You are on page 1of 4

Lab work on detection and estimation theory, SiCoM 3A 2013 This lab worked is based on a paper published in 2002

whose reference is O.Michel, A.Ferrari : A brief introduction to detection problems and illustration, in [Astronomy with high contrast imaging, EAS publication series, 2002.

1
1.1

Studying a toy model for direct detection of extra-solar planet


Problem statement and notations

A set of N elementary contiguous photo-detectors is considered. All the detectors are supposed to be perfectly identical to each other and independent from each other (the response to a photon or a photonic ux does aect the cell receiving the photons, but not its neighbors). Such eects as diusion and blooming are not considered here. The eect of the coronographic setup and the eventual adaptive optic are assumed to be perfect. For sake of conciseness in the notations and without loss of generality, a one-dimensional image is considered in the sequel. Hence, the toy problem under consideration resumes to testing the following hypothesis: H0 : n = 1 . . . N, xn Poiss(f ) H1 : n = n1 , . . . , n1 + r 1, xn Poiss(f + hA(n (n1 + otherwise: xn Poiss(f )
r 1 2 )))

(1) (2)

where f is the background noise intensity, (h, n1 ) are the brightness and position of the planet and A(n) is the optical instrument response function (the PSF is usually considered an Airy function) of width r1 . We will consider herein a telescope of 100m2 operating at a wavelength of 10m and in a bandwith of 1m. According to these values, every second an average of 102 photons will be received from the planet and 108 photon will be recieved from the star. We assume that the coronograph reduces the star ux by a factor 104 , that the optical instrument response function spreads over 10 pixels, typical values for the zodiacal and exo-zodiacal background noise, and quantum eciency of the CCD. These asumptions lead to typical optimistic values of h = 10T , f = 104 T and r = 10 pixels, where T stands for the exposure time.

1.2

Detection algorithm with known parameters

In the sequel, both the position n1 of the planet and the ux h received from it are assumed to be known.
1x n

Poiss(f ) if P (xn = k) = exp(f ) f ) k!

Let X1:N be the N -dimensional joined observation of the N pixels, X1:N = [x1 , . . . , x N ]T .

def

1. Express the likelihood function of X1:N under H0 and the alternative hypothesis H1 . 2. Express the Neyman Pearson statistics for the problem.
h << 1 (this assumption will be kept in the sequel), show that 3. Assuming f the simplied test corresponds to the classical matched lter.

4. What is the distribution of the likelihood ratio for the simple case where A(n) = 1, for e.g. 4 n 5? TO DO With the values given above, simulate a set of say 100 N -dimensional observations under each hypothesis, for T = 1minute, and T = 20 minutes. Plot the experimental PFA and PD obtained for this simulation. What can you conclude about the exposure time required to achieve reasonable performances? The limiting distribution of a standardized Poisson distribution when the parameter goes to innity as a unit normal distribution suggests to replace the signal model by: H0 : n = 1 . . . N, xn N (f, f ) H1 : n = n1 , . . . , n1 + r 1, xn N (f + hA(.), f + hA(.)) otherwise: xn N (f, f ) (3) (4)

for the derivation of the test statistics distribution. 1. What is now tyhe statistical distribution of the test under each hypothesis? 2. Express the PFA and PD as a function of Q(t) = p =
(r 1)/2 n=(r 1)/2 1 2 t

expx

/2

dx and

A(n).

TO DO Compare your expressions for PFA and PD with the simulation results previously obtained.

Detection algorithm with unknown parameters

The optimal NPD gives a reference to which suboptimal detectors can be compared. However, it requires a priori knowledge of the parameters n1 and h (the parameters f is assumed to be known). Otherwise, it can be replaced by the GLRT obtained by the ratio of the supremum of the likelihood function with respect to the unknown parameters under both hypothesis. 2

Here, the unknown parameters n1 and h are estimated by maximizing the log-likelihood (n1 , h). 1. Prove that h may be obtained by solving
n1 +r 1

0 = 1 +
n=n1

1 xn A(n (n1 + r 2 )) r 1 f + hA(n (n1 + 2 ))

(5)

2. Compute 2 (n1 , h)/ 2 h and show that it is always strictly negative if at least one variable xn on [n1 , n1 + r 1] is non-zero. As the rst derivative of (n1 , h) possesses r negative singularities, the equation (n1 , h)/h = 0 has at most one positive solution. For a given n1 , maximizing the likelihood with respect to h resumes to identify the unique positive root of the r-th order polynomial equation:
(r 1)/2

(h) = P (h) +
n=(r 1)/2

xn1 +n+(r1)/2 Qn (h)

(6)

where polynomials P (h) and Qn (h) are dened as:


(r 1)/2

P (h) = 1

(f + hA(n))
n=(r 1)/2 (r 1)/2

(7)

Qn (h) = A(n)
q =(r 1)/2,q =n

(f + hA(q ))

(8)

P (h) and Qn (h) do not depend on n1 or the data and can therefore be computed only once. Note that if the larger root of this polynomial equation is negative, necessarily the value h = 0 maximizes (n1 , h) over [0, ]. Computation of the roots of (h) is not be a problem here, the values of r being r = 10. The estimation of h can also be performed for the model derived in the previous section. The main drawback of this solution is that it leads to the computation of the roots of a polynom with order 2r and no direct technique for the proper root selection. TO DO Implement the detection scheme that can now be summarized as follows (n1 ) from 1. For n1 {1, . . . , N r}: compute h (n1 ) = max{0, root((h)) h (n1 )) for n1 2. The ML of n1 , n1 , is the global maximum of LP (X1:N ; h {1, . . . , N r}. (n1 ). 3. The ML estimate of h is h = h 3

4. The GLRT is:

LP (Xn1 :n1 +r1 ; h)


H1

H0

Report the result obtained on a single plot showing the estimated positions n1 on the x-axis, and thee stimated h on the y-axis. The simulations generated in the rst part of the lab work may be used for this question. Comment your results.

You might also like