Pulsar Detection and Parameter Estimation with MCMC - Six Parameters Nelson Christensen Physics and...
-
Upload
jaidyn-rimer -
Category
Documents
-
view
216 -
download
0
Transcript of Pulsar Detection and Parameter Estimation with MCMC - Six Parameters Nelson Christensen Physics and...
Pulsar Detection and Parameter Estimation with MCMC - Six
Parameters
Nelson ChristensenPhysics and Astronomy
Carleton College
GWDAWDecember 2003
Collaborators
Rejean Dupuis, John Veitch, Graham Woan Department of Physics and AstronomyGlasgow University
Renate Meyer, Richard UmstaetterDepartment of StatisticsUniversity of Auckland
Markov Chain Monte Carlo
● Computational Bayesian Technique● Metropolis-Hastings Routine● Estimate Parameters and Generate Summary
Statistics (PDFs, cross correlation, etc)● 6 Unknown Parameters (so far): h
0, f,
df/dt● Initial Application SN1987a: Location known but
other parameters not
Starting Point
The Starting point for the MCMC is the same likelihood and priors used in the LSC time domain (bayesian) search
Remember Rejean's talk
Time domain method (Rejean)
• The phase evolution can be removed by heterodyning to dc.– Heterodyne (multiply by e-i (t)) calibrated time domain data from detectors.– This process reduces a potential GW signal h(t) to a slow varying
complex signal y(t) which reflects the beam pattern of the interferometer.– By means of averaging and filtering, we calculate an estimate of this
signal y(t) every 1 minute which we call Bk.
The Bk’s are our data which we compare with the model
y t =14Fƒ t h0 1ƒ cos 2 i e i 2? 0 i
2F X žtŸh0 cos i e i2 ? 0
Bayesian analysis (Rejean)
A Bayesian approach is used to determine the joint posterior distribution of the probability of the unknown parameters,
a=(h0f,df/dt), via the likelihood:
p˜a# Bk p
˜a p Bk #
˜a
Posterior Prior Likelihood
Model
p Bk #˜a exp
k
Bk y t k ;˜a
2 ?k2
2
¿
Bk’s are processed data Noise estimate
Metropolis Hasting Routine
Generate a chain of parameter values
PDFs generated from the distribution of chain values
Quasi-Random (somewhat intelligent) walk through parameter space
4 or 5 parameter search is straightforward h
0, , f
6th parameter df/dt causes problems
Distributions for f and df/dt are VERY narrow Hard to find
Reparametrization New Variables
fstart
=f + (1/2)(df/dt)Tstart
fend
=f + (1/2)(df/dt)Tend
S2 Injected Signals: Analysis using a Metropolis-Hastings Markov chain
Monte Carlo Routine
Here we present results for analysis of signals injected during S2.
Injected Signal Parameters
Signal 2:Unknown Parameters of Injected Signal:
h0 = 20.0 (± calibration errors, in some units) = 0.0 = 0.0 cos()=0.0
df/dt = -5.0e-9
Rejean put in an offset in a f to make it non-zero. f = 6.17e-4
H1 signal 524 minutes of data, 458 minutes for H2, and 394 minutes for L1.
One Parameter Hiccup
Parameter of the phase : Injected with wrong phase (-), and with factor of 2 for gravity wave this resulted in an effective value of
= -/4
H1 Parameter Estimation Result
H1 Parameter Estimation Result
H1 Parameter Estimation Result
Mean h0 1.99e+01 psi -1.62e-02 phase -7.495e-01 cosiota 4.20e-02 df 6.19e-04 dfdot -5.09e-09
Quantiles for each variable: 2.5% h0 1.55e+01 psi -1.17e-01phase -1.40e+00cosiota -9.12e-02df 6.06e-04 dfdot -5.66e-09
97.5%h0 2.41e+01psi 8.37e-02phase -2.96e-02cosiota 1.82e-01df 6.31e-04dfdot -4.48e-09
H2 Parameter Estimation Result
H2 Parameter Estimation Result
H2 Parameter Estimation Result
Mean h0 1.71e+01 psi -1.63e-02 phase -7.09e-01 cosiota 4.01e-02 df 6.19e-04 dfdot -5.09e-09
Quantiles for each variable: 2.5% h0 1.13e+01 psi -1.90e-01phase -1.48e+00cosiota -2.02e-01df 5.95e-04 dfdot -5.83e-09
97.5%h0 2.34e+01psi 1.31e-01phase 5.00e-01cosiota 2.69e-01df 6.34e-04dfdot -3.94e-09
Tested Code on Synthesized Data
f 0.007 Hzdf/dt -2.5e-10d2f/dt2 0.0h0 “varied” 0.4 1.0 0.5RA 1.23DEC 0.32110 days worth of data
Noise k = 1.0
Estimate of h0
Estimate of df
Estimate of df/dt
Goal of our Work
SN1987a Location known, but all other parameters unknown.
Heterodyne 1/60 Hz bandwidth: 5 Hz search with 300 processes.
Also useful for radio observed pulsar: work in concert with time domain search.
John Veitch (Glasgow) is presently running code on E10 and S3 injections.
Delayed Rejection
ABSTRACT
Presented is a Markov chain Monte Carlo technique for finding a laser interferometer detected pulsar signal and estimating 6 unknown parameters, including the pulsar frequency and frequency derivative. The technique used is called Delayed Rejection in Reversible Jump Metropolis-Hastings. This method will be explained, as well as noting how a simple extension to multiple computer processors will allow for a wider frequency band search. The goal of this research is to search for a possible (but radio quiet) source at a know location; for example, SN1987A. The code has been successfully demonstrated with synthesized data, and with LIGO S2 injected signals. The results of the code on these artificial signals will be presented.
The Details of the Technique
This talk will describe the details of how the 6 parameter MCMC search works. However, I have not written this yet. The results of the technique
using S2 injected signals are presented here for LSC review. Results on synthesized and ficticious data are not included here (but will eventually make it
into the talk).