Beruflich Dokumente
Kultur Dokumente
2. PROBLEM FORMULATION Note that since both noise and signal are assumed tempo-
Consider a spectrum monitor equipped with M antennas rally iid, the likelihood is given by the product of the indi-
which is to monitor a given frequency channel. The received vidual pdfs, i.e.,
signals are downconverted and sampled at the Nyquist rate.
We assume no synchronization with any potentially present 1
exp −K tr CR−1 ,
f (Y | R) = (5)
primary signal. Primary transmission, if present, has an πM K det(R) K
Note that the model above assumes that the channel from where L(ν(P ), K) is a penalty function depending on the
the primary transmitter to the spectrum monitor is frequency- number of degrees of freedom of the model ν(P ) and on K,
flat in the RF channel bandwidth, and that it remains con- and R̂ML denotes the Maximum Likelihood (ML) estimate
stant for the duration of the sensing time. We restrict of R for fixed P .
our analysis to both signal and noise following a zero-mean
Gaussian distribution. The reasons for adopting a Gaussian It is interesting to note that the penalty function is required
model for the primary signal are, (i) the Gaussian pdf for to obtain a meaningful estimate of the rank of the covariance
a signal is the least informative one when the noise is as- matrix of the primary signal. If we ignore the penalty func-
sumed Gaussian as well, (ii) the Gaussian model is accurate tion related to the rank, from the likelihood expression (5)
for primary system using multicarrier modulation when the it is easy to see that f (Y | R) is maximized for R = C. This
can be always achieved if we allow a rank P large enough.1 Algorithm 1: Iterative estimation of HΣ and Σ via alter-
nating optimization.
The penalty function L(ν(P ), K) can be determined via in- Input: Starting point α(0) and C.
formation theoretic or Bayesian criteria, such as the BIC, Output: ML estimates of HΣ and Σ.
MDL or AIC presented previously. The model selection Initialize: n = 0;
schemes proposed in [14], [18] and [5] fit into the general repeat
model (7) with different expressions for f (Y | R̂ML ) and for Compute Σ−1 (n) = diag(αn ) ;
the penalty function L(ν(P ), K). (n+1)
Obtain CΣ = Σ−1 −1
(n) CΣ(n) and its EVD;
(n+1)
Compute HΣ from (10) (fixed Σ−1
(n) );
3.1 Maximum Likelihood Estimation Solve (12) to obtain α(n+1) (fixed HΣ
(n+1)
);
We shall assume without loss of generality that H has full Update n = n + 1;
rank. From (4) we have that R corresponds to a rank-P until Convergence;
matrix plus a diagonal matrix.
The ML estimation problem maxH,Σ2 f (Y | R) can be rewrit- Note that, given the trace term in (11), the matrix CT ⊙R−1Σ
ten as is positive semidefinite. Hence, the problem (12) is convex
with respect to the parameter vector α and, therefore, it can
minimize log det RΣ − log det Σ−2 + tr CΣ R−1
, (9) be efficiently solved using any convex optimization solver.
Σ
HΣ ,Σ
subject to [Σ]i,i ≥ 0. that the Lagrangian of the cost function and its gradient are
respectively given by
T
Defining the vector α = [Σ−1 ]1,1 , . . . , [Σ−1 ]M,M , the
M
X M
X
trace term in (11) can be reorganized to obtain an equiv- J(α) = αT (CT ⊙ R−1
Σ )α − log α2i − λi αi , (15)
i=1 i=1
1
For example, in the extreme case P = M , H = C1/2 and ∂J(α) 1 λi
= 2 αT (CT ⊙ R−1
Σ )ei − − . (16)
Σ2 = 0, obtaining R = HH H + Σ2 = C. ∂αi αi 2
Equating (16) to zero and rearranging terms we obtain 1.2 1.2
1
α̂T (CT ⊙ R̂−1
Σ )(α̂i ei ) = 1 + λi αi , (17)
2 1 1
Pr { P̂ = P }
Pr { P̂ = P }
ness condition, λi αi = 0 for all i. Using this property and 0.8 0.8
adding together the equations (17) for i = 1, . . . , M we ob-
tain 0.6 0.6
T T
α̂ (C ⊙ R̂−1
Σ )α̂ = M. (18)
0.4 0.4
Alternating Alternating
Therefore, we have that the quasi-ML estimate R̂qML fullfils Chiani et al.
Wax et al.
Chiani et al.
Wax et al.
0.2 0.2
0 1 2 3 4 0 1 2 3 4
tr CR̂−1
qML = tr CΣ̂
−1 −1 −1
R̂Σ Σ̂ (19) P P
(a) (b)
= α̂T (CT ⊙ R̂−1
Σ )α̂ = M. (20)
Finally, noting that (20) does not depend on P , we may Figure 1: Probability of correct estimation versus
rewrite (14) as the rank of the primary signal. (a) iid noise (b)
non-iid noise.
P̂ = arg min − log det(R̂qML ) + L(ν(P ), K)/K . (21)
P
Unless otherwise specified, the noise level at each antenna is
In our problem, the number of free-adjusted parameters fixed for each experiment, and for each Monte Carlo realiza-
ν(P ) can be computed as follows: the diagonal matrix Σ2 tion the entries of the channel matrix H are independently
presents M real parameters and the P eigenvalues and eigen- drawn from a complex Gaussian distribution (thus obtain-
vectors conforming the H matrix correspond to P +P (2M − ing a Rayleigh fading scenario) and scaled so that the SNR
(P + 1)) real parameters. Then the total number of free real is constant during the experiment:
parameters is given by
tr(HHH )/P
ν(P ) = M + P (2M − P ). (22) SNR (dB) = 10 log10 . (25)
tr(Σ2 )/M