Post on 10-Jul-2015
S.Klimenko, G060091-00-Z, March 20, 2006, LSC meeting
First results from the likelihood pipelineS.Klimenko (UF), I.Yakushin (LLO), A.Mercer (UF),G.Mitselmakher (UF)
● network WaveBurst pipeline● LIGO-Virgo project 1b results● Coherent energy, x-correlation● S4 results● Summary&Plans
S.Klimenko, G060091-00-Z, March 20, 2006, LSC meeting
Why one more method for burst detection?
● The network WaveBurst pipeline is based on
likelihood method (PRD 72, 122002, 2005)
a coherent method for burst detection
allows reconstruction of GW waveforms and
source coordinates
works for arbitrary alignment of detectors
S.Klimenko, G060091-00-Z, March 20, 2006, LSC meeting
Search algorithm● works in time-frequency domain● likelihood formalism independent for each TF
sample greatly simplifies calculation of network response
● likelihood time-frequency maps● time shifts use 2D TF time delay filter
pick any pixel with fixed (f,t) and calculate pixel amplitudes
a2(t,f,δ) and a3(t,f,δ) time-shifting planes 2 & 3 by delay δ
The amplitudes a1, a2(t,f,δ) and a3(t,f,δ) are used to calculate
pixel likelihood L(t,f,δ)
f
t
f
t
S.Klimenko, G060091-00-Z, March 20, 2006, LSC meeting
L1: 1.1e-22
V1: 1.3e-22
H1: 1.5e-22
likelihood
Time-frequency maps
sg250Hz, τ=0.02sec, θ=20, φ=150
apply threshold on likelihood and reconstruct clusters
S.Klimenko, G060091-00-Z, March 20, 2006, LSC meeting
Coherent WaveBurst
● Similar approach as for incoherent WaveBurst● Uses most of existing WaveBurst functionality
data conditioning:wavelet transform,
(rank statistics),pixel selection
channel 1
data conditioning:wavelet transform,
(rank statistics),pixel selection
channel 2
Likelihood TF map
event generation
data conditioning:wavelet transform,
(rank statistics),pixel selection
channel 3,…
S.Klimenko, G060091-00-Z, March 20, 2006, LSC meeting
Likelihood sky maps
● xxx
sg250Hz, τ=0.02sec, θ=20, φ=150, L1/H1/V1
S.Klimenko, G060091-00-Z, March 20, 2006, LSC meeting
LIGO-Virgo simulated noise ● Likelihood of triggers on the output of nWB
different types of injections (SG,GAUSS,SN) background -- 101 time shifts (T=8726400sec)
2L is the totaldetectedenergy
S.Klimenko, G060091-00-Z, March 20, 2006, LSC meeting
Periodic table of burst algorithms
62 / 3277 / 3946 / 1876 / 3559 / 2345 / 20QT
34 / 1319 / 319 / 760 / 1739 / 1033 / 9PF
11 / 238 / 614 / 452 / 1336 / 926 / 6KW
----55 / 2376 / 3067 / 1956 / 14PC
32 / 1360 / 1932 / 1366 / 2343 / 1252 / 16EGC
1 / 020 / 317 / 855 / 1743 / 1328 / 8MF
6 / 225 / 328 / 1260 / 2047 / 1433 / 9ALF
738262--6341WBN
sg820or2 / 3pl
sg235or2 / 3pl
gauss4or2 / 3pl
gauss1or2 /3pl
A2B4G1or2 / 3pl
A1B2G1or2 / 3pl
false alarm rate 1-3 µHz3pl – triple coincidenceor2 – OR of detector pairs
Fraction (%) of detected events
S.Klimenko, G060091-00-Z, March 20, 2006, LSC meeting
maximum likelihood statistic
● Likelihood for Gaussian noise with variance σ2 k and GW waveform u:
xk[i] – detector output, Fk – antenna patterns
● Events reconstructed by the pipeline are coincident in time by construction what is definition of a coincidence?
[ ]( )[ ]∑∑ −−=i k
kkkk
iixixL 222
][][2
1 ξσ
xkxkk FhFh += ++ξdetector response -
NEL −=2total
energynoise (null)
energydetected (signal)
energy
S.Klimenko, G060091-00-Z, March 20, 2006, LSC meeting
real world analysis
● In ideal world where detector noise is always Gaussian: likelihood statistic is sufficient.
● In the real world where data is always contaminated with glitches: - the signal consistency statistics are required
● Real world analysis strategy
use L
as detection statistic
usepower
null streamcoherent energy
as signalconsistency
statistics
S.Klimenko, G060091-00-Z, March 20, 2006, LSC meeting
Likelihood distribution
Ligo-Virgo noise S4 data
S.Klimenko, G060091-00-Z, March 20, 2006, LSC meeting
Detection performance as ETG
3.18
3.71
849
1.65
2.04
361
2.15
2.49
553
1.12
1.29
235 105315310070sgQ9
4.010.941.232.70nWB
5.001.131.324.60WB
tuned to FA rate 20µHz
by setting threshold
on power ineach detector
S.Klimenko, G060091-00-Z, March 20, 2006, LSC meeting
Coherent energy● Likelihood is a quadratic form:
● xi – whitened data streams● Cij depend on antenna patterns and detector
sensitivity. For constraint likelihood:
g – network sensitivity
jijiji
ijji EECxxL ≠= +== ∑,
2
−=
+= ××++
≠×+
ji
ji
ji
jiji
i
i
i
iii
FFFF
gC
FF
gC
σσσσσσ 2
1 ,
2
12
2
2
2
incoherent coherent
S.Klimenko, G060091-00-Z, March 20, 2006, LSC meeting
X-correlation for misaligned detectors
● S-statistic
● Cauchy-statistic
network global x-correlation coefficient
● Similar to r-statistic the KS test can be applied
coherentull
coherent
ii
ji ij
net EN
E
EE
EC
+=
−=
∑∑≠
22
ji
ji
jjii
ij
xx
xx
EE
Er →=
22
2
ji
ji
jjii
ij
xx
xx
EEE
Ec
+→
−−=
arbitrary detectors aligned detectors
S.Klimenko, G060091-00-Z, March 20, 2006, LSC meeting
Network x-correlation ● powerful signal consistency test different from the
NULL stream
SG injections
S4 background triggers
S.Klimenko, G060091-00-Z, March 20, 2006, LSC meeting
End-to-End pipeline
4.6
3.9
849
2.2
2.1
361
2.9
2.6
553
1.5
1.3
235 105315310070sgQ9
6.11.21.53.5WBN
5.41.21.44.7WB+R
● In addition to excess power cuts apply selection cuts based on the likelihood x-correlation termsworks for arbitrary detector alignment
● False alarm (100 time lags) 0.2 mkHz ● Results are very preliminary
no DQ flags were appliedpost-processing election cuts are not tuned
S.Klimenko, G060091-00-Z, March 20, 2006, LSC meeting
Pipeline performance
● Search over the entire sky with good angular resolution can be computationally very intensive
● nWB pipeline performance for all sky search with 1o resolution (65000 sky locations) and 101 time lags cit 2.2GHz 64 bit obteron – 1. 0 sec/sec Llo 3.4GHz 32 bit Xeon - 1.6 sec/sec
● nWB performance is comparable to the incoherent WaveBurst pipeline typical run time on LLO cluster for S4 data is 1 day
S.Klimenko, G060091-00-Z, March 20, 2006, LSC meeting
Summary
● network WaveBurst pipeline (nWB) based on likelihood analysis is operational.LIGO-Virgo project 1b data: nWB looks good in “ideal
world” S4 data: results consistent with the standard WB+R
pipelineExcellent computational performance
● x-correlation coefficients can be defined for arbitrary networks which offers a waveform consistency test complementary to Null stream
S.Klimenko, G060091-00-Z, March 20, 2006, LSC meeting
Plans
● Finalize tuning of selection cuts & S4 data analysis● study coordinate and waveform reconstruction● try algorithm on LIGO-GEO and LIGO-Virgo data● run online coherent search on S5 data● Produce likelihood statistic for triggered population
search (with S.Mohanty)● perform BH-BH merger search