PowerPoint Template
Download
Report
Transcript PowerPoint Template
Gravitational Wave
Data Analysis
Probability and Statistics
Junwei Cao (曹军威) and Junwei Li (李俊伟)
Tsinghua University
Gravitational Wave Summer School
Kunming, China, July 2009
Drawback of Matched Filtering
Two premises of matched filtering
Premise 1: A given signal is present in data
stream
Premise 2: The form of h(t) is known
Premises are impractical
Repeat matched filtering with many
different filters
A number of “events” are extracted
“events” indicate that in the detector
happened something, which deserves further
scrutiny
Definition of Probability
Consider a set S with subsets A, B…
Define probability P as a real function
For every A in S, P(A)≥0
For disjoint subsets (i.e. A∩B= ),
P(A∪B)=P(A)+P(B)
P(S)=1
Further more, conditional probability
P( A B)
P( A | B)
P( B)
Two Approaches of Probability
Frequentist (also called classical)
A, B, … are the outcome of a repeatable
experiment, and P(A) is defined as the
frequency of occurrence of A
In considering the conditional probability, such
as P(data | hypothesis), one is never allowed to
think about the probability that the parameters
take a given a value, nor of the probability that a
hypothesis is correct
Two Approaches of Probability(contd.)
Bayesian approach
Bayes’ theorem
P( B | A) P( A)
P( A | B)
P( B)
(3.1)
P( B) P( B | Ai ) P( Ai )
(3.2)
i
P( A | B)
P ( B | A) P( A)
P( B | A ) P( A )
i
i
i
(3.3)
Prior and Posterior Probability
P(hypothesis | data) P(data | hypothesis ) P(hypothesis)
Posterior probability
Likelihood function Prior probability
(3.4)
Which One to be Chosen
Depend on the type of experiment
Elementary particle physics is suited for the
classical approach since it is the physicist that
controls the parameters of the experiment
In astrophysics, the sources can be rare, and
each one is very interesting individually, e.g. a
single BH-BH binary coalesces
Before Parameter Estimation
A number of free parameters
A family of possible templates
Denoted generically as h(t ; )
1 ,..., N is a collection of parameters
A family of optimal filters
Denoted generically as K (t; )
~
~
Determined by eq. 2.12, K ( f ; ) ~ h( f ; ) / Sn ( f )
Must discretize the θ-space
For some template, the SNR exceeds
a predefined threshold, indicating a
detection
Parameter Estimation
How to reconstruct parameters of the
source
Assume that n(t) is stationary and
Gaussian
Corresponding Gaussian probability
distribution for n(t)
p(n0 ) N exp (n0 | n0 ) / 2
(3.5)
This is the probability that n(t) has a
given realization n0(t)
Likelihood Function
Assumption
s(t ) h(t;t ) n0 (t )
θt is the true value of the parameters θ
Likelihood function for s(t), by
plugging n0=s-h(θt) into 3.5
1
( s | t ) N exp ( s h(t ) | s h(t ))
2
(3.6)
Introduce ht≡h(θt)
1
1
( s | t ) N exp (ht | s) (ht | ht ) ( s | s)
2
2
(3.7)
Posterior Probability Distribution
(t ) to
Introduce a prior probability
eq. 3.7
1
p (t | s ) Np (t ) exp (ht | s ) (ht | ht )
2
(0)
p
(t ) can be an un-flat prior in θt
(0)
(3.8)
Estimator: a rule for assigning , the
most probable value of θt
Consistency
The bias b E ( ) t
Efficiency
Robustness
Maximum Likelihood Estimator
First, the prior probability is flat
Max. posterior Max. likelihood ( s | t )
Denote the maximum likelihood estimator by
ML ( s)
Simpler to maximize log
1
log ( s | t ) (ht | s ) (ht | ht )
2
(3.9)
Define i / ti
(i ht | s) (i ht | ht ) 0
(3.10)
Maximum Posterior Probability
Maximize the full posterior probability
Takes into account the prior probability distribution
Non-trivial prior information
Example
Two-dimensional parameters space (θ1,θ2)
~
Only interested in θ1
p(1 | s) p(1 , 2 | s)
Drawback
An ambiguity on the value of the most probable
value of θ1
Unable to minimize the error on θ1 determination
Bayes Estimator
Most probable value of parameters
Bi ( s) i p( | s)d
(3.11)
Errors
i i j j
B B (s) B (s) p( | s)d
ij
(3.12)
The “operational” meaning
i
i
(
s
)
B Is the value of , averaged over an
ensemble of same outputs
Drawback: computational costs
Gravitational Wave
Data Analysis
Junwei Cao
[email protected]
http://ligo.org.cn