Y. Liu, C. Li / Information Sciences 326 (2016) 334–349 337
3.2. Bayesian estimation
Maximum a posterior (MAP) estimation is a typical method for Bayesian estimation. To obtain the MAP estimator, it is as-
sumed that the vector of parameters is a random vector instead of a deterministic unknown constant vector. Then, its posterior
distribution is estimated using the Bayes’ rule. The applicability of the Bayes’ rule to complex-valued data is due to (i) the pdf
of a complex random variable or vector is still a real-valued function; (ii) from the probabilistic point of view, the Bayes’s rule
considers the probability of occurrence, which means that all of the variable types (real, complex, continuous, discrete etc.) fall
elegantly in the same unifying framework.
According to the Bayes’ rule, we have
p(x|y) =
p
(y|x)p(x)
p(y)
=
p
(y|x)p(x)
C
D
p(y|x)p(x)dx
,
(13)
where p(y|x) is the likelihood, p(x) is a prior distribution of x and p(y) is called evidence, which is a normalizing factor for ensuring
its integral equal to unity.
Consequently, the Bayesian MAP estimator can be taken by
arg max
x
log p(x|y) = arg max
x
[log (p(y|x) + log p(x)], (14)
where the likelihood p(y|x) can be computed by (12).
Once the posterior p(x|y) is obtained, the estimate of each component of parameter x
(d)
can be obtained using the alternative
Bayesian estimate
ˆ
x
(d)
ˆ
x
(d)
=
C
x
(d)
p(x
(d)
|y)dx
(d)
, (15)
where the marginal density of x
(d)
is computed by
p(x
(d)
|y) =
C
(D−1)
p(x|y)dx
(1)
...dx
(d−1)
dx
(d+1)
...dx
(D)
. (16)
However, the computation of the multivariable integral in (16) is computationally difficult and thus the analytic solution of
(16) is usually intractable. Alternatively, the MCMC provides an efficient approach to simulatively solve this problem, and it is
introduced in the following section [13,29,32].
4. Complex Markov chain Monte Carlo
Markov chain Monte Carlo provides an algorithm for sampling from a certain probability distribution by constructing ran-
dom sequences as realizations of parameters with their limit distribution equal to the desired posterior. In the conventional
real-valued MCMC, even if the parameters are complex, the estimates are usually represented by their equivalent real-valued
composite vectors or by the polar coordinates with real-valued amplitude and phase [4], and the sampling is performed in
real parameter space. Yet, this is inefficient if the considered system is widely-linear or nonlinear. Due to the advantages of
complex-valued representation as introduced in the Introduction, we develop two complex MCMC sampling algorithms using
the Metropolis–Hastings (MH) sampling [15,22,23] and the differential evolution (DE) [28,31] in this section. Analog to the real
MCMC sampling algorithms, we first present the complex MCMC sampling algorithms and then analyze their convergence prop-
erties. It is noted that, the following study is performed based on the assumption that the observation signal y is proper. Under
this assumption, y is independent of y
∗
, and the pdfs required in Bayesian estimation can be characterized by their corresponding
Hermitian covariance matrices.
4.1. Complex Metropolis–Hastings sampling
We first consider the Metropolis–Hastings sampling [15,22]. Analog to the real MH sampling, we develop a complex MH sam-
pling algorithm, in which the proposal candidates are generated from a ‘feasible complex parameter set’. We call this algorithm
complex Metropolis–Hastings (CMH) sampling.
4.1.1. Algorithm description
Let x
k
denote the complex-valued sample of the Markov chain at iteration k, which is a D-dimensional complex vector. In the
next iteration k + 1, a new sample, x
k+1
, is generated from the previous sample, x
k
, by first generating a ‘candidate’ sample, ζ
k+1
subject to a proposal complex distribution γ ( · ).
Then, we determine whether this candidate is accepted with probability
α(ζ
k+1
|x
k
) = min
1,
p
(ζ
k+1
|y)
p(x
k
|y)
·
γ(x
k
|ζ
k+1
)
γ(ζ
k+1
|x
k
)
, (17)
where p
(ζ
k+1
|y) and p(x
k
|y) are computed according to the Bayes’ rule (13). Note that as the probability of a complex random
vector is a nonnegative real scalar, the acceptance probability
α( · ) is also a real-valued scalar within [0, 1]. If accepted, ζ
k+1