HAL Id: hal-01421517
https://hal.parisnanterre.fr//hal-01421517
Submitted on 4 Dec 2017
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Cramér-Rao bound and statistical resolution limit investigation for near-field source localization
Tao Bao, Mohammed Nabil El Korso, Habiba Hafdallah Ouslimani
To cite this version:
Tao Bao, Mohammed Nabil El Korso, Habiba Hafdallah Ouslimani. Cramér-Rao bound and statistical
resolution limit investigation for near-field source localization. Digital Signal Processing, Elsevier,
2016, 48, pp.137-147. �10.1016/j.dsp.2015.09.019�. �hal-01421517�
Cramér–Rao bound and statistical resolution limit investigation for near-field source localization
Tao Bao
a, Mohammed Nabil El Korso
b,∗ , Habiba Hafdallah Ouslimani
ba DepartmentofElectronicsEngineering,NorthwesternPolytechnicalUniversity,Xi’an710129,China bParisOuestUniversity,LEMEEA4416,50ruedeSèvres,92410Villed’Avray,France
1. Introduction
Therehasbeentremendous interestintheinvestigation ofar- ray signal processing during the last three decades due to its importance in radar, sonar, seismology and communication sys- tems[2–9].One canclassify sourcelocalization, inthecontextof array processing, intothe so-called far-field andnear-field cases.
Thecontext offar-fieldsourcesassumesplanar wavefrontsofthe propagating waves as they reach the array. Nevertheless, when thesources arelocated intheso-callednear-fieldregion, thisas- sumptionisnolongervalid.Moreprecisely,thesourcewavefronts mustbeassumedspherical,whichrequiresboththerangeandthe direction-of-arrival (DOA) parameter estimation to carry out the localizationprocedure[10–12].Onecanfindseveralestimational- gorithmsadaptedforthenear-fieldsourcelocalization[13–17]and somerelatedperformanceanalysisinordertoquantifytheestima- torperformanceintermsofthemeansquareerror(MSE)[18–22].
However, to fully characterize estimator performance, one has, additionally, to perform the resolvability analysis of two closely spacedsources.Recently,someworksconsideredbothfarfieldand near field scenario. In [23–25] the authors studied, respectively, theasymptoticresolvabilityoffar-field,deterministic,discreteand stochasticsourcesignals, whereas,in[26] theauthorsfocused on
✩ ThisworkwaspartiallypresentedduringtheInternationalConferenceSAM-12 [1].
*
Correspondingauthor.E-mailaddress:m.elkorso@u-paris10.fr(M.N. El Korso).
polarized far-field sources and in [27], the authors studied the angular resolution limit for far-field sources in the presence of modeling errors. Moreover, in [28], the authors provide a novel methodologytoderiveanapproximationoftheresolutionlimit.
Nevertheless,tothebestofourknowledge,noworkshave been done toquantifytheresolvabilityoftwo closelyspacednear-field sources in the non-uniform linear array (NULA) context for the general case1 (general case, means that the time varying source signals, the noisevariance, the rangeandthebearing areall un- knownparameters).Consequently, thegoalofthispaperis tofill thislackbyaddressingthefollowingquestion:“Whatis,inanear- field source scenario, the minimum source separation required, underwhichtwonear-fieldsourcescanstillbecorrectlyresolved?”
Tothis end, one commontool is thedistance resolutionlimit (DRL)oftwocloselyspacedsources,definedastheminimumdis- tance withrespect to the sources for which allows a correct re- solvability.TheDRLcanbederived basedononeofthefollowing three fundamental approaches: i) the analysis of the mean null spectrum[30],ii)thedetectiontheory[19,23](usingabinaryhy- pothesis test anda proper generalizedlikelihood ratio test [27]), oriii)theestimationaccuracy(namely, basedontheCramér–Rao
1 In[29],theauthorsstudiedtheresolvabilityinthenearfieldcontextforthe specialoptimisticcaseinwhichthesourcesignalsandthenoisevarianceareas- sumedknown,andtheonlyunknownparametersaretherangeandthebearing.
Whereas,in[28],theauthorsderivedtheresolutionlimitwithrespecttothebear- ingparameteronly,withoutconsideringtherangeasaparameterofinterestaffect- ingtheresolution.
bound(CRB)tool).Moreprecisely,in[31],Smithproposedthefol- lowingDRLcriterion:“Twosourcesareresolvableiftheseparation betweenthem,denotedbydNF,islessthanthestandarddeviation oftheseparationestimation”.Sincethestandarddeviationcanbe approximatedusingtheCRBby
√
CRB
(
dNF)
[31],thus,Smith’scri- terion statesthat the DRL is givenasthe particular value ofthe distancedNF betweentwo sources which isequal to√
CRB
(
dNF)
. ThismeansthattheDRListhesolutionofthefollowingequation DRL= √
CRB
(
DRL)
.ItisworthnotingthatthisDRL,whichisbasedontheCRB,is widely andcommonly usedfor thefollowing reasons: i) it takes thecouplingbetweentheparametersintoaccount,ii)itavoidsthe well-knowndrawbackofthemeannullspectrumapproachwhich is valid fora specific given high-resolution algorithm and hence lacks generality[24],andfinally,iii) itis relatedtothe detection theorybasedapproachthankstoatranslatorfactorwhichisgiven in[11].
In order to derive the DRL, we first investigate and develop explicitclosed-form expressions ofthe deterministic CRBfortwo closely spaced near-field spaced sources (which, to the best of our knowledge, has not been derived2). Then, we resort Smith’s criterion to derive the DRL. Thus, we deduce and analyze the relationship between the minimum distance of resolvability and the minimumsignal-to-noise ratio (SNR) requiredto resolve two closelyspacednear-fieldsources.Finally,weproposeafast,nearly optimal,arraydesignproceduretoenhancethecapacityofresolv- ability fora given constraint(number of sensors, array aperture, etc.).Numericalsimulationsaregiventoassessthevalidity ofthe proposedexpressions(forboththedeterministicCRBandtheDRL) andtheenhancementduetotheproposedarraygeometrydesign.
Forthe restpart ofthis paper,the following notation will be used.A lowercasebold letterdenotes avector, andan uppercase onea matrix.Vectors areby defaultincolumnorientationunless specified.Upper scripts
·
T and·
H denote,respectively, the trans- pose and the trans-conjugate of a matrix. The operators diag( · )
, adiag(·)
,·
and{·}
denote,respectively,thediagonaloperator, theanti-diagonal operator, theEuclidean normandthe realpart.tr
{·}
anddet{·}
denotethetraceandthedeterminantofamatrix;whereas,
and⊗
denotetheHadamardandtheKroneckerprod- ucts,respectively. IL denotestheidentitymatrixofsize L×
L.[·]
i and[·]
i,j represent, respectively, the i-thvector element and the(
i,
j)
-thmatrixelement.Finally,ε ˆ
denotesanyasymptoticallyun- biasedestimatorofε
.This paper is structured as follows. Section 2 introduces the datamodelandtheunderlyingassumptions.InSection 3,wede- rivetheCRBofaNULAfororthogonalandnon-orthogonalsources.
We deduced the DRL in Section 4. Discussionof the results and conclusionarepresentedinSections5and6,respectively.
2. Systemmodels
Assume a linear (possibly non-uniform) array composed of N sensorswithlocationsd0
·
d,d1·
d,· · ·
,dN−1·
dthatreceivesasignal emittedbytwonear-field,narrow-bandsourcess1(
t)
ands2(
t)
,as showninFig. 1.Letdn·
ddenotethelocationofthe(
n+
1)
-thsen- sor,inwhichdistheunitandthereferencesensoristhefirstone (i.e.,d0=
0).Furthermore,forsake ofsimplicity,the sensorloca- tionsareexpressedasmultiplicationofacommonunitbase-lined.Thismeansthatd0,d1,
· · ·
,dN−1areintegers.Theobservedsignal, xn(
t)
,atthet-thsnapshotofthe(
n+
1)
-thsensorisgivenbyxn
(
t) =
s1(
t)
ejτn1+
s2(
t)
ejτn2+
vn(
t),
n
= 0 , 1 ,· · · ,
N− 1 ,
t= 1 , 2 ,· · · ,
T(1)
2 Moreprecisely,in[18–22],onlytheonesinglesourcecasewasunderconsider- ation.
Fig. 1.Localizationoftwosourcesinthenear-fieldregionusinganon-uniformlin- eararray.Symbols•and◦representthepositionofasensorandthepositionofa missingsensor,respectively.
wherevn
(
t)
denotesacomplexcircularwhiteGaussiannoisewith zero-mean and unknown varianceσ
2, which is assumed to be uncorrelated both temporally and spatially. The i-th source sig- nal, with a carrier frequency equal to f0 is given by si(
t) = α
i(
t)
ej(2πf0t+ψi(t)),i=
1,
2,whereα
i(
t)
andψ
i(
t)
aretherealam- plitude andtheshiftphaseofthe i-thsource,respectively,and T is thenumber ofsnapshots.Moreover, the time delayτ
ni associ- atedwiththesignalpropagationtime fromthei-thsourcetothe(
n+
1)
-thsensorisrepresentedas[23]τ
ni= 2 π
riλ
1 +
d2nd2r2i
− 2d
ndsin θ
iri
− 1
,
n= 0 , 1 ,· · · ,
N− 1 (2)
whereλ
isthesignalwavelength,riandθ
i∈ [
0, π /
2]
aretherange andthe bearingof thei-th source,respectively. Itis well known thatiftherangeisinsidetheso-calledFresnelregion[17],i.e.,if0 . 62
d3N−1d3λ
1/2<
ri< 2d
2 N−1d2
λ (3)
then,thetimedelay
τ
ni canbeapproximatedbyτ
ni≈ ω
idn+ φ
id2n+
O d2r2i
(4)
in whichω
i andφ
i are the so-calledelectricangles for the i-th sourcewhichareconnectedtothephysicalparametersbythefol- lowingrelationshipsω
i= − 2 π
dλ sin (θ
i) (5)
and
φ
i= π
d2
λ
ricos
2(θ
i) (6)
Then,basedonapproximationoftimedelayin(4),theobservation atthet-thsnapshotofthe
(
n+
1)
-thsensorbecomesxn
(
t) =
s1(
t)
ejω1dn+φ1d2n
+
s2(
t)
ejω2dn+φ2d2n
+
vn(
t) (7)
Consequently,thet-thobservationvectorcanbeexpressedas x(
t) = [
x0(
t) · · ·
xN−1(
t)]
T=
A(
p1,
p2)
s(
t) +
v(
t) (8)
where s(
t) = [
s1(
t),
s2(
t)]
T, v(
t) = [
v0(
t), · · · ,
vN−1(
t)]
T, pi= [ ω
i, φ
i]
T for i=
1,
2 and A(
p1,
p2) = [
a( ω
1, φ
1),
a( ω
2, φ
2) ]
. The(
n+
1)
-thelementofthesteeringvectora( ω
i, φ
i)
isgivenby[
a( ω
i, φ
i)]
n+1= e
j(ωidn+φidn2). (9)
Weassumethattheunknownparametervectorisε =
p1T,
pT2, ψ
1T,ψ
T2, α
1T, α
2T, σ
2 T(10)
where
ψ
i= [ ψ
i(
1), · · · , ψ
i(
T) ]
T,α
i= [ α
i(
1), · · · , α
i(
T) ]
T. It is worth noting that the unknown vector parameterε
is assumed tobe deterministic anditssize dependsonthe numberofsnap- shotsT.Underthedeterministicmodel,thejointprobabilitydistribution function(pdf)ofthefullobservations
χ = [
xT(
1) · · ·
xT(
T) ]
T,under i.i.d.assumption,iswrittenasfollowsp
( χ | ε ) = 1 π
NTdet (
R)
e−(χ−μ)HR−1(χ−μ)
(11)
inwhichR
= σ
2INT andμ = [
sT( 1 )
AT(
p1,
p2), · · · ,
sT(
T)
AT(
p1,
p2) ]
T. (12)
3. Cramér–RaobounddefinitionandderivationLetE
(ˆ ε − ε )(ˆ ε − ε )
Tbethecovariancematrixofanestimate of
ε
.Letusassumethatε ˆ
isanasymptotically unbiasedestimate ofthetrueparametervectorε
,andletus,inthefollowing,define theCRBforthe consideredmodel[32].The covarianceinequality principlestatesthatunder asymptoticconditions[33],wehaveMSE
[ˆ ε ]
i=
E[ˆ ε ]
i− [ ε ]
i2
≥ [
CRB( ε )]
i,i(13)
wheretheCRBisgivenastheinverseoftheso-calledFisherinfor- mationmatrix(FIM)asfollowsCRB
( ε ) =
FIM−1( ε ) (14)
Sinceweare workingwithacomplexcircularGaussian obser- vation model, the i-th row,k-th columnelement ofthe FIM for theunknownrealparameter vector,givenin(10),canbe written as[34]:
[
CRB−1( ε ) ]
i,k= [
FIM( ε )]
i,k=
trR−1
∂
R∂ [ ε ]
iR−1
∂
R∂ [ ε ]
k+ 2
∂ μ
H∂ [ ε ]
iR−1
∂ μ
∂ [ ε ]
k=
NTσ
4∂ σ
2∂ [ ε ]
i∂ σ
2∂ [ ε ]
k+
2
σ
2∂ μ
H∂ [ ε ]
i∂ μ
∂ [ ε ]
k(15)
3.1.FIMderivation
From(15)we caneasily seethat thenoise variance
σ
2 isde- coupledfromtheothersparameters.Thus,forsimplicityandwith- out loss of generality, we compute the FIM without taking into accountthe noise variance.Consequently, thevariance parameter isomitted fromthe vector parameterε
. So that the structureof theFIMisasfollows(forproof,pleaserefertoAppendix A):FIM
( ε ) = 2 σ
2⎡
⎢ ⎢
⎢ ⎢
⎢ ⎢
⎣
Fp1p1 Fp1p2
Fp2p1 Fp2p2
Fp1ψ1 Fp1ψ2
0
Fp1α2Fp2ψ1 Fp2ψ2 Fp2α1
0
Fψ1p1 Fψ1p2Fψ2p1 Fψ2p2
0
Fα1p2 Fα2p10
FL L
⎤
⎥ ⎥
⎥ ⎥
⎥ ⎥
⎦
(16)
where3 L= [ ψ
T1, ψ
T2, α
1T, α
T2]
T.3 Forsakeofsimplicityandtoavoidanymisunderstanding,weusedthefollowing notation,Fuv,wherethelowerscriptu v denotestheconsideredpartoftheFIM whichcorrespondstothederivationaccordingtothevectorparametersuandvas shownin(7),respectively.
3.2. Analyticalinversion
First,wenotethatthe4T
×
4T matrixFL L in(16)canbepar- titionedintodiagonalsub-blocksasFL L
=
⎡
⎢ ⎢
⎣
Fψ1ψ1 Fψ1ψ2 Fψ2ψ1 Fψ2ψ2
0
Fψ1α2Fψ2α1
0 0
Fα1ψ2Fα2ψ1
0
Fα1α1 Fα1α2 Fα2α1 Fα2α2
⎤
⎥ ⎥
⎦ (17)
inwhich,theexpressionofeachsub-blockisgiveninAppendix A.
Second,the FL L matrixis fullyinverted byapplyingtheSchur complement[35]formulatwiceasfollows:
1) Using theSchur complementwithrespect tothe lower right block matrixin(17)representedby thedottedlinepartition, oneobtains
Fα1α1 Fα1α2 Fα2α1 Fα2α2 −1=
S−1−
S−1Fα1α2F−α11α1−
F−α11α1Fα2α1S−1 F−α11α1+
F−α11α1Fα2α1S−1Fα1α2F−α11α1(18)
inwhichS
=
Fα2α2−
Fα1α2F−α11α1Fα2α1. (19)
2) Now,weusetheSchurcomplementwithrespecttothelower rightblockmatrixin(17)representedbythesolid lineparti- tionandweobtainF−L L1
=
⎡
⎣
Fψ1ψ1 Fψ1ψ2 0 Fψ1α2
Fψ2ψ1 Fψ2ψ2 Fψ2α1 0 0 Fα1ψ2 Fα1α1 Fα1α2
Fα2ψ1 0 Fα2α1 Fα2α2
⎤
⎦
−1
=
A−BD−1C−1
−
A−BD−1C−1 BD−1
−D−1C
A−BD−1C−1
D−1+D−1C
A−BD−1C−1 BD−1
(20)
whereA
=
Fψ1ψ1 Fψ1ψ2 Fψ2ψ1 Fψ2ψ2,
B=
0
Fψ1α2Fψ2α1
0
,
C=
0
Fα1ψ2 Fα2ψ10
and
D=
Fα1α1 Fα1α2Fα2α1 Fα2α2
.
3) Finally,having theinverse ofFLL matrixone can apply again thismethodtocomputeanalyticallytheSchurcomplementof thelowerrightblockofthematrixFIM
( ε )
in(16).Aftertediouscalculation,weareabletoexpressinclosedform the Schur complement of the matrix FLL, which corresponds to CRB−P P1:
CRB−P P1
= 2 σ
2⎡
⎢ ⎢
⎣
( 1 , 1 , 2 ) ( 1 ) ( 1 ) ( 1 ) ( 1 , 2 , 4 ) ( 2 ) ( 1 ) ( 2 , 1 , 2 ) ( 2 )
( 2 ) ( 2 ) ( 2 , 2 , 4 )
⎤
⎥ ⎥
⎦
(21)
inwhich P= [
pT1,
p2T]
T andfunctions, ,
andaregivenin Appendix B.
Consequently,insteadofinvertinga
(
4T+
4) × (
4T+
4)
matrix toobtainCRBpipk,wecaninvertjusta2×
2 matrixasfollowsCRBpipk
=
σ22
ϒ
−1(
i) for
i=
k, ∀
i= 1 , 2
−
σ22ϒ
−1( 1 )ℵ ( 2 ) for
i=
k, ∀
i,
k= 1 , 2 (22)
whereℵ (
i) = 1
(
i, 1 , 2 )(
i, 2 , 4 ) −
2(
i)
(
i, 2 , 4 ) − (
i)
−(
i) (
i, 1 , 2 )
, =
( 1 ) ( 2 )
and ϒ(
i) =
(
i, 1 , 2 ) (
i) (
i) (
i, 2 , 4 )
− ℵ ( 3 −
i),
which,finally,leadstothefollowingresult CRBpipi
= σ
22
1
h
(
i, 1 )
h(
i, 2 ) −
m(
i)
2 h(
i, 2 ) −
m( 1 )
−
m( 1 )
h(
i, 1 )
,
∀
i= 1 , 2 (23)
where
h
(
i,
j) = (
i,
j, 2
j)
+ (
j)(( 3 −
i, 3 −
j, 2 ( 3 −
j))(
j) − ( 3 −
i))
2( 3 −
i) − ( 3 −
i, 1 , 2 )( 3 −
i, 2 , 4 ) + (( 3 −
i,
j, 2
j) − ( 3 −
i)(
j))
2
( 3 −
i) − ( 3 −
i, 1 , 2 )( 3 −
i, 2 , 4 ) , (24)
andm
(
i) = (
i) + ( 1 )(( 3 −
i, 2 , 4 )) − ( 3 −
i)( 2 ))
2( 3 −
i) − ( 3 −
i, 1 , 2 )( 3 −
i, 2 , 4 ) + (( 3 −
i, 1 , 2 )( 2 ) − ( 3 −
i))
2
( 3 −
i) − ( 3 −
i, 1 , 2 )( 3 −
i, 2 , 4 ) . (25)
3.3. CRBforthephysicalparametervectorκ
Evenifthemodel(7)isusuallyusedinarraysignalprocessing, itsCRBrelatedto
ε
doesnotbringusphysicalinformation.Then, it is interesting to analyzethis CRB regardingthe bearingθ
and ranger.Thatis,thenewunknownphysicalparametervectorbecomes
κ =
g( ε ) =
θ
1,
r1, θ
2,
r2,ψ
1T, ψ
T2, α
T1, α
T2T
(26)
FromCRB(ε
),wededuceCRB(κ
)byusingthefollowingformula (see[36,p.45])CRB
( κ ) = ∂
g( ε )
∂ ε
CRB( ε ) ∂
g( ε )
∂ ε
T(27)
wheretheJacobianmatrixisgivenby∂
g( ε )
∂ ε = λ π
d⎡
⎢ ⎢
⎢ ⎢
⎢ ⎢
⎢ ⎣
−2 cos1(θ1)
0 0 0 0
r1sin(θ1) cos2(θ1)
−r21
dcos2(θ1)
0 0 0
0 0
−2 cos1(θ2)
0 0
0 0
r2sin(θ2)cos2(θ2)
−r22 dcos2(θ2)
0
0 0 0 0
πλdI4T⎤
⎥ ⎥
⎥ ⎥
⎥ ⎥
⎥ ⎦
(28)
Using(27)andtheJacobianmatrixabove,weobtainCRB (θ
i) = σ
22
λ
2h(
i, 2 ) 4 π
2d2cos
2(θ
i)
h
(
i, 1 )
h(
i, 2 ) −
m(
i)
2, (29)
andCRB (
ri) = σ
22 λ2m(i)2
h(i,2)d2sin2(θi)+2k(i)ridsin(θi)+h(i,1)r2i
π
2d4cos4(θi)h(i,1)h(i,2)−m(i)2 .
(30)
3.4. Thespecialcaseoftwoorthogonalsources
Inthissubsection,weconsidertheparticularcaseoftwonear- field orthogonal sources.Thiskindofsignal iscommonlyusedin radarandradiocommunicationtofacilitatethetaskofseparating signals [37]. From theabove results, we deducethe CRBfortwo near-field orthogonal sources. The assumption of two orthogonal sourcesmeansthat[38]
T t=1sH1
(
t)
s2(
t) = 0 (31)
Substituting(31)in(18),weobtain
=
02×2(32)
Consequently, the CRBpipk for near-field orthogonal sources is givenby
CRBpipk
=
σ22
ℵ (
i) for
i=
k, ∀
i= 1 , 2
0
for
i=
k, ∀
i,
k= 1 , 2 (33)
whichleadstothefollowingresultsCRB ( ω
i) = σ
22
(
i, 2 , 4 )
(
i, 1 , 2 )(
i, 2 , 4 ) −
2(
i) , (34) CRB (φ
i) = σ
22
(
i, 1 , 2 )
(
i, 1 , 2 )(
i, 2 , 4 ) −
2(
i) , (35) CRB (θ
i) = σ
22
λ
2(
i, 2 , 4 ) 4 π
2d2cos
2(θ
i)
(
i, 1 , 2 )(
i, 2 , 4 ) −
2(
i) , (36)
andCRB (
ri) =
σ22 λ2r2i
(i,2,4)d2sin2(θi)+2(i)ridsin(θi)+(i,1,2)r2i π2d4cos4(θi)
(i,1,2)(i,2,4)−2(i) .
(37)
4. TheminimumSNRrequiredtoresolvetwocloselyspaced near-fieldsourcesThe aimof this section is to derive a closed-form expression oftheminimumSNR,denotedbySNRmin,requiredtoresolvetwo closely spaced near-field sources. Byusing Smith’scriterion [39], the DRL is theparticular value of the distancedNF between two sources forwhich
√
CRB
(
dNF)
isequaltodNF,wherethedistance betweentwonearfieldsourcesisdefinedasdNF
=
r21
+
r22− 2r
1r2cos (θ
2− θ
1). (38)
Consequently,thenewunknownphysicalparametervectorisj
( κ ) =
r21+
r22−
2r1r2cos(θ
2− θ
1),
r1, θ
2,
r2, ψ
1T, ψ
T2, α
1T, α
T2 T(39)
Asin(27),weobtainCRB(
dNF)
fromthefollowingformulaCRB (
dNF) = ∂
j( κ )
∂ κ
CRB( κ ) ∂
j( κ )
∂ κ
T1,1
(40)
From (21), (22) and using the relation (27), one notices that CRB( κ ) =
CRBσ(2κ) does not depend onσ
2. Denoting CRB(
dNF) =
CRB(dNF)
σ2 equation(40)canbewrittenas
CRB (
dNF) =
∂
j( κ )
∂ κ
CRB( κ ) ∂
j( κ )
∂ κ
T1,1
(41)
Fig. 2.DRLvs.theratioofSNRsinthecaseofaULAof6sensors,100snapshots andSNR1+SNR2=30 dB.
Consequently,oneobtains
CRB (
dNF) =
4 i,j=1∂
j( κ )
∂ κ
1,i
∂
j( κ )
∂ κ
1,j
[
CRB( κ )]
i,j(42)
where
∂
j( κ )
∂ κ
1,1
= − ∂
j( κ )
∂ κ
1,3
= −
r1r2sin (θ
2− θ
1)
dNF
, (43)
and
∂
j( κ )
∂ κ
1,2
= − ∂
j( κ )
∂ κ
1,4
=
r1−
r2cos (θ
2− θ
1)
dNF. (44)
LetusdefinetheSNRas
SNR = α
12
+ α
22
σ
2. (45)
Furthermore, from (40) and (42), one has CRB
(
dNF) =
CRBσ(d2NF). Since,CRB(
dNF)
isσ
2 independentandtheSNRminisgivenbythe SNRforwhichthefollowingequationholdsσ
2CRB ( DRL ) =
r21+
r22− 2r
1r2cos (θ
2− θ
1) (46)
Consequently,from(45)and(46),oneobtainsSNR
min=
α
12+ α
22CRB ( DRL )
r21
+
r22− 2r
1r2cos (θ
2− θ
1) (47)
Remark. Weshould precise that usingthe definitionofthe total SNR inthe deterministic case(i.e.,SNR
=
SNR1+
SNR2, inwhich SNR1=
||ασ12||2 and SNR2=
||ασ22||2) to characterize the SNRmin we losethe informationrelated to theratio ofSNRs, i.e., SNRSNR12. Con- sequently,one has to take a precaution using the SNRmin, since, foragivenDRLonemayrequireahigherSNRminiftheratio SNRSNR1
2
is far from 1 as represented in Fig. 2. Indeed, Fig. 2 shows the variation ofthe DRL w.r.t. SNRSNR1
2 subjectto a fixed total SNR (i.e., SNR1
+
SNR2=
constant).Furthermore,wecan notice,bysimula- tion,thatthesmallestSNRminpossibleforagivenDRLisobtained inthecaseof SNRSNR12
=
1.Fig. 3.Comparisonbetweennumericalandanalytical CRB(ω1)andCRB(φ1)with (θ1,r1)=(30◦,10λ)and(θ2,r2)=(60◦,20λ).Linesandmarkersrepresentthenu- mericalandanalyticalCRBs,respectively.
5. Numericalinvestigation:CRB,DRLandgeometrydesign NumericalresultsarepresentedinthissectionforaNULAwith sixsensors(withthefollowinglocations:0, 3
·
d,4·
d,5·
d,6·
d, 9·
d) thatreceives asignal emittedbytwo sourcesintheFresnel region inwhich their amplitudes andphasesare generated from arealizationofan i.i.d.Gaussiandistributionwithzeromeanand unit variance unlessspecified. The number ofsnapshots is given by T=
100.5.1. NumericalanalysisoftheCRB
In this section, we analyze the proposed closed-form expres- sions giveninSection 3by comparing theanalytical resultswith thenumerical resultstoillustrate thevalidity ofourderived CRB closedformexpressions.Moreprecisely,
•
In Fig. 3,we compare the analyticalCRBs givenin (23)with thenumericalexactCRBs givenin(6).Thisfigure showsthat theproposedclosed-formexpressionsareinagoodagreement withthetruenumericalCRBs,whichvalidatesourderivedex- pressions.Furthermore,we have thesame resultsfor CRB(θ )
andCRB(
r)
withrespectto(29)and(30).•
From(24),wenoticethat∀
i=
1,
2,h(
i,
1)
is O(
N3)
,however h(
i,
2)
is O(
N8)
,which meansthat theestimation oftheso- calledsecondelectricalangleφ
iismoreaccuratethanestimat- ingtheso-calledfirstelectrical angleω
i.Thisiscorroborated byFig. 3.•
From (22) and using the inversion lemma matrix, we can quantify the effect of the estimation accuracy of one source ontheotheronebythefollowingrelationshipCRBp1p1
= σ
22 ℵ ( 1 ) + ℵ ( 1 )
CRBp2p2ℵ ( 1 )
= σ
22 ℵ ( 1 )(
I2×2+ 2
σ
2CRBp2p2
ℵ ( 1 )) (48)
andinthesamewayweobtain CRBp2p2
= σ
22 ℵ ( 2 )(
I2×2+ 2
σ
2CRBp1p1
ℵ ( 2 )) (49)
Thus,inthegeneralcase,theestimationerroroftheelectrical anglesofonesourceaffectstheelectricalanglesestimationof
Fig. 4.VariationoftheCRBw.r.t.r1withfixedr2=2 m andforequalsourcepower.
theother source.Whereas, from(48) and(49),we note that thisisnotthecaseifthesourcesareorthogonal,meaningthat theestimationaccuracyofonesourceneitherdependsonnor affectstheestimation accuracyofthe other source.However, itis obviousthat CRBpipi is affected by theposition ofboth sources, moreprecisely, by the distancebetweenthe electric anglesofbothsources(because
(
i,
k,
p)
dependsonω
andφ
evenintheorthogonalcase,andthusCRBpipi dependson thedistancebetweentheelectricalangles).•
Unlikethecaseofonesource,theCRBfortwosourcesarenot phase-invariant. However, we notice that for two orthogonal sources the CRBbecomes phase-invariant. Numericalsimula- tionshowstheimprovementduetoorthogonality.•
For a sufficient number of sensors, CRBθiθi and CRBriri are O(
1/
N3)
.•
CRBriri isbearing-invariantandrange-invariant.•
Forλ,
ri∝
d,thedependenceontherangeis O(
r2i)
,meaning that nearer thesource is, the better the estimation(keeping inmindtheFresnelconstraints).ThisiscorroboratedbyFig. 4.Thedependenceoftherangeonthebearingis O
(
1/
cos4(θ
i))
. Forθ
i closetoπ /
2, we observe that CRBriri goes toinfinity butfasterthanCRBθiθi (cf.Fig. 5forCRB(θ
1)).•
Fig. 6 showsthat, fora different value of bearing andfor a fixedSNR,thehigherthefrequencyis,thelowertheCRBriri. 5.2. CorrelationfactorInthefollowing,weinvestigatetheeffectofthecorrelationfac- torontheDRL.Wedefinethecorrelationfactor
ξ
betweenthetwo sourcesas[40]ξ =
sH1s2 s1s2(50)
inwhichsi
= [
si(
1), . . . ,
si(
T) ]
T denotesthei-thsignalvector.We usethe same simulation parameters asin Fig. 3,expect the fact thesignalsourcessatisfy(50)foragivenξ
.From(34)to(37),we knowthattheorthogonalitybetweentwosourcesimpliestheun- coupling property. Then, since the sources are decoupled,ξ =
0, the estimation will be improved (at least for any efficient algo- rithm). In Fig. 7, we compare the correlation effect of the real amplitude ofξ
on the DRL. Note that, fora given SNR, DRL be- comesgreaterasthevalueoftherealamplitudeofξ
increases.In otherwords,thesmallerthecorrelationbetweenbothsources, theFig. 5.CRB(θ1)vs.SNRforr1=10λ,(θ2,r2)=(60◦,20λ)andvaluesofθ1asshown inthelegend.LinesandmarkersrepresenttheanalyticalandnumericalCRBs,re- spectively.
Fig. 6.CRB(r1)vs.SNRfor(θ1,r1)=(30◦,10λ),(θ2,r2)=(60◦,20λ)andvaluesof f0asshowninthelegend.Linesandmarkersrepresenttheanalyticalandnumeri- calCRBs,respectively.
better theresolutionlimit performance.Meanwhile, DRLachieves its lowestvalue when
ξ =
0,whichmeansthattwo signalsenjoy aminimum(best)resolutionlimitbetweenthemiftheyhavethe orthogonalrelationship.5.3. SNRminbehaviorw.r.t.centralangle
Letusdefinethecentralangleas
θ
c= θ
1+ θ
22 (51)
InFig. 8,we haveplottedthe SNRmin requiredtoresolves two closelyspacedsourcesfordifferentvaluesof
θ
c.Itshouldbenoted that fora givenresolutionlimit,asthecentralangleθ
c increases, theminimumsignaltonoiseratiorequiredincreasesalso.5.4. Minimumresolutionlimitboundary
In practical applications, one mayneed the knowledge of the physicalboundary, whichdelimitstheresolvabilityregion (mean-
Fig. 7.The effect of the amplitude of the correlation factor on the DRL.
Fig. 8.TheSNRminversusthedistancecurvesfromanon-uniformarrayforvarious centralangles.
ing the region for which two near field sources are resolvable).
Althoughthisquestion is worth consideration,to thebest ofour knowledge, no work has been done on it in the literature. In thissubsection, the minimum resolution limit boundary for two sourcesisinvestigatedanddiscussedfordifferentSNRs.
Followingsimilarstepsasthoseemployedin Section4,weob- taintheminimumresolutionlimitboundaryoftwosourcesfordif- ferentSNRvalue.SuchboundaryisrepresentedinFig. 9,inwhich thestarsymbolrepresentsthecoordinatesofthefirstsourcew.r.t.
thereferencesensor
(
xo,
yo) = (
0,
0)
,inwhich(θ
1,
r1) = (
45◦,
6λ)
,α
12=
1 andα
22=
1. For different SNR, the second source must be outside this boundary to ensure a correct resolvability givenbySmith’scriterion.Asexpected,onenoticesthatforhigher SNR, one obtains a small boundary. Furthermore, we can notice thatthisboundaryismorebearingsensitivethanrangesensitive.Assaidbefore,thisboundaryisaphysicallimitation.Nevertheless, thislattercanbecompensatedbyaproperarraygeometrydesign whichistheaimofthenextsection.
5.5.Effectofsensors’position
Theremainingpartofthissectionisdevotedtothenumerical analysisoftheeffectofsensorarraygeometryontheDRL.Assume
(θ
1,
r1) = (
30◦,
6λ)
,(θ
2,
r2) = (
60◦,
10λ)
andα
12= α
22=
1.Fig. 9.DRL boundary for 6-sensors.
Table 1
Differentarraygeometricconfigurations.
Array type Geometric configuration
Type 1 • • • • • ◦ ◦ ◦ ◦ •
Type 2 • ◦ • ◦ • ◦ • ◦ • •
Type 3 • ◦ ◦ ◦ ◦ • • • • •
Type 4 • ◦ ◦ • • • • ◦ ◦ •
Type 5 • • • • • • • • • •
KeepthearrayapertureA
=
9 (inunitofd= λ/
2)fixedandN=
6, thereare10positionsavailableforsixsensors.InFig. 10,weplotthe DRLwhichcomparestheimpact ofthe different sensorarray geometry w.r.t. SNRmin.Five types ofarray configurationsareconsidered,asshowninTable 1.Type1totype 4arelistedinTable 1withN
=
6 sensors,Type5isanULAcon- figurationwithN=
10 sensors.From comparisonofsimulation results,one canfurther notice that
•
Forthesame array aperture,sameSNR andsame numberof sensors,theDRLisgreatlyaffectedbythearraygeometriccon- figuration.•
By comparing the results given from Type 1 to Type 4, the configurationthat putstwo sensors attheextremity andthe restinthemiddleseemstobethebestgeometryconfiguration (atleast,itisthecaseforA=
9 and N=
6 sensors).Fig. 10.TherequiredSNRmintoresolvetwounknowncloselysourcesfordifferent arraygeometries.
Fig. 11.TherequiredSNRmintoresolvetwounknowncloselysourcesforthebest&
worstperformanceofdifferentsensors.
Table 2
Thebestandworstarraygeometricconfigurationsfordifferentnumberofsensors.
Array type Geometric configuration SNRmin
Type 6 • ◦ ◦ ◦ ◦ • ◦ ◦ ◦ • lowest SNRmin
Type 7 • ◦ ◦ ◦ • • ◦ ◦ ◦ • lowest SNRmin
Type 8 • • ◦ ◦ ◦ ◦ ◦ ◦ • • highest SNRmin
Type 9 • ◦ ◦ ◦ • • • ◦ ◦ • lowest SNRmin
Type 10 • • • ◦ ◦ ◦ ◦ ◦ • • highest SNRmin
Type 11 • ◦ ◦ • • • • ◦ ◦ • lowest SNRmin
Type 12 • • • ◦ ◦ ◦ ◦ • • • highest SNRmin
Type 13 • • • • ◦ ◦ ◦ • • • highest SNRmin
•
Type4,withsixsensors, isslightlyworsethan theULAcon- figurationwith10sensors.Furthermore, to better comprehend the resolvability perfor- mance as a function of array geometric configuration, consider Fig. 11,wherewecomparetheminimumsignaltonoiseratiowith respecttothebestandworstperformance ofdifferentnumberof sensors. Theconsidered type ofsensor geometryconfigurationis listinTable 2.
From thisfigure, we can see that the Type 11represents the best achievable performance, and Type 8 has the worst perfor- mance. Furthermore, on one hand, we can notice that with the same array aperture, Type6 (with3 sensors)andType13 (with 7 sensors)have,approximately,the sameperformance (Type 6is only 0.2 dB worse than the Type 13); and, on the other hand, one can notice that Type6 (with3 sensors)has a better perfor- manceinregard toType12(whichhas6sensors)by1.5 dB.This meansthatdependingonthearraygeometry,onecanhavebetter performance using a lessnumber of sensors. The relatively poor performance ofthetype8,10,12and13isduetothesensorge- ometry configuration.Consequently, withthe same aperture,one canoptimallydesignthearrayusingtheproposedmethodbelow, inordertoobtainbetterperformanceevenwithlesssensors.
5.6. Sensorsarraygeometrydesign
Asdiscussedabove,theresolvabilityisnotonlyconstrainedby thenumberofsensors,butalsobythearray geometry,whichaf- fectstheoptimalresolutioncapability.
In the following, we give a solution to design an optimalar- rayinordertoenhancetheresolvabilityand/ortocompensatethe performancelossduetothereductionorfailureofcertainsensors.
The optimalmethod isgiven by minimizing SNRmin over sen- sors’ positionamongallpossibilities usingthederived expression (47).However, thisschemeisnot feasiblefora modelwithlarge apertureandsmallnumberofsensors.Moreprecisely,foraperture ofsize A and N sensors,onehas CNA+1 possibilities.Toovercome thisdrawbackweproposeasub-optimalfastprocedureasfollows.
Step 1:We placetwo sensors ateach extremitytoobtain the largestaperturethatleadstothebestresolution.
Step 2:We placeonly onesensor by minimizing SNRmin with respecttothe A
−
2 remainingpositions.Step 3: We iterate the second step by placing the
(
n+
1)
-th sensoratoncebyminimizingsequentiallytheSNRminwithrespect tothe A−
n+
1 remainingpositions.In thefollowingpart,weshow that theproposed sub-optimal method produces a nearly optimal result with low complexity.
More precisely, the complexity cost ratioof the optimal method overtheproposedsub-optimalmethodisgivenby
f
(
N,
A) =
Nopt Nprop=
2(
A+
1) !
N
!(
A−
N+
1)!(
A(
A−
1) − (
A−
N+
1)(
A−
N+
2)) (52)
where Nopt meansthenumberofrequirediterationsbyusingthe optimalmethodandNprop meansthatbyusingtheproposed sub- optimalmethod.In the examples studied in this section, we consider several cases: Case 1) N
=
4, A=
7; Case 2) N=
5, A=
8, so that we maycomparetheresultsofthetwomethods,asshowninFig. 12.One notes that our proposed method is nearly identical to the optimal method.Inaddition, from(52),we have f
(
4,
7) =
6 and f(
5,
8) =
7,whichmeansthattheproposedmethodissixtoseven times fasterthan the optimalone. Ifwe consider a moreimpor- tantaperture,e.g.,forN=
6, A=
25,wenoticethattheproposed method decreasessignificantly the complexity cost by f(
6,
25) =
2558 times compared with the optimal one with the same per- formance.Consequently,wenoticethattheproposedmethodisof lowcomplexity,especiallyforthedesignofsparsearrays.Oneshouldnotethatfortheabovediscussion,weassumethat derived closed formexpression of SNRmin is given. If no expres- sion ofthe SNRmin is given, one has,first,to numericallyinverse