Statistical analysis of the covariance m
ARTICLE IN PRESS
Signal Processing 90 (2010) 1165–1175
Contents lists available at ScienceDirect
Signal Processing
journal homepage: www.elsevier.com/locate/sigpro
Statistical analysis of the covariance matrix MLE in
K-distributed clutter
Frédéric Pascal a,, Alexandre Renaux b
a
b
SONDRA/Supelec, 3 rue Joliot-Curie, F-91192 Gif-sur-Yvette Cedex, France
Laboratoire des Signaux et Systemes (L2S), Université Paris-Sud XI (UPS), CNRS, SUPELEC, Gif-Sur-Yvette, France
a r t i c l e in fo
abstract
Article history:
Received 22 January 2009
Received in revised form
13 July 2009
Accepted 26 September 2009
Available online 17 October 2009
In the context of radar detection, the clutter covariance matrix estimation is an
important point to design optimal detectors. While the Gaussian clutter case has been
extensively studied, the new advances in radar technology show that non-Gaussian
clutter models have to be considered. Among these models, the spherically invariant
random vector modelling is particularly interesting since it includes the K-distributed
clutter model, known to fit very well with experimental data. This is why recent results
in the literature focus on this distribution. More precisely, the maximum likelihood
estimator of a K-distributed clutter covariance matrix has already been derived. This
paper proposes a complete statistical performance analysis of this estimator through its
consistency and its unbiasedness at finite number of samples. Moreover, the closedform expression of the true Cramér–Rao bound is derived for the K-distribution
covariance matrix and the efficiency of the maximum likelihood estimator is
emphasized by simulations.
& 2009 Elsevier B.V. All rights reserved.
Keywords:
Covariance matrix estimation
Cramér–Rao bound
Maximum likelihood
K-distribution
Spherically invariant random vector
Notations
The notational convention adopted is as follows: italic
indicates a scalar quantity, as in A; lower case boldface
indicates a vector quantity, as in a; upper case boldface
indicates a matrix quantity, as in A. RefAg and I mfAg
are the real and the imaginary parts of A, respectively.
The complex conjugation, the matrix transpose operator,
and the conjugate transpose operator are indicated by , T ,
and H , respectively. The j th element of a vector a is
denoted aðjÞ . The n th row and m th column element of the
matrix A will be denoted by An;m. jAj and TrðAÞ are the
determinant and the trace of the matrix A, respectively.
denotes the Kronecker product. J J denotes any matrix
norm. The operator vec(A) stacks the columns of the
matrix A one under another into a single column vector.
The operator vech(A), where A is a symmetric matrix, does
the same things as vec(A) with the upper triangular
portion excluded. The operator veckðAÞ of a skewsymmetric matrix (i.e., AT ¼ A) does the same thing as
vechðAÞ by omitting the diagonal elements. The identity
matrix, with appropriate dimensions, is denoted I and the
zero matrix is denoted 0. E½ denotes the expectation
a:s:
operator. - stands for the almost sure convergence and
Pr
- stands for the convergence in probability. A zero-mean
complex circular Gaussian distribution with covariance
matrix A is denoted CN ð0; AÞ. A gamma distribution with
shape parameter k and scale parameter y is denoted
Gðk; yÞ. A complex m-variate K-distribution with parameters k, y, and covariance matrix A is denoted Km ðk; y; AÞ.
A central chi-square distribution with k degrees of freedom is denoted w2 ðkÞ. A uniform distribution with
boundaries a and b is denoted U ½a;b .
1. Introduction
Corresponding author. Tel.: þ33 147405320; fax: þ33 147402199.
E-mail addresses: frederic.pascal@supelec.fr (F. Pascal),
alexandre.renaux@lss.supelec.fr (A. Renaux).
0165-1684/$ - see front matter & 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.sigpro.2009.09.029
The Gaussian assumption makes sense in many
applications, e.g., sources localization in passive sonar,
ARTICLE IN PRESS
1166
F. Pascal, A. Renaux / Signal Processing 90 (2010) 1165–1175
radar detection where thermal noise and clutter are
generally modelled as Gaussian processes. In these
contexts, Gaussian models have been thoroughly investigated in the framework of statistical estimation and
detection theory (see, e.g., [1–3]). They have led to
attractive algorithms such as the stochastic maximum
likelihood method [4,5] or Bayesian estimators.
However, the assumption of Gaussian noise is not
always valid. For instance, due to the recent evolution of
radar technology, one can cite the area of space time
adaptive processing-high resolution (STAP-HR) where the
resolution is such that the central limit theorem cannot
be applied anymore since the number of backscatters is
too small. Equivalently, it is known that reflected signals
can be very impulsive when they are collected by a low
grazing angle radar [6,7]. This is why, in the last decades,
the radar community has been very interested in
problems dealing with non-Gaussian clutter modelling
(see, e.g., [8–11]).
One of the most general non-Gaussian noise model is
provided by spherically invariant random vectors (SIRV)
which are a compound processes [12–14]. More precisely,
an SIRV is the product of a Gaussian random vector (the
so-called speckle) with the square root of a non-negative
random scalar variable (the so-called texture). In other
words, a noise modelled as an SIRV is a non-homogeneous
Gaussian process with random power. Thus, these kind of
processes are fully characterized by the texture and the
unknown covariance matrix of the speckle. One of the
major challenging difficulties in SIRV modelling is to
estimate these two unknown quantities [15]. These
problems have been investigated in [16] for the texture
estimation while [17,18] have proposed different estimation procedures for the covariance matrix. Moreover, the
knowledge of these estimates accuracy is essential in
radar detection since the covariance matrix and the
texture are required to design the different detection
schemes.
In this context, this paper focuses on parameters
estimation performance where the clutter is modelled
by a K-distribution. A K-distribution is an SIRV, with a
gamma distributed texture depending on two real positive
parameters a and b. Consequently, a K-distribution
depends on a, b and on the covariance matrix M. This
model choice is justified by the fact that a lot of
operational data experimentations have shown the good
agreement between real data and the K-distribution
model (see [7,19–22] and references herein).
This K-distribution model has been extensively studied
in the literature. First, concerning the parameters estimation problem, [23,24] have estimated the gamma distribution parameters assuming that M is equal to the
identity matrix, [17] has proposed a recursive algorithm
for the covariance matrix M estimation assuming a and
b known and [25] has used a parameter-expanded
expectation-maximization (PX-EM) algorithm for the
covariance matrix M estimation and for a parameter n
assuming n ¼ a ¼ 1=b. Note also that estimation schemes
in K-distribution context can be found in [26,27] and
references herein. Second, concerning the statistical
performance of these estimators, it has been proved in
[28] that the recursive scheme proposed by [17] converges
and has a unique solution which is the maximum
likelihood (ML) estimator. Consequently, this estimator
has become very attractive. In order to evaluate the
ultimate performance in terms of mean square error, [23]
has derived the true Cramér–Rao Bound (CRB) for the
parameters of the gamma texture (namely, a and b)
assuming M equal to the identity matrix. Gini [29] has
derived the modified CRB on the one-lag correlation
coefficient of M where the parameters of the gamma
texture are assumed to be nuisance parameters. Concerning the covariance matrix M, a first approach for the true
CRB study, which is known to be tighter than the modified
one, has been proposed in [25] whatever the texture
distribution. However, note that, for the particular case of
a gamma distributed texture, the analysis of [25] involves
several numerical integrations and no useful information
concerning the structure of the Fisher information matrix
(FIM) is given. Finally, classical covariance matrix estimators are compared in [30] in the more general context of
SIRV.
The knowledge of an accurate covariance matrix
estimate is of the utmost interest in context of radar
detection since this matrix is always involved in the
detector expression [30]. Therefore, the goal of this
contribution is twofold. First, the covariance matrix ML
estimate statistical analysis is provided in terms of
consistency and bias. Second, the closed-form expression
of the true CRB for the covariance matrix M is given and is
analyzed. Finally, through a discussion and simulation
results, classical estimation procedures in Gaussian and
SIRV contexts are compared.
The paper is organized as follows. Section 2 presents
the problem formulation while Sections 3 and 4 contain
the main results of this paper: the ML estimate statistical
performance in terms of consistency and bias and the
derivation of the true CRB. Finally, Section 5 gives
simulations which validate theoretical results.
2. Problem formulation
In radar detection, the basic problem consists in
detecting if a known signal corrupted by an additive
clutter is present or not. In order to estimate the clutter
parameters before detection, it is generally assumed that
K signal-free independent measurements, traditionally
called the secondary data ck , k ¼ 1; . . . ; K are available.
As stated in the introduction, one considers a clutter
modelled thanks to a K-distribution denoted
ck Km ða; ð2=bÞ2 ; MÞ:
From the SIRV definition, ck can be written as
pffiffiffiffiffi
ck ¼ tk xk ;
ð1Þ
ð2Þ
where tk is gamma distributed with parameters a and
ð2=bÞ2 , i.e., tk Gða; ð2=bÞ2 Þ and, where xk is a complex
circular zero-mean m-dimensional Gaussian vector with
covariance matrix E½xk xH
k ¼ M independent of tk . For
identifiability considerations, M is normalized according
to TrðMÞ ¼ m (see [17]). Note that the parameter a
represents the spikiness of the clutter. Indeed, when a is
ARTICLE IN PRESS
F. Pascal, A. Renaux / Signal Processing 90 (2010) 1165–1175
high the clutter tends to be Gaussian and, when a is small,
the tail of the clutter becomes heavy.
The probability density function (PDF) of a random
variable tk distributed according to Gða; ð2=bÞ2 Þ is given by
!
!a
b2
pðtk Þ ¼
4
tak 1
b2
exp tk ;
GðaÞ
4
ð3Þ
where GðaÞ is the gamma function defined by
Z þ1
GðaÞ ¼
xa1 expðxÞ dx:
0
ð4Þ
From Eq. (2), the PDF of ck can be written
!
Z þ1
1
cH
ck
1
kM
pðtk Þ dtk ;
exp
pðck ; M; a; bÞ ¼
tm
pm jMj
tk
0
k
ð5Þ
which is equal to
pðck ; M; a; bÞ ¼
aþm
b
1
ðcH
ck ÞðamÞ=2
kM
aþm1 m
2
jMjGð Þ
p
a
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Kma b cH
M1 ck ;
k
ð6Þ
where Kn ðÞ is the modified Bessel function of the second
kind of order n [31].
Gini et al. have derived the ML estimator as the
solution of the following equation [17]:
^ ML
M
K
1X
H
^ 1
¼
cm ðcH
k M ML ck Þck ck ;
K k¼1
ð7Þ
where the function cm ðqÞ is defined as
pffiffiffi
b Kam1 ðb qÞ
cm ðqÞ ¼ pffiffiffi
ð8Þ
pffiffiffi :
2 q Kam ðb qÞ
^ ML has to be normalized as
Note that the ML estimate M
^ ML Þ ¼ m. Finally, it has been shown in [28] that
M : TrðM
the solution to Eq. (7) exists and is unique for the
aforementioned normalization.
^ ML
3. Statistical analysis of M
^ ML
This section is devoted to the statistical analysis of M
in terms of consistency and bias.
3.1. Consistency
^ of M is said to be consistent if
An estimator M
Pr
^ MJ - 0;
JM
K-þ1
where K is the number of secondary data ck ’s used to
estimate M.
^ ML is a consistent
^ ML consistency). M
Theorem 3.1 (M
estimate of M.
^ ML will be denoted MðKÞ
^
Proof. In the sequel, M
to show
^ ML and the number K
explicitly the dependence between M
of xk0 s. Let us define the function fK;M such that
8
D!D;
>
>
<
K
1X
ð9Þ
fK;M :
1
A!
cm ðcH
ck Þck cH
>
kA
k;
>
K k¼1
:
1167
where cm ðÞ is defined by Eq. (8), where D ¼ fA 2
Mm ðCÞjAH ¼ A; A positive definite matrix} with Mm ðCÞ ¼
fm m matrices with elements in Cg, and where C is the
^
set of complex scalar. As MðKÞ
is a fixed point of function
fK;M , it is the unique zero, which respects the constraint
^
TrðMðKÞÞ
¼ m, of the following function:
(
D!D;
gK :
A!gK ðAÞ ¼ A fK;M ðAÞ:
^
To prove the consistency of MðKÞ,
Theorem 5.9 of
[32, p. 46] will be used. First, the strong law of large
numbers (SLLN) gives
a:s:
gK ðAÞ - gðAÞ;
8A 2 D;
K-þ1
where
gðAÞ ¼ A E½cm ðcH A1 cÞccH
8A 2 D;
ð10Þ
for cKm ða; ð2=bÞ2 ; MÞ.
Let us now apply the change of variable y ¼ A1=2 c. We
obtain
yKm ða; ð2=bÞ2 ; A1=2 MA1=2 Þ
and
gðAÞ ¼ A1=2 ðI E½cm ðyH yÞyyH ÞA1=2
8A 2 D;
and
gK ðAÞ ¼ A1=2 I
8A 2 D;
!
K
1X
1=2
H
cm ðyH
y
Þy
y
:
k k k k A
K k¼1
Let us verify the hypothesis of Theorem 5.9 of
[32, p. 46]. We have to prove that for every e40,
Pr
ðH1 Þ : supfJgK ðAÞ gðAÞJg - 0;
K-þ1
A2D
ðH2 Þ :
inf
A:JAMJZe
fJgðAÞJg40 ¼ gðMÞ:
For every A 2 D, we have
1 X
K
H
H
H
JgK ðAÞ gðAÞJ ¼ JAJ
ðcm ðyH
y
Þy
y
E
½c
ðy
yÞyy
Þ
:
m
k k k k
K
k¼1
Since E½cm ðyH yÞyyH o þ 1, one can apply the SLLN to
H
the K i.i.d. variables cm ðyH
k yk Þyk yk , with same first order
moment. This ensures ðH1 Þ.
Moreover, the function cm ðcH A1 cÞ is strictly decreasing
w.r.t. A. Consequently, E½cm ðcH A1 cÞccH too. This implies
that E½cm ðcH A1 cÞccH aA, except for A ¼ M. This ensures
ðH2 Þ.
Finally, Theorem 5.9 of [32, p. 46] concludes the proof
Pr
^ ML M. &
and M
K-þ1
3.2. Bias
This subsection provides an analysis of the bias B
^ ML Þ ¼ E½M
^ ML M.
defined by BðM
^ ML ). M
^ ML is an unbiased
Theorem 3.2 (Unbiasedness of M
estimate of M at finite distance (i.e., at finite number K).
ARTICLE IN PRESS
1168
F. Pascal, A. Renaux / Signal Processing 90 (2010) 1165–1175
^ ML will be denoted M
^
Proof. For the sake of simplicity, M
in this part. By applying the following change of variable,
yk ¼ M1=2 ck , to Eq. (7), one has
where pðc1 ; . . . ; cK ; hÞ is the likelihood function of the
observations ck , k ¼ 1; . . . ; K. Concerning our model, the
parameter vector is
K
X
^ 1 y ÞM1=2 y yH M1=2 ;
^ ¼1
cm ðyH
M
k
k k
kT
K k¼1
h ¼ ðvechT ðRefMgÞ veckT ðI mfMgÞÞT ;
^ 1=2 :
T^ ¼ M1=2 MM
Therefore,
K
1X
^ 1 y Þy yH :
cm ðyH
T^ ¼
k k k
kT
K k¼1
T^ is thus the unique estimate (see [28, Theorem III.1]) of
^ ¼ m. Its statistic is clearly
the identity matrix, with TrðTÞ
independent of M since the yk ’s are i.i.d. SIRVs with a
gamma distributed texture and identity matrix for the
Gaussian
covariance
matrix.
In
other
words,
yk Km ða; ð2=bÞ2 ; IÞ.
Moreover, for any unitary matrix U,
K
1X
^ H 1 zk Þzk zH ;
cm ðzH
k ðUTU Þ
k
K k¼1
where zk ¼ Uyk are also i.i.d. and distributed as
^ H has the same distribution as T.
^
Km ða; ð2=bÞ2 ; IÞ and UTU
Consequently,
^ ¼ UE½TU
^ H
E½T
for any unitary matrix U:
^ is different from 0, Lemma A.1 of [30] ensures
Since E½T
^ 1=2 ,
^
that E½T ¼ gI for g 2 R. Remind that T^ ¼ M1=2 MM
^ ¼ TrðMÞ ¼ m, one
^ ¼ gM. Moreover, since TrðMÞ
then E½M
has
^ ¼ TrðEðMÞÞ
^ ¼ gTrðMÞ ¼ gm;
m ¼ EðTrðMÞÞ
ð13Þ
With the parametrization of Eq. (13), the structure of
CRB becomes
!1
F1;1 F1;2
;
ð14Þ
CRBh ¼
F2;1 F2;2
where
^ H¼
UTU
m2 1:
ð11Þ
where the Fi;j ’s are the elements of the FIM given by
"
#
@2 ln pðc1 ; . . . ; cK ; hÞ
;
F1;1 ¼ E
T
@vechðRefMgÞ@vech ðRefMgÞ
mðm þ 1Þ mðm þ 1Þ
;
ð15Þ
2
2
"
#
@2 ln pðc1 ; . . . ; cK ; hÞ
;
F1;2 ¼ FT2;1 ¼ E
T
@vechðRefMgÞ@veck ðI mfMgÞ
mðm þ 1Þ mðm 1Þ
;
ð16Þ
2
2
"
#
@2 ln pðc1 ; . . . ; cK ; hÞ
;
F2;2 ¼ E
T
@veckðI mfMgÞ@veck ðI mfMgÞ
mðm 1Þ mðm 1Þ
:
ð17Þ
2
2
Since the ck ’s are i.i.d. random vectors, one have from
Eq. (6),
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
K
Y
baþm ðcHk M1 ck ÞðamÞ=2
Kma ðb cH
pðc1 ; . . . ; cK ; hÞ ¼
M1 ck Þ:
k
aþm1 m
2
p jMjGðaÞ
k¼1
ð18Þ
Consequently, the log-likelihood function can be written as
ln pðc1 ; . . . ; cK jhÞ ¼ Kln
which implies that g ¼ 1.
^ is an unbiased estimate of M, for any
In conclusion, M
number K of secondary data. &
baþm
2aþm1 pm GðaÞ
K
X
þ
k¼1
!
KlnðjMjÞ
1
lnððcH
ck ÞðamÞ=2 Kma ðb
kM
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
M1 ck ÞÞ:
cH
k
4.1. Result
ð19Þ
3.3. Comments
Theorems 3.1 and 3.2 show the attractiveness of the
estimator (7) in terms of statistical properties, i.e.,
consistency and unbiasedness. Note also that this estimator is robust since the unbiasedness property is at finite
number of samples. In the next section, the Cramér–Rao
bound is derived for the observation model (2).
4. Cramér–Rao bound
In this Section, the Cramér–Rao bound w.r.t. M is
derived. The CRB gives the best variance that an unbiased
estimator can achieve. The proposed bound will be
compared to the mean square error of the previously
studied ML estimator (Eq. (7)) in the next section.
The CRB for a parameter vector h is given by
2
1
@ ln pðc1 ; . . . ; cK ; hÞ
CRBh ¼ E
;
ð12Þ
@h @hT
The next subsections will show that the CRB w.r.t. h, in
this context, is given by
!
T
1
0
1 ðH FHÞ
;
ð20Þ
CRBh ¼
K
0
ðKT FKÞ1
where the matrices H and K are constant transformation
matrices filled with ones and zeros such that vecðAÞ ¼
HvechðAÞ and vecðAÞ ¼ KveckðAÞ, with A a skew-symmetric matrix, and where
2ða þ 1Þ jða; mÞ
F ¼ ðMT MÞ1
mþ1
8
ðMT MÞ1=2 ðI þ vecðIÞvecT ðIÞÞðMT MÞ1=2 ;
ð21Þ
where jða; mÞ is given by Eq. (F.3).
Remarks.
jða; mÞ is a constant which does not depend on b since
b2 tk0 Gða; 4Þ (see Eq. (F.3)).
ARTICLE IN PRESS
F. Pascal, A. Renaux / Signal Processing 90 (2010) 1165–1175
The first term of the right hand side of Eq. (21) is the
Gaussian FIM (i.e., when tk ¼ 1 8k). Indeed, using
Eqs. (24), (49) and (30),1 the Gaussian CRB, denoted
GCRBh , is straightforwardly obtained as
!
T
T
1
1
0
1 ðH ðM MÞ HÞ
GCRBh ¼
:
K
0
ðKT ðMT MÞ1 KÞ1
ð22Þ
By identification with Eq. (21), it means that
2ða þ 1Þ g
¼ 0:
lim
a-1
8
mþ1
Eq. (23) is given by
1
@2 lnððcH
ck ÞðamÞ=2 Kma ðb
kM
T
@vechðRefMgÞ@vech ðRefMgÞ
2
¼
@ f ðzÞ
2
ð28Þ
T
2
T
þ ib @veck ðImfMgÞKT ðMT MÞ1 vecðck cH
kÞ
ð29Þ
and
T
M1 ÞH@vechðRefMgÞ
2
T
þ 2b @veck ðI mfMgÞKT ððMT ck cTk MT Þ
M1 ÞK@veckðI mfMgÞ
ð30Þ
and (details are given in Appendix C)
pffiffiffi
@f ðzÞ
1 Kmaþ1 ð zÞ
pffiffiffi Þ
d1 ðzÞ ¼
¼ pffiffiffi
@z
2 z Kma ð z
4.2.1. Analysis of F1;1
With Eq. (19) one has
ð31Þ
and
@2 ln pðc1 ; . . . ; cK ; hÞ
T
@vechðRefMgÞ@vech ðRefMgÞ
@ lnðjMjÞ
T
@vechðRefMgÞ@vech ðRefMgÞ
@2 f ðzÞ
@z2
pffiffiffi
pffiffiffi
1
Kmaþ1 ð zÞ 2
K
ð zÞ
pffiffiffi Þ pffiffiffi ma1 pffiffiffi Þ
:
¼
1þ
4z
Kma ð z
Kma ð z
z
d2 ðzÞ ¼
2
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
H 1 c ÞÞ
K @2 lnððcH M1 c ÞðamÞ=2 K
X
ma ðb ck M
k
k
k
:
T
@vechðRefMgÞ@vech ðRefMgÞ
k¼1
ð23Þ
The first part of the right hand side of Eq. (23) is given
ð32Þ
Consequently, the structure of Eq. (28) becomes
2
T
@2 f ðzÞ ¼ 2b d1 ðzÞ@vech ðRefMgÞHT P1 H@vechðRefMgÞ
2
T
þ 2b d1 ðzÞ@veck ðI mfMgÞKT P1 K@veckðI mfMgÞ
by
4
T
4
T
þ b d2 ðzÞ@vech ðRefMgÞHT P2 H@vechðRefMgÞ
@2 lnðjMjÞ
þ b d2 ðzÞ@veck ðI mfMgÞKT P2 K@veckðI mfMgÞ;
T
@vechðRefMgÞ@vech ðRefMgÞ
ð24Þ
thanks to Appendix A, Eq. (A.3), and thanks to the fact that
T
@veck ðI mfMgÞ
¼0
@vechðRefMgÞ
ð27Þ
@z ¼ b @vech ðRefMgÞHT ðMT MÞ1 vecðck cH
kÞ
To make the reading easier, only the outline of the CRB
derivation is given below. All the details are reported into
the different appendices.
¼ KHT ðMT MÞ1 H
:
Note that
@ @f ðzÞ
@2 f ðzÞ
@f ðzÞ 2
@2 f ðzÞ ¼
@z @z þ
@z @z ¼
@ z;
@z
@z
@z
@z2
2
K
T
@vechðRefMgÞ@vech ðRefMgÞ
@2 z ¼ 2b @vech ðRefMgÞHT ððMT ck cTk MT Þ
4.2. Outline of the proof
þ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
cH
M1 ck ÞÞ
k
with (details are given in Appendix B)
Consequently, due to the structure of I þ vecðIÞvecT ðIÞ,
the FIM for K-distributed observations is given by the
Gaussian FIM minus a sparse matrix depending on a, M
and m.
¼ K
1169
ð25Þ
ð33Þ
where
P1 ¼ ðMT ck cTk MT Þ M1 ;
ð34Þ
T
1
H
H
P2 ¼ ðMT MÞ1 vecðck cH
k Þvec ðck ck ÞðM MÞ :
ð35Þ
Therefore, Eq. (27) is given by
and
@2 f ðzÞ
T
@vech ðRefMgÞ
¼ I:
@vechðRefMgÞ
ð26Þ
Through the remain of the paper, let us set z ¼
pffiffiffi
b2 cHk M1 ck and f ðzÞ ¼ lnððz=b2 ÞðamÞ=2 Kma ð zÞÞ. The
second term (inside the sum) of the right-hand side of
T
@vechðRefMgÞ@vech ðRefMgÞ
2
4
¼ 2b d1 ðzÞHT P1 H þ b d2 ðzÞHT P2 H
ð36Þ
due to Eqs. (25) and (26).
Using Eqs. (23), (24), and (36), F1;1 is given by
2
F1;1 ¼ KHT ððMT MÞ1 þ 2b E½d1 ðzÞP1
4
þ b E½d2 ðzÞP2 ÞH;
2
1
1
With b ¼ 1, z ¼ cH
ck which is the term to be derived w.r.t. h in
kM
the exponential term of the multivariate Gaussian distribution.
ð37Þ
where d1 ðzÞ and d2 ðzÞ are defined by Eqs. (31) and (32),
respectively.
ARTICLE IN PRESS
1170
F. Pascal, A. Renaux / Signal Processing 90 (2010) 1165–1175
ðMT MÞ1=2 ðI þ vecðIÞvecT ðIÞÞðMT MÞ1=2 H;
The two expectation operators involved in the previous
equation can be detailed as follows:
E½d1 ðzÞP1 ¼ 12ðMT CMT Þ M1
ð38Þ
and
E½d2 ðzÞP2 ¼ 14ðMT MÞ1 ðW þ N !ÞðMT MÞ1 ;
ð39Þ
where
8
pffiffiffi
1 Kmaþ1 ð zÞ T
>
>
>
p
ffiffi
ffi
p
ffiffi
ffi
;
c
C
¼
E
c
>
>
z Kma ð zÞ k k
>
>
>
>
>
1
>
H
H
H
>
W
¼
E
c
Þvec
ðc
c
Þ
;
vecðc
>
k
k
<
k
k
z
pffiffiffi
>
2 K
ð zÞ
>
>
N ¼ E 3=2 maþ1pffiffiffi vecðck cHk ÞvecH ðck cHk Þ ;
>
>
z
>
Kma ð zÞ
>
>
pffiffiffi
pffiffiffi
>
>
>
1 Kma1 ð zÞKmaþ1 ð zÞ
>
H
H
>
p
ffiffi
ffi
vecðck cH
!
¼
E
:
k Þvec ðck ck Þ:
2
z
Km
a ð zÞ
4.2.2. Analysis of F2;2
The analysis of F2;2 is similar to the one used for F1;1.
Indeed, one has to calculate
@2 ln pðc1 ; . . . ; cK ; hÞ
T
@veckðI mfMgÞ@veck ðI mfMgÞ
2
¼ K
þ
C¼
2
b2
MT :
ð41Þ
Concerning W, one has
"
#
H
H
vecðck cH
1
k Þvec ðck ck Þ
;
W¼ 2E
cH
M1 ck
b
k
T
ð44Þ
where the expectation is taken under a complex
K-distribution Km ða; ð2=bÞ2 ; MÞ.
H
H
The closed-form expression of E vecðck cH
k Þvec ðck ck Þ=
1
H
ck M ck under a complex K-distribution is given in
Appendix E. One finds
a
b4 m þ 1
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
cH
M1 ck ÞÞ
k
@veckðI mfMgÞ@veck ðI mfMgÞ
k¼1
:
ð48Þ
Using Eq. (A.3), one has
K
@2 lnðjMjÞ
T
@veckðI mfMgÞ@veck ðI mfMgÞ
¼ KKT ðMT MÞ1 K
@veckðI mfMgÞ
ð42Þ
where the expectation is taken under a complex
K-distribution Km ða 1; ð2=bÞ2 ; MÞ.
Concerning !, one has
"
#
pffiffiffi
pffiffiffi
H
H
1
K
ð zÞKmaþ1 ð zÞ vecðck cH
k Þvec ðck ck Þ
ffiffiffi
;
! ¼ 2 E ma1 2 p
Kma ð zÞ
cH
M1 ck
b
k
8
T
ð49Þ
due to Eq. (25) and due to
where the expectation is taken under a complex
K-distribution Km ða; ð2=bÞ2 ; MÞ.
Concerning N, one has
"
#
H
H
vecðck cH
1
k Þvec ðck ck Þ
;
ð43Þ
N¼
E
2
cH
M1 ck
ða 1Þb
k
W¼
@ lnðjMjÞ
@veckðI mfMgÞ@veck ðI mfMgÞ
K @2 lnððcH M1 c ÞðamÞ=2 K
X
ma ðb
k
k
ð40Þ
After some calculus detailed in Appendix D, one finds
ð47Þ
where jða; mÞ is given by Eq. (F.3).
ðMT=2 M1=2 ÞðI þ vecðIÞvecT ðIÞÞðMT=2 M1=2 Þ
ð45Þ
T
@veck ðI mfMgÞ
¼ I:
ð50Þ
By using the same notation as for the derivation of F1;1
and by using Eq. (33), one obtains for the second term on
the right hand side of Eq. (48) as
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1
@2 lnððcH
ck ÞðamÞ=2 Kma ðb cH
M1 ck ÞÞ
kM
k
T
@veckðI mfMgÞ@veck ðI mfMgÞ
2
4
¼ 2b d1 ðzÞKT P1 K þ b d2 ðzÞKT P2 K;
ð51Þ
where P1 , P2 , d1 ðzÞ and d2 ðzÞ are defined by Eqs. (34), (35),
(31), and (32), respectively. Therefore, the structure of F2;2
is the same as the structure of F1;1 except that one
replaces the matrix H by the matrix K.
4.2.3. Analysis of F1;2 ¼ FT2;1
Due to the structure of Eq. (A.3) and Eq. (33), and since
T
the derivation is w.r.t. @vechðRefMgÞ and @veck ðI mfMgÞ, it
is clear that
F1;2 ¼ FT2;1 ¼ 0
ð52Þ
by using Eq. (25).
This concludes the proof of the CRB derivation.
5. Simulation results
and
N¼
8
1
ðMT=2 M1=2 ÞðI þ vecðIÞvecT ðIÞÞðMT=2 M1=2 Þ:
b4 m þ 1
ð46Þ
The structure of ! is analyzed in Appendix F.
Consequently, Eq. (37) is reduced to
2ða þ 1Þ jða; mÞ
F1;1 ¼ KHT ðMT MÞ1
mþ1
8
In this section, some simulations are provided in order
to illustrate the proposed previous results in terms of
consistency, bias, and variance analysis (throughout the
CRB). While no mathematical proof is given in other
sections, we show the efficiency of the MLE meaning that
the CRB is achieved by the variance of the MLE.
The results presented are obtained for complex
K-distributed clutter with covariance matrix M randomly
chosen.
ARTICLE IN PRESS
F. Pascal, A. Renaux / Signal Processing 90 (2010) 1165–1175
All the results are presented different values of
pffiffiffi
parameter a and b ¼ 2 a following the scenario of [17].
The size of each vector ck is m ¼ 3. Remember that the
parameter a represents the spikiness of the clutter (when
a is high the clutter tends to be Gaussian and, when a is
small, the tail of the clutter becomes heavy). The norm
used for consistency and bias is the L2 norm.
1171
5.1. Consistency
Fig. 1 presents results of MLE consistency for 1000
Monte Carlo runs per each value of K. For that purpose, a
^ KÞ ¼ JM
^ MJ versus the number K of ck ’s is
plot of DðM;
presented for each estimate. It can be noticed that the
^ KÞ tends to 0 when K tends to 1 for
above criterion DðM;
^ KÞ versus the number of secondary data for different values of a.
Fig. 1. DðM;
^ KÞ versus the number of secondary data for different values of a.
Fig. 2. CðM;
ARTICLE IN PRESS
1172
F. Pascal, A. Renaux / Signal Processing 90 (2010) 1165–1175
0.09
Empirical MSE α =1
CRB α = 1
Empirical MSE α = 0.1
CRB α = 0.1
Gaussian CRB
0.08
0.07
MSE
0.06
0.05
0.04
0.03
0.02
0.01
100
200
300
400
500
600
700
Number of secondary data
800
900
1000
Fig. 3. GCRB, CRB and empirical variance of the MLE for different values of a.
each estimate. Moreover, note that the parameter a has
very few influence on the convergence speed which
highlights the robustness of the MLE.
5.2. Bias
Fig. 2 shows the bias of each estimate for the different
values of a. The number of Monte Carlo runs is given in
the legend of the figure.
For that purpose, a plot of the
___
^ KÞ ¼ JM
^ MJ versus
the number K of ck ’s is
criterion CðM;
___
^ is defined as the empirical
presented for each estimate. M
^
mean of the quantities MðiÞ
obtained from I Monte Carlo
runs. For each iteration i, a new set of K secondary data ck
^
is generated to compute MðiÞ.
It can be noticed that,
as enlightened by the previous theoretical analysis, the
^ tends to 0 whatever the value of K. Furthermore,
bias of M
one sees again the weak influence of the parameter a on
the unbiasedness of the MLE.
(at finite number of samples) have been proved. In order
to analyze the variance of the estimator, the Cramér–Rao
lower bound is derived. The Fisher information matrix in
this case is simply the Fisher information matrix of the
Gaussian case plus a term depending on the tail of the
K-distribution. Simulation results have been proposed to
illustrate these theoretical analyses. These results have
shown the efficiency of the estimator and the weak
influence of the spikiness parameter in terms of consistency and bias.
Appendix A. Derivation of @2 lnfjMjg
To find this term and several other, we will use the
following results [33]:
@TrðXÞ ¼ Trð@XÞ;
ðA:1aÞ
@vecðXÞ ¼ vecð@XÞ;
ðA:1bÞ
5.3. CRB and MSE
@A1 ¼ A1 @AA1 ;
ðA:1cÞ
The CRB and empirical variance of the MLE for 10 000
Monte Carlo runs are plotted in Fig. 3. For comparison, we
also plot the Gaussian CRB (i.e., when tk ¼ 1 8k) given by
Eq. (22). Although it is not mathematically proved in this
paper, one observes, in Fig. 3, the efficiency of the MLE
even for impulsive noise (a small).
@jAj ¼ jAj TrðA1 @AÞ;
ðA:1dÞ
6. Conclusion
In this paper, a statistical analysis of the maximum
likelihood estimator of the covariance matrix of a complex
multivariate K-distributed process has been proposed.
More particularly, the consistency and the unbiasedness
@ðA
BÞ ¼ @ðAÞ
BþA
@ðBÞ
where
¼ or
ðA:1eÞ
@lnðjMjÞ ¼ TrðM1 @MÞ;
ðA:1fÞ
@ð@MÞ ¼ 0;
ðA:1gÞ
TrðABÞ ¼ TrðBAÞ
ðA:1hÞ
M1 M1 ¼ ðM MÞ1 ;
ðA:1iÞ
TrðAH BÞ ¼ vecH ðAÞvecðBÞ;
ðA:1jÞ
ARTICLE IN PRESS
F. Pascal, A. Renaux / Signal Processing 90 (2010) 1165–1175
vecðABCÞ ¼ ðCT AÞvecðBÞ:
@lnðjMjÞ ¼ TrðM
2
1
@MM1 ck Þ
@z ¼ b TrðcH
kM
from ðA:1fÞ
ð@MÞÞ
Concerning @2 z, one has
ðA:1kÞ
By using these properties, one has
1
2
2
from ðA:1aÞðA:1eÞðA:1gÞðA:1cÞ
¼ vecH ðM1 ð@MÞM1 Þvecð@MÞ
¼ b
2
¼ 2b Trð@MM
from ðA:1kÞ
ðM MÞ
1
M1 Þ@vecðMÞ
1
ck cH
Þ
kM
from ðA:1cÞðA:1hÞ
from ðA:1jÞðA:1kÞðA:1bÞ
2
T
@2 z ¼ 2b @vech ðRefMgÞHT ððMT ck cTk MT ÞM1 ÞH@vechðRefMgÞ
@vecðRefMg þ iI mfMgÞ
2
T
þ2b @veck ðI mfMgÞKT ððMT ck cTk MT Þ
M1 ÞK@veckðI mfMgÞ:
@vecH ðiI mfMgÞðMT MÞ1 @vecðRefMgÞ
@vecH ðRefMgÞðMT MÞ1 @vecðiI mfMgÞ
@vecH ðiI mfMgÞðMT MÞ1 @vecðiI mfMgÞ:
ðA:2Þ
Since vecH ðiI mfMgÞ ¼ ivecT ðI mfMgÞ and vecH ðRefMgÞ
¼ vecT ðRefMgÞ, @2 lnðjMjÞ is reduced to
@2 lnðjMjÞ ¼ @vecT ðRefMgÞðMT MÞ1 @vecðRefMgÞ
@vecT ðI mfMgÞðMT MÞ1 @vecðI mfMgÞ
¼ @vech ðRefMgÞH
from ðA:1eÞðA:1gÞ
By letting M ¼ RefMg þ iI mfMg, one obtains
¼ @vecH ðRefMgÞðMT MÞ1 @vecðRefMgÞ
T
@MM
1
¼ 2b @vecH ðMÞððMT ck cTk MT Þ
By letting M ¼ RefMg þ iI mfMg in Eq. (A.3), one has
@ lnðjMjÞ ¼ @vec ðRefMg þ iI mfMgÞ
1
2
¼ @vecH ðMÞðMT MÞ1 @vecðMÞ
from ðA:1bÞðA:1iÞ:
H
from ðA:1aÞ
1
TrðcH
Þ@MM1 ck
k @ðM
1
þ cH
@M@ðM1 Þck Þ
kM
from ðA:1jÞ
¼ vecH ð@MÞðMT M1 ÞH vecð@MÞ
T
ðB:2Þ
1
@2 z ¼ b TrðcH
@MM1 Þck Þ
k @ðM
@2 lnðjMjÞ ¼ TrðM1 ð@MÞM1 @MÞ
2
1173
T
ðMT MÞ1 H@vechðRefMgÞ
T
@veck ðI mfMgÞKT ðMT MÞ1 K@veckðI mfMgÞ;
ðA:3Þ
where the matrix H and K are constant transformation
matrices filled with ones and zeros such that vecðAÞ ¼
HvechðAÞ and vecðAÞ ¼ KveckðAÞ.
Appendix B. Derivation of @z and @2 z
ðB:3Þ
Appendix C. Derivation of @f ðzÞ=@z and @2 f ðzÞ=@z2
Concerning the first derivative of f ðzÞ, one has
pffiffiffi
@f ðzÞ
@
2
¼ lnððz=b ÞðamÞ=2 Kma ð zÞÞ
@z
@z
pffiffiffi
am
1
@Kma ð zÞ
pffiffiffi
¼
þ
:
2z
@z
Kma ð zÞ
ðC:1Þ
Since @Kn ðyÞ=@y ¼ Knþ1 ðyÞ þ ðn=yÞKn ðyÞ [34]. It follows
that
pffiffiffi
pffiffiffi
pffiffiffi
@Kma ð zÞ
1
ma
Kma ð zÞ:
ðC:2Þ
¼ pffiffiffi Kmaþ1 ð zÞ þ
2z
@z
2 z
Plugging Eq. (C.2) in Eq. (C.1), one obtains
pffiffiffi
@f ðzÞ
1 Kmaþ1 ð zÞ
pffiffiffi :
¼ pffiffiffi
@z
2 z Kma ð zÞ
ðC:3Þ
Concerning the second derivative of f ðzÞ, one has
pffiffiffi
@ f ðzÞ
1 @
1 K
ð zÞ
pffiffiffi maþ1 pffiffiffi Þ
¼
2
2 @z
@z
z Kma ð z
!
pffiffiffi
1 Kmaþ1 ð zÞ
1 @ Kmaþ1 ðyÞ
pffiffiffi
: ðC:4Þ
¼ 3=2
4z @y Kma ðyÞ
y¼pffiffiz
4z
Kma ð zÞ
2
By using the properties from Eq. (A.1a) to (A.1k), one
has for @z
2
2
1
1
ck Þ ¼ b @TrðcH
ck Þ
@z ¼ b @ðcH
kM
kM
2
1
¼ b TrðcH
Þck Þ
k @ðM
2
from ðA:1aÞ
1
TrðcH
@MM1 ck Þ
kM
from ðA:1cÞ
1
¼ b Trð@MM1 ck cH
Þ
kM
from ðA:1hÞ
¼ b
2
2
1
¼ b vecH ð@MÞvecðM1 ck cH
Þ
kM
2
Since @Kn ðyÞ=@y ¼ Kn1 ðyÞ ðn=yÞKn ðyÞ [34]. It follows
that
@ Kmaþ1 ðyÞ
@y Kma ðyÞ
y¼pffiffiz
from ðA:1jÞ
pffiffiffi
pffiffiffi
pffiffiffi
maþ1
pffiffiffi
Kmaþ1 ð zÞÞKma ð zÞ
ðKma ð zÞ þ
z
pffiffiffi
¼
2
Km
a ð zÞ
pffiffiffi
pffiffiffi
pffiffiffi
ma
Kmaþ1 ð zÞðKma1 ð zÞ þ pffiffiffi Kma ð zÞÞ
z
pffiffiffi
:
þ
2
Km
a ð zÞ
H
¼ b @vec ðMÞ
ðMT MÞ1 vecðck cH
kÞ
from ðA:1bÞðA:1kÞðA:1iÞ:
By letting M ¼ RefMg þ iI mfMg, one obtains
2
@z ¼ b @vecT ðRefMgÞðMT MÞ1 vecðck cH
kÞ
2
þ ib @vecT ðI mfMgÞðMT MÞ1 vecðck cH
kÞ
2
T
¼ b @vech ðRefMgÞHT ðMT MÞ1 vecðck cH
kÞ
2
T
þ ib @veck ðI mfMgÞKT ðMT MÞ1 vecðck cH
k Þ:
ðB:1Þ
ðC:5Þ
Plugging Eq. (C.5) in Eq. (C.4), one obtains
pffiffiffi
pffiffiffi
pffiffiffi
@2 f ðzÞ
1
1 Kmaþ1 ð zÞ 1 Kmaþ1 ð zÞKma1 ð zÞ
pffiffiffi
pffiffiffi
¼
:
þ 3=2
2
2
4z 2z
@z
Kma ð zÞ 4z
Kma ð zÞ
ðC:6Þ
ARTICLE IN PRESS
1174
F. Pascal, A. Renaux / Signal Processing 90 (2010) 1165–1175
Appendix D. Derivation of matrix C
since
xH
k xk ¼
The matrix C is given by
pffiffiffi
1 Kmaþ1 ð zÞ T
Gða 1Þ T
pffiffiffi ck ck ¼
E ½ck cH
k ;
2GðaÞ
z Kma ð zÞ
C ¼ E pffiffiffi
¼
ðD:1Þ
Kmaþ1 ðb
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
M1 ck Þ;
cH
k
ðD:2Þ
which is a complex K-distribution Km ða 1; ð2=bÞ2 ; MÞ.
Then
C¼
¼
1
1
ET ½ck cH
ET ½tk xk xH
k ¼
k
2ða 1Þ
2ða 1Þ
1
E½tk ET ½xk xH
k ;
2ða 1Þ
ðD:3Þ
2
where E½tk ¼ ða 1Þð2=bÞ since tk follows a Gamma
T
distribution Gða 1; ð2=bÞ2 Þ and ET ½xk xH
k ¼ M since xk is
a complex normal random vector (independent of tk ) with
zero mean and covariance matrix M. Consequently,
C ¼ ð2=b2 ÞMT .
E
M1 ck
cH
k
"
E
¼ ðM
1=2
M
Þ
#
H
H
vecðyk yH
k Þvec ðyk yk Þ
ðMT=2 M1=2 Þ;
yH
y
k k
ðE:1Þ
ðE:2Þ
where E½tk ¼ að2=bÞ2 and where
"
#
H
H
vecðxk xH
k Þvec ðxk xk Þ
R¼E
:
ðE:3Þ
xH
x
k k
qffiffiffiffiffiffi
Let us set xðjÞ
¼ r2j expðiyj Þ for j ¼ 1; . . . ; m. Note that
k
r2j w2 ð2Þ is independent of yj U ½0;2p . Consequently, the
elements of the matrix R can be rewritten as
2qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi3
Rk;l ¼ E4
r2p r2q r2p0 r2q0
Pm
j¼1
r2j
5E½expðiðyp yp0 þ yq0 yq ÞÞ
r2p r2p0 expðiðyp yp0 ÞÞ:
ðE:5Þ
(1) p ¼ p0 ¼ q ¼ q0 , i.e., l ¼ k ¼ p þ mðp 1Þ,
(2) p ¼ p0 , q ¼ q0 and paq, i.e., l ¼ k ¼ p þ mðq 1Þ,
(3) p ¼ q, p0 ¼ q0 and pap0 , i.e., l ¼ p0 þ mðp0 1Þ and
k ¼ p þ mðp 1Þ.
Consequently, the non-zero elements of Rk;l are given by
P
(1) Rpþmðp1Þ;pþmðp1Þ ¼ E½ðr2p Þ2 = m
r2 ¼ 4=ðm þ 1Þ,
Pj¼1 j2
r ¼ 2=ðm þ 1Þ,
(2) Rpþmðq1Þ;pþmðq1Þ ¼ E½r2p r2p0 = m
Pj¼1 j 2
(3) Rpþmðp1Þ;p0 þmðp0 1Þ ¼ E½r2p r2q = m
j¼1 rj ¼ 2=ðm þ 1Þ,
H
H
H 1
and the matrix E½vecðck cH
ck can be
k Þvec ðck ck Þ=ck M
written as
"
#
2
H
H
vecðck cH
2a
2
k Þvec ðck ck Þ
¼
E
mþ1 b
cH M1 ck
k
ðMT=2 M1=2 Þ
ðI þ vecðIÞvecT ðIÞÞðMT=2 M1=2 Þ:
1
H
ck where ck Km ða 1; ð2=bÞ2 ; MÞ is
vecH ðck cH
k Þ=ck M
given by
"
#
H
H
vecðck cH
2ða 1Þ 2 2
k Þvec ðck ck Þ
E
¼
mþ1 b
cH
M1 ck
k
ðMT=2 M1=2 Þ
ðI þ vecðIÞvecT ðIÞÞðMT=2 M1=2 Þ:
ðE:7Þ
Appendix F. Analysis of !
pffiffiffiffiffi
where yk Km ða; ð2=bÞ2 ; IÞ. Since yk ¼ tk xk , where
2
tk Gða; ð2=bÞ Þ is independent of xk CN ð0; IÞ, one has
"
#
H
H
vecðck cH
k Þvec ðck ck Þ
E
¼ ðMT=2 M1=2 ÞE½tk R
1
cH
M
c
k
k
ðMT=2 M1=2 Þ;
qffiffiffiffiffiffiffiffiffiffiffiffi
In the same way, the expression of E½vecðck cH
kÞ
In this appendix, we derive the expression of E½vecðck cH
kÞ
1
2
H
Þ=c
M
c
where
c
K
ð
a
;
ð2=
b
Þ
;
MÞ.
The
case
vecH ðck cH
m
k
k
k
k
where ck Km ða 1; ð2=bÞ2 ; MÞ will be, of course, straightforward. Let us set the following change of variable:
ck ¼ M1=2 yk . One obtains from Eq. (A.1k)
T=2
r2j and ½vecðxk xHk Þn
ðE:6Þ
Appendix E. Derivation of
H
H
H 1
E½vecðck cH
k Þvec ðck ck Þ=ck M ck
"
#
H
H
vecðck cH
k Þvec ðck ck Þ
j¼1
Note that E½expðiðyp yp0 þ yq0 yq ÞÞa0 if and only if
where the last expectation is taken under the distribution
baþm1 jMj1
pðck Þ ¼ aþm2
ðcH M1 ck Þðam1Þ=2
2
pm Gða 1Þ k
m
X
Let us set the following change of variable: ck ¼
pffiffiffiffiffi
M1=2 tk xk , where tk Gða; ð2=bÞ2 Þ is independent of
xk CN ð0; IÞ, one has
!¼
~ ðMT=2 M1=2 Þ;
ðMT=2 M1=2 Þ!
where
2
!~ ¼ E6
4
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
3
tk Kma1 ðb tk xHk xk ÞKmaþ1 ðb tk xHk xk Þ vecðxk xH ÞvecH ðxk xH Þ7
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
H
2
Km
a ðb tk xk xk Þ
Let us set
xðjÞ
k
k
xH
x
k k
k
5:
ðF:1Þ
qffiffiffiffiffiffi
2
2
2
¼ rj expðiyj Þ for j ¼ 1; . . . ; m. rj w ð2Þ is
independent of yj U ½0;2p . Consequently, due to Eq. (E.5),
~ can be rewritten as
the elements of the matrix !
U~
ðE:4Þ
1
b2
k;l
2
6
¼ E4tk0
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
ffi3
P
Pm 2ffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2
Kma1 b tk0 m
r2p r2q r2p0 r2q0 7
j¼1 rj Kmaþ1 b tk0
j¼1 rj
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pm 2 5
Pm 2ffi
2
j¼1 rj
Km
a b tk0
j¼1 rj
E½expðiðyp yp0 þ yq0 yq ÞÞ:
ðF:2Þ
ARTICLE IN PRESS
F. Pascal, A. Renaux / Signal Processing 90 (2010) 1165–1175
As before, E½expðiðyp yp0 þ yq0 yq ÞÞa0 if and only if
(1) p ¼ p0 ¼ q ¼ q0 , i.e., l ¼ k ¼ p þ mðp 1Þ,
(2) p ¼ p0 , q ¼ q0 and paq, i.e., l ¼ k ¼ p þ mðq 1Þ,
(3) p ¼ q, p0 ¼ q0 and pap0 , i.e., l ¼ p0 þ mðp0 1Þ and
k ¼ p þ mðp 1Þ.
Consequently, the non-zero elements of U~ k;l are given by
~
(1) U pþmðp1Þ;pþmðp1Þ
2
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
P
~
(2) U pþmðq1Þ;pþmðq1Þ
2
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
3
ðr2p Þ2 7
Pm 2 5;
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
P
~
0 þmðp0 1Þ
(3) U pþmðp1Þ;p
2
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
r2p r2p0 7
Pm 2 5;
j¼1 rj
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
P
r2p r2q 7
Pm 2 5:
j¼1 rj
6
¼ E4tk0
6
¼ E4tk0
6
¼ E4tk0
Note
Kma1 b
Kma1 b
6
¼ b E4tk0
j¼1
r2j Kmaþ1 b tk0
m
j¼1
r2j
tk0
Pm
j¼1
r2j Kmaþ1 b tk0
m
j¼1
r2j
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pm 2ffi
2
Km
a b tk0
j¼1 rj
Kma1 b
tk0
Pm
j¼1
r2j Kmaþ1 b tk0
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pm 2ffi
2
Km
a b tk0
j¼1 rj
that
2
Pm
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pm 2ffi
2
Km
a b tk0
j¼1 rj
1 ~
2 U pþmðp1Þ;pþmðp1Þ
jða; mÞ 2
tk0
m
j¼1
r2j
j¼1
rj
3
3
U~ pþmðq1Þ;pþmðq1Þ ¼ U~ pþmðp1Þ;p0 þmðp0 1Þ ¼
8p 8p0 8q 8q0 . Consequently, only
3
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P
Pm 2ffi
2
2 2
Kma1 b tk0 m
j¼1 rj Kmaþ1 ðb tk0
j¼1 rj Þ ðrp Þ
7
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pm 2 5
Pm 2ffi
2
j¼1 rj
Km
a ðb tk0
j¼1 rj Þ
ðF:3Þ
has to be computed and the matrix ! can be written as
!¼
jða; mÞ
4
2b
ðMT=2 M1=2 ÞðI þ vecðIÞvecT ðIÞÞ
ðMT=2 M1=2 Þ:
Note that
b2 tk0 Gða; 4Þ.
ðF:4Þ
jða; mÞ is independent of b since,
References
[1] S.M. Kay, Fundamentals of Statistical Signal Processing—Detection
Theory, vol. 2, Prentice-Hall, PTR, Englewood Cliffs, New Jersey,
1998.
[2] H.L. Van Trees, Detection Estimation and Modulation Theory, Part I,
Wiley, New York, 1968.
[3] L.L. Scharf, D.W. Lytle, Signal detection in Gaussian noise of
unknown level: an invariance application, IEEE Trans. Inf. Theory
17 (1971) 404–411.
[4] S. Haykin, Array Signal Processing, Prentice-Hall Signal Processing
Series, Prentice-Hall, Englewood Cliffs, New Jersey, 1985.
[5] H.L. Van Trees, Detection, Estimation and Modulation Theory, Part
IV: Optimum Array Processing, Wiley, New York, 2002.
[6] J.B. Billingsley, Ground clutter measurements for surface-sited
radar, Technical Report 780, MIT, February 1993.
[7] A. Farina, F. Gini, M.V. Greco, L. Verrazzani, High resolution sea
clutter data: a statistical analysis of recorded live data, IEE Proc. Part
F 144 (3) (1997) 121–130.
[8] E. Conte, M. Longo, Characterization of radar clutter as a
spherically invariant random process, IEE Proc. Part F 134 (2)
(1987) 191–197.
1175
[9] K.J. Sangston, K.R. Gerlach, Coherent detection of radar targets in a
non-Gaussian background, IEEE Trans. Aerosp. Electron. Syst. 30
(1994) 330–340.
[10] F. Gini, Sub-optimum coherent radar detection in a mixture of Kdistributed and Gaussian clutter, IEE Proc. Radar Sonar Navigation
144 (1) (1997) 39–48.
[11] F. Gini, M.V. Greco, M. Diani, L. Verrazzani, Performance analysis of two
adaptive
radar
detectors
against
non-Gaussian
real
sea
clutter data, IEEE Trans. Aerosp. Electron. Syst. 36 (4) (2000) 1429–1439.
[12] B. Picinbono, Spherically invariant and compound Gaussian
stochastic processes, IEEE Trans. Inf. Theory (1970) 77–79.
[13] K. Yao, A representation theorem and its applications to spherically
invariant random processes, IEEE Trans. Inf. Theory 19 (5) (1973)
600–608.
[14] M. Rangaswamy, D.D. Weiner, A. Ozturk, Non-Gaussian vector
identification using spherically invariant random processes, IEEE
Trans. Aerosp. Electron. Syst. 29 (1) (1993) 111–124.
[15] M. Rangaswamy, Statistical analysis of the nonhomogeneity
detector for non-Gaussian interference backgrounds, IEEE Trans.
Signal Process. 53 (6) (2005) 2101–2111.
[16] E. Jay, J.-P. Ovarlez, D. Declercq, P. Duvaut, Bord: Bayesian optimum
radar detector, Signal Process. 83 (6) (2003) 1151–1162.
[17] F. Gini, M.V. Greco, Covariance matrix estimation for CFAR detection
in correlated heavy tailed clutter, Signal Process. 82 (12) (2002)
1847–1859 (special section on SP with Heavy Tailed Distributions).
[18] E. Conte, A. DeMaio, G. Ricci, Recursive estimation of the covariance
matrix of a compound-Gaussian process and its application to
adaptive CFAR detection, IEEE Trans. Signal Process. 50 (8) (2002)
1908–1915.
[19] S. Watts, Radar detection prediction in sea clutter using the
compound K-distribution model, IEE Proc. Part F 132 (7) (1985)
613–620.
[20] T. Nohara, S. Haykin, Canada east coast trials and the K-distribution,
IEE Proc. Part F 138 (2) (1991) 82–88.
[21] S. Watts, Radar performance in K-distributed sea clutter, in: RTO
SET Symposium on ‘‘Low Grazing Angle Clutter: Its Characterisation, Measurement and Application’’, RTO MP-60, 2000, pp.
372–379.
[22] I. Antipov, Statistical analysis of northern Australian coastline sea
clutter data, Technical Report, Defence Science and Technology
Organisation, DSTO, Available at: /http://www.dsto.defence.gov.
au/corporate/reports/DSTO-TR-1236.pdfS, November 2001.
[23] D. Blacknell, Comparison of parameter estimators for K-distribution, IEE Proc. Radar Sonar Navigation 141 (1) (1994) 45–52.
[24] W.J.J. Roberts, S. Furui, Maximum likelihood estimation of Kdistribution parameters via the expectation-maximization algorithm, IEEE Trans. Signal Process. 48 (12) (2000) 3303–3306.
[25] J. Wang, A. Dogandzic, A. Nehorai, Maximum likelihood estimation
of compound-Gaussian clutter and target parameters, IEEE Trans.
Signal Process. 54 (10) (2006) 3884–3898.
[26] D.R. Raghavan, R.S., N.B. Pulsone, A generalization of the adaptive
matched filter receiver for array detection in a class of a nongaussian interference, in: Proceedings of the ASAP Workshop,
Lexinton, MA, 1996, pp. 499–517.
[27] N.B. Pulsone, Adaptive signal detection in non-gaussian interference, Ph.D. Thesis, Northeastern University, Boston, MA, May 1997.
[28] Y. Chitour, F. Pascal, Exact maximum likelihood estimates for SIRV
covariance matrix: existence and algorithm analysis, IEEE Trans.
Signal Process. 56 (10) (2008) 4563–4573.
[29] F. Gini, A radar application of a modified Cramér–Rao bound:
parameter estimation in non-Gaussian clutter, IEEE Trans. Signal
Process. 46 (7) (1998) 1945–1953.
[30] F. Pascal, P. Forster, J.-P. Ovarlez, P. Larzabal, Performance analysis of
covariance matrix estimates in impulsive noise, IEEE Trans. Signal
Process. 56 (6) (2008) 2206–2217.
[31] M. Abramowitz, I. Stegun, Handbook of Mathematical Functions,
National Bureau of Standard, AMS 55, 1964.
[32] A.W. van der Vaart, Asymptotic Statistics, Cambridge University
Press, Cambridge, 1998.
[33] K.B. Petersen, M.S. Pedersen, The matrix cookbook, Available at:
/http://matrixcookbook.comS, February 2008.
[34] M. Abramowitz, I.A. Stegun, Handbook of Mathematical Functions
with Formulas, Graphs and Mathematical Tables, Dover, New York,
1972.
Signal Processing 90 (2010) 1165–1175
Contents lists available at ScienceDirect
Signal Processing
journal homepage: www.elsevier.com/locate/sigpro
Statistical analysis of the covariance matrix MLE in
K-distributed clutter
Frédéric Pascal a,, Alexandre Renaux b
a
b
SONDRA/Supelec, 3 rue Joliot-Curie, F-91192 Gif-sur-Yvette Cedex, France
Laboratoire des Signaux et Systemes (L2S), Université Paris-Sud XI (UPS), CNRS, SUPELEC, Gif-Sur-Yvette, France
a r t i c l e in fo
abstract
Article history:
Received 22 January 2009
Received in revised form
13 July 2009
Accepted 26 September 2009
Available online 17 October 2009
In the context of radar detection, the clutter covariance matrix estimation is an
important point to design optimal detectors. While the Gaussian clutter case has been
extensively studied, the new advances in radar technology show that non-Gaussian
clutter models have to be considered. Among these models, the spherically invariant
random vector modelling is particularly interesting since it includes the K-distributed
clutter model, known to fit very well with experimental data. This is why recent results
in the literature focus on this distribution. More precisely, the maximum likelihood
estimator of a K-distributed clutter covariance matrix has already been derived. This
paper proposes a complete statistical performance analysis of this estimator through its
consistency and its unbiasedness at finite number of samples. Moreover, the closedform expression of the true Cramér–Rao bound is derived for the K-distribution
covariance matrix and the efficiency of the maximum likelihood estimator is
emphasized by simulations.
& 2009 Elsevier B.V. All rights reserved.
Keywords:
Covariance matrix estimation
Cramér–Rao bound
Maximum likelihood
K-distribution
Spherically invariant random vector
Notations
The notational convention adopted is as follows: italic
indicates a scalar quantity, as in A; lower case boldface
indicates a vector quantity, as in a; upper case boldface
indicates a matrix quantity, as in A. RefAg and I mfAg
are the real and the imaginary parts of A, respectively.
The complex conjugation, the matrix transpose operator,
and the conjugate transpose operator are indicated by , T ,
and H , respectively. The j th element of a vector a is
denoted aðjÞ . The n th row and m th column element of the
matrix A will be denoted by An;m. jAj and TrðAÞ are the
determinant and the trace of the matrix A, respectively.
denotes the Kronecker product. J J denotes any matrix
norm. The operator vec(A) stacks the columns of the
matrix A one under another into a single column vector.
The operator vech(A), where A is a symmetric matrix, does
the same things as vec(A) with the upper triangular
portion excluded. The operator veckðAÞ of a skewsymmetric matrix (i.e., AT ¼ A) does the same thing as
vechðAÞ by omitting the diagonal elements. The identity
matrix, with appropriate dimensions, is denoted I and the
zero matrix is denoted 0. E½ denotes the expectation
a:s:
operator. - stands for the almost sure convergence and
Pr
- stands for the convergence in probability. A zero-mean
complex circular Gaussian distribution with covariance
matrix A is denoted CN ð0; AÞ. A gamma distribution with
shape parameter k and scale parameter y is denoted
Gðk; yÞ. A complex m-variate K-distribution with parameters k, y, and covariance matrix A is denoted Km ðk; y; AÞ.
A central chi-square distribution with k degrees of freedom is denoted w2 ðkÞ. A uniform distribution with
boundaries a and b is denoted U ½a;b .
1. Introduction
Corresponding author. Tel.: þ33 147405320; fax: þ33 147402199.
E-mail addresses: frederic.pascal@supelec.fr (F. Pascal),
alexandre.renaux@lss.supelec.fr (A. Renaux).
0165-1684/$ - see front matter & 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.sigpro.2009.09.029
The Gaussian assumption makes sense in many
applications, e.g., sources localization in passive sonar,
ARTICLE IN PRESS
1166
F. Pascal, A. Renaux / Signal Processing 90 (2010) 1165–1175
radar detection where thermal noise and clutter are
generally modelled as Gaussian processes. In these
contexts, Gaussian models have been thoroughly investigated in the framework of statistical estimation and
detection theory (see, e.g., [1–3]). They have led to
attractive algorithms such as the stochastic maximum
likelihood method [4,5] or Bayesian estimators.
However, the assumption of Gaussian noise is not
always valid. For instance, due to the recent evolution of
radar technology, one can cite the area of space time
adaptive processing-high resolution (STAP-HR) where the
resolution is such that the central limit theorem cannot
be applied anymore since the number of backscatters is
too small. Equivalently, it is known that reflected signals
can be very impulsive when they are collected by a low
grazing angle radar [6,7]. This is why, in the last decades,
the radar community has been very interested in
problems dealing with non-Gaussian clutter modelling
(see, e.g., [8–11]).
One of the most general non-Gaussian noise model is
provided by spherically invariant random vectors (SIRV)
which are a compound processes [12–14]. More precisely,
an SIRV is the product of a Gaussian random vector (the
so-called speckle) with the square root of a non-negative
random scalar variable (the so-called texture). In other
words, a noise modelled as an SIRV is a non-homogeneous
Gaussian process with random power. Thus, these kind of
processes are fully characterized by the texture and the
unknown covariance matrix of the speckle. One of the
major challenging difficulties in SIRV modelling is to
estimate these two unknown quantities [15]. These
problems have been investigated in [16] for the texture
estimation while [17,18] have proposed different estimation procedures for the covariance matrix. Moreover, the
knowledge of these estimates accuracy is essential in
radar detection since the covariance matrix and the
texture are required to design the different detection
schemes.
In this context, this paper focuses on parameters
estimation performance where the clutter is modelled
by a K-distribution. A K-distribution is an SIRV, with a
gamma distributed texture depending on two real positive
parameters a and b. Consequently, a K-distribution
depends on a, b and on the covariance matrix M. This
model choice is justified by the fact that a lot of
operational data experimentations have shown the good
agreement between real data and the K-distribution
model (see [7,19–22] and references herein).
This K-distribution model has been extensively studied
in the literature. First, concerning the parameters estimation problem, [23,24] have estimated the gamma distribution parameters assuming that M is equal to the
identity matrix, [17] has proposed a recursive algorithm
for the covariance matrix M estimation assuming a and
b known and [25] has used a parameter-expanded
expectation-maximization (PX-EM) algorithm for the
covariance matrix M estimation and for a parameter n
assuming n ¼ a ¼ 1=b. Note also that estimation schemes
in K-distribution context can be found in [26,27] and
references herein. Second, concerning the statistical
performance of these estimators, it has been proved in
[28] that the recursive scheme proposed by [17] converges
and has a unique solution which is the maximum
likelihood (ML) estimator. Consequently, this estimator
has become very attractive. In order to evaluate the
ultimate performance in terms of mean square error, [23]
has derived the true Cramér–Rao Bound (CRB) for the
parameters of the gamma texture (namely, a and b)
assuming M equal to the identity matrix. Gini [29] has
derived the modified CRB on the one-lag correlation
coefficient of M where the parameters of the gamma
texture are assumed to be nuisance parameters. Concerning the covariance matrix M, a first approach for the true
CRB study, which is known to be tighter than the modified
one, has been proposed in [25] whatever the texture
distribution. However, note that, for the particular case of
a gamma distributed texture, the analysis of [25] involves
several numerical integrations and no useful information
concerning the structure of the Fisher information matrix
(FIM) is given. Finally, classical covariance matrix estimators are compared in [30] in the more general context of
SIRV.
The knowledge of an accurate covariance matrix
estimate is of the utmost interest in context of radar
detection since this matrix is always involved in the
detector expression [30]. Therefore, the goal of this
contribution is twofold. First, the covariance matrix ML
estimate statistical analysis is provided in terms of
consistency and bias. Second, the closed-form expression
of the true CRB for the covariance matrix M is given and is
analyzed. Finally, through a discussion and simulation
results, classical estimation procedures in Gaussian and
SIRV contexts are compared.
The paper is organized as follows. Section 2 presents
the problem formulation while Sections 3 and 4 contain
the main results of this paper: the ML estimate statistical
performance in terms of consistency and bias and the
derivation of the true CRB. Finally, Section 5 gives
simulations which validate theoretical results.
2. Problem formulation
In radar detection, the basic problem consists in
detecting if a known signal corrupted by an additive
clutter is present or not. In order to estimate the clutter
parameters before detection, it is generally assumed that
K signal-free independent measurements, traditionally
called the secondary data ck , k ¼ 1; . . . ; K are available.
As stated in the introduction, one considers a clutter
modelled thanks to a K-distribution denoted
ck Km ða; ð2=bÞ2 ; MÞ:
From the SIRV definition, ck can be written as
pffiffiffiffiffi
ck ¼ tk xk ;
ð1Þ
ð2Þ
where tk is gamma distributed with parameters a and
ð2=bÞ2 , i.e., tk Gða; ð2=bÞ2 Þ and, where xk is a complex
circular zero-mean m-dimensional Gaussian vector with
covariance matrix E½xk xH
k ¼ M independent of tk . For
identifiability considerations, M is normalized according
to TrðMÞ ¼ m (see [17]). Note that the parameter a
represents the spikiness of the clutter. Indeed, when a is
ARTICLE IN PRESS
F. Pascal, A. Renaux / Signal Processing 90 (2010) 1165–1175
high the clutter tends to be Gaussian and, when a is small,
the tail of the clutter becomes heavy.
The probability density function (PDF) of a random
variable tk distributed according to Gða; ð2=bÞ2 Þ is given by
!
!a
b2
pðtk Þ ¼
4
tak 1
b2
exp tk ;
GðaÞ
4
ð3Þ
where GðaÞ is the gamma function defined by
Z þ1
GðaÞ ¼
xa1 expðxÞ dx:
0
ð4Þ
From Eq. (2), the PDF of ck can be written
!
Z þ1
1
cH
ck
1
kM
pðtk Þ dtk ;
exp
pðck ; M; a; bÞ ¼
tm
pm jMj
tk
0
k
ð5Þ
which is equal to
pðck ; M; a; bÞ ¼
aþm
b
1
ðcH
ck ÞðamÞ=2
kM
aþm1 m
2
jMjGð Þ
p
a
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Kma b cH
M1 ck ;
k
ð6Þ
where Kn ðÞ is the modified Bessel function of the second
kind of order n [31].
Gini et al. have derived the ML estimator as the
solution of the following equation [17]:
^ ML
M
K
1X
H
^ 1
¼
cm ðcH
k M ML ck Þck ck ;
K k¼1
ð7Þ
where the function cm ðqÞ is defined as
pffiffiffi
b Kam1 ðb qÞ
cm ðqÞ ¼ pffiffiffi
ð8Þ
pffiffiffi :
2 q Kam ðb qÞ
^ ML has to be normalized as
Note that the ML estimate M
^ ML Þ ¼ m. Finally, it has been shown in [28] that
M : TrðM
the solution to Eq. (7) exists and is unique for the
aforementioned normalization.
^ ML
3. Statistical analysis of M
^ ML
This section is devoted to the statistical analysis of M
in terms of consistency and bias.
3.1. Consistency
^ of M is said to be consistent if
An estimator M
Pr
^ MJ - 0;
JM
K-þ1
where K is the number of secondary data ck ’s used to
estimate M.
^ ML is a consistent
^ ML consistency). M
Theorem 3.1 (M
estimate of M.
^ ML will be denoted MðKÞ
^
Proof. In the sequel, M
to show
^ ML and the number K
explicitly the dependence between M
of xk0 s. Let us define the function fK;M such that
8
D!D;
>
>
<
K
1X
ð9Þ
fK;M :
1
A!
cm ðcH
ck Þck cH
>
kA
k;
>
K k¼1
:
1167
where cm ðÞ is defined by Eq. (8), where D ¼ fA 2
Mm ðCÞjAH ¼ A; A positive definite matrix} with Mm ðCÞ ¼
fm m matrices with elements in Cg, and where C is the
^
set of complex scalar. As MðKÞ
is a fixed point of function
fK;M , it is the unique zero, which respects the constraint
^
TrðMðKÞÞ
¼ m, of the following function:
(
D!D;
gK :
A!gK ðAÞ ¼ A fK;M ðAÞ:
^
To prove the consistency of MðKÞ,
Theorem 5.9 of
[32, p. 46] will be used. First, the strong law of large
numbers (SLLN) gives
a:s:
gK ðAÞ - gðAÞ;
8A 2 D;
K-þ1
where
gðAÞ ¼ A E½cm ðcH A1 cÞccH
8A 2 D;
ð10Þ
for cKm ða; ð2=bÞ2 ; MÞ.
Let us now apply the change of variable y ¼ A1=2 c. We
obtain
yKm ða; ð2=bÞ2 ; A1=2 MA1=2 Þ
and
gðAÞ ¼ A1=2 ðI E½cm ðyH yÞyyH ÞA1=2
8A 2 D;
and
gK ðAÞ ¼ A1=2 I
8A 2 D;
!
K
1X
1=2
H
cm ðyH
y
Þy
y
:
k k k k A
K k¼1
Let us verify the hypothesis of Theorem 5.9 of
[32, p. 46]. We have to prove that for every e40,
Pr
ðH1 Þ : supfJgK ðAÞ gðAÞJg - 0;
K-þ1
A2D
ðH2 Þ :
inf
A:JAMJZe
fJgðAÞJg40 ¼ gðMÞ:
For every A 2 D, we have
1 X
K
H
H
H
JgK ðAÞ gðAÞJ ¼ JAJ
ðcm ðyH
y
Þy
y
E
½c
ðy
yÞyy
Þ
:
m
k k k k
K
k¼1
Since E½cm ðyH yÞyyH o þ 1, one can apply the SLLN to
H
the K i.i.d. variables cm ðyH
k yk Þyk yk , with same first order
moment. This ensures ðH1 Þ.
Moreover, the function cm ðcH A1 cÞ is strictly decreasing
w.r.t. A. Consequently, E½cm ðcH A1 cÞccH too. This implies
that E½cm ðcH A1 cÞccH aA, except for A ¼ M. This ensures
ðH2 Þ.
Finally, Theorem 5.9 of [32, p. 46] concludes the proof
Pr
^ ML M. &
and M
K-þ1
3.2. Bias
This subsection provides an analysis of the bias B
^ ML Þ ¼ E½M
^ ML M.
defined by BðM
^ ML ). M
^ ML is an unbiased
Theorem 3.2 (Unbiasedness of M
estimate of M at finite distance (i.e., at finite number K).
ARTICLE IN PRESS
1168
F. Pascal, A. Renaux / Signal Processing 90 (2010) 1165–1175
^ ML will be denoted M
^
Proof. For the sake of simplicity, M
in this part. By applying the following change of variable,
yk ¼ M1=2 ck , to Eq. (7), one has
where pðc1 ; . . . ; cK ; hÞ is the likelihood function of the
observations ck , k ¼ 1; . . . ; K. Concerning our model, the
parameter vector is
K
X
^ 1 y ÞM1=2 y yH M1=2 ;
^ ¼1
cm ðyH
M
k
k k
kT
K k¼1
h ¼ ðvechT ðRefMgÞ veckT ðI mfMgÞÞT ;
^ 1=2 :
T^ ¼ M1=2 MM
Therefore,
K
1X
^ 1 y Þy yH :
cm ðyH
T^ ¼
k k k
kT
K k¼1
T^ is thus the unique estimate (see [28, Theorem III.1]) of
^ ¼ m. Its statistic is clearly
the identity matrix, with TrðTÞ
independent of M since the yk ’s are i.i.d. SIRVs with a
gamma distributed texture and identity matrix for the
Gaussian
covariance
matrix.
In
other
words,
yk Km ða; ð2=bÞ2 ; IÞ.
Moreover, for any unitary matrix U,
K
1X
^ H 1 zk Þzk zH ;
cm ðzH
k ðUTU Þ
k
K k¼1
where zk ¼ Uyk are also i.i.d. and distributed as
^ H has the same distribution as T.
^
Km ða; ð2=bÞ2 ; IÞ and UTU
Consequently,
^ ¼ UE½TU
^ H
E½T
for any unitary matrix U:
^ is different from 0, Lemma A.1 of [30] ensures
Since E½T
^ 1=2 ,
^
that E½T ¼ gI for g 2 R. Remind that T^ ¼ M1=2 MM
^ ¼ TrðMÞ ¼ m, one
^ ¼ gM. Moreover, since TrðMÞ
then E½M
has
^ ¼ TrðEðMÞÞ
^ ¼ gTrðMÞ ¼ gm;
m ¼ EðTrðMÞÞ
ð13Þ
With the parametrization of Eq. (13), the structure of
CRB becomes
!1
F1;1 F1;2
;
ð14Þ
CRBh ¼
F2;1 F2;2
where
^ H¼
UTU
m2 1:
ð11Þ
where the Fi;j ’s are the elements of the FIM given by
"
#
@2 ln pðc1 ; . . . ; cK ; hÞ
;
F1;1 ¼ E
T
@vechðRefMgÞ@vech ðRefMgÞ
mðm þ 1Þ mðm þ 1Þ
;
ð15Þ
2
2
"
#
@2 ln pðc1 ; . . . ; cK ; hÞ
;
F1;2 ¼ FT2;1 ¼ E
T
@vechðRefMgÞ@veck ðI mfMgÞ
mðm þ 1Þ mðm 1Þ
;
ð16Þ
2
2
"
#
@2 ln pðc1 ; . . . ; cK ; hÞ
;
F2;2 ¼ E
T
@veckðI mfMgÞ@veck ðI mfMgÞ
mðm 1Þ mðm 1Þ
:
ð17Þ
2
2
Since the ck ’s are i.i.d. random vectors, one have from
Eq. (6),
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
K
Y
baþm ðcHk M1 ck ÞðamÞ=2
Kma ðb cH
pðc1 ; . . . ; cK ; hÞ ¼
M1 ck Þ:
k
aþm1 m
2
p jMjGðaÞ
k¼1
ð18Þ
Consequently, the log-likelihood function can be written as
ln pðc1 ; . . . ; cK jhÞ ¼ Kln
which implies that g ¼ 1.
^ is an unbiased estimate of M, for any
In conclusion, M
number K of secondary data. &
baþm
2aþm1 pm GðaÞ
K
X
þ
k¼1
!
KlnðjMjÞ
1
lnððcH
ck ÞðamÞ=2 Kma ðb
kM
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
M1 ck ÞÞ:
cH
k
4.1. Result
ð19Þ
3.3. Comments
Theorems 3.1 and 3.2 show the attractiveness of the
estimator (7) in terms of statistical properties, i.e.,
consistency and unbiasedness. Note also that this estimator is robust since the unbiasedness property is at finite
number of samples. In the next section, the Cramér–Rao
bound is derived for the observation model (2).
4. Cramér–Rao bound
In this Section, the Cramér–Rao bound w.r.t. M is
derived. The CRB gives the best variance that an unbiased
estimator can achieve. The proposed bound will be
compared to the mean square error of the previously
studied ML estimator (Eq. (7)) in the next section.
The CRB for a parameter vector h is given by
2
1
@ ln pðc1 ; . . . ; cK ; hÞ
CRBh ¼ E
;
ð12Þ
@h @hT
The next subsections will show that the CRB w.r.t. h, in
this context, is given by
!
T
1
0
1 ðH FHÞ
;
ð20Þ
CRBh ¼
K
0
ðKT FKÞ1
where the matrices H and K are constant transformation
matrices filled with ones and zeros such that vecðAÞ ¼
HvechðAÞ and vecðAÞ ¼ KveckðAÞ, with A a skew-symmetric matrix, and where
2ða þ 1Þ jða; mÞ
F ¼ ðMT MÞ1
mþ1
8
ðMT MÞ1=2 ðI þ vecðIÞvecT ðIÞÞðMT MÞ1=2 ;
ð21Þ
where jða; mÞ is given by Eq. (F.3).
Remarks.
jða; mÞ is a constant which does not depend on b since
b2 tk0 Gða; 4Þ (see Eq. (F.3)).
ARTICLE IN PRESS
F. Pascal, A. Renaux / Signal Processing 90 (2010) 1165–1175
The first term of the right hand side of Eq. (21) is the
Gaussian FIM (i.e., when tk ¼ 1 8k). Indeed, using
Eqs. (24), (49) and (30),1 the Gaussian CRB, denoted
GCRBh , is straightforwardly obtained as
!
T
T
1
1
0
1 ðH ðM MÞ HÞ
GCRBh ¼
:
K
0
ðKT ðMT MÞ1 KÞ1
ð22Þ
By identification with Eq. (21), it means that
2ða þ 1Þ g
¼ 0:
lim
a-1
8
mþ1
Eq. (23) is given by
1
@2 lnððcH
ck ÞðamÞ=2 Kma ðb
kM
T
@vechðRefMgÞ@vech ðRefMgÞ
2
¼
@ f ðzÞ
2
ð28Þ
T
2
T
þ ib @veck ðImfMgÞKT ðMT MÞ1 vecðck cH
kÞ
ð29Þ
and
T
M1 ÞH@vechðRefMgÞ
2
T
þ 2b @veck ðI mfMgÞKT ððMT ck cTk MT Þ
M1 ÞK@veckðI mfMgÞ
ð30Þ
and (details are given in Appendix C)
pffiffiffi
@f ðzÞ
1 Kmaþ1 ð zÞ
pffiffiffi Þ
d1 ðzÞ ¼
¼ pffiffiffi
@z
2 z Kma ð z
4.2.1. Analysis of F1;1
With Eq. (19) one has
ð31Þ
and
@2 ln pðc1 ; . . . ; cK ; hÞ
T
@vechðRefMgÞ@vech ðRefMgÞ
@ lnðjMjÞ
T
@vechðRefMgÞ@vech ðRefMgÞ
@2 f ðzÞ
@z2
pffiffiffi
pffiffiffi
1
Kmaþ1 ð zÞ 2
K
ð zÞ
pffiffiffi Þ pffiffiffi ma1 pffiffiffi Þ
:
¼
1þ
4z
Kma ð z
Kma ð z
z
d2 ðzÞ ¼
2
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
H 1 c ÞÞ
K @2 lnððcH M1 c ÞðamÞ=2 K
X
ma ðb ck M
k
k
k
:
T
@vechðRefMgÞ@vech ðRefMgÞ
k¼1
ð23Þ
The first part of the right hand side of Eq. (23) is given
ð32Þ
Consequently, the structure of Eq. (28) becomes
2
T
@2 f ðzÞ ¼ 2b d1 ðzÞ@vech ðRefMgÞHT P1 H@vechðRefMgÞ
2
T
þ 2b d1 ðzÞ@veck ðI mfMgÞKT P1 K@veckðI mfMgÞ
by
4
T
4
T
þ b d2 ðzÞ@vech ðRefMgÞHT P2 H@vechðRefMgÞ
@2 lnðjMjÞ
þ b d2 ðzÞ@veck ðI mfMgÞKT P2 K@veckðI mfMgÞ;
T
@vechðRefMgÞ@vech ðRefMgÞ
ð24Þ
thanks to Appendix A, Eq. (A.3), and thanks to the fact that
T
@veck ðI mfMgÞ
¼0
@vechðRefMgÞ
ð27Þ
@z ¼ b @vech ðRefMgÞHT ðMT MÞ1 vecðck cH
kÞ
To make the reading easier, only the outline of the CRB
derivation is given below. All the details are reported into
the different appendices.
¼ KHT ðMT MÞ1 H
:
Note that
@ @f ðzÞ
@2 f ðzÞ
@f ðzÞ 2
@2 f ðzÞ ¼
@z @z þ
@z @z ¼
@ z;
@z
@z
@z
@z2
2
K
T
@vechðRefMgÞ@vech ðRefMgÞ
@2 z ¼ 2b @vech ðRefMgÞHT ððMT ck cTk MT Þ
4.2. Outline of the proof
þ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
cH
M1 ck ÞÞ
k
with (details are given in Appendix B)
Consequently, due to the structure of I þ vecðIÞvecT ðIÞ,
the FIM for K-distributed observations is given by the
Gaussian FIM minus a sparse matrix depending on a, M
and m.
¼ K
1169
ð25Þ
ð33Þ
where
P1 ¼ ðMT ck cTk MT Þ M1 ;
ð34Þ
T
1
H
H
P2 ¼ ðMT MÞ1 vecðck cH
k Þvec ðck ck ÞðM MÞ :
ð35Þ
Therefore, Eq. (27) is given by
and
@2 f ðzÞ
T
@vech ðRefMgÞ
¼ I:
@vechðRefMgÞ
ð26Þ
Through the remain of the paper, let us set z ¼
pffiffiffi
b2 cHk M1 ck and f ðzÞ ¼ lnððz=b2 ÞðamÞ=2 Kma ð zÞÞ. The
second term (inside the sum) of the right-hand side of
T
@vechðRefMgÞ@vech ðRefMgÞ
2
4
¼ 2b d1 ðzÞHT P1 H þ b d2 ðzÞHT P2 H
ð36Þ
due to Eqs. (25) and (26).
Using Eqs. (23), (24), and (36), F1;1 is given by
2
F1;1 ¼ KHT ððMT MÞ1 þ 2b E½d1 ðzÞP1
4
þ b E½d2 ðzÞP2 ÞH;
2
1
1
With b ¼ 1, z ¼ cH
ck which is the term to be derived w.r.t. h in
kM
the exponential term of the multivariate Gaussian distribution.
ð37Þ
where d1 ðzÞ and d2 ðzÞ are defined by Eqs. (31) and (32),
respectively.
ARTICLE IN PRESS
1170
F. Pascal, A. Renaux / Signal Processing 90 (2010) 1165–1175
ðMT MÞ1=2 ðI þ vecðIÞvecT ðIÞÞðMT MÞ1=2 H;
The two expectation operators involved in the previous
equation can be detailed as follows:
E½d1 ðzÞP1 ¼ 12ðMT CMT Þ M1
ð38Þ
and
E½d2 ðzÞP2 ¼ 14ðMT MÞ1 ðW þ N !ÞðMT MÞ1 ;
ð39Þ
where
8
pffiffiffi
1 Kmaþ1 ð zÞ T
>
>
>
p
ffiffi
ffi
p
ffiffi
ffi
;
c
C
¼
E
c
>
>
z Kma ð zÞ k k
>
>
>
>
>
1
>
H
H
H
>
W
¼
E
c
Þvec
ðc
c
Þ
;
vecðc
>
k
k
<
k
k
z
pffiffiffi
>
2 K
ð zÞ
>
>
N ¼ E 3=2 maþ1pffiffiffi vecðck cHk ÞvecH ðck cHk Þ ;
>
>
z
>
Kma ð zÞ
>
>
pffiffiffi
pffiffiffi
>
>
>
1 Kma1 ð zÞKmaþ1 ð zÞ
>
H
H
>
p
ffiffi
ffi
vecðck cH
!
¼
E
:
k Þvec ðck ck Þ:
2
z
Km
a ð zÞ
4.2.2. Analysis of F2;2
The analysis of F2;2 is similar to the one used for F1;1.
Indeed, one has to calculate
@2 ln pðc1 ; . . . ; cK ; hÞ
T
@veckðI mfMgÞ@veck ðI mfMgÞ
2
¼ K
þ
C¼
2
b2
MT :
ð41Þ
Concerning W, one has
"
#
H
H
vecðck cH
1
k Þvec ðck ck Þ
;
W¼ 2E
cH
M1 ck
b
k
T
ð44Þ
where the expectation is taken under a complex
K-distribution Km ða; ð2=bÞ2 ; MÞ.
H
H
The closed-form expression of E vecðck cH
k Þvec ðck ck Þ=
1
H
ck M ck under a complex K-distribution is given in
Appendix E. One finds
a
b4 m þ 1
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
cH
M1 ck ÞÞ
k
@veckðI mfMgÞ@veck ðI mfMgÞ
k¼1
:
ð48Þ
Using Eq. (A.3), one has
K
@2 lnðjMjÞ
T
@veckðI mfMgÞ@veck ðI mfMgÞ
¼ KKT ðMT MÞ1 K
@veckðI mfMgÞ
ð42Þ
where the expectation is taken under a complex
K-distribution Km ða 1; ð2=bÞ2 ; MÞ.
Concerning !, one has
"
#
pffiffiffi
pffiffiffi
H
H
1
K
ð zÞKmaþ1 ð zÞ vecðck cH
k Þvec ðck ck Þ
ffiffiffi
;
! ¼ 2 E ma1 2 p
Kma ð zÞ
cH
M1 ck
b
k
8
T
ð49Þ
due to Eq. (25) and due to
where the expectation is taken under a complex
K-distribution Km ða; ð2=bÞ2 ; MÞ.
Concerning N, one has
"
#
H
H
vecðck cH
1
k Þvec ðck ck Þ
;
ð43Þ
N¼
E
2
cH
M1 ck
ða 1Þb
k
W¼
@ lnðjMjÞ
@veckðI mfMgÞ@veck ðI mfMgÞ
K @2 lnððcH M1 c ÞðamÞ=2 K
X
ma ðb
k
k
ð40Þ
After some calculus detailed in Appendix D, one finds
ð47Þ
where jða; mÞ is given by Eq. (F.3).
ðMT=2 M1=2 ÞðI þ vecðIÞvecT ðIÞÞðMT=2 M1=2 Þ
ð45Þ
T
@veck ðI mfMgÞ
¼ I:
ð50Þ
By using the same notation as for the derivation of F1;1
and by using Eq. (33), one obtains for the second term on
the right hand side of Eq. (48) as
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1
@2 lnððcH
ck ÞðamÞ=2 Kma ðb cH
M1 ck ÞÞ
kM
k
T
@veckðI mfMgÞ@veck ðI mfMgÞ
2
4
¼ 2b d1 ðzÞKT P1 K þ b d2 ðzÞKT P2 K;
ð51Þ
where P1 , P2 , d1 ðzÞ and d2 ðzÞ are defined by Eqs. (34), (35),
(31), and (32), respectively. Therefore, the structure of F2;2
is the same as the structure of F1;1 except that one
replaces the matrix H by the matrix K.
4.2.3. Analysis of F1;2 ¼ FT2;1
Due to the structure of Eq. (A.3) and Eq. (33), and since
T
the derivation is w.r.t. @vechðRefMgÞ and @veck ðI mfMgÞ, it
is clear that
F1;2 ¼ FT2;1 ¼ 0
ð52Þ
by using Eq. (25).
This concludes the proof of the CRB derivation.
5. Simulation results
and
N¼
8
1
ðMT=2 M1=2 ÞðI þ vecðIÞvecT ðIÞÞðMT=2 M1=2 Þ:
b4 m þ 1
ð46Þ
The structure of ! is analyzed in Appendix F.
Consequently, Eq. (37) is reduced to
2ða þ 1Þ jða; mÞ
F1;1 ¼ KHT ðMT MÞ1
mþ1
8
In this section, some simulations are provided in order
to illustrate the proposed previous results in terms of
consistency, bias, and variance analysis (throughout the
CRB). While no mathematical proof is given in other
sections, we show the efficiency of the MLE meaning that
the CRB is achieved by the variance of the MLE.
The results presented are obtained for complex
K-distributed clutter with covariance matrix M randomly
chosen.
ARTICLE IN PRESS
F. Pascal, A. Renaux / Signal Processing 90 (2010) 1165–1175
All the results are presented different values of
pffiffiffi
parameter a and b ¼ 2 a following the scenario of [17].
The size of each vector ck is m ¼ 3. Remember that the
parameter a represents the spikiness of the clutter (when
a is high the clutter tends to be Gaussian and, when a is
small, the tail of the clutter becomes heavy). The norm
used for consistency and bias is the L2 norm.
1171
5.1. Consistency
Fig. 1 presents results of MLE consistency for 1000
Monte Carlo runs per each value of K. For that purpose, a
^ KÞ ¼ JM
^ MJ versus the number K of ck ’s is
plot of DðM;
presented for each estimate. It can be noticed that the
^ KÞ tends to 0 when K tends to 1 for
above criterion DðM;
^ KÞ versus the number of secondary data for different values of a.
Fig. 1. DðM;
^ KÞ versus the number of secondary data for different values of a.
Fig. 2. CðM;
ARTICLE IN PRESS
1172
F. Pascal, A. Renaux / Signal Processing 90 (2010) 1165–1175
0.09
Empirical MSE α =1
CRB α = 1
Empirical MSE α = 0.1
CRB α = 0.1
Gaussian CRB
0.08
0.07
MSE
0.06
0.05
0.04
0.03
0.02
0.01
100
200
300
400
500
600
700
Number of secondary data
800
900
1000
Fig. 3. GCRB, CRB and empirical variance of the MLE for different values of a.
each estimate. Moreover, note that the parameter a has
very few influence on the convergence speed which
highlights the robustness of the MLE.
5.2. Bias
Fig. 2 shows the bias of each estimate for the different
values of a. The number of Monte Carlo runs is given in
the legend of the figure.
For that purpose, a plot of the
___
^ KÞ ¼ JM
^ MJ versus
the number K of ck ’s is
criterion CðM;
___
^ is defined as the empirical
presented for each estimate. M
^
mean of the quantities MðiÞ
obtained from I Monte Carlo
runs. For each iteration i, a new set of K secondary data ck
^
is generated to compute MðiÞ.
It can be noticed that,
as enlightened by the previous theoretical analysis, the
^ tends to 0 whatever the value of K. Furthermore,
bias of M
one sees again the weak influence of the parameter a on
the unbiasedness of the MLE.
(at finite number of samples) have been proved. In order
to analyze the variance of the estimator, the Cramér–Rao
lower bound is derived. The Fisher information matrix in
this case is simply the Fisher information matrix of the
Gaussian case plus a term depending on the tail of the
K-distribution. Simulation results have been proposed to
illustrate these theoretical analyses. These results have
shown the efficiency of the estimator and the weak
influence of the spikiness parameter in terms of consistency and bias.
Appendix A. Derivation of @2 lnfjMjg
To find this term and several other, we will use the
following results [33]:
@TrðXÞ ¼ Trð@XÞ;
ðA:1aÞ
@vecðXÞ ¼ vecð@XÞ;
ðA:1bÞ
5.3. CRB and MSE
@A1 ¼ A1 @AA1 ;
ðA:1cÞ
The CRB and empirical variance of the MLE for 10 000
Monte Carlo runs are plotted in Fig. 3. For comparison, we
also plot the Gaussian CRB (i.e., when tk ¼ 1 8k) given by
Eq. (22). Although it is not mathematically proved in this
paper, one observes, in Fig. 3, the efficiency of the MLE
even for impulsive noise (a small).
@jAj ¼ jAj TrðA1 @AÞ;
ðA:1dÞ
6. Conclusion
In this paper, a statistical analysis of the maximum
likelihood estimator of the covariance matrix of a complex
multivariate K-distributed process has been proposed.
More particularly, the consistency and the unbiasedness
@ðA
BÞ ¼ @ðAÞ
BþA
@ðBÞ
where
¼ or
ðA:1eÞ
@lnðjMjÞ ¼ TrðM1 @MÞ;
ðA:1fÞ
@ð@MÞ ¼ 0;
ðA:1gÞ
TrðABÞ ¼ TrðBAÞ
ðA:1hÞ
M1 M1 ¼ ðM MÞ1 ;
ðA:1iÞ
TrðAH BÞ ¼ vecH ðAÞvecðBÞ;
ðA:1jÞ
ARTICLE IN PRESS
F. Pascal, A. Renaux / Signal Processing 90 (2010) 1165–1175
vecðABCÞ ¼ ðCT AÞvecðBÞ:
@lnðjMjÞ ¼ TrðM
2
1
@MM1 ck Þ
@z ¼ b TrðcH
kM
from ðA:1fÞ
ð@MÞÞ
Concerning @2 z, one has
ðA:1kÞ
By using these properties, one has
1
2
2
from ðA:1aÞðA:1eÞðA:1gÞðA:1cÞ
¼ vecH ðM1 ð@MÞM1 Þvecð@MÞ
¼ b
2
¼ 2b Trð@MM
from ðA:1kÞ
ðM MÞ
1
M1 Þ@vecðMÞ
1
ck cH
Þ
kM
from ðA:1cÞðA:1hÞ
from ðA:1jÞðA:1kÞðA:1bÞ
2
T
@2 z ¼ 2b @vech ðRefMgÞHT ððMT ck cTk MT ÞM1 ÞH@vechðRefMgÞ
@vecðRefMg þ iI mfMgÞ
2
T
þ2b @veck ðI mfMgÞKT ððMT ck cTk MT Þ
M1 ÞK@veckðI mfMgÞ:
@vecH ðiI mfMgÞðMT MÞ1 @vecðRefMgÞ
@vecH ðRefMgÞðMT MÞ1 @vecðiI mfMgÞ
@vecH ðiI mfMgÞðMT MÞ1 @vecðiI mfMgÞ:
ðA:2Þ
Since vecH ðiI mfMgÞ ¼ ivecT ðI mfMgÞ and vecH ðRefMgÞ
¼ vecT ðRefMgÞ, @2 lnðjMjÞ is reduced to
@2 lnðjMjÞ ¼ @vecT ðRefMgÞðMT MÞ1 @vecðRefMgÞ
@vecT ðI mfMgÞðMT MÞ1 @vecðI mfMgÞ
¼ @vech ðRefMgÞH
from ðA:1eÞðA:1gÞ
By letting M ¼ RefMg þ iI mfMg, one obtains
¼ @vecH ðRefMgÞðMT MÞ1 @vecðRefMgÞ
T
@MM
1
¼ 2b @vecH ðMÞððMT ck cTk MT Þ
By letting M ¼ RefMg þ iI mfMg in Eq. (A.3), one has
@ lnðjMjÞ ¼ @vec ðRefMg þ iI mfMgÞ
1
2
¼ @vecH ðMÞðMT MÞ1 @vecðMÞ
from ðA:1bÞðA:1iÞ:
H
from ðA:1aÞ
1
TrðcH
Þ@MM1 ck
k @ðM
1
þ cH
@M@ðM1 Þck Þ
kM
from ðA:1jÞ
¼ vecH ð@MÞðMT M1 ÞH vecð@MÞ
T
ðB:2Þ
1
@2 z ¼ b TrðcH
@MM1 Þck Þ
k @ðM
@2 lnðjMjÞ ¼ TrðM1 ð@MÞM1 @MÞ
2
1173
T
ðMT MÞ1 H@vechðRefMgÞ
T
@veck ðI mfMgÞKT ðMT MÞ1 K@veckðI mfMgÞ;
ðA:3Þ
where the matrix H and K are constant transformation
matrices filled with ones and zeros such that vecðAÞ ¼
HvechðAÞ and vecðAÞ ¼ KveckðAÞ.
Appendix B. Derivation of @z and @2 z
ðB:3Þ
Appendix C. Derivation of @f ðzÞ=@z and @2 f ðzÞ=@z2
Concerning the first derivative of f ðzÞ, one has
pffiffiffi
@f ðzÞ
@
2
¼ lnððz=b ÞðamÞ=2 Kma ð zÞÞ
@z
@z
pffiffiffi
am
1
@Kma ð zÞ
pffiffiffi
¼
þ
:
2z
@z
Kma ð zÞ
ðC:1Þ
Since @Kn ðyÞ=@y ¼ Knþ1 ðyÞ þ ðn=yÞKn ðyÞ [34]. It follows
that
pffiffiffi
pffiffiffi
pffiffiffi
@Kma ð zÞ
1
ma
Kma ð zÞ:
ðC:2Þ
¼ pffiffiffi Kmaþ1 ð zÞ þ
2z
@z
2 z
Plugging Eq. (C.2) in Eq. (C.1), one obtains
pffiffiffi
@f ðzÞ
1 Kmaþ1 ð zÞ
pffiffiffi :
¼ pffiffiffi
@z
2 z Kma ð zÞ
ðC:3Þ
Concerning the second derivative of f ðzÞ, one has
pffiffiffi
@ f ðzÞ
1 @
1 K
ð zÞ
pffiffiffi maþ1 pffiffiffi Þ
¼
2
2 @z
@z
z Kma ð z
!
pffiffiffi
1 Kmaþ1 ð zÞ
1 @ Kmaþ1 ðyÞ
pffiffiffi
: ðC:4Þ
¼ 3=2
4z @y Kma ðyÞ
y¼pffiffiz
4z
Kma ð zÞ
2
By using the properties from Eq. (A.1a) to (A.1k), one
has for @z
2
2
1
1
ck Þ ¼ b @TrðcH
ck Þ
@z ¼ b @ðcH
kM
kM
2
1
¼ b TrðcH
Þck Þ
k @ðM
2
from ðA:1aÞ
1
TrðcH
@MM1 ck Þ
kM
from ðA:1cÞ
1
¼ b Trð@MM1 ck cH
Þ
kM
from ðA:1hÞ
¼ b
2
2
1
¼ b vecH ð@MÞvecðM1 ck cH
Þ
kM
2
Since @Kn ðyÞ=@y ¼ Kn1 ðyÞ ðn=yÞKn ðyÞ [34]. It follows
that
@ Kmaþ1 ðyÞ
@y Kma ðyÞ
y¼pffiffiz
from ðA:1jÞ
pffiffiffi
pffiffiffi
pffiffiffi
maþ1
pffiffiffi
Kmaþ1 ð zÞÞKma ð zÞ
ðKma ð zÞ þ
z
pffiffiffi
¼
2
Km
a ð zÞ
pffiffiffi
pffiffiffi
pffiffiffi
ma
Kmaþ1 ð zÞðKma1 ð zÞ þ pffiffiffi Kma ð zÞÞ
z
pffiffiffi
:
þ
2
Km
a ð zÞ
H
¼ b @vec ðMÞ
ðMT MÞ1 vecðck cH
kÞ
from ðA:1bÞðA:1kÞðA:1iÞ:
By letting M ¼ RefMg þ iI mfMg, one obtains
2
@z ¼ b @vecT ðRefMgÞðMT MÞ1 vecðck cH
kÞ
2
þ ib @vecT ðI mfMgÞðMT MÞ1 vecðck cH
kÞ
2
T
¼ b @vech ðRefMgÞHT ðMT MÞ1 vecðck cH
kÞ
2
T
þ ib @veck ðI mfMgÞKT ðMT MÞ1 vecðck cH
k Þ:
ðB:1Þ
ðC:5Þ
Plugging Eq. (C.5) in Eq. (C.4), one obtains
pffiffiffi
pffiffiffi
pffiffiffi
@2 f ðzÞ
1
1 Kmaþ1 ð zÞ 1 Kmaþ1 ð zÞKma1 ð zÞ
pffiffiffi
pffiffiffi
¼
:
þ 3=2
2
2
4z 2z
@z
Kma ð zÞ 4z
Kma ð zÞ
ðC:6Þ
ARTICLE IN PRESS
1174
F. Pascal, A. Renaux / Signal Processing 90 (2010) 1165–1175
Appendix D. Derivation of matrix C
since
xH
k xk ¼
The matrix C is given by
pffiffiffi
1 Kmaþ1 ð zÞ T
Gða 1Þ T
pffiffiffi ck ck ¼
E ½ck cH
k ;
2GðaÞ
z Kma ð zÞ
C ¼ E pffiffiffi
¼
ðD:1Þ
Kmaþ1 ðb
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
M1 ck Þ;
cH
k
ðD:2Þ
which is a complex K-distribution Km ða 1; ð2=bÞ2 ; MÞ.
Then
C¼
¼
1
1
ET ½ck cH
ET ½tk xk xH
k ¼
k
2ða 1Þ
2ða 1Þ
1
E½tk ET ½xk xH
k ;
2ða 1Þ
ðD:3Þ
2
where E½tk ¼ ða 1Þð2=bÞ since tk follows a Gamma
T
distribution Gða 1; ð2=bÞ2 Þ and ET ½xk xH
k ¼ M since xk is
a complex normal random vector (independent of tk ) with
zero mean and covariance matrix M. Consequently,
C ¼ ð2=b2 ÞMT .
E
M1 ck
cH
k
"
E
¼ ðM
1=2
M
Þ
#
H
H
vecðyk yH
k Þvec ðyk yk Þ
ðMT=2 M1=2 Þ;
yH
y
k k
ðE:1Þ
ðE:2Þ
where E½tk ¼ að2=bÞ2 and where
"
#
H
H
vecðxk xH
k Þvec ðxk xk Þ
R¼E
:
ðE:3Þ
xH
x
k k
qffiffiffiffiffiffi
Let us set xðjÞ
¼ r2j expðiyj Þ for j ¼ 1; . . . ; m. Note that
k
r2j w2 ð2Þ is independent of yj U ½0;2p . Consequently, the
elements of the matrix R can be rewritten as
2qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi3
Rk;l ¼ E4
r2p r2q r2p0 r2q0
Pm
j¼1
r2j
5E½expðiðyp yp0 þ yq0 yq ÞÞ
r2p r2p0 expðiðyp yp0 ÞÞ:
ðE:5Þ
(1) p ¼ p0 ¼ q ¼ q0 , i.e., l ¼ k ¼ p þ mðp 1Þ,
(2) p ¼ p0 , q ¼ q0 and paq, i.e., l ¼ k ¼ p þ mðq 1Þ,
(3) p ¼ q, p0 ¼ q0 and pap0 , i.e., l ¼ p0 þ mðp0 1Þ and
k ¼ p þ mðp 1Þ.
Consequently, the non-zero elements of Rk;l are given by
P
(1) Rpþmðp1Þ;pþmðp1Þ ¼ E½ðr2p Þ2 = m
r2 ¼ 4=ðm þ 1Þ,
Pj¼1 j2
r ¼ 2=ðm þ 1Þ,
(2) Rpþmðq1Þ;pþmðq1Þ ¼ E½r2p r2p0 = m
Pj¼1 j 2
(3) Rpþmðp1Þ;p0 þmðp0 1Þ ¼ E½r2p r2q = m
j¼1 rj ¼ 2=ðm þ 1Þ,
H
H
H 1
and the matrix E½vecðck cH
ck can be
k Þvec ðck ck Þ=ck M
written as
"
#
2
H
H
vecðck cH
2a
2
k Þvec ðck ck Þ
¼
E
mþ1 b
cH M1 ck
k
ðMT=2 M1=2 Þ
ðI þ vecðIÞvecT ðIÞÞðMT=2 M1=2 Þ:
1
H
ck where ck Km ða 1; ð2=bÞ2 ; MÞ is
vecH ðck cH
k Þ=ck M
given by
"
#
H
H
vecðck cH
2ða 1Þ 2 2
k Þvec ðck ck Þ
E
¼
mþ1 b
cH
M1 ck
k
ðMT=2 M1=2 Þ
ðI þ vecðIÞvecT ðIÞÞðMT=2 M1=2 Þ:
ðE:7Þ
Appendix F. Analysis of !
pffiffiffiffiffi
where yk Km ða; ð2=bÞ2 ; IÞ. Since yk ¼ tk xk , where
2
tk Gða; ð2=bÞ Þ is independent of xk CN ð0; IÞ, one has
"
#
H
H
vecðck cH
k Þvec ðck ck Þ
E
¼ ðMT=2 M1=2 ÞE½tk R
1
cH
M
c
k
k
ðMT=2 M1=2 Þ;
qffiffiffiffiffiffiffiffiffiffiffiffi
In the same way, the expression of E½vecðck cH
kÞ
In this appendix, we derive the expression of E½vecðck cH
kÞ
1
2
H
Þ=c
M
c
where
c
K
ð
a
;
ð2=
b
Þ
;
MÞ.
The
case
vecH ðck cH
m
k
k
k
k
where ck Km ða 1; ð2=bÞ2 ; MÞ will be, of course, straightforward. Let us set the following change of variable:
ck ¼ M1=2 yk . One obtains from Eq. (A.1k)
T=2
r2j and ½vecðxk xHk Þn
ðE:6Þ
Appendix E. Derivation of
H
H
H 1
E½vecðck cH
k Þvec ðck ck Þ=ck M ck
"
#
H
H
vecðck cH
k Þvec ðck ck Þ
j¼1
Note that E½expðiðyp yp0 þ yq0 yq ÞÞa0 if and only if
where the last expectation is taken under the distribution
baþm1 jMj1
pðck Þ ¼ aþm2
ðcH M1 ck Þðam1Þ=2
2
pm Gða 1Þ k
m
X
Let us set the following change of variable: ck ¼
pffiffiffiffiffi
M1=2 tk xk , where tk Gða; ð2=bÞ2 Þ is independent of
xk CN ð0; IÞ, one has
!¼
~ ðMT=2 M1=2 Þ;
ðMT=2 M1=2 Þ!
where
2
!~ ¼ E6
4
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
3
tk Kma1 ðb tk xHk xk ÞKmaþ1 ðb tk xHk xk Þ vecðxk xH ÞvecH ðxk xH Þ7
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
H
2
Km
a ðb tk xk xk Þ
Let us set
xðjÞ
k
k
xH
x
k k
k
5:
ðF:1Þ
qffiffiffiffiffiffi
2
2
2
¼ rj expðiyj Þ for j ¼ 1; . . . ; m. rj w ð2Þ is
independent of yj U ½0;2p . Consequently, due to Eq. (E.5),
~ can be rewritten as
the elements of the matrix !
U~
ðE:4Þ
1
b2
k;l
2
6
¼ E4tk0
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
ffi3
P
Pm 2ffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2
Kma1 b tk0 m
r2p r2q r2p0 r2q0 7
j¼1 rj Kmaþ1 b tk0
j¼1 rj
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pm 2 5
Pm 2ffi
2
j¼1 rj
Km
a b tk0
j¼1 rj
E½expðiðyp yp0 þ yq0 yq ÞÞ:
ðF:2Þ
ARTICLE IN PRESS
F. Pascal, A. Renaux / Signal Processing 90 (2010) 1165–1175
As before, E½expðiðyp yp0 þ yq0 yq ÞÞa0 if and only if
(1) p ¼ p0 ¼ q ¼ q0 , i.e., l ¼ k ¼ p þ mðp 1Þ,
(2) p ¼ p0 , q ¼ q0 and paq, i.e., l ¼ k ¼ p þ mðq 1Þ,
(3) p ¼ q, p0 ¼ q0 and pap0 , i.e., l ¼ p0 þ mðp0 1Þ and
k ¼ p þ mðp 1Þ.
Consequently, the non-zero elements of U~ k;l are given by
~
(1) U pþmðp1Þ;pþmðp1Þ
2
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
P
~
(2) U pþmðq1Þ;pþmðq1Þ
2
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
3
ðr2p Þ2 7
Pm 2 5;
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
P
~
0 þmðp0 1Þ
(3) U pþmðp1Þ;p
2
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
r2p r2p0 7
Pm 2 5;
j¼1 rj
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
P
r2p r2q 7
Pm 2 5:
j¼1 rj
6
¼ E4tk0
6
¼ E4tk0
6
¼ E4tk0
Note
Kma1 b
Kma1 b
6
¼ b E4tk0
j¼1
r2j Kmaþ1 b tk0
m
j¼1
r2j
tk0
Pm
j¼1
r2j Kmaþ1 b tk0
m
j¼1
r2j
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pm 2ffi
2
Km
a b tk0
j¼1 rj
Kma1 b
tk0
Pm
j¼1
r2j Kmaþ1 b tk0
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pm 2ffi
2
Km
a b tk0
j¼1 rj
that
2
Pm
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pm 2ffi
2
Km
a b tk0
j¼1 rj
1 ~
2 U pþmðp1Þ;pþmðp1Þ
jða; mÞ 2
tk0
m
j¼1
r2j
j¼1
rj
3
3
U~ pþmðq1Þ;pþmðq1Þ ¼ U~ pþmðp1Þ;p0 þmðp0 1Þ ¼
8p 8p0 8q 8q0 . Consequently, only
3
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P
Pm 2ffi
2
2 2
Kma1 b tk0 m
j¼1 rj Kmaþ1 ðb tk0
j¼1 rj Þ ðrp Þ
7
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pm 2 5
Pm 2ffi
2
j¼1 rj
Km
a ðb tk0
j¼1 rj Þ
ðF:3Þ
has to be computed and the matrix ! can be written as
!¼
jða; mÞ
4
2b
ðMT=2 M1=2 ÞðI þ vecðIÞvecT ðIÞÞ
ðMT=2 M1=2 Þ:
Note that
b2 tk0 Gða; 4Þ.
ðF:4Þ
jða; mÞ is independent of b since,
References
[1] S.M. Kay, Fundamentals of Statistical Signal Processing—Detection
Theory, vol. 2, Prentice-Hall, PTR, Englewood Cliffs, New Jersey,
1998.
[2] H.L. Van Trees, Detection Estimation and Modulation Theory, Part I,
Wiley, New York, 1968.
[3] L.L. Scharf, D.W. Lytle, Signal detection in Gaussian noise of
unknown level: an invariance application, IEEE Trans. Inf. Theory
17 (1971) 404–411.
[4] S. Haykin, Array Signal Processing, Prentice-Hall Signal Processing
Series, Prentice-Hall, Englewood Cliffs, New Jersey, 1985.
[5] H.L. Van Trees, Detection, Estimation and Modulation Theory, Part
IV: Optimum Array Processing, Wiley, New York, 2002.
[6] J.B. Billingsley, Ground clutter measurements for surface-sited
radar, Technical Report 780, MIT, February 1993.
[7] A. Farina, F. Gini, M.V. Greco, L. Verrazzani, High resolution sea
clutter data: a statistical analysis of recorded live data, IEE Proc. Part
F 144 (3) (1997) 121–130.
[8] E. Conte, M. Longo, Characterization of radar clutter as a
spherically invariant random process, IEE Proc. Part F 134 (2)
(1987) 191–197.
1175
[9] K.J. Sangston, K.R. Gerlach, Coherent detection of radar targets in a
non-Gaussian background, IEEE Trans. Aerosp. Electron. Syst. 30
(1994) 330–340.
[10] F. Gini, Sub-optimum coherent radar detection in a mixture of Kdistributed and Gaussian clutter, IEE Proc. Radar Sonar Navigation
144 (1) (1997) 39–48.
[11] F. Gini, M.V. Greco, M. Diani, L. Verrazzani, Performance analysis of two
adaptive
radar
detectors
against
non-Gaussian
real
sea
clutter data, IEEE Trans. Aerosp. Electron. Syst. 36 (4) (2000) 1429–1439.
[12] B. Picinbono, Spherically invariant and compound Gaussian
stochastic processes, IEEE Trans. Inf. Theory (1970) 77–79.
[13] K. Yao, A representation theorem and its applications to spherically
invariant random processes, IEEE Trans. Inf. Theory 19 (5) (1973)
600–608.
[14] M. Rangaswamy, D.D. Weiner, A. Ozturk, Non-Gaussian vector
identification using spherically invariant random processes, IEEE
Trans. Aerosp. Electron. Syst. 29 (1) (1993) 111–124.
[15] M. Rangaswamy, Statistical analysis of the nonhomogeneity
detector for non-Gaussian interference backgrounds, IEEE Trans.
Signal Process. 53 (6) (2005) 2101–2111.
[16] E. Jay, J.-P. Ovarlez, D. Declercq, P. Duvaut, Bord: Bayesian optimum
radar detector, Signal Process. 83 (6) (2003) 1151–1162.
[17] F. Gini, M.V. Greco, Covariance matrix estimation for CFAR detection
in correlated heavy tailed clutter, Signal Process. 82 (12) (2002)
1847–1859 (special section on SP with Heavy Tailed Distributions).
[18] E. Conte, A. DeMaio, G. Ricci, Recursive estimation of the covariance
matrix of a compound-Gaussian process and its application to
adaptive CFAR detection, IEEE Trans. Signal Process. 50 (8) (2002)
1908–1915.
[19] S. Watts, Radar detection prediction in sea clutter using the
compound K-distribution model, IEE Proc. Part F 132 (7) (1985)
613–620.
[20] T. Nohara, S. Haykin, Canada east coast trials and the K-distribution,
IEE Proc. Part F 138 (2) (1991) 82–88.
[21] S. Watts, Radar performance in K-distributed sea clutter, in: RTO
SET Symposium on ‘‘Low Grazing Angle Clutter: Its Characterisation, Measurement and Application’’, RTO MP-60, 2000, pp.
372–379.
[22] I. Antipov, Statistical analysis of northern Australian coastline sea
clutter data, Technical Report, Defence Science and Technology
Organisation, DSTO, Available at: /http://www.dsto.defence.gov.
au/corporate/reports/DSTO-TR-1236.pdfS, November 2001.
[23] D. Blacknell, Comparison of parameter estimators for K-distribution, IEE Proc. Radar Sonar Navigation 141 (1) (1994) 45–52.
[24] W.J.J. Roberts, S. Furui, Maximum likelihood estimation of Kdistribution parameters via the expectation-maximization algorithm, IEEE Trans. Signal Process. 48 (12) (2000) 3303–3306.
[25] J. Wang, A. Dogandzic, A. Nehorai, Maximum likelihood estimation
of compound-Gaussian clutter and target parameters, IEEE Trans.
Signal Process. 54 (10) (2006) 3884–3898.
[26] D.R. Raghavan, R.S., N.B. Pulsone, A generalization of the adaptive
matched filter receiver for array detection in a class of a nongaussian interference, in: Proceedings of the ASAP Workshop,
Lexinton, MA, 1996, pp. 499–517.
[27] N.B. Pulsone, Adaptive signal detection in non-gaussian interference, Ph.D. Thesis, Northeastern University, Boston, MA, May 1997.
[28] Y. Chitour, F. Pascal, Exact maximum likelihood estimates for SIRV
covariance matrix: existence and algorithm analysis, IEEE Trans.
Signal Process. 56 (10) (2008) 4563–4573.
[29] F. Gini, A radar application of a modified Cramér–Rao bound:
parameter estimation in non-Gaussian clutter, IEEE Trans. Signal
Process. 46 (7) (1998) 1945–1953.
[30] F. Pascal, P. Forster, J.-P. Ovarlez, P. Larzabal, Performance analysis of
covariance matrix estimates in impulsive noise, IEEE Trans. Signal
Process. 56 (6) (2008) 2206–2217.
[31] M. Abramowitz, I. Stegun, Handbook of Mathematical Functions,
National Bureau of Standard, AMS 55, 1964.
[32] A.W. van der Vaart, Asymptotic Statistics, Cambridge University
Press, Cambridge, 1998.
[33] K.B. Petersen, M.S. Pedersen, The matrix cookbook, Available at:
/http://matrixcookbook.comS, February 2008.
[34] M. Abramowitz, I.A. Stegun, Handbook of Mathematical Functions
with Formulas, Graphs and Mathematical Tables, Dover, New York,
1972.