Any correspondence concerning this service should be sent to the repository
administrator:
staff-oatao@listes-diff.inp-toulouse.fr
O
pen
A
rchive
T
OULOUSE
A
rchive
O
uverte (
OATAO
)
OATAO is an open access repository that collects the work of Toulouse researchers and
makes it freely available over the web where possible.
This is an author-deposited version published in:
http://oatao.univ-toulouse.fr/
Eprints ID: 16657
To cite this version:
Chaumette, Eric and Vincent, François and Besson, Olivier On
ordered normally distributed vector parameter estimates. (2015) Signal Processing, vol.
115. pp. 20-26. ISSN 0165-1684
On ordered normally distributed vector parameter estimates
Eric Chaumette, François Vincent, Olivier Besson
University of Toulouse-ISAE, Department of Electronics, Optronics and Signal, 10 Avenue Edouard Belin, 31055 Toulouse, France
Keywords:
Multivariate normal distribution Order statistics
Mean square error
Maximum likelihood estimator
a b s t r a c t
The ordered values of a sample of observations are called the order statistics of the sample and are among the most important functions of a set of random variables in probability and statistics. However the study of ordered estimates seems to have been overlooked in maximum-likelihood estimation. Therefore it is the aim of this communication to give an insight into the relevance of order statistics in maximum-likelihood estimation by providing a second-order statistical prediction of ordered normally distributed estimates. Indeed, this second-order statistical prediction allows to refine the asymptotic perfor-mance analysis of the mean square error (MSE) of maximum likelihood estimators (MLEs) of a subset of the parameters. A closer look to the bivariate case highlights the possible impact of estimates ordering on MSE, impact which is not negligible in (very) high resolution scenarios.
1. Introduction
The ordered values of a sample of observations are called the order statistics of the sample: if
θ
¼θ
1;θ
2; …;θ
MT is avector1 of M real valued random variables, then
θ
M ð Þ¼
θ
ð Þ1;θ
ð Þ2; …;θ
ð ÞMT
denotes the vector of order statistics induced by
θ
whereθ
ð Þ1rθ
ð Þ2r⋯rθ
ð ÞM [1,2]. Order statis-tics and extremes (smallest and largest values) are among themost important functions of a set of random variables in probability and statistics. There is natural interest in studying the highs and lows of a sequence, and the other order statistics help in understanding the concentration of probability in a distribution, or equivalently, the diversity in the population represented by the distribution. Order statistics are also useful in statistical inference, where estimates of parameters are often based on some suitable functions of the order statistics vector (robust location estimates, detection of outliers, cen-sored sampling, characterizations, goodness of fit, etc.). How-ever the study of ordered estimates seems to have been
overlooked in maximum-likelihood estimation[4], which is
at first sight a little bit surprising. Indeed, if x denotes the
random observation vector and p x ;
Θ
denotes theprob-ability density function (p.d.f.) of x depending on a vector of P
real unknown parameters
Θ
¼Θ
1; …;Θ
P
to be estimated, then many estimation problems in this setting lead to estima-tion algorithms yielding ordered estimates b
θ
ð ÞM induced by a vector bθ
of M estimates formed from a subset of the whole setof P estimates b
Θ
. Among all various possible instances ofthis setting, the most studied in signal processing is that of separating the components of data formed from a linear superposition of individual signals and noise (nuisance). For the sake of illustration, let us consider the following simplified
http://dx.doi.org/10.1016/j.sigpro.2015.02.026
E-mail addresses:eric.chaumette@isae.fr(E. Chaumette),
francois.vincent@isae.fr(F. Vincent),olivier.besson@isae.fr(O. Besson).
1
The notational convention adopted is as follows: italic indicates a scalar quantity, as in a; lower case boldface indicates a column vector quantity, as in a; upper case boldface indicates a matrix quantity, as in A. The n-th row and m-th column element of the matrix A will be denoted by An;mor Að Þn;m. The n-th coordinate of the column vector a will be denoted by an or að Þn. The matrix/vector transpose is indicated by a
superscript T as in AT. For two vectors a and b, aZb means that ab is
positive componentwise. MRðN; PÞ denotes the vector space of real
matrices with N rows and P columns. 1m:m þ lM denotes the M-dimensional
vector with all components set to 0 except components from m to m þ l set to 1. 1Mdenotes the M-dimensional vector with all components set to
1. IMARMM denotes the identity matrix. 1f gA denotes the indicator function of the event A. EΘ½g xð Þ ¼Rg xð Þp x; Θdx denotes the statistical expectation of the vector of functions gð Þ with respect to x parameterized byθ.
example:
xt
Θ
¼ Aθ
stþnt;Θ
T¼θ
T; sT1; …; sTTT
ð1Þ
where 1rt rT, T is the number of independent observations,
xtis the vector of samples of size N, M is the number of signal
sources, st is the vector of complex amplitudes of the M
sources for the tth observation, A
θ
¼ aθ
1 ; …; aθ
Mand að Þ is a vector of N parametric functions depending on a single parameterθ
, ntare Gaussian complex circular noisesindepen-dent of the M sources. Since(1)is invariant over permutation of signal sources, i.e. for any permutation matrix PiARMM: xt
Θ
¼ Aθ
PiðPistÞþnt;it is well known that (1) is an ill-posed unidentifiable
estimation problem which can be regularized, i.e. trans-formed into a well-posed and identifiable estimation pro-blem, by imposing the ordering of the unknown parameters
θm
:θ
9θ
1; …;θ
M
;
θ
1o⋯oθ
M; and of their estimates as well: bθ
9 bθ
ð ÞM. Therefore in the MSE sense, the correctstatistical prediction is given by the computation of EΘ
½ðb
θ
ð Þmθ
mÞ2; 1rmrM. Unfortunately, the correct statistical prediction cannot be obtained from scratch since most of the available results in the open literature on order statistics[1–3]have been derived the other way round, i.e. they request
the knowledge of the distribution of b
θ
. The distribution of bθ
can be obtained from a priori information on the problem at hand or may have been derived in some regions of operation of the observation model. For instance, it has been known for a while that, under reasonably general conditions on the observation model[4,7], the ML estimates are asymptotically Gaussian distributed when the number of independent observation tends to infinity. Additionally, if the observation model is linear Gaussian as in(1), some additional asymptotic regions of operation yielding Gaussian MLEs have also been identified: at finite number of independent observations[5–9] or when the number of samples and the number of
independent observations increase without bound at the
same rate, i.e. N; T-1, N=T-c, 0oco1[10]. Nevertheless
a close look at the derivations of these results reveals an implicit hypothesis: the asymptotic condition of operation considered yields resolvable estimates[11,12], what prevents from estimates re-ordering. Therefore, under this implicit hypothesis b
θ
ð ÞM ¼ bθ
. However when the condition of opera-tion degrades, distribuopera-tion spread and/or locaopera-tion bias of each bθ
m increase and the hypothesis of resolvable estimates doesnot hold any longer yielding observation samples for which b
θ
ð ÞMa bθ
[11,12].Therefore it is the aim of this communication to give an insight into the relevance of order statistics in maximum-likelihood estimation by providing a second-order statistical prediction of ordered normally distributed estimates. This second-order statistical prediction allows to refine the asymp-totic performance analysis of the MSE of MLEs of a subset
θ
of the parameters setΘ
.2Indeed, in the setting of a multivariatenormal distribution with mean vector
μ
bθ and covariance
matrix Cb
θ, b
θ
NMμ
bθ; Cbθ
, with p.d.f. denoted
pNMðb
θ
;μ
bθ; CbθÞ, the most general statistical characterization, i. e. including distribution and moments, have been derived for an exchangeable multivariate normal random vector[1,13,14],that is a normal distribution with a common mean
μ
, acommon variance
σ
2 and a common correlation coefficientρ
: bθ
NMμ
1M;σ
2 1ρ
IMþρ
1M1TM
, with
ρ
A 0; 1½½ . If the focus is on distribution, then the most general result has been released recently in[3]where the exact distribution of linear combinations of order statistics (L-statistics) [15] of arbitrary dependent random variables has been derived (see also[16]for the joint distribution of order statistics in a set of univariate or bivariate observations). In particular,[3]examines the case where the random variables have a joint elliptically contoured distribution and the case where the random vari-ables are exchangeable. Arellano-Vallea and Genton[3] inves-tigate also the particular L-statistics that simply yield a set of order statistics, and study their joint distribution. Unfortu-nately, general derivations of closed form expressions for moments and cumulants of L-statistics were beyond the scopeof [3] and were left for future research. However in the
particular case of a multivariate normal distribution, it is possible to obtain closed forms for first and second order moments of its order statistics directly, i.e. without explicitly computing the order statistics distribution (see Section 2). These closed forms not only generalize the earlier work from the exchangeable case to the general case providing a second-order statistical prediction of L-statistics from multivariate normal distribution but are also required to characterize the MSE of normally distributed vector parameter estimates. Indeed, since it is always assumed that
θ
has distinctcompo-nents, any sensible estimation technique of
θ
must preservethis resolvability requirement and yield distinct mean values, leading to asymptotically non-exchangeable multivariate nor-mal random vector.
2. Second-order statistical prediction of ordered normally distributed estimates
First, note that b
θ
ð ÞMAPerðbθ
Þ, where Perðbθ
Þ ¼ fbθ
i¼ Piθ
b; i ¼ 1; …; M!g is the collection of random vectors bθ
icorre-sponding to the M! different permutations of the components
of b
θ
. Here PiARMM are permutation matrices with PiaPj for all iaj. LetΔ
ARðM 1ÞM be the difference matrix such thatΔθ
¼θ
2θ
1;θ
3θ
2; …;θ
Mθ
M 1 T , i.e., the mth row ofΔ
is dTm þ 1d T m, m ¼ 1; …; M 1, where d1; …; dMare the M-dimensional unit basis vectors. Let Si¼ fb
θ
:Δ
bθ
iZ0g where bθ
i NMμ
i; Ci
,
μ
i¼ Piμ
bθ, Ci¼ Pi CbθPTi.Let P Dð Þ be the probability of an event D. As the set of
eventsf gSi Mi ¼ 1! is a partition ofRM, whatever the real valued function fð Þ, by the theorem of total probability we have; E f b
θ
ð Þm; bθ
ð Þl h i ¼X M! i ¼ 1 E f bθ
ð Þm; bθ
ð Þl jSi h i P Sð Þi 2Note that these results are also applicable to other estimators, such as M-estimators, Bayesian estimators (MAP, MMSE), as long as their distribution is normal.
that is E f b
θ
ð Þm; bθ
ð Þl h i ¼X M! i ¼ 1 E fθ
bi m; bθ
i l jSi h i P Sð Þ:i ð2ÞHowever, from a computational point of view, it is wiser to express(2)as E f b
θ
ð Þm; bθ
ð Þl h i ¼XM! i ¼ 1 E fθ
bi m; bθ
i l jUi h i P Uð Þi ð3Þ where Ui¼ ubi: buiZΔμ
i and ui¼Δ
θ
biμ
i NM 1 0;Δ
CiΔ
T. Then a smart exploitation of(3)(see Appendix
for details) yields E b
θ
ð Þm h i ¼X M! i ¼ 1α
mjiPiþβ
T mjiei ð4Þ E bθ
2ð Þm ¼X M! i ¼ 1σ
2 mjiþα
2 mji Piþ2α
mjiβ
T mjieiþβ
T mjiRiβ
mji ð5Þ E bθðm þ lÞθbð Þm h i ¼1 2 E bθ 2 m þ l ð Þ þE bθ2ð Þm E bθðm þ lÞ bθð Þm 2 ð6Þ E θbðm þ lÞ bθð Þm 2 ¼ 1m:m þ l 1 M 1 T XM! i ¼ 1 ΔμiΔμi T Piþ 2eiΔμi TþR i 0 @ 1 A 0 @ 1 A1m:m þ l 1 M 1 ð7Þ Pi¼ P Uð Þ; ei i¼ E ui1Ui ; Ri¼ E uið ÞuiT1Ui h i ð8Þ where lZ1,α
mji¼ dTmμ
i,μ
i¼ Piμ
b θ, Ci¼ PiCbθP T i,β
mji¼ ðΔ
CiΔ
T Þ 1Δ
C idm,σ
2mji¼ dTmðCbθCiΔ
TΔ
CiΔ
T 1Δ
CiÞ dm. Finally, the second order statistical prediction of any L-statistics, that is linear combinations of the vector of order statistics,bz ¼ Lbθ
ð ÞM, LAMRðN; MÞ, is given byμbz¼ Lμbθð ÞM; Cbz¼ LCbθð ÞM LT; μbθ M ð Þ ¼ E bθð ÞM h i Cbθð ÞM ¼ E bθð ÞMθb T M ð Þ E bθð ÞM h i E bθð ÞM h iT
and can be computed from expressions(4)–(7) of E b
θ
ð Þmh i , E b
θ
ðm þ lÞbθ
ð Þm h i , E bθ
2ð Þmwhen the L-statistics derive from multivariate normal distribution.
2.1. Cramér–Rao bound for ordered normally distributed
estimates
Let CRBΘ
θ
denote the Cramér–Rao bound (CRB) forunbiased estimates of
θ
[4,7]. Then (4) allows for thecomputation of the ordered estimates bias bΘ
θ
m¼
EΘ
θ
bð Þmh i
θ
m which may be different from the (a priori)theoretical bias bΘ
θ
m¼ EΘθ
bmh i
θ
m, leading to theCRB for ordered estimates given by[4,7]
CRBbΘ
θ
¼ bΘθ
bTΘθ
þ IMþ∂ bΘθ
∂θ
T ! CRBΘθ
IMþ∂ bΘθ
∂θ
T !T ; ð9Þwhich may be different from the (a priori) theoretical CRBbΘ
θ
.2.2. Implementation
Let Di be the diagonal matrix such that ð ÞDim;m¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð
Δ
CiΔ
T Þm;m; q
, m ¼ 1; …; M 1. The computational cost of
(4)–(7) can be reduced by reformulatingPi, ei, Ri (8)as Pi¼ P Vð Þ; ei i¼ DiEbvijVi ; Ri¼ DiE bvibv T ijVi h i Di; ð10Þ where Vi¼ bvi: bviZ Di 1
Δμ
i n o and bvi NM 1ð0; Di 1 ðΔ
CiΔ
T ÞD 1i Þ is a vector of correlated standard normal
random variables satisfyingP bvi
m
Z10
o10 23. Thus, from a numerical point of view, the distribution support of each bvi
m can be restricted to the interval ½10; 10
leading to E bvi mjVi r10P Vð Þ and Ei bvi m bvi ljVi r 102 P Vð Þ. Therefore, sincei P Vð Þri min 1r m r M 1 P bvi mZ D 1 i
Δμ
i m n o ; from a numerical point of view:if (mA 1; M 1½ j D 1 i
Δμ
i mr 10, then: P Vð Þ ¼ 0; E bvi ijVi ¼ 0; E bvibv T ijVi h i ¼ 0; if 8mA 1; M 1½ ; D 1 iΔμ
i mZ10, then: P Vð Þ ¼ 1; E bvi ijVi¼ E bvi ¼ 0; ptE bvibv T ijVi h i ¼ E bvibv T i h i ¼ D 1 iΔ
CiΔ
T Di 1:In any other case, P Vð Þ, Ei bvijVi
and E bvibv T ijVi h i can be
computed by resorting to algorithms proposed by Genz[17]
for numerical evaluation of multivariate normal
distribu-tions and moments over domains included in 10½ ; 10M
. 3. The bivariate (two sources) case
Let C9Cb θ,
μ
9μ
bθ, dμ
¼μ
2μ
1, and dbθ
¼ u1¼ bθ
2 bθ
1. As M¼2 then u2¼ u1,Δμ
1¼ dμ
¼Δμ
2, P1þP2¼ 1, 111:1¼ 1, e1e2¼ E u½ ¼ 0, R1 1þR2¼ E u21 ¼σ
2 dbθ,Δ
C1Δ
T ¼Δ
C2Δ
T ¼σ
2 dbθ¼ C1;1þC2;22C1;2. Moreover, if C ¼σ
2ðð1ρ
ÞI2þρ
121T2Þ, thenσ
dbθ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2σ
21ρ
qand a few additional calculations yield E b
θ
ð Þm h i ¼ E bθ
m h i þ 1ð Þm hð Þτ
σ
dbθ; E bθ
2ð Þm ¼ E bθ
2m þ 1ð Þm hð Þτ
μ
2þμ
1σ
dbθ; Var bθ
ð Þm h i ¼ Var bθ
m h iσ
2 dbθhð Þτ
ðτ
þhð Þτ
ÞMSE b
θ
ð Þm h i ¼ Var bθ
m h iσ
2 dbθτ
hð Þ;τ
ð11Þ whereτ
¼ dμ
=σ
dbθ, h yð Þ ¼ E v1fvZ yg h iyP vZyð Þ and
MSE b
θ
ð Þm h i 9E bθ
ð Þmμ
m 2 . An interesting feature is the MSE shrinkage factor:ξ τ
;ρ
¼MSE bθ
ð Þmh i
Var b
θ
mh i ¼ 12 1
ρ
τ
hð Þ:τ
ð12ÞThe computation of the derivatives vector∂
ξ τ
;ρ
=∂τ
;ρ
shows that, for a given
ρ
,ξ τ
;ρ
admits a global minimumwhich is the solution of E v1fvZ yg
h i
2yP vZyð Þ ¼ 0, y ¼
τ
=qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi2 1ρ
, that is, after a numerical resolution, yC0:61 and leading toτ
C0:863pffiffiffiffiffiffiffiffiffiffiffi1ρ
,ξ τ
;ρ
Z10:21
ρ
Z0:6 (seeFig. 1). Therefore the minimum of
ξ τ
;ρ
is 0.6 ( 2.2 dB) and is reached for couplesτ
;ρ
belonging to the setτ
;ρ
-1þjτ
C1:22
. This result is applicable
to any instantiation of (1) for which b
θ
is normal withCbθ¼
σ
2 1ρ
I2þ
ρ
121T2
, as for example, the asymptotic
−50 −45 −40 −35 −30 −25 −20 −15 −10 −5 0 5 10 15 20 −2.3 −2.2 −2.1 −2 −1.9 −1.8 −1.7 −1.6 −1.5 −1.4 −1.3 −1.2 −1.1 −1 −0.9 −0.8 −0.7 −0.6 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 (dB) τ (dB)
Mean Square Error shrinkage factor
ρ = −1 ρ = −0.5 ρ = 0 ρ = 0.5 ρ = 1
behaviour of the MLEs of the direction of arrival (DoA) of two sources of equal power impinging on a uniform linear array (ULA) with symmetric DoAs relative to boresight, or two cisoids (tones) of equal power with opposite frequen-cies, whatever the observation model is conditional or
unconditional[5]. Indeed in these two cases CRBΘ
θ
¼σ
2 1ρ
I2þ
ρ
121T2
[4,5,7]and a priori, a violation of the CRB not attributable to the variance of the empirical MSE
is expected in some scenarios, unless the ad hoc CRB(9)is
computed. However, a closer look to Fig. 1 reveals that
τ
¼ dμ
=σ
dbθ must be inferior to 1015=20C6 in order toexhibit a measurable MSE shrinkage effect. As the normality of estimates and convergence to the CRB is obtained only
in asymptotic conditions where
σ
dbθ≪1, this means that d
θ
¼ dμ
≪1 as well, that is the scenario considered is a (very) high resolution scenario. To highlight the possible impact of estimates ordering on MSE in (very) high resolution scenarios,we consider the estimation of two tones M ¼ 2ð Þ of equal
power with opposite frequencies where the observation model (1) is deterministic: a
θ
T¼ 1; e j2πθ; …; ej2πðN 1ÞθÞ, N¼8, T¼2, dθ
¼ 1=12N, Cnt¼ I2, Cst¼ SNR=N 1 þ1 8 I2 18121T2Þ where SNR is the signal to noise ratio measured at the
20 25 30 35 40 45 50 55 60 65 −75 −70 −65 −60 −55 −50 −45 −40 −35 −30 −25 −20 −15
Signal to Noise Ratio (dB)
dB
Mean Square Error of MLEs
Simu CRB Theo
output of the frequency matched filter. Fig. 2 displays the
empirical and theoretical MSE of ordered estimates (11)
averaged over the two frequencies (since equal by symmetry) and the associated unbiased CRB.Fig. 3displays the empirical
and theoretical MSE shrinkage factor(12)averaged over the
two frequencies as well. The empirical MSE are assessed with 104Monte-Carlo trials and a frequency step
δθ
¼ 1=12N 164. As expected there is a perfect match between the theoretical results and the empirical results in the asymptotic region
(SNR Z 55 dB) where Gaussianity is reached by MLEs,
highlighting the non-negligible impact ( 0.6 dB) of estimates ordering on high resolution scenarios. As the SNR values enter
the threshold area, the MLEs start to explore the side lobe peaks of the matched filter producing outliers and loosing
gradually their Gaussian distribution [4]. Because of this
phenomenon, as the SNR values enter the threshold area, the results derived in this section must be regarded as an approximation which accuracy is somewhat difficult to quan-tify in general. Indeed, it would require to recompute the second-order statistical prediction of ordered estimates taking into account the probability of outlier[4], which is far beyond the scope of this communication. We can simply notice, but without generality, that in our example, the approximation is
quite good for SNRZ30 dB.
20 25 30 35 40 45 50 55 60 65 −2 −1 0 1 2 3 4 5 6 7 8 9 10
Signal to Noise Ratio (dB)
dB
Ratio Mean Square Error/CRB of MLEs
Simu Theo
4. Conclusion
In this communication we have provided a second-order statistical prediction of second-ordered normally distributed estimates which allows to refine the asymptotic
perfor-mance analysis of the MSE of MLEs of a subset
θ
of theparameters set
Θ
. Analytical expressions of the MSEs ofthe whole parameters set
Θ
after re-ordering of the subsetθ
should be the next topic to be addressed.Appendix As the vector b
ξ
m;i¼ ððbθ
iÞm; bu T iÞ T, bu i¼Δ
ðbθ
iμ
iÞ NM 1ð0;Δ
CiΔ
TÞ, is a M-dimensional normal vector result-ing from a bijective affine transformation of b
θ
i, then wehave[1]
pNMbξm;i;μm;i; Cm;i¼ pN1 bθi
mjbui;μmji;σ 2 mji pNM 1 bui; 0;ΔCiΔ T where
μ
mji¼ d T mμ
iþCiΔ
TΔ
CiΔ
T 1 bui andσ
2 mji¼ d T m CbθCiΔ
TΔ
CiΔ
T 1Δ
Ci dm. Let Ui¼ ubi: buiZΔμ
i , then whatever the real valued function f ð Þ:E f
θ
bi m jSi h i ¼ E f bθ
i m jUi h i ¼ E E f bθ
i m jbui h i jUi h iIn the particular cases where f b
θ
ð Þm¼ b
θ
kð Þm; kA 1; 2f g: Eθ
bi mjbui h i ¼μ
mji¼α
mjiþβ
T mjiubi Eθ
bi 2 mjbui ¼σ
2 mjiþμ
2 mji¼σ
2mjiþα
2 mji þ2α
mjiβ
T mjibuiþβ
T mjiubibu T iβ
mjiwhere
α
mji¼ dTmμ
i,β
mji¼Δ
CiΔ
T 1Δ
Cidm, and (2) can finally be expressed as E bθ
ð Þm h i ¼X M! i ¼ 1α
mjiP Uð Þþiβ
T mjiEubi1Ui E bθ
2ð Þm ¼XM! i ¼ 1σ
2 mjiþα
2 mji P Uð Þþ2iα
mjiβ
T mjiEubi1Ui þβ
TmjiE ubi bui T 1Ui h iβ
mji ð13Þ From(2), 8 mA 1; M½ , 8ljmþlA 1; M½ : Eθ
bðm þ lÞ bθ
ð Þm 2 ¼X M! i ¼ 1 Eθ
bi m þ l bθ
i m 2 jSi " # P Sð Þi Additionally, 8 lZ1jmþlA 1; M½ : bθ
i m þ l bθ
i m¼ 1 m:m þ l 1 M 1 TΔ
bθ
i therefore bθ
i m þ l bθ
i m 2 ¼ 1m:m þ l 1 M 1 TΔ
bθ
iΔ
bθ
T i1m:m þ l 1M 1 Finally E bθ
ðm þ lÞ bθ
ð Þm 2 ¼X M! i ¼ 1 1mM 1:m þ l 1 T EΔ
θ
biΔ
θ
b T ijSi 1m:m þ l 1M 1 P Sð Þi ð14Þ Last, as EΔ
θ
biΔ
θ
b T ijSi ¼ E buibu T ijUi h i þEbuijUiΔμ
i T þΔμ
iE buijUi T þΔμ
iΔμ
i T an equivalent expression of(14)is(7). References[1]Y.L. Tong, The Multivariate Normal Distribution, Springer, New York, 1990.
[2]H.A. David, H.N. Nagaraja (Eds.), Order Statistics, 3rd ed. Wiley, New York, 2003.
[3]R.B. Arellano-Vallea, M.G. Genton, On the exact distribution of linear combinations of order statistics from dependent random variables, J. Multivar. Anal. 98 (2007) 1876–1894.
[4]H.L. van Trees, Optimum Array Processing, Wiley-Interscience, New-York, 2002.
[5]P. Stoica, A. Nehorai, Performances study of conditional and uncon-ditional direction of arrival estimation, IEEE Trans. Signal Process. 38 (10) (1990) 1783–1795.
[6]M. Viberg, B. Ottersten, Sensor array processing based on subspace fitting, IEEE Trans. Signal Process. 39 (5) (1991) 1110–1121. [7]B. Ottersten, M. Viberg, P. Stoica, A. Nehorai, Exact and large sample
maximum likelihood techniques for parameter estimation and detection in array processing, in: S. Haykin, J. Litva, T.J. Shepherd (Eds.), Radar Array Processing, Springer-Verlag, Berlin, Germany, 1993, pp. 99–151. (Chapter 4).
[8]J. Li, R.T. Compton, Maximum likelihood angle estimation for signals with known waveforms, IEEE Trans. Signal Process. 41 (9) (1993) 2850–2862.
[9]A. Renaux, P. Forster, E. Chaumette, P. Larzabal, On the high snr cml estimator full statistical characterization, IEEE Trans. Signal Process. 54 (12) (2006) 4840–4843.
[10] X. Mestre, P. Vallet, P. Loubaton, On the resolution probability of conditional and unconditional maximum likelihood doa estimation, in: Proceedings of the Eurasip EUSIPCO, 2013.
[11]M.P. Clark, On the resolvability of normally distributed vector parameter estimates, IEEE Trans. Signal Process. 43 (12) (1995) 2975–2981.
[12]C. Ren, M.N. El Korso, J. Galy, E. Chaumette, P. Larzabal, A. Renaux, On the accuracy and resolvability of vector parameter estimates, IEEE Trans. Signal Process. 62 (14) (2014) 3682–3694.
[13]I. Olkin, M.A.G. Viana, Correlation analysis of extreme observations from a multivariate normal distribution, J. Am. Stat. Assoc. 90 (1995) 1373–1379.
[14]N. Loperfido, A note on skew-elliptical distributions and linear functions of order statistics, Stat. Probab. Lett. 78 (2009) 3184–3186.
[15]J.R.M. Hosking, L-moments: analysis and estimation of distributions using linear combinations of order statistics, J. R. Stat. Soc. 52 (1) (1990) 105–124.
[16]H.M. Barakat, Multivariate order statistics based on dependent and nonidentically distributed random variables, J. Multivar. Analysis 100 (2009) 81–90.
[17]A. Genz, Numerical computation of multivariate normal probabil-ities, J. Comput. Graph. Stat. 1 (1992) 141–149.