Update on Rolling Cascade Search

1
Update on Rolling Cascade Search

• Brennan Hughey
• UW-Madison
• 3-26-04

2
Overview of Rolling Cascade Search
Rolling search scans entire year (2001) of data
for a significant clumping of events
(inconsistent with poissonian background) Utilitz
es Cascade channel (and high-energy cascade-like
muons), so is not directionally dependent and has
effective volumes greater than the volume of the
detector at high energies Broken power law
spectrum for signal Monte Carlo and 15-second
bins selected to be consistent with expectations
for Gamma Ray Bursts
3
Rolling Search
Time
4
Data Reduction
Step 1 High Energy Filter (1 of data
remains) Step 2 Cut on ratio of
Ndird/Nhits Step 3 Support Vector Machine
output cut based on 8 variable input and using
SVMlight (Replaces PawMLP neural network cut)
Goal of data reduction obtain greatest chance
of observing a signal, Signal is defined as a
number of events in a bin such that there is a
less than 5 chance of that number of events
ocurring in 300 days given poissonian background
5
Ndird shows good separation, but there is
a significant tail of signal events with a large
number of total hits
_____ real data (background) _ _ _ _TEA
Monte Carlo (cascade signal) ............dCorsik
a (background MC)
6
Dividing by nhits removes the high energy
tail. Initial hard cut taken at .18 to reduce
data but stay well away from signal, since MC and
data show some significant disagreement
_____ real data (background) _ _ _ _TEA
Monte Carlo (cascade signal) ............dCorsik
a (background MC)
7
Cut Variables in Support Vector Machine
_____ real data (background) _ _ _ _TEA
Monte Carlo (cascade signal) ............dCorsik
a (background MC) _._._._.E-1 Nusim (high
energy muons)
8
Cut Variables in Support Vector Machine
_____ real data (background) _ _ _ _TEA
Monte Carlo (cascade signal) ............dCorsik
a (background MC) _._._._.E-1 Nusim (high
energy muons)
9
Cut Variables in Support Vector Machine
_____ real data (background) _ _ _ _TEA
Monte Carlo (cascade signal) ............dCorsik
a (background MC) _._._._.E-1 Nusim (high
energy muons)
10
Cut Variables in Support Vector Machine
_____ real data (background) _ _ _ _TEA
Monte Carlo (cascade signal) ............dCorsik
a (background MC) _._._._.E-1 Nusim (high
energy muons)
11
Cut Variables in Support Vector Machine
_____ real data (background) _ _ _ _TEA
Monte Carlo (cascade signal) ............dCorsik
a (background MC) _._._._.E-1 Nusim (high
energy muons)
12
Cut Variables in Support Vector Machine
_____ real data (background) _ _ _ _TEA
Monte Carlo (cascade signal) ............dCorsik
a (background MC) _._._._.E-1 Nusim (high
energy muons)
13
_____ real data (background) _ _ _ _TEA
Monte Carlo (cascade signal) ............dCorsik
a (background MC)
Positive values are events the Support Vector
Machine identified as signal. Plot shows test
data, not original training data.
Cost factor (favoring identification as
background or signal) and constant in
mathematical kernel function can be varied
14
Average remaining background events around 11 per
day total acceptance rate on the order of
10-6 Retention of signal MC is around 55.
Assuming Poissonian background, there is less
than a 5 chance of seeing 3 events in 15 seconds
during 300 days without signal at this background
level Assuming Poissonian distribution of signal
events, odds of seeing 3 signal events in a 15
second window at this cut level are better than
is possible for any situation requiring 4 or more
events for a significant detection.
15
One can obtain nearly identical results from the
Paw neural network by cutting at .98, but this
requires cutting into the sharply spiked
signal peak, which means considerable systematic
uncertainty
_____ real data (background)
_ _ _ _TEA Monte Carlo (cascade signal)
............dCorsika (background MC)
16
Cascade Effective Volume
physical volume of detector
17
Neutrino Effective Area
Pre-filter events circles Cascade
channel squares Muon channel
10
10
Muons blocked by Earth increasingly at higher
energies (Earth shadowing effect) Also, range of
muon does not increase linearly due to
Bremstrahlung Cascade range does continue to
increase change in slope can be attributed to
interaction cross sections
10
10
Effective Area (cm2)
10
10
10
10
10
10
10
10
10
10
10
10
Energy (GeV)
18
2 event coincidence for signal detection?
Possibility 1 further background rejection -
requires approximately 1 event every 5 days - too
much training data needed - reduces signal
retention Possibility 2 shortening the time
window of search - on the order of 15 ms at
current cut level - difficult to demonstrate that
data is poissonian at this level - deadtime
becomes a significant factor......
19
Deadtime
time (milliseconds)
time (milliseconds)
Nchgt160 (high energy events)
All Nch
Number of events vs. Dt (in milliseconds) between
event and next event for all events and events
with Nchannel gt 160