Design, Implementation and Test of a new Feature Extractor for the ...
Design, Implementation and Test of a new Feature Extractor for the ...
Design, Implementation and Test of a new Feature Extractor for the ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
<strong>Design</strong>, <strong>Implementation</strong> <strong>and</strong> <strong>Test</strong><br />
<strong>of</strong> a New <strong>Feature</strong> <strong>Extractor</strong><br />
<strong>for</strong> <strong>the</strong> IceCube Neutrino Observatory<br />
von<br />
Marius Wallraff<br />
Diplomarbeit in P H Y S I K<br />
vorgelegt der<br />
Fakultät für Ma<strong>the</strong>matik, In<strong>for</strong>matik und<br />
Naturwissenschaften<br />
der Rheinisch-Westfälischen Technischen Hochschule Aachen<br />
im<br />
März 2010<br />
angefertigt am<br />
III. Physikalischen Institut B<br />
Pr<strong>of</strong>. Dr. Christopher Wiebusch
NewFeatu r eExtracto r
Abstract<br />
The IceCube Neutrino Observatory at South Pole consists<br />
<strong>of</strong> digital optical modules (DOMs) deep down in <strong>the</strong><br />
ice equipped with photomultipliers to capture Čerenkov<br />
light induced by muons <strong>and</strong> o<strong>the</strong>r particles. These<br />
DOMs digitize <strong>the</strong> analogue photomultiplier signals <strong>and</strong><br />
send <strong>the</strong> resulting wave<strong>for</strong>ms to <strong>the</strong> surface. The large<br />
amount <strong>of</strong> in<strong>for</strong>mation has to be condensed <strong>for</strong> later particle<br />
track <strong>and</strong> energy reconstructions.<br />
This <strong>the</strong>sis presents a <strong>new</strong> framework – <strong>the</strong> New<strong>Feature</strong><strong>Extractor</strong>,<br />
NFE – to extract <strong>the</strong> arrival times <strong>and</strong><br />
<strong>the</strong> number <strong>of</strong> photons. Four algorithms have been implemented<br />
in this framework to analyze different types<br />
<strong>of</strong> wave<strong>for</strong>ms. Their per<strong>for</strong>mance is tested by comparison<br />
between simulated data <strong>and</strong> experimental data, <strong>and</strong><br />
by comparison with earlier algorithms, which are also<br />
analyzed conceptually.
Contents<br />
Abstract<br />
List <strong>of</strong> Figures<br />
List <strong>of</strong> Tables<br />
i<br />
vi<br />
ix<br />
1 Introduction 1<br />
2 Neutrino Astrophysics 5<br />
2.1 Neutrinos in Comparison with O<strong>the</strong>r Messenger Particles . . . . . . . . . . 6<br />
2.1.1 Photons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6<br />
2.1.2 Cosmic Rays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7<br />
2.1.3 Gravitational Waves . . . . . . . . . . . . . . . . . . . . . . . . . . 12<br />
2.2 Neutrino Production Processes . . . . . . . . . . . . . . . . . . . . . . . . . 12<br />
2.2.1 Nuclear Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12<br />
2.2.2 Thermal Cooling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12<br />
2.2.3 Cosmic Ray Interactions . . . . . . . . . . . . . . . . . . . . . . . . 13<br />
2.3 Air Showers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14<br />
2.3.1 Atmospheric Muons . . . . . . . . . . . . . . . . . . . . . . . . . . . 16<br />
2.3.2 Atmospheric Neutrinos . . . . . . . . . . . . . . . . . . . . . . . . . 16<br />
3 Neutrino Detection 17<br />
3.1 Neutrino Interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18<br />
3.2 Lepton Propagation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19<br />
3.2.1 Electron Propagation . . . . . . . . . . . . . . . . . . . . . . . . . . 21<br />
3.2.2 Muon Propagation . . . . . . . . . . . . . . . . . . . . . . . . . . . 21<br />
3.2.3 Tau Propagation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22<br />
3.3 Čerenkov Detectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23<br />
4 The IceCube Neutrino Observatory 25<br />
4.1 Layout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26<br />
4.2 Ice Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28<br />
4.3 Digital Optical Modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30<br />
4.4 Signal Digitization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33<br />
4.4.1 ATWD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33<br />
4.4.2 FADC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34<br />
4.4.3 SLC Chargestamps . . . . . . . . . . . . . . . . . . . . . . . . . . . 35<br />
4.5 Data Structure <strong>and</strong> Data Rate . . . . . . . . . . . . . . . . . . . . . . . . . 35<br />
4.6 S<strong>of</strong>tware . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36<br />
4.6.1 DOMcalibrator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36<br />
iii
CONTENTS<br />
5 <strong>Feature</strong> Extraction in IceCube 37<br />
5.1 <strong>Feature</strong><strong>Extractor</strong> . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38<br />
5.2 Pulse<strong>Extractor</strong> . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41<br />
5.3 SLCHit<strong>Extractor</strong> . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41<br />
5.4 New<strong>Feature</strong><strong>Extractor</strong> . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42<br />
5.4.1 Main Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . 42<br />
5.4.2 Program Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . 43<br />
5.4.3 Time Offset Constants . . . . . . . . . . . . . . . . . . . . . . . . . 44<br />
5.4.4 Wave<strong>for</strong>ms Without Pulses . . . . . . . . . . . . . . . . . . . . . . . 45<br />
5.4.5 Pulse Merger . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45<br />
6 Algorithms Implemented in NFE 47<br />
6.1 Pre-evaluation Algorithm “Eva” . . . . . . . . . . . . . . . . . . . . . . . . 48<br />
6.2 Extraction Algorithm “Simple” . . . . . . . . . . . . . . . . . . . . . . . . 49<br />
6.3 Extraction Algorithm “BayesUnfold” . . . . . . . . . . . . . . . . . . . . . 50<br />
6.3.1 Differences <strong>of</strong> FE’s <strong>and</strong> PE’s <strong>Implementation</strong>s . . . . . . . . . . . . 54<br />
6.4 Extraction Algorithm “SLCHE” . . . . . . . . . . . . . . . . . . . . . . . . 55<br />
7 Per<strong>for</strong>mance Optimization 57<br />
7.1 Calibration Using Monte-Carlo Data . . . . . . . . . . . . . . . . . . . . . 58<br />
7.1.1 Pre-evaluation Algorithm "Eva" . . . . . . . . . . . . . . . . . . . . 58<br />
7.1.2 Extraction Algorithm "Simple" . . . . . . . . . . . . . . . . . . . . . 62<br />
7.1.3 Extraction Algorithm "BayesUnfold" . . . . . . . . . . . . . . . . . 65<br />
7.1.4 Extraction Algorithm "SLCHE" . . . . . . . . . . . . . . . . . . . . 67<br />
7.2 Verification Using Experimental Data . . . . . . . . . . . . . . . . . . . . . 71<br />
8 Per<strong>for</strong>mance <strong>Test</strong>s 79<br />
8.1 Extraction <strong>of</strong> Simple Pulses with “Simple” <strong>and</strong> “BayesUnfold” . . . . . . . 80<br />
8.2 Extraction <strong>of</strong> Exotic <strong>Feature</strong>s . . . . . . . . . . . . . . . . . . . . . . . . . 90<br />
8.3 Comparison with O<strong>the</strong>r <strong>Feature</strong> <strong>Extractor</strong>s . . . . . . . . . . . . . . . . . . 97<br />
8.3.1 <strong>Feature</strong><strong>Extractor</strong> in Multi-Pulse Mode . . . . . . . . . . . . . . . . 97<br />
8.3.2 <strong>Feature</strong><strong>Extractor</strong> in Single-Pulse Mode . . . . . . . . . . . . . . . . 112<br />
8.4 SLCHit<strong>Extractor</strong> . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113<br />
8.5 Runtime Per<strong>for</strong>mance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118<br />
9 Summary And Outlook 121<br />
A Bayesian Unfolding 125<br />
A.1 Formal Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125<br />
A.2 Adaption to IceCube’s Wave<strong>for</strong>ms . . . . . . . . . . . . . . . . . . . . . . . 125<br />
B Cascade Pulse Tagging 127<br />
iv
CONTENTS<br />
C Specific Problems <strong>and</strong> Anomalies 129<br />
C.1 ATWD FADC Time Offset Caused Double Extraction . . . . . . . . . . . . 129<br />
C.2 <strong>Implementation</strong> <strong>of</strong> <strong>the</strong> Second Single-Pulse Extraction Algorithm in <strong>Feature</strong><strong>Extractor</strong><br />
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129<br />
C.3 Missing Simulation <strong>of</strong> <strong>the</strong> daq_baseline in DOMsimulator . . . . . . . . . 130<br />
C.4 Time Offset in SLCHit<strong>Extractor</strong> . . . . . . . . . . . . . . . . . . . . . . . . 130<br />
Acknowledgements<br />
Erklärung / Declaration<br />
References<br />
I<br />
III<br />
V<br />
v
List <strong>of</strong> Figures<br />
2.1 Differential particle flux <strong>for</strong> major cosmic ray components. . . . . . . . . . 8<br />
2.2 All-particle cosmic ray energy spectrum from air shower measurements. . . 9<br />
2.3 Second <strong>and</strong> first order Fermi acceleration. . . . . . . . . . . . . . . . . . . 10<br />
2.4 Cosmic ray deflection angles near <strong>the</strong> GZK cut<strong>of</strong>f. . . . . . . . . . . . . . . 11<br />
2.5 Total number <strong>of</strong> neutrinos produced in stars as function <strong>of</strong> star mass. . . . 13<br />
2.6 Cosmic rays <strong>and</strong> neutrinos hitting <strong>the</strong> Earth. . . . . . . . . . . . . . . . . . 15<br />
2.7 Vertical atmospheric muon <strong>and</strong> muon neutrino differential fluxs. . . . . . . 15<br />
2.8 Zenith distribution <strong>of</strong> atmospheric muons <strong>and</strong> atmospheric muon neutrinos. 16<br />
3.1 Neutrino nucleon cross-sections as a function <strong>of</strong> energy. . . . . . . . . . . . 18<br />
3.2 Signatures <strong>of</strong> charged current interactions <strong>of</strong> neutrinos in Čerenkov media. 20<br />
3.3 Depth development <strong>of</strong> electron initiated cascades. . . . . . . . . . . . . . . 20<br />
3.4 Effective radiation length modified by <strong>the</strong> LPM effect. . . . . . . . . . . . . 21<br />
3.5 Expected track length <strong>for</strong> muons <strong>and</strong> taus in Antarctic ice. . . . . . . . . . 22<br />
3.6 Emission <strong>of</strong> Čerenkov radiation. . . . . . . . . . . . . . . . . . . . . . . . . 23<br />
4.1 IceCube surface geometry with deployment season in<strong>for</strong>mation. . . . . . . 26<br />
4.2 Three-dimensional sketch <strong>of</strong> IceCube with DeepCore. . . . . . . . . . . . . 27<br />
4.3 Effective scattering length <strong>and</strong> absorption length in deep Antarctic ice. . . 29<br />
4.4 Sketch <strong>of</strong> an IceCube Digital Optical Module (DOM). . . . . . . . . . . . . 31<br />
4.5 Wave<strong>for</strong>ms illustrating <strong>the</strong> toroidal droop effect <strong>for</strong> OT <strong>and</strong> NT DOMs. . . 32<br />
4.6 Comparison <strong>of</strong> different ATWD <strong>and</strong> FADC SPE shape parametrizations. . 34<br />
5.1 Sketch illustrating <strong>Feature</strong><strong>Extractor</strong>’s two single-pulse extraction algorithms. 39<br />
6.1 Sketch illustrating NFE’s pre-evaluation algorithm “Eva”. . . . . . . . . . . 48<br />
6.2 Sketch illustrating NFE’s extraction algorithm “Simple”. . . . . . . . . . . 49<br />
6.3 Sketch illustrating Bayesian Unfolding. . . . . . . . . . . . . . . . . . . . . 51<br />
6.4 FE, PE, <strong>and</strong> NFE pulse definition algorithms <strong>for</strong> deconvoluted distributions. 53<br />
6.5 Residual plot <strong>for</strong> <strong>the</strong> width parametrization used in “BayesUnfold”. . . . . 54<br />
6.6 Sketch comparing NFE’s extraction algorithm “SLCHE” to SLCHit<strong>Extractor</strong>. 55<br />
7.1 Two examples <strong>for</strong> wave<strong>for</strong>ms with pulses extracted by NFE. . . . . . . . . 59<br />
7.2 Two examples <strong>for</strong> wave<strong>for</strong>ms with pulses extracted by NFE. . . . . . . . . 60<br />
7.3 Two examples <strong>for</strong> wave<strong>for</strong>ms with pulses extracted by NFE. . . . . . . . . 61<br />
7.4 Time residuals <strong>of</strong> NFE extracted pulses <strong>for</strong> different “Simple” charge thresholds.<br />
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63<br />
7.5 Charge-time correlation <strong>of</strong> pulses from NFE extraction algorithm “Simple”. 65<br />
7.6 Effects <strong>of</strong> “BayesUnfold”’s variable number <strong>of</strong> iterations – time residuals. . 68<br />
7.7 Effects <strong>of</strong> “BayesUnfold”’s variable number <strong>of</strong> iterations – charges. . . . . . 69<br />
7.8 Effects <strong>of</strong> “BayesUnfold”’s variable number <strong>of</strong> iterations – numbers <strong>of</strong> pulses. 70<br />
7.9 Effects <strong>of</strong> “BayesUnfold”’s optimized deconvolution starting distribution. . 71<br />
7.10 Time residuals <strong>and</strong> charges <strong>of</strong> “SLCHE” <strong>and</strong> NFE FADC pulses. . . . . . . 72<br />
7.11 Time differences <strong>of</strong> ATWD <strong>and</strong> FADC NFE pulses <strong>for</strong> MC <strong>and</strong> experimental<br />
data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73<br />
vi
LIST OF FIGURES<br />
7.12 Ratio <strong>of</strong> total FADC <strong>and</strong> ATWD wave<strong>for</strong>m charge <strong>for</strong> MC <strong>and</strong> experimental<br />
data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74<br />
7.13 Charges <strong>of</strong> <strong>the</strong> first ATWD <strong>and</strong> FADC NFE pulses <strong>for</strong> MC <strong>and</strong> experimental<br />
data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75<br />
7.14 Differences between <strong>the</strong> total charges <strong>of</strong> ATWD <strong>and</strong> FADC pulses <strong>for</strong> MC<br />
<strong>and</strong> experimental data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76<br />
7.15 Comparison <strong>of</strong> ATWD <strong>and</strong> FADC pulses <strong>for</strong> MC <strong>and</strong> experimental data. . 77<br />
8.1 Time residuals <strong>for</strong> <strong>the</strong> first pulses from simple MC wave<strong>for</strong>ms extracted by<br />
“Simple” <strong>and</strong> “BayesUnfold”. . . . . . . . . . . . . . . . . . . . . . . . . . . 81<br />
8.2 Numbers <strong>of</strong> pulses from simple MC wave<strong>for</strong>ms extracted by “Simple” <strong>and</strong><br />
“BayesUnfold”. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82<br />
8.3 Example <strong>for</strong> a wave<strong>for</strong>m with a high baseline caused by incomplete simulation. 83<br />
8.4 Effect <strong>of</strong> erroneous baselines on “Simple”’s <strong>and</strong> “BayesUnfold”’s numbers<br />
<strong>of</strong> pulses from simple MC wave<strong>for</strong>ms. . . . . . . . . . . . . . . . . . . . . . 84<br />
8.5 Charges <strong>of</strong> <strong>the</strong> first pulses from simple MC wave<strong>for</strong>ms extracted by “Simple”<br />
<strong>and</strong> “BayesUnfold”. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85<br />
8.6 Differences between <strong>the</strong> total charges <strong>of</strong> all pulses from simple MC wave<strong>for</strong>ms<br />
extracted by “Simple” <strong>and</strong> “BayesUnfold”. . . . . . . . . . . . . . . . 86<br />
8.7 Effect <strong>of</strong> erroneous baselines on “Simple”’s <strong>and</strong> “BayesUnfold”’s total charge<br />
from simple MC wave<strong>for</strong>ms. . . . . . . . . . . . . . . . . . . . . . . . . . . 87<br />
8.8 Numbers <strong>of</strong> pulses from simple experimental wave<strong>for</strong>ms extracted by “Simple”<br />
<strong>and</strong> “BayesUnfold”. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88<br />
8.9 Differences between <strong>the</strong> total charges <strong>of</strong> all pulses from simple experimental<br />
wave<strong>for</strong>ms extracted by “Simple” <strong>and</strong> “BayesUnfold”. . . . . . . . . . . . . 89<br />
8.10 Example wave<strong>for</strong>ms to demonstrate NFE’s option En<strong>for</strong>cePulse. . . . . . . 91<br />
8.11 Example wave<strong>for</strong>ms <strong>of</strong> exotic or difficult features. . . . . . . . . . . . . . . 92<br />
8.12 Example wave<strong>for</strong>ms <strong>of</strong> exotic or difficult features. . . . . . . . . . . . . . . 93<br />
8.13 Example wave<strong>for</strong>ms <strong>of</strong> exotic or difficult features. . . . . . . . . . . . . . . 94<br />
8.14 Example wave<strong>for</strong>ms <strong>of</strong> exotic or difficult features. . . . . . . . . . . . . . . 95<br />
8.15 Example <strong>of</strong> a bright wave<strong>for</strong>m from Markus Voge’s catalog. . . . . . . . . . 97<br />
8.16 Time residuals <strong>of</strong> <strong>the</strong> first pulses extracted by FE <strong>and</strong> NFE from MC;<br />
multi-pulse online-filtering settings. . . . . . . . . . . . . . . . . . . . . . . 98<br />
8.17 Differences between <strong>the</strong> times <strong>of</strong> <strong>the</strong> first pulses extracted by FE <strong>and</strong> NFE<br />
from MC <strong>and</strong> experimental data; multi-pulse online-filtering settings. . . . 100<br />
8.18 Charges <strong>of</strong> <strong>the</strong> first pulses extracted by FE <strong>and</strong> NFE from MC <strong>and</strong> experimental<br />
data; multi-pulse online-filtering settings. . . . . . . . . . . . . . . 101<br />
8.19 Example wave<strong>for</strong>ms <strong>for</strong> lower charges in NFE than in FE. . . . . . . . . . 102<br />
8.20 Differences <strong>of</strong> <strong>the</strong> total charges <strong>of</strong> <strong>the</strong> pulses extracted by FE <strong>and</strong> NFE<br />
from MC <strong>and</strong> experimental data; multi-pulse online-filtering settings. . . . 103<br />
8.21 Example <strong>of</strong> saturated wave<strong>for</strong>ms in different calibrations with pulses from<br />
NFE <strong>and</strong> FE. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104<br />
vii
LIST OF FIGURES<br />
8.22 Numbers <strong>of</strong> pulses extracted by FE <strong>and</strong> NFE from MC <strong>and</strong> experimental<br />
data; multi-pulse online-filtering settings. . . . . . . . . . . . . . . . . . . . 106<br />
8.23 Time residuals <strong>of</strong> <strong>the</strong> first pulses extracted by FE <strong>and</strong> NFE from MC;<br />
multi-pulse <strong>of</strong>fline-processing settings. . . . . . . . . . . . . . . . . . . . . . 107<br />
8.24 Charges <strong>of</strong> <strong>the</strong> first pulses extracted by FE <strong>and</strong> NFE from MC <strong>and</strong> experimental<br />
data; multi-pulse <strong>of</strong>fline-processing settings. . . . . . . . . . . . . . 108<br />
8.25 Numbers <strong>of</strong> pulses extracted by FE <strong>and</strong> NFE from MC <strong>and</strong> experimental<br />
data; multi-pulse <strong>of</strong>fline-processing settings. . . . . . . . . . . . . . . . . . 109<br />
8.26 Charge per pulse ratio <strong>of</strong> simulated data <strong>and</strong> experimental data <strong>for</strong> FE <strong>and</strong><br />
NFE. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111<br />
8.27 Time residuals <strong>of</strong> <strong>the</strong> first pulses extracted by FE <strong>and</strong> NFE from MC;<br />
single-pulse settings. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112<br />
8.28 Charges <strong>of</strong> <strong>the</strong> first pulses extracted by FE <strong>and</strong> NFE from MC <strong>and</strong> experimental<br />
data; single-pulse settings. . . . . . . . . . . . . . . . . . . . . . . . 114<br />
8.29 Differences <strong>of</strong> <strong>the</strong> total charges <strong>of</strong> <strong>the</strong> pulses extracted by FE <strong>and</strong> NFE<br />
from MC <strong>and</strong> experimental data; single-pulse settings. . . . . . . . . . . . . 115<br />
8.30 Time residuals <strong>of</strong> pulses extracted by SLCHit<strong>Extractor</strong> <strong>and</strong> “SLCHE” from<br />
MC. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116<br />
8.31 Charges <strong>of</strong> pulses extracted by SLCHit<strong>Extractor</strong> <strong>and</strong> “SLCHE” from MC<br />
<strong>and</strong> experimental data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117<br />
B.1 Tagging <strong>of</strong> pulses caused by cascades during muon propagation. . . . . . . 127<br />
C.1 Comparison <strong>of</strong> <strong>Feature</strong><strong>Extractor</strong>’s documented <strong>and</strong> implemented second<br />
single-pulse extraction algorithms. . . . . . . . . . . . . . . . . . . . . . . . 130<br />
C.2 Effect <strong>of</strong> <strong>the</strong> simulation <strong>of</strong> <strong>the</strong> daq_baseline on extracted pulses. . . . . . . 131<br />
viii
List <strong>of</strong> Tables<br />
4.1 Overview over different digitizers resp. I3Wave<strong>for</strong>m source types. . . . . . . 33<br />
7.1 Default values used <strong>for</strong> pre-evaluation algorithm “Eva”. . . . . . . . . . . . 62<br />
7.2 Default values used <strong>for</strong> extraction algorithm “Simple”. . . . . . . . . . . . . 62<br />
7.3 Parameter values <strong>of</strong> <strong>the</strong> SPE pulse parametrizations employed. . . . . . . . 66<br />
7.4 Default values used <strong>for</strong> <strong>the</strong> extraction algorithm “BayesUnfold”. . . . . . . 66<br />
7.5 Default values used <strong>for</strong> extraction algorithm “SLCHE”. . . . . . . . . . . . 67<br />
8.1 Runtimes <strong>for</strong> different feature extractors <strong>and</strong> datasets. . . . . . . . . . . . . 119<br />
ix
LIST OF TABLES<br />
x
CHAPTER I<br />
Introduction
1 INTRODUCTION<br />
Astronomy challenged <strong>and</strong> inspired mankind since prehistoric times. In its beginning,<br />
it was strongly connected to astrology <strong>and</strong> served as basis <strong>for</strong> cults <strong>and</strong> religious beliefs.<br />
With <strong>the</strong> introduction <strong>of</strong> agricultural techniques <strong>and</strong> <strong>the</strong> rise <strong>of</strong> great civilizations, many<br />
practical applications emerged independently across <strong>the</strong> globe. Astronomy enabled <strong>the</strong><br />
creation <strong>of</strong> reliable calendars to <strong>for</strong>ecast seasonal climate changes, provided means to<br />
learn about <strong>the</strong> topography <strong>of</strong> Earth, allowed navigation on open sea, <strong>and</strong> more generally<br />
motivated ma<strong>the</strong>matical <strong>and</strong> scientific progress.<br />
The ancient astronomers were limited to observations that could be undertaken with<br />
<strong>the</strong>ir bare eyes. This did not change until Hans Lippershey invented <strong>the</strong> first known<br />
telescope in 1608. Utilizing <strong>the</strong> light-ga<strong>the</strong>ring power <strong>of</strong> <strong>the</strong>se <strong>new</strong> instruments, famous<br />
astronomers like Galileo Galilei <strong>and</strong> Johannes Kepler soon made <strong>new</strong> discoveries <strong>and</strong> laid<br />
<strong>the</strong> foundation <strong>for</strong> modern astronomy <strong>and</strong> astrophysics. During <strong>the</strong> following centuries<br />
constant improvements to <strong>the</strong> instruments were made, but it was until <strong>the</strong> 19th century<br />
that fundamentally <strong>new</strong> techniques were devised: Photography allowed precise long-time<br />
observations, spectrography made it possible to analyze <strong>the</strong> composition <strong>and</strong> radial velocity<br />
<strong>of</strong> distant objects. While especially <strong>the</strong> latter provided valuable <strong>new</strong> observables to<br />
astrophysics, researchers were still depending on visible light. This limitation was eliminated<br />
in <strong>the</strong> middle <strong>of</strong> <strong>the</strong> 20th century with <strong>the</strong> establishment <strong>of</strong> radio astronomy. In<br />
<strong>the</strong> following decades, many parts <strong>of</strong> <strong>the</strong> electromagnetic spectrum were discovered to be<br />
valuable means <strong>for</strong> <strong>the</strong> exploration <strong>of</strong> <strong>the</strong> visible universe, such as far <strong>and</strong> near infrared,<br />
ultraviolet, X-rays, <strong>and</strong> gamma rays.<br />
In parallel to this late development, Theodor Wulf, Domenico Pacini <strong>and</strong> Victor Hess<br />
found <strong>the</strong> first strong evidences <strong>for</strong> extra-terrestial particles entering <strong>the</strong> Earth’s atmosphere,<br />
<strong>the</strong> so-called cosmic rays. In <strong>the</strong> following decades, many experiments were conducted<br />
to analyze <strong>the</strong> underlying mechanism, <strong>the</strong> composition <strong>and</strong> origins <strong>of</strong> <strong>the</strong>se cosmic<br />
rays, <strong>and</strong> a <strong>new</strong> field <strong>of</strong> physics was born: particle physics. Today, most particle physics experimentalists<br />
work at accelerator experiments which allow high-precision measurements<br />
in controlled environments <strong>and</strong> extremely high luminosities. However, cosmic rays remain<br />
to be an interesting subject <strong>for</strong> both astrophysics <strong>and</strong> particle physics as <strong>the</strong>y are both<br />
astronomical in<strong>for</strong>mation carriers complementary to electromagnetic radiation <strong>and</strong> reach<br />
energies several orders <strong>of</strong> magnitude higher than those currently produced at particle<br />
accelerators.<br />
The advancement <strong>of</strong> detectors <strong>for</strong> <strong>the</strong> different branches <strong>of</strong> multi-messenger astronomy<br />
<strong>and</strong> astroparticle physics has proven to be an important aspect <strong>of</strong> <strong>the</strong> ef<strong>for</strong>ts made<br />
to underst<strong>and</strong> <strong>the</strong> structure <strong>of</strong> matter <strong>and</strong> <strong>the</strong> Universe. This <strong>the</strong>sis is written in <strong>the</strong><br />
hope to be a small contribution to this endeavor by designing, implementing <strong>and</strong> finally<br />
testing a <strong>new</strong> feature extractor <strong>for</strong> <strong>the</strong> IceCube Neutrino Observatory – <strong>the</strong> New<strong>Feature</strong><strong>Extractor</strong>,<br />
NFE. NFE is a package which analyses <strong>the</strong> recorded photomultiplier signals <strong>and</strong><br />
extracts <strong>the</strong> physics in<strong>for</strong>mation; this is <strong>the</strong> number <strong>and</strong> arrival time <strong>of</strong> photons hitting<br />
<strong>the</strong> photomultipliers.<br />
2
Chapter 2 gives a short motivation <strong>and</strong> explains <strong>the</strong> basics <strong>of</strong> neutrino astrophysics<br />
with its advantages <strong>and</strong> challenges. Chapter 3 goes more into detail <strong>and</strong> explains <strong>the</strong><br />
neutrinos’ interactions with baryonic matter, <strong>the</strong> propagation <strong>of</strong> <strong>the</strong> resulting particles<br />
through matter <strong>and</strong> how <strong>the</strong>se properties can be exploited <strong>for</strong> neutrino detection. Chapter<br />
4 describes <strong>the</strong> IceCube Neutrino Observatory, with emphasis on its hardware <strong>and</strong> <strong>the</strong><br />
data acquisition. Chapter 5’s topic is feature extraction in IceCube. The three most<br />
important existing feature extractor packages are introduced, <strong>and</strong> <strong>the</strong> design decisions <strong>for</strong><br />
New<strong>Feature</strong><strong>Extractor</strong> are discussed. Chapter 6 illustrates <strong>the</strong> different algorithms used<br />
by NFE; <strong>the</strong> corresponding per<strong>for</strong>mance optimization analyses are explained in chapter<br />
7. Per<strong>for</strong>mance tests are presented in chapter 8. Finally chapter 9 gives a summary <strong>and</strong><br />
provides an outlook on future developments <strong>and</strong> tasks.<br />
3
1 INTRODUCTION<br />
4
CHAPTER II<br />
Neutrino Astrophysics
2 NEUTRINO ASTROPHYSICS<br />
2.1 Neutrinos in Comparison with O<strong>the</strong>r Messenger Particles<br />
Neutrino astronomy is one <strong>of</strong> <strong>the</strong> modern branches <strong>of</strong> observational astronomy. Most <strong>of</strong><br />
<strong>the</strong> early neutrino detectors were originally designed <strong>for</strong> astrophysics (e.g., <strong>the</strong> Homestake<br />
experiment <strong>for</strong> measuring <strong>the</strong> solar neutrino flux) or particle physics (like KamiokaNDE<br />
<strong>for</strong> <strong>the</strong> investigation <strong>of</strong> proton decay). Supernova SN1987A proved neutrino telescopes<br />
to be valuable <strong>for</strong> both observational astronomy <strong>and</strong> astroparticle physics. Among o<strong>the</strong>r<br />
applications, extra-terrestial neutrinos can be used to test astrophysical models <strong>and</strong> to<br />
constrain values <strong>for</strong> neutrino properties which cannot be accessed directly by accelerator<br />
experiments.<br />
A big challenge is <strong>the</strong> very small cross-section <strong>of</strong> neutrino interactions (see section 3.1)<br />
combined with a strong background <strong>of</strong> atmospheric particles caused by air showers (see<br />
section 2.3). However, this small cross-section is also a major advantage <strong>of</strong> neutrino<br />
astronomy as it allows to probe astrophysical regions obstructed by ei<strong>the</strong>r <strong>for</strong>eground<br />
objects or by <strong>the</strong> object itself. Fur<strong>the</strong>rmore, because <strong>of</strong> different production processes<br />
neutrinos carry in<strong>for</strong>mation that is complementary to that obtainable by o<strong>the</strong>r types <strong>of</strong><br />
telescopes, making neutrinos a valuable part <strong>of</strong> multi-messenger astronomy.<br />
2.1.1 Photons<br />
The electromagnetic spectrum <strong>of</strong>fers much diversity in <strong>the</strong> obtainable data. The wavelength<br />
λ is coupled with <strong>the</strong> photon’s energy E by<br />
λ = c ν = hc<br />
E<br />
1.2398 eV m<br />
≈ ,<br />
E<br />
where ν denotes <strong>the</strong> frequency <strong>of</strong> <strong>the</strong> wave; because <strong>of</strong> this, photons <strong>of</strong> different energies<br />
interact with different objects. While <strong>for</strong> example visible light is absorbed by interstellar<br />
dust, <strong>the</strong> dust grains reemit this light brightly in infrared frequency b<strong>and</strong>s. The HI line –<br />
also known as 21 centimeter line – is <strong>the</strong> spectral line corresponding to <strong>the</strong> hyperfine level<br />
transitions <strong>of</strong> atomic hydrogen <strong>and</strong> can be observed by microwave telescopes. Because <strong>of</strong><br />
its small width its doppler shifts can be used to calculate velocity maps <strong>of</strong> <strong>the</strong> interstellar<br />
medium. From roughly 200 keV upwards, gamma-ray observatories monitor <strong>the</strong> sky <strong>for</strong><br />
high-energy objects <strong>and</strong> events such as supernovae remnants, gamma-ray bursts or active<br />
galactic nuclei (AGNs).<br />
One major disadvantage <strong>of</strong> photons is <strong>the</strong>ir high scattering <strong>and</strong> absorbtion probability.<br />
In many waveb<strong>and</strong>s, parts <strong>of</strong> <strong>the</strong> sky are obstructed by interstellar matter, especially by<br />
<strong>for</strong>mations known as dark nebulas such as <strong>the</strong> Great Rift, or by <strong>the</strong> Galactic Center.<br />
Fur<strong>the</strong>rmore, many parts <strong>of</strong> <strong>the</strong> spectrum are absorbed by <strong>the</strong> Earth’s atmosphere, raising<br />
<strong>the</strong> need <strong>for</strong> extra-terrestial detectors, which are however limited in mass <strong>and</strong> <strong>the</strong>reby<br />
not suitable <strong>for</strong> ultra-high-energy surveys as <strong>the</strong> low flux at <strong>the</strong>se energies requires large<br />
effective areas.<br />
6
2.1 Neutrinos in Comparison with O<strong>the</strong>r Messenger Particles<br />
2.1.2 Cosmic Rays<br />
Hadronic cosmic rays are massive particles originating from outer space. Their composition<br />
is energy dependent; at low energy about 90% <strong>of</strong> <strong>the</strong> particles are protons, 9% are<br />
helium nuclei <strong>and</strong> <strong>the</strong> rest consists <strong>of</strong> heavier nuclei, neutrons <strong>and</strong> anti-protons (see figure<br />
2.1).[3] Cosmic Rays also have leptonic components; e.g., about one additional percent<br />
consists <strong>of</strong> electrons. Depending on <strong>the</strong> definition used, neutrinos <strong>and</strong> gamma rays are<br />
also included. For <strong>the</strong> rest <strong>of</strong> this section, ”cosmic rays“ is used synonymously <strong>for</strong> charged<br />
hadronic cosmic rays unless stated o<strong>the</strong>rwise.<br />
Cosmic rays can be divided into two categories: Primary cosmic rays are directly emitted<br />
from astronomical objects such as stars, supernovae or AGNs, while secondary cosmic<br />
rays emerge from collisions between primary ones <strong>and</strong> interstellar matter.Some elements<br />
such as Li, Be, B, F, Sc, <strong>and</strong> V (which are rare in stars because no stable isotopes are<br />
generated during stellar nucleosyn<strong>the</strong>sis) as well as anti-protons are unlikely to be primary<br />
cosmic ray particles. However, during propagation <strong>the</strong> composition changes because <strong>of</strong><br />
nuclear interactions. Thus, <strong>the</strong> abundance <strong>of</strong> <strong>the</strong> a<strong>for</strong>ementioned particles can be used<br />
to differentiate between <strong>the</strong> two categories <strong>and</strong> <strong>the</strong>reby to learn about <strong>the</strong> interstellar<br />
medium.<br />
The energy spectrum <strong>of</strong> cosmic rays above 10 GeV follows an at least tw<strong>of</strong>old broken<br />
power law:<br />
dN<br />
dE ∝ E−α ,<br />
with different spectral indices α at different energy regions. Between about 10 GeV <strong>and</strong><br />
<strong>the</strong> so-called knee at about 4·10 6 GeV, <strong>the</strong> spectral index is 2.68 ± 0.02[4], between <strong>the</strong><br />
knee <strong>and</strong> <strong>the</strong> disputable second knee near 3·10 8 GeV it is 3.02 ± 0.03[5], from <strong>the</strong>re up to<br />
<strong>the</strong> ankle at 5·10 9 GeV its value is 3.16 ± 0.08[5] <strong>and</strong> at even higher energies it flattens<br />
to 2.81 ± 0.03[6]. Above <strong>the</strong> GZK cut<strong>of</strong>f energy <strong>of</strong> 5·10 10 GeV <strong>the</strong> spectrum is assumed<br />
to steepen dramatically because <strong>of</strong> interactions <strong>of</strong> <strong>the</strong> cosmic rays with cosmic microwave<br />
background photons,<br />
p + γ → ∆ + → p + π 0 <strong>and</strong><br />
p + γ → ∆ + → n + π + .<br />
The nucleons emerge with about 80% <strong>of</strong> <strong>the</strong> initial proton energy, <strong>the</strong> effective mean free<br />
path is <strong>of</strong> <strong>the</strong> order <strong>of</strong> 13 Mpc.[7] Recent HiRes <strong>and</strong> Auger results show strong evidence<br />
<strong>for</strong> a cut<strong>of</strong>f, however it is unresolved if this is caused by <strong>the</strong> GZK effect.[6]<br />
Low energy cosmic rays up to about 10 GeV are modified by solar flares. The origin<br />
<strong>of</strong> high-energy cosmic rays is unknown. Top-down models propose <strong>the</strong>m to be decay<br />
products <strong>of</strong> topological defects <strong>for</strong>med in <strong>the</strong> very early universe; however, <strong>the</strong>se models<br />
7
2 NEUTRINO ASTROPHYSICS<br />
Figure 2.1: Differential particle flux <strong>for</strong> <strong>the</strong> major cosmic ray components.[1] Shown are <strong>the</strong><br />
ten elements which have <strong>the</strong> highest abundance in <strong>the</strong> solar system.[2]<br />
8
2.1 Neutrinos in Comparison with O<strong>the</strong>r Messenger Particles<br />
10 5<br />
Knee<br />
E 2.7 F(E) [GeV 1.7 m −2 s −1 sr −1 ]<br />
10 4<br />
Grigorov<br />
JACEE<br />
MGU<br />
TienShan<br />
Tibet07<br />
Akeno<br />
CASA/MIA<br />
Hegra<br />
Flys Eye<br />
Agasa<br />
HiRes1<br />
HiRes2<br />
Auger SD<br />
Auger hybrid<br />
Kascade<br />
10 3 10 14 10 15<br />
2nd Knee<br />
Ankle<br />
10 13 10 16 10 17 10 18 10 19 10 20<br />
E [eV]<br />
Figure 2.2: All-particle cosmic ray spectrum at high energies from air shower measurements,<br />
weighted with E 2.7 . The shaded area was probed by direct measurements.[1]<br />
are disfavored because <strong>the</strong>y predict a photon flux that has already been ruled out by<br />
EGRET <strong>and</strong> Auger.[8] Bottom-up models on <strong>the</strong> o<strong>the</strong>r h<strong>and</strong> explain high-energy cosmic<br />
rays by acceleration <strong>of</strong> lower energy ones, e.g., by processes called Fermi acceleration.<br />
Second order Fermi acceleration relies on <strong>the</strong> r<strong>and</strong>om elastic scattering <strong>of</strong> charged particles<br />
at moving magnetic inhomogenities such as sparse plasma clouds, see figure 2.3. If <strong>the</strong><br />
particle enters <strong>the</strong> inhomogenity head-on <strong>and</strong> is reflected, its energy increases. Likewise it<br />
loses energy if it is reflected at a receding inhomogenity, but assuming a r<strong>and</strong>om velocity<br />
distribution this is less likely to happen. The resulting relative energy gain per reflection<br />
as a function <strong>of</strong> <strong>the</strong> entry angle θ 1 , <strong>the</strong> exit angle θ 2 <strong>and</strong> <strong>the</strong> cloud’s speed v c ≡ cβ c <strong>for</strong><br />
relativistic particles (β p ≈ 1) is<br />
ξ := ∆E<br />
E = 1 − β c cos(θ 1 ) + β c cos(θ 2 ) − βc 2 cos(θ 1 ) cos(θ 2 )<br />
− 1. (2.1)<br />
1 − βc<br />
2<br />
Averaging equation (2.1) over θ 1 <strong>and</strong> θ 2 leads to<br />
ξ = 1 + 1 3 β2 c<br />
1 − β 2 c<br />
− 1 = 4 3 β2 c + O(β 4 c ). (2.2)<br />
The speed <strong>of</strong> galactic plasma clouds is only <strong>of</strong> <strong>the</strong> order <strong>of</strong> β c = 10 −4 , <strong>the</strong>re<strong>for</strong>e second<br />
order Fermi acceleration is too ineffective to explain <strong>the</strong> highest energy cosmic rays.[3]<br />
This does not apply to first order Fermi acceleration: Instead <strong>of</strong> considering plasma clouds,<br />
first order Fermi acceleration takes place at large <strong>and</strong> sufficiently plain shock fronts such as<br />
9
2 NEUTRINO ASTROPHYSICS<br />
E + DE<br />
E + DE<br />
v g<br />
v s<br />
q 2<br />
q 1<br />
v c<br />
q 2<br />
q1<br />
plasma<br />
cloud<br />
E<br />
E<br />
upstream<br />
(unshocked)<br />
downstream<br />
(shocked)<br />
Figure 2.3: Left: Second order Fermi acceleration at magnetic plasma clouds.<br />
Right: First order Fermi acceleration at shock fronts.<br />
those originating from supernovae (see figure 2.3). If a shock front traveling at β s through<br />
resting unshocked plasma (upstream) accelerates <strong>the</strong> plasma to β g < β s (downstream),<br />
equation (2.1) holds <strong>for</strong> particles passing <strong>the</strong> shock front from <strong>the</strong> upstream side as <strong>the</strong>ir<br />
average speeds adapt to <strong>the</strong> average speed <strong>of</strong> <strong>the</strong> shocked plasma after enough elastic<br />
scattering processes, with β c := β g . As <strong>the</strong> particles are now on average comoving with<br />
<strong>the</strong> shocked plasma, <strong>the</strong>y will gain <strong>the</strong> same amount <strong>of</strong> energy by a second crossing <strong>of</strong> <strong>the</strong><br />
shock front, with <strong>the</strong> unshocked plasma approaching at β g in <strong>the</strong> shocked plasma’s rest<br />
frame. Averaging over <strong>the</strong> angles yields<br />
ξ = 1 + 4 3 β g + 4 9 β2 g<br />
1 − β 2 g<br />
− 1 = 4 3 β g + O(β 2 g). (2.3)<br />
First order Fermi acceleration is substantially more efficient than <strong>the</strong> second order variant<br />
not only because <strong>of</strong> ξ being proportional to <strong>the</strong> first order <strong>of</strong> β g but also because <strong>of</strong> larger<br />
average speeds (β g ≈ 10 −1 ) <strong>and</strong> higher repetition rates – <strong>the</strong> same shock front can be<br />
passed many times.<br />
The escape probability can be estimated as P esc = 4(β s − β g ).[3] Using this, <strong>the</strong><br />
expected spectral index <strong>for</strong> non-relativistic shocks at <strong>the</strong> source is<br />
α = P esc<br />
ξ<br />
+ 1 ≈ 3(β s − β g )<br />
β g<br />
+ 1. (2.4)<br />
For strong shocks in a monatomic gas or plasma this evaluates to α ≈ 2.[3] For highly<br />
relativistic shocks, <strong>the</strong> spectral index approaches 2.3.[9] The observed cosmic ray spectrum<br />
with α ≈ 2.7 can be understood with (first order) Fermi acceleration if <strong>the</strong> effect <strong>of</strong> <strong>the</strong><br />
propagation is considered: Certain propagation models such as <strong>the</strong> leaky box model predict<br />
a s<strong>of</strong>tening <strong>of</strong> <strong>the</strong> spectrum because higher energy particles are more likely to escape <strong>the</strong><br />
galaxy in which <strong>the</strong>y are accelerated.<br />
Cosmic rays with energies up to <strong>the</strong> knee are assumed to originate from galactic accelerators<br />
such as supernovae remnants, while ultra-high energies above <strong>the</strong> ankle are<br />
10
2.1 Neutrinos in Comparison with O<strong>the</strong>r Messenger Particles<br />
Figure 2.4: Ratio A <strong>of</strong> <strong>the</strong> sky with cosmic ray deflection angle δ > δ th at 4·10 10 GeV, extrapolated<br />
to <strong>the</strong> propagation distance d = 500 Mpc under <strong>the</strong> assumption that<br />
A(δ th , d) = A 0 (δ th · ( d 0<br />
d<br />
) α ). Between 70 Mpc <strong>and</strong> 110 Mpc, α equals 0.8, however it<br />
might drop to lower values because <strong>the</strong> magnetic field’s orientation differs between<br />
galaxy filaments.[10]<br />
attributed to extra-galactic sources which still have to be determined. The spectral s<strong>of</strong>tening<br />
in <strong>the</strong> “lower leg” energy region might be explained by propagation models or cut-<strong>of</strong>f<br />
energies as mentioned above. The transition between galactic <strong>and</strong> extra-galactic cosmic<br />
rays remains unknown.<br />
Cosmic rays are interesting messenger particles because <strong>the</strong>y carry in<strong>for</strong>mation about<br />
<strong>the</strong> composition <strong>of</strong> <strong>the</strong>ir sources (primary cosmic rays) <strong>and</strong> <strong>the</strong> properties <strong>of</strong> interstellar<br />
matter (secondary cosmic rays). However, only ultra-high-energetic cosmic rays can be<br />
used to directionally locate <strong>the</strong> source objects: Primary cosmic rays are charged, thus<br />
<strong>the</strong>y are subject to electro-magnetic deflection. A particle’s gyroradius is given by<br />
r = p ⊥<br />
qB ≈ 3.3 m E e T<br />
GeV q B<br />
(2.5)<br />
with p ⊥ being <strong>the</strong> particle’s momentum perpendicular to <strong>the</strong> magnetic field ⃗ B. For most<br />
parts <strong>of</strong> <strong>the</strong> sky, <strong>the</strong> average deflection angle <strong>for</strong> particles near <strong>the</strong> GZK cut<strong>of</strong>f does not<br />
exceed a few degrees over at least 500 Mpc, see figure 2.4.<br />
11
2 NEUTRINO ASTROPHYSICS<br />
2.1.3 Gravitational Waves<br />
The carrier <strong>of</strong> <strong>the</strong> gravitational <strong>for</strong>ce called graviton has not yet been directly detected<br />
<strong>and</strong> might never be due to fundamental difficulties such as extremely small cross-sections,<br />
very low fluxes <strong>and</strong> a very strong neutrino background.[11] Still, gravitation can be used<br />
<strong>for</strong> astronomical means by searching <strong>for</strong> gravitational waves, which are distortions <strong>of</strong> <strong>the</strong><br />
spacetime traveling at <strong>the</strong> speed <strong>of</strong> light. They can only be emitted by sources with timedependent<br />
quadrupole or higher-order moments in <strong>the</strong>ir stress-energy tensor; c<strong>and</strong>idates<br />
are closely rotating or colliding objects such as neutron stars or black holes.<br />
Up to September 2009 no gravitational waves have been identified, but competitive<br />
upper limits have been published by various earth-bound detectors.[12] Space-based detectors<br />
such as LISA are planned to extend <strong>the</strong> observed frequency range within <strong>the</strong> next few<br />
decades. Compared to neutrino observatories, <strong>the</strong>y are more limited concerning <strong>the</strong> possible<br />
source objects, but <strong>the</strong>y can deliver complementary in<strong>for</strong>mation <strong>and</strong> provide tests <strong>for</strong><br />
general relativity <strong>and</strong> possibly o<strong>the</strong>r models <strong>of</strong> gravity. Both gravitational waves <strong>and</strong> neutrinos<br />
are able to propagate through regions that are not transparent to electro-magnetic<br />
radiation.<br />
2.2 Neutrino Production Processes<br />
2.2.1 Nuclear Processes<br />
Stars in <strong>the</strong>ir nuclear fusion phase continously produce large numbers <strong>of</strong> neutrinos. For<br />
stars with masses from 2 up to at least 20 solar masses (M ⊙ ), <strong>the</strong> total number <strong>of</strong> neutrinos<br />
produced by fusion until <strong>the</strong> end <strong>of</strong> <strong>the</strong> helium burning phase in <strong>the</strong> core <strong>of</strong> <strong>the</strong> star is<br />
approximately proportional to <strong>the</strong> star’s initial mass (figure 2.5). Later fusion phases<br />
that can occur in stars heavier than 4 solar masses contribute only about one additional<br />
tenth to this number as none <strong>of</strong> <strong>the</strong> main processes besides <strong>the</strong> silicon burning require<br />
proton-neutron-conversions or similar weak interactions.[14]<br />
Some neutrinos originate from processes with well-defined initial states <strong>and</strong> two-body final<br />
states such as electron capture. These neutrinos are emitted at discrete energies, while<br />
o<strong>the</strong>rs are produced with many-body final states (e.g. by beta decay) <strong>and</strong> have continuous<br />
energy spectra. Independent <strong>of</strong> this, <strong>the</strong>ir highest energies are <strong>of</strong> <strong>the</strong> order <strong>of</strong> some MeV<br />
because this is <strong>the</strong> typical energy scale <strong>of</strong> nuclear processes.<br />
2.2.2 Thermal Cooling<br />
In addition to <strong>the</strong> neutrinos directly produced by nuclear fusion, <strong>the</strong> extreme temperatures<br />
<strong>and</strong> densities inside stars allow <strong>the</strong> emmitance <strong>of</strong> <strong>the</strong>rmal or cooling neutrinos by vari-<br />
12
{1{<br />
2.2 Neutrino Production Processes<br />
10 58<br />
pp chain<br />
CNO cycle<br />
10 57<br />
Neutrinos<br />
10 56<br />
10 55<br />
− − ignition <strong>of</strong> He burning<br />
____ end <strong>of</strong> He burning<br />
10 54<br />
1 10<br />
M/Mo<br />
Figure 2.5: Total number <strong>of</strong> neutrinos produced until <strong>the</strong> ignition (dashed line) <strong>and</strong> until <strong>the</strong><br />
end (solid line) <strong>of</strong> <strong>the</strong> helium burning phase in population I stars as a function <strong>of</strong><br />
<strong>the</strong> initial star mass in solar masses.[13]<br />
ous processes.[15] The resulting neutrinos’ energies directly depend on <strong>the</strong> temperature.<br />
Silicon burning is <strong>the</strong> last <strong>and</strong> hottest fusion phase <strong>and</strong> takes place at about 3.5·10 6 K;<br />
stars below about 4 M ⊙ never reach <strong>the</strong> last fusion phases but instead gravitationally<br />
collapse to white dwarfs when <strong>the</strong>ir fuel is burnt <strong>and</strong> heat up to well over 100·10 6 K.[16]<br />
Proto-neutron stars’ core temperatures can approach 1·10 12 K <strong>for</strong> a couple <strong>of</strong> seconds,<br />
quickly cooling down to below 100 GK. Using this in<strong>for</strong>mation <strong>and</strong> <strong>the</strong> Boltzmann constant<br />
k B = 86.17 eV/MK, <strong>the</strong> cooling neutrinos’ energies can be estimated to almost never<br />
exceed several MeV.<br />
2.2.3 Cosmic Ray Interactions<br />
The most probable source <strong>of</strong> high-energy (> 1GeV) neutrinos are interactions <strong>of</strong> cosmic<br />
rays. Various reaction channels exist, among <strong>the</strong> most important are those with charged<br />
pion production<br />
p + p → X + π +<br />
p + p → n + ∆ ++ → n + p + π +<br />
p + p → p + ∆ + → p + n + π +<br />
p + n → p + ∆ 0 → p + p + π −<br />
p + γ → ∆ + → n + π + 13
2 NEUTRINO ASTROPHYSICS<br />
<strong>and</strong> <strong>for</strong> higher energies those with kaon production.<br />
The resulting mesons <strong>the</strong>mselves decay quickly by<br />
K + → µ + + ν µ (63.55%)<br />
K + → π + + π 0 (20.66%)<br />
K + → π + + π + + π − (5.59%)<br />
K + → π 0 + e + + ν e (5.07%)<br />
K 0 S → π + + π − (69.20%)<br />
K 0 L → π ± + e ∓ + ( ν ) e (40.55%)<br />
K 0 L → π ± + µ ∓ + ( ν ) µ (27.04%)<br />
K 0 L → π + + π − + π 0 (12.54%)<br />
π + → µ + + ν µ (99.99%).<br />
Only decay modes with likely neutrino output <strong>and</strong> branching ratio ≥ 5% have been listed,<br />
modes <strong>for</strong> negatively charged particles exist correspondingly.[17] The neutrino production<br />
efficiency is lower in regions <strong>of</strong> high density because <strong>the</strong> particles might interact with o<strong>the</strong>r<br />
particles be<strong>for</strong>e <strong>the</strong>y decay.<br />
The decay <strong>of</strong> charged pions to electrons is highly suppressed in favor <strong>of</strong> <strong>the</strong> decay to<br />
muons because <strong>the</strong> latters’ helicity can be changed by boosts more easily. The muons<br />
almost exclusively decay into electrons by µ + → e + + ν e + ν µ . Tau neutrino production<br />
is dominated by <strong>the</strong> decay <strong>of</strong> D s<br />
± mesons, which <strong>the</strong>mselves have a low production crosssection.<br />
There<strong>for</strong>e, <strong>the</strong> ratios between <strong>the</strong> expected neutrino flavors near <strong>the</strong> interaction<br />
point can be roughly estimated to be ( ( ν ) e : ( ν ) µ : ( ν ) τ) = (1 : 2 : 0).<br />
2.3 Air Showers<br />
High energy cosmic rays <strong>and</strong> gamma rays which enter <strong>the</strong> Earth’s atmosphere generate<br />
a large amount <strong>of</strong> particles in <strong>for</strong>m <strong>of</strong> atmospheric cascades called air showers, see figure<br />
2.6. Particularly important are atmospheric muons <strong>and</strong> atmospheric neutrinos. Their<br />
underst<strong>and</strong>ing is vital <strong>for</strong> all neutrino telescopes; <strong>the</strong>y pose both signal <strong>for</strong> some analyses<br />
<strong>and</strong> <strong>the</strong> major background source <strong>for</strong> <strong>the</strong> o<strong>the</strong>rs.<br />
The reactions discussed in section 2.2.3 can also occur in <strong>the</strong> upper atmosphere. The<br />
energy spectrum <strong>for</strong> atmospheric muons <strong>and</strong> neutrinos follows a power law <strong>of</strong> E −3.7 at<br />
lower energies; <strong>the</strong> reason is that a meson’s probability to interact with <strong>the</strong> atmosphere<br />
instead <strong>of</strong> decaying increases with <strong>the</strong> meson’s lifetime τ = γτ 0 ∝ E. At high energies a<br />
prompt decay component from charmed mesons supervenes, hardening <strong>the</strong> spectrum to<br />
<strong>the</strong> primary cosmic ray spectrum <strong>of</strong> E −2.7 (figure 2.7).[19]<br />
14
339298621 5634138942<br />
01234 563416249522.3 Air Showers<br />
Figure 2.6: Two extraterrestial muon neutrinos (dashed lines, entering from <strong>the</strong> left) <strong>and</strong> two<br />
cosmic rays hitting <strong>the</strong> Earth. One neutrino passes undetectable, <strong>the</strong> o<strong>the</strong>r interacts<br />
inside <strong>the</strong> detector. The upper air shower produces an atmospheric neutrino<br />
which is detected <strong>and</strong> an atmospheric muon which is absorbed by <strong>the</strong> Earth, <strong>the</strong><br />
lower air shower’s atmospheric muon is detected.<br />
±<br />
p<strong>and</strong> K decay<br />
±<br />
p<strong>and</strong> K decay<br />
secondary decay<br />
prompt charmed<br />
meson decay<br />
prompt charmed<br />
meson decay<br />
secondary decay<br />
Figure 2.7: Vertical atmospheric muon (left) <strong>and</strong> muon neutrino (right) differential fluxs<br />
(including anti-particles) weighted with E −3 . Dashed lines indicate alternative<br />
models.[18]<br />
15
2 NEUTRINO ASTROPHYSICS<br />
2.3.1 Atmospheric Muons<br />
Events Passing<br />
10 8 40<br />
10 7<br />
35<br />
10 6<br />
30<br />
10 5<br />
25<br />
10 4<br />
20<br />
15<br />
10 3<br />
10<br />
10 2<br />
5<br />
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 0 -1 -0.9 -0<br />
Despite <strong>the</strong>ir short mean lifetime <strong>of</strong> about<br />
2.2 µs, high-energetic muons produced in<br />
<strong>the</strong> upper atmosphere at a height <strong>of</strong> typically<br />
20 km can reach <strong>the</strong> Earth’s surface<br />
because <strong>of</strong> <strong>the</strong> time dilation or length contraction<br />
(depending on <strong>the</strong> frame <strong>of</strong> reference)<br />
explained by special relativity. With<br />
<strong>the</strong>ir high penetration potential which is<br />
fur<strong>the</strong>r discussed in section 3.2.2 <strong>the</strong>y are<br />
a major component <strong>of</strong> <strong>the</strong> background <strong>for</strong><br />
neutrino telescopes. This background can<br />
be significantly reduced by “looking downwards”,<br />
i. e. zenith angle cuts, as muons –<br />
cos(θ)<br />
in contrast to neutrinos – are not able to<br />
pass more than a few kilometers through Figure 2.8: Zenith distribution <strong>of</strong> atmospheric<br />
ÙÖ ÄØ Ì muons ÞÒØ (solid) ÒÐ <strong>and</strong>×ØÖÙØÓÒ atmospheric Ó ÅÆ ØÖ<br />
<strong>the</strong> Earth; see figure 2.8.<br />
ÖÓÑ ÓÛÒÓÒ Ó×Ñ muon neutrinos ÖÝ ÑÙÓÒ׺ (dashed), Ì × simulated<br />
ÒÐ<strong>for</strong> ×ØÖÙØÓÒ IceCube’s Óprecursor<br />
ÙÔÛÖ ÖÓÒ×ØÖÙØ <br />
ÐÒ ×ÓÛ× ØÖ<br />
ÊØ Ì ÞÒØ<br />
2.3.2 Atmospheric Neutrinos<br />
ÒØ× Ø ×ØØ×ØÐ AMANDA.[20] ÔÖ×ÓÒ Ó Ø ØÑÓ×ÔÖ ÒÙØÖÒÓ ×Ñ<br />
For most analyses, atmospheric neutrinos are background. Except <strong>for</strong> <strong>the</strong>ir energy spectrum<br />
which differs from <strong>the</strong> expected signal spectrum, <strong>the</strong>y are fundamentally indistinguishable<br />
from neutrinos <strong>of</strong> extraterrestial origin. However, down-going atmospheric<br />
neutrinos can be tagged by searching <strong>for</strong> coincident muons in <strong>the</strong> detector.[21]<br />
10 4 Data<br />
Atmospheric MC<br />
10 3<br />
Downgoing MC<br />
10 2<br />
10<br />
1<br />
0 2.5 5 7.5 10 12.5 15 17.5 20 22.5 25<br />
Analyses which involve atmospheric neutrinos as signal include neutrino oscillation<br />
studies. For those, atmospheric neutrinos are <strong>of</strong> special interest because <strong>of</strong> <strong>the</strong>ir long<br />
baselines (up to <strong>the</strong> diameter <strong>of</strong> Earth) compared to reactor neutrinos or artificial neutrino<br />
beams from particle accelerators, making <strong>the</strong>se measurements complementary.<br />
10 -1<br />
Quality Cut<br />
Data / MC<br />
6<br />
5<br />
4<br />
3<br />
2<br />
1<br />
0<br />
0 2.5<br />
ÙÖ ÚÒØ ÕÙÐØݺ ÄØ È××Ò ÖØ× Ó ÚÒØ× ÓÚ <br />
ÖÓÙÒ Å¸ ØÑÓ×ÔÖ ÒÙØÖÒÓ Å Ò ÜÔÖÑÒØРغ<br />
ÖØÓ Ø»Åº<br />
16<br />
¿¼
CHAPTER III<br />
Neutrino Detection
3 NEUTRINO DETECTION<br />
Figure 3.1: Cross-sections <strong>for</strong> scattering <strong>of</strong> neutrinos (solid) <strong>and</strong> antineutrinos (dashed) from<br />
isoscalar nucleons 1 at energies between 10 3 GeV <strong>and</strong> 10 12 GeV. The results are<br />
based on HERA particle distribution function measurements <strong>and</strong> have been extrapolated<br />
above 10 7 GeV (CTEQ4-DIS).[22][23]<br />
3.1 Neutrino Interactions<br />
Neutrinos are electrically neutral leptons with spin 1 . They only interact via weak <strong>and</strong><br />
2<br />
gravitational <strong>for</strong>ces <strong>and</strong> come in three weak eigenstates called flavors, named after <strong>the</strong><br />
charged lepton with which <strong>the</strong> respective neutrino eigenstate can interact by a charged<br />
current interaction.<br />
According to <strong>the</strong> St<strong>and</strong>ard Model, neutrinos are massless particles; however it is widely accepted<br />
that this is not accurate since many experiments verified <strong>the</strong> existence <strong>of</strong> neutrino<br />
oscillations, which can occur because <strong>the</strong> neutrino mass eigenstates are pairwise different<br />
<strong>and</strong> not congruent with <strong>the</strong> weak eigenstates.<br />
Because <strong>of</strong> <strong>the</strong>ir low cross-sections, neutrinos can travel over huge distances through ordinary<br />
matter almost without attenuation up to very high energies (figure 3.1). Not until<br />
about 10 6 GeV, <strong>the</strong> Earth’s core starts to become opaque to neutrinos.<br />
The relevant processes involving neutrinos are neutral current (NC) <strong>and</strong> charged current<br />
(CC) interactions.<br />
Neutral current interactions – i.e., interactions which are mediated by a Z 0 boson – are<br />
1 The particle distribution function (PDF) <strong>of</strong> an isoscalar nucleon (an hadronic single state with isospin<br />
I = I 3 = 0) is <strong>the</strong> arithmetic mean between those <strong>of</strong> a proton (I 3 = + 1 2 ) <strong>and</strong> a neutron (I 3 = − 1 2<br />
), or<br />
half <strong>the</strong> PDF <strong>of</strong> a deuteron, which is <strong>the</strong> true isospin singlet. As we can neglect electrical charges, it is a<br />
good approximation <strong>of</strong> <strong>the</strong> average nucleon PDF.<br />
18
3.2 Lepton Propagation<br />
<strong>the</strong> same <strong>for</strong> <strong>the</strong> three flavors <strong>and</strong> constitute about 10% <strong>of</strong> <strong>the</strong> charged current crosssections.[22]<br />
The most important NC interactions are those with <strong>the</strong> nucleons N because<br />
<strong>the</strong>y have much larger cross-sections than interactions with <strong>the</strong> matter’s electrons:<br />
ν a + N → ν a + X, a ∈ {e, µ, τ}<br />
X is <strong>the</strong> hadronic rest <strong>of</strong> <strong>the</strong> nucleon which will trigger a hadronic cascade. The neutrino<br />
escapes with a large fraction <strong>of</strong> its initial energy.<br />
Charged current interactions are mediated by W ± bosons <strong>and</strong> differ <strong>for</strong> <strong>the</strong> different flavors<br />
mainly in <strong>the</strong> produced leptons a ± . As above, interactions with <strong>the</strong> nucleons make up <strong>the</strong><br />
largest part:<br />
ν a + N → a − + X<br />
ν a + N → a + + X<br />
The initial direction <strong>of</strong> <strong>the</strong> lepton is well aligned with <strong>the</strong> neutrino’s track; <strong>the</strong> average<br />
angle mismatch can be estimated as ψ = 0.7° · ( )<br />
10 3 GeV 0.7.[24]<br />
E ν<br />
For both NC <strong>and</strong> CC up to about 10 7 GeV, <strong>the</strong> cross-sections <strong>for</strong> antineutrinos are below<br />
<strong>the</strong>ir neutrino counterparts because <strong>of</strong> <strong>the</strong> valence quark distribution. This effect becomes<br />
negligible <strong>for</strong> higher energies where <strong>the</strong> PDFs are dominated by sea quarks (<strong>and</strong> gluons,<br />
which, however, do not interact via <strong>the</strong> weak <strong>for</strong>ce).<br />
Gravitational effects can largely be ignored. Highly relativistic neutrinos independent<br />
<strong>of</strong> <strong>the</strong>ir mass are subject to gravitational lensing just as photons; in fact, <strong>the</strong>y can even<br />
be lensed stronger by certain objects as <strong>the</strong>y are able to pass most electromagnetically<br />
opaque regions without absorption.[25]. However, in most cases <strong>the</strong> angle <strong>of</strong> deflection is<br />
very small compared to <strong>the</strong> angular resolution <strong>of</strong> current neutrino telescopes.<br />
The probability <strong>for</strong> a detector signal being caused by direct gravitational interactions is<br />
negligible in <strong>the</strong> St<strong>and</strong>ard Model but increases under assumptions such as <strong>the</strong> existence<br />
<strong>of</strong> compact extra dimensions. In <strong>the</strong> latter case, one could expect an increased ratio <strong>of</strong><br />
elastic scattering (at impact parameters b larger than <strong>the</strong> Schwarzschild radius r S ) <strong>and</strong><br />
cascades without primary lepton through black-hole creation (at b < r S ).[26]<br />
3.2 Lepton Propagation<br />
The charged lepton produced in charged current interactions is <strong>the</strong> most important signature<br />
<strong>for</strong> most types <strong>of</strong> neutrino detectors. Common to all flavors is <strong>the</strong> hadronic cascade at<br />
<strong>the</strong> interaction point, which contains roughly 40% <strong>of</strong> <strong>the</strong> neutrino energy. The signatures<br />
<strong>of</strong> <strong>the</strong> primary leptons inside <strong>the</strong> detector vary <strong>for</strong> <strong>the</strong> different flavors (see figure 3.2).<br />
19
3 NEUTRINO DETECTION<br />
hadronic<br />
cascade<br />
electromagnetic<br />
cascade<br />
Èerenkov radiation<br />
<br />
e <br />
<br />
Figure 3.2: Sketch illustrating <strong>the</strong> signatures <strong>of</strong> charged current interactions <strong>of</strong> neutrinos in<br />
Čerenkov media as described in sections 3.2 <strong>and</strong> 3.3. Note that every hadronic<br />
cascade also has an electromagnetic component.<br />
Figure 3.3: Depth development <strong>of</strong> electron initiated cascades in numbers <strong>of</strong> particles vs. depth<br />
in units <strong>of</strong> radiation lengths in ice. Shown are simulated curves <strong>for</strong> <strong>the</strong> sum <strong>of</strong><br />
electrons <strong>and</strong> positrons (blue), <strong>the</strong> excess <strong>of</strong> electrons (red), <strong>and</strong> <strong>the</strong>ory predictions<br />
(dashed, Approximation-B) at 1, 10, <strong>and</strong> 100 TeV (down to up).[27]<br />
20
3.2 Lepton Propagation<br />
Figure 3.4: Effective radiation length λ ∼ = X LPM modified by <strong>the</strong> LPM effect.[28]<br />
3.2.1 Electron Propagation<br />
Electrons in dense matter quickly lose energy through bremsstrahlung. The generated<br />
photons convert to <strong>new</strong> electrons <strong>and</strong> positrons via pair production, resulting in an exponential<br />
growth <strong>of</strong> <strong>the</strong> particle number until <strong>the</strong> photons’ energies fall below 2m e <strong>and</strong> <strong>the</strong><br />
leptons’ energy falls below <strong>the</strong> threshold at which ionization losses become dominating,<br />
called critical energy. For ice, this cricital energy is E e crit = 81 MeV.[29] Since <strong>the</strong> photon<br />
<strong>and</strong> electron free mean paths are <strong>of</strong> <strong>the</strong> same order <strong>of</strong> magnitude, <strong>the</strong> shower depth X at<br />
a given initial energy E 0 up to about 1 PeV can be approximated by<br />
X = X 0 log 2<br />
( E0<br />
E crit<br />
)<br />
with <strong>the</strong> radiation length X 0 (36.08 g/cm 2 in ice). As <strong>the</strong> cascade is highly boosted<br />
<strong>and</strong> <strong>the</strong>re<strong>for</strong>e directed, its length can be estimated to be L = 2Xρ, where ρ is <strong>the</strong><br />
density <strong>of</strong> <strong>the</strong> material. ( For deep Antarctic ice (ρ = 0.924 g/cm 3 ), this evaluates to<br />
E e<br />
)<br />
L = 0.67 m log 0<br />
2 81 MeV , which is <strong>of</strong> <strong>the</strong> order <strong>of</strong> 10 m.<br />
Starting at about 1 PeV <strong>the</strong> radiation length X 0 has to be replaced by <strong>the</strong> effective<br />
radiation length X LPM to take account <strong>for</strong> <strong>the</strong> LPM effect: The <strong>for</strong>mation length <strong>for</strong><br />
bremsstrahlung <strong>and</strong> pair production interactions increases with <strong>the</strong> particle’s energy, resulting<br />
in <strong>the</strong> suppression <strong>of</strong> <strong>the</strong>se interactions <strong>and</strong> <strong>the</strong>re<strong>for</strong>e in significantly longer cascades<br />
(about 200 m at 10 EeV, see figure 3.4).[28]<br />
3.2.2 Muon Propagation<br />
Because <strong>of</strong> <strong>the</strong> muon’s higher mass m µ = 207m e , its energy loss due to bremsstrahlung is<br />
much smaller. Up to <strong>the</strong> critical energy (E µ crit ≈ 500 GeV in ice) ionization losses are most<br />
important, at higher energies electron pair production <strong>and</strong> radiation losses dominate. The<br />
21
3 NEUTRINO DETECTION<br />
expected track length l/km<br />
50<br />
40<br />
30<br />
20<br />
10<br />
muon<br />
tau<br />
0<br />
10 2 10 3 10 4 10 5 10 6 10 7 10 8 10 9<br />
initial energy E 0 /GeV<br />
Figure 3.5: Expected track length <strong>for</strong> muons <strong>and</strong> tau particles in Antarctic ice according to<br />
<strong>the</strong> models explained in section 3.2.1 <strong>and</strong> section 3.2.2.<br />
total energy loss is approximately given by<br />
− dE<br />
dx<br />
= a + bE, (3.1)<br />
where a(E) is <strong>the</strong> ionization loss given by <strong>the</strong> Be<strong>the</strong>-Bloch equation <strong>and</strong> b(E)E a parametrization<br />
<strong>for</strong> <strong>the</strong> higher-energy processes; both parameters are fairly constant <strong>for</strong><br />
E > E µ crit.[30] For ice, this yields approximately<br />
− dE<br />
dx = 0.25 GeV/m + 4.3·10−4 E/m, [31]<br />
leading to an expected track length <strong>of</strong> L = 2300 m · ln ( )<br />
1 + E 0<br />
E (figure 3.5), which is<br />
µ<br />
crit<br />
small compared to <strong>the</strong> length a muon can travel in its mean decay time at its critical<br />
energy, l µ ≈ 3 Mm. Since <strong>the</strong> energy losses are stochastical, actual values vary, <strong>and</strong> <strong>of</strong>ten<br />
much energy is lost at once, triggering a cascade along <strong>the</strong> muon track. The scattering <strong>of</strong><br />
high-energy muons is negligible.[31]<br />
3.2.3 Tau Propagation<br />
Much heavier than muons (m τ = 16.8m µ ), tau particles lose even less energy when propagating<br />
through matter. Equation (3.1) holds <strong>for</strong> <strong>the</strong>m as well, with a being very similar<br />
<strong>and</strong> b ≈ 3.6·10 −5 /m.[31] Because <strong>of</strong> <strong>the</strong> tau’s short mean lifetime τ τ = 0.29 ps, <strong>the</strong>ir detector<br />
signature is determined by <strong>the</strong> hadronic cascade at <strong>the</strong> neutrino interaction point<br />
<strong>and</strong> <strong>the</strong> decay <strong>of</strong> <strong>the</strong> tau which usually triggers a second cascade. The branching ratio <strong>for</strong><br />
τ − → µ − + ν µ + ν τ is 17.4%, <strong>the</strong> remaining decays spawn at least one electron or charged<br />
22
3.3 Čerenkov Detectors<br />
C<br />
Figure 3.6: Sketch illustrating <strong>the</strong> emission <strong>of</strong> Čerenkov radiation (blue) by a superluminal<br />
particle (track in red).<br />
meson.[17] The interaction points are seperated by <strong>the</strong> tau decay length l τ = 49 m · E0<br />
PeV<br />
(figure 3.5). Up to a few PeV, <strong>the</strong> resulting cascades overlap <strong>and</strong> are indistinguishable.<br />
For higher energies, both cascades are seperated <strong>and</strong> connected by a faint Čerenkov track,<br />
<strong>for</strong>ming a characteristic signature called double bang (illustrated in figure 3.2). If one <strong>of</strong><br />
<strong>the</strong>se cascades lies outside <strong>the</strong> detector volume, <strong>the</strong> resulting signature is called lollipop<br />
(resp. inverted lollipop).<br />
3.3 Čerenkov Detectors<br />
One <strong>of</strong> <strong>the</strong> most effective methods to detect neutrinos is <strong>the</strong> use <strong>of</strong> Čerenkov radiation.<br />
Čerenkov radiation is emitted whenever a charged particle travels through dielectric matter<br />
(<strong>the</strong> Čerenkov medium) at a speed v = βc higher than <strong>the</strong> material’s phase speed <strong>of</strong><br />
light c m = c , where n denotes <strong>the</strong> material’s refractive index (n = 1.32 <strong>for</strong> ice). The<br />
n<br />
particle travelling through <strong>the</strong> medium polarizes <strong>the</strong> surrounding atoms, which emit light<br />
when <strong>the</strong>y return into equilibrium. If <strong>the</strong> speed <strong>of</strong> <strong>the</strong> particle exceeds c m , <strong>the</strong> interferences<br />
become constructive <strong>and</strong> a wave front <strong>of</strong> light is sent out at a distinctive angle<br />
θ C = arccos (n −1 β −1 ) called Čerenkov angle (see figure 3.6). The light’s spectrum is continous<br />
<strong>and</strong> its intensity is approximately linear in frequency, resulting in a blue appearance<br />
to <strong>the</strong> human eye. At x-ray frequencies <strong>the</strong> spectrum is cut-<strong>of</strong>f as <strong>the</strong> refraction index<br />
becomes smaller than one.<br />
The rate <strong>of</strong> emitted photons can be calculated with <strong>the</strong> simplified Frank-Tamm <strong>for</strong>mula:<br />
∫<br />
dN<br />
dx = λmax<br />
2παz 2 (<br />
1 − 1 )<br />
dλ, (3.2)<br />
λ min λ 2 n 2 β 2<br />
where α ≈ 1 is <strong>the</strong> fine-structure constant <strong>and</strong> z <strong>the</strong> particle’s electric charge. The<br />
137<br />
wavelength range given by λ min . . . λ max is determined by <strong>the</strong> optical properties <strong>of</strong> <strong>the</strong><br />
23
3 NEUTRINO DETECTION<br />
Čerenkov medium.<br />
Čerenkov media have to be transparent <strong>for</strong> blue to ultraviolet light. Additionally, <strong>for</strong><br />
neutrino detection <strong>the</strong>y need to have reasonably high density <strong>and</strong> be available cheaply in<br />
large quantities. Air is not sufficiently dense, so neutrino interactions in air are ra<strong>the</strong>r<br />
unlikely; it can still be used <strong>for</strong> air shower detectors. H 2 O, both liquid or solid, is well<br />
suited <strong>for</strong> large detectors. Deep <strong>and</strong> clear lake or sea water generally involves less scattering<br />
(up to a factor <strong>of</strong> 10[32]) but higher absorption (about factor 2) than bubble free ice.<br />
Additional challenges <strong>for</strong> water neutrino telescopes are <strong>the</strong> variable environment <strong>and</strong> a<br />
strong background <strong>of</strong> bioluminescence from various life<strong>for</strong>ms as well as radioactive decays<br />
<strong>of</strong> unstable isotopes such as 40 K.[33][34] The optical properties <strong>of</strong> deep Antarctic ice will<br />
be examined in section 4.2.<br />
24
CHAPTER IV<br />
The IceCube Neutrino Observatory
4 THE ICECUBE NEUTRINO OBSERVATORY<br />
4.1 Layout<br />
The IceCube Neutrino Observatory is a combined air shower <strong>and</strong> neutrino detector located<br />
at <strong>the</strong> Amundsen-Scott South Pole Station at <strong>the</strong> geographic South Pole. It consists <strong>of</strong><br />
three main parts, InIce, DeepCore, <strong>and</strong> IceTop (figure 4.2):<br />
InIce is an underground Čerenkov detector which uses about one cubic kilometer <strong>of</strong><br />
Antarctic ice as neutrino interaction <strong>and</strong> Čerenkov medium, making it <strong>the</strong> World’s largest<br />
neutrino detector at <strong>the</strong> time <strong>of</strong> writing. In its final configuration, it will consist <strong>of</strong> 80<br />
long cables called strings with 60 digital optical modules (DOMs) each. The strings are<br />
lowered into hot-water-drilled holes in <strong>the</strong> ice sheet, with <strong>the</strong> lowest DOMs positioned<br />
around 2450 meters below <strong>the</strong> surface <strong>and</strong> a vertical spacing <strong>of</strong> about 17 m, adding up<br />
to 1000 m <strong>of</strong> instrumented length. After deployment <strong>of</strong> each string, <strong>the</strong> water inside <strong>the</strong><br />
hole refreezes, making recovery <strong>and</strong> maintenance impossible but also optically coupling<br />
<strong>the</strong> DOM to <strong>the</strong> surrounding ice. The horizontal spacing between <strong>the</strong> strings is approximately<br />
125 m in a hexagonal pattern, spanning approximately one square kilometer across<br />
(figure 4.1). Because <strong>of</strong> this geometry <strong>and</strong> <strong>the</strong> ice properties elaborated in section 4.2,<br />
InIce’s lower energy threshold is about 100 GeV.<br />
Figure 4.1: IceCube surface geometry with distances between <strong>the</strong> strings; <strong>the</strong> lighter a circle’s<br />
color, <strong>the</strong> earlier <strong>the</strong> string has been deployed. Black circles are strings scheduled<br />
<strong>for</strong> deployment in summer 2010/2011, small dots are IceTop tanks.[35]<br />
26
4.1 Layout<br />
Figure 4.2: Three-dimensional sketch <strong>of</strong> IceCube, with DeepCore marked in green. Also shown<br />
are <strong>the</strong> dust layer at (2000 . . . 2100) m depth (see section 4.2), <strong>the</strong> ground below<br />
<strong>the</strong> ice sheet at about 2850 m depth, <strong>and</strong> an image <strong>of</strong> <strong>the</strong> Eiffel Tower <strong>for</strong> size<br />
comparison.[36]<br />
27
4 THE ICECUBE NEUTRINO OBSERVATORY<br />
DeepCore is a low-energy extension to this detector. Located at its lower center, it consists<br />
<strong>of</strong> 6 extra strings with 60 high-efficiency DOMs each (section 4.3). In contrast to <strong>the</strong><br />
st<strong>and</strong>ard strings, <strong>the</strong>y are more densly instrumented in two groups <strong>of</strong> 10 <strong>and</strong> 50 DOMs<br />
each with an inner-group spacing <strong>of</strong> 10 m <strong>and</strong> 7 m respectively, <strong>and</strong> a gap <strong>of</strong> 257 m between<br />
<strong>the</strong> two groups to take advantage <strong>of</strong> <strong>the</strong> clear ice in <strong>the</strong>se regions (section 4.2). These<br />
strings have been deployed in a half-sized hexagon centered at st<strong>and</strong>ard string 36. The<br />
two st<strong>and</strong>ard strings 79 <strong>and</strong> 80 have not been deployed at <strong>the</strong>ir original positions <strong>and</strong> have<br />
been proposed to be installed inside <strong>the</strong> DeepCore volume to make it even more densly<br />
instrumented. The energy threshold <strong>of</strong> <strong>the</strong> combined InIce+DeepCore detector will be<br />
around 10 GeV. In its task as low-energy extension, DeepCore supersedes IceCube’s now<br />
decomissioned precursor, <strong>the</strong> Antarctic Muon And Neutrino Detector Array (AMANDA),<br />
which took data from 1994 to 2009 <strong>and</strong> which was integrated into IceCube in 2005.<br />
IceTop is an air shower Čerenkov detector which in its final configuration will comprise<br />
<strong>of</strong> 160 polyethylene tanks at <strong>the</strong> ice sheet’s surface which encompass each a volume <strong>of</strong><br />
2.5 m 3 <strong>of</strong> bubble-free ice observed by two DOMs. It will be capable <strong>of</strong> reliably detecting<br />
air showers with a primary particle energy threshold <strong>of</strong> about 150 TeV.[37] Besides conducting<br />
interesting experiments by itself, it is planned to use it <strong>for</strong> background rejection<br />
in neutrino analyses.<br />
For this <strong>the</strong>sis InIce <strong>and</strong> DeepCore <strong>for</strong>m <strong>the</strong> relevant part <strong>of</strong> IceCube as IceTop uses<br />
its own feature extraction algorithms. There<strong>for</strong>e “InIce+DeepCore” <strong>and</strong> “IceCube” will<br />
be used interchangeably in <strong>the</strong> following chapters if not stated o<strong>the</strong>rwise.<br />
In <strong>the</strong> Antartic winter, temperatures regularly fall below −65 °C, rendering impossible<br />
<strong>the</strong> operation <strong>of</strong> most machines. Because <strong>of</strong> this, IceCube had different string configurations<br />
<strong>for</strong> data-taking during <strong>the</strong> months in which deployment had to be suspended.<br />
Construction <strong>of</strong> Icecube began with one experimental string during 2005, followed by eight<br />
additional ones in <strong>the</strong> 2005-2006 Antarctic summer. This configuration – following <strong>the</strong><br />
same naming scheme as <strong>the</strong> later ones – was called IC9. Toge<strong>the</strong>r with <strong>the</strong> thirteen strings<br />
deployed during <strong>the</strong> summer <strong>of</strong> 2006-2007, it <strong>for</strong>med IC22. During <strong>the</strong> next summer eighteen<br />
strings were added, resulting in <strong>the</strong> detector configuration IC40. In <strong>the</strong> summer<br />
between 2008 <strong>and</strong> 2009 eighteen more st<strong>and</strong>ard strings <strong>and</strong> one DeepCore string were<br />
deployed (IC58+1 or IC59), as well as twenty during <strong>the</strong> 2009-2010 deployment season<br />
(IC73+6 or IC79). With DeepCore already completed, <strong>the</strong> final detector will be fully<br />
available in 2011.<br />
4.2 Ice Properties<br />
IceCube’s <strong>and</strong> especially DeepCore’s geometry are heavily influenced by <strong>the</strong> properties<br />
<strong>of</strong> <strong>the</strong> Antarctic ice. The experiment’s location was chosen because <strong>of</strong> <strong>the</strong> deepness <strong>and</strong><br />
purity <strong>of</strong> <strong>the</strong> naturally available ice <strong>and</strong> because <strong>of</strong> <strong>the</strong> good infrastructure provided by<br />
28
4.2 Ice Properties<br />
Figure 4.3: Inverse effective scattering length (left) <strong>and</strong> inverse absorption length (right) vs<br />
depth in deep Antarctic ice measured with pulsed sources at different wavelengths.<br />
A, B, C <strong>and</strong> D denote major dust peaks. Values <strong>for</strong> deep ice (below D) might be<br />
overestimated, see section 4.2.[39]<br />
<strong>the</strong> Amundsen-Scott South Pole Station.[38] The two most important optical properties<br />
are <strong>the</strong> effective scattering length λ e <strong>and</strong> <strong>the</strong> absorption length λ a .<br />
The scattering length λ s <strong>for</strong> IceCube is defined as <strong>the</strong> mean free path between scattering<br />
processes. The average angle between <strong>the</strong> unscattered path <strong>and</strong> <strong>the</strong> scattered path is<br />
〈cos(θ)〉 ≈ 0.94 ≫ 0. Because <strong>of</strong> this, <strong>the</strong> effective scattering length is defined as<br />
λ e := λ s<br />
n ∑<br />
i=0<br />
〈cos(θ)〉 i<br />
n→∞<br />
−−−−→<br />
λ s<br />
1 − 〈cos(θ)〉 .<br />
This is an estimate <strong>for</strong> <strong>the</strong> length after which a beam <strong>of</strong> light has become fairly isotropic.[39]<br />
The scattering inside <strong>the</strong> upper ice layers is dominated by microscopic trapped air bubbles.<br />
Because <strong>of</strong> <strong>the</strong> enormous pressure at high depths, <strong>the</strong> bubbles are compressed steadily until<br />
at about 1400 m scattering at dust particles begins to prevail (figure 4.3). Below, <strong>the</strong>re<br />
are regions <strong>of</strong> higher dust concentrations which have been designated as dust peaks A to<br />
D. The same structures have been identified by ice core measurements in o<strong>the</strong>r regions<br />
<strong>and</strong> correlate well with cold periods (stadials) during <strong>the</strong> current interglacial period <strong>of</strong><br />
<strong>the</strong> current ice age. The strongest <strong>of</strong> those peaks – peak D, also known as <strong>the</strong> dust layer<br />
– is approximately 65000 years old <strong>and</strong> features very short effective scattering lengths <strong>of</strong><br />
less than 10 m; because <strong>of</strong> this, DeepCore was designed with DOMs above <strong>and</strong> below this<br />
region, but without DOMs within.<br />
The absorption length – defined as <strong>the</strong> length after which <strong>the</strong> light intensity has<br />
dropped by a factor <strong>of</strong> e −1 – is significantly higher than <strong>the</strong> effective scattering length<br />
(about 100 m compared to about 30 m, see figure 4.3). At green wavelengths <strong>the</strong> main<br />
contribution is <strong>the</strong> absorption by <strong>the</strong> ice itself; at shorter wavelenghts dust particles<br />
dominate <strong>and</strong> <strong>the</strong> dust peaks can be identified accordingly.<br />
29
4 THE ICECUBE NEUTRINO OBSERVATORY<br />
The deep ice starting at 2100 m below <strong>the</strong> dust layer has not yet been studied as well<br />
as above because <strong>the</strong> ice properties were calculated mostly based on data measured with<br />
AMANDA. Only three AMANDA strings reached below <strong>the</strong> dust layer, <strong>the</strong> deepest being<br />
string 11 with a maximum depth <strong>of</strong> 2136 m. As well-defined light signals are needed <strong>for</strong><br />
this type <strong>of</strong> measurement, various calibrated light sources have been used: AMANDA<br />
optical modules (OM s) had diffusor balls attached which were coupled with a frequency<br />
doubled Nd-YAG laser at <strong>the</strong> surface; two nitrogen lasers were deployed near <strong>the</strong> center<br />
<strong>of</strong> <strong>the</strong> detector; eight OMs were equipped with flasher boards containing six blue LEDs<br />
each, <strong>and</strong> one string was equipped with ultraviolet LED flasher boards <strong>for</strong> every OM.<br />
Additionally to <strong>the</strong>se pulsed light sources, two continous sources with wider spectrum<br />
were deployed inside <strong>the</strong> ice.<br />
All IceCube DOMs contain similar flasher boards with twelve LEDs each, see section<br />
4.3. Recent IceCube measurements have shown that <strong>the</strong> deep ice might be clearer than<br />
previously thought, with λ e ≈ 50 m <strong>and</strong> λ a ≈ 200 m at blue wavelengths (405 nm).<br />
The resulting propagation length λ p acts as damping constant in <strong>the</strong> fluence development<br />
<strong>of</strong> light sources in <strong>the</strong> diffuse regime, i.e., after a few scattering processes:<br />
λ p :=<br />
√<br />
λe λ a<br />
3<br />
with<br />
F (d 5λ e ) ∝ 1 d e− d<br />
λp .[39]<br />
4.3 Digital Optical Modules<br />
The main task <strong>of</strong> IceCube’s DOMs is to capture Čerenkov light <strong>and</strong> to convert it into<br />
digitized wave<strong>for</strong>ms, which are <strong>the</strong>n sent to <strong>the</strong> surface. They consist <strong>of</strong> a photomultiplier<br />
tube (PMT), a flasher board, a mainboard containing most <strong>of</strong> <strong>the</strong> electronics, a delay<br />
board, <strong>and</strong> a high-voltage board, housed inside a 32.5 cm diameter glass pressure sphere.<br />
The photomultiplier is embedded into RTV (room temperature vulcanizing) silicone gel<br />
to optically couple it to <strong>the</strong> glass, <strong>and</strong> it is shielded against Earth’s magnetic field by a<br />
cage <strong>of</strong> mu-metal (figure 4.4).<br />
The flasher board lies in <strong>the</strong> DOM’s upper hemisphere with six pairs <strong>of</strong> blue (405 nm)<br />
LEDs mounted outwards in 60° intervals. Each pair consists <strong>of</strong> one LED on <strong>the</strong> lower side<br />
<strong>of</strong> <strong>the</strong> board mounted horizontally <strong>and</strong> one LED on <strong>the</strong> upper side tilted approximately<br />
40° upwards. As pointed out in section 4.2, <strong>the</strong> flashers can be used to determine or<br />
validate ice properties <strong>and</strong> to calibrate <strong>the</strong> detector.<br />
The high-voltage board is required to power <strong>the</strong> PMT. The PMT’s signal is decoupled<br />
by a toroidal trans<strong>for</strong>mer instead <strong>of</strong> a capacitor; <strong>the</strong> circuit’s low capacity <strong>of</strong> about 5 pF<br />
compared to <strong>the</strong> about 1 µF needed <strong>for</strong> a conventional capacitor coupling reduces <strong>the</strong> risk<br />
<strong>of</strong> discharges damaging <strong>the</strong> mainboard <strong>and</strong> delivers a more stable long-time signal quality<br />
as capacitors degrade over time. Fur<strong>the</strong>rmore, tests have proven <strong>the</strong> trans<strong>for</strong>mer coupling<br />
30
4.3 Digital Optical Modules<br />
Figure 4.4: Sketch <strong>of</strong> an IceCube Digital Optical Module (DOM).[40]<br />
to better filter out <strong>the</strong> power supply’s noises. However, as <strong>the</strong> trans<strong>for</strong>mer acts as high-pass<br />
filter, it causes droop: The signal’s voltage falls <strong>of</strong>f right after <strong>the</strong> initial peak, distorting<br />
<strong>the</strong> wave<strong>for</strong>m <strong>and</strong> making it undershoot <strong>the</strong> <strong>for</strong>mer baseline. This effect depends strongly<br />
both on <strong>the</strong> surrounding temperature T <strong>and</strong> on <strong>the</strong> toroid used. DOMs built be<strong>for</strong>e<br />
2006 – called old toroid DOMs or OT – are highly affected by droop, while <strong>new</strong> toroid<br />
DOMs (NT) are almost unaffected (figure 4.5). The individual toroid types are known<br />
<strong>and</strong> <strong>the</strong> temperatures can be measured with an on-board sensor; <strong>the</strong>re<strong>for</strong>e <strong>the</strong> droop can<br />
be corrected <strong>for</strong> during calibration (section 4.6.1), using <strong>the</strong> dual-tau parametrization[42]<br />
in which <strong>the</strong> DOM’s transient response ˜δ(t) to a signal δ(t) is modelled as<br />
˜δ(t) = δ(t) − N<br />
(<br />
4.3 e − t<br />
τ 1<br />
− 3.3 e − t<br />
τ 2<br />
)<br />
, τ 1 (T ) := t 1 +<br />
t 2<br />
, τ<br />
1 + e − T 2 := 0.75 τ 1 .<br />
Tc<br />
The parameters N, t 1 , t 2 , <strong>and</strong> T C have to be determined empirically. While <strong>the</strong> NT<br />
DOMs significantly reduce <strong>the</strong> droop problems, <strong>the</strong>y also possess differently shaped, wider<br />
wave<strong>for</strong>ms (figure 4.6), which is important <strong>for</strong> feature extraction, i. e., <strong>the</strong> analyzation <strong>of</strong><br />
<strong>the</strong> wave<strong>for</strong>ms captured by <strong>the</strong> DOMs, see section 5.<br />
The delay board features a high-quality serpentine circuit that delays <strong>the</strong> throughgoing<br />
signal <strong>for</strong> 75 ns. Only two <strong>of</strong> <strong>the</strong> three signal paths are delayed, a third path leads<br />
to <strong>the</strong> trigger system to decide whe<strong>the</strong>r to launch this DOM, i.e., to digitalize wave<strong>for</strong>m<br />
31
4 THE ICECUBE NEUTRINO OBSERVATORY<br />
Figure 4.5: Wave<strong>for</strong>ms illustrating <strong>the</strong> toroidal droop effect <strong>for</strong> OT (upper) <strong>and</strong> NT (lower)<br />
DOMs, showing an uncalibrated (raw) wave<strong>for</strong>m on <strong>the</strong> left <strong>and</strong> <strong>the</strong> corresponding<br />
droop-corrected one on <strong>the</strong> right. For <strong>the</strong> OT graphs 100 wave<strong>for</strong>ms have<br />
been averaged, <strong>the</strong> NT graphs show an individual wave<strong>for</strong>m; both were taken at<br />
−55 °C.[41]<br />
<strong>and</strong> to check <strong>for</strong> local coincidence, see section 4.4.3. The trigger threshold is typically set<br />
to 0.25 PE (photoelectron charges).<br />
From <strong>the</strong> remaining two paths, one is distributed among three amplifiers (×16, ×2 <strong>and</strong><br />
×0.25) which supply <strong>the</strong> analog transient wave<strong>for</strong>m digitizer (ATWD, section 4.4.1) with<br />
data, <strong>and</strong> <strong>the</strong> o<strong>the</strong>r one supplies <strong>the</strong> fast analog digital converter (FADC, section 4.4.2).<br />
Last but not least, <strong>the</strong> PMT itself is relevant <strong>for</strong> feature extraction as it is <strong>the</strong> component<br />
which generates <strong>the</strong> signal. It is a Hamamatsu R7081-02 with a diameter <strong>of</strong> 25 cm<br />
<strong>and</strong> a peak quantum efficiency <strong>of</strong> 25% around 410 nm. It houses ten dynodes leading to<br />
a gain factor <strong>of</strong> about 10 7 <strong>and</strong> has a dark count rate below 400 Hz between −60 °C <strong>and</strong><br />
0 °C, mostly caused by 40 K decay in <strong>the</strong> glass. The DeepCore high-efficiency DOMs differ<br />
in that <strong>the</strong>y have an improved photocathode, leading to a higher quantum efficiency <strong>of</strong><br />
up to 33%.[43]<br />
Three effects can lead to erroneous pulses: Photons may bypass <strong>the</strong> photocathode <strong>and</strong><br />
directly hit <strong>the</strong> first dynode ahead <strong>of</strong> <strong>the</strong> electrons, triggering an early electron cascade<br />
from <strong>the</strong>re on. Despite missing one multiplication process, <strong>the</strong>se prepulses exceed a signal<br />
32
4.4 Signal Digitization<br />
Table 4.1: Overview over different digitizers resp. I3Wave<strong>for</strong>m source types.<br />
source n bins T bin T total resolution U satu<br />
ns ns bit mV<br />
ATWD ch0 128 3.3 422 10 100<br />
ATWD ch1 128 3.3 422 10 800<br />
ATWD ch2 128 3.3 422 10 7500<br />
FADC 256 25 6400 10 70<br />
SLC 3 25 75 10 70<br />
strength <strong>of</strong> 0.25 PE <strong>for</strong> about 0.5% <strong>of</strong> all 50 PE pulses at 25 °C according to measurements<br />
by Hamamatsu.[44] They typically occur 32 ns prior to <strong>the</strong> real pulse.[45]<br />
Late pulses can occur if photoelectrons are backscattered at <strong>the</strong> first dynode. If <strong>the</strong>y hit<br />
<strong>the</strong> first dynode at <strong>the</strong> second approach, <strong>the</strong>y will trigger a pulse which lags behind by up<br />
to 66 ns, which is twice <strong>the</strong> transit time between photocathode <strong>and</strong> first dynode, or even<br />
more <strong>for</strong> multiple backscattering. Their typical rate is 1.5% <strong>for</strong> pulses as above.<br />
Finally, afterpulses are caused by ionized molecules <strong>of</strong> <strong>the</strong> remaining gas or luminescence<br />
photons from <strong>the</strong> dynodes hitting <strong>the</strong> photocathode. They typically occur <strong>for</strong> 2.0% <strong>of</strong><br />
<strong>the</strong> a<strong>for</strong>ementioned pulses <strong>and</strong> are frequently found a few microseconds after <strong>the</strong> real<br />
pulse.[44][46]<br />
4.4 Signal Digitization<br />
The two digitizers ATWD <strong>and</strong> FADC mentioned in section 4.3 act as different sources <strong>for</strong><br />
raw wave<strong>for</strong>ms. These raw wave<strong>for</strong>ms still show undesirable characteristics such as droop,<br />
<strong>and</strong> have to be calibrated be<strong>for</strong>e <strong>the</strong>y can be used <strong>for</strong> feature extraction, see section 4.6.1.<br />
The most important parameters <strong>of</strong> <strong>the</strong>se digitization circuits are listed in table 4.1.<br />
4.4.1 ATWD<br />
The analog transient wave<strong>for</strong>m digitizer is an application-specific integrated circuit (ASIC)<br />
located on <strong>the</strong> DOM’s mainboard. Three amplifiers feed <strong>the</strong> PMT signals into three <strong>of</strong><br />
<strong>the</strong> ATWD’s input channels, leading to different saturation voltages U satu . If a trigger<br />
condition is met, channel 0 – <strong>the</strong> one with <strong>the</strong> lowest gain – is digitized first. All channels<br />
feature a resolution <strong>of</strong> 10 bit ̂= 1024 counts, i.e., discrete values; if <strong>the</strong> maximum value <strong>of</strong><br />
one <strong>of</strong> <strong>the</strong> first two channels exceeds 768 counts, <strong>the</strong> next higher channel is digitized as<br />
well. For channel 2, U satu = 7.5 V is greater than <strong>the</strong> PMT’s maximum voltage <strong>of</strong> 5 V, so<br />
all ATWD channels toge<strong>the</strong>r span <strong>the</strong> full dynamic range <strong>of</strong> <strong>the</strong> PMT. The combination<br />
33
4 THE ICECUBE NEUTRINO OBSERVATORY<br />
q / PE<br />
0.3<br />
<strong>Feature</strong><strong>Extractor</strong><br />
NFE ATWD OT<br />
NFE ATWD NT<br />
NFE FADC<br />
0.2<br />
0.1<br />
0 5 10 15 20<br />
bin<br />
Figure 4.6: Comparison <strong>of</strong> ATWD <strong>and</strong> FADC single photoelectron (SPE) wave<strong>for</strong>m shape<br />
parametrizations used in different feature extractors. The three NFE parametrizations<br />
are based on a study by Christopher Wendt.[47] Note that <strong>the</strong> bin length<br />
differs <strong>for</strong> ATWD <strong>and</strong> FADC, so <strong>the</strong> shapes are not to scale time-wise.<br />
<strong>of</strong> <strong>the</strong> different channels is per<strong>for</strong>med by <strong>the</strong> IceCube s<strong>of</strong>tware module DOMcalibrator<br />
ei<strong>the</strong>r on a bin-to-bin or on a global wave<strong>for</strong>m basis depending on <strong>the</strong> configuration; <strong>the</strong><br />
three channels’ combined effective resolution is about 14 bit.<br />
The average wave<strong>for</strong>m shape <strong>of</strong> a single photoelectron (SPE) signal depends on <strong>the</strong> DOM’s<br />
toroid, see figure 4.6.<br />
The conversion time amounts to 29 µs during which <strong>the</strong> ATWD can not capture fur<strong>the</strong>r<br />
wave<strong>for</strong>ms. To compensate <strong>for</strong> <strong>the</strong>se dead times, <strong>the</strong> mainboard contains two different<br />
ATWD chips called ATWD-A <strong>and</strong> -B, processing <strong>the</strong> signal alternatingly. This significantly<br />
reduces <strong>the</strong> dead time: When a DOM is triggered (launched), both <strong>the</strong> FADC <strong>and</strong><br />
one <strong>of</strong> <strong>the</strong> ATWD chips start capturing. The DOM can not be retriggered until <strong>the</strong> FADC<br />
has done capturing; it is ready again after 6.45 µs (section 4.4.2). If <strong>the</strong> DOM is launched<br />
again in <strong>the</strong> remaining 22.55 µs during which <strong>the</strong> ATWD chip is digitizing <strong>the</strong> wave<strong>for</strong>m,<br />
<strong>the</strong> FADC will start capturing toge<strong>the</strong>r with <strong>the</strong> o<strong>the</strong>r ATWD chip.[46]<br />
4.4.2 FADC<br />
The fast analog digital converter has a larger bin length resulting in a coarser but longer<br />
digitized wave<strong>for</strong>m (table 4.1). Its main usage is to capture features which are too late or<br />
too long to appear in <strong>the</strong> ATWD wave<strong>for</strong>m(s).<br />
The FADC’s time resolution is not sufficient to resolve SPE-like PMT features. Because<br />
34
4.5 Data Structure <strong>and</strong> Data Rate<br />
<strong>of</strong> this, a 180 ns pulse shaper is used to broaden <strong>the</strong> input signals. The widening allows<br />
to estimate <strong>the</strong> feature’s arrival time with sub-bin precision by <strong>the</strong> distribution <strong>of</strong> <strong>the</strong><br />
captured charge in several bins. The FADC has negligible dead time (two clock cycles<br />
equaling 50 ns), <strong>and</strong> is <strong>the</strong> digitizer used to provide SLC charge stamps.[46]<br />
4.4.3 SLC Chargestamps<br />
Prior to <strong>the</strong> IC59 data taking season in 2009, DOMs were only read out in case <strong>of</strong> multiple<br />
DOMs fulfilling a hard local coincidence (HLC) condition: To reduce data caused by noise,<br />
a DOM’s ATWD <strong>and</strong> FADC wave<strong>for</strong>ms were only transmitted to <strong>the</strong> surface if one <strong>of</strong> <strong>the</strong><br />
neighboring or – depending on <strong>the</strong> configuration – one <strong>of</strong> <strong>the</strong> next-to-neighboring DOMs<br />
also triggered within one microsecond.<br />
Since IC59, DOM launches that do not fulfill HLC are kept irrespectively if o<strong>the</strong>r DOMs<br />
triggered HLC during a small time window. To reduce <strong>the</strong> amount <strong>of</strong> data caused by <strong>the</strong>se<br />
so-called s<strong>of</strong>t local coincidence (SLC) launches, SLC charge stamps are stored instead <strong>of</strong><br />
full ATWD <strong>and</strong> FADC wave<strong>for</strong>ms. These charge stamps are condensed FADC wave<strong>for</strong>ms;<br />
instead <strong>of</strong> 256 bins <strong>the</strong>y contain only <strong>the</strong> highest <strong>of</strong> <strong>the</strong> first 16 bins <strong>and</strong> its two direct<br />
neighbors. In most cases, this will yield a concave wave<strong>for</strong>m (in which <strong>the</strong> middle bin will<br />
be <strong>the</strong> highest one), but in some cases FADC bin 0 or 15 might hold <strong>the</strong> maximum. In<br />
<strong>the</strong>se cases, <strong>the</strong> value <strong>of</strong> bin -1 (which can <strong>and</strong> will be accessed by <strong>the</strong> DAQ) resp. bin 16<br />
might be even larger, leading to an irregular SLC charge stamp in which <strong>the</strong> middle bin<br />
is not <strong>the</strong> highest.[48]<br />
The comparatively low saturation voltage <strong>of</strong> <strong>the</strong> FADC does not hamper SLC charge<br />
stamps because <strong>of</strong> two reasons: First <strong>the</strong> brightness required to saturate FADC would<br />
almost inevitably trigger nearby DOMs <strong>and</strong> consequently HLC, <strong>and</strong> secondly <strong>the</strong>re is<br />
a third trigger condition called self-local coincidence, fully launching every DOM with<br />
exceptionally strong PMT signals independently <strong>of</strong> HLC.[46]<br />
4.5 Data Structure <strong>and</strong> Data Rate<br />
The start times <strong>of</strong> DOMs fulfilling a HLC condition are sent to <strong>the</strong> IceCube Laboratory<br />
on <strong>the</strong> surface <strong>and</strong> analyzed by triggers. If <strong>the</strong>y meet <strong>the</strong> specified criteria <strong>of</strong> at least<br />
one trigger condition, <strong>the</strong> launches are aggregated into an event. All wave<strong>for</strong>ms inside an<br />
event are calibrated, features are extracted <strong>and</strong> used by different algorithms to quickly<br />
reconstruct a muon track or a cascade, which is <strong>the</strong>n assessed by different online-filters.<br />
If at least one <strong>of</strong> <strong>the</strong>se filters decides that this event is worth to be kept, it remains in <strong>the</strong><br />
data stream which is written into binary .i3 files that are sent to nor<strong>the</strong>rn hemisphere<br />
collaboration members per satellite <strong>for</strong> <strong>of</strong>fline-processing. Additionally all data is recorded<br />
on tapes. Up to several hundreds successive .i3 files are summarized into a single detector<br />
run normally lasting about 8 h, during which <strong>the</strong> detector configuration is assumed to be<br />
constant.<br />
35
4 THE ICECUBE NEUTRINO OBSERVATORY<br />
A DOM’s in-ice data rate <strong>for</strong> HLC launches is about 10 Hz, its SLC launch rate<br />
corresponds to <strong>the</strong> dark count rate <strong>of</strong> about 300 Hz. The rate <strong>of</strong> events passing <strong>the</strong><br />
triggers (global trigger rate) is <strong>of</strong> <strong>the</strong> order <strong>of</strong> 4000 Hz, <strong>the</strong> global filter rate lies below<br />
100 Hz.<br />
4.6 S<strong>of</strong>tware<br />
The IceCube s<strong>of</strong>tware is built within a highly modular framework called IceTray[49]. Data<br />
processing <strong>of</strong> all kinds is organized serially in one-way streams. IceTray itself provides<br />
base classes <strong>for</strong> modules (e. g. I3Module) <strong>and</strong> services (I3Service), frames (I3Frame)<br />
which are <strong>the</strong> containers in which all data belonging to an individual IceCube event is<br />
saved, a definition <strong>of</strong> physical units (I3Units), Python scripting support <strong>and</strong> much more.<br />
Various projects can be loaded to provide fur<strong>the</strong>r functionality by adding modules <strong>and</strong><br />
services, such as dataclasses as st<strong>and</strong>ardized interface <strong>for</strong> various IceCube specific objects,<br />
DOMcalibrator <strong>for</strong> wave<strong>for</strong>m calibration (section 4.6.1), tools <strong>for</strong> filtering, visualization<br />
as well as feature extraction <strong>and</strong> reconstruction. The New<strong>Feature</strong><strong>Extractor</strong> (NFE, I3NFE)<br />
presented in this <strong>the</strong>sis is one <strong>of</strong> <strong>the</strong>se projects.<br />
IceTray <strong>and</strong> selected projects are bundeled to <strong>for</strong>m different meta-projects which facilitate<br />
easy installation <strong>of</strong> s<strong>of</strong>tware environments <strong>for</strong> specialized tasks such as online-filtering,<br />
simulation (IceSim), reconstruction (IceRec), or analysis (e. g. Offline-S<strong>of</strong>tware).<br />
The s<strong>of</strong>tware is written in C++ almost exclusively, however since IceTray v3 was released<br />
in 2009, modules can be written in Python as well, which was previously only used <strong>for</strong><br />
steering scripts.<br />
4.6.1 DOMcalibrator<br />
Calibration encompasses many different processes in both <strong>the</strong> detector maintenance <strong>and</strong><br />
data taking processes. This section deals with <strong>the</strong> wave<strong>for</strong>m calibration conducted by <strong>the</strong><br />
IceCube s<strong>of</strong>tware module DOMcalibrator (I3DOMcalibrator).<br />
The DOMcalibrator’s tasks are to convert <strong>the</strong> raw wave<strong>for</strong>ms (I3DOMLaunchs) captured<br />
by <strong>the</strong> PMT in units <strong>of</strong> counts to calibrated wave<strong>for</strong>ms (I3Wave<strong>for</strong>ms) in units <strong>of</strong> mV with<br />
a near-zero baseline, to combine <strong>the</strong> three ATWD channels, <strong>and</strong> to compensate <strong>for</strong> known<br />
signal distortions <strong>and</strong> time delays.<br />
Per-DOM values <strong>for</strong> <strong>the</strong> signal transit time, a baseline <strong>of</strong>fset <strong>and</strong> <strong>the</strong> gain are obtained by<br />
monthly DOMCal detector runs <strong>and</strong> stored in an online database among o<strong>the</strong>r important<br />
in<strong>for</strong>mation such as droop correction parameters (see section 4.3). For use in <strong>of</strong>flineprocessing,<br />
GCD files (geometry, calibration, detector status) belonging to individual<br />
runs (or simulated datasets) are used instead.[41]<br />
36
CHAPTER V<br />
<strong>Feature</strong> Extraction in IceCube
5 FEATURE EXTRACTION IN ICECUBE<br />
<strong>Feature</strong> extraction in IceCube refers to <strong>the</strong> analysis <strong>of</strong> digitized calibrated wave<strong>for</strong>ms.<br />
It aims <strong>for</strong> <strong>the</strong> determination <strong>of</strong> <strong>the</strong> number <strong>and</strong> time distribution <strong>of</strong> photons<br />
observed by IceCube’s DOMs. The main challenge is to deconvolute <strong>the</strong> PMT’s <strong>and</strong> in<br />
case <strong>of</strong> FADC <strong>the</strong> pulse shaper’s smearing functions with robust <strong>and</strong> efficient algorithms.<br />
Dataclasses provides two classes <strong>for</strong> <strong>the</strong> feature extraction output: <strong>the</strong> I3RecoHit <strong>and</strong><br />
<strong>the</strong> I3RecoPulse.<br />
An I3RecoHit stores in<strong>for</strong>mation about time <strong>and</strong> source. It was designed to represent<br />
a single photon triggering <strong>the</strong> PMT <strong>and</strong> <strong>the</strong>re<strong>for</strong>e corresponds to a deposited charge<br />
<strong>of</strong> exactely 1 PE. Similar hits were used in AMANDA’s reconstruction chain, however<br />
in IceCube hits are largely deprecated in favor <strong>of</strong> pulses; <strong>of</strong> <strong>the</strong> four feature extraction<br />
modules presented in this <strong>the</strong>sis, only <strong>Feature</strong><strong>Extractor</strong> <strong>and</strong> SLCHit<strong>Extractor</strong> support<br />
hits.<br />
I3RecoPulse also provides in<strong>for</strong>mation about time <strong>and</strong> source, but additionally stores<br />
<strong>the</strong> deposited charge in units <strong>of</strong> PE <strong>and</strong> <strong>the</strong> width <strong>of</strong> <strong>the</strong> pulse, <strong>the</strong>reby <strong>of</strong>fering more<br />
in<strong>for</strong>mation, which depends less on design decisions; <strong>for</strong> example I3RecoHit creation is<br />
ambigious in case <strong>of</strong> three distinct but close 0.7 PE features because <strong>the</strong>ir total charge<br />
corresponds to only two hits, or in case <strong>of</strong> a single 1.5 PE feature.<br />
Both classes <strong>of</strong>fer a data member called hitID whose purpose is not well-defined. For<br />
all four feature extraction modules presented in this <strong>the</strong>sis, it redundantly reflects <strong>the</strong><br />
hit’s or pulse’s position in <strong>the</strong>ir respective series 2 . This hitID is put to better use in<br />
merged I3RecoPulseSeries as it can be used to trace back a pulse, see section 5.4.5.<br />
All <strong>of</strong> <strong>the</strong> series <strong>of</strong> one event are finally organized in a STL map using <strong>the</strong>ir DOM<br />
<strong>and</strong> string numbers combined in OMKeys as key value. The resulting I3RecoHitSeries-<br />
Map/I3RecoPulseSeriesMap can <strong>the</strong>n be accessed by o<strong>the</strong>r modules <strong>and</strong> services fur<strong>the</strong>r<br />
down <strong>the</strong> data stream.<br />
For <strong>the</strong> following chapters, let w i , i = 0 . . . L, be <strong>the</strong> values <strong>of</strong> <strong>the</strong> bins <strong>of</strong> a given<br />
calibrated wave<strong>for</strong>m, with L = 127 <strong>for</strong> ATWD, L = 255 <strong>for</strong> FADC <strong>and</strong> L = 2 <strong>for</strong> SLC.<br />
5.1 <strong>Feature</strong><strong>Extractor</strong><br />
The <strong>Feature</strong><strong>Extractor</strong> (I3<strong>Feature</strong><strong>Extractor</strong>, FE) mainly written by Dmitry Chirkin[50]<br />
is <strong>the</strong> feature extraction module currently used <strong>for</strong> IceCube ATWD <strong>and</strong> FADC wave<strong>for</strong>ms.<br />
Developement began in 2003 <strong>for</strong> fat-reader, an alternative IceCube s<strong>of</strong>tware suite. A first<br />
release <strong>for</strong> IceTray under SVN 3 control was made in 2005 <strong>for</strong> <strong>the</strong> use with data from<br />
2 Series is IceCube’s name <strong>for</strong> st<strong>and</strong>ard C++ STL vectors that contain related objects such as all<br />
hits/pulses extracted from a single DOM <strong>for</strong> a distinct event.<br />
3 SVN aka Subversion is <strong>the</strong> version-control system used <strong>for</strong> <strong>the</strong> central IceCube source code repository.<br />
38
sum <strong>of</strong> all bins above<br />
baseline + error<br />
extrapolation <strong>of</strong><br />
maximum slope<br />
to baseline<br />
width<br />
width:<br />
half <strong>of</strong> <strong>the</strong> number<br />
<strong>of</strong> bins above<br />
first pulse‘s half<br />
maximum height<br />
threshold<br />
5.1 <strong>Feature</strong><strong>Extractor</strong><br />
parabola fit to maximum bin<br />
charge:<br />
parabola maximum<br />
× pulse width<br />
extrapolation <strong>of</strong> first<br />
local maximum slope<br />
above threshold<br />
to baseline<br />
charge:<br />
sum <strong>of</strong> all bins<br />
extrapolation <strong>of</strong><br />
maximum slope<br />
to baseline<br />
width:<br />
proportional to<br />
charge / maximum<br />
threshold<br />
width<br />
Figure 5.1: Sketch illustrating <strong>Feature</strong><strong>Extractor</strong>’s first single-pulse extraction algorithms to<br />
left <strong>and</strong> its second one to <strong>the</strong> right. Shown is a part <strong>of</strong> a calibrated wave<strong>for</strong>m in<br />
arbitrary units; baseline correction has been omitted <strong>for</strong> reasons <strong>of</strong> clarity.<br />
x<br />
x<br />
IceCube’s first deployed string. Since <strong>the</strong>n, <strong>new</strong> algorithms <strong>and</strong> options were added to<br />
<strong>the</strong> module. It is used <strong>for</strong> practically all physics analyses up to early 2010.<br />
For ATWD <strong>the</strong> <strong>Feature</strong><strong>Extractor</strong> <strong>of</strong>fers two single-pulse extraction algorithms that<br />
define at most one pulse per wave<strong>for</strong>m as well as two more sophisticated multi-pulse<br />
x<br />
extraction algorithms; <strong>for</strong> FADC it uses a fifth algorithm.[50]<br />
x<br />
x<br />
x<br />
The first single-pulse algorithm extracts only one pulse given by <strong>the</strong> largest feature<br />
in <strong>the</strong> wave<strong>for</strong>m, <strong>and</strong> its corresponding charge, see figure 5.1. It searches <strong>the</strong> maximum <strong>of</strong><br />
<strong>the</strong> wave<strong>for</strong>m’s slopes ∆w i := w i+1 −w i from <strong>the</strong> beginning up to <strong>the</strong> wave<strong>for</strong>m’s maximum<br />
bin w max (or to <strong>the</strong> first saturated bin if any). The intersection <strong>of</strong> <strong>the</strong> extrapolation <strong>of</strong> this<br />
slope with <strong>the</strong> baseline is used to define <strong>the</strong> leading-edge time. The algorithm <strong>the</strong>n fits a<br />
parabola through <strong>the</strong> wave<strong>for</strong>m’s maximum <strong>and</strong> its two surrounding bins; <strong>the</strong> difference<br />
between <strong>the</strong> position <strong>of</strong> <strong>the</strong> maximum <strong>and</strong> <strong>the</strong> leading-edge time from above defines <strong>the</strong><br />
width <strong>of</strong> <strong>the</strong> pulse. This width multiplied with <strong>the</strong> parabola’s maximum defines <strong>the</strong> charge<br />
<strong>of</strong> <strong>the</strong> pulse.<br />
This is a very fast algorithm, but it lacks robustness. For a given wave<strong>for</strong>m, <strong>the</strong> maximum<br />
<strong>of</strong> <strong>the</strong> wave<strong>for</strong>m’s slopes does not neccessarily belong to <strong>the</strong> feature with <strong>the</strong> highest<br />
amplitude; this might be due to overlapping features where one feature’s tail flattens<br />
<strong>the</strong> next feature’s leading edge, due to binning effects, or due to statistical fluctuations.<br />
In <strong>the</strong>se cases, <strong>the</strong> extracted width <strong>and</strong> <strong>the</strong>reby <strong>the</strong> extracted charge can be extremly<br />
overestimated, <strong>and</strong> <strong>the</strong> time does not match <strong>the</strong> largest feature. Also, <strong>for</strong> most track<br />
reconstruction methods <strong>the</strong> first feature’s time is more valuable than <strong>the</strong> time <strong>of</strong> <strong>the</strong><br />
largest feature, because it is least affected by scattering. There<strong>for</strong>e it is disadvantageous<br />
to extract only <strong>the</strong> maximum pulse’s leading edge time as done by this algorithm.<br />
To make it more robust, <strong>the</strong> algorithm can be configured to use <strong>the</strong> second single-pulse<br />
algorithm’s charge estimate instead; however, <strong>the</strong> whole algorithm is largely superseded<br />
by <strong>the</strong> second single-pulse algorithm.<br />
39
5 FEATURE EXTRACTION IN ICECUBE<br />
The second single-pulse algorithm extracts <strong>the</strong> time <strong>of</strong> <strong>the</strong> first feature’s leading<br />
edge along with <strong>the</strong> wave<strong>for</strong>m’s total integrated charge. Precisely, it searches <strong>for</strong> <strong>the</strong><br />
first local maximum <strong>of</strong> <strong>the</strong> ∆w i <strong>for</strong> all consecutive pairs (w i , w i+1 ) above a configurable<br />
threshold <strong>and</strong> computes <strong>the</strong> position <strong>of</strong> baseline-crossing <strong>of</strong> its extrapolated slope. The<br />
pulse’s charge q is defined as sum over all wave<strong>for</strong>m bins, <strong>and</strong> <strong>the</strong> width is defined by<br />
q<br />
const·<br />
w max<br />
as illustrated in figure 5.1 (contradicting to <strong>the</strong> documentation[50], see section<br />
C.2).<br />
Like <strong>the</strong> first single-pulse extraction algorithm, this second one is fast, however it is<br />
not robust regarding prepulses (see section 4.3): As prepulses are rare <strong>and</strong> only have<br />
a small amplitude, <strong>the</strong>y normally have little impact on reconstruction <strong>and</strong> might not<br />
even pass <strong>the</strong> trigger or extraction threshold. However, <strong>for</strong> many-PE features <strong>the</strong>y can<br />
cause a comparably tiny peak ahead <strong>of</strong> <strong>the</strong> large feature. In <strong>the</strong>se cases, <strong>the</strong> algorithm<br />
attributes <strong>the</strong> whole charge to <strong>the</strong> prepulse, potentially increasing its weight in <strong>the</strong> later<br />
track reconstruction heavily.<br />
Because <strong>of</strong> its speed <strong>and</strong> <strong>the</strong> low rate <strong>of</strong> prepulses <strong>the</strong> algorithm is none<strong>the</strong>less well-suited<br />
<strong>for</strong> online-filtering, <strong>for</strong> which it is currently used.[51]<br />
The first multi-pulse algorithm is <strong>the</strong> sophisticated but time-consuming ROOTfit<br />
algorithm. Using libraries <strong>of</strong> <strong>the</strong> eponymous ROOT framework[52], it fits increasing numbers<br />
<strong>of</strong> SPE-like pulses to <strong>the</strong> wave<strong>for</strong>m until ei<strong>the</strong>r <strong>the</strong> χ 2 -measured goodness <strong>of</strong> <strong>the</strong> fit<br />
does not increase anymore, or until <strong>the</strong> maximum number <strong>of</strong> pulses has been reached.<br />
<strong>Feature</strong><strong>Extractor</strong>’s SPE pulse shape parametrization is<br />
w(t) = q ((<br />
1 − e<br />
−τ 2) e −τ + w d (t) ) , (5.1)<br />
c 1<br />
(<br />
√ ( ( ( ))<br />
w d (t) = −c 2 e −c 3 τ<br />
1 − e −τ − 1 πe 1 2<br />
2 erf τ +<br />
1<br />
2) ) − erf 1<br />
, τ = t−t 0<br />
,<br />
2<br />
σ<br />
where q, t 0 , <strong>and</strong> σ are fitting-parameters <strong>for</strong> <strong>the</strong> pulse’s charge, leading-edge time, <strong>and</strong><br />
width, respectively; erf denotes <strong>the</strong> error function, <strong>and</strong> c 1 . . . c 3 are constants. The term<br />
w d is an approach to deal with droop; if it is to be included, <strong>the</strong> droop correction in<br />
DOMcalibrator has to be disabled. The parametrization is shown in figure 4.6 with <strong>the</strong><br />
values used in <strong>the</strong> second multi-pulse algorithm (q = 1 PE, t 0 = 0, σ = 2 ns), <strong>and</strong> with<br />
<strong>the</strong> droop correction term disabled (i.e., c 2 = 0).<br />
This algorithm is nei<strong>the</strong>r recommended nor maintained anymore, <strong>and</strong> it is too slow <strong>for</strong><br />
large-scale reconstruction.<br />
The second multi-pulse algorithm is based upon <strong>the</strong> method <strong>of</strong> Bayesian Unfolding.<br />
Being reasonably fast <strong>and</strong> capable <strong>of</strong> determining photon arrival times in complex<br />
features, it is used <strong>for</strong> practically all analyses that rely on multi-pulse extraction, e. g.,<br />
cascade reconstruction <strong>and</strong> <strong>of</strong>fline-processing muon track reconstruction.[51]<br />
40
5.2 Pulse<strong>Extractor</strong><br />
An algorithm using <strong>the</strong> same underlying method is implemented in <strong>the</strong> New<strong>Feature</strong><strong>Extractor</strong><br />
NFE developed within this <strong>the</strong>sis. There<strong>for</strong>e <strong>the</strong> method <strong>and</strong> <strong>the</strong> differences between<br />
<strong>the</strong> two implementations are discussed in section 6.3.<br />
The FADC algorithm is a multi-pulse algorithm similar to <strong>the</strong> second single-pulse<br />
ATWD algorithm: It searches <strong>the</strong> wave<strong>for</strong>m from start to end <strong>for</strong> a local slope maximum<br />
<strong>of</strong> pairs (w i , w i+1 ) with w i+1 above threshold <strong>and</strong> extrapolates <strong>the</strong> resulting line to <strong>the</strong><br />
baseline to define <strong>the</strong> pulse’s time. For <strong>the</strong> calculation <strong>of</strong> <strong>the</strong> charge it sums up <strong>the</strong> bin<br />
contents from i+1 on until ei<strong>the</strong>r a bin is below <strong>the</strong> threshold or until a bin is higher than<br />
its predecessor after <strong>the</strong> feature has fallen below half its previous maximum value, in both<br />
cases excluding <strong>the</strong> determining bin. The width is defined as <strong>the</strong> number <strong>of</strong> bins summed<br />
up. The algorithm <strong>the</strong>n continues searching <strong>for</strong> <strong>the</strong> next pulse. Incomplete pulses at <strong>the</strong><br />
end <strong>of</strong> wave<strong>for</strong>ms are not extracted.<br />
Besides <strong>the</strong> obvious tasks <strong>of</strong> a feature extractor, <strong>Feature</strong><strong>Extractor</strong> also per<strong>for</strong>ms some<br />
wave<strong>for</strong>m calibration tasks such as baseline finding, transit time correction, or droop<br />
correction. While <strong>the</strong> employed methods were superior to <strong>the</strong> ones used in DOMcalibrator<br />
at <strong>the</strong> time <strong>of</strong> <strong>the</strong>ir introduction, <strong>the</strong> good quality <strong>of</strong> recent DOMcalibrator versions<br />
caused most <strong>of</strong> <strong>the</strong>se features to become obsolete. Their remaining in <strong>the</strong> code impairs<br />
its readability[53], introduces an overload to <strong>the</strong> maintenance <strong>of</strong> <strong>the</strong> code <strong>and</strong> increases<br />
<strong>the</strong> risk <strong>of</strong> misconfiguration.<br />
5.2 Pulse<strong>Extractor</strong><br />
Pulse<strong>Extractor</strong> (I3Pulse<strong>Extractor</strong>) is a rewritten <strong>and</strong> condensed version <strong>of</strong> <strong>Feature</strong><strong>Extractor</strong>.<br />
Created mainly by Dmitry Chirkin <strong>and</strong> Christopher Wendt in 2009, <strong>the</strong> only<br />
algorithm it includes is a modified variant <strong>of</strong> <strong>the</strong> original <strong>Feature</strong><strong>Extractor</strong>’s Bayesian<br />
Unfolding algorithm. This algorithm can be applied to both ATWD <strong>and</strong> FADC if <strong>the</strong><br />
module is run in different instances. The module does not <strong>of</strong>fer configuration options<br />
besides names <strong>of</strong> <strong>the</strong> input <strong>and</strong> output objects.[54]<br />
The Pulse<strong>Extractor</strong>’s advantages are easy maintainability <strong>and</strong> high ease <strong>of</strong> use at <strong>the</strong><br />
expense <strong>of</strong> adaptivity <strong>and</strong> extraction per<strong>for</strong>mance (see section 6.3.1).<br />
5.3 SLCHit<strong>Extractor</strong><br />
The SLCHit<strong>Extractor</strong> (I3SLCHit<strong>Extractor</strong>) was written in 2009 by Andreas Groß to be<br />
<strong>the</strong> first module to extract hits from SLC charge stamps, which are recorded since IC59.<br />
The algorithm defines <strong>the</strong> pulse time as t = t 0 + (i max − 1) · T bin − c 1<br />
w imax−1<br />
w max<br />
− c 2 , where t 0<br />
41
5 FEATURE EXTRACTION IN ICECUBE<br />
denotes <strong>the</strong> charge stamp’s startTime <strong>and</strong> <strong>the</strong>reby <strong>the</strong> time corresponding to w 0 , whereas<br />
i max denotes <strong>the</strong> bin with <strong>the</strong> highest value, w max ; this is illustrated on <strong>the</strong> right side <strong>of</strong><br />
figure 6.6. The charge is defined as q = wmax<br />
c 3<br />
.<br />
The constants c 1 <strong>and</strong> c 3 are obtained by comparison <strong>of</strong> SLCHit<strong>Extractor</strong>’s pulses <strong>for</strong> SLC<br />
charge stamps generated from full FADC wave<strong>for</strong>ms with <strong>Feature</strong><strong>Extractor</strong>’s pulses. The<br />
time <strong>of</strong>fset c 2 has been ab<strong>and</strong>oned recently in <strong>the</strong> belief that it was caused by a general<br />
ATWD FADC time <strong>of</strong>fset, which has been taken care <strong>of</strong> in DOMcalibrator; see appendix<br />
C.4 <strong>for</strong> fur<strong>the</strong>r discussion.[55]<br />
5.4 New<strong>Feature</strong><strong>Extractor</strong><br />
The New<strong>Feature</strong><strong>Extractor</strong> (NFE) – written by <strong>the</strong> author <strong>of</strong> this <strong>the</strong>sis – is <strong>the</strong> latest <strong>of</strong><br />
<strong>the</strong> four feature extractors covered in this document <strong>and</strong> constitutes its main topic.<br />
NFE is special in that it employs multiple algorithms <strong>for</strong> <strong>the</strong> same event, choosing an<br />
approriate algorithm according to <strong>the</strong> wave<strong>for</strong>m’s complexity; this concept is explained in<br />
section 5.4.2. The algorithms <strong>the</strong>mselves are discussed in chapter 6.<br />
5.4.1 Main Characteristics<br />
Well-documented code con<strong>for</strong>ming <strong>the</strong> IceCube Coding St<strong>and</strong>ards 4<br />
Documentation <strong>for</strong> IceCube users <strong>and</strong> developers alike exists in <strong>for</strong>m <strong>of</strong> wiki pages <strong>and</strong><br />
an internal report created out <strong>of</strong> passages <strong>of</strong> this <strong>the</strong>sis. It contains general in<strong>for</strong>mation<br />
about <strong>the</strong> project’s structure <strong>and</strong> <strong>the</strong> employed algorithms. Fur<strong>the</strong>r in-detail in<strong>for</strong>mation<br />
<strong>for</strong> developers is found in doxygen mark-up source code comments that explain class<br />
structures including purpose <strong>of</strong> methods <strong>and</strong> data members, <strong>and</strong> in st<strong>and</strong>ard C++ inline<br />
comments explaining programming decisions <strong>and</strong> purpose <strong>of</strong> small code passages. The<br />
project also provides example scripts to make <strong>the</strong> use <strong>of</strong> this project as easy as possible.<br />
Focus on <strong>the</strong> main task <strong>of</strong> feature extraction<br />
NFE relies on <strong>the</strong> wave<strong>for</strong>ms being calibrated prior by DOMcalibrator or any o<strong>the</strong>r module.<br />
It intentionally does not aim to improve <strong>for</strong> example <strong>the</strong> baseline level, but instead <strong>the</strong><br />
author communicated potential shortcomings <strong>of</strong> o<strong>the</strong>r modules to <strong>the</strong> respective authors.<br />
This preserves IceTray’s data flow concept <strong>and</strong> prevents double treatment <strong>of</strong> problems or<br />
o<strong>the</strong>r unintended side-effects later in <strong>the</strong> reconstruction chain. This also results in simpler<br />
configuration <strong>of</strong> NFE.<br />
4 http://s<strong>of</strong>tware.icecube.wisc.edu/OFFLINE-SOFTWARE-V02-02-03/codingst<strong>and</strong>ards.html<br />
42
5.4 New<strong>Feature</strong><strong>Extractor</strong><br />
Ease <strong>of</strong> use<br />
To make <strong>the</strong> transition from o<strong>the</strong>r feature extractors to NFE as easy as possible without<br />
sacrificing configurability <strong>and</strong> flexibility, all NFE modules <strong>and</strong> services require only a minimum<br />
<strong>of</strong> configuration. This is implemented by reasonable default values which give good<br />
results under almost all circumstances. Thoroughly tested defaults also help minimizing<br />
unintentional misconfigurations.<br />
Modular code organization<br />
The code is organized into IceTray modules <strong>and</strong> services with specific purposes to improve<br />
readability, exp<strong>and</strong>ability <strong>and</strong> possible bug tracking. Modular organization also allows <strong>the</strong><br />
creation <strong>of</strong> unit tests to thouroughly check passages <strong>of</strong> code. While initially providing good<br />
means to test <strong>the</strong> code <strong>for</strong> errors during development, <strong>the</strong> main advantage <strong>of</strong> unit tests<br />
is <strong>the</strong>ir ability to test <strong>the</strong> code <strong>for</strong> bugs introduced later by seemingly unrelated code<br />
changes. The tests toge<strong>the</strong>r with <strong>the</strong> provided example scripts are run prior to releasing<br />
<strong>new</strong> versions <strong>of</strong> <strong>the</strong> project or <strong>of</strong> meta-projects containing it.<br />
High per<strong>for</strong>mance<br />
NFE is intended to succeed <strong>Feature</strong><strong>Extractor</strong> <strong>and</strong> SLCHit<strong>Extractor</strong> not only <strong>for</strong> <strong>of</strong>flineprocessing,<br />
i.e., processing <strong>of</strong> stored data in <strong>the</strong> No<strong>the</strong>rn Hemisphere, but also <strong>for</strong> onlineprocessing,<br />
<strong>for</strong> which per<strong>for</strong>mance is a major issue. Besides choosing efficient algorithms,<br />
a <strong>new</strong> approach was taken in using different algorithms <strong>for</strong> different types <strong>of</strong> wave<strong>for</strong>ms;<br />
this is elaborated in section 5.4.2.<br />
5.4.2 Program Structure<br />
Most <strong>of</strong> IceCube’s DOM launches are caused by single photons whose corresponding wave<strong>for</strong>ms<br />
are considerably regular. To take advantage <strong>of</strong> this, <strong>the</strong> NFE framework allows<br />
to use different extraction algorithms <strong>for</strong> wave<strong>for</strong>ms <strong>of</strong> varying complexity. The time<br />
saved during <strong>the</strong> extraction <strong>of</strong> simple wave<strong>for</strong>ms can be spent <strong>for</strong> sophisticated <strong>and</strong> timeconsuming<br />
algorithms to improve <strong>the</strong> results <strong>for</strong> more complex wave<strong>for</strong>ms. Additionally,<br />
this allows to use algorithms which excel in extracting certain wave<strong>for</strong>ms but fail <strong>for</strong> o<strong>the</strong>r<br />
wave<strong>for</strong>m categories, or to use <strong>the</strong> same algorithm with different settings <strong>for</strong> different categories.<br />
To decide which algorithm to use <strong>for</strong> a given wave<strong>for</strong>m, a pre-evaluation algorithm is initially<br />
called to quickly assess <strong>the</strong> wave<strong>for</strong>m. Depending on <strong>the</strong> results, <strong>the</strong> NFE framework<br />
(i.e., <strong>the</strong> module I3NFE) passes <strong>the</strong> wave<strong>for</strong>m <strong>and</strong> all relevant data to <strong>the</strong> extraction algorithm<br />
configured <strong>for</strong> this category. Currently, three categories have been implemented,<br />
but this number can be exp<strong>and</strong>ed; <strong>the</strong>y are simple <strong>for</strong> ATWD or FADC wave<strong>for</strong>ms with<br />
simple wave<strong>for</strong>ms, complex <strong>for</strong> all remaining ATWD <strong>and</strong> FADC wave<strong>for</strong>ms, <strong>and</strong> slc <strong>for</strong><br />
SLC charge stamps; <strong>the</strong> classification <strong>of</strong> <strong>the</strong>se categories lies in <strong>the</strong> responsibility <strong>of</strong> <strong>the</strong><br />
43
5 FEATURE EXTRACTION IN ICECUBE<br />
pre-evaluation algorithm (e. g. “Eva”, section 6.1).<br />
All algorithms are implemented as services <strong>and</strong> inherit from a common pre-evaluation or<br />
extraction algorithm base class. This modular design – inspired by <strong>the</strong> IceCube reconstruction<br />
framework gulliver by David Boersma[56] – simplifies maintenance <strong>and</strong> facilitates<br />
easy implementation <strong>of</strong> <strong>new</strong> algorithms without breaking existing code.<br />
NFE’s framework is also responsible <strong>for</strong> calculating <strong>the</strong> gain constant needed to translate<br />
wave<strong>for</strong>ms in mV to wave<strong>for</strong>ms in units <strong>of</strong> photoelectron charges (PE). Providing this<br />
gain constant to <strong>the</strong> algorithms allows to define thresholds independently <strong>of</strong> hardware or<br />
firmware thresholds that might change between different firmware or simulation s<strong>of</strong>tware<br />
versions. This eliminates one reason <strong>for</strong> recalibration <strong>of</strong> configuration settings; in addition<br />
PE are more practical <strong>for</strong> setting thresholds as <strong>the</strong>y are closer related to <strong>the</strong> physics processes.<br />
The same approach was taken in Pulse<strong>Extractor</strong>, while <strong>Feature</strong><strong>Extractor</strong> defines<br />
thresholds in fractions <strong>of</strong> <strong>the</strong> trigger’s SPE discriminator threshold.<br />
The exact conversion factor <strong>for</strong> <strong>the</strong> translation <strong>of</strong> mV to PE is T bin (g · Z · e) −1 . The average<br />
PMT gain g (SPEmean) <strong>and</strong> <strong>the</strong> DOM’s impedance Z (FrontEndImpedance), which<br />
is dominated by <strong>the</strong> toroid, are provided by ei<strong>the</strong>r <strong>the</strong> GCD file or <strong>the</strong> online database.<br />
Ano<strong>the</strong>r non-obvious task <strong>for</strong> <strong>the</strong> framework is to optionally add an extra-info series<br />
map <strong>and</strong> an algorithm-info series map to <strong>the</strong> frame. Those objects mirror <strong>the</strong><br />
I3RecoPulseSeriesMap’s structure, which means that <strong>for</strong> every pulse <strong>of</strong> every DOM with<br />
entries in <strong>the</strong> pulse map <strong>the</strong>se objects contain one integer each. The extra-info integer is<br />
a bit mask containing in<strong>for</strong>mation about whe<strong>the</strong>r <strong>the</strong> pulse was cut <strong>of</strong>f in <strong>the</strong> beginning,<br />
cut <strong>of</strong>f in <strong>the</strong> end, or if it was saturated. The latter in<strong>for</strong>mation is obtained from DOMcalibrator’s<br />
StatusCompounds which are added to calibrated wave<strong>for</strong>ms but are largely<br />
ignored by o<strong>the</strong>r feature extractors. The algorithm-info contains <strong>the</strong> unique ID <strong>of</strong> <strong>the</strong><br />
algorithm that was used to extract <strong>the</strong> corresponding pulse.<br />
To keep <strong>the</strong> code simple <strong>and</strong> to <strong>of</strong>fer more configurability when needed, all data sources<br />
(ATWD, FADC <strong>and</strong> SLC, resp. all different calibrated wave<strong>for</strong>m series) are attached to<br />
different instances <strong>of</strong> <strong>the</strong> I3NFE module. Algorithm instances (services) can be used by an<br />
unlimited number <strong>of</strong> modules. The resulting pulse series maps can optionally be merged<br />
into a single map by <strong>the</strong> pulse merger (section 5.4.5).<br />
5.4.3 Time Offset Constants<br />
One drawback <strong>of</strong> <strong>the</strong> application <strong>of</strong> different feature extraction algorithms within <strong>the</strong><br />
same event is <strong>the</strong> need <strong>for</strong> <strong>the</strong> introduction <strong>of</strong> time <strong>of</strong>fset calibration constants. Those are<br />
required to align <strong>the</strong> pulses to <strong>the</strong> same start time independently <strong>of</strong> <strong>the</strong> algorithm <strong>and</strong><br />
data source used. The two major objections against <strong>the</strong>se constants are that <strong>the</strong>y might<br />
require maintenance <strong>and</strong> that <strong>the</strong>y should ideally not be required.<br />
The first objection only arises <strong>for</strong> features <strong>of</strong> different sources, specifically features which<br />
44
5.4 New<strong>Feature</strong><strong>Extractor</strong><br />
occur in both ATWD <strong>and</strong> FADC. It can be mitigated by deducing <strong>the</strong> time <strong>of</strong>fsets from<br />
well-calibrated wave<strong>for</strong>ms so that <strong>the</strong>y only compensate <strong>of</strong>fsets introduced by <strong>the</strong> algorithms,<br />
but not by a potentially imperfect calibration.<br />
The second objection only holds under <strong>the</strong> assumption that SPE pulses are sufficiently<br />
similar <strong>for</strong> all sources. Even if a single method could be used <strong>for</strong> all sources, <strong>the</strong> different<br />
pulse shapes would require subsequent alignment (see figure 4.6).<br />
5.4.4 Wave<strong>for</strong>ms Without Pulses<br />
Wave<strong>for</strong>ms might not contain any features which pass a given threshold. NFE provides<br />
three alternatives to deal with this situation: DOMs without pulses can ei<strong>the</strong>r be excluded<br />
from <strong>the</strong> pulse series map, or <strong>the</strong>y can be added with an empty pulse series, or NFE can<br />
<strong>for</strong>ce <strong>the</strong> algorithms to extract at least one pulse, respectively NFE can pass <strong>the</strong> wave<strong>for</strong>m<br />
to an algorithm which is capable <strong>of</strong> finding at least one pulse if en<strong>for</strong>ced.<br />
The important difference between <strong>the</strong> first two alternatives is that <strong>the</strong>y lead to different<br />
values <strong>for</strong> <strong>the</strong> per-event quantity NChan – i.e., <strong>the</strong> number <strong>of</strong> DOMs hit during <strong>the</strong> given<br />
event, <strong>of</strong>ten used <strong>for</strong> energy estimation <strong>and</strong> quality cuts –, because it is defined as <strong>the</strong><br />
number <strong>of</strong> entries in <strong>the</strong> pulse series map. There<strong>for</strong>e, <strong>the</strong> underlying question is if one is<br />
to trust <strong>the</strong> hardware trigger or <strong>the</strong> feature extraction to decide whe<strong>the</strong>r a DOM was hit.<br />
The physical implications <strong>of</strong> this choice have to be discussed in a larger scale than it is<br />
possible in this <strong>the</strong>sis.<br />
5.4.5 Pulse Merger<br />
The pulse merger (I3NFEPulseMerger) is a st<strong>and</strong>-alone IceTray module designed to join<br />
up to three pulse series maps into a single one. Usually <strong>the</strong>se maps correspond to <strong>the</strong> different<br />
pulse sources ATWD, FADC <strong>and</strong> SLC individually extracted by NFE, but differing<br />
usage is supported.<br />
The input maps are arranged by priority; by default, ATWD has <strong>the</strong> highest priority <strong>and</strong><br />
SLC has <strong>the</strong> lowest. The primary map is copied completely, pulses from <strong>the</strong> o<strong>the</strong>r two<br />
maps are added successively if <strong>the</strong>y don’t overlap with pulses from higher priority input<br />
maps – <strong>the</strong>y may, however, overlap with pulses from <strong>the</strong> same input map. An extra time<br />
window can be configured to be required between <strong>the</strong> ending (defined as start time plus<br />
width) <strong>of</strong> one pulse <strong>and</strong> <strong>the</strong> start <strong>of</strong> ano<strong>the</strong>r pulse to reduce <strong>the</strong> probability <strong>of</strong> double<br />
extraction, i.e., extraction <strong>of</strong> <strong>the</strong> same feature by different sources (compare appendix<br />
C.1).<br />
The NFE pulse merger differs from similar modules in that it supports merging <strong>of</strong> maps<br />
with extra-info <strong>and</strong> algorithm-info.<br />
45
5 FEATURE EXTRACTION IN ICECUBE<br />
46
CHAPTER VI<br />
Algorithms Implemented in NFE
6 ALGORITHMS IMPLEMENTED IN NFE<br />
maximum<br />
threshold<br />
x<br />
feature<br />
threshold<br />
x<br />
x<br />
Figure 6.1: Sketch illustrating NFE’s pre-evaluation algorithm “Eva”. If a wave<strong>for</strong>m only contains<br />
features like <strong>the</strong> first one, it is evaluated as simple; however if it contains<br />
features which are too long, too high, or too close toge<strong>the</strong>r, <strong>the</strong> whole wave<strong>for</strong>m is<br />
classified to be complex.<br />
x<br />
Up to now, one pre-evaluation algorithm (“Eva”) <strong>and</strong> three extraction algorithms have<br />
been implemented. Of <strong>the</strong> latter three, “Simple” was designed mainly <strong>for</strong> <strong>the</strong> wave<strong>for</strong>m<br />
category <strong>of</strong> <strong>the</strong> same name, “BayesUnfold” is capable <strong>of</strong> h<strong>and</strong>ling well all types <strong>of</strong> ATWD<br />
<strong>and</strong> FADC wave<strong>for</strong>ms, <strong>and</strong> “SLCHE” extracts SLC chargestamps<br />
x<br />
exclusively; all four<br />
algorithms are part <strong>of</strong> <strong>the</strong> default configuration<br />
x<br />
<strong>of</strong> <strong>the</strong> New<strong>Feature</strong><strong>Extractor</strong>.<br />
Currently “BayesUnfold” is <strong>the</strong> only algorithm capable <strong>of</strong> finding at least one pulse per<br />
wave<strong>for</strong>m if en<strong>for</strong>ced (see section 5.4.4), so it needs to be used if En<strong>for</strong>cePulse is set;<br />
to accomplish this, a special En<strong>for</strong>ceAlgorithmServiceName can be set, which is only<br />
necessary if “BayesUnfold” is not used <strong>for</strong> ei<strong>the</strong>r simple or complex wave<strong>for</strong>ms.<br />
6.1 Pre-evaluation Algorithm “Eva”<br />
As pointed out in section 5.4.2, <strong>the</strong> purpose <strong>of</strong> pre-evaluation algorithm is to quickly assess<br />
a wave<strong>for</strong>m <strong>and</strong> to decide whe<strong>the</strong>r it belongs to <strong>the</strong> simple, complex, or slc category.<br />
The pre-evaluation algorithm “Eva” accomplishes this by four simple checks:<br />
If a calibrated wave<strong>for</strong>m’s source is SLC, it is directly assigned to <strong>the</strong> slc category. An<br />
ATWD or FADC wave<strong>for</strong>m is scanned once; if its features are ei<strong>the</strong>r too high, too long<br />
or too close toge<strong>the</strong>r, <strong>the</strong> wave<strong>for</strong>m is marked as complex (see figure 6.1). More precisely,<br />
if one wave<strong>for</strong>m bin’s value w i exceeds <strong>the</strong> threshold w max , or if l max successive bins<br />
exceed <strong>the</strong> threshold w feat , or if <strong>the</strong> gap between bins exceeding w feat is smaller than <strong>the</strong><br />
allowed minimum distance <strong>of</strong> d min bins, <strong>the</strong>n <strong>the</strong> scanning is stopped <strong>and</strong> <strong>the</strong> wave<strong>for</strong>m is<br />
categorized as complex.<br />
The four parameters w max , w feat , l max , <strong>and</strong> d min can be configured individually <strong>for</strong><br />
48
6.2 Extraction Algorithm “Simple”<br />
quadratic interpolation <strong>for</strong> leading edge times<br />
linear interpolation <strong>for</strong> trailing edge times<br />
charge:<br />
sum over bins above<br />
feature threshold<br />
plus boundary bins<br />
detection threshold w ,<br />
detect<br />
rejects flat features<br />
widths<br />
feature threshold w , feat<br />
defines feature bounds<br />
Figure 6.2: Sketch illustrating NFE’s extraction algorithm “Simple”. In <strong>the</strong> shown wave<strong>for</strong>m,<br />
three pulses are identified. The charge compensation has been omitted <strong>for</strong> reasons<br />
<strong>of</strong> clarity. Note that “Simple” was not designed <strong>for</strong> complex features like <strong>the</strong> third.<br />
ATWD <strong>and</strong> FADC to take into account <strong>the</strong> different characteristics <strong>of</strong> <strong>the</strong>se two sources.<br />
6.2 Extraction Algorithm “Simple”<br />
The extraction algorithm “Simple” is a fast threshold based algorithm particularly suitable<br />
<strong>for</strong> simple features. To define a feature, a threshold w feat is used. Starting at <strong>the</strong><br />
wave<strong>for</strong>m’s beginning, <strong>the</strong> algorithm searches <strong>for</strong> bins exceeding w feat . If it finds one in bin<br />
i, <strong>the</strong> potential pulse’s leading edge time t is defined as <strong>the</strong> time t parab where a parabola<br />
through <strong>the</strong> points (i − 1, w i−1 ) to (i + 1, w i+1 ) crosses <strong>the</strong> threshold w feat plus a quadratic<br />
charge compensation term <strong>and</strong> a constant time <strong>of</strong>fset:<br />
⎧<br />
t = t parab − t q + t <strong>of</strong>fset with<br />
⎨c P0 (q − c P1 ) 2 <strong>for</strong> q < c P1 ,<br />
t q =<br />
⎩0 o<strong>the</strong>rwise,<br />
(6.1)<br />
with charge q <strong>and</strong> configurable constants c P0 <strong>and</strong> c P1 . While using a parabola <strong>for</strong> interpolation<br />
mainly improves <strong>the</strong> time resolution at a given charge level, <strong>the</strong> charge compensation<br />
term accounts <strong>for</strong> <strong>the</strong> fact that smaller pulses cross <strong>the</strong> threshold later, so <strong>the</strong>y are not<br />
redundant (compare section 7.1).<br />
If <strong>the</strong> wave<strong>for</strong>m starts with a bin surpassing <strong>the</strong> threshold (i. e., i = 0), t parab in equation<br />
(6.1) is not well-defined <strong>and</strong> gets replaced by t lin , which is <strong>the</strong> time <strong>of</strong> <strong>the</strong> intersection <strong>of</strong><br />
w feat with <strong>the</strong> linear extrapolation <strong>of</strong> <strong>the</strong> slope between <strong>the</strong> points (0, w 0 ) <strong>and</strong> (1, w 1 ). If<br />
t lin is more than 10 ns ahead <strong>of</strong> <strong>the</strong> wave<strong>for</strong>m’s start time t 0 , it is replaced by t 0 because<br />
it has to be assumed that <strong>the</strong> wave<strong>for</strong>m starts well within a feature <strong>and</strong> <strong>the</strong> extrapolation<br />
could be too far <strong>of</strong>f. Analogously, if <strong>the</strong> threshold is passed by <strong>the</strong> last bin (i = L), t parab is<br />
replaced by t lin defined by <strong>the</strong> last two points, (L−1, w L−1 ) <strong>and</strong> (L, w L ). Lastly, this linear<br />
fallback method is used with <strong>the</strong> points (i − 1, w i−1 ) <strong>and</strong> (i, w i ) if <strong>the</strong> parabola’s point <strong>of</strong><br />
49
6 ALGORITHMS IMPLEMENTED IN NFE<br />
intersection lies more than one bin length away from bin i, i. e., if |t parab − iT bin | > T bin ; if<br />
this happens, it must be due to numerical instability which arises <strong>for</strong> nearly colinear points<br />
(catastrophic cancellation). It remains to be an open task to rearrange <strong>the</strong> equation used<br />
to determine t parab to become more stable; this should be possible using Viète’s <strong>for</strong>mulae,<br />
but it will not improve <strong>the</strong> algorithm because <strong>the</strong> fallback method is well-suited in case<br />
<strong>of</strong> colinear points.<br />
The time calculation based on <strong>the</strong> intersection with <strong>the</strong> threshold w feat instead <strong>of</strong> <strong>the</strong><br />
baseline w = 0 was chosen to reduce <strong>the</strong> impact <strong>of</strong> wave<strong>for</strong>m binning effects; extrapolations<br />
down to <strong>the</strong> baseline possess a greater lever <strong>and</strong> <strong>the</strong>reby larger error potential than<br />
interpolations between two known points <strong>of</strong> a relatively smooth curve. A disadvantage is<br />
<strong>the</strong> crossing point’s obvious dependence on <strong>the</strong> pulse charge, or <strong>the</strong> need <strong>for</strong> two constants<br />
to compensate <strong>for</strong> it.<br />
The trailing edge time t end is required to calculate <strong>the</strong> width t − t end <strong>of</strong> <strong>the</strong> pulse;<br />
“Simple” defines it as <strong>the</strong> point <strong>of</strong> threshold crossing determined by a linear interpolation<br />
between <strong>the</strong> points (i+j −1, w i+j−1 ) <strong>and</strong> (i+j, w i+j ), where bin i+j is <strong>the</strong> first one to fall<br />
below w feat <strong>for</strong> j > 0. No fur<strong>the</strong>r corrections are applied because <strong>the</strong> achieved precision<br />
<strong>for</strong> <strong>the</strong> width is sufficient. If <strong>the</strong> wave<strong>for</strong>m ends prematurely, <strong>the</strong> wave<strong>for</strong>m’s ending time<br />
is taken instead <strong>and</strong> <strong>the</strong> pulse is marked as cut <strong>of</strong>f in <strong>the</strong> extra-info series.<br />
The pulse’s charge is defined as <strong>the</strong> total charge contained in <strong>the</strong> bins i − 1 to i + j,<br />
q = ∑ i+j<br />
k=i−1 w k. A potential charge related to a non-zero baseline is not substracted as<br />
NFE explicitly relies on a correct baseline from DOMcalibrator.<br />
If a pulse is added to <strong>the</strong> pulse series depends on whe<strong>the</strong>r <strong>the</strong> feature passes a second<br />
threshold w detect <strong>and</strong> whe<strong>the</strong>r q ≥ q min . These two criteria can be disabled by setting<br />
w detect = w feat <strong>and</strong> q min = 0, respectively. However having at least one <strong>of</strong> <strong>the</strong>m active<br />
allows w feat to be set to a substantially lower value without erroneously extracting baseline<br />
fluctuations (see figure 6.2).<br />
The two thresholds <strong>and</strong> <strong>the</strong> two charge compensation constants can be set individually<br />
<strong>for</strong> ATWD <strong>and</strong> FADC, <strong>the</strong> time <strong>of</strong>fset constant can be set <strong>for</strong> ATWD OT, ATWD NT<br />
<strong>and</strong> FADC, <strong>and</strong> q min is <strong>the</strong> same <strong>for</strong> all sources, leading to a total <strong>of</strong> twelve configuration<br />
parameters.<br />
6.3 Extraction Algorithm “BayesUnfold”<br />
“BayesUnfold” (BU ) uses <strong>the</strong> method <strong>of</strong> Bayesian Unfolding described by G. D’Agostini<br />
in his paper A multidimensional unfolding method based on Bayes’ <strong>the</strong>orem.[57] It is also<br />
employed in <strong>Feature</strong><strong>Extractor</strong> <strong>and</strong> Pulse<strong>Extractor</strong>; this section explains <strong>the</strong> implementation<br />
in NFE, section 6.3.1 shows <strong>the</strong> differences to FE’s <strong>and</strong> PE’s implementations, <strong>and</strong><br />
appendix A gives a more <strong>for</strong>mal approach to <strong>the</strong> unfolding itself.<br />
50
6.3 Extraction Algorithm “BayesUnfold”<br />
Figure 6.3: Sketch illustrating Bayesian Unfolding as employed in FE, PE <strong>and</strong> NFE; see text<br />
<strong>for</strong> description. Note that <strong>the</strong> deconvoluted wave<strong>for</strong>m is not drawn to scale, <strong>the</strong><br />
unfolding is charge conserving.<br />
The underlying idea behind unfolding techniques is to undo smearing <strong>and</strong> distortion<br />
effects caused by <strong>the</strong> experiment’s hardware. Afterwards,<br />
x<br />
x it is easier to reconstruct <strong>the</strong><br />
arrival times <strong>of</strong> photons at <strong>the</strong> photocathode.<br />
Besides <strong>the</strong> calibrated wave<strong>for</strong>m to extract, <strong>the</strong> unfolding algorithm requires samples <strong>of</strong><br />
generic SPE pulses in <strong>the</strong> same binning. It <strong>the</strong>n deconvolutes <strong>the</strong> wave<strong>for</strong>m iteratively,<br />
moving <strong>the</strong> charge into those bins in which SPE-like pulses with <strong>the</strong> given charge must<br />
have occured to cause <strong>the</strong> actual wave<strong>for</strong>m with maximum probability. An illustration<br />
<strong>for</strong> this is given in figure 6.3: The wave<strong>for</strong>m to <strong>the</strong> left side is <strong>the</strong> superposition <strong>of</strong> five<br />
SPE-like features, <strong>the</strong> deconvoluted distribution to <strong>the</strong> right side is a histogram <strong>of</strong> <strong>the</strong><br />
same length <strong>and</strong> total charge, but <strong>the</strong> charge has been moved to <strong>the</strong> bins at <strong>the</strong> beginnings<br />
<strong>of</strong> <strong>the</strong> individual features. For more details, see appendix A.<br />
x<br />
Starting Distribution One <strong>of</strong> <strong>the</strong> unfolding method’s parameters is <strong>the</strong> starting distribution<br />
<strong>for</strong> <strong>the</strong> deconvoluted distribution, u i, 0 , i = 0 . . . L. As D’Agostini points out,<br />
<strong>the</strong> method “gives <strong>the</strong> best results (in terms <strong>of</strong> its ability to reproduce <strong>the</strong> true distribution)<br />
if one makes a realistic guess about <strong>the</strong> distribution that <strong>the</strong> true values follow, but,<br />
in case <strong>of</strong> total ignorance, satisfactory results are obtained even starting from a uni<strong>for</strong>m<br />
distribution”.[57]<br />
The SPE samples are zero everywhere but in a small region behind <strong>the</strong> position <strong>of</strong> <strong>the</strong><br />
hit. Because <strong>of</strong> this, a combination <strong>of</strong> an uni<strong>for</strong>m distribution with <strong>the</strong> shifted wave<strong>for</strong>m<br />
provides a good initial guess:<br />
u i, 0 = 1 2<br />
w tot<br />
L + 1 2 w k, k := (i + 1) mod L, w tot = ∑ w i (6.2)<br />
The shift <strong>of</strong> one bin length is introduced to minimize a possible bias to later pulse times<br />
<strong>and</strong> to speed up <strong>the</strong> unfolding, because <strong>the</strong> charge always needs to be shifted to <strong>the</strong><br />
beginning <strong>of</strong> <strong>the</strong> features. The uni<strong>for</strong>m term is preserved because <strong>the</strong> algorithm can not<br />
increase bin contents which are zero at any time during <strong>the</strong> deconvolution process, <strong>and</strong><br />
does so only slowly <strong>for</strong> very low bins contents.<br />
51
6 ALGORITHMS IMPLEMENTED IN NFE<br />
Number <strong>of</strong> Iterations An important parameter <strong>of</strong> <strong>the</strong> algorithm is <strong>the</strong> number <strong>of</strong><br />
iterations n iter . Steering <strong>the</strong> number <strong>of</strong> iterations adaptively can both reduce computation<br />
time <strong>and</strong> improve <strong>the</strong> quality <strong>of</strong> extraction <strong>for</strong> “tricky” wave<strong>for</strong>ms by giving <strong>the</strong>m more<br />
CPU time <strong>and</strong> <strong>the</strong>reby more iterations.<br />
Moreover, terminating <strong>the</strong> unfolding procedure at <strong>the</strong> right time is <strong>the</strong> suppression <strong>of</strong> <strong>the</strong><br />
method’s inherent amplification <strong>of</strong> statistical fluctuations (positive feedback). D’Agostini<br />
suggests to smooth <strong>the</strong> deconvoluted distribution after every iteration step to both speed<br />
up <strong>the</strong> convergence <strong>and</strong> to decrease <strong>the</strong> amount <strong>of</strong> artifacts. However, this does not seem<br />
to be applicable in this case since <strong>the</strong> target deconvolution should be spiky. 5<br />
The “BayesUnfold” algorithm in NFE employs a conjunction <strong>of</strong> two stopping conditions:<br />
Both <strong>the</strong> change <strong>of</strong> <strong>the</strong> highest bin <strong>and</strong> <strong>the</strong> sum <strong>of</strong> all bins above half <strong>the</strong> pulse charge<br />
threshold q min must be small <strong>for</strong> two successive iteration steps (with <strong>the</strong> constraint n iter ≥<br />
10):<br />
max ({u i, ñ })<br />
max ({u i, ñ−1 }) < ∆u min<br />
∧<br />
∑<br />
i: u i, ñ > 0.5 q min<br />
u i, ñ<br />
∑<br />
i: u i, ñ−1 > 0.5 q min<br />
u i, ñ−1<br />
< 0.3 ∆u min ∀ ñ ∈ {n iter−1 , n iter }<br />
(6.3)<br />
Both conditions depend only on relatively few bins which reduces <strong>the</strong> impact <strong>of</strong> <strong>the</strong> a<strong>for</strong>ementioned<br />
fluctuations. The first condition ensures that <strong>the</strong> highest pulse is extracted<br />
precisely, <strong>and</strong> with it <strong>the</strong> o<strong>the</strong>r pulses as <strong>the</strong>y received as many iterations. The second<br />
condition ensures that no charge associated to pulses is lost. The factor <strong>of</strong> 0.3 has been<br />
determined empirically during this <strong>the</strong>sis, using <strong>the</strong> unit test’s extra output <strong>for</strong> idealized<br />
pulses. Additionally, a maximum number <strong>of</strong> iterations n max can be specified.<br />
Pulse Definition When <strong>the</strong> unfolding procedure ends after n iter iterations, <strong>the</strong> features<br />
in <strong>the</strong> deconvoluted distribution u i := u i, niter have to be extracted to define <strong>the</strong> pulses.<br />
“BayesUnfold” uses three subsequent bins to define a pulse (figure 6.4). This is motivated<br />
by two facts: First, even in an ideal case, SPE-pulse-like features have to be unfolded into<br />
two subsequent bins in a deterministic ratio if <strong>the</strong>y occured at a different time than that<br />
<strong>of</strong> a bin – i.e., “between” two bins. Secondly, only a finite number <strong>of</strong> iterations (∼ 30) is<br />
conducted, so SPE-pulse-like features whose charge should have been moved into a single<br />
bin in fact end up in this main bin <strong>and</strong> <strong>the</strong> surrounding ones.<br />
The deconvoluted distribution is scanned from its beginning to its end. If <strong>for</strong> a given i u i<br />
exceeds 0.5 q min , “BayesUnfold” checks whe<strong>the</strong>r u i+1 > u i ; if so, i is incremenetd by one.<br />
The three bins defining a pulse are u i−1 , u i , <strong>and</strong> u i+1 . The algorithm <strong>the</strong>n checks whe<strong>the</strong>r<br />
<strong>the</strong> third bin u i+1 should be accounted completely to <strong>the</strong> current pulse’s charge q or if it<br />
5 It might be worthwhile to try out advanced smoothing methods, e. g. a multi-resolution Savitzky-<br />
Golay filter as described in http://www.er.ams.eng.osaka-u.ac.jp/Paper/2006/Norbert06a.pdf.<br />
52
6.3 Extraction Algorithm “BayesUnfold”<br />
Figure 6.4: Sketch illustrating <strong>the</strong> algorithms used by <strong>the</strong> different feature extractors to define<br />
pulses out <strong>of</strong> <strong>the</strong> unfolded distribution; areas <strong>of</strong> <strong>the</strong> same color belong to a single<br />
pulse. Dark dashed lines indicate <strong>the</strong> charge thresholds, light dashed lines indicate<br />
<strong>the</strong> threshold boundaries in which <strong>the</strong> pulse definition does not change.<br />
left: NFE; middle: FE/PE with <strong>the</strong> same threshold;<br />
right: FE/PE with a higher threshold; striped areas apply to FE only.<br />
needs to be shared with a later pulse (see figure 6.4):<br />
⎧<br />
q = u i−1 + u i + f u i+1 with<br />
⎨ u i<br />
u<br />
f := i +u i+2<br />
u i+2 ≥ min ({u i , u i+1 })<br />
⎩1 o<strong>the</strong>rwise<br />
If <strong>the</strong> total charge q exceeds q min , <strong>the</strong> pulse is accepted (i.e., written to <strong>the</strong> series) <strong>and</strong><br />
fu i+1 is substracted from bin i + 1 to prevent <strong>the</strong> charge from being accounted twice.<br />
The pulse’s leading edge time is defined according to <strong>the</strong> charge distribution:<br />
(<br />
t = t i + T bin −<br />
u i−1<br />
+ f<br />
u i−1 + u i<br />
)<br />
u i+1<br />
u i + u i+1<br />
The pulse width T is obtained by a charge-dependent parametrization <strong>of</strong> <strong>the</strong> SPE pulse<br />
parametrization’s width at a height <strong>of</strong> 0.05 PE per bin:<br />
T = T bin max ({1, c T1 ln (c T2 q − c T3 )}) (6.4)<br />
The constants c T1 , c T2 , <strong>and</strong> c T3 are hard-coded <strong>for</strong> ATWD OT, ATWD NT, <strong>and</strong> FADC<br />
respectively. The parametrization error is usually well below 1%, see figure 6.5. At <strong>the</strong><br />
beginning <strong>and</strong> at <strong>the</strong> end <strong>of</strong> <strong>the</strong> wave<strong>for</strong>m, <strong>the</strong> algorithm switches to linear calculations<br />
<strong>and</strong> marks pulses as cut <strong>of</strong>f.<br />
In summary, “BayesUnfold” <strong>of</strong>fers three configuration parameters: <strong>the</strong> maximum number<br />
<strong>of</strong> iterations, <strong>the</strong> stopping parameter ∆u min , <strong>and</strong> <strong>the</strong> charge threshold q min . O<strong>the</strong>r<br />
parameters, such as <strong>the</strong> SPE pulse samples, <strong>the</strong> width parametrization, <strong>and</strong> <strong>the</strong> time<br />
<strong>of</strong>fsets are hard-coded, but can easily be changed.<br />
53
6 ALGORITHMS IMPLEMENTED IN NFE<br />
0,01<br />
relative error DT / T<br />
0<br />
−0,01<br />
−0,02<br />
2 4 6 8 10 12 14 16 18 20<br />
q / PE<br />
NFE ATWD OT<br />
NFE ATWD NT<br />
NFE FADC<br />
Figure 6.5: Residual plot comparing <strong>the</strong> width parametrizations used in “BayesUnfold” to <strong>the</strong><br />
true SPE pulse parametrizations’ widths in dependence <strong>of</strong> <strong>the</strong> pulse’s charge.<br />
6.3.1 Differences <strong>of</strong> FE’s <strong>and</strong> PE’s <strong>Implementation</strong>s<br />
<strong>Feature</strong><strong>Extractor</strong> <strong>and</strong> Pulse<strong>Extractor</strong> both use <strong>the</strong> same SPE pulse parametrization (shown<br />
in figure 4.6) which differs significantly from NFE’s; <strong>the</strong>y do not discriminate between old<br />
toroid <strong>and</strong> <strong>new</strong> toroid DOMs. The unfolding <strong>and</strong> pulse definition algorithms <strong>of</strong> FE <strong>and</strong><br />
PE are identical except <strong>for</strong> <strong>the</strong>ir h<strong>and</strong>ling <strong>of</strong> pulses below <strong>the</strong> threshold <strong>and</strong> <strong>the</strong> <strong>Feature</strong><strong>Extractor</strong>’s<br />
time refinement: <strong>Feature</strong><strong>Extractor</strong> uses <strong>the</strong> time obtained from one <strong>of</strong> its<br />
single-pulse extraction algorithms to replace <strong>the</strong> time <strong>of</strong> <strong>the</strong> nearest Bayesian Unfolding<br />
pulse.<br />
Both extractors conduct a fixed number <strong>of</strong> unfolding iterations (20) <strong>and</strong> define pulses<br />
by two bins compared to <strong>the</strong> three bins per pulse used by NFE. Because <strong>of</strong> this, <strong>the</strong>y<br />
tend to split up features into many pulses <strong>and</strong> <strong>the</strong>re<strong>for</strong>e require a comparably high charge<br />
threshold to reduce <strong>the</strong> splitting. The only difference with non-negligible impact between<br />
<strong>the</strong> FE <strong>and</strong> PE algorithms is how <strong>the</strong>y h<strong>and</strong>le charge in unfolded bins that fall below this<br />
threshold. While Pulse<strong>Extractor</strong> omits <strong>the</strong>m <strong>and</strong> <strong>the</strong>reby loses some <strong>of</strong> <strong>the</strong> pulses’ charge,<br />
<strong>Feature</strong><strong>Extractor</strong> compares <strong>the</strong> sum <strong>of</strong> all pulses’ charges to <strong>the</strong> charge computed by <strong>the</strong><br />
selected single-pulse extraction algorithm (preferably <strong>the</strong> second one, as it computes <strong>the</strong><br />
total charge which is needed here) <strong>and</strong> multiplies every pulse’s charge with this ratio to<br />
ensure that <strong>the</strong> wave<strong>for</strong>m’s charge is conserved. The difference between <strong>the</strong>se two methods<br />
is depicted as striped areas in figure 6.4.<br />
The advantage <strong>of</strong> <strong>Feature</strong><strong>Extractor</strong>’s method is that no pulse charge is lost, <strong>the</strong> disadvantage<br />
is an increased sensitivity to wrong baselines, e. g. those caused by faulty droop<br />
correction. A disadvantage common to FE <strong>and</strong> PE is that <strong>the</strong> high threshold causes<br />
a shift in <strong>the</strong> pulse’s time if from a pulse originally consisting <strong>of</strong> two bins only one is<br />
incorporated, because <strong>the</strong> time depends on <strong>the</strong> charge’s center <strong>of</strong> gravity.<br />
54
6.4 Extraction Algorithm “SLCHE”<br />
parabola through all three points<br />
time: point <strong>of</strong> maximum + <strong>of</strong>fset<br />
charge: const × area below parabola<br />
width: parametrization T(q)<br />
time: pre-maximum bin + const × ratio <strong>of</strong><br />
maximum bin value to predecessor bin value<br />
charge: const × maximum bin value<br />
width: three bin lengths<br />
Figure 6.6: Sketch comparing NFE’s extraction algorithm “SLCHE” (left) to SLCHit<strong>Extractor</strong><br />
(right). Shown are three different digitizations <strong>of</strong> <strong>the</strong> same SPE-like feature (dashed<br />
brown lines), <strong>the</strong> third being an irregular chargestamp.<br />
6.4 Extraction Algorithm “SLCHE”<br />
“SLCHE” – named after <strong>the</strong> SLCHit<strong>Extractor</strong> on which it was based originally – is an<br />
exclusive SLC extraction algorithm. For regular charge stamps, i. e., those in which<br />
<strong>the</strong> middle bin w 1 is <strong>the</strong> highest (section 4.4.3), <strong>the</strong> algorithm computes <strong>the</strong> position <strong>of</strong><br />
<strong>the</strong> maximum <strong>of</strong> <strong>the</strong> parabola defined by <strong>the</strong> three points (0, w 0 ) . . . (2, w 2 ), as well as<br />
<strong>the</strong> area A enclosed by it <strong>and</strong> <strong>the</strong> baseline. To calculate <strong>the</strong> pulse time, <strong>the</strong> time t pmax<br />
corresponding to <strong>the</strong> parabola’s maximum is shifted by a constant negative time <strong>of</strong>fset<br />
t <strong>of</strong>fset :<br />
t = t pmax + t <strong>of</strong>fset , with t pmax = t 0 + T bin<br />
−3w 0 + 4w 1 − w 2<br />
−2w 0 + 4w 1 − 2w 2<br />
.<br />
The charge is computed by q = c q A, where c q is a configurable constant. It is used to<br />
estimate <strong>the</strong> original feature’s width, employing <strong>the</strong> parametrization in equation (6.4)<br />
<strong>and</strong> <strong>the</strong>reby exploiting that full FADC wave<strong>for</strong>ms <strong>and</strong> SLC chargestamps have <strong>the</strong> same<br />
shape. Figure 6.6 visualizes this algorithm at <strong>the</strong> two leftmost chargestamps.<br />
Irregular wave<strong>for</strong>ms are not necessarily concave (see section 4.4.3), <strong>the</strong>re<strong>for</strong>e <strong>the</strong> calculations<br />
<strong>of</strong> t pmax <strong>and</strong> A fail <strong>for</strong> <strong>the</strong>m. Also, two or more photons hitting a DOM with<br />
a small delay could cause wave<strong>for</strong>ms that are approximately flat, leading to erroneously<br />
high charge estimates.<br />
Because <strong>of</strong> this, a fallback method is implemented:<br />
t = t i + t <strong>of</strong>fset , <strong>and</strong> q = c q<br />
〈 A<br />
w 1<br />
〉<br />
w i ,<br />
55
6 ALGORITHMS IMPLEMENTED IN NFE<br />
with i = 0 if w 0 > w 1 , <strong>and</strong> o<strong>the</strong>rwise i = 2 if w 2 > w 1 or w 0 + w 2 ≥ 1.8 w 1 . 〈 〉<br />
A<br />
w 1<br />
denotes<br />
<strong>the</strong> mean A to w 1 ratio <strong>for</strong> regular charge stamps, determined empirically.<br />
This fallback charge calculation is equivalent to <strong>the</strong> regular charge calculation 〈 per<strong>for</strong>med 〉 in<br />
SLCHit<strong>Extractor</strong> <strong>and</strong> could be expressed with a single constant ˜c q = c A<br />
q w 1<br />
. However,<br />
<strong>the</strong> more complicated looking <strong>for</strong>mulation with two multiplicative constants decouples<br />
<strong>the</strong>se constants <strong>and</strong> <strong>the</strong>reby simplifies configuration. Apart from c q , 〈 〉<br />
A<br />
w 1<br />
, <strong>and</strong> t<strong>of</strong>fset ,<br />
<strong>the</strong>re are no o<strong>the</strong>r configuration constants.<br />
56
CHAPTER VII<br />
Per<strong>for</strong>mance Optimization
7 PERFORMANCE OPTIMIZATION<br />
Prior to <strong>the</strong> design <strong>of</strong> <strong>the</strong> algorithm, many wave<strong>for</strong>ms were studied by eye (e.g., figure<br />
7.1, visualized using a custom IceTray Python module <strong>and</strong> matplotlib) to get a feeling<br />
<strong>for</strong> pulse shapes, baseline fluctuations, <strong>and</strong> potential problems arising with <strong>the</strong> feature<br />
extraction. Most <strong>of</strong> <strong>the</strong> hard-coded <strong>and</strong> default parameters were based on this experience<br />
<strong>and</strong> proved sustainable <strong>and</strong> robust under fur<strong>the</strong>r investigation.<br />
Important per<strong>for</strong>mance observables <strong>for</strong> <strong>the</strong> feature extraction are <strong>the</strong> time difference<br />
∆t between <strong>the</strong> photon hitting <strong>the</strong> DOM <strong>and</strong> <strong>the</strong> extracted pulse, <strong>the</strong> charge per pulse,<br />
<strong>the</strong> number <strong>of</strong> pulses, <strong>the</strong> total charge per wave<strong>for</strong>m, <strong>and</strong> <strong>the</strong> number <strong>of</strong> wave<strong>for</strong>ms in<br />
which no pulse is found. Measurements <strong>of</strong> <strong>the</strong>se observables have been conducted using<br />
custom Monte-Carlo datasets simulated <strong>for</strong> <strong>the</strong> calibration <strong>of</strong> NFE. These datasets (2595<br />
<strong>and</strong> 3071) include <strong>the</strong> Monte-Carlo hit in<strong>for</strong>mation which is usually discarded after <strong>the</strong><br />
wave<strong>for</strong>m simulation. This allows to directly evaluate <strong>the</strong> time resolution.[58][59] If not<br />
stated o<strong>the</strong>rwise, plots are based dataset 3071. Un<strong>for</strong>tunately, all simulated datasets<br />
include erroneous data caused by bugs partly found as a consequence <strong>of</strong> this <strong>the</strong>sis; this<br />
includes high baselines <strong>for</strong> ATWD <strong>and</strong> wrong impedances in <strong>the</strong> GCD files, see appendix<br />
C.3. The impact <strong>of</strong> <strong>the</strong>se bugs is discussed in <strong>the</strong> appropriate sections. Independently,<br />
<strong>the</strong> calibrations <strong>and</strong> tests presented here will be repeated when fixed datasets become<br />
available.<br />
After all parameters were fixed, <strong>the</strong> measurements were repeated with experimental data<br />
to verify <strong>the</strong> parameters’ correctness.<br />
7.1 Calibration Using Monte-Carlo Data<br />
The time <strong>of</strong>fsets (section 5.4.3) <strong>of</strong> all algorithms have been adjusted to comply with pulses<br />
extracted by “BayesUnfold” from ATWD <strong>new</strong> toroid wave<strong>for</strong>ms, as <strong>the</strong> author assumes<br />
<strong>the</strong>m to be <strong>the</strong> most reliable pulses on average. A constant common time <strong>of</strong>fset <strong>of</strong> all<br />
algorithms to Monte-Carlo hits is irrelevant as long as it is <strong>of</strong> <strong>the</strong> order ten nanoseconds;<br />
it does not affect <strong>the</strong> track reconstructions as this only depends on time differences between<br />
pulses, <strong>and</strong> it is far too small to affect later high-level analyses. The policy chosen is to<br />
extract <strong>the</strong> pulse start times instead <strong>of</strong> <strong>the</strong> Monte-Carlo hit times, which are about 11 ns<br />
too late in comparison <strong>and</strong> roughly tag a SPE pulse’s maximum, see figure 7.1, figure 7.2,<br />
<strong>and</strong> figure 7.3.<br />
7.1.1 Pre-evaluation Algorithm “Eva”<br />
First, <strong>the</strong> parameters <strong>of</strong> <strong>the</strong> pre-evaluation algorithm “Eva” <strong>and</strong> <strong>the</strong>reby <strong>the</strong> categorization<br />
<strong>of</strong> wave<strong>for</strong>ms were adjusted, as <strong>the</strong> per<strong>for</strong>mance <strong>of</strong> o<strong>the</strong>r algorithms depends on <strong>the</strong>m.<br />
They were determined using individual wave<strong>for</strong>ms such as those in figure 7.1; <strong>the</strong>y have<br />
been verified using <strong>the</strong> SPE parametrizations employed in “BayesUnfold” (equation (7.1),<br />
figure 4.6).<br />
58
7.1 Calibration Using Monte-Carlo Data<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
0.99<br />
NFE_ATWDPulses<br />
OMKey(5,41)<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
14900 15000 15100 15200 15300 15400 15500 15600 15700<br />
NFE_FADCPulses<br />
time / ns<br />
0.4<br />
0.8<br />
0.88<br />
0.57<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
14900 15000 15100 15200 15300 15400 15500 15600 15700<br />
NFE_ATWDPulses<br />
OMKey(12,33)<br />
1.51<br />
0.58<br />
1.64<br />
0.73 0.70<br />
0.60<br />
0.6<br />
0.4<br />
0.2<br />
0.0<br />
13400 13500 13600 13700 13800 13900 14000 14100 14200<br />
NFE_FADCPulses<br />
time / ns<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
1.60<br />
1.93 0.69 0.61 1.92<br />
1.58<br />
0.74<br />
0.0<br />
13400 13500 13600 13700 13800 13900 14000 14100 14200<br />
Figure 7.1: Two examples <strong>for</strong> calibrated ATWD (blue, narrow) <strong>and</strong> FADC (red, wide) wave<strong>for</strong>ms;<br />
only a part <strong>of</strong> <strong>the</strong> FADC wave<strong>for</strong>m is shown. Pulses are indicated by dashed<br />
full-length vertical lines with <strong>the</strong> charge q given at top <strong>of</strong> <strong>the</strong> lines in units <strong>of</strong> PE;<br />
horizontal bars indicate <strong>the</strong> pulse widths. Lines in <strong>the</strong> upper image each correspond<br />
to ATWD pulses, lines in lower image to FADC.<br />
Large solid or dotted ticks at <strong>the</strong> bottom axis indicate Monte-Carlo hit in<strong>for</strong>mation<br />
if available.<br />
59
7 PERFORMANCE OPTIMIZATION<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
0.8<br />
0.7<br />
1.63<br />
NFE_ATWDPulses<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
OMKey(69,33)<br />
17500 17600 17700 17800 17900 18000 18100 18200 18300<br />
NFE_FADCPulses<br />
time / ns<br />
0.8<br />
1.60<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0.0<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
17500 17600 17700 17800 17900 18000 18100 18200 18300<br />
1.51<br />
0.58<br />
1.64<br />
0.73 0.70<br />
NFE_ATWDPulses<br />
0.60<br />
13400 13500 13600 13700 13800 13900 14000 14100 14200<br />
NFE_FADCPulses<br />
time / ns<br />
1.60<br />
0.74<br />
1.93 0.69 0.61 1.92<br />
1.58<br />
OMKey(12,33)<br />
0.0<br />
13400 13500 13600 13700 13800 13900 14000 14100 14200<br />
Figure 7.2: Two examples <strong>for</strong> calibrated ATWD (blue, narrow) <strong>and</strong> FADC (red, wide) wave<strong>for</strong>ms;<br />
only a part <strong>of</strong> <strong>the</strong> FADC wave<strong>for</strong>m is shown. Pulses are indicated by dashed<br />
full-length vertical lines with <strong>the</strong> charge q given at top <strong>of</strong> <strong>the</strong> lines in units <strong>of</strong> PE;<br />
horizontal bars indicate <strong>the</strong> pulse widths. Lines in <strong>the</strong> upper image each correspond<br />
to ATWD pulses, lines in lower image to FADC.<br />
Large solid or dotted ticks at <strong>the</strong> bottom axis indicate Monte-Carlo hit in<strong>for</strong>mation<br />
if available.<br />
60
7.1 Calibration Using Monte-Carlo Data<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
0.15<br />
0.10<br />
0.05<br />
0.00<br />
0.05<br />
0.15<br />
0.10<br />
0.05<br />
0.00<br />
0.05<br />
0.28<br />
NFE_ATWDPulses<br />
OMKey(11,54)<br />
67400 67500 67600 67700 67800 67900 68000 68100 68200<br />
NFE_FADCPulses<br />
time / ns<br />
0.17<br />
67400 67500 67600 67700 67800 67900 68000 68100 68200<br />
NFE_ATWDPulses<br />
OMKey(87,37)<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
1.0 0.95 1.44<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0.0<br />
67600 67800 68000 68200<br />
NFE_FADCPulses<br />
time / ns<br />
1.0 0.78 1.18 2.50 1.03<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0.0<br />
67600 67800 68000 68200<br />
Figure 7.3: Two examples <strong>for</strong> calibrated ATWD (blue, narrow) <strong>and</strong> FADC (red, wide) wave<strong>for</strong>ms;<br />
only a part <strong>of</strong> <strong>the</strong> FADC wave<strong>for</strong>m is shown. Pulses are indicated by dashed<br />
full-length vertical lines with <strong>the</strong> charge q given at top <strong>of</strong> <strong>the</strong> lines in units <strong>of</strong> PE;<br />
horizontal bars indicate <strong>the</strong> pulse widths. Lines in <strong>the</strong> upper image each correspond<br />
to ATWD pulses, lines in lower image to FADC.<br />
Large solid or dotted ticks at <strong>the</strong> bottom axis indicate Monte-Carlo hit in<strong>for</strong>mation<br />
if available.<br />
61
7 PERFORMANCE OPTIMIZATION<br />
Table 7.1: Default values used <strong>for</strong> pre-evaluation algorithm “Eva”.<br />
parameter name in code value<br />
ATWD<br />
FADC<br />
w max SimpleThreshold 0.40 PE 0.60 PE<br />
w feat <strong>Feature</strong>Threshold 0.08 PE 0.10 PE<br />
l max <strong>Feature</strong>MaxLength 6 bins 5 bins<br />
d min <strong>Feature</strong>MinDistance 4 bins 3 bins<br />
Table 7.2: Default values used <strong>for</strong> extraction algorithm “Simple”.<br />
parameter name in code value<br />
ATWD<br />
FADC<br />
w feat <strong>Feature</strong>Threshold 0.04 PE 0.08 PE<br />
w detect DetectionThreshold 0.04 PE 0.08 PE<br />
q min minCharge 0.15 PE 0.15 PE<br />
c P0 QTCorrelationP0 1.6 ns 17.722 ns<br />
c P1 QTCorrelationP1 1.7 1.2345<br />
t <strong>of</strong>fset NT DeltaTNewToroid −0.57 ns −10.03 ns<br />
t <strong>of</strong>fset OT DeltaTOldToroid −1.83 ns −10.03 ns<br />
The values were adjusted in such a way that narrow <strong>and</strong> well-seperated SPE-pulse-like<br />
features up to 1.5 PE in <strong>the</strong> spikiest possible binning are still considered to be simple<br />
(table 7.1). This is tailored to <strong>the</strong> extraction algorithm “Simple” as it is not capable <strong>of</strong><br />
splitting features, <strong>for</strong> example to account a feature’s tail to <strong>the</strong> associated peak if it is<br />
superimposed by ano<strong>the</strong>r feature. While SPE features with charges as high as 1.5 PE are<br />
quite common (e.g., figure 7.1, first pulse in DOM 12-33, or figure 7.2, DOM 69-33), very<br />
similar features can be caused by two or more coincident photons (thid pulse in DOM<br />
12-33); however as long as those are sufficiently close in time, this does not contradict a<br />
pulse’s definition.<br />
7.1.2 Extraction Algorithm “Simple”<br />
The extraction algorithm “Simple”’s feature thresholds w feat were set sufficiently high not<br />
to be surpassed by baseline fluctuations. For lower feature thresholds, <strong>the</strong> charge threshold<br />
q min would still reliably prevent individual r<strong>and</strong>om fluctuations to be extracted as<br />
pulses. However, fluctuations at <strong>the</strong> beginning <strong>of</strong> a real feature could impede <strong>the</strong> leading<br />
edge detection.<br />
62
7.1 Calibration Using Monte-Carlo Data<br />
10 7<br />
10 6<br />
width: 0.81 0.81<br />
mean: 11.19 11.20<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
40 20 0 20 40<br />
t MC −t pulse /ns<br />
10 7<br />
10 6<br />
width: 0.81 0.81<br />
mean: 11.19 11.20<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
40 20 0 20 40<br />
t MC −t pulse /ns<br />
Figure 7.4: Distribution <strong>of</strong> <strong>the</strong> time residuals <strong>of</strong> NFE extracted pulses <strong>for</strong> different charge<br />
thresholds; shown is <strong>the</strong> time difference between <strong>the</strong> first Monte-Carlo hit <strong>and</strong> <strong>the</strong><br />
first extracted pulse per wave<strong>for</strong>m.<br />
Blue areas are NFE with default settings (q min = 0.15 PE), green lines to <strong>the</strong><br />
left have q min = 0.2 PE <strong>for</strong> algorithm “Simple”, <strong>and</strong> green lines to <strong>the</strong> right have<br />
q min = 0.25 PE. Red lines indicate Gaussian fits to all bins (excluding underflow<br />
<strong>and</strong> overflow), with its width <strong>and</strong> its mean specified in <strong>the</strong> plot.<br />
63
7 PERFORMANCE OPTIMIZATION<br />
Tuning <strong>of</strong> <strong>the</strong> threshold q min requires a decision about whe<strong>the</strong>r some more false pulses outweigh<br />
a better detection capability <strong>for</strong> very small pulses. Generally, early falsely detected<br />
pulses (false pulses) can have a relatively high impact on <strong>the</strong> later track reconstruction,<br />
however <strong>the</strong> number <strong>of</strong> false first pulses extracted using <strong>the</strong> current low q min = 0.15 PE<br />
is deemed to be acceptable: For <strong>the</strong> about 4.1 million wave<strong>for</strong>ms evaluated in figure 7.4,<br />
only about 200 first pulses precede <strong>the</strong> expected time by more than 10 ns. For this optimization<br />
it has to be considered that increasing <strong>the</strong> threshold only marginally reduces<br />
<strong>the</strong> number <strong>of</strong> false pulses, but has a relatively high impact on true pulses:<br />
For <strong>the</strong> sample in figure 7.4, q min = 0.2 PE eliminates about 50 early false pulses <strong>of</strong> about<br />
23 million pulses in total, but almost triples <strong>the</strong> number <strong>of</strong> instances where <strong>the</strong> first hit<br />
was missed. This is seen by <strong>the</strong> underflow bin, which rise from about 11000 entries to<br />
about 31000 entries; q min = 0.25 PE eliminates additional 10 early false pulses at <strong>the</strong> cost<br />
<strong>of</strong> more than 40000 additional missed first hits. In general, an as low as possible setting<br />
is preferrable.<br />
In cases where only one feature is present inside <strong>the</strong> wave<strong>for</strong>m, <strong>the</strong> option En<strong>for</strong>cePulse<br />
in conjunction with “BayesUnfold” could be used to extract <strong>the</strong> pulse none<strong>the</strong>less if it<br />
is missed; however, <strong>the</strong> rising <strong>of</strong> <strong>the</strong> plateau at <strong>the</strong> top in figure 7.4 with increasing q min<br />
shows that <strong>of</strong>ten o<strong>the</strong>r features are extracted instead, preventing <strong>the</strong> extraction <strong>of</strong> <strong>the</strong><br />
valuable first feature (possibly direct primary lepton Čerenkov light).<br />
This calibration will be repeated with future <strong>new</strong> Monte-Carlo data, because a significant<br />
fraction <strong>of</strong> <strong>the</strong> fake pulses seems to be caused by artifacts from an incomplete simulation,<br />
see figure C.2 in appendix C.<br />
By default <strong>the</strong> detection threshold is disabled (w detect = w feat ). A different setting<br />
would hamper <strong>the</strong> detection <strong>of</strong> very flat features like those seen in figure 7.3 <strong>for</strong> DOM<br />
11-54, <strong>and</strong> q min is sufficiently suppressing noise hits as can be seen at <strong>the</strong> low fake pulse<br />
rate discussed above.<br />
The parameters c P0 <strong>and</strong> c P1 <strong>for</strong> <strong>the</strong> charge-time correlation compensation term t q in<br />
equation (6.1) were obtained by a fit to <strong>the</strong> two-dimensional charge-time histograms in figure<br />
7.5. While <strong>the</strong> agreement is not perfect, especially <strong>for</strong> very low charges, <strong>the</strong> quadratic<br />
model is robust, <strong>and</strong> <strong>the</strong> fit lies within <strong>the</strong> natural scattering <strong>of</strong> <strong>the</strong> distribution. For <strong>the</strong><br />
analyzed dataset, <strong>the</strong> time resolution measured by a Gaussian fit to <strong>the</strong> distribution seen<br />
in figure 7.4 (excluding <strong>the</strong> underflow bin) improves from 1.04 ns to 0.81 ns <strong>for</strong> ATWD<br />
<strong>and</strong> from 4.68 ns to 3.34 ns <strong>for</strong> FADC.<br />
The double structure in figure 7.5 also shows that <strong>the</strong> ATWD time distribution is heavily<br />
influenced by <strong>the</strong> erroneous GCD in<strong>for</strong>mation (geometry, calibratoin, detector status;<br />
section 4.6.1), which prevents correct categorization <strong>of</strong> old toroid <strong>and</strong> <strong>new</strong> toroid DOMs.<br />
The time resolution will probably improve fur<strong>the</strong>r as soon as this general calibration problem<br />
is fixed. A rough estimate <strong>of</strong> this bug’s impact can be obtained by extracting pulses<br />
exclusively from old toroid DOMs because much fewer NT DOMs are reported to have<br />
an old toroid than <strong>the</strong> o<strong>the</strong>r way around. The Gaussian width is 0.68 ns in this case.<br />
The same test applied to FADC pulses yields no improvements in time resolution because<br />
64
7.1 Calibration Using Monte-Carlo Data<br />
20<br />
20<br />
15<br />
9000<br />
15<br />
9000<br />
7500<br />
7500<br />
10<br />
10<br />
t MC −t pulse /ns<br />
5<br />
6000<br />
4500<br />
t MC −t pulse /ns<br />
5<br />
6000<br />
4500<br />
0<br />
3000<br />
0<br />
3000<br />
5<br />
1500<br />
5<br />
1500<br />
10<br />
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0<br />
q pulse1<br />
0<br />
10<br />
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0<br />
q pulse1<br />
0<br />
25<br />
2000<br />
25<br />
2000<br />
t MC −t pulse /ns<br />
20<br />
15<br />
10<br />
5<br />
0<br />
1750<br />
1500<br />
1250<br />
1000<br />
t MC −t pulse /ns<br />
20<br />
15<br />
10<br />
5<br />
0<br />
1750<br />
1500<br />
1250<br />
1000<br />
5<br />
750<br />
5<br />
750<br />
10<br />
500<br />
10<br />
500<br />
15<br />
250<br />
15<br />
250<br />
20<br />
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0<br />
q pulse1<br />
0<br />
20<br />
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0<br />
q pulse1<br />
0<br />
Figure 7.5: Charge-time correlation <strong>of</strong> pulses from NFE extraction algorithm “Simple” only;<br />
left: uncorrected, right: default c P0 <strong>and</strong> c P1 ; top: ATWD, bottom: FADC.<br />
<strong>the</strong> toroid’s effect is smeared out by <strong>the</strong> pulse shaper. Never<strong>the</strong>less, fur<strong>the</strong>r tests will be<br />
undertaken to check if it can be worthwile to differentiate between FADC OT <strong>and</strong> NT as<br />
soon as improved Monte-Carlo datasets become available.<br />
7.1.3 Extraction Algorithm “BayesUnfold”<br />
As pointed out in section 6.3, “BayesUnfold” requires generic templates <strong>of</strong> SPE pulses in<br />
<strong>the</strong> same binning as <strong>of</strong> <strong>the</strong> wave<strong>for</strong>ms. Basis <strong>for</strong> <strong>the</strong>se samples was Christopher Wendt’s<br />
parametrization, calculated from flasher run data (figure 4.6):[47]<br />
w SPE (t) = a 1<br />
(<br />
e<br />
−a 2 t+a 3<br />
+ e a 4t−a 5<br />
) −8<br />
, (7.1)<br />
with <strong>the</strong> parameter values specified in table 7.3. With <strong>the</strong> given parametrization, <strong>the</strong> time<br />
<strong>of</strong>fsets <strong>of</strong> both <strong>the</strong> binning <strong>and</strong> <strong>the</strong> parametrization itself are still arbitrary. The binning<br />
was chosen not to be shifted against <strong>the</strong> arbitrary <strong>of</strong>fset <strong>of</strong> <strong>the</strong> parametrization. For<br />
ATWD NT/OT <strong>the</strong> samples start at bin 1 <strong>and</strong> end with bin 12 <strong>and</strong> 11, respectively; <strong>the</strong><br />
FADC sample reaches from bin 2 to <strong>the</strong> end <strong>of</strong> bin 10. These selections include 99.47%,<br />
65
7 PERFORMANCE OPTIMIZATION<br />
Table 7.3: Parameter values <strong>of</strong> <strong>the</strong> SPE pulse parametrization employed, <strong>and</strong><br />
<strong>of</strong> <strong>the</strong> pulse width parametrization, equations (7.1) <strong>and</strong> (6.4).<br />
parameter<br />
value<br />
ATWD NT ATWD OT FADC<br />
a 1 4.422419 mV 4.240862 mV 8.793769 mV<br />
a 2 0.6537139 ns −1 0.7588095 ns −1 0.8369602 ns −1<br />
a 3 0.1049991 0.0743131 0.3809843<br />
a 4 0.08669082 ns −1 0.09235075 ns −1 0.13469828 ns −1<br />
a 5 0.01385802 0.00904426 0.06131466<br />
T bin · c T1 6.46351 ns 5.96069 ns 29.9302 ns<br />
c T2 30.856 PE −1 31.983 PE −1 53.063 PE −1<br />
c T3 4.8980 4.5038 6.3877<br />
Table 7.4: Default configuration values <strong>and</strong> hard-coded parameters<br />
used <strong>for</strong> <strong>the</strong> extraction algorithm “BayesUnfold”.<br />
parameter name in code value<br />
n max maxIterations<br />
minRelativeChange<br />
40<br />
0.012 PE<br />
speThreshold<br />
0.25 PE<br />
∆u min<br />
q min<br />
ATWD<br />
FADC<br />
L SPE NT NewToroid::speLength 11 8<br />
L SPE OT OldTOroid::speLength 10 8<br />
t <strong>of</strong>fset NT NewToroid::timeOffset 0 ns 6.66 ns<br />
t <strong>of</strong>fset OT OldToroid::timeOffset −1.82 ns 6.66 ns<br />
66
7.1 Calibration Using Monte-Carlo Data<br />
Table 7.5: Default values used <strong>for</strong> extraction algorithm “SLCHE”.<br />
parameter name in code value<br />
c q chargeCalibConst 1.1584<br />
〈 A<br />
w 1<br />
〉<br />
· Tbin meanParabolaArea 53.09 ns<br />
t <strong>of</strong>fset slcDeltaT −50.14 ns<br />
99.43%, <strong>and</strong> 99.60% <strong>of</strong> <strong>the</strong> SPE pulses’ charges, yet <strong>the</strong>y are sufficiently short to permit<br />
unfolding with high CPU efficiency.<br />
The parameters <strong>for</strong> <strong>the</strong> termination condition in equation (6.3) <strong>for</strong> <strong>the</strong> number <strong>of</strong><br />
iterations were obtained from idealized SPE pulses using <strong>the</strong> unit test’s extra output<br />
(option -as), <strong>and</strong> are listed in table 7.4. Comparisons between results with fixed <strong>and</strong><br />
variable numbers <strong>of</strong> iterations show that <strong>the</strong> time resolution improves slightly (2%) <strong>for</strong><br />
<strong>the</strong> latter, see figure 7.6; it also tends to extract more charge per pulse (figure 7.7) <strong>and</strong> in<br />
total . In general <strong>the</strong> extraction <strong>of</strong> more charge is not necessarily advantageous because<br />
it might originate from a wrong baseline. However, in this case it can be assumed that<br />
<strong>the</strong> extra charge belongs to <strong>the</strong> pulses <strong>and</strong> got lost because <strong>the</strong> unfolding was stopped<br />
too early. Thus <strong>the</strong> adaptive stopping condition is considered to be worthwile, see also<br />
section 8.5.<br />
Never<strong>the</strong>less, <strong>the</strong> extraction results could probably be fur<strong>the</strong>r improved by fine-tuning: As<br />
an example, <strong>the</strong> number <strong>of</strong> pulses per wave<strong>for</strong>m shown in figure 7.8 decreases on average<br />
when using a fixed n iter = 20, <strong>and</strong> it decreases fur<strong>the</strong>r <strong>for</strong> n iter = 30. This indicates that a<br />
significant fraction <strong>of</strong> <strong>the</strong> extra pulses which are extracted with <strong>the</strong> default settings might<br />
be caused by excessive splitting.<br />
After <strong>the</strong> termination conditions were optimized during development, <strong>the</strong> improvement<br />
gained by using <strong>the</strong> optimized starting condition instead <strong>of</strong> a flat one mostly vanished.<br />
The gain is only 0.8 iterations instead <strong>of</strong> 3 be<strong>for</strong>e optimization, see figure 7.9. Still, <strong>the</strong><br />
optimized starting condition is kept as <strong>the</strong>re are no drawbacks.<br />
7.1.4 Extraction Algorithm “SLCHE”<br />
For <strong>the</strong> calibration <strong>of</strong> <strong>the</strong> paramters <strong>of</strong> “SLCHE”, <strong>the</strong> algorithm has been applied to<br />
Monte-Carlo SLC charge stamps <strong>for</strong> which full wave<strong>for</strong>ms were available. The results<br />
were compared to pulses extracted with one <strong>of</strong> <strong>the</strong> o<strong>the</strong>r algorithms. Despite <strong>of</strong> ATWD<br />
wave<strong>for</strong>ms providing a better time resolution, FADC pulses were chosen <strong>for</strong> <strong>the</strong> comparison<br />
in order to minimize systematical errors due to differences between ATWD <strong>and</strong> FADC,<br />
as SLC charge stamps are generated from FADC wave<strong>for</strong>ms. Also, only wave<strong>for</strong>ms from<br />
DOMs which did not fulfill <strong>the</strong> hard local coincidence condition were used to get a more<br />
67
7 PERFORMANCE OPTIMIZATION<br />
10 7<br />
10 6<br />
width: 0.81 0.83<br />
mean: 11.11 11.07<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
40 30 20 10 0 10 20 30 40<br />
t MC −t pulse /ns<br />
10 7<br />
10 6<br />
width: 0.81 0.80<br />
mean: 11.11 11.12<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
40 30 20 10 0 10 20 30 40<br />
t MC −t pulse /ns<br />
Figure 7.6: Plots illustrating <strong>the</strong> effects <strong>of</strong> “BayesUnfold”’s variable number <strong>of</strong> iterations. Both<br />
simple <strong>and</strong> complex ATWD wave<strong>for</strong>ms have been extracted toge<strong>the</strong>r by BU.<br />
Shown are <strong>the</strong> distributions <strong>of</strong> <strong>the</strong> time residuals <strong>for</strong> <strong>the</strong> default BU (variable n iter ,<br />
〈n iter 〉 = 19.8; blue areas), <strong>and</strong> <strong>for</strong> BU with a fixed number <strong>of</strong> iterations (green<br />
lines; n iter = 20 on <strong>the</strong> top, n iter = 20 on <strong>the</strong> bottom). Red lines indicate Gaussian<br />
fits to all bins (excluding underflow <strong>and</strong> overflow).<br />
68
7.1 Calibration Using Monte-Carlo Data<br />
500000<br />
width: 0.34 0.34<br />
mean: 0.91 0.90<br />
400000<br />
300000<br />
entries<br />
200000<br />
100000<br />
0<br />
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0<br />
q pulse1<br />
500000<br />
width: 0.34 0.34<br />
mean: 0.91 0.91<br />
400000<br />
300000<br />
entries<br />
200000<br />
100000<br />
0<br />
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0<br />
q pulse1<br />
Figure 7.7: Plots illustrating <strong>the</strong> effects <strong>of</strong> “BayesUnfold”’s variable number <strong>of</strong> iterations. Both<br />
simple <strong>and</strong> complex ATWD wave<strong>for</strong>ms have been extracted toge<strong>the</strong>r by BU.<br />
Shown are <strong>the</strong> distributions <strong>of</strong> <strong>the</strong> charges <strong>of</strong> first pulses extracted by <strong>the</strong> default<br />
BU (variable n iter , 〈n iter 〉 = 19.8; blue areas), <strong>and</strong> <strong>for</strong> BU with a fixed number <strong>of</strong><br />
iterations (green lines; n iter = 20 on <strong>the</strong> top, n iter = 20 on <strong>the</strong> bottom). Red lines<br />
indicate Gaussian fits to all bins (excluding underflow <strong>and</strong> overflow).<br />
69
7 PERFORMANCE OPTIMIZATION<br />
10 7<br />
10 6<br />
width: 0.37<br />
mean: 0.04<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
10 5 0 5 10 15 20<br />
∆n pulses<br />
10 7<br />
10 6<br />
width: 0.36<br />
mean: 0.09<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
10 5 0 5 10 15 20<br />
∆n pulses<br />
Figure 7.8: Plots illustrating <strong>the</strong> effects <strong>of</strong> “BayesUnfold”’s variable number <strong>of</strong> iterations. Both<br />
simple <strong>and</strong> complex ATWD wave<strong>for</strong>ms have been extracted toge<strong>the</strong>r by BU.<br />
Shown are <strong>the</strong> distributions <strong>of</strong> <strong>the</strong> differences between <strong>the</strong> numbers <strong>of</strong> pulses extracted<br />
by <strong>the</strong> default BU (variable n iter , 〈n iter 〉 = 19.8) <strong>and</strong> BU with a fixed<br />
number <strong>of</strong> iterations (n iter = 20 on <strong>the</strong> top, n iter = 20 on <strong>the</strong> bottom). Positive<br />
values correspond to more pulses <strong>for</strong> default BU.<br />
70
7.2 Verification Using Experimental Data<br />
40000<br />
0cm<br />
60000<br />
35000<br />
30000<br />
25000<br />
50000<br />
40000<br />
entries<br />
20000<br />
entries<br />
30000<br />
15000<br />
20000<br />
10000<br />
5000<br />
10000<br />
0<br />
10 15 20 25 30 35 40<br />
number <strong>of</strong> iterations<br />
0<br />
10 15 20 25 30 35 40<br />
number <strong>of</strong> iterations<br />
Figure 7.9: Effects <strong>of</strong> “BayesUnfold”’s optimized deconvolution starting distribution:<br />
Shown is <strong>the</strong> distribution <strong>of</strong> <strong>the</strong> number <strong>of</strong> iterations n iter <strong>for</strong> complex wave<strong>for</strong>ms<br />
from ATWD (left) <strong>and</strong> FADC (right) <strong>for</strong> <strong>the</strong> default starting distribution (solid<br />
lines) <strong>and</strong> a uni<strong>for</strong>m one (dashed lines).<br />
specific sample <strong>for</strong> <strong>the</strong> calibration.<br />
The resulting parameter values are listed in table 7.5. An illustration <strong>of</strong> <strong>the</strong> good agreement<br />
between FADC <strong>and</strong> SLC in both time <strong>and</strong> charge can be seen in figure 7.10.<br />
7.2 Verification Using Experimental Data<br />
After all parameters have been adjusted to work well with simulated data, most analyses<br />
were repeated using experimental data to verify both <strong>the</strong> correctness <strong>of</strong> <strong>the</strong>se parameters<br />
<strong>and</strong> <strong>the</strong> correctness <strong>of</strong> <strong>the</strong> simulation itself. The dataset used is run 113587 (IC59, 2009-<br />
04-21), filtered to level 1 which means that it contains <strong>the</strong> events that passed <strong>the</strong> online<br />
filtering.<br />
The timing tests can not be repeated without modification because <strong>the</strong> true hit in<strong>for</strong>mation<br />
is not available <strong>for</strong> experimental data. Instead, <strong>the</strong> relative time distribution<br />
<strong>for</strong> different sources (e.g., ATWD <strong>and</strong> FADC) can be compared to <strong>the</strong> one obtained from<br />
Monte-Carlo simulations.<br />
The results <strong>of</strong> this test can be seen in figure 7.11. The small time misalignement <strong>of</strong><br />
about 4 ns between <strong>the</strong> pulses from ATWD <strong>and</strong> FADC can be attributed to a time <strong>of</strong>fset in<br />
<strong>the</strong> experimental dataset itself, caused by an old DOMcal version. This <strong>of</strong>fset is compensated<br />
<strong>for</strong> in DOMcalibrator by manually shifting FADC wave<strong>for</strong>ms by <strong>the</strong> recommended<br />
value <strong>of</strong> −15 ns. Still, <strong>the</strong> Monte-Carlo dataset already uses <strong>the</strong> <strong>new</strong> DOMcal version <strong>for</strong><br />
which <strong>the</strong> problem was fixed; <strong>the</strong>re<strong>for</strong>e, it is deemed to be more trustworthy. Final time<br />
<strong>of</strong>fset calibration values <strong>for</strong> NFE will be obtained using future IC79 Monte-Carlo data,<br />
which can be verified by future IC79 experimental data.<br />
71
7 PERFORMANCE OPTIMIZATION<br />
10 7<br />
10 6<br />
width: 3.04 2.25<br />
mean: 10.70 10.69<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
80 60 40 20 0 20 40 60<br />
t MC −t pulse /ns<br />
300000<br />
width: 0.34 0.33<br />
mean: 0.78 0.78<br />
250000<br />
200000<br />
entries<br />
150000<br />
100000<br />
50000<br />
0<br />
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0<br />
q pulse1<br />
Figure 7.10: Distributions <strong>of</strong> <strong>the</strong> time residuals <strong>and</strong> charges <strong>of</strong> “SLCHE” (green lines) <strong>and</strong><br />
FADC (blue areas) pulses from <strong>the</strong> same wave<strong>for</strong>ms, with <strong>the</strong> latter extracted by<br />
NFE with En<strong>for</strong>cePulse set.<br />
72
7.2 Verification Using Experimental Data<br />
250000<br />
width: 3.12<br />
mean: 0.39<br />
200000<br />
150000<br />
entries<br />
100000<br />
50000<br />
0<br />
6 4 2 0 2 4 6<br />
∆t pulse /ns<br />
250000<br />
width: 3.68<br />
mean: -3.46<br />
200000<br />
150000<br />
entries<br />
100000<br />
50000<br />
0<br />
10 8 6 4 2 0 2<br />
∆t pulse /ns<br />
Figure 7.11: Distributions <strong>of</strong> <strong>the</strong> time differences <strong>of</strong> ATWD <strong>and</strong> FADC NFE pulses <strong>for</strong> Monte-<br />
Carlo (top) <strong>and</strong> experimental data (bottom). Positive values indicate earlier<br />
times in <strong>the</strong> FADC.<br />
73
7 PERFORMANCE OPTIMIZATION<br />
1200<br />
1000<br />
Old Toroid DOMs<br />
New Toroid DOMs<br />
total<br />
350000<br />
300000<br />
Old Toroid DOMs<br />
New Toroid DOMs<br />
total<br />
800<br />
250000<br />
entries<br />
600<br />
400<br />
entries<br />
200000<br />
150000<br />
100000<br />
200<br />
50000<br />
0<br />
1 0 1 2 3 4<br />
FADC charge / ATWD charge<br />
0<br />
1 0 1 2 3 4<br />
FADC charge / ATWD charge<br />
Figure 7.12: Ratio <strong>of</strong> <strong>the</strong> total (integrated) charges <strong>of</strong> <strong>the</strong> ATWD wave<strong>for</strong>m <strong>and</strong> <strong>the</strong> corresponding<br />
fraction <strong>of</strong> <strong>the</strong> FADC wave<strong>for</strong>m.<br />
Left side: simulated dataset 3071; right side: experimental dataset L1 113587.<br />
The distributions <strong>of</strong> <strong>the</strong> charge <strong>of</strong> <strong>the</strong> first pulse <strong>for</strong> ATWD <strong>and</strong> FADC disagree significantly<br />
between Monte-Carlo data <strong>and</strong> <strong>the</strong> experimental data (figure 7.13): Most importantly<br />
<strong>the</strong> mean values <strong>for</strong> Monte-Carlo differ by more than 0.1 PE while those <strong>for</strong><br />
experimental data seem to be well aligned. This is likely a problem <strong>of</strong> <strong>the</strong> simulation:<br />
The ratio <strong>of</strong> <strong>the</strong> total (integrated) charge <strong>of</strong> <strong>the</strong> ATWD wave<strong>for</strong>m <strong>and</strong> <strong>the</strong> corresponding<br />
fraction <strong>of</strong> <strong>the</strong> FADC wave<strong>for</strong>m differs largely between Monte-Carlo <strong>and</strong> experimental<br />
data, see figure 7.12. A part <strong>of</strong> this disagreement can be attributed to a bug in <strong>the</strong><br />
daq_baseline simulation, see appendix C.3 <strong>and</strong> figure C.2.<br />
Considered individually, <strong>the</strong> first ATWD <strong>and</strong> FADC pulses’ charges match well <strong>for</strong> experimental<br />
data. The tails in <strong>the</strong> FADC charge distributions towards values higher than<br />
2 PE can be explained by <strong>the</strong> FADC’s inferior time resolution; <strong>the</strong> separation <strong>of</strong> features<br />
that can be distinguished as SPE-like in ATWD is <strong>of</strong>ten rendered impossible in FADC,<br />
so multiple-PE features are more likely.<br />
The distributions <strong>of</strong> <strong>the</strong> differences between <strong>the</strong> cumulative charge <strong>of</strong> all ATWD <strong>and</strong><br />
all FADC pulses per wave<strong>for</strong>m in figure 7.14 support <strong>the</strong> hypo<strong>the</strong>sis <strong>of</strong> a wrong simulation;<br />
while <strong>the</strong> differences peak at 0 PE <strong>for</strong> experimental data, Monte-Carlo contains on<br />
average about 0.14 PE more charge in ATWD than in FADC.<br />
The distributions show tails towards high FADC charges similar to <strong>the</strong> tails already observed<br />
in <strong>the</strong> distributions <strong>of</strong> <strong>the</strong> charges <strong>of</strong> first pulses (see above); however, in <strong>the</strong><br />
cumulative distributions <strong>the</strong>y do not originate from inseparable pulses because it is irrelevant<br />
if <strong>the</strong> charge is distributed between multiple pulses. Instead, <strong>the</strong>y are caused by late<br />
pulses that are present in <strong>the</strong> FADC wave<strong>for</strong>m, but not in <strong>the</strong> ATWD wave<strong>for</strong>m. Evidence<br />
<strong>for</strong> this is <strong>the</strong> shallow secondary peak whose distance to <strong>the</strong> main peak is approximately<br />
one mean FADC SPE charge <strong>for</strong> ei<strong>the</strong>r Monte-Carlo or experimental data respectively.<br />
The bottom plots in figure 7.15 show <strong>the</strong> distributions <strong>of</strong> <strong>the</strong> number <strong>of</strong> pulses per<br />
wave<strong>for</strong>m; <strong>the</strong>y can not be compared directly because <strong>the</strong> datasets differ too much: The<br />
74
7.2 Verification Using Experimental Data<br />
500000<br />
width: 0.36 0.36<br />
mean: 0.93 0.81<br />
400000<br />
300000<br />
entries<br />
200000<br />
100000<br />
0<br />
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0<br />
q pulse1<br />
600000<br />
width: 0.39 0.43<br />
mean: 0.90 0.94<br />
500000<br />
400000<br />
entries<br />
300000<br />
200000<br />
100000<br />
0<br />
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0<br />
q pulse1<br />
Figure 7.13: Distributions <strong>of</strong> <strong>the</strong> charges <strong>of</strong> <strong>the</strong> first pulses <strong>of</strong> ATWD (blue areas) <strong>and</strong> FADC<br />
(green lines) NFE pulses <strong>for</strong> Monte-Carlo (top) <strong>and</strong> experimental data (bottom);<br />
red lines indicate Gaussian fits.<br />
75
7 PERFORMANCE OPTIMIZATION<br />
90000<br />
80000<br />
width: 0.11<br />
mean: 0.14<br />
70000<br />
60000<br />
entries<br />
50000<br />
40000<br />
30000<br />
20000<br />
10000<br />
0<br />
1.5 1.0 0.5 0.0 0.5<br />
∆q pulses<br />
160000<br />
140000<br />
width: 0.11<br />
mean: -0.02<br />
120000<br />
100000<br />
entries<br />
80000<br />
60000<br />
40000<br />
20000<br />
0<br />
1.5 1.0 0.5 0.0 0.5<br />
∆q pulses<br />
Figure 7.14: Distributions <strong>of</strong> <strong>the</strong> differences between <strong>the</strong> total charges <strong>of</strong> all ATWD <strong>and</strong> FADC<br />
NFE pulses per wave<strong>for</strong>m <strong>for</strong> Monte-Carlo (top) <strong>and</strong> experimental data (bottom).<br />
Positive values indicate higher charges in ATWD; red lines indicate Gaussian fits.<br />
76
7.2 Verification Using Experimental Data<br />
10 7<br />
10 6<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
0 50 100 150 200<br />
n pulses<br />
10 7<br />
10 6<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
0 20 40 60 80 100 120 140 160<br />
n pulses<br />
Figure 7.15: Distributions <strong>of</strong> <strong>the</strong> numbers <strong>of</strong> ATWD (blue areas) <strong>and</strong> FADC (green lines) NFE<br />
pulses <strong>for</strong> Monte-Carlo (top) <strong>and</strong> experimental data (bottom).<br />
77
7 PERFORMANCE OPTIMIZATION<br />
particle energy in <strong>the</strong> simulated dataset follows an E −1 neutrino energy spectrum (although<br />
<strong>the</strong> dataset is untriggered, favoring low-energy events with few pulses), while <strong>the</strong><br />
experimental dataset follows <strong>the</strong> atmospheric muon energy spectrum <strong>of</strong> E −3.7 .<br />
Still <strong>the</strong> plots are plausible: The triple peak structure in <strong>the</strong> Monte-Carlo ATWD <strong>and</strong><br />
<strong>the</strong> small break around bin 128 in FADC originate from fully illuminated wave<strong>for</strong>ms, i.e.,<br />
wave<strong>for</strong>ms with “BayesUnfold”’s maximum number <strong>of</strong> pulses; 64 <strong>for</strong> ATWD <strong>and</strong> 128 <strong>for</strong><br />
FADC 6 . Higher entries <strong>and</strong> <strong>the</strong> third peak respectively originate from DOMs which were<br />
launched twice because <strong>of</strong> ongoing events (see section 4.4.2).<br />
Despite <strong>the</strong> problems arising due to <strong>the</strong> insufficient simulation, <strong>the</strong> comparisons <strong>of</strong><br />
ATWD data with FADC data are promising; <strong>the</strong> important features are qualitatively<br />
understood, <strong>and</strong> <strong>the</strong> agreements are expected to improve with <strong>the</strong> next release <strong>of</strong> <strong>the</strong><br />
simulation s<strong>of</strong>tware.<br />
6 Technically, “BayesUnfold” can extract 128 pulses from ATWD <strong>and</strong> 256 from FADC. However, an<br />
appropriate wave<strong>for</strong>m is extremly unlikely.<br />
78
CHAPTER VIII<br />
Per<strong>for</strong>mance <strong>Test</strong>s
8 PERFORMANCE TESTS<br />
8.1 Extraction <strong>of</strong> Simple Pulses with “Simple” <strong>and</strong> BU<br />
The use <strong>of</strong> <strong>the</strong> algorithm “Simple” instead <strong>of</strong> “BayesUnfold” (BU) <strong>for</strong> pulses that have been<br />
classified as simple by <strong>the</strong> pre-evaluation algorithm (see section 6.1) not only improves<br />
NFE’s CPU efficiency (see section 8.5), but also improves <strong>the</strong> extraction results because<br />
“Simple” only needs to extract those wave<strong>for</strong>ms <strong>for</strong> which it was designed. “Simple”’s<br />
lower charge threshold allows <strong>the</strong> algorithm to find more small true pulses while extracting<br />
slightly less false pulses than BU in <strong>the</strong> same simple sample.<br />
The leftmost bin <strong>of</strong> figure 8.2 shows that “Simple” does not have ATWD wave<strong>for</strong>ms<br />
without pulses compared to about 100000 <strong>for</strong> BU; <strong>the</strong> latter corresponds to 1.1% <strong>of</strong> all<br />
tested wave<strong>for</strong>ms. In FADC, <strong>the</strong>re are 221 compared to about 170000, which corresponds<br />
to 1.3% <strong>of</strong> all tested wave<strong>for</strong>ms. Fur<strong>the</strong>rmore, <strong>the</strong> lower plateau to <strong>the</strong> left side in <strong>the</strong><br />
distribution <strong>of</strong> <strong>the</strong> time residuals <strong>of</strong> <strong>the</strong> first pulses in figure 8.1 indicates that <strong>the</strong> number<br />
<strong>of</strong> first true features missed in favor <strong>of</strong> a later pulse is lower by about a factor <strong>of</strong> 5 in<br />
ATWD; it is at <strong>the</strong> same level <strong>for</strong> “Simple” <strong>and</strong> BU in FADC. The time resolution <strong>of</strong><br />
“Simple” measured by a Gaussian fit to <strong>the</strong> peak (figure 8.1) is worse than that <strong>of</strong> BU by<br />
a factor <strong>of</strong> 4% resp. 8%; this trade-<strong>of</strong>f is deemed to be acceptable because <strong>the</strong> decline is<br />
small compared to a wave<strong>for</strong>m bin length.<br />
The average number <strong>of</strong> pulses extracted by “Simple” is slightly larger compared to BU<br />
(figure 8.2). The reason is that while “Simple” is not able to split features into multiple<br />
pulses, its charge threshold is lower (0.15 PE vs. 0.20 PE). The inability to split can be<br />
considered advantageous in this case because simple pulses are SPE-like enough to not<br />
require splitting, <strong>and</strong> excessive splitting is not desired.<br />
The large number <strong>of</strong> wave<strong>for</strong>ms with more than ten pulses from “Simple” is caused by <strong>the</strong><br />
missing simulation <strong>of</strong> <strong>the</strong> daq_baseline (see appendix C.3). An example <strong>for</strong> a wave<strong>for</strong>m<br />
with extraordinarily high baseline caused by this bug can be seen in figure 8.3. Figure 8.4<br />
shows <strong>the</strong> same distribution as figure 8.2, but with “Simple”’s detection threshold w detect<br />
set to “Eva”’s feature threshold w feat = 0.08 PE. Due to this change <strong>the</strong> excessive pulses<br />
are cut away at <strong>the</strong> cost <strong>of</strong> losing many true pulses from wave<strong>for</strong>ms with a more normal<br />
baseline: The number <strong>of</strong> wave<strong>for</strong>ms without pulses increases from 0 to almost 200000<br />
(leftmost bin).<br />
The charge <strong>of</strong> <strong>the</strong> first pulse extracted by “Simple” is about 2% higher <strong>for</strong> ATWD <strong>and</strong><br />
about 2% lower <strong>for</strong> FADC on average than <strong>the</strong> charge <strong>of</strong> <strong>the</strong> first BU pulse, as can be<br />
seen in figure 8.5; this is considered to be good agreement.<br />
Finally, <strong>the</strong> difference <strong>of</strong> <strong>the</strong> total pulse charges extracted per wave<strong>for</strong>m by both algorithms<br />
peaks near to zero <strong>and</strong> features a small width <strong>of</strong> 0.1 PE if <strong>the</strong> pulses caused by<br />
<strong>the</strong> incomplete simulation are cut away, compare figure 8.6 to figure 8.7.<br />
These results will change slightly <strong>for</strong> <strong>new</strong> Monte-Carlo datasets which incorporate<br />
<strong>the</strong> simulation <strong>of</strong> <strong>the</strong> daq_baseline; however significant deviations from <strong>the</strong> predicted<br />
80
8.1 Extraction <strong>of</strong> Simple Pulses with “Simple” <strong>and</strong> “BayesUnfold”<br />
10 7<br />
10 6<br />
width: 0.73 0.76<br />
mean: 10.99 11.09<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
40 20 0 20 40<br />
t MC −t pulse /ns<br />
10 7<br />
10 6<br />
width: 2.85 3.08<br />
mean: 11.16 10.82<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
80 60 40 20 0 20 40 60<br />
t MC −t pulse /ns<br />
Figure 8.1: Distribution <strong>of</strong> <strong>the</strong> time residuals <strong>of</strong> <strong>the</strong> first pulses extracted from simple Monte-<br />
Carlo wave<strong>for</strong>ms by “BayesUnfold” (blue areas) <strong>and</strong> “Simple” (green lines): Shown<br />
is <strong>the</strong> difference between <strong>the</strong> Monte-Carlo hit time <strong>and</strong> <strong>the</strong> time <strong>of</strong> <strong>the</strong> first pulse;<br />
red lines indicate Gaussian fits; top: ATWD, bottom: FADC.<br />
81
8 PERFORMANCE TESTS<br />
10 7<br />
10 6<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
0 5 10 15 20 25<br />
n pulses<br />
10 7<br />
10 6<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
0 5 10 15 20 25 30 35 40<br />
n pulses<br />
Figure 8.2: Distribution <strong>of</strong> <strong>the</strong> numbers <strong>of</strong> pulses extracted from simple Monte-Carlo wave<strong>for</strong>ms<br />
by “BayesUnfold” (blue areas) <strong>and</strong> “Simple” (green lines); top: ATWD,<br />
bottom: FADC.<br />
82
8.1 Extraction <strong>of</strong> Simple Pulses with “Simple” <strong>and</strong> “BayesUnfold”<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
1.30<br />
NFE_ATWDPulses_bu<br />
10400 10600 10800 11000 11200<br />
NFE_ATWDPulses_simple<br />
time / ns<br />
1.33<br />
0.18<br />
0.65<br />
0.16<br />
0.35<br />
0.27 0.16<br />
0.39<br />
0.41 0.48 0.18 0.21<br />
0.23<br />
0.37 0.18<br />
10400 10600 10800 11000 11200<br />
0.35<br />
OMKey(9,55)<br />
Figure 8.3: Example wave<strong>for</strong>m <strong>for</strong> an erroneously high baseline caused by incomplete simulation.<br />
Dotted lines indicate ATWD pulses extracted by “BayesUnfold” (top) <strong>and</strong><br />
“Simple” (bottom), with charge given in units <strong>of</strong> PE.<br />
83
8 PERFORMANCE TESTS<br />
10 7<br />
10 6<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
0 2 4 6 8 10 12 14<br />
n pulses<br />
Figure 8.4: Effect <strong>of</strong> erroneous baselines on <strong>the</strong> distributions <strong>of</strong> <strong>the</strong> number <strong>of</strong> pulses extracted<br />
from simple Monte-Carlo wave<strong>for</strong>ms by “BayesUnfold” (blue areas) <strong>and</strong> “Simple”<br />
(green lines). The excessive pulses caused by high baselines were cut away by a<br />
higher feature threshold in “Simple”; compare to figure 8.2.<br />
84
8.1 Extraction <strong>of</strong> Simple Pulses with “Simple” <strong>and</strong> “BayesUnfold”<br />
1000000<br />
width: 0.32 0.35<br />
mean: 0.90 0.92<br />
800000<br />
600000<br />
entries<br />
400000<br />
200000<br />
0<br />
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0<br />
q pulse1<br />
900000<br />
800000<br />
width: 0.32 0.34<br />
mean: 0.81 0.79<br />
700000<br />
600000<br />
entries<br />
500000<br />
400000<br />
300000<br />
200000<br />
100000<br />
0<br />
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0<br />
q pulse1<br />
Figure 8.5: Distribution <strong>of</strong> <strong>the</strong> charges <strong>of</strong> <strong>the</strong> first pulses extracted from simple Monte-Carlo<br />
wave<strong>for</strong>ms by “BayesUnfold” (blue areas) <strong>and</strong> “Simple” (green lines); red lines<br />
indicate Gaussian fits; top: ATWD, bottom: FADC.<br />
85
8 PERFORMANCE TESTS<br />
10 7<br />
10 6<br />
width: 0.07<br />
mean: -0.08<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
10 5 0 5 10<br />
∆q pulses<br />
10 7<br />
10 6<br />
width: 0.09<br />
mean: -0.03<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
10 5 0 5 10<br />
∆q pulses<br />
Figure 8.6: Distribution <strong>of</strong> <strong>the</strong> differences between <strong>the</strong> total charges <strong>of</strong> all pulses extracted<br />
from simple Monte-Carlo wave<strong>for</strong>ms by “BayesUnfold” <strong>and</strong> “Simple”: Positive<br />
values correspond to higher values in BU; red lines indicate Gaussian fits; top:<br />
ATWD, bottom: FADC.<br />
86
8.1 Extraction <strong>of</strong> Simple Pulses with “Simple” <strong>and</strong> “BayesUnfold”<br />
10 7<br />
10 6<br />
width: 0.10<br />
mean: -0.02<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
10 5 0 5 10<br />
∆q pulses<br />
Figure 8.7: Effect <strong>of</strong> erroneous baselines on <strong>the</strong> distribution <strong>of</strong> <strong>the</strong> difference between <strong>the</strong> total<br />
charges <strong>of</strong> all pulses extracted from simple Monte-Carlo wave<strong>for</strong>ms by “BayesUnfold”<br />
<strong>and</strong> “Simple”; positive values indicate higher total charges <strong>for</strong> BU; <strong>the</strong> red<br />
line indicates a Gaussian fit. The excessive pulses caused by high baselines were<br />
cut away by a higher feature threshold in “Simple”; compare to figure 8.6.<br />
87
8 PERFORMANCE TESTS<br />
10 7<br />
10 6<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
0 2 4 6 8 10 12 14 16 18<br />
n pulses<br />
Figure 8.8: Distribution <strong>of</strong> <strong>the</strong> numbers <strong>of</strong> pulses extracted from simple experimental ATWD<br />
wave<strong>for</strong>ms by “BayesUnfold” (blue areas) <strong>and</strong> “Simple” (green lines). Experimental<br />
data is unaffected by <strong>the</strong> daq_baseline bug, compare to figure 8.2.<br />
88
8.1 Extraction <strong>of</strong> Simple Pulses with “Simple” <strong>and</strong> “BayesUnfold”<br />
10 7<br />
10 6<br />
width: 0.08<br />
mean: -0.08<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
10 5 0 5 10<br />
∆q pulses<br />
Figure 8.9: Distribution <strong>of</strong> <strong>the</strong> differences between <strong>the</strong> total charges <strong>of</strong> all pulses extracted<br />
from simple experimental ATWD wave<strong>for</strong>ms by “BayesUnfold” <strong>and</strong> “Simple”:<br />
Positive values correspond to higher values in “BayesUnfold”; <strong>the</strong> red line indicates<br />
a Gaussian fit. Experimental data is unaffected by <strong>the</strong> daq_baseline bug, compare<br />
to figure 8.6.<br />
89
8 PERFORMANCE TESTS<br />
distributions are not expected because <strong>the</strong> distributions <strong>of</strong> total charge <strong>and</strong> number <strong>of</strong><br />
pulses <strong>for</strong> experimental data (in which this bug does not occur) show virtually no excessive<br />
pulses; compare figure 8.2 to figure 8.8 <strong>and</strong> figure 8.6 to figure 8.9.<br />
If some <strong>of</strong> <strong>the</strong> excessive pulses will remain in <strong>the</strong> <strong>new</strong> data regardless, ei<strong>the</strong>r “Simple”’s<br />
detection threshold or “Eva”’s feature threshold will be adjusted; lowering <strong>the</strong> latter helps<br />
because <strong>the</strong>n <strong>the</strong> affected wave<strong>for</strong>ms will be marked as complex, <strong>and</strong> BU is more robust<br />
towards an increased baseline as can be seen in figure 8.3.<br />
8.2 Extraction <strong>of</strong> Exotic <strong>Feature</strong>s<br />
The extraction <strong>of</strong> exotic features can be examined best by manually checking individual<br />
wave<strong>for</strong>ms; o<strong>the</strong>rwise problems might be concealed by <strong>the</strong> higher statistics <strong>of</strong> normal<br />
wave<strong>for</strong>ms. Also, certain tests are hard to automatize, e. g., timing distributions <strong>of</strong> nonfirst<br />
pulses, because <strong>the</strong> assignement <strong>of</strong> pulses to Monte-Carlo hits is not trivial.<br />
Figure 8.10 shows an example <strong>for</strong> <strong>the</strong> per<strong>for</strong>mance <strong>of</strong> NFE <strong>for</strong> very small features: Normally<br />
NFE does not extract any pulses because only <strong>the</strong> middle bin <strong>of</strong> <strong>the</strong> ATWD feature<br />
exceeds “Simple”’s threshold w feat , <strong>the</strong>re<strong>for</strong>e <strong>the</strong> extracted charge is about 0.12 PE < q min .<br />
With En<strong>for</strong>cePulse set, <strong>the</strong> feature is <strong>the</strong>n relayed to “BayesUnfold” which first tries to<br />
extract pulses normally. If this fails, BU defines a pulse centered at <strong>the</strong> deconvoluted<br />
distribution’s maximum bin <strong>and</strong> accepts it regardless <strong>of</strong> its charge.<br />
Most <strong>of</strong> <strong>the</strong> time, this works well: The only difference between <strong>the</strong> blue distribution in<br />
figure 7.4 <strong>and</strong> green distribution in figure 8.16 is that <strong>for</strong> <strong>the</strong> latter En<strong>for</strong>cePulse is set.<br />
The number <strong>of</strong> wave<strong>for</strong>ms <strong>for</strong> which <strong>the</strong> first extracted pulse is more than 50 ns away<br />
from <strong>the</strong> Monte-Carlo hit is reduced from about 11000 to about 5000 (underflow bin).<br />
The drawback is an increased number <strong>of</strong> false early pulses; an example <strong>for</strong> this can be<br />
seen in figure 8.10: The ATWD feature is extracted successfully, but <strong>the</strong> FADC pulse (<strong>for</strong><br />
which En<strong>for</strong>cePulse was set independently) might be caused by noise.<br />
Final quantitative tests can only be conducted when simulations with proper baselines<br />
become available (see appendix C.3).<br />
Besides many r<strong>and</strong>omly chosen wave<strong>for</strong>ms like <strong>the</strong> ones shown in figure 7.1, a small<br />
catalogue <strong>of</strong> exotic features provided by Markus Voge[60] was examined, along with o<strong>the</strong>r<br />
wave<strong>for</strong>ms from <strong>the</strong> same dataset; see figure 8.11 to figure 8.15.<br />
For comparison, <strong>Feature</strong><strong>Extractor</strong>’s pulses are also shown.<br />
Figure 8.11, top (DOM 50-24):<br />
Noise artifacts such as this one appear in experimental data relatively <strong>of</strong>ten ( 1%); both<br />
feature extractors are robust concerning this effect; it might however interfere with FE’s<br />
charge calculation as single pulse’s charge seems to be overestimated – extraction by eye<br />
yields about 0.38 PE, NFE finds 0.35 PE, FE 0.60 PE.<br />
90
8.2 Extraction <strong>of</strong> Exotic <strong>Feature</strong>s<br />
NFE_MergedPulses<br />
OMKey(47,48)<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
0.15<br />
0.10<br />
0.05<br />
0.00<br />
0.05<br />
0.15<br />
0.10<br />
0.05<br />
0.00<br />
0.05<br />
19100 19200 19300 19400 19500 19600 19700 19800 19900<br />
FE_Pulses<br />
time / ns<br />
0.23<br />
19100 19200 19300 19400 19500 19600 19700 19800 19900<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
0.15<br />
0.10<br />
0.05<br />
0.00<br />
0.05<br />
0.15<br />
0.10<br />
0.05<br />
0.00<br />
0.05<br />
0.15<br />
0.15<br />
NFE_MergedPulses<br />
OMKey(47,48)<br />
19100 19200 19300 19400 19500 19600 19700 19800 19900<br />
FE_Pulses<br />
time / ns<br />
0.23<br />
19100 19200 19300 19400 19500 19600 19700 19800 19900<br />
Figure 8.10: Example wave<strong>for</strong>ms to demonstrate NFE’s option En<strong>for</strong>cePulse, shown is a<br />
wave<strong>for</strong>m with pulses (dotted lines with charges in PE) extracted by NFE (upper<br />
image each) <strong>and</strong> FE (lower image each):<br />
Top: default NFE (En<strong>for</strong>cePulse = False);<br />
bottom: same settings bar En<strong>for</strong>cePulse = True <strong>for</strong> both ATWD <strong>and</strong> FADC.<br />
91
8 PERFORMANCE TESTS<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
0.30<br />
0.25<br />
0.20<br />
0.15<br />
0.10<br />
0.05<br />
0.00<br />
0.05<br />
0.30<br />
0.25<br />
0.20<br />
0.15<br />
0.10<br />
0.05<br />
0.00<br />
0.05<br />
0.35<br />
NFE_MergedPulses<br />
OMKey(50,24)<br />
9900 10000 10100 10200 10300 10400 10500 10600 10700<br />
FE_Pulses<br />
time / ns<br />
0.60<br />
9900 10000 10100 10200 10300 10400 10500 10600 10700<br />
NFE_MergedPulses<br />
OMKey(18,4)<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0.0<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0.38<br />
1.10<br />
0.76<br />
1.45<br />
15500 15600 15700 15800 15900 16000 16100 16200 16300<br />
FE_Pulses<br />
time / ns<br />
0.32<br />
0.82<br />
0.40<br />
0.69<br />
0.83<br />
0.63<br />
0.31<br />
1.51<br />
1.44<br />
0.0<br />
15500 15600 15700 15800 15900 16000 16100 16200 16300<br />
Figure 8.11: Example wave<strong>for</strong>ms <strong>of</strong> exotic or difficult features from IC59 experimental run<br />
113912, partially from Markus Voge’s catalog <strong>of</strong> exotic wave<strong>for</strong>ms. All wave<strong>for</strong>ms<br />
are shown with default NFE merged pulses (ATWD+FADC; top) <strong>and</strong> FE pulses<br />
(IC59 multi-pulse online-filtering settings; bottom), indicated by dotted lines with<br />
<strong>the</strong> charge given in units <strong>of</strong> PE.<br />
92
8.2 Extraction <strong>of</strong> Exotic <strong>Feature</strong>s<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
1.4<br />
1.2<br />
1.0<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0.0<br />
1.4<br />
1.2<br />
1.0<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0.0<br />
0.60<br />
0.81<br />
0.51<br />
0.70<br />
0.52<br />
1.02<br />
NFE_MergedPulses<br />
0.94<br />
0.65<br />
0.59<br />
9900 10000 10100 10200 10300 10400 10500 10600 10700<br />
FE_Pulses<br />
time / ns<br />
0.54<br />
0.45 0.42<br />
0.40<br />
0.710.34<br />
0.34<br />
0.47<br />
0.22<br />
0.81<br />
0.24<br />
0.67<br />
0.41<br />
0.20<br />
9900 10000 10100 10200 10300 10400 10500 10600 10700<br />
1.18<br />
OMKey(54,1)<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
0.79<br />
0.92<br />
0.44<br />
1.43<br />
1.41<br />
0.45<br />
NFE_MergedPulses<br />
OMKey(54,2)<br />
10000 10100 10200 10300 10400 10500 10600 10700<br />
FE_Pulses<br />
time / ns<br />
0.36<br />
1.62<br />
0.48<br />
0.72<br />
1.02<br />
10000 10100 10200 10300 10400 10500 10600 10700<br />
Figure 8.12: Example wave<strong>for</strong>ms <strong>of</strong> exotic or difficult features from IC59 experimental run<br />
113912, partially from Markus Voge’s catalog <strong>of</strong> exotic wave<strong>for</strong>ms. All wave<strong>for</strong>ms<br />
are shown with default NFE merged pulses (ATWD+FADC; top) <strong>and</strong> FE pulses<br />
(IC59 multi-pulse online-filtering settings; bottom), indicated by dotted lines with<br />
<strong>the</strong> charge given in units <strong>of</strong> PE.<br />
93
8 PERFORMANCE TESTS<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
1.00<br />
0.32<br />
0.28<br />
NFE_MergedPulses<br />
1.15<br />
18300 18400 18500 18600 18700 18800 18900 19000 19100<br />
FE_Pulses<br />
time / ns<br />
0.7 0.24<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
1.03<br />
0.35<br />
0.91<br />
0.46<br />
0.23<br />
OMKey(47,51)<br />
18300 18400 18500 18600 18700 18800 18900 19000 19100<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
0.26<br />
1.32<br />
NFE_MergedPulses<br />
OMKey(65,13)<br />
11300 11400 11500 11600 11700 11800 11900 12000<br />
FE_Pulses<br />
time / ns<br />
1.08<br />
11300 11400 11500 11600 11700 11800 11900 12000<br />
Figure 8.13: Example wave<strong>for</strong>ms <strong>of</strong> exotic or difficult features from IC59 experimental run<br />
113912, partially from Markus Voge’s catalog <strong>of</strong> exotic wave<strong>for</strong>ms. All wave<strong>for</strong>ms<br />
are shown with default NFE merged pulses (ATWD+FADC; top) <strong>and</strong> FE pulses<br />
(IC59 multi-pulse online-filtering settings; bottom), indicated by dotted lines with<br />
<strong>the</strong> charge given in units <strong>of</strong> PE.<br />
94
8.2 Extraction <strong>of</strong> Exotic <strong>Feature</strong>s<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
0.25<br />
0.58<br />
0.44<br />
1.35<br />
NFE_MergedPulses<br />
17500 17600 17700 17800 17900 18000 18100 18200 18300<br />
FE_Pulses<br />
time / ns<br />
0.75<br />
1.62<br />
0.36<br />
OMKey(83,17)<br />
17500 17600 17700 17800 17900 18000 18100 18200 18300<br />
NFE_MergedPulses<br />
OMKey(38,48)<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
0.4<br />
0.19<br />
0.65<br />
0.45<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
18000 18100 18200 18300 18400 18500 18600 18700 18800<br />
FE_Pulses<br />
time / ns<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
0.39<br />
0.34<br />
18000 18100 18200 18300 18400 18500 18600 18700 18800<br />
Figure 8.14: Example wave<strong>for</strong>ms <strong>of</strong> exotic or difficult features from IC59 experimental run<br />
113912, partially from Markus Voge’s catalog <strong>of</strong> exotic wave<strong>for</strong>ms. All wave<strong>for</strong>ms<br />
are shown with default NFE merged pulses (ATWD+FADC; top) <strong>and</strong> FE pulses<br />
(IC59 multi-pulse online-filtering settings; bottom), indicated by dotted lines with<br />
<strong>the</strong> charge given in units <strong>of</strong> PE.<br />
95
8 PERFORMANCE TESTS<br />
Figure 8.11, bottom (DOM 18-04):<br />
This wave<strong>for</strong>m was included in Markus Voge’s initial catalog because <strong>the</strong> first feature<br />
was not extracted properly with <strong>the</strong> <strong>of</strong>fline-reconstruction settings <strong>of</strong> <strong>Feature</strong><strong>Extractor</strong><br />
(see section 8.3.1); in its online-filtering settings that are shown here, FE extracts this<br />
feature without problems, but tends to split SPE-like features excessively. NFE’s results<br />
are reasonable.<br />
Figure 8.12, top (DOM 54-01):<br />
In this example, both extractors miss <strong>the</strong> tiny feature at 10 100 ns. Extraction by eye<br />
yields about 0.1 PE.<br />
Figure 8.12, bottom (DOM 54-02):<br />
For this ATWD wave<strong>for</strong>m, both extractors agree well. The ExclusionTime introduced<br />
in <strong>Feature</strong><strong>Extractor</strong> to prevent double extraction (see appendix C.1) circumvents <strong>the</strong><br />
extraction <strong>of</strong> <strong>the</strong> FADC feature present in NFE.<br />
Figure 8.13, top (DOM 47-51):<br />
Here both extractors recognize <strong>the</strong> first feature as multi-PE hit with a small charge in<br />
<strong>the</strong> first peak (perhabs a prepulse). The middle feature is extracted by NFE exclusively<br />
despite <strong>the</strong> fact that it technically lies below FE’s charge threshold: FE splits <strong>the</strong> last<br />
feature into three parts <strong>of</strong> which one only contains 0.23 PE – <strong>the</strong> missed feature contains<br />
about 0.28 PE.<br />
Figure 8.13, bottom (DOM 65-13):<br />
This is an example <strong>for</strong> a very small first feature. It is extracted by NFE, but not by<br />
FE. Also, NFE extracts more charge from <strong>the</strong> late ATWD feature (1.32 PE instead <strong>of</strong><br />
0.26 PE); extraction by eye yields well about 1.2 PE.<br />
Figure 8.14, top (DOM 83-17):<br />
Ano<strong>the</strong>r example <strong>for</strong> a small early feature. Again it is missed in <strong>Feature</strong><strong>Extractor</strong>, but<br />
here <strong>the</strong> missing charge is distributed among <strong>the</strong> extracted pulses.<br />
Figure 8.14, bottom (DOM 38-48):<br />
The first ATWD feature was too small to be recognized by one <strong>of</strong> <strong>the</strong> extractors; however<br />
in contrast to FE, NFE’s PulseMerger is capable <strong>of</strong> appending FADC pulses at times<br />
where ATWD wave<strong>for</strong>ms are available if <strong>the</strong>y do not clash with ATWD pulses.<br />
Figure 8.15 shows a typical bright wave<strong>for</strong>m, i. e. a wave<strong>for</strong>m with high integrated<br />
charge. The two feature extractors give roughly comparable results besides <strong>Feature</strong><strong>Extractor</strong>’s<br />
more pronounced pulse splitting. Noteworthy is <strong>the</strong> gap between <strong>the</strong> end <strong>of</strong> <strong>the</strong><br />
ATWD wave<strong>for</strong>m <strong>and</strong> <strong>the</strong> first FADC pulse <strong>for</strong> FE; <strong>the</strong> amount <strong>of</strong> charge lost due to this<br />
is estimated by eye to about 7.6 PE <strong>of</strong> which NFE extracts 5.6 PE.<br />
One idea <strong>for</strong> a more quantitative analysis <strong>for</strong> bright wave<strong>for</strong>ms is to compare <strong>the</strong> original<br />
wave<strong>for</strong>m to a wave<strong>for</strong>m generated based on <strong>the</strong> extracted pulses, <strong>for</strong> example with a<br />
96
8.3 Comparison with O<strong>the</strong>r <strong>Feature</strong> <strong>Extractor</strong>s<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
5<br />
4<br />
3<br />
2<br />
1<br />
0<br />
5<br />
4<br />
3<br />
2<br />
1<br />
0.79 1.27 1.43<br />
1.07 1.12 0.74<br />
0.74<br />
2.05<br />
0.71<br />
1.82<br />
1.79<br />
3.12 1.42<br />
0.50<br />
1.82<br />
1.98<br />
0.90<br />
3.68<br />
0.62<br />
0.63<br />
0.45<br />
0.64 2.56 1.88 0.69<br />
4.84<br />
0.39<br />
0.48<br />
0.49<br />
1.20 1.66<br />
2.48 3.20 0.40<br />
0.95<br />
1.66<br />
2.05<br />
0.71<br />
1.12<br />
1.81<br />
NFE_MergedPulses<br />
4.31<br />
1.30<br />
15400 15600 15800 16000 16200<br />
FE_Pulses<br />
time / ns<br />
1.81 0.66 1.07<br />
1.15<br />
0.86 0.44 0.51<br />
0.68 1.20<br />
0.48<br />
0.56 0.75 0.77 1.13 1.71 0.81 0.95<br />
0.74 0.65<br />
1.27<br />
0.72 0.54<br />
0.79<br />
0.84<br />
0.73<br />
0.60<br />
0.43<br />
0.67<br />
0.56<br />
0.64<br />
0.77<br />
0.84<br />
0.81<br />
0.35 0.72<br />
1.63<br />
0.85<br />
0.79<br />
1.04<br />
0.70<br />
0.77 0.80<br />
1.27<br />
0.68<br />
0.57<br />
1.55<br />
0.49<br />
1.34 0.78<br />
0.48<br />
0.36 2.23 1.29 0.69<br />
2.31 0.84 0.74<br />
0.57<br />
1.07<br />
0.71<br />
0.63 0.62<br />
1.21<br />
0.60<br />
1.63 1.05<br />
0.74<br />
OMKey(26,10)<br />
2.25<br />
0.81<br />
1.11<br />
2.0<br />
0<br />
15400 15600 15800 16000 16200<br />
Figure 8.15: Example <strong>of</strong> a typical bright wave<strong>for</strong>m from Markus Voge’s IC59 exotic wave<strong>for</strong>m<br />
catalog. It is shown with default NFE merged pulses (ATWD+FADC; top) <strong>and</strong><br />
FE pulses (IC59 multi-pulse online-filtering settings; bottom), indicated by dotted<br />
lines with <strong>the</strong> charge given in units <strong>of</strong> PE.<br />
Kolmogorov-Smirnov test, however this has yet to be done systematically.<br />
Reliable extraction <strong>of</strong> features where <strong>Feature</strong><strong>Extractor</strong> has difficulties bodes well; however<br />
this is no guaranty <strong>for</strong> bug free <strong>and</strong> unproblematic operation. There<strong>for</strong>e tests such as<br />
<strong>the</strong>se must be continued in future to identify problems <strong>of</strong> NFE or to discover unexpected<br />
changes in <strong>the</strong> low-level reconstruction chain.<br />
8.3 Comparison with O<strong>the</strong>r <strong>Feature</strong> <strong>Extractor</strong>s<br />
The comparison <strong>of</strong> NFE’s results to that <strong>of</strong> o<strong>the</strong>r feature extractors is in<strong>for</strong>mative concerning<br />
deficiencies <strong>of</strong> ei<strong>the</strong>r <strong>of</strong> <strong>the</strong>m. Knowledge <strong>of</strong> <strong>the</strong>se deficiencies is important <strong>for</strong><br />
fur<strong>the</strong>r improvements <strong>and</strong> <strong>for</strong> <strong>the</strong> study <strong>of</strong> systematical errors in low-level reconstruction.<br />
8.3.1 <strong>Feature</strong><strong>Extractor</strong> in Multi-Pulse Mode<br />
<strong>Feature</strong><strong>Extractor</strong> is <strong>the</strong> extractor used <strong>for</strong> ATWD <strong>and</strong> FADC wave<strong>for</strong>ms in <strong>the</strong> IC59<br />
online-filtering <strong>and</strong> <strong>of</strong>fline-processing. The ability to test it with an independent project<br />
is one <strong>of</strong> <strong>the</strong> main benefits <strong>of</strong> having multiple feature extractors.<br />
The tests were conducted using both FE’s IC59 multi-pulse online <strong>and</strong> <strong>of</strong>fline settings[51].<br />
The <strong>for</strong>mer will be discussed in detail; afterwards, <strong>the</strong> differences <strong>of</strong> <strong>the</strong> <strong>of</strong>fline-processing<br />
97
8 PERFORMANCE TESTS<br />
10 7<br />
10 6<br />
width: 1.23 0.81<br />
mean: 0.53 11.19<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
40 20 0 20 40<br />
t MC −t pulse /ns<br />
Figure 8.16: Distributions <strong>of</strong> <strong>the</strong> time residuals <strong>of</strong> <strong>the</strong> first pulses extracted by FE (multi-pulse<br />
online-filtering settings, blue area) <strong>and</strong> NFE (with En<strong>for</strong>cePulse set, green line)<br />
from Monte-Carlo data; red lines indicate Gaussian fits.<br />
settings will be explained.<br />
For better comparability, NFE’s option En<strong>for</strong>cePulse was set because <strong>the</strong> analog is set<br />
<strong>for</strong> FE. Both extractors are configured to extract features in both ATWD <strong>and</strong> FADC.<br />
IC59 online-filtering settings<br />
The distribution <strong>of</strong> <strong>the</strong> time residuals in figure 8.16 shows that NFE has a better time<br />
resolution (Gaussian width <strong>of</strong> σ NFE = 0.81 ns instead <strong>of</strong> σ FE = 1.23 ns). NFE extracts<br />
19 fake early pulses instead <strong>of</strong> a single one <strong>for</strong> <strong>Feature</strong><strong>Extractor</strong> from about 4.1 million<br />
wave<strong>for</strong>ms in total. This is balanced by <strong>the</strong> fact that <strong>for</strong> NFE only about 5500 first pulses<br />
do not match <strong>the</strong> first Monte-Carlo hit, while this happens in about 90000 wave<strong>for</strong>ms <strong>for</strong><br />
<strong>Feature</strong><strong>Extractor</strong>; this can be seen in <strong>the</strong> histogram entries in <strong>the</strong> lower plateau <strong>and</strong> <strong>the</strong><br />
underflow bin to <strong>the</strong> left <strong>of</strong> figure 8.16.<br />
The time <strong>of</strong>fset <strong>of</strong> about 11 ns between NFE’s pulses <strong>and</strong> Monte-Carlo hit times <strong>and</strong><br />
<strong>the</strong> absence <strong>of</strong> such an <strong>of</strong>fset in <strong>Feature</strong><strong>Extractor</strong>’s pulse times is explained by <strong>the</strong> fact<br />
that <strong>Feature</strong>Extrator is used with its option PMTTransit set to 2 to have it subtract <strong>the</strong><br />
PMT transit time. By design NFE relies on DOMcalibrator’s time calibration instead,<br />
98
8.3 Comparison with O<strong>the</strong>r <strong>Feature</strong> <strong>Extractor</strong>s<br />
see section 7.1. This time <strong>of</strong>fset can clearly be seen in <strong>the</strong> figures in section 8.2.<br />
The distribution <strong>of</strong> <strong>the</strong> differences between <strong>the</strong> times <strong>of</strong> <strong>the</strong> first pulses <strong>for</strong> Monte-Carlo<br />
data (figure 8.17) has a Gaussian width <strong>of</strong> σ ∆ = 0.91 ns. Assuming <strong>the</strong> individual times<br />
to follow Gaussian distributions, this yields a correlation coefficient<br />
ρ = σ2 FE + σ 2 NFE − σ 4 ∆<br />
2σ FE σ NFE<br />
= 0.74,<br />
which is an encouraging result <strong>for</strong> both extractors because <strong>the</strong> distributions’ widths <strong>the</strong>mselves<br />
are significantly smaller than <strong>the</strong> ATWD wave<strong>for</strong>m bin length (T bin = 3.3 ns).<br />
In contrast to NFE, <strong>Feature</strong><strong>Extractor</strong> is not sensitive to <strong>the</strong> daq_baseline bug explained<br />
in section 7.2 <strong>and</strong> appendix C.3 because <strong>of</strong> its own implementation <strong>of</strong> baseline<br />
correction. There<strong>for</strong>e, <strong>the</strong> mean <strong>of</strong> <strong>the</strong> charges <strong>of</strong> <strong>the</strong> first pulse extracted by <strong>Feature</strong><strong>Extractor</strong><br />
is <strong>the</strong> same <strong>for</strong> Monte-Carlo <strong>and</strong> experimental data, although it is fur<strong>the</strong>r away<br />
from 1 PE compared to NFE, see figure 8.18.<br />
Moreover, FE’s distributions <strong>of</strong> <strong>the</strong> first charges are closer to a Gaussian distribution <strong>for</strong><br />
very low pulse charges. However, in a r<strong>and</strong>om sample <strong>of</strong> 37 wave<strong>for</strong>ms <strong>for</strong> which NFE extracted<br />
less than 0.2 PE <strong>for</strong> <strong>the</strong> first pulse, 11 <strong>of</strong> <strong>the</strong>se pulses were caused by features that<br />
were missed by FE in favor <strong>of</strong> a later feature. The remaining 26 pulses were extracted<br />
from <strong>the</strong> only distinct feature <strong>of</strong> each wave<strong>for</strong>m, triggering <strong>the</strong> extractors’ methods to<br />
extract at least one pulse. For almost all <strong>of</strong> <strong>the</strong>se features FE’s charges were between<br />
50% <strong>and</strong> 100% higher than those <strong>of</strong> NFE, while <strong>the</strong> latter approximately agreed with <strong>the</strong><br />
charges extracted by eye; see figure 8.19 <strong>for</strong> an extreme example. The reason <strong>for</strong> this<br />
effect might be that FE <strong>of</strong>ten raises <strong>the</strong> baseline <strong>and</strong> <strong>the</strong>n scales up <strong>the</strong> pulse charge to<br />
match <strong>the</strong> wave<strong>for</strong>m’s integraged charge as described in section 6.3.1. Thus, despite <strong>the</strong><br />
strange shape <strong>of</strong> <strong>the</strong> distribution in figure 8.18, <strong>the</strong> charges <strong>of</strong> <strong>the</strong> first pulses extracted<br />
by NFE are plausible.<br />
The distributions <strong>of</strong> <strong>the</strong> differences <strong>of</strong> <strong>the</strong> cumulative charge <strong>of</strong> all pulses per wave<strong>for</strong>m<br />
between FE <strong>and</strong> NFE in figure 8.20 show a higher correlation between <strong>the</strong> extractors (i.<br />
e., smaller width) <strong>for</strong> simulated data than <strong>for</strong> experimental data. The estimation <strong>of</strong> <strong>the</strong><br />
correlation coefficient using <strong>the</strong> fitted Gaussian distributions fails as it yields ρ = 1.02<br />
<strong>and</strong> ρ = 1.01, respectively.<br />
The distributions peak near 0 PE, but show tails with extreme values in both directions.<br />
These deviations are partly due to problems with <strong>the</strong> calibration <strong>of</strong> saturated ATWD wave<strong>for</strong>ms;<br />
<strong>the</strong> IC59 online-filtering settings <strong>of</strong> DOMcalibrator have <strong>the</strong> SaturationLevel set<br />
to 1022; this means that DOMcalibrator switchs to higher ATWD channels only <strong>for</strong> bins<br />
that reach <strong>the</strong> digitizer’s maximum value. However, because <strong>of</strong> noise wave<strong>for</strong>ms <strong>of</strong>ten fluctuate<br />
to slightly lower values even if <strong>the</strong> channel is saturated. This leads to very uneven<br />
wave<strong>for</strong>ms <strong>of</strong> which one can be seen in figure 8.21. Setting <strong>the</strong> level lower (<strong>for</strong> example to<br />
a value <strong>of</strong> 900) solves <strong>the</strong> problem; this was already done <strong>for</strong> <strong>the</strong> IC59 <strong>of</strong>fline-processing<br />
<strong>and</strong> is planned <strong>for</strong> future uses as well.[61]<br />
Besides <strong>the</strong>se calibration problems, <strong>the</strong> tails in figure 8.20 towards higher charges from<br />
99
8 PERFORMANCE TESTS<br />
900000<br />
800000<br />
width: 0.91<br />
mean: -10.59<br />
700000<br />
600000<br />
entries<br />
500000<br />
400000<br />
300000<br />
200000<br />
100000<br />
0<br />
16 14 12 10 8 6 4<br />
∆t pulse /ns<br />
900000<br />
800000<br />
width: 1.15<br />
mean: -10.16<br />
700000<br />
600000<br />
entries<br />
500000<br />
400000<br />
300000<br />
200000<br />
100000<br />
0<br />
16 14 12 10 8 6 4<br />
∆t pulse /ns<br />
Figure 8.17: Distributions <strong>of</strong> <strong>the</strong> times <strong>of</strong> <strong>the</strong> first pulses extracted by FE (multi-pulse onlinefiltering<br />
settings) <strong>and</strong> NFE (with En<strong>for</strong>cePulse set). Positive values correspond<br />
to later times <strong>for</strong> FE; red lines indicate Gaussian fits.<br />
Top: Monte-Carlo data; bottom: experimental data.<br />
100
8.3 Comparison with O<strong>the</strong>r <strong>Feature</strong> <strong>Extractor</strong>s<br />
600000<br />
width: 0.29 0.36<br />
mean: 0.81 0.93<br />
500000<br />
400000<br />
entries<br />
300000<br />
200000<br />
100000<br />
0<br />
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0<br />
q pulse1<br />
700000<br />
600000<br />
width: 0.31 0.39<br />
mean: 0.81 0.89<br />
500000<br />
entries<br />
400000<br />
300000<br />
200000<br />
100000<br />
0<br />
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0<br />
q pulse1<br />
Figure 8.18: Distributions <strong>of</strong> <strong>the</strong> charges <strong>of</strong> <strong>the</strong> first pulses extracted by FE (multi-pulse onlinefiltering<br />
settings, blue area) <strong>and</strong> NFE (with En<strong>for</strong>cePulse set, green line); red<br />
lines indicate Gaussian fits.<br />
Top: Monte-Carlo data; bottom: experimental data.<br />
101
8 PERFORMANCE TESTS<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
0.30<br />
0.25<br />
0.20<br />
0.15<br />
0.10<br />
0.05<br />
0.00<br />
0.05<br />
0.30<br />
0.25<br />
0.20<br />
0.15<br />
0.10<br />
0.05<br />
0.00<br />
0.05<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
0.16<br />
NFE_MergedPulses<br />
OMKey(3,16)<br />
11800 12000 12200 12400 12600<br />
FE_Pulses<br />
time / ns<br />
0.42<br />
11800 12000 12200 12400 12600<br />
NFE_MergedPulses<br />
OMKey(47,40)<br />
0.16<br />
0.61<br />
9800 10000 10200 10400 10600<br />
FE_Pulses<br />
time / ns<br />
0.88<br />
9800 10000 10200 10400 10600<br />
Figure 8.19: Examples <strong>for</strong> wave<strong>for</strong>ms in which NFE (upper image each) extracts significantly<br />
less charge than FE (lower image each). Pulses are indicated by dotted lines with<br />
<strong>the</strong> charge given in units <strong>of</strong> PE.<br />
102
8.3 Comparison with O<strong>the</strong>r <strong>Feature</strong> <strong>Extractor</strong>s<br />
10 7<br />
10 6<br />
width: 0.12<br />
mean: -0.09<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
40 20 0 20 40<br />
∆q pulses<br />
10 7<br />
10 6<br />
width: 0.26<br />
mean: 0.03<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
40 20 0 20 40<br />
∆q pulses<br />
Figure 8.20: Distributions <strong>of</strong> <strong>the</strong> differences between <strong>the</strong> total charges <strong>of</strong> all pulses extracted<br />
by FE (multi-pulse online-filtering settings) <strong>and</strong> NFE (with En<strong>for</strong>cePulse set).<br />
Positive values correspond to higher charges <strong>for</strong> FE; red lines indicate Gaussian<br />
fits.<br />
Top: Monte-Carlo data; bottom: experimental data.<br />
103
8 PERFORMANCE TESTS<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
ATWD (blue) <strong>and</strong> FADC (red) amplitudes / PE<br />
100<br />
80<br />
60<br />
40<br />
20<br />
0<br />
100<br />
80<br />
60<br />
40<br />
20<br />
0<br />
100<br />
80<br />
60<br />
40<br />
20<br />
0<br />
100<br />
80<br />
60<br />
40<br />
20<br />
0<br />
23.41<br />
331.57<br />
9.02<br />
1.78<br />
41.98<br />
17.79<br />
3.14<br />
15.52<br />
6.35<br />
20.62<br />
12.33<br />
12.98<br />
7.66<br />
4.37<br />
6.01<br />
5.26<br />
6.21<br />
NFE_MergedPulses<br />
8.02<br />
2.92<br />
0.76<br />
1.55<br />
1.59<br />
6.23<br />
1.22<br />
0.29<br />
12000 12100 12200 12300 12400<br />
FE_Pulses<br />
time / ns<br />
30.68<br />
25.61<br />
44.67<br />
250.20<br />
55.18<br />
12000 12100 12200 12300 12400<br />
0.36<br />
0.75<br />
NFE_MergedPulses<br />
1.87<br />
0.26<br />
0.49<br />
18.26<br />
42.32<br />
3.07<br />
29.75<br />
4.40<br />
0.54<br />
419.15<br />
8.75<br />
0.84 6.51<br />
0.51<br />
23.47<br />
5.40<br />
140.35 14.61 9.05<br />
1.71 1.37 0.94 0.53<br />
6.37 1.71<br />
60.92<br />
17.12<br />
10.35<br />
0.44 2.05<br />
8.16<br />
34.48<br />
0.39<br />
6.37<br />
1.21<br />
OMKey(37,26)<br />
12000 12100 12200 12300 12400<br />
FE_Pulses<br />
time / ns<br />
43.88<br />
26.38<br />
118.37<br />
147.56<br />
134.07<br />
109.03<br />
75.19<br />
12000 12100 12200 12300 12400<br />
1.55<br />
0.45<br />
0.42<br />
0.79<br />
0.88<br />
0.74<br />
0.89<br />
1.44<br />
1.06<br />
OMKey(37,26)<br />
0.98<br />
1.80<br />
0.79<br />
Figure 8.21: Example <strong>of</strong> saturated wave<strong>for</strong>ms with pulses from NFE (upper image each) <strong>and</strong><br />
FE (lower image each). Pulses are indicated by dotted lines with <strong>the</strong> charge<br />
given in units <strong>of</strong> PE. The wave<strong>for</strong>ms have been calibrated with <strong>the</strong> IC59<br />
online-filtering settings <strong>for</strong> DOMcalibrator (top) <strong>and</strong> its <strong>of</strong>fline-processing settings<br />
(SaturationLevel = 900, bottom), respectively.<br />
104
8.3 Comparison with O<strong>the</strong>r <strong>Feature</strong> <strong>Extractor</strong>s<br />
NFE (negative values) are mostly caused by saturated wave<strong>for</strong>ms; <strong>for</strong> those, <strong>Feature</strong><strong>Extractor</strong>’s<br />
option TinyThreshold causes pulses below 5% <strong>of</strong> <strong>the</strong> charge <strong>of</strong> <strong>the</strong> highest pulse<br />
to be ignored. This can be seen in figure 8.21, where FE does not accept pulses during<br />
most <strong>of</strong> <strong>the</strong> wave<strong>for</strong>m. TinyThreshold is set to zero in <strong>the</strong> IC59 <strong>of</strong>fline-reconstruction<br />
settings.[51]<br />
The tails towards higher charges from FE are probably caused by FADC wave<strong>for</strong>ms that<br />
are heavily affected by droop. FE successfully employs its own methods to correct <strong>for</strong><br />
this effect. This has to be investigated fur<strong>the</strong>r with <strong>the</strong> aim to include <strong>the</strong>se methods in<br />
DOMcalibrator.<br />
<strong>Feature</strong><strong>Extractor</strong> <strong>of</strong>ten produces many more pulses in ATWD wave<strong>for</strong>ms because it<br />
defines a pulse from two bins or from one bin <strong>of</strong> its deconvoluted distribution while NFE<br />
uses three bins each, see section 6.3.1: FE extracts up to 127 pulses per ATWD wave<strong>for</strong>m,<br />
while NFE usually extracts 64 at most. For FADC on <strong>the</strong> o<strong>the</strong>r h<strong>and</strong> <strong>Feature</strong><strong>Extractor</strong><br />
extracts comparatively few pulses because its FADC algorithm is not capable <strong>of</strong> splitting<br />
long features, while NFE usually extracts up to 128 pulses from a saturated FADC wave<strong>for</strong>m.<br />
This can be verified in figure 8.22, where <strong>the</strong> FE’s distribution <strong>for</strong> Monte-Carlo has its<br />
first bend at about 150 pulses per wave<strong>for</strong>m <strong>and</strong> <strong>the</strong>n quickly falls <strong>of</strong>f, whereas NFE has<br />
a first bend at about 50 pulses per wave<strong>for</strong>m, but does not fall <strong>of</strong>f quickly because <strong>of</strong> high<br />
numbers <strong>of</strong> FADC pulses. Very high numbers <strong>of</strong> pulses in Monte-Carlo data are caused<br />
by DOMs which were launched twice; in current experimental data, first launch cleaning<br />
cuts away wave<strong>for</strong>ms from later launches.<br />
In summary, <strong>Feature</strong><strong>Extractor</strong> finds less pulses on average because <strong>of</strong> its FADC algorithm,<br />
<strong>and</strong> also because TinyThreshold rejects low pulses in very bright wave<strong>for</strong>ms.<br />
IC59 <strong>of</strong>fline-processing settings<br />
The recommended settings <strong>for</strong> IC59 <strong>of</strong>fline-processing have been used in connection<br />
with <strong>the</strong> recommended settings <strong>for</strong> DOMcalibrator, i. e., SaturationLevel = 900.[62][51]<br />
For <strong>Feature</strong><strong>Extractor</strong>, <strong>the</strong> changes are:<br />
• ADCThreshold = 1.1 −→ ADCThreshold = 1.8<br />
The charge threshold to decide whe<strong>the</strong>r to accept a pulse or not was increased to<br />
take account <strong>for</strong> a recent change in <strong>the</strong> DOMs’ discriminator thresholds. The effect<br />
<strong>of</strong> this change is that <strong>for</strong> pulses that were previously split excessively now many<br />
fractions fall below <strong>the</strong> threshold; this results in a higher charge per pulse because<br />
<strong>the</strong> remaining pulses are rescaled (see section 6.3.1). The drawbacks are a higher<br />
number <strong>of</strong> missed pulses <strong>and</strong> more unphysical redistribution <strong>of</strong> charge between pulses<br />
<strong>of</strong> <strong>the</strong> same wave<strong>for</strong>m.<br />
105
8 PERFORMANCE TESTS<br />
10 7<br />
10 6<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
0 50 100 150 200 250<br />
n pulses<br />
10 7<br />
10 6<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
0 50 100 150<br />
n pulses<br />
Figure 8.22: Distributions <strong>of</strong> <strong>the</strong> numbers <strong>of</strong> pulses extracted by FE (multi-pulse onlinefiltering<br />
settings, blue area) <strong>and</strong> NFE (with En<strong>for</strong>cePulse set, green line).<br />
Top: Monte-Carlo data; bottom: experimental data.<br />
106
8.3 Comparison with O<strong>the</strong>r <strong>Feature</strong> <strong>Extractor</strong>s<br />
10 7<br />
10 6<br />
width: 0.90 0.81<br />
mean: 10.18 11.19<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
40 20 0 20 40<br />
t MC −t pulse /ns<br />
Figure 8.23: Distributions <strong>of</strong> <strong>the</strong> time residuals <strong>of</strong> <strong>the</strong> first pulses extracted by FE (multipulse<br />
<strong>of</strong>fline-processing settings, blue area) <strong>and</strong> NFE (with En<strong>for</strong>cePulse set,<br />
green line) from Monte-Carlo data; red lines indicate Gaussian fits.<br />
107
8 PERFORMANCE TESTS<br />
600000<br />
width: 0.32 0.36<br />
mean: 0.88 0.93<br />
500000<br />
400000<br />
entries<br />
300000<br />
200000<br />
100000<br />
0<br />
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0<br />
q pulse1<br />
600000<br />
width: 0.35 0.39<br />
mean: 0.96 0.90<br />
500000<br />
400000<br />
entries<br />
300000<br />
200000<br />
100000<br />
0<br />
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0<br />
q pulse1<br />
Figure 8.24: Distributions <strong>of</strong> <strong>the</strong> charges <strong>of</strong> <strong>the</strong> first pulses extracted by FE (multi-pulse <strong>of</strong>flineprocessing<br />
settings, blue area) <strong>and</strong> NFE (with En<strong>for</strong>cePulse set, green line); red<br />
lines indicate Gaussian fits.<br />
Top: Monte-Carlo data; bottom: experimental data.<br />
108
8.3 Comparison with O<strong>the</strong>r <strong>Feature</strong> <strong>Extractor</strong>s<br />
10 7<br />
10 6<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
0 50 100 150 200 250<br />
n pulses<br />
10 7<br />
10 6<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
0 50 100 150 200<br />
n pulses<br />
Figure 8.25: Distributions <strong>of</strong> <strong>the</strong> numbers <strong>of</strong> pulses extracted by FE (multi-pulse <strong>of</strong>flineprocessing<br />
settings, blue area) <strong>and</strong> NFE (with En<strong>for</strong>cePulse set, green line).<br />
Top: Monte-Carlo data; bottom: experimental data.<br />
109
8 PERFORMANCE TESTS<br />
• ExclusionSize = 5 −→ ExclusionSize = 1<br />
This parameters governs <strong>the</strong> deadtime between ATWD <strong>and</strong> FADC in units <strong>of</strong> FADC<br />
bin lengths during which no FADC pulses are accepted because <strong>the</strong>ir ATWD counterpart<br />
might already have been extracted (double extraction, see appendix C.1);<br />
it was largely overestimated in <strong>the</strong> old settings. The consequence <strong>of</strong> this change is<br />
higher total charge <strong>and</strong> more accurate extraction per<strong>for</strong>mance.<br />
• TinyThreshold = 0.05 −→ TinyThreshold = 0.00<br />
TinyThreshold was originally introduced to prevent <strong>the</strong> extraction <strong>of</strong> pulses from<br />
droop artifacts <strong>and</strong> prepulses. However, it hampers <strong>the</strong> accurate extraction <strong>of</strong> wave<strong>for</strong>ms<br />
(see figure 8.21) <strong>and</strong> can cause direct hits to be ignored. This was judged to<br />
be more important, so <strong>the</strong> threshold was disabled. The results are more accurate<br />
extraction <strong>and</strong> higher total charge, because <strong>the</strong> charge attributed to pulses which<br />
do not pass TinyThreshold is <strong>for</strong>feit.<br />
• PMTTransit = 2 −→ PMTTransit = -1<br />
This parameter influences <strong>the</strong> timing distribution by adding a time <strong>of</strong>fset <strong>and</strong> correcting<br />
<strong>for</strong> a correlation between <strong>the</strong> PMT voltage <strong>and</strong> <strong>the</strong> pulse time; however,<br />
observed individually, disabeling <strong>the</strong> correction improves <strong>the</strong> time resolution measured<br />
by <strong>the</strong> Gaussian width from 1.23 ns (not shown) to 0.90 ns (figure 8.23).<br />
The <strong>of</strong>fline-processing settings increase <strong>Feature</strong><strong>Extractor</strong>’s average charge per pulse<br />
towards 1 PE (figure 8.24: 0.88 PE <strong>for</strong> Monte-Carlo <strong>and</strong> 0.96 PE <strong>for</strong> experimental data<br />
instead <strong>of</strong> 0.81 PE <strong>for</strong> each in figure 8.18). However, <strong>the</strong> <strong>new</strong> settings also approximately<br />
double <strong>the</strong> number <strong>of</strong> missed first pulses to about 177000 in 4.1 million wave<strong>for</strong>ms (4.3%),<br />
see figure 8.23. Besides, <strong>the</strong> time resolution increases from 1.23 ns to 0.90 ns.<br />
The total number <strong>of</strong> pulses (figure 8.25) decreases slightly because <strong>of</strong> <strong>the</strong> lower ADCThreshold,<br />
<strong>and</strong> <strong>the</strong>re are more wave<strong>for</strong>ms with exeptionally high numbers <strong>of</strong> pulses ( 100) because<br />
<strong>of</strong> <strong>the</strong> disabled TinyThreshold.<br />
In total, <strong>the</strong> changes were carefully compiled by many people <strong>and</strong> provide <strong>for</strong> an improvement<br />
upon <strong>the</strong> older online-filtering settings, yet <strong>the</strong>y also increase <strong>the</strong> number <strong>of</strong><br />
missed pulses <strong>and</strong> do not generally improve <strong>the</strong> charge per pulse:<br />
A related test originally conducted by Juanan Aguilar[63] compares <strong>the</strong> agreement <strong>of</strong> <strong>the</strong><br />
ratio <strong>of</strong> charge per pulse between realistically simulated data (CORSIKA) <strong>and</strong> current<br />
experimental data. The results can be seen in figure 8.26. The simulated data uses <strong>the</strong><br />
old discriminator thresholds, so it has to be extracted with ADCTreshold = 1.1 (onlinefiltering<br />
settings). In contrast, <strong>the</strong> experimental data was taken with <strong>the</strong> <strong>new</strong> thresholds,<br />
so <strong>the</strong> <strong>of</strong>fline-processing settings are recommended; <strong>the</strong> ratio <strong>for</strong> experimental data is<br />
shown <strong>for</strong> both <strong>Feature</strong><strong>Extractor</strong> settings.<br />
As expected, <strong>the</strong> agreement is better <strong>for</strong> <strong>the</strong> <strong>of</strong>fline-processing settings <strong>of</strong> <strong>Feature</strong><strong>Extractor</strong>,<br />
however it is worse <strong>for</strong> ratios near 0.3 PE per pulse.<br />
For NFE, <strong>the</strong> agreement is good <strong>for</strong> regions <strong>of</strong> high abundance. The disagreement <strong>for</strong><br />
110
8.3 Comparison with O<strong>the</strong>r <strong>Feature</strong> <strong>Extractor</strong>s<br />
10 6<br />
10 5<br />
exp data (Run114060), onl.<br />
exp data (Run114060), <strong>of</strong>fl.<br />
corsika (1628), onl.<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
1.0 0.5 0.0 0.5 1.0 1.5 2.0<br />
log 10 (q tot PE −1 / n pulses )<br />
10 6<br />
10 5<br />
exp data (Run114060)<br />
corsika (1628)<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
1.0 0.5 0.0 0.5 1.0 1.5 2.0<br />
log 10 (q tot PE −1 / n pulses )<br />
Figure 8.26: Charge per pulse ratio <strong>for</strong> simulated data (dotted line) <strong>and</strong> experimental data<br />
(solid lines) <strong>for</strong> FE (top, online <strong>and</strong> <strong>of</strong>fline settings) <strong>and</strong> NFE (bottom, default<br />
settings).<br />
111
8 PERFORMANCE TESTS<br />
10 7<br />
10 6<br />
width: 1.23 0.81<br />
mean: 0.55 11.19<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
40 20 0 20 40<br />
t MC −t pulse /ns<br />
Figure 8.27: Distributions <strong>of</strong> <strong>the</strong> time residuals <strong>of</strong> <strong>the</strong> first pulses extracted by FE (singlepulse<br />
settings, blue area) <strong>and</strong> NFE (with En<strong>for</strong>cePulse set, green line) from<br />
Monte-Carlo data; red lines indicate Gaussian fits.<br />
low charges is probably caused by <strong>the</strong> missing simulation <strong>of</strong> <strong>the</strong> daq_baseline; <strong>the</strong> disagreement<br />
<strong>for</strong> ratios higher than 3 PE per pulse has to be examined. Both datasets have<br />
been extracted using <strong>the</strong> same (default) settings; NFE is not directly influenced by DOM<br />
discriminator threshold changes because all thresholds are defined in units <strong>of</strong> PE.<br />
8.3.2 <strong>Feature</strong><strong>Extractor</strong> in Single-Pulse Mode<br />
NFE’s results were also compared to <strong>Feature</strong><strong>Extractor</strong>’s IC59 single-pulse extraction results<br />
because this was <strong>the</strong> mode <strong>of</strong> operation <strong>for</strong> <strong>the</strong> IC59 online-filtering muon track<br />
reconstruction; FE is configured to use its second single-pulse extraction algorithm <strong>and</strong><br />
to only extract ATWD wave<strong>for</strong>ms.[51] For best comparability, En<strong>for</strong>cePulse was set <strong>for</strong><br />
NFE <strong>and</strong> no FADC pulses were extracted; still NFE’s distributions shown in this section<br />
closely resemble those in section 8.3.1 because <strong>the</strong> missing FADC pulses barely affect <strong>the</strong><br />
first pulse.<br />
The distribution <strong>of</strong> <strong>the</strong> time residuals <strong>of</strong> <strong>the</strong> first pulses (figure 8.27) <strong>for</strong> FE is similar<br />
to those obtained by FE in its multi-pulse mode; this is not surprising because FE uses<br />
<strong>the</strong> single-pulse extraction time to replace <strong>the</strong> time <strong>of</strong> <strong>the</strong> Bayesian Unfolding pulse closest<br />
112
8.4 SLCHit<strong>Extractor</strong><br />
to it (section 6.3.1), <strong>and</strong> <strong>the</strong> single-pulse extraction algorithm extracts <strong>the</strong> first feature if<br />
can find. The distribution <strong>of</strong> <strong>the</strong> time residuals shows more early fake pulses; however,<br />
<strong>the</strong> shapes <strong>and</strong> widths <strong>of</strong> <strong>the</strong> distributions <strong>of</strong> <strong>the</strong> time differences between FE <strong>and</strong> NFE<br />
are effectively <strong>the</strong> same (±0.01 ns, not shown).<br />
In figure 8.28, <strong>the</strong> distributions <strong>of</strong> <strong>the</strong> charges <strong>of</strong> <strong>the</strong> first pulses have a strong tail<br />
towards high values because <strong>of</strong> <strong>the</strong> algorithm’s inability to split features. Fur<strong>the</strong>rmore<br />
<strong>the</strong>y are less stable regarding <strong>the</strong> transition from Monte-Carlo to experimental data: Their<br />
mean increases from 0.87 PE to 1.04 PE, <strong>and</strong> <strong>the</strong> tail becomes significantly stronger <strong>for</strong><br />
experimental data; this is probably caused by <strong>the</strong> wrong charge simulation (figure 7.12)<br />
<strong>and</strong> <strong>the</strong> datasets’ different energy spectra.<br />
The distribution <strong>of</strong> <strong>the</strong> differences <strong>of</strong> <strong>the</strong> total charge per wave<strong>for</strong>m <strong>for</strong> simulated data<br />
in figure 8.29 reveals an on average higher total charge <strong>for</strong> NFE, despite <strong>the</strong> fact that<br />
FE calculates <strong>the</strong> integrated charge; this is due to <strong>the</strong> wrong baseline simulation, which<br />
is <strong>of</strong>ten fixed by <strong>Feature</strong><strong>Extractor</strong>’s own baseline correction algorithm. For experimental<br />
data, FE’s integrated charge is higher than NFE’s charge because <strong>the</strong> latter is extracted<br />
from features only.<br />
8.4 SLCHit<strong>Extractor</strong><br />
SLCHit<strong>Extractor</strong> is currently used <strong>for</strong> <strong>the</strong> IC59 SLC charge stamp extraction, which<br />
motivates tests <strong>of</strong> its per<strong>for</strong>mance <strong>and</strong> its use as a benchmark <strong>for</strong> NFE; see appendix C.4.<br />
SLCHit<strong>Extractor</strong>’s time resolution is better than “SLCHE”’s with a Gaussian width<br />
<strong>of</strong> 1.91 ns instead <strong>of</strong> 2.25 ns, see figure 8.30; both <strong>of</strong> <strong>the</strong>m have better resolutions than <strong>the</strong><br />
(more flexible) NFE default combination <strong>of</strong> FADC algorithms (∼ 3 ns, figure 7.10), <strong>and</strong><br />
both resolutions are well below even one ATWD wave<strong>for</strong>m bin length, so both are deemed<br />
to be excellent.<br />
SLCHit<strong>Extractor</strong>’s pulse times match those <strong>of</strong> FE’s ATWD pulses (if FE’s transit time<br />
correction is activated), “SLCHE”’s instead match <strong>the</strong> time <strong>of</strong> NFE’s FADC pulses (see<br />
section 7.1.4).<br />
The charge distribution <strong>of</strong> SLCHit<strong>Extractor</strong> is less stable than that <strong>of</strong> “SLCHE” (figure<br />
8.31); it shifts by 0.07 PE between Monte-Carlo <strong>and</strong> experimental data. It also shows an<br />
unexplained excess in its 0.8 PE bin. Compared to “SLCHE” <strong>the</strong> mean <strong>of</strong> <strong>the</strong> distribution<br />
is closer to 1 PE because it was calibrated to match <strong>the</strong> distribution <strong>of</strong> <strong>Feature</strong><strong>Extractor</strong><br />
in its <strong>of</strong>fline-reconstruction settings.[64] Correspondignly, “SLCHE” matches <strong>the</strong> average<br />
FADC charge extracted by NFE from simulated data.<br />
113
8 PERFORMANCE TESTS<br />
500000<br />
width: 0.38 0.36<br />
mean: 0.87 0.93<br />
400000<br />
300000<br />
entries<br />
200000<br />
100000<br />
0<br />
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0<br />
q pulse1<br />
300000<br />
width: 0.51 0.39<br />
mean: 1.04 0.90<br />
250000<br />
200000<br />
entries<br />
150000<br />
100000<br />
50000<br />
0<br />
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0<br />
q pulse1<br />
Figure 8.28: Distributions <strong>of</strong> <strong>the</strong> charges <strong>of</strong> <strong>the</strong> first pulses extracted by FE (single-pulse settings,<br />
blue area) <strong>and</strong> NFE (with En<strong>for</strong>cePulse set, green line); red lines indicate<br />
Gaussian fits.<br />
Top: Monte-Carlo data; bottom: experimental data.<br />
114
8.4 SLCHit<strong>Extractor</strong><br />
10 7<br />
10 6<br />
width: 0.13<br />
mean: -0.06<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
40 20 0 20 40<br />
∆q pulses<br />
10 7<br />
10 6<br />
width: 0.22<br />
mean: 0.14<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
50 40 30 20 10 0 10 20 30<br />
∆q pulses<br />
Figure 8.29: Distributions <strong>of</strong> <strong>the</strong> differences between <strong>the</strong> total charges <strong>of</strong> all pulses extracted<br />
by FE (single-pulse settings) <strong>and</strong> NFE (with En<strong>for</strong>cePulse set). Positive values<br />
correspond to higher charges <strong>for</strong> FE; red lines indicate Gaussian fits.<br />
Top: Monte-Carlo data; bottom: experimental data.<br />
115
8 PERFORMANCE TESTS<br />
10 7<br />
10 6<br />
width: 1.91 2.25<br />
mean: -0.03 10.70<br />
10 5<br />
10 4<br />
entries<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 -1<br />
80 60 40 20 0 20 40 60<br />
t MC −t pulse /ns<br />
Figure 8.30: Distributions <strong>of</strong> <strong>the</strong> time residuals <strong>of</strong> <strong>the</strong> first pulses extracted by SLCHit<strong>Extractor</strong><br />
(blue area) <strong>and</strong> NFE’s “SLCHE” (green line) from Monte-Carlo data; red<br />
lines indicate Gaussian fits.<br />
116
8.4 SLCHit<strong>Extractor</strong><br />
700000<br />
600000<br />
width: 0.43 0.33<br />
mean: 0.96 0.78<br />
500000<br />
entries<br />
400000<br />
300000<br />
200000<br />
100000<br />
0<br />
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0<br />
q pulse1<br />
1400000<br />
1200000<br />
width: 0.46 0.35<br />
mean: 1.03 0.79<br />
1000000<br />
entries<br />
800000<br />
600000<br />
400000<br />
200000<br />
0<br />
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0<br />
q pulse1<br />
Figure 8.31: Distributions <strong>of</strong> <strong>the</strong> charges <strong>of</strong> <strong>the</strong> first pulses extracted by SLCHit<strong>Extractor</strong><br />
(blue area) <strong>and</strong> NFE’s “SLCHE” (green line); red lines indicate Gaussian fits.<br />
Top: Monte-Carlo data; bottom: experimental data.<br />
117
8 PERFORMANCE TESTS<br />
8.5 Runtime Per<strong>for</strong>mance<br />
NFE was designed to be viable <strong>for</strong> real-time online processing. Since computing capacity<br />
at <strong>the</strong> South Pole is limited, efficient algorithms <strong>and</strong> data structures were used to generate<br />
fast code.<br />
The pr<strong>of</strong>iler callgrind/KCacheGrind was used to analyze <strong>the</strong> code. As expected,<br />
<strong>the</strong> most CPU time consuming part is “BayesUnfold”; <strong>the</strong> fraction <strong>of</strong> <strong>the</strong> time spent<br />
at this algorithm <strong>for</strong> ATWD with default NFE settings is 94.5% <strong>for</strong> E −1 Monte-Carlo<br />
data, <strong>and</strong> 89.6% <strong>for</strong> experimental data. Of this time, “BayesUnfold”’s unfolding method<br />
GetUnfolding() takes <strong>the</strong> highest share by far (93.9% <strong>of</strong> <strong>the</strong> total time <strong>for</strong> Monte-Carlo,<br />
89.0% <strong>for</strong> experimental data). Thus, <strong>the</strong> focus in code optimization was on this unfolding<br />
method.<br />
Most <strong>of</strong> <strong>the</strong> remaining time (4.8% resp. 9.8%) is spent at I3Frame::Get() to read in<br />
<strong>the</strong> input data, where calibration in<strong>for</strong>mations constitute <strong>the</strong> largest share <strong>of</strong> <strong>the</strong> time<br />
requirements (> 80%).<br />
For “BayesUnfold”’s GetUnfolding(), <strong>the</strong> number <strong>of</strong> calls to expensive functions or<br />
oper<strong>and</strong>s such as double::/ was minimized <strong>and</strong> temporary objects were avoided if possible.<br />
The use <strong>of</strong> <strong>the</strong> iterative stopping condition saves roughly 10 iterations on average<br />
(see section 7.1.3) while it costs less than 10% <strong>of</strong> <strong>the</strong> CPU time to check <strong>the</strong> breaking<br />
conditions <strong>for</strong> every iteration. Ano<strong>the</strong>r tweak <strong>for</strong> BU was <strong>the</strong> introduction <strong>of</strong> templated<br />
structs containing <strong>the</strong> source specific parameters from table 7.3 in <strong>for</strong>m <strong>of</strong> static constants;<br />
this allows <strong>the</strong> compiler to optimize <strong>the</strong> code more effectively <strong>and</strong> led to a gain in<br />
speed <strong>of</strong> about 19%.<br />
In comparison with o<strong>the</strong>r feature extractors, NFE shows superior runtime per<strong>for</strong>mance,<br />
mostly because <strong>of</strong> <strong>the</strong> use <strong>of</strong> multiple algorithms, see table 8.1. The systematically shorter<br />
times <strong>for</strong> experimental data are caused by its s<strong>of</strong>ter energy spectrum, since low-energy<br />
events <strong>of</strong>fer less wave<strong>for</strong>ms to extract.<br />
Switching from <strong>Feature</strong><strong>Extractor</strong> to NFE <strong>for</strong> online processing would reduce <strong>the</strong> average<br />
3.76<br />
time required per event to about ≈ 35% <strong>and</strong> hence would free up many <strong>of</strong> <strong>the</strong><br />
8.47+2.24<br />
20 . . . 25 CPUs currently needed <strong>for</strong> feature extraction at <strong>the</strong> South Pole. If future muon<br />
track reconstruction scripts still employ algorithms that depend on integrated charge, a<br />
small <strong>and</strong> fast module can easily be created to join all pulses <strong>of</strong> each wave<strong>for</strong>m.<br />
118
8.5 Runtime Per<strong>for</strong>mance<br />
Table 8.1: Runtimes <strong>for</strong> different feature extractors <strong>and</strong> datasets; errors are about ±3%.<br />
Module Sources runtime per event<br />
Simulated data,<br />
Dataset 3071<br />
t ms −1<br />
Experimental data,<br />
Run 113587, L1<br />
t ms −1<br />
FE multi-pulse ATWD+FADC 33.89 8.47<br />
FE single-pulse ATWD 7.89 2.24<br />
PE ATWD 18.49 5.39<br />
PE FADC 37.89 10.66<br />
SLCHit<strong>Extractor</strong> SLC < 0.50 < 0.27<br />
NFE ATWD 4.19 1.21<br />
NFE FADC 9.06 2.55<br />
NFE ATWD+FADC 13.25 3.76<br />
NFE SLC < 0.50 < 0.27<br />
NFE PulseMerger — ≪ 0.50 ≪ 0.27<br />
NFE En<strong>for</strong>cePulse ATWD 4.21 1.19<br />
NFE En<strong>for</strong>cePulse FADC 9.45 2.61<br />
NFE En<strong>for</strong>cePulse ATWD+FADC 13.87 3.77<br />
NFE BU only ATWD 20.13 5.70<br />
NFE BU only FADC 30.21 8.30<br />
119
8 PERFORMANCE TESTS<br />
120
CHAPTER IX<br />
Summary And Outlook
9 SUMMARY AND OUTLOOK<br />
Within this <strong>the</strong>sis a <strong>new</strong> feature extraction package <strong>for</strong> recorded photomultiplier signals<br />
in <strong>the</strong> IceCube Neutrino Observatory at South Pole was developed. Its task is to search<br />
<strong>the</strong> wave<strong>for</strong>ms captured by <strong>the</strong> digital optical modules <strong>for</strong> signals caused by photons.<br />
This in<strong>for</strong>mation is made available to o<strong>the</strong>r s<strong>of</strong>tware modules by extracting it from <strong>the</strong><br />
wave<strong>for</strong>ms’ features into pulses.<br />
Existing feature extractor algorithms have been analyzed conceptually <strong>and</strong> with respect<br />
to <strong>the</strong>ir per<strong>for</strong>mances. A concept <strong>for</strong> a <strong>new</strong> feature extractor was designed <strong>and</strong> key<br />
characteristics were defined: The <strong>new</strong> feature extractor – called NFE – should have modular,<br />
maintainable <strong>and</strong> well-documented code, it should be easy to use, flexible enough<br />
to cover all feature extraction dem<strong>and</strong>s, it should be reasonably fast, <strong>and</strong> provide good<br />
extraction per<strong>for</strong>mance in terms <strong>of</strong> miss rate, noise rate, <strong>and</strong> charge <strong>and</strong> time resolutions.<br />
A <strong>new</strong> technique <strong>of</strong> dynamically choosing an appropriate algorithm according to <strong>the</strong><br />
wave<strong>for</strong>m’s complexity was designed <strong>and</strong> implemented. The implemented algorithms are<br />
“Eva” to quickly decide which extraction algorithm to use on <strong>the</strong> wave<strong>for</strong>m, “Simple”<br />
to extract pulses from wave<strong>for</strong>ms with exclusively SPE-like features, “BayesUnfold” to<br />
extract complex features, <strong>and</strong> “SLCHE” to extract pulses from SLC chargestamps. The<br />
free parameters <strong>of</strong> <strong>the</strong>se algorithms were calibrated using simulated datasets.<br />
The per<strong>for</strong>mance <strong>of</strong> <strong>the</strong> <strong>new</strong> feature extractor was tested under various conditions.<br />
The technique <strong>of</strong> dynamically choosing an appropriate algorithm proved to be successfull;<br />
it improves <strong>the</strong> extraction quality <strong>and</strong> drastically speeds up <strong>the</strong> process. <strong>Test</strong>ing NFE<br />
with individual wave<strong>for</strong>ms yields good results, in many cases NFE seems to per<strong>for</strong>m better<br />
than <strong>the</strong> currently used <strong>Feature</strong><strong>Extractor</strong> especially at small features.<br />
Several characteristic distributions <strong>of</strong> resulting quantities <strong>for</strong> different feature extractors<br />
were both tested individually <strong>and</strong> compared to each o<strong>the</strong>r. NFE’s distributions are promising<br />
<strong>and</strong> support <strong>the</strong> positive impression gained from <strong>the</strong> individual wave<strong>for</strong>m checks. Finally,<br />
<strong>the</strong> CPU efficiency was found to be significantly superior to that <strong>of</strong> o<strong>the</strong>r feature<br />
extractors.<br />
On March <strong>the</strong> 1 st , 2010, NFE passed <strong>the</strong> collaboration’s code review with positive<br />
remarks. As a consequence <strong>the</strong> project will be <strong>of</strong>ficially released soon after all comments<br />
have been incorporated. Fortunately, a <strong>new</strong> version <strong>of</strong> <strong>the</strong> simulation meta-project with<br />
various bugfixes will be released soon <strong>and</strong> can be used to verify or recalibrate <strong>the</strong> free<br />
parameters <strong>of</strong> NFE’s algorithms; this is required as errors in <strong>the</strong> current simulated datasets<br />
affect feature extraction.<br />
For now, some open issues remain: More tests have to be conducted concerning saturated<br />
wave<strong>for</strong>ms. Especially <strong>the</strong> effect <strong>of</strong> transistor droop in <strong>the</strong> optical modules has to be<br />
adressed; droop correction is already per<strong>for</strong>med by <strong>the</strong> s<strong>of</strong>tware module DOMcalibrator,<br />
however sometimes droop effects remain. The original <strong>Feature</strong><strong>Extractor</strong> seems to employ<br />
a powerful method <strong>of</strong> fur<strong>the</strong>r repair wave<strong>for</strong>ms, which could probably be implemented<br />
into DOMcalibrator after exhaustive tests.<br />
122
Ano<strong>the</strong>r open issue are tests with flasher runs. Usually, experimental data does not <strong>of</strong>fer<br />
in<strong>for</strong>mation about true hit times, so one has to rely on simulated datasets, accepting systematical<br />
errors. Flasher runs <strong>of</strong>fer an alternative because <strong>the</strong> time at which <strong>the</strong> flasher –<br />
i. e., a light source inside <strong>the</strong> detector – was activated is well-known. There<strong>for</strong>e <strong>the</strong>y can<br />
be used to verify <strong>the</strong> parameter settings <strong>and</strong> to check <strong>the</strong> extraction per<strong>for</strong>mance.<br />
Fur<strong>the</strong>rmore, NFE lacks pybindings, i. e., an interface to directly access <strong>the</strong> algorithms<br />
from a possibly interactive Python session. Pybindings can be useful <strong>for</strong> <strong>the</strong> verification<br />
<strong>of</strong> <strong>the</strong> low-level reconstruction <strong>and</strong> will be part <strong>of</strong> a future release.<br />
With its release, NFE will be open <strong>for</strong> analyses by <strong>the</strong> collaboration to decide whe<strong>the</strong>r it<br />
should replace or complement <strong>the</strong> existing feature extractors in <strong>the</strong> <strong>of</strong>ficial reconstruction<br />
data chain. Regardless <strong>of</strong> this decision, much was learned about IceCube’s low-level<br />
reconstruction, <strong>and</strong> some improvements were made.<br />
123
9 SUMMARY AND OUTLOOK<br />
124
APPENDIX A<br />
Bayesian Unfolding<br />
A.1 Formal Approach<br />
Formally, one considers as given a histogram whose n E entries n Ei are interpreted as<br />
numbers <strong>of</strong> effects E i which are caused by a not necessarily equal number n C <strong>of</strong> causes C j .<br />
With n tot := ∑ n Ei one can define <strong>the</strong> probability <strong>for</strong> effect E i to occur as P (E i ) := n E i<br />
n tot<br />
.<br />
Fur<strong>the</strong>rmore, <strong>the</strong> probability <strong>for</strong> a certain C j to cause each <strong>of</strong> <strong>the</strong> effects is given by<br />
P (E i |C j ).<br />
By applying Bayes’ Theorem, one obtaines<br />
P (C j |E i ) = P (E i|C j ) · P (C j )<br />
∑<br />
j P (E i |C j ) · P (C j )<br />
Using this <strong>and</strong> <strong>the</strong> pairwise disjointness <strong>of</strong> <strong>the</strong> E i , one can compute <strong>the</strong> probabilities<br />
P (C j ) t+1 = ∑ i<br />
P (C j |E i ) t · P (E i ) = ∑ i<br />
P (E i |C j ) · P (C j ) t<br />
∑k P (E i |C k ) · P (C k ) t<br />
P (E i ) ∀ t ∈ N<br />
iteratively, which <strong>the</strong>n can easily be trans<strong>for</strong>med into <strong>the</strong> most probable numbers <strong>of</strong> causes<br />
that occured, n Cj = P (C j ) · n tot .<br />
A.2 Adaption to IceCube’s Wave<strong>for</strong>ms<br />
The entries <strong>of</strong> <strong>the</strong> wave<strong>for</strong>m (with negative values set to zero) are <strong>the</strong> numbers <strong>of</strong> effects<br />
n Ei , <strong>the</strong> numbers <strong>of</strong> causes n Cj correspond to <strong>the</strong> charge (actually ∆t times charge, but<br />
∆t is constant <strong>and</strong> can be considered later on) belonging to a pulse originating in bin j,<br />
using <strong>the</strong> same binning <strong>for</strong> both effects <strong>and</strong> causes <strong>and</strong> <strong>the</strong>re<strong>for</strong>e n C = n E .<br />
P (E i |C j ) denotes <strong>the</strong> normalized single photo electron (SPE) pulse shape which <strong>for</strong> ATWD<br />
is given by Christopher Wendt’s parametrization[47] (see section 7.1.3)<br />
f : R ≥0 → [0, 1] : t ↦→ c<br />
(<br />
e − x−x 0<br />
b 1 + e x−x 0<br />
b 2<br />
) −8<br />
.<br />
125
A<br />
BAYESIAN UNFOLDING<br />
According to this parametrization, <strong>the</strong> first L bins contain over 99.4% <strong>of</strong> <strong>the</strong> pulse’s<br />
charge, <strong>the</strong>re<strong>for</strong>e <strong>the</strong> computation can be significantly sped up by omitting most addends<br />
<strong>of</strong> both sums:<br />
P (C j ) t+1 =<br />
j+L−1 ∑<br />
i=j<br />
P (E i |C j ) · P (C j ) t<br />
∑ ik=i−L+1<br />
P (E i |C k ) · P (C k ) t<br />
P (E i )<br />
Computation can fur<strong>the</strong>r be accelerated by saving S m := P (E m |C 0 ) ≡ P (E m+j |C j )<br />
instead <strong>of</strong> calculating P (E i |C j ) <strong>for</strong> all values <strong>of</strong> j; <strong>the</strong> resulting equation is<br />
P (C j ) t+1 =<br />
j+L−1 ∑<br />
i=j<br />
S i−j · P (C j ) t<br />
∑ ik=i−L+1<br />
S i−k · P (C k ) t<br />
P (E i )<br />
Finally, as we are interested in n Cj , we cancel out n tot :<br />
n Cj , t+1 = n Cj , t<br />
j+L−1 ∑<br />
i=j<br />
S i−j · n Ei<br />
∑ ik=i−L+1<br />
S i−k · n Ck , t<br />
126
APPENDIX B<br />
Cascade Pulse Tagging<br />
This <strong>the</strong>sis’ initial topic was to tag pulses originating from cascades to reconstruct locations<br />
<strong>of</strong> strong stochastic energy losses at muon propagation, with <strong>the</strong> aim to improve<br />
both energy reconstruction <strong>and</strong> track reconstruction (see figure B.1).<br />
This turned out to be impractical because <strong>of</strong> low Čerenkov luminosity, prepulses, <strong>and</strong><br />
too much scattering, but also demonstrated some limitations <strong>of</strong> <strong>Feature</strong><strong>Extractor</strong> such as<br />
excessive pulse splitting (section 6.3.1), which were one <strong>of</strong> <strong>the</strong> motivations <strong>for</strong> <strong>the</strong> creation<br />
<strong>of</strong> NFE. The low-level work also revealed two <strong>of</strong> <strong>the</strong> bugs found during <strong>the</strong> work on this<br />
<strong>the</strong>sis.<br />
U<br />
U<br />
t<br />
DOM<br />
t<br />
particle track<br />
Čerenkov cone<br />
U<br />
cascade light<br />
t<br />
Figure B.1: This <strong>the</strong>sis’ initial topic <strong>and</strong> one <strong>of</strong> <strong>the</strong> motivations to create NFE: Tagging <strong>of</strong><br />
pulses caused by cascades during muon propagation. If <strong>the</strong> wave<strong>for</strong>ms were as<br />
clear as <strong>the</strong> idealized ones in this illustration, cascades could be located <strong>and</strong> used<br />
<strong>for</strong> direction <strong>and</strong> energy reconstruction refinement.<br />
127
B<br />
CASCADE PULSE TAGGING<br />
128
APPENDIX C<br />
Specific Problems <strong>and</strong> Anomalies<br />
Many aspects <strong>of</strong> this <strong>the</strong>sis required checks <strong>of</strong> low-level observables or single wave<strong>for</strong>ms,<br />
owing to which several previously unknown problems or anomalies were found in <strong>the</strong><br />
projects involved. Some <strong>of</strong> those relevant to this <strong>the</strong>sis are explained below; this is by no<br />
means meant to be <strong>of</strong>fensive.<br />
C.1 ATWD FADC Time Offset Caused Double Extraction<br />
Up to very recent versions <strong>of</strong> <strong>the</strong> calibration tool DOMcal, <strong>the</strong>re was a time <strong>of</strong>fset between<br />
<strong>the</strong> ATWD <strong>and</strong> FADC wave<strong>for</strong>ms <strong>of</strong> about 15 ns <strong>for</strong> experimental data <strong>and</strong> 34 ns<br />
<strong>for</strong> simulated data. This did not only corrupt <strong>the</strong> late (i. e. FADC) pulse times, but also<br />
caused pulses extracted by FE just preparatory to <strong>the</strong> end <strong>of</strong> <strong>the</strong> ATWD wave<strong>for</strong>m to<br />
be extracted again in <strong>the</strong> FADC wave<strong>for</strong>m, effectively doubling <strong>the</strong>ir impact because FE<br />
automatically merged all ATWD pulses with all FADC pulses found outside <strong>the</strong> AWTD<br />
wave<strong>for</strong>m’s timespan.<br />
This double extraction problem was solved by <strong>the</strong> respective projects’ authors by introducing<br />
a time shift in DOMcalibrator <strong>and</strong> an exclusion time window in FE (25 ns by<br />
default), during which no FADC pulses are extracted right after <strong>the</strong> end <strong>of</strong> <strong>the</strong> ATWD<br />
wave<strong>for</strong>m.<br />
C.2 <strong>Implementation</strong> <strong>of</strong> <strong>the</strong> Second Single-Pulse Extraction Algorithm<br />
in <strong>Feature</strong><strong>Extractor</strong><br />
The implementation <strong>of</strong> <strong>the</strong> second single-pulse extraction algorithm in <strong>Feature</strong><strong>Extractor</strong><br />
(section 5.1) differs from <strong>the</strong> documentation that ships with <strong>the</strong> source code[50] <strong>and</strong> from<br />
its depiction in presentations[53] (figure C.1).<br />
The pulse width is not determined by <strong>the</strong> number <strong>of</strong> bins that pass half <strong>of</strong> <strong>the</strong> first pulse’s<br />
amplitude, but by a constant multiplied with <strong>the</strong> extracted charge, <strong>and</strong> divided by <strong>the</strong><br />
maximum entry <strong>of</strong> <strong>the</strong> wave<strong>for</strong>m.<br />
More importantly, <strong>the</strong> charge is not defined as charge above threshold, but as total integrated<br />
charge which includes baseline fluctuations: In <strong>the</strong> IC59 settings[51], <strong>the</strong> baseline<br />
is estimated by taking <strong>the</strong> average <strong>of</strong> <strong>the</strong> first three bins, <strong>and</strong> <strong>the</strong> resulting value is substracted<br />
from all bins if it does not exceed a hard-coded threshold; still, time-dependent<br />
129
sum <strong>of</strong> all bins above<br />
baseline + error<br />
C<br />
extrapolation <strong>of</strong><br />
maximum slope<br />
to baseline<br />
SPECIFIC PROBLEMS AND ANOMALIES<br />
width<br />
width:<br />
half <strong>of</strong> <strong>the</strong> number<br />
<strong>of</strong> bins above<br />
first pulse‘s half<br />
maximum height<br />
threshold<br />
extrapolation <strong>of</strong> first<br />
local maximum slope<br />
above threshold<br />
to baseline<br />
width:<br />
half <strong>of</strong> <strong>the</strong> number<br />
<strong>of</strong> bins above<br />
first pulse‘s half<br />
maximum height<br />
charge:<br />
sum <strong>of</strong> all bins<br />
above threshold<br />
extrapolation <strong>of</strong> first<br />
local maximum slope<br />
above threshold<br />
to baseline<br />
width:<br />
constant times<br />
charge / maximum<br />
charge:<br />
sum <strong>of</strong> all bins<br />
first pulse‘s half<br />
maximum height<br />
threshold<br />
threshold<br />
Figure C.1: Sketch illustrating<br />
parabola fit to maximum bin charge:<br />
<strong>the</strong> differences between<br />
extrapolation<br />
<strong>the</strong><br />
<strong>of</strong> firstdocumentation (left) <strong>and</strong> implementation<br />
(right) parabola <strong>of</strong> <strong>Feature</strong><strong>Extractor</strong>’s maximum<br />
local maximum secondslope<br />
single-pulse extraction algorithm.<br />
× pulse width<br />
above threshold<br />
Baseline detection has been omittedto <strong>for</strong> baseline reasons <strong>of</strong> clarity.<br />
x<br />
x<br />
width:<br />
proportional to<br />
charge / maximum<br />
charge:<br />
sum <strong>of</strong> all bins<br />
baseline shifts caused by droop or bad estimations due to <strong>the</strong> low statistics <strong>of</strong> <strong>the</strong> first<br />
extrapolation <strong>of</strong><br />
three binsmaximum can cause<br />
slope<br />
deviations <strong>of</strong> <strong>the</strong> total charge.<br />
to baseline<br />
This behaviour might be responsible <strong>for</strong> <strong>the</strong> overestimation <strong>of</strong> <strong>the</strong> charge <strong>of</strong> low pulses<br />
threshold<br />
which is described in section 8.3.1 <strong>and</strong> illustrated in figure 8.19, because this algorithm is<br />
width<br />
used in <strong>Feature</strong><strong>Extractor</strong>’s multi-pulse settings x to obtain <strong>the</strong> total charge that is used to<br />
rescale <strong>the</strong> pulses <strong>of</strong> its Bayesian Unfolding algorithm (see section 6.3.1).<br />
x<br />
x<br />
x<br />
C.3 Missing Simulation <strong>of</strong> <strong>the</strong> daq_baseline in DOMsimulator<br />
If available, DOMcalibrator uses <strong>the</strong> daq_baseline stored in <strong>the</strong> calibration data to calculate<br />
a wave<strong>for</strong>m’s average baseline. Up to its current release however, DOMsimulator<br />
does not simulate this baseline. This leads to an overcompensation by DOMcalibrator<br />
<strong>and</strong> <strong>the</strong>reby to a wrong baseline (e. g., figure 8.3), which in turn affects <strong>the</strong> extracted<br />
charge (figure C.2). Moreover, <strong>the</strong> pulse times are affected by <strong>the</strong> increased baseline, <strong>and</strong><br />
more fake pulses are extracted.<br />
After being in<strong>for</strong>med about an apparent mismatch between simulated <strong>and</strong> experimental<br />
data baselines, <strong>the</strong> DOMsimulator’s <strong>and</strong> DOMcalibrator’s maintainer Stijn Buitink<br />
quickly tracked <strong>and</strong> fixed <strong>the</strong> bug <strong>for</strong> <strong>the</strong> next release.<br />
C.4 Time Offset in SLCHit<strong>Extractor</strong><br />
In its initial release, SLCHit<strong>Extractor</strong> had a hard-coded time <strong>of</strong>fset calibration parameter<br />
c 2 (see section 6.4), which was used to align SLCHit<strong>Extractor</strong>’s pulse times with FE’s<br />
ATWD pulse times. It has been ab<strong>and</strong>oned later when <strong>the</strong> ATWD FADC time <strong>of</strong>fset was<br />
fixed (appendix C.1), because <strong>the</strong> times matched well without this <strong>of</strong>fset. However, this is<br />
only true <strong>for</strong> <strong>Feature</strong><strong>Extractor</strong>’s FE59 online-filtering settings, or more precisely only <strong>for</strong><br />
130
C.4 Time Offset in SLCHit<strong>Extractor</strong><br />
40000<br />
35000<br />
Old Toroid DOMs<br />
New Toroid DOMs<br />
total<br />
30000<br />
25000<br />
Old Toroid DOMs<br />
New Toroid DOMs<br />
total<br />
30000<br />
25000<br />
20000<br />
entries<br />
20000<br />
entries<br />
15000<br />
15000<br />
10000<br />
10000<br />
5000<br />
5000<br />
0<br />
1 0 1 2 3 4<br />
FADC charge / ATWD charge<br />
0<br />
1 0 1 2 3 4<br />
FADC charge / ATWD charge<br />
30000<br />
width: 0.32 0.31<br />
mean: 0.98 0.81<br />
30000<br />
width: 0.32 0.31<br />
mean: 0.92 0.81<br />
25000<br />
25000<br />
20000<br />
20000<br />
entries<br />
15000<br />
entries<br />
15000<br />
10000<br />
10000<br />
5000<br />
5000<br />
0<br />
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0<br />
q pulse1<br />
0<br />
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0<br />
q pulse1<br />
10 7<br />
10 6<br />
width: 0.71 2.95<br />
mean: 11.92 25.94<br />
10 7<br />
10 6<br />
width: 0.70 2.95<br />
mean: 11.78 25.94<br />
10 5<br />
10 5<br />
10 4<br />
10 4<br />
entries<br />
10 3<br />
entries<br />
10 3<br />
10 2<br />
10 2<br />
10 1<br />
10 1<br />
10 0<br />
10 0<br />
10 -1<br />
40 20 0 20 40<br />
t MC −t pulse /ns<br />
10 -1<br />
40 20 0 20 40<br />
t MC −t pulse /ns<br />
Figure C.2: Effect <strong>of</strong> <strong>the</strong> simulation <strong>of</strong> <strong>the</strong> daq_baseline on pulses extracted with NFE; custom<br />
simulation using 500 non-relativistic monopoles as light sources, <strong>the</strong> original<br />
scripts were provided by Thorsten Glüsenkamp.<br />
Left: data without baseline simulation; right: data with proper simulation.<br />
Top row: ratio between <strong>the</strong> total ATWD integrated charge <strong>and</strong> <strong>the</strong> corresponding<br />
FADC charge;<br />
middle row: <strong>the</strong> first pulse’s charge per wave<strong>for</strong>m <strong>for</strong> ATWD (blue area) <strong>and</strong><br />
FADC (green line);<br />
bottom row: <strong>the</strong> first pulse’s time per wave<strong>for</strong>m; <strong>the</strong> ATWD FADC time <strong>of</strong>fset<br />
was not applied.<br />
131
C<br />
SPECIFIC PROBLEMS AND ANOMALIES<br />
pulses extracted with PMTTransit = 2 (section 8.3.1), <strong>and</strong> fur<strong>the</strong>rmore it is coincidence:<br />
w<br />
There is no reason why <strong>the</strong> time t = t 0 + (i max − 1) · 25 ns − c imax−1 1 w max<br />
(section 6.4)<br />
w<br />
should match <strong>the</strong> true leading edge, because <strong>the</strong> term c imax−1 1 w max<br />
only compensates <strong>for</strong> <strong>the</strong><br />
correlation (slope) between <strong>the</strong> ratio <strong>of</strong> <strong>the</strong> first two bins <strong>and</strong> <strong>the</strong> pulse time (which is<br />
only well-defined <strong>for</strong> regular charge stamps). This time <strong>of</strong>fset can be seen in <strong>the</strong> sketch in<br />
figure 6.6, which is up to scale.<br />
The total time <strong>of</strong>fset <strong>of</strong> SLC pulses compared to FE’s ATWD pulses in IC59 <strong>of</strong>flineprocessing<br />
data is composed <strong>of</strong> <strong>the</strong> about 11 ns between FE ATWD pulses <strong>for</strong> PMTTransit<br />
= 2 <strong>and</strong> PMTTransit = -1, <strong>and</strong> <strong>the</strong> 15 ns caused by <strong>the</strong> original ATWD FADC time <strong>of</strong>fset,<br />
which was erroneously not applied to SLC charge stamps.<br />
132
Acknowledgements<br />
Without contributions <strong>of</strong> many people, writing this <strong>the</strong>sis would not have been possible.<br />
First I would like to thank Pr<strong>of</strong>. Dr. Christopher Wiebusch, who gave me <strong>the</strong> opportunity<br />
to work on this highly interesting project. He invited me into his workgroup <strong>and</strong><br />
provided <strong>new</strong> ideas as well as neverending enthusiasm.<br />
Special thanks go to Dr. David Boersma, whose great advice <strong>and</strong> invaluable experience<br />
especially in s<strong>of</strong>tware development <strong>and</strong> IceCube reconstruction made this project possible.<br />
Moreover, I would like to thank Pr<strong>of</strong>. Dr. Martin Erdmann <strong>for</strong> reviewing this <strong>the</strong>sis<br />
as second referee.<br />
Many thanks go to Anne Schukraft, Sebastian Euler, Matthias Schunck, Thomas<br />
Krings, <strong>and</strong> Jan-Patrick Hülß <strong>for</strong> <strong>the</strong>ir extensive pro<strong>of</strong> reading, <strong>and</strong> to <strong>the</strong>m <strong>and</strong> <strong>the</strong><br />
rest <strong>of</strong> <strong>the</strong> IceCube Aachen workgroup <strong>for</strong> many interesting discussions, a great working<br />
atmosphere <strong>and</strong> generally a good time. This particularly includes <strong>the</strong> Kinderzimmer in<br />
both its <strong>for</strong>mer <strong>and</strong> its current lineup.<br />
Special credits go to Thomas Krings, who shared his initial L A TEX templates to start<br />
a common framework <strong>for</strong> future <strong>the</strong>ses in Aachen. I would also like to thank Matthias<br />
Schunk <strong>for</strong> <strong>the</strong> simulation <strong>of</strong> specialized Monte-Carlo datasets needed <strong>for</strong> tests <strong>and</strong> calibration,<br />
<strong>and</strong> Thorsten Glüsenkamp <strong>for</strong> providing me with his simulation scripts.<br />
Fur<strong>the</strong>rmore, many thanks go to <strong>the</strong> rest <strong>of</strong> <strong>the</strong> IceCube Collaboration, in particular<br />
to Dmitry Chirkin, Christopher Wendt, Alex Olivas, Stijn Buitink, Andreas Groß, Cécile<br />
Portello-Roucelle, Dennis Diederix, Markus Voge, <strong>and</strong> Fabian Clevermann <strong>for</strong> many interesting<br />
discussions, <strong>and</strong> also to Thorsten Stezelberger from LBNL <strong>for</strong> his investigation<br />
on <strong>the</strong> SLC firmware.<br />
Finally I want to express my gratitude towards my family <strong>and</strong> my friends <strong>for</strong> <strong>the</strong>ir<br />
enduring support!<br />
I
Erklärung<br />
Hiermit erkläre ich, dass ich die vorliegende Arbeit selbständig verfasst und keine <strong>and</strong>eren<br />
als die angegebenen Quellen und Hilfsmittel verwendet habe.<br />
Aachen, den 02. März 2010<br />
Declaration<br />
I hereby certify that this document has been composed by myself, <strong>and</strong> describes my own<br />
work, unless o<strong>the</strong>rwise acknowledged in <strong>the</strong> text.<br />
Aachen, March <strong>the</strong> 2 nd , 2010<br />
III
References<br />
[1] Amsler, C. et al.: “Review <strong>of</strong> Particle Physics – Astrophysics <strong>and</strong> Cosmology”.<br />
Physics Letters B, vol. 667(1-5), pp. 212 – 260, 2008. ISSN 0370-2693. doi:<br />
DOI:10.1016/j.physletb.2008.07.028. Review <strong>of</strong> Particle Physics.<br />
URL http://dx.doi.org/10.1016/j.physletb.2008.07.028<br />
[2] Simpson, J. A.: “Elemental <strong>and</strong> Isotopic Composition <strong>of</strong> <strong>the</strong> Galactic Cosmic Rays”.<br />
Annual Review <strong>of</strong> Nuclear <strong>and</strong> Particle Science, vol. 33(1), pp. 323–382, 1983.<br />
doi:10.1146/annurev.ns.33.120183.001543.<br />
URL http://arjournals.annualreviews.org/doi/abs/10.1146/annurev.ns.<br />
33.120183.001543<br />
[3] Gaisser, T. K.: Cosmic Rays <strong>and</strong> Particle Physics. Cambridge Univ. Press, Cambridge,<br />
1990.<br />
[4] Amenomori, M. et al.: “The cosmic-ray energy spectrum around <strong>the</strong> knee measured<br />
by <strong>the</strong> Tibet-III air-shower array”. Nuclear Physics B - Proceedings Supplements,<br />
vol. 175-176, pp. 318 – 321, 2008. ISSN 0920-5632. doi:DOI:10.1016/j.nuclphysbps.<br />
2007.11.021. Proceedings <strong>of</strong> <strong>the</strong> XIV International Symposium on Very High Energy<br />
Cosmic Ray Interactions.<br />
URL http://www.sciencedirect.com/science/article/B6TVD-4RJ49K5-26/2/<br />
6859b5dcee40103ac6be83912a4b9f55<br />
[5] Nagano, M., Teshima, M., Matsubara, Y., Dai, H. Y., Hara, T., Hayashida, N.,<br />
Honda, M., Ohoka, H., <strong>and</strong> Yoshida, S.: “Energy spectrum <strong>of</strong> primary cosmic rays<br />
above 10 17 eV determined from extensive air shower experiments at Akeno”. Journal<br />
<strong>of</strong> Physics G: Nuclear <strong>and</strong> Particle Physics, vol. 18(2), pp. 423–442, 1992.<br />
URL http://stacks.iop.org/0954-3899/18/423<br />
[6] Cherry, M. L.: “An abrupt slowdown <strong>for</strong> particles on <strong>the</strong> fast track”. Physics, vol. 1,<br />
9, Aug 2008. doi:10.1103/Physics.1.9.<br />
URL http://physics.aps.org/articles/v1/9<br />
[7] Greisen, K.: “End to <strong>the</strong> Cosmic-Ray Spectrum?” Phys. Rev. Lett., vol. 16(17), pp.<br />
748–750, Apr 1966. doi:10.1103/PhysRevLett.16.748.<br />
URL http://prl.aps.org/abstract/PRL/v16/i17/p748_1<br />
[8] Drees, M.: “The Top-Down Interpretation <strong>of</strong> Ultra-High Energy Cosmic Rays”. Journal<br />
<strong>of</strong> <strong>the</strong> Physical Society <strong>of</strong> Japan, vol. 77SB(Supplement B), pp. 16–18, 2008.<br />
doi:10.1143/JPSJS.77SB.16.<br />
URL http://jpsj.ipap.jp/link?JPSJS/77SB/16/<br />
V
[9] Kachelrieß, M. <strong>and</strong> Semikoz, D.: “Reconciling <strong>the</strong> ultra-high energy cosmic ray<br />
spectrum with Fermi shock acceleration”. Physics Letters B, vol. 634(2-3), pp. 143 –<br />
147, 2006. ISSN 0370-2693. doi:DOI:10.1016/j.physletb.2006.01.009.<br />
URL http://www.sciencedirect.com/science/article/B6TVN-4J5D6G3-8/2/<br />
f2184636c2d405bc6da9f940becf5269<br />
[10] Dolag, K., Grasso, D., Springel, V., <strong>and</strong> Tkachev, I.: “Mapping deflections <strong>of</strong> extragalactic<br />
ultrahigh-energy cosmic rays in magnetohydrodynamic simulations <strong>of</strong> <strong>the</strong><br />
local universe”. JETP Letters, vol. 79(12), pp. 583–587, June 2004. ISSN 0021-3640<br />
(Print) 1090-6487 (Online). doi:10.1134/1.1790011.<br />
URL http://www.springerlink.com/content/wn2117762165xg21/<br />
[11] Rothman, T. <strong>and</strong> Boughn, S.: “Can gravitons be detected?” Found. Phys., vol. 36,<br />
pp. 1801–1825, 2006. doi:10.1007/s10701-006-9081-9.<br />
URL http://arxiv.org/abs/gr-qc/0601043<br />
[12] Collaboration, T. L. S. <strong>and</strong> Collaboration, T. V.: “Searches <strong>for</strong> gravitational waves<br />
from known pulsars with S5 LIGO data”, September 2009.<br />
URL http://arxiv.org/abs/0909.3583v1<br />
[13] Brocato, E., Castellani, V., Degl’Innocenti, S., Fiorentini, G., <strong>and</strong> Raimondo, G.:<br />
“Stars as galactic neutrino sources”. Astron. Astrophys., vol. 333, p. 910, 1998.<br />
URL http://arxiv.org/abs/astro-ph/9711269<br />
[14] Woosley, S. E., Heger, A., <strong>and</strong> Weaver, T. A.: “The evolution <strong>and</strong> explosion <strong>of</strong><br />
massive stars”. Rev. Mod. Phys., vol. 74(4), pp. 1015–1071, Nov 2002. doi:10.1103/<br />
RevModPhys.74.1015.<br />
URL http://rmp.aps.org/abstract/RMP/v74/i4/p1015_1<br />
[15] Nicolas Chamel, P. H.: “Physics <strong>of</strong> Neutron Star Crusts”. Living Reviews in Relativity,<br />
vol. 11(10), 2008.<br />
URL http://www.livingreviews.org/lrr-2008-10<br />
[16] Hansen, B.: “The astrophysics <strong>of</strong> cool white dwarfs”. Physics Reports, vol. 399(1),<br />
pp. 1 – 70, 2004. ISSN 0370-1573. doi:DOI:10.1016/j.physrep.2004.07.001.<br />
URL http://www.sciencedirect.com/science/article/B6TVP-4D3B39C-1/2/<br />
8143f54436eb55b9ee72bf541e93349e<br />
[17] Amsler, C. et al.: “Review <strong>of</strong> Particle Physics”. Phys. Lett., vol. B667, p. 1, 2008.<br />
doi:10.1016/j.physletb.2008.07.018. And 2009 partial update <strong>for</strong> <strong>the</strong> 2010 edition.<br />
URL http://pdglive.lbl.gov/listings1.brl?quickin=Y<br />
[18] Pasquali, L., Reno, M. H., <strong>and</strong> Sarcevic, I.: “Secondary decays in atmospheric charm<br />
contributions to <strong>the</strong> flux <strong>of</strong> muons <strong>and</strong> muon neutrinos”. Astroparticle Physics,<br />
vol. 9(3), pp. 193 – 202, 1998. ISSN 0927-6505. doi:DOI:10.1016/S0927-6505(98)<br />
VI
00019-X.<br />
URL http://dx.doi.org/10.1016/S0927-6505(98)00019-X<br />
[19] Thunman, M., Ingelman, G., <strong>and</strong> Gondolo, P.: “Charm production <strong>and</strong> high energy<br />
atmospheric muon <strong>and</strong> neutrino fluxes”. Astroparticle Physics, vol. 5(3-4), pp. 309 –<br />
332, 1996. ISSN 0927-6505. doi:DOI:10.1016/0927-6505(96)00033-3.<br />
URL http://www.sciencedirect.com/science/article/B6TJ1-3VPSFK6-C/2/<br />
d99a97b5b0b4a7e077c1dbf6ebe646e8<br />
[20] Ahrens, J. et al.: “IceCube Preliminary <strong>Design</strong> Document”. Tech. Rep., The IceCube<br />
Collaboration, Oct 2001.<br />
URL http://www.icecube.wisc.edu/science/publications/pdd/pddwhole.php<br />
[21] Schonert, S., Gaisser, T. K., Resconi, E., <strong>and</strong> Schulz, O.: “Vetoing atmospheric<br />
neutrinos in a high energy neutrino telescope”. Physical Review D, vol. 79, p. 043009,<br />
2009.<br />
URL doi:10.1103/PhysRevD.79.043009<br />
[22] G<strong>and</strong>hi, R., Quigg, C., Reno, M. H., <strong>and</strong> Sarcevic, I.: “Neutrino interactions at<br />
ultrahigh energies”. Phys. Rev. D, vol. 58(9), p. 093009, Sep 1998. doi:10.1103/<br />
PhysRevD.58.093009.<br />
URL http://prd.aps.org/abstract/PRD/v58/i9/e093009<br />
[23] Reno, M. H.: “High energy neutrino cross sections”. Nuclear Physics B - Proceedings<br />
Supplements, vol. 143, p. 407, 2005. doi:doi:10.1016/j.nuclphysbps.2005.01.137.<br />
URL http://arxiv.org/abs/hep-ph/0410109<br />
[24] Neunhöffer, T.: Die Entwicklung eines neuen Verfahrens zur Suche nach kosmischen<br />
Neutrino-Punktquellen mit dem AMANDA-Neutrinoteleskop. Shaker, 2004.<br />
URL http://icecube.berkeley.edu/manuscripts/<br />
[25] Escribano, R., Frère, J. M., Monderen, D., <strong>and</strong> Elewyck, V. V.: “Insights on neutrino<br />
lensing”. Physics Letters B, vol. 512(1-2), pp. 8 – 17, 2001. ISSN 0370-2693.<br />
doi:DOI:10.1016/S0370-2693(01)00686-4.<br />
URL http://www.sciencedirect.com/science/article/B6TVN-43CTFPP-3/2/<br />
2777c795ad8bb667d593d14700f378a9<br />
[26] Illana, J. I., Masip, M., <strong>and</strong> Meloni, D.: “Probing TeV gravity at neutrino telescopes”.<br />
In Proc. <strong>of</strong> <strong>the</strong> First Workshop on Exotic Physics with Neutrino Telescopes (edited<br />
by de los Heros, C.). Uppsala, Sep 2006.<br />
URL http://arxiv.org/abs/hep-ph/0612305<br />
[27] Alvarez-Muniz, J. <strong>and</strong> Zas, E.: “Calculations <strong>of</strong> radio pulses from High Energy Showers”.<br />
AIP CONF.PROC., vol. 579, p. 117, 2001. doi:doi:10.1063/1.1398165.<br />
URL http://arxiv.org/abs/astro-ph/0103369<br />
VII
[28] Voigt, B.: Sensitivity <strong>of</strong> <strong>the</strong> IceCube detector <strong>for</strong> ultra-high energy electronneutrino<br />
events. Ph.D. <strong>the</strong>sis, Humboldt-Universität zu Berlin, Ma<strong>the</strong>matisch-<br />
Naturwissenschaftliche Fakultät I, Nov 2008.<br />
URL http://edoc.hu-berlin.de/docviews/abstract.php?id=29421<br />
[29] Besson, D. <strong>and</strong> The Rice Collaboration: “Modeling <strong>of</strong> high-energy electromagnetic<br />
showers in ice”. In International Cosmic Ray Conference, vol. 3 <strong>of</strong> International<br />
Cosmic Ray Conference, pp. 1179–+, 2001.<br />
URL http://adsabs.harvard.edu/abs/2001ICRC....3.1179B<br />
[30] Amsler, C. et al.: “Review <strong>of</strong> Particle Physics – Experimental Methods <strong>and</strong> Colliders”.<br />
Physics Letters B, vol. 667(1-5), pp. 261 – 315, 2008. ISSN 0370-2693. doi:DOI:<br />
10.1016/j.physletb.2008.07.029. Review <strong>of</strong> Particle Physics.<br />
URL http://dx.doi.org/10.1016/j.physletb.2008.07.029<br />
[31] Albuquerque, I., Burdman, G., <strong>and</strong> Chacko, Z.: “Neutrino Telescopes as a Direct<br />
Probe <strong>of</strong> Supersymmetry Breaking”. Phys. Rev. Lett., vol. 92(22), p. 221802, Jun<br />
2004. doi:10.1103/PhysRevLett.92.221802. And corresponding presentation Looking<br />
<strong>for</strong> SUSY in <strong>the</strong> Ice at TeV Particle Astrophysics converence.<br />
URL http://www-astro-<strong>the</strong>ory.fnal.gov/Conferences/TeV/Albuquerque.pdf<br />
[32] Katz, U. et al.: “KM3NeT – Conceptual <strong>Design</strong> Report <strong>for</strong> a Deep Sea Research Infrastructure<br />
Incorporating a Very Large Volume Neutrino Telescope in <strong>the</strong> Mediterranean<br />
Sea”. Tech. Rep., The KM3NeT Consortium, April 2008.<br />
URL http://www.km3net.org/CDR/CDR-KM3NeT.pdf<br />
[33] Halzen, F.: “Status <strong>of</strong> Neutrino Astronomy: The Quest <strong>for</strong> Kilometer-Scale Instruments”.<br />
COMMENTS NUCL.PART.PHYS., vol. 22, p. 155, 1997.<br />
URL http://lanl.arxiv.org/abs/astro-ph/9701029<br />
[34] Montaruli, T.: “Neutrino Astronomy in <strong>the</strong> Ice”. Nuclear Physics B – Proceedings<br />
Supplements, vol. 188, pp. 239–244, March 2009. Proceedings <strong>of</strong> <strong>the</strong> Neutrino<br />
Oscillation Workshop.<br />
URL http://arxiv.org/abs/0901.2664<br />
[35] Woschnagg, K.: “IC79 Geometry Figure”. IceCube Internal Wiki, 2010.<br />
URL http://wiki.icecube.wisc.edu/index.php/Geometry_figures<br />
[36] Vevea, D.: “Array-PublicationDL”. IceCube Internal Gallery, 2009.<br />
URL http://gallery.icecube.wisc.edu/internal/v/graphics/sketchup<br />
[37] Ruzybayev, B., Hussain, S., Xu, C., Gaisser, T., <strong>and</strong> <strong>the</strong> IceCube Collaboration:<br />
“Small air showers in IceTop”. In Proceedings <strong>of</strong> <strong>the</strong> 31 s t ICRC. Łódź, July 2009.<br />
URL http://www.srl.utu.fi/AuxDOC/kocharov/ICRC2009/pdf/icrc0737.pdf<br />
VIII
[38] The IceCube Collaboration: “Live at <strong>the</strong> South Pole”. IceCube Public Website, June<br />
2009.<br />
URL http://www.icecube.wisc.edu/info/life.php<br />
[39] Ackermann, M. et al.: “Optical properties <strong>of</strong> deep glacial ice at <strong>the</strong> South Pole”.<br />
Journal <strong>of</strong> Geophysical Research - Atmospheres, vol. 111(D13), pp. D13203+, July<br />
2006. ISSN 0148-0227. doi:10.1029/2005JD006687.<br />
URL http://dx.doi.org/10.1029/2005JD006687<br />
[40] The IceCube Collaboration: “Anatomy <strong>of</strong> a DOM”. IceCube Public Gallery, Nov<br />
2006.<br />
URL http://gallery.icecube.wisc.edu/external/4-cons-doms/DOM-Picture.<br />
png.html<br />
[41] Portello-Roucelle, C.: “DOMCalibrator”. IceCube Virtual Meeting 2009, July 2009.<br />
URL http://wiki.icecube.wisc.edu/index.php/Agenda_Day_1%2C_Session_2<br />
[42] Wendth, C.: “Droop Correction – Dual τ Model”. IceCube Collaboration Meeting,<br />
Oct 2006.<br />
URL https://docushare.icecube.wisc.edu/dsweb/Get/Document-30244/<br />
Droop-dual-tau-Zeu<strong>the</strong>n2006_wendt.pdf<br />
[43] Wiebusch, C. <strong>and</strong> <strong>the</strong> IceCube Collaboration: “Physics Capabilities <strong>of</strong> <strong>the</strong> IceCube<br />
DeepCore Detector”. In Proceedings <strong>of</strong> <strong>the</strong> 31 s t ICRC. Łódź, July 2009.<br />
URL http://arxiv.org/PS_cache/arxiv/pdf/0907/0907.2263v1.pdf<br />
[44] Hamamatsu: Photomultiplier Tube R7081-02 Data Sheet, Nov 2003.<br />
URL https://docushare.icecube.wisc.edu/dsweb/Get/Document-6637/<br />
R7081-02%20data%20sheet.pdf<br />
[45] The IceCube Collaboration: “Prepulse Data”. IceCube Internal Wiki, May 2007.<br />
Based on Chris Wendt’s measurements.<br />
URL http://wiki.icecube.wisc.edu/index.php/Prepulse_Data<br />
[46] The IceCube Collaboration: “The IceCube data acquisition system: Signal capture,<br />
digitization, <strong>and</strong> timestamping”. Nuclear Instruments <strong>and</strong> Methods in Physics<br />
Research Section A, vol. 601(3), pp. 294 – 316, 2009. ISSN 0168-9002. doi:<br />
DOI:10.1016/j.nima.2009.01.001. Revision 1.3.<br />
URL http://www.sciencedirect.com/science/article/B6TJM-4VBMNCJ-4/2/<br />
282b1e53516b6eab577fa652971a8fd9<br />
[47] Wendt, C.: “DOM SPE Wave<strong>for</strong>m Shape”. IceCube Internal Wiki, 2009.<br />
URL http://icecube.wisc.edu/~chwendt/dom-spe-wave<strong>for</strong>m-shape/<br />
[48] Stezelberger, T.: “private conversation”, Aug 2009. Lawrence Berkeley National<br />
Laboratory.<br />
IX
[49] The IceCube Collaboration: “IceTray”, Feb 2010.<br />
URL http://s<strong>of</strong>tware.icecube.wisc.edu/<strong>of</strong>fline-s<strong>of</strong>tware.trunk/projects/<br />
icetray/index.html<br />
[50] Chirkin, D., Klein, S. et al.: “<strong>Feature</strong><strong>Extractor</strong> V02-03-00 Source Code”. IceCube<br />
SVN Source Code Repository, January 2010.<br />
URL http://code.icecube.wisc.edu/projects/icecube/browser/projects/<br />
<strong>Feature</strong><strong>Extractor</strong>/releases/V02-03-00<br />
[51] The IceCube Collaboration: “St<strong>and</strong>ard Processing Scripts, V10-01-00”. IceCube<br />
SVN Source Code Repository, Jan 2010.<br />
URL<br />
http://code.icecube.wisc.edu/projects/icecube/browser/<br />
meta-projects/std-processing/releases/10-01-00/scripts/IC59/level1_<br />
CalibrateAndExtractPulses.py<br />
[52] ROOT Development Team: “ROOT”. http://root.cern.ch/drupal/, 2010.<br />
[53] Panknin, S.: “The <strong>Feature</strong> <strong>Extractor</strong>”. IceTray Seminar, Oct 2008.<br />
URL<br />
http://nuastro-zeu<strong>the</strong>n.desy.de/e13/e63159/e27/e689/e693/<br />
infoboxContent722/fe.pdf<br />
[54] Chirkin, D. <strong>and</strong> Wendth, C.: “Pulse<strong>Extractor</strong> Source Code”. IceCube SVN Source<br />
Code Repository, January 2010.<br />
URL http://code.icecube.wisc.edu/projects/icecube/browser/s<strong>and</strong>box/<br />
Pulse<strong>Extractor</strong><br />
[55] Groß, A.: “SLCHit<strong>Extractor</strong> V00-01-00 Source Code”. IceCube SVN Source Code<br />
Repository, January 2010.<br />
URL http://code.icecube.wisc.edu/projects/icecube/browser/projects/<br />
SLCHit<strong>Extractor</strong>/releases/V00-01-00<br />
[56] Boersma, D. J.: “Gulliver”. IceCube Internal Wiki, Oct 2009.<br />
URL http://wiki.icecube.wisc.edu/index.php/Gulliver<br />
[57] D’Agostini, G.: “A multidimensional unfolding method based on Bayes’ <strong>the</strong>orem”.<br />
Nuclear Instruments <strong>and</strong> Methods in Physics Research Section A, vol. 362(2-3), pp.<br />
487 – 498, Mar 1995. ISSN 0168-9002. doi:DOI:10.1016/0168-9002(95)00274-X.<br />
URL http://www.sciencedirect.com/science/article/B6TJM-3YRNX0H-5K/2/<br />
3e3a92555a7955c7f4ab989fa99baef7<br />
[58] The IceCube Collaboration: “IC59 NuMu E −1 dataset 2595”. Simulation Production,<br />
Sep 2009.<br />
URL http://internal.icecube.wisc.edu/simulation/dataset/2595<br />
[59] The IceCube Collaboration: “IC59 NuMu E −1 dataset 3071”. Simulation Production,<br />
Feb 2010.<br />
URL http://internal.icecube.wisc.edu/simulation/dataset/3071<br />
X
[60] Voge, M.: “IC59 Wave<strong>for</strong>ms”, Dec 2009. https://docushare.icecube.wisc.edu/<br />
dsweb/Get/Document-52517/MarkusVoge_09-12-15.pdf.<br />
[61] Merck, M.: “DOMCalibrator Problems”. IceCube Internal Wiki, Jul 2009.<br />
URL http://wiki.icecube.wisc.edu/index.php/DOMCalibrator_Problems<br />
[62] Groß, A.: “IC59 L2 processing”. IceCube Internal Wiki, Oct 2009.<br />
URL http://wiki.icecube.wisc.edu/index.php/IC59_L2_processing<br />
[63] Aguilar, J.: “IC59 L2 processing/FADC usage”. IceCube Internal Wiki, Dec 2009.<br />
URL http://wiki.icecube.wisc.edu/index.php/IC59_L2_processing/FADC_<br />
usage<br />
[64] Groß, A.: “SLCHitExtraction”. IceCube Internal Wiki, Apr 2009.<br />
URL http://wiki.icecube.wisc.edu/index.php/SLCHitExtraction<br />
XI