNE

• Eric Prebys, FNAL/BooNE Collaboration

2
The MiniBooNE Collaboration
University of Alabama Los Alamos
National Laboratory Bucknell University
Louisiana State University University of
Cincinnati University of
Michigan University of Colorado
Princeton University Columbia University
Saint Marys University of Minnesota Embry
Riddle University Virginia Polytechnic
Institute Fermi National Accelerator Laboratory
Western Illinois University Indiana
University Yale University
3
Outline

• State of neutrino mixing measurements
• History and background
• Without LSND
• LSND and Karmen
• Experiment
• Beam
• Detector
• Calibration and cross checks
• Analysis
• Reconstruction
• Blindness
• Errors and fitting
• Unblinding
• Results
• Interpretation

4
The Neutrino Problem

• 1968 Experiment in the Homestake Mine first
  observes neutrinos from the Sun, but there are
  far fewer than predicted. Possibilities
• Experiment wrong?
• Solar Model wrong? (? believed by most not
  involved)
• Enough created, but maybe oscillated (or decayed
  to something else) along the way.
• 1987 Also appeared to be too few atmospheric
  muon neutrinos. Less uncertainty in prediction.
  Similar explanation.
• Both results confirmed by numerous experiments
  over the years.
• 1998 SuperKamiokande observes clear oscillatory
  behavior in signals from atmospheric neutrinos.
  For most, this establishes neutrino oscillations
  beyond a reasonable doubt

Solar Problem
Atmospheric Problem
5
Theory of Neutrino Oscillations

• Neutrinos are produced and detected as weak
  eigenstates (ne ,nm, or nt ).
• These can be represented as linear combination of
  mass eigenstates.
• If the above matrix is not diagonal and the
  masses are not equal, then the net weak flavor
  content will oscillate as the neutrinos
  propagate.
• Example if there is mixing between the ne and
  nmthen the probability that a ne will be
  detected as a nm after a distance L is

Mass eigenstates
Flavor eigenstates
Distance in km
Energy in GeV
Only measure magnitude of the difference of the
squares of the masses.
6
Probing Neutrino Mass Differences
Accelerators use p decay to directly probe nm ? ne
Reactors use use disappearance to probe ne ? ?
Reactors
Cerenkov detectors directly measure nm and ne
content in atmospheric neutrinos. Fit to ne?nm ?
nt mixing hypotheses
Also probe with new generation of long
baseline accelerator and reactor experiments,
MINOS, T2K, etc
Solar neutrino experiments typically measure the
disappearance of ne.
7
Best Three Generation Picture (all experiments
but LSND)
8
The LSND Experiment (1993-1998)
mix
30 m
Energy 20-50 MeV

• Signature
• Cerenkov ring from electron
• Delayed g from neutron capture

9
LSND Result
Excess Signal
Best fit
(Soudan, Kamiokande, MACRO, Super-K)
(Homestake, SAGE, GALLEX, Super-K SNO, KamLAND)

• Only exclusive appearance result to date
• Problem Dm2 1 eV2 not consistent with other
  results with simple three generation mixing

10
Possibilities

• 4 neutrinos?
• We know from Z lineshape there are only 3 active
  flavors
• Sterile?
• CP or CPT Violation?
• More exotic scenarios?
• LSND Wrong?
• Cant throw it out just because people dont like
  it.

11
Karmen II Experiment not quite enough
Combined

• Pulse 800 MeV proton beam (ISIS)
• 17.6 m baseline
• 56 tons of liquid scintillator
• Factor of 7 less statistical reach than LSND
• -gt NO SIGNAL
• Combined analysis still leaves an allowed region

12
Role of MiniBooNE

• Boo(ster) N(eutrino) E(xperiment)
• Full BooNE would have two detectors
• Primary Motivation Absolutely confirm or refute
  LSND result
• Optimized for L/E 1
• Higher energy beam -gt Different systematics than
  LSND
• E 30 MeV -gt 500 MeV
• L 30 m -gt 500 m
• Timeline
• Proposed 12/97
• Approved 1998
• Began Construction 10/99
• Completed 5/02
• First Beam 8/02
• Began to run concurrently with NuMI 3/05
• Oscillation results 4/07

13
MiniBooNE Neutrino Beam (not to scale)

• 8 GeV Protons
• 8E16 p/hr max
• 1 detected neutrino/minute
• L/E 1

Little Muon Counter (LMC) to understand K flux
500m dirt
FNALBooster
Be Targetand Horn
50 m Decay Region
Detector
14
Detector

• 950,000 l of pure mineral oil
• 1280 PMTs in inner region
• 240 PMTs outer veto region
• Light produced by Cerenkov radiation and
  scintillation

• Triggers
• All beam spills
• Cosmic ray triggers
• Laser/pulser triggers
• NuMI Trigger
• First off-axis experiment
• Supernova trigger

Light barrier
15
Neutrino Detection/Particle ID
Important Background!!!
16
Selecting Neutrino Events

• Collect data from -5 to 15 usec around each beam
  spill trigger.
• Identify individual events within this window
  based on PMT hits clustered in time.

No cuts
Veto hits lt 6
Veto hitslt6tank hitsgt200
1600 ns spill
Time (ns)
Time (ns)
Time (ns)
17
Delivering Protons

• Requirements of MiniBooNE greatly exceed the
  historical performance of the 30 year old 8 GeV
  Booster, pushes
• Average repetition rate
• Above ground radiation
• Radiation damage and activation of accelerator
  components
• Intense Program to improvethe Booster
• Shielding
• Loss monitoring and analysis
• Lattice improvements (result of Beam Physics
  involvement)
• Collimation system
• Very challenging to continue to operate 8 GeV
  line during NuMI/MINOS operation
• Once believed impossible
• Element of labs Proton Plan
• Goal to continue to deliver roughly 2E20 protons
  per years to the 8 GeV program even as NuMI
  intensities ramp up.

18
Small but Hungry

• MiniBooNE began taking data in Fall of 2002, and
  has currently taken more protons than all other
  users in history of Fermilab combined (including
  NuMI and pBar)

Total protons (1019)
6.3x1020 Protons in n mode This analysis
(5.580.12)x1020 protons In anti-n mode since
1/06
19
Modeling neutrino flux

• Production
• GEANT4 model of target, horn, and beamline
• MARS for protons and neutrons
• Sanford-Wang fit to production data for p and K
• Mesons allowed to decay in model of decay pipe.
• Retain neutrinos which point at target
• Data Constrained by HARP Experiment

p ? m nm
K? m nm
m ? e nm ne K? p e ne

• Intrinsic ne ?ne sources
• m ? e ?nm ne (52)
• K ? p0 e ne (29)
• K0 ? p e ne (14)
• Other ( 5)

ne/nm 0.5
Antineutrino content 6
20
nm Interactions

• Cross sections
• Based on NUANCE 3 Monte Carlo
• Use NEUT and NEUEN as cross checks
• Theoretical input
• Llewellyn-Smith free nucleon cross sections
• Rein-Sehgal resonant and coherent cross-sections
• Bodek-Yang DIS at low-Q2
• Standard DIS parametrization at high Q2
• Fermi-gas model
• Final state interaction model
• Constraining NUANCE
• Observed nm data
• MAeff -- effective axial mass
• EloSF -- Pauli Blocking parameter
• From electron scattering data
• Eb -- binding energy
• pf -- Fermi momentum
• K2K and other experiments are also better
  explained by these modifications

21
Predicted Event Rates (NUANCE)
D. Casper, NPS, 112 (2002) 161
22
Characterizing the Detector

• Full GEANT 3.21 model of detector
• Detailed (39-parameter!!) optical model of oil
• Detailed model of PMT response
• MC events produced at the raw data level and
  reconstructed in the same way as real events

23
Calibrating the Detector
24
Events Producing Pions

CCp Easy to tag due to 3 subevents. Not a
substantial background to the oscillation
analysis.
?
?
25
??
?
N
N
?
NCp0 The p0 decays to 2 photons, which can look
electron-like mimicking the signal... lt1 of
p0 contribute to background.
?
8
?0
?
N
N
(also decays to a single photon with 0.56
probability)
25
Bottom line signal and background

• If the LSND best fit is accurate, only about a
  third of our observed rate will come from
  oscillations
• Backgrounds come from both intrinsic ne and
  misidentified nm

Energy distribution can help separate
26
Analysis Philosophy

• The golden signal in the detector is the
  Charged Current Quasi-Elastic (CCQE) event

• With a particular mass hypothesis, the neutrino
  energy can be reconstructed from the observed
  partcle with

• If the LSND signal were due to neutrino
  oscillations, we would expect an excess in the
  energy range of 300-1500 MeV
• The signal is therefore an event with
• A high probability of being a ne CCQE event
• A reconstructed energy in the range 300 to 1500
  MeV

27
Two Analyses

• Track Based analysis
• Use detailed tracking information to form a
  particle ID liklihood.
• Backgrounds weighted by observed events outside
  of box.
• Boosting
• Particle ID based on feeding a large amount of
  event information into a boosted decision tree
  (details to follow)
• Box defined by the boosted score and the mass
  range.
• Backgrounds determined by models which are
  largely constrained by the data.

28
Common Event Cuts
data MC

• gt200 tank hits
• lt6 veto hits

• Rlt500 cm (algorithm dependent)

29
Blindness

• In general, the two analyses have sequestered
  (hidden) data which has
• A reconstructed neutrino energy in the range 300
  to 1500 MeV
• A high probability of being an electron
• For historical reasons, official box based on
  Boosting analysis cuts
• This leaves the vast majority (99) of the data
  available for analysis
• Individual open boxes were allowed, provided it
  could be established that an oscillation would
  not lead to a significant signal.
• In addition, detector level data (hit times, PMT
  spectra, etc) were available for all events.
• Determination of primary analysis based on
  final sensitivity before opening the box

30
Track-based analysis
ne/nm separation
Reconstructed radius cubed
31
NCp0 separation
32
Testing e-p0 separation using data
1 subevent log(Le/Lm)gt0 (e-like) log(Le/Lp)lt0
(p-like) massgt50 (high mass)
signal
invariant mass
BLIND
log(Le/Lp)
33
Evaluating Sidebands
1 subevent log(Le/Lm)gt0 (e-like) log(Le/Lp)lt0
(p-like) masslt200 (low mass)
Next look here....
c2 Prob for masslt50 MeV (most signal-like)
69
34
Summary of Track-based Cuts
Precuts
Log(Le/Lm) Log(Le/Lp) invariant mass
Backgrounds after cuts
35
Boosted Decision Trees

• Fundamental variables are subjected to a series
  of cuts, each of which classifies the data as
  signal or background.
• However it is classified at each step, the data
  are still subjected to subsequent cuts.
• In the end, the number of times the event is
  classified as background is subtracted from the
  number of times its classified as signal,
  leading to a final score.
• The algorithm is trained on Monte Carlo, with
  both the cut values and their order optimized to
  maximize signal to background.
• Widely used outside of physics
• Example Arditi and Pulket Predicting the
  Outcome of Construction Litigation Using Boosted
  Decision Trees, J. Comp. Civ. Eng., vol 19, iss.
  4 (2005)

• Byron P. Roe, et al.,
• NIM A543 (2005) 577.

36
Some analysis variables
Resolutions vertex 24 cm direction 3.8º energy
14
Reconstructed quantities which are inputs to EnQE
nm CCQE
nm CCQE
Evisible
UZ cosqz
37
Example of a Decision Tree
Variable 1
(Nsignal/Nbkgd)
bkgd-like
signal-like
Variable 2
9755/23695
bkgd-like
sig-like
Variable 3
30,245/16,305
1906/11828
7849/11867
sig-like
bkgd-like
20455/3417
9790/12888
etc.
38
Box Cuts on Energy and Boosting Score
Sideband
Box
ne-like
background-like
39
Test MC with Sideband Information
40
Efficiencies from Boosted Decision Trees
Efficiency after precuts
signal
background
41
Background
nm mis-id
intrinsic ne
(TB analysis)
42
Sources of Uncertainty
Checked or Constrained by MB data
Further reduced by tying ne to nm
Track Based /Boosted Decision Tree error in
Source of Uncertainty On ne background
Flux from p/m decay 6.2 / 4.3 v
v Flux from K decay 3.3 / 1.0 v
v Flux from K0 decay 1.5 / 0.4 v
v Target and beam models 2.8 / 1.3 v
n-cross section 12.3 / 10.5
v v NC p0 yield 1.8 / 1.5 v
External interactions (Dirt) 0.8 / 3.4
v Optical model 6.1 / 10.5
v v DAQ electronics model 7.5 / 10.8
v
43
Overall normalization
BDT
Normalization energy dependence of both
background and signal
Predict
From the nm CCQE events
Data/MC Boosted Decision Tree 1.22 0.29
Track Based 1.32 0.26
Tying the ne background and signal prediction to
the nm flux constrains this analysis to a
strict nm ? ne appearance-only search
44
K and K0 Backgrounds
At high energies, above signal range nm and
ne -like events are largely due to kaon decay
signal range
Signal examples Dm20.4 eV2 Dm20.7 eV2 Dm21.0
eV2
45
Using low and high energy bins to constrain
backgrounds
In both analyses, high energy bins constrain ne
background
TB
signal range
up to 3000 MeV
BDT
In Boosted Decision Tree analysis Low energy bin
(200ltEnQElt300 MeV) constrains nm mis-ids p0,
D?Ng, dirt ...
signal
46
We constrain p0 production using data from our
detector
This reduces the error on predicted mis-identified
p0s
Reweighting improves agreement in other
variables, e.g.?
Because this constrains the D resonance rate, it
also constrains the rate of D?Ng
47
External Sources of Background
Dirt Events
n interactions outside of the detector Ndata/NMC
0.99 0.15
Event Type of Dirt after PID cuts
EnhancedBackgroundCuts
Cosmic Rays
Measured from out-of-beam data 2.1 0.5 events
48
Summary of Backgrounds
49
Constraining the Measurement

• Track-Based
• Re-weight MC predictions to match measured nm
  spectrum, taking into account correlations.
• Boosted Decision Tree
• Include the nm/ne correlations in the error
  matrix
• Systematic and statistical uncertainties are
  included in M

Binned in energy
50
Example Cross Section Uncertainties
(Many are common to nm and ne and cancel in the
fit)
MAQE, elosf 6, 2 (stat bkg only) QE ?
norm 10 QE ? shape function of
E? ??e/?? QE ? function of E? NC ?0 rate
function of ?0 mom MAcoh, coh ?????25
? ? N??rate function of ? mom 7 BF EB, pF
9 MeV, 30 MeV ??s
10 MA1? 25 MAN?
40 DIS ? 25
determined from MiniBooNE ?? QE data
determined from MiniBooNE ?? NC ?0 data
determined from other experiments
51
Multisims

• The large number of parameters that go into the
  modeling of the experiment make it impossible to
  vary each one individually.
• In addition, there are large correlations between
  them, some non-linear.
• Therefore, for a particular class of model, ALL
  variables are varied randomly within their
  allowed limits, each set of values capturing one
  possible reality in what is termed a
  multisim.
• The results are used to reweight the predicted
  value for each energy bin.

1000 multisims for K production
70 multisims for the Optical Model
standard MC
number of multisims
52
Error Matrix Elements

• N is number of events passing cuts
• MC is standard monte carlo
• a represents a given multisim
• M is the total number of multisims
• i,j are EnQE bins

BDT
Total error matrix is sum from each source.
TB ne-only total error matrix BDT nm-ne total
error matrix
53
Sensitivity Comparison
Dc21.64

• As previously agreed, based on this criterion,
  the track-based algorithm was chosen as the
  primary analysis.

54
Comparison with the 5E20 goal as of 2003 Run Plan
55
(No Transcript)
56
Box Opening Procedure

1. Fit sequestered data to oscillation hypothesis,
   returning no fit parameters and only c2 values
   related to oscillation-insensitive diagnostic
   variables.
2. Return plots of those variables.
3. Return c2 of fit to EnQE spectrum
4. Return best fit parameters and plot EnQE
5. Report results 2 weeks later

closed
?
open
57
Step 1

• Fit insensitive variables
• 12 for track-based
• 46 for boosted decision-tree
• All looked good except c2 for visible energy
  spectrum
• Probability 1
• No obvious problem, but background known to be
  high at low energy
• Decide to move low energy cut from 300-475 MeV

EnQEgt 475 MeV We agreed to report events
over the original full range EnQEgt 300 MeV,
58
Second pass at Step 1

• c2s look reasonable
• Go on!

c2 probabilities returned
BDT
TB (EnQEgt475 MeV)
12 variables
46 variables
59
Step 2

• Example Plots

Examples of what we saw
Evisible
c2 Prob 59
c2 Prob 28
fitted energy (MeV)
TB (EnQEgt475 MeV)
BDT
60
Step 3

• Return c2 probabilities
• Track-based 99
• Boosting 52
• Decision proceed to step 4.

61
Step 4 Open the Box

• The track-based nm-gtne oscillation experiment
• Counting experiment 475ltEnQElt1250 MeV
• Data 380 events
• Expectation 358 ?19 (stat) ? 35 (sys) events

significance 0.55 s
No signal!!
62
Track-based energy dependent fit results
Data are in good agreement with background
prediction.
Error bars are diagnonals of error matrix. Fit
errors for gt475 MeV Normalization 9.6 Energy
scale 2.3
Best Fit (dashed) (sin22q, Dm2) (0.001, 4 eV2)
63
The Results
A limit on nm? ne oscillations
c2 probability, null hypothesis 93
Energy fit 475ltEnQElt3000 MeV
64
Full Range
96 17 20 events above background, for
300ltEnQElt475MeV Deviation 3.7s
to Egt475 MeV
Background-subtracted
65
Best fit to the full gt 300 MeV range
Best Fit (dashed) (sin22q, Dm2) (1.0, 0.03
eV2) c2 Probability 18
Well-excluded by Chooz and others
Examples in LSND allowed range

66
Boosted Decision Tree Analysis
Counting Experiment 300ltEnQElt1600 MeV data
971 events expectation 1070 ?33 (stat) ? 225
(sys) events significance -0.38 s
67
Boosted Decision Tree EnQE data/MC comparison
68
Boosted Decision Tree
No evidence for nm? ne appearance-only
oscillations.
Energy-fit analysis solid TB dashed
BDT Independent analyses are in good agreement.
69
Interpreting the limit

• 1) There are various ways
• to present limits
• Single sided raster scan
• (historically used, presented here)
• Global scan
• Unified approach
• (most recent method)
• 2) This result must be
• folded into an
• LSND-Karmen
• joint analysis.
• We will present a full joint analysis soon.

Church, et al., PRD 66, 013001
70
Compatibility with LSND
Maximum Joint Probability
Dm2 (eV2)
MiniBooNE is incompatible with a nm?ne
appearance only interpretation of LSND at 98 CL
71
Conclusions

• MiniBooNE has released the first results of two,
  independent, blind analyses which search for the
  oscillation of nm to ne
• Within the energy range defined prior to
  unblinding the analysis, both results are
  consistent with background.
• We therefore report a limit, which is
  inconsistent with an interpretation of the LSND
  result as two neutrino oscillation.
• One analysis observes a low energy excess,
  outside of the fit region, which will be
  investigated by the MiniBooNE and SciBooNE
  experiments.

72
The Bottom Line
The observed reconstructed energy
distribution is inconsistent with a nm?ne
appearance-only model
Therefore we set a limit on nm?ne appearance
73
Backup Slides
74
Within the energy range defined by this
oscillation analysis, the event rate is
consistent with background.
The observed low energy deviation is under
investigation.
75
We Thank
DOE and NSF
The Fermilab Divisions and Staff
76
Conclusions and Outlook

• MiniBooNE has been running for over three years,
  and continues to run well in the NuMI era
• The analysis tools are well developed and being
  refined to achieve the quality necessary to
  release the result of our blind analysis
• Recent results for CCQE and CCPiP give us
  confidence on our understanding of the detector
  and data.
• Look forward to many interesting results in 2006

77
Backup Slides
78
Sterile Neutrinos Astrophysics Constraints
(courtesy M. Shaevitz)

• Constraints on the number of neutrinos from BBN
  and CMB
• Standard model gives Nn2.60.4 constraint
• If 4He systematics larger, then Nn4.02.5
• If neutrino lepton asymmetry or non-equilibrium,
  then the BBN limit can be evaded.K. Abazajian
  hep-ph/0307266G. Steigman hep-ph/0309347
• One result of this is that the LSND result is
  not yet ruled out by cosmological observations.
  Hannestad astro-ph/0303076

• Bounds on the neutrino masses also depend on the
  number of neutrinos (active and sterile)
• Allowed Smi is 1.4 (2.5) eV 4 (5) neutrinos

79
Reactor Experimental Results

• Single reactor experiments (Chooz, Bugey, etc).
  Look for ne disappearance all negative
• KamLAND (single scintillator detector looking at
  ALL Japanese reactors) ne disappearance
  consistent with mixing.

80
K2K

• First long baseline accelerator experiment
• Beam from KEK PS to Kamiokande, 250 km away
• Look for nm disappearance (atmospheric problem)
• Results consistent with mixing

No mixing
Allowed Mixing Region
Best fit
81
A MiniBooNE LSND Compatibility Test

• For each ?m2, determine the MB and LSND
  measurement
• zMB ? ?zMB, zLSND ? ?zLSND
• where z sin2(2?) and ?z is the 1? error
• For each ?m2, form ?2 between MB and LSND
  measurement
• Find z0 that minimizes ?2
• (weighted average of two measurements) and this
  gives ?2min
• Find probability of ?2min for 1 dof
• this is the joint compatibility probability for
  this ?m2

82
Modeling Production of Secondary Pions

• HARP (CERN)
• 5 l Beryllium target
• 8.9 GeV proton beam momentum

Data are fit to a Sanford-Wang parameterization.
HARP collaboration, hep-ex/0702024
83
Everybody Loves a Mystery

• 32 Sterile neutrinos
• Sorel, Conrad, and Shaevitz (hep-ph/0305255)
• MaVaN 31
• Hung (hep-ph/0010126)
• Sterile neutrinos
• Kaplan, Nelson, and Weiner (hep-ph/0401099)
• Explain Dark Energy?
• CPT violation and 31 neutrinos
• Barger, Marfatia Whisnant (hep-ph/0308299)
• Explain matter/antimatter asymmetry
• Lorentz Violation
• Kostelecky Mewes (hep-ph/0406035)
• Extra Dimensions
• Pas, Pakvasa, Weiler (hep-ph/0504096)
• Sterile Neutrino Decay
• Palomares-Ruiz, Pascoli Schwetz (hep-ph/0505216)

84
Three Generation Mixing (Driven by experiments
listed)

• General Mixing Parameterization

CP violating phase

• Almost diagonal
• Third generation weakly coupled to first two
• Wolfenstein Parameterization

• Mixing large
• No easy simplification
• Think of mass and weak eigenstates as totally
  separate

85
SNO Solar Neutrino Result

• Looked for Cerenkov signals in a large detector
  filled with heavy water.
• Focus on 8B neutrinos
• Used 3 reactions
• ned?ppe- only sensitive to ne
• nxd?pnnx equally sensitive to ne ,nm ,nt
• nx e-? nx e- 6 times more sensitive to ne
  than nm ,nt d
• Consistent with initial full SSM flux of nes
  mixing to nm ,nt
• Completely consistent with three generation
  mixing

Just SNO
SNOothers
86

• The goal of both analyses
• minimize background
• maximize signal efficiency.
• Signal range is approximately
• 300 MeV lt EnQE lt 1500 MeV
• One can then either
• look for a total excess
• (counting expt)
• fit for both an excess and
• energy dependence
• (energy fit)

MiniBooNE signal examples Dm20.4 eV2 Dm20.7
eV2 Dm21.0 eV2
87
Each event is characterized by 7 reconstructed
variables vertex (x,y,z), time, energy,
and direction (q,f)?(Ux, Uy, Uz). Resolutions
vertex 22 cm direction 2.8?
energy 11
nm CCQE events
2 subevents Veto Hitslt6 Tank Hitsgt200
88
Testing e-p0 separation using data
1 subevent log(Le/Lm)gt0 (e-like) log(Le/Lp)lt0
(p-like) massgt50 (high mass)
signal
invariant mass
BLIND
log(Le/Lp)
89
Rejecting p0-like events
Using log(Le/Lp)
ne CCQE
nm NCp0
Cuts were chosen to maximize nm ? ne sensitivity
90
Rejecting muon-like events Using log(Le/Lm)
log(Le/Lm)gt0 favors electron-like hypothesis
Note photon conversions are electron-like. This
does not separate e/p0. Separation is clean at
high energies where muon-like events are
long. Analysis cut was chosen to maximize the
nm ? ne sensitivity
91
nm constraint on intrinsic ne from p decay
chains

• Measure the nm flux
• Kinematics allows
• connection to the p flux

p ? m nm
En (GeV)
En 0.43 Ep
m ? e nm ne
Ep (GeV)

• Once the p flux is known,
• the m flux is determined

92
How the constraints enter...
Two Approaches
TB Reweight MC prediction to match measured nm
result (accounting for systematic error
correlations)

• BDT include the correlations of nm to ne in the
  error matrix

Systematic (and statistical) uncertainties are
included in (Mij)-1
(i,j are bins of EnQE)
93
Analysis 2 Boosted Decision Trees (BDT)
Philosophy
Construct a set of low-level analysis variables
which are used to make a series of cuts to
classify the events.
This algorithm represents an independent cross
check of the Track Based Analysis
94
How the constraints enter...
Two Approaches
TB Reweight MC prediction to match measured nm
result (accounting for systematic error
correlations)

• BDT include the correlations of nm to ne in the
  error matrix

Systematic (and statistical) uncertainties are
included in (Mij)-1
(i,j are bins of EnQE)
95
Handling uncertainties in the analyses
What we begin with...
... what we need
For a given source of uncertainty, Errors on a
wide range of parameters in the underlying model
For a given source of uncertainty, Errors in
bins of EnQE and information on the
correlations between bins
96
Using Multisims to convert from errors on
parameters to errors in EnQE bins For each
error source, Multisims are generated within
the allowed variations by reweighting the
standard Monte Carlo. In the case of the OM,
hit-level simulations are used.
1000 multisims for K production
70 multisims for the Optical Model
standard MC
number of multisims
Number of events passing cuts in bin
500ltEnQElt600 MeV
97
Step 1 Convert the Fundamental
information into Analysis Variables
Fundamental information from
PMTs Analysis Hit Position Charge
Hit Timing variables Energy ? ? Time
sequence ? ? Event shape ? ? ? Physics
? ? ?
Physics p0 mass, EnQE, etc.
98
Step 2 Reduce Analysis Variables to a Single
PID Variable
Boosted Decision Trees
A procedure that combines many weak
classifiers to form a powerful committee
hit level (charge, time, position)
analysis variables
One single PID score
99
Plans
A paper on this analysis will be posted to the
archive and to the MiniBooNE webpage after 5
CT today. Many more papers supporting this
analysis will follow, in the very near
future nm CCQE production p0
production MiniBooNE-LSND-Karmen joint analysis
We are pursuing further analyses of the
neutrino data, including... an analysis which
combines TB and BDT, more exotic models for the
LSND effect. MiniBooNE is presently taking data
in antineutrino mode.
100
A set of decision trees can be developed, each
re-weighting the events to enhance
identification of backgrounds misidentified by
earlier trees (boosting) For each tree,
the data event is assigned 1 if it is
identified as signal, -1 if it is identified as
background. The total for all trees is combined
into a score
Background- like
negative
positive
signal-like
101
Box Opening Procedure
Progress cautiously, in a step-wise fashion

• After applying all analysis cuts
• Fit sequestered data to an oscillation
  hypothesis, returning no fit parameters.
• Return the c2 of the data/MC comparison for a set
  of diagnostic variables.
• 2. Open up the plots from step 1. The Monte
  Carlo has unreported signal.
• Plots chosen to be useful diagnostics, without
  indicating if signal was added.
• 3. Report the c2 for a fit to EnQE , without
  returning fit parameters.
• Compare EnQE in data and Monte Carlo, returning
  the fit parameters.
• At this point, the box is open (March 26, 2007)
• 5. Present results two weeks later.

102
As we show distributions in EnQE, keep in mind
that error bars are the diagonals of the error
matrix. The effect of correlations between EnQE
bins is not shown, however EnQE bin-to-bin
correlations improve the sensitivity to
oscillations, which are based on an
energy-dependent fit.
103
Step 3
Report the c2 for a fit to EnQE across full
energy range
TB (EnQEgt475 MeV) c2 Probability of fit 99
BDT analysis c2 Probability of fit 52
Leading to...
Step 4
Open the box...
104
Accommodating a Positive Signal

• We know from LEP that there are only 3 active,
  light neutrino flavors.
• If MiniBooNE confirms the LSND results, it might
  be evidence for the existence of sterile
  neutrinos

105
Example Optical Model Uncertainties
39 parameters must be varied Allowed variations
are set by the Michel calibration sample
To understand allowed variations, we ran 70
hit-level simulations, with differing
parameters. ?Multisims
106
Other Single Photon Sources
Neutral Current n N ? n N g
negligible
From Efrosinin, hep-ph/0609169, calculation
checked by Goldman, LANL
Charged Current
lt 6 events _at_ 95 CL
n N ? m N g
where the presence of the g leads to
mis-identification
Use events where the m is tagged by the michel
e-, study misidentification using BDT algorithm.
107
Step 4 Counting experiment result
The Track-based nm?ne Appearance-only Result
Counting Experiment 475ltEnQElt1250 MeV
data 380 events expectation 358 ?19 (stat)
? 35 (sys) events
significance 0.55 s
108
Standard Backups
109
SW and FS
We must model the production of pions and
kaons in the Geant4 Beam Monte Carlo
Meson-production data are fit to two independent
models
The best fit parameterization is used in the
Geant4 Monte Carlo. The other result is used to
estimate the parameterization error
110
low Q2
For nuclear targets there has been a deviation in
the Q2 distribution...
K2K Carbon SciFi detector
K2K H2O 1kt detector
MiniBooNE sees this effect too
We can fit our data if we adjust parameters in
our nuclear model including Pauli Blocking
(accounting for energy level differences
when converting C to N)
111
Hit level info
Hit Level information Timing Charge
Position Example of how these can be used
112
mu-e sep, data
Testing muon-electron separation on data
m
p0
All-but-signal events, 1 subevent only
CCQE Box
2 subevent requirement leads to well-tagged nm
sample
8 of muons capture on C, resulting in 1 subevent
113
Cosmic Rays
Cosmic Rays
Measured from our strobe data, using BDT
analysis 2.1 0.5 events in the beam window
Loosening the cut, these events populate upper
region of tank
114
History
History
From start of run, we had initial open boxes and
signal-blind information on all events for
studies. August, 2006, we opened
All-but-Signal box December, 2006, we began a
series of side-band studies February 16, 2007,
we did a blind c2 analysis of the
reconstructed energy distributions, Based on
this we set the final cuts March 26, 2007, we
looked at the signal region.
115
Independence
The Independence of the two analyses

• Independent
• Event reconstruction
• Particle identification
• Cut optimization
• nm CCQE analysis used to predict backgrounds
• Shared
• Beam and event simulation
• Data calibration
• Precuts
• Input systematic errors

The result is different strengths BDT has
better background constraint, Likelihood has
less sensitivity to detector-related systematics.
116
Boxes
Blindness and Algorithm Development
To test the algorithms, define specific event
sets (boxes) with lt 1s signal in an
energy-based analysis
Initial Boxes 0.25 random sample -- an
unbiased cross check michel electrons --
electron subevents from cosmic rays CCQE -- nm
CCQE events used to constrain the flux CCp --
nm CC production with p used to constrain cross
section NCp0 -- nm neutral pion production used
n-e elastic -- electron events for MC
comparison dirt -- External-to-tank neutrino
events putting energy in the tank, used to
measure rate from entering background. High
Energy Events -- all events with energy that is
above the 90 CL signal region, Engt1.4
GeV, used as a cross-check Second Step
All-but-signal box -- explicitly sequester the
signal
117
FSI
Final State Interaction Uncertainties (pion
absorption and charge exchange) are from external
data

• ?? absorption 25
• (inside target nucleus)
• ?? charge exchange 30
• (inside target nucleus)
• ??N ? NN rate 100
• ??? absorption 25
• (in transit thru CH2, wrt GCALOR)
• ??? charge exchange 50
• (in transit thru CH2, wrt GCALOR)

118
HE Box
TB
? p0 ? neK ? nem ? other
ne events for EnQEgt1500 MeV lower than
prediction in both analyses, but within
error (only stat error shown)
119
Norm
Normalization In our Run Plan, we reported a
1.5 data/MC normalization factor. A series of
corrections led to the present ratio of 1.2 The
important changes to the rate prediction
were 17.5 from modeling beam optics to
reflect measurement 16.2 from HARP p
measurement and p-Be cross-sections 3.1 from
CCQE cross section tuning (MAeff, elosf, Eb,
pf) -3.5 from adjustments to the inelastic
cross section -6.0 from secondary
reinteractions in beamline
120
Sense, 300 and 475
Sensitivity for EnQEgt475 MeV and gt 300 MeV
121
300 limit
Fit to the gt300 Energy Range
Best Fit (dashed) (sin22q, Dm2) (1.0, 0.03
eV2) c2 Probability 18
While the c2 is acceptable, there is a systematic
shift in the energy dependence from the
appearance-only model.
122
pi0 momentum bins
Constraining p0 Misidentification
Both analyses use the same method Use the fact
that p0 peaks are visible across a wide range of
momentum in the p0 box The MC p0 rate (flux ?
xsec) is reweighted to match the measurement