an author's
https://oatao.univ-toulouse.fr/26816
https://doi.org/10.1016/j.sigpro.2020.107792
Medina, Daniel and Vilà Valls, Jordi and Chaumette, Eric and Vincent, François and Closas, Pau Cramér-Rao bound
for a mixture of real- and integer-valued parameter vectors and its application to the linear regression model. (2021)
Signal Processing, 179. 107792. ISSN 0165-1684
Cramér-Rao
bound
for
a
mixture
of
real-
and
integer-valued
parameter
vectors
and
its
application
to
the
linear
regression
model
Daniel Medina
a,∗, Jordi Vilà-Valls
b, Eric Chaumette
b, François Vincent
b, Pau Closas
c a Institute of Communications and Navigation, German Aerospace Center (DLR), Germanyb ISAE-SUPAERO/University of Toulouse, DEOS Dept., Toulouse, France
c Electrical and Computer Engineering Dept., Northeastern University, Boston, MA, USA
Keywords: Cramér-Rao bound McAulay-Seidman bound
Mixed real-integer parameter vector estimation
Linear regression GNSS
Ambiguity resolution
a
b
s
t
r
a
c
t
Performancelowerboundsareknowntobeafundamentaldesigntoolinparametricestimationtheory.A plethoraofdeterministicboundsexistintheliterature,rangingfromthegeneralBarankinboundtothe well-knownCramér-Raobound(CRB),thelatterprovidingtheoptimalmeansquareerrorperformance oflocallyunbiasedestimators.Inthiscontribution,weare interestedintheestimationofmixed real-andinteger-valuedparametervectors.Weproposeaclosed-formlowerboundexpressionleveragingon thegeneralCRBformulation,beingthelimitingformoftheMcAulay-Seidmanbound.Suchformulation isthekeypointtotakeintoaccountinteger-valuedparameters.Asaparticularcaseofthegeneralform, weprovideclosed-formexpressionsfortheGaussianobservationmodel.Onenoteworthypointisthe as-sessmentoftheasymptoticefficiencyofthemaximumlikelihoodestimatorforalinearregressionmodel withmixedparametervectorsand knownnoisecovariancematrix,thuscomplementingtheratherrich literatureonthattopic.Arepresentativecarrier-phasebasedprecisepositioningexampleisprovidedto supportthediscussionandshowtheusefulnessoftheproposedlowerbound.
1. Introduction
Integer parameter estimation appears inmany signal process-ing, biology and communications problems, to name a few. For instance, consider a multi-hypothesis testing problem where we want to identify the received signal over a (finite) set of possi-ble transmitted signals, then a solution is to maximize the log-likelihood function over the (integer) set of candidates. Another problem involvingestimation ofinteger quantities,jointly witha real-valued vector, isthat ofcarrier-phasebasedprecise position-ing in the context of Global Navigation Satellite Systems (GNSS) receivers.Inthegeodesyandnavigationcommunity,awellknown estimation approach is referred to asReal Time Kinematic (RTK) positioning[1].Carrier-phase measurementshaveanunknown in-teger part,referred to asthe ambiguity,to be estimatedin order toachievecm-levelaccuracyonthereal-valuedunknownposition ofthereceiver.TheframeworkthatunderpinspreciseGNSScarrier phase-basedambiguityresolutionisthetheoryofintegeraperture
∗Corresponding author.
E-mail addresses: daniel.ariasmedina@dlr.de (D. Medina), jordi.vila-valls@isae.fr
(J. Vilà-Valls), eric.chaumette@isae.fr (E. Chaumette),closas@ece.neu.edu (P. Closas).
estimation [2,3], which also applies to other carrier phase-based interferometrictechniques,suchasVeryLongBaseline Interferom-etry (VLBI) [4], Interferometric Synthetic Aperture Radar (InSAR)
[5],orunderwateracousticcarrierphase-basedpositioning[6]. Regardlessoftheestimationproblemaddressed, when design-ing andassessing estimators it is of fundamental importance to knowtheminimumachievableperformance,thatis,toobtaintight performance lower bounds (LBs). In general, in estimation prob-lems we are interested in minimal performance bounds in the meansquared error(MSE) sense, which providethe best achiev-ableperformanceontheestimationofparametersofasignal cor-ruptedbynoise.TherearetwomaincategoriesofLBs,deterministic
andBayesian[7].While theformerconsiders thatthe parameters tobeestimatedaredeterministic andevaluate thelocallybest es-timatorperformance, thelatterconsiderrandom parameterswith a givena priori probability andevaluate theglobally best estima-torbehavior.Inthiscontributionweareinterestedindeterministic parameterestimation,thusonlythefirstclasswillbediscussed.
It is worth saying that such LBs have been proved to be ex-tremelyuseful,notonlyforcharacterizinganestimatorasymptotic performance, butalsoforsystem design[7–9].The mostpopular LB is the well-known Cramér-Rao Bound (CRB)derived for real-valuedparametervector,mainlydueto:i)itssimplicityof calcula-https://doi.org/10.1016/j.sigpro.2020.107792
tion,forinstanceusingtheSlepian-Bangs’formula[10];ii)itisthe lowest bound onthe MSE ofanyunbiased estimator(i.e.,it con-siders localunbiasedness at the vicinity ofany selected parame-ters’value);andiii)itisasymptoticallyattainedbymaximum like-lihoodestimators(MLEs)undercertainconditions(i.e.,high signal-to-noise ratio (SNR) [11] and/or large numberof snapshots[12]), that is, MLEs are asymptotically efficient. Inherent limitations of such CRBsare their inabilityto:predict thethresholdphenomena; provide tightboundsincertaincases[13];anddeal with integer-valuedparameterestimation,whichisthecontributionofthis ar-ticle.
SincetheseminalCRBworks,severaldeterministicboundshave beenproposedintheliterature[14–22]toprovidecomputable ap-proximationsoftheBarankinbound(BB)[23],whichisthetightest (greatest) LBforanyabsolutemoment ofordergreater than1 of unbiasedestimators.Infact,theBBconsidersuniformunbiasedness
(i.e., unbiasedness over an interval ofparameter values including the selected value),resulting in a much strongerrestriction than thelocalunbiasednessconditionoftheCRB,butnotadmittingan analyticsolutioningeneral.
In thiscontribution,inorder toobtain a CRB-likeclosed-form expression forthe estimationofmixedparametervectors, includ-ing both real-andinteger-valuedparameters, we leverageonthe McAulay-Seidman bound (MSB)[16].The MSBis the BB approxi-mationobtainedfromadiscretizationoftheBarankinuniform un-biasednessconstraint,usingasetofselectedvaluesofthe param-etervector, so-calledtestpoints.TheMSByields toageneral def-inition ofthe CRB, expressedas a limitingform ofthe MSB. We derive aclosed-formgeneralCRBexpressionformixedparameter vectorsandprovideitsparticularclosed-formfortheGaussian ob-servation model,which encompassesthe well known conditional andunconditionalobservationmodels[24].
On another note, one must keep in mind that inmany prob-lemsofpracticalinterest,includingthegeneral(nonlinear)caseof theproblemunderconsideration,noevidenceoftheachievability ofa givenLBby realizableestimatorsexists [7,8,13].Thus,froma practicalperspective,theLBconsideredmaybetoooptimisticand unable to represent the actual performance of anyestimator. To circumventtheunavailabilityofLBachievabilityresults,asolution relatesto thederivation ofanupperbound toprovidea comple-mentaryvision tothat oftheLB.Unfortunately, upperboundson theMSEofunbiasedestimatesdonotgenerallyexistifthe obser-vationspaceisunbounded.Nonetheless,upperboundsonthe sta-tistical performance (notnecessarily theMSE) mayexistfor spe-cific estimators (not necessarily unbiased) in specific estimation problems [1,25–27]. Inparticular,forthe mixedintegerlinear re-gression model, a rich literature on the statistical performances of various estimators is alreadyavailable (see [1] and references therein), andan upperbound onthe probability that theMLE of thereal-valuedparametervectorliesinacertainregionexists[25]. Remarkably,aLBontheambiguitysuccessrate(probabilityof cor-rect estimationoftheinteger-valuedparametervector)doesexist aswell andappearsto be thecornerstonetoprove thatthe MLE of the real-valued parametervector is efficientin the asymptotic regime (high SNR) when the noise covariance matrix is known, thus completingtheliterature onthattopic(see Section3for de-tails). Asausecase,we considerthelinearregressionproblemin the contextofGNSSprecisepositioning tosupportthediscussion andshowtheusefulnessoftheproposedperformanceLB.
The article is organized as follows: Section 2 provides back-groundondeterministicLBsandtheir derivationasanorm mini-mizationproblem,mainlyfocusedontheCRBasthelimitingform oftheMSB. Section 3detailsthederivation ofthenewbound, in the general case and for the Gaussian observation model. It es-tablishesthe asymptoticefficiencyoftheMLE fora linear regres-sionmodelwithmixedparametervectorsandknownnoise
covari-ance matrix, and sketches possible generalizations and outlooks. Theseresultsare thenparticularized foralinearregression prob-lem,servingasmotivatingexampleanddiscussedinSection4.The paperconcludeswithadiscussionoftheresultsinSection5. 2. BackgroundonMcAulay-SeidmanandCramér-Raobounds forareal-valuedparametervector
2.1. TheMcAulay-Seidmanbound
Lety1bearandomreal-valuedobservationsvectorand
⊂ RM
the observation space. Denote by p(y;
θ
)p(y|θ
) the pdf of the observationsconditionalonanunknowndeterministicreal-valued parametervectorθ
∈⊂ RK.LetL
2
(
)
be thereal vectorspace ofsquare integrablefunctionsover.Ifweconsideranestimator
g
θ
0(
y)
∈LN2
(
)
ofg(θ
0),whereθ
0 isaselectedvalueofthe pa-rameterθ
andg(
θ
)
=(
g1(
θ
)
,...,gN(
θ
))
isareal-valued functionvector,thentheMSEmatrixwrites,
MSEθ0
gθ
0=E y;θ0 gθ
0(
y)
− gθ
0(
·)
. (1) By noticingthat (1)is a Gram matrix associated withthe scalar producth(
y)
|
l(
y)
θ0=Ey;θ0[h(
y)
l(
y)
],thesearch fora LBontheMSE (1) (w.r.t. the Löwner ordering for positive symmetric ma-trices [28]) can be performed with two equivalent fundamental results: the generalization of the Cauchy-Schwartz inequality to Gram matrices (generally referred to as the "covariance inequal-ity”[18]) and the minimization ofa norm under linear constraints (LCs)[17,19,20].We shallpreferthe“norm minimization” formas its userequiresexplicitlytheselection ofappropriate constraints, which then determine the value of the LB on the MSE matrix, hence providing a clear understanding of the hypotheses associ-ated with the different LBs on the MSE. To avoid the trivial so-lutiong
(
θ
0)
(
y)
=g(
θ
0)
, some constraintsmustbe added.In that perspective,Barankin[23]introducedtheuniformunbiasedness for-mulation,Ey;θ
g
θ
0(
y)
=gθ
,∀
θ
∈, (2a)
leadingtotheBarankinbound(BB), min g
(
θ0)
∈LN 2() MSEθ0 gθ
0s.t. E y;θ g
θ
0(
y)
=gθ
,∀
θ
∈, (2b) which does not admit an analytic solution in general. The McAulay-Seidman bound(MSB)isthecomputable BB approxima-tion obtained from a discretization of the uniform unbiasedness
constraint(2a).Let
{
θ}
L{
θ}
[1,L]=
{
θ
1,...,θ
L}
∈Lbeasubsetof Lselectedvaluesof
θ
(a.k.a.testpoints). Then,anyunbiased esti-matorg(
θ
0)
(
y)
verifying(2a)mustcomplywiththefollowing sub-1 Italic indicates a scalar quantity, as in a ; lower case boldface indicates a col-umn 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,m or [ A ] n,m . The n th coordinate of the column vector a will be denoted by a n or
[ a ] n . The matrix/vector transpose is indicated by a superscript ( · ) as in A . | A | is
the determinant of the square matrix A . [ A B ] and A B
denote the matrix resulting from the horizontal and the vertical concatenation of A and B , respectively. I M is the
identity matrix of dimension M . 1 M is a M -dimensional vector with all components
equal to one. For two matrices A and B, A ≥ B means that A − B is positive semi- definite (Löwner partial ordering) . · denotes a norm. If θ= [ θ1 , θ2 , . . . , θK ] ,
then: ∂θ∂ = ∂θ∂1 , ∂θ∂2 , . . . , ∂θ∂K, ∂θ∂ = ∂ ∂θ1 , ∂θ∂2 , . . . , ∂θ∂K and ∂h(θ0,y) ∂θ = ∂ h(θ,y) ∂θ θ 0 . p ( y ; θ) p ( y | θ) denotes the probability density function (pdf) of y parameterized by
θ. E y;θ[ g ( y ) ] denote the statistical expectation of the vector of functions g ( · ) with
setofLLCs, Ey;θl
g
θ
0(
y)
=gθ
l, 1≤ l≤ L, (3a)whichcanberecastas Ey;θ0
υ
θ0 y;θ
L gθ
0(
y)
− gθ
0 =⎡
⎢
⎣
gθ
1− gθ
0 . . . gθ
L− gθ
0⎤
⎥
⎦
V , (3b) whereυ
θ0(
y;{
θ}
L)
=(
υ
θ0(
y;θ
1)
,...,υ
θ0(
y;θ
L))
,υ
θ0(
y;θ
)
=p
(
y;θ
)
/p(
y;θ
0)
is the vector of likelihood ratios associated to {θ
}L.TheLLCs(3b)yields theapproximationof(2b)proposedbyMcAulayandSeidman[20], min g
(
θ0)
∈LN 2() MSEθ0 gθ
0s.t.
=V, (4a)
anddefinestheMSB(Lemma1in[29])[16,20] Ey;θ0
gθ
0(
y)
− gθ
0(
·)
≥g
θ
0R−1 υθ0g
θ
0,g
θ
0g
θ
0,θ
L =gθ
1− gθ
0 ... gθ
L− gθ
0, Rυθ0 Rυθ0θ
L =Ey;θ0υ
θ0 y;θ
Lυ
θ0 y;θ
L, (4b) a generalization of the Hammersley-Chapman-Robbins bound (HaChRB)previouslyintroducedin[15,30]for2testpoints(
L=2)
.2.2. CRBasalimitingformoftheMSB
TheCRBcanbedefinedforanyabsolutemoment(greaterthan 1) asthelimitingformoftheHaChRB [15,30],asshowedin[23]. Theextensiontothemultidimensionalreal-valuedparameterscase forthe MSE(i.e.,absolutemomentoforder2) wasintroduced in
[16],allowing to define the CRBasthe limitingformofthe MSB
(4b).Consideringthesubsetoftestpoints
θ
1+K =θ
0,θ
0+i 1dθ
1,...,θ
0+iKdθ
K underdθ
k=0, 1≤ k≤ K,whereikisthekthcolumnoftheidentitymatrixIK,leadsto υθ0 y;θ1+K =1 p(y;θ0+i1dθ1) p (y;θ0) ... p(y;θ0+i KdθK) p (y;θ0) , g θ0=0 gθ0+i 1dθ1 − gθ0 ... gθ0+i K dθK − gθ0,
and,withd
θ
=(
dθ
1,...,dθ
K)
,yieldsto(seeAppendixA)g
θ
0R−1 υθ0g
θ
0=g
θ
0,dθ
Fθ
0,dθ
−1g
θ
0,dθ
, (5a) Fθ
0,dθ
=E y;θ0⎡
⎢
⎢
⎢
⎣
⎛
⎜
⎜
⎜
⎝
p(
y;θ0+i 1dθ1)
−p(
y;θ0)
dθ1p(
y;θ0)
. . . p(
y;θ0+i KdθK)
−p(
y;θ0)
dθKp(
y;θ0)
⎞
⎟
⎟
⎟
⎠
(
.)
⎤
⎥
⎥
⎥
⎦
(5b)g
θ
0,dθ
= g(
θ0+i 1dθ1)
−g(
θ0)
dθ1 ... g(
θ0+i KdθK)
−g(
θ0)
dθK , (5c) whichresultsinageneraldefinitionoftheCRBg|θ(θ
0)asCRBg | θ θ0= lim sup { d θ1=0, ... ,d θK=0 } →0 gθ0,dθFθ0,dθ−1g θ0,dθ. (6a)
If
θ
0∈⊂ RKandg(
θ
)andp(y;θ
)areC1atθ
0,then(6a)yields thewellknownFisherInformationMatrix(FIM)F(θ
)andtheusual CRBexpression Fθ
=Ey;θ"
∂
lnpy;θ
∂θ
(
.)
#
, (6b) CRBg|θθ
0=∂
gθ
0∂θ
Fθ
0−1$
∂
gθ
0∂θ
%
. (6c)3. CRBforamixtureofreal-valuedandinteger-valued parameters
Leveragingthe MSBandCRBresultspresentedintheprevious
Section2,wederiveinthissectionaLBfordeterministic parame-tervectorestimation,wheresuchvectorcontainsbothreal-valued andinteger-valued parameters.A generalresultis provided, then particularizedforthecaseofGaussianobservations.
3.1. GeneralCRBformixedparametervectors
The main result derived in this article is summarized in the formof Theorem1. Acorollary follows,which simplifiesthe for-merinaparticularclassofmodels.
Theorem1(General CRBformixedparametervectors). Assumea setof observations y∈
⊂ RM and anunknown deterministic real-valuedparametervector
θ
∈⊂ RKwhere
θ
=[ω
,z],ω
∈RKω,z∈ZKz ,Kω+Kz=K.Thosequantitiesarerelatedthroughastatistical
modelof theformy|
θ
~ p(y|θ
),whichis available.Then,theMSEof any unbiased estimator of a function gθ
0∈L2
(
)
for a selectedvalueoftheparameter
θ
0 islowerboundedbyCRBg|θ
θ
0=g
θ
0Fθ
0−1g
θ
0, (7) withg
θ
0=∂g(
θ0)
∂ω gθ
1− gθ
0 ... gθ
2Kz− gθ
0 (8) Fθ
0=Fω|θ
θ
0 Hθ
0 Hθ
0 MS z|θθ
0 , (9)wherethetestpoints
θ
2Kz aredefinedasθ
j=θ
0+(
−1)
j−1i Kω +j+1 2 , 1≤ j≤ 2Kz, (10) thatis,θ
1,θ
2,...,θ
2Kz−1,θ
2Kz =θ
0+i Kω +1,θ
0− iKω +1,...,θ
0+iK,θ
0− iK .ThedifferenttermsinF
θ
0aregivenbyFω|θ
θ
0=E y;θ0"
∂
lnpy;θ
0∂ω
(
.)
#
, (11a) Hθ
0=E y;θ0"
∂
lnpy;θ
0∂ω
t2Kz#
(11b) =hθ
0,θ
1 hθ
0,θ
2 ... hθ
0,θ
2Kz, (11c) MSz|θθ
0=E y;θ0 t2Kzt2Kz − 12Kz12Kz, (11d) 3wheret2Kz isdefinedas t2Kz
υ
θ0 y;θ
2Kz =$
py;θ
1 py;θ
0, py;θ
2 py;θ
0,..., py;θ
2Kz py;θ
0%
. (11e) Proof. First, notice that in the real-valued parameter case, that is, ifθ
0k∈R, and both g(
θ
) and p(y;θ
) are C1 atθ
k0, then,the constraints associated to the following two test points,
θ
0+i kdθ
k,θ
0+ik(
−dθ
k)
=θ
0+i kdθ
k,θ
0− ikdθ
k , Ey;θ0⎡
⎣
⎛
⎝
p(
y;θ0+i kdθk)
−p(
y;θ0)
dθkp(
y;θ0)
p(
y;θ0−i kdθk)
−p(
y;θ0)
(−dθk)p(
y;θ0)
⎞
⎠
gθ
0(
y)
− gθ
0⎤
⎦
=⎡
⎣
g(
θ0+i kdθk)
−g(
θ0)
dθk g(
θ0−i kdθk)
−g(
θ0)
(−dθk)⎤
⎦
(12)aim at the same single constraint in the limiting case where
d
θ
k→0,dθ
k=0, Ey;θ0"
∂
lnpy;θ
0∂θ
k gθ
0(
y)
− gθ
0#
=∂
gθ
0∂θ
k . (13)However, this phenomenon isunlikely tohappen if
θ
0k ∈Z inthe
limitingcasewhered
θ
k→0,dθ
k=0,since(12)thenbecomesEy;θ0
⎡
⎣
⎛
⎝
p(
y;θ0+i k)
−p(
y;θ0)
p(
y;θ0)
p(
y;θ0−i k)
−p(
y;θ0)
p(
y;θ0)
⎞
⎠
gθ
0(
y)
− gθ
0⎤
⎦
="
g
θ
0+i k − gθ
0 gθ
0− i k − gθ
0#
, (14) where py;θ
0+i k − py;θ
0/py;θ
0 and py;θ
0− i k− p
y;θ
0/py;θ
0 are unlikely to be linearly dependent (i.e., notice that Fθ
0,dθ
in (5b) must be invertible to compute the CRBg|θ(θ
0) in (6a)). Therefore, in most cases, the combination of LCs (13) and (14) yields, from Lemma 1 in [29], the generaldefinition(7)ofCRBg|θ(θ
0) wherethedifferent terms inFθ
0aregivenby Fω|θθ
0=E y;θ0"
∂
lnpy;θ
0∂ω
(
.)
#
, (15a) Hθ
0=E y;θ0"
∂
lnpy;θ
0∂ω
(
t2Kz− 12Kz)
#
, (15b) MSz|θθ
0=E y;θ0(
t2Kz− 12Kz)
(
t2Kz− 12Kz)
, (15c)and where H(
θ
0) and MSz|θ(
θ
0) can also been expressed as (11b)and(11d).Corollary1. Ifg
(
θ
)
=θ
,matrixθ(
θ
0)inTheorem1simplifiestoθ
θ
0 = i1 ... iKω iKω +1 −iKω +1 ... iK −iK = (Kz=3)⎡
⎢
⎣
IKω 0 0 0 0 0 0 0 1 −1 0 0 0 0 0 0 0 1 −1 0 0 0 0 0 0 0 1 −1⎤
⎥
⎦
. (16)3.2. TheGaussianobservationmodel
Let us consider an M-dimensional Gaussian real vector y with mean my=m
θ
and covariance matrix Cy=Cθ
: y∼NM
m
θ
,Cθ
andpy;θ
=py; mθ
,Cθ
suchthatp
y; mθ
,Cθ
= e− 1 2(
y−m(
θ)
)
C−1(
θ)
(
y−m(
θ)
)
√ 2π
M&
C
θ
. (17) Ifwedefine Ci j=Cθ
i−1+Cθ
j−1− Cθ
0−1−1, (18a) mi j =Cθ
i−1mθ
i+Cθ
j−1mθ
j− Cθ
0−1mθ
0, (18b)δ
i j =mθ
iCθ
i−1mθ
i+mθ
jCθ
j−1mθ
j − mθ
0Cθ
0−1mθ
0, (18c) thenwecanobtainthedifferentcomponentsrequiredtocompute the CRB (7)(see Appendix B fordetailed derivation of MSz|θ(θ
0) andh(θ
0,θ
j))as MSz|θθ
0 i, j='
Ci jC
θ
0 Cθ
iCθ
je 1 2(
mi j)
Ci jmi j−δi j − 1, (18d) hθ
0,θ
j k=⎛
⎜
⎜
⎜
⎜
⎝
1 2tr(
∂C(
θ0)
−1 ∂ωk Cθ
0− Cθ
j−∂C(
θ0)
−1 ∂ωk ×mθ
j− mθ
0mθ
j− mθ
0 +∂m∂ω(
θk0)
Cθ
0−1mθ
j− mθ
0⎞
⎟
⎟
⎟
⎟
⎠
(18e) Fω|θθ
0 k,l =∂
mθ
0∂ω
k C−1θ
0∂
mθ
0∂ω
l +1 2tr$
C−1θ
0∂
Cθ
0∂ω
k C−1θ
0∂
Cθ
0∂ω
l%
, (18f) where(18f)istheSlepian-Bangsformula[31,p.47].Inthefollowing,forsakeoflegibility,
θ
denoteseitherthe vec-tor of unknown parameters or a selected vector valueθ
0= [ω
0,z0].3.3. AsymptoticallyefficientestimatorsfortheGaussianlinear conditionalobservationmodel(g
(
θ
)
=θ
)Acaseof particularinterest is theGaussian linear conditional signalmodel,alsoknownasthemixed-integermodel[1,Ch.23],
y=B
ω
+Az+n, n∼ NM(
0,Cn)
,θ
=[ω
,z],ω
∈RKω , z∈ZKz, (19)wherethenoise covariancematrixCnis knownandthe
parame-tervectorofinterestisg
θ
=θ
.Forinstance,inGNSSRTKprecise positioning,theM-vectorycontainsthepseudorangeand carrier-phaseobservables,theKz-vectorztheintegerambiguities,andthereal-valuedKω-vector
ω
theremainingunknown parameters,such as, forexample, position coordinates,atmospheric delay parame-ters (troposphere, ionosphere) and clock parameters. The theory that underpins theresolution of(19) in themaximum likelihoodsenseisthetheoryofintegerinference[2,3][1,Ch.23].Thesearch for the MLE of a selected value
θ
for the mixed-integer model(19)canbecastasaminimizationproblemovermixedinteger-real parameters, )
θ
=arg min ω∈RKω ,z∈ZKz**
y− Dθ
**
2C n, D= B A. (20)A closed-form solution to (20) is not known,due to the integer natureofz.Instead,athree-stepdecompositionoftheproblemis typicallyconsidered[32],andtheresultingminimizationproblems aresequentiallyresolvedas[33]
min ω∈RKω ,z∈ZKz
**
y− Dθ
**
2C n=**
**
y− D(
¯ω
¯z+**
**
2 Cn (21a) +min z∈ZKz¯z− z 2 Cω¯ (21b) + min ω∈RKω )ω
(
z)
−ω
2 C)ω( z) , (21c)where)
ω
(
z)
=ω
¯− Cω¯,¯zC−1¯z(
¯z− z)
.Thefirstterm(21a)corresponds to the MLE solution where z is treated as a real valued vector (instead of aninteger valuedvector).The output ofthisestimate¯
θ
=[ω
¯,¯z]isreferredtoasfloatsolutionanditsassociated co-variancematrixis Cθ¯=Cω¯ Cω¯,¯z C¯z,ω¯ C¯z =DC−1 n D −1 ,
which, by exploiting the four-blocks matrix inversion expression
[28],leadsto C)ω(z)=Cω¯− Cω¯,¯zC−1¯z C¯z,ω¯ =
BC−1 n B −1 .The second term (21b) in the decomposition corresponds to the integer-least-square(ILS),forwhichanintegersolution)zisfound. Finally,thethirdterm(21c)isthe fixedsolution,consistingon en-hancingtheestimates
ω
¯ upontheestimatedintegervector)z )ω
=)ω
(
)z)
=ω
¯ − Cω¯,¯zC−1¯z(
¯z−)z)
. (22)The improvement in )
ω
accuracy is dueto constrainingthe float solution ¯z to a more restrictive class of estimators. Three differ-ent classes of estimators have beendeveloped for mixedinteger models [34], andforeach one the optimalestimators have been identified:i)theclassofinteger(I)estimators[35];ii)theclassof integer-aperture (IA) estimators[36];andiii)the classof integer-equivariant (IE)estimators[37].Thefirstclassisthemost restric-tive class. Thisis due to the fact that the outcomes ofany esti-mator within this class are required to be integers. Well-known examples of estimators fromthis class are integerrounding ()zR),integerbootstrapping()zB)andtheoptimalsolutionso-called
inte-ger least-squares ()zLS) which is the MLE. The most relaxedclass
istheclassofIE-estimators.Theseestimatorsarereal-valued,and theyonlyobeytheintegerremove-restoreprinciple.Animportant estimator in thisclass is the best IE-estimator ()zBIE) since it has
the smallestvariance,even smallerthan thevariance ofthe float solution.TheclassofIA-estimatorsisasubsetoftheIE-estimators butitencompassestheclassofI-estimators.TheIA-estimatorsare ofahybridnature inthesense thattheir outcomesareeither in-tegers or real. It is also worth noting that distributional results are readily available [25,38].Interestingly enough, integer round-ing,integerbootstrapping,andintegerleast-squaresestimators()zR,
)zB,)zLS)areuniformlyunbiased[39]underGaussianadditivenoise
(19),leadingtoanuniformlyunbiasedestimator
ω
)(22),sincethen E[)ω
]=E[ω
¯]=ω
.Thus theproposed CRBθ|θ(θ
) (7)isarelevantLB fortheGaussianlinearconditionalsignalmodel(19)andC)ω=C)ω()z)≥ CRBω|θ
θ
,)z∈{
)zR,)zB,)zLS}
. Firstly, as [1, (23.54)] P(
)zLS=z)
≥ P(
)zB=z)
≥ P(
)zR=z)
, and [1, (23.23)] lim tr(Cn)→0 P(
)zR=z)
=1,then lim tr(Cn)→0 C)zLS= lim tr(Cn)→0 C)zB= lim tr(Cn)→0 C)zR=0.Thus, for any )z∈
{
)zR,)zB,)zLS}
, since [25, (29)] C)ω()z)=C)ω(z)+ Cω¯,¯zC−1¯z C)zC−1¯z C¯z,ω¯, lim tr(Cn)→0 Cω)()z)=BC−1n B −1 .Secondly,sinceitiswell knownthat addingunknownparameters leadstoanequalorhigherCRB,then(25a)
C)ω()z)≥ CRBω|θ
θ
≥ F−1ω|θθ
=BC−1 n B −1 .Therefore,forany)z∈
{
)zR,)zB,)zLS}
,lim tr(Cn)→0 Cω)()z)=tr(limC n)→0 CRBω|θ
θ
=BC−1 n B −1 ,which proves that )zR,)zB and)zLS are asymptotically efficient
es-timators. Last, since)zBIE is also uniformly unbiased with a MSE
lessthan oratthemostequaltotheMSEof)zLS[37,(24)],itisan
asymptoticallyefficientestimatoraswell.
3.4. Generalizationsandoutlooks
TheproposedCRBformixedparametervectors(7)hasbeen de-rived in thecontext of “standard” deterministicestimation prob-lemsforwhichaclosed-formexpressionofp(y;
θ
)isavailable.In the context of “non standard” deterministic estimation problems (see[40]andreferencestherein),p(y;θ
)resultsfromthe marginal-ization of an hybrid p.d.f. dependingon both randomθ
r∈RPrand deterministic (
θ
) parameters, i.e., py;θ
=,RP r py,θ
r|
θ
dθ
r,which ismathematically intractable andprevents fromusing the proposed standardCRB (7). Fortunately,anyLB deriving fromthe MSB, as for instance (7), can be transformed into two variants, namely the so-called “modified” LB [40, Section III] and “non-standard” LB [40,Section IV]fitted tonon-standard deterministic estimation. Thus, two CRB variants for mixed parameter vectors canbe readilyintroduced inthecontext of“nonstandard” deter-ministic estimation. Since the proposed CRB (7) can also be re-gardedasaCRBdealingwithrestrictingthesetofpossiblevalues of some of the unknown parameters, namely the integer-valued parameters,itisworthnoticingthat(7)canalsotakeintoaccount continuous restriction on the set of possible values of the real-valued parameters. When the continuous restriction is described by a set of P equality constraints, f
(
ω
)
=0∈RP, 1≤ P≤ Kω− 1, wherethematrix
∂
f(
ω
)
/∂
ω
∈RP×Kω hasfullrowrank(P),which is equivalentto requiringthat the constraintsare not redundant, it leads to the so-calledconstrained CRB [41–44]. In the caseof mixedparametervectors,it amountstoreplacetheLCs(13)with[29] Ey;θ
"
$
U(
ω
)
∂
lnp y;θ
∂ω
%
g
θ
(
y)
− gθ
#
=$
∂
gθ
∂ω
U(
ω
)
%
,where U
(
ω
)
∈RKω×(Kω−P) is a basis of ker∂f(ω) ∂ω, which leads to update the definition of Fω|θ(
θ
) (11a), H(θ
) (11b) andg(
θ
)(8) as follows: Fω|θ(
θ
) → U(ω
)Fω|θ(θ
)U(ω
), H(θ
) → U(ω
)H(θ
), andg
θ
=∂∂ωg(
θ)
U(
ω
)
gθ
1− gθ
0 ... gθ
2Kz − gθ
0. 54. ExampleofGaussianlinearregressionproblem:GNSSRTK precisepositioning
Inthissection,weexemplifytheaforementionedresultswitha particularexampleofGaussianlinearconditionalsignalmodel,aka thelinearregressionproblemwithmixedrealandinteger param-eters,thatistheGNSSRTKprecisepositioningproblem.
4.1. SignalmodelforGNSSRTKprecisepositioning
RTK is a GNSS-based relative positioning method, where the position of a target is determined with respect to a base sta-tion of known coordinates [45]. RTK exploits the use of code and carrier-phase pseudorange observations. Carrier-phase obser-vationsarecharacterizedbyanoise(typically)twoordersof mag-nitude lower than code pseudorange measurement, but they are ambiguous by an unknown numberof integer ambiguities.Thus, the achievement of high precision positioning requires the esti-mation ofthe dynamicalparameters ofthe target alongwiththe unknown integerambiguitieswithinaprocessreferred toas Inte-gerAmbiguityResolution(IAR)[33].Weassume thatM+1 satel-lites are tracked simultaneously by thebase andtarget GNSS re-ceivers. First, single-differencingthe observations,i.e., subtracting the observations atthe target from the base stations, eliminates the atmosphericandsatellite-related delays.Then, the process of double-differencing,i.e., subtracting thesingle-difference observa-tions withrespect toa reference satellite, removesthe effects of receiversclockoffsets.TheresultingGNSSRTKmodelcanbe mod-eledasy∼ NM
m(
θ
)
,Cnwith m(
θ
)
=Dω
z , D=[BA], B=B B , A=
A 0 , Cn=
Cφ 0 0 Cρ , (23)
where y=
φ
,ρ
is theobservation vector, composed by the double-differencecarrier-phaseandcodeobservations,denotedasφ
,ρ
∈RM respectively, whose corresponding covariance matricesare Cφ andCρ.
ω
∈R3 isthetarget receivertobasestation base-linevector;andz∈ZM isthevectorofunknownintegerambigui-ties. ThematrixBistheso-calledgeometrymatrixwhichis com-posed ofthe unit line-of-sight vectorspointingfromthereceiver to eachsatellite. Aisthediagonal matrixwiththe wavelengthof the carrier-phase measurements [32,46]. The covariance matrices CφandCρaredefinedasCφ=2
σ
2φTW−1TandCρ=2
σ
ρ2TW−1T,where
σ
φ andσ
ρ are thezenith-referenced undifferenced phase and code standard deviations [46], T=[IM − 1M] is thedouble-differencing matrix, W=diag
(
w1,...,wM+1)
is a diagonal values andwi isthesatellite elevation-dependentweight.Asitisformu-lated, the RTK problem can be seen to fit the linear regression problem discussed earlier in Section 4. In practice, the solution forthe mixedrealandinteger model(20)isgenerallysolved via themixedreal-integerregression((21a)–(21c)).Thus, theRTK po-sitioning modelconstitutesa practicalexampleofGaussian linear regressionwithmixedreal-andinteger-valuedparameters.
4.2. CRBforGNSSRTKprecisepositioning
Asaforementioned,intheRTKproblem,theGaussianlinear ob-servationmodelreads
y=
φ
ρ
=Dθ
+n, n∼ NM(
0,Cn)
,B=[b1...bKω ], A=[a1...aKz]. (24)From theresultspresentedinSection 3.1,werecall thattheCRB, referredtoasCRBReal/Integerinthefollowing,isgivenby
CRBθ|θ
θ
=θ
θ
Fθ
−1θ
θ
, Fθ
=Fω|θ
θ
Hθ
Hθ
MSz|θθ
,where
θ(
θ
)isgivenby(16),andwehavetocomputeFθ
using the generalGaussian modelequations givenin Section 3.2. Since Cndoesnotdependonθ
,(18a)–(18c))becomeCi j=C n, mi j=C−1n
mθ
i+mθ
j− mθ
,δ
i j=mθ
iC−1 n mθ
i+mθ
jC−1 n mθ
j − mθ
C−1n mθ
, [MS]i, j=e(
m(
θ)
−m(
θi)
−m(
θj)
)
C−1 nm(
θ)
+m(
θi)
C−1nm(
θj)
.TheseequationsallowcomputationoftheelementsintheCRB for-mula.Particularly,computingFω|θ(
θ
)for1≤ k,k≤ Kω becomes Fω|θθ
k,k=∂
mθ
∂ω
k C−1n∂
mθ
∂ω
k =bk bk C−1n
bk bk , suchthat Fω|θ
θ
=BC−1 n B. (25a) Letθ
j=θ
+(
−1)
j−1i Kω+-j+1 2 . andθ
i=θ
+(
−1)
i−1i Kω+/i +1 2 0. Then, MSz|θθ
i, j=e(
m(
θ)
−m(
θi)
−m(
θj)
)
C−1n m(
θ)
+m(
θi)
C−1 nm(
θj)
− 1 =e(
θ−θi−θj)
DC−1nDθ+(
θi)
DC−1 nD(
θj)
− 1, (25b)andfor1≤ k≤ Kω,wehavethat
hθ
,θ
j k=∂
mθ
∂ω
k C−1n mθ
j− mθ
=bk bk C−1n D
θ
j−θ
, whichleadsto Hθ
=BC−1 n D iKω +1 −iKω +1 ... iK −iK . (25c)It is worth remembering that relaxing the condition on the integer-valued part ofthe parameters’ vector, and assuming that bothparameters arereal-valued,
ω
∈RKω, z∈RKz ,then the stan-dardCRBisgivenbytheinverseofthefollowingFIM,Fθ|θ
θ
=DC−1n D, (26)
andisreferredtoasCRBRealinthefollowing. 4.3. Illustrativeexample
To illustrate the validity of the proposed LB, a realistic GNSS RTK experiment was simulated.Particularly, the receiver-satellite geometryconsideredisillustratedinFig.1(M+1=13satellites), underawiderangeofprecision levelsfortheundifferencedcode observations –preserving the noise ofcarrier-phase (
φ
) measure-mentstwoordersofmagnitudelowerthancode(ρ
)observable–. Toevaluate theLSperformance,the root(total)MSE (RMSE), ob-tained from104 Monte Carlo runs, forboth 3D positionand the 12(M)integerambiguitieswasconsideredasameasureof perfor-mance.Fig.2(left)showsthe3D position(
ω
) RMSEoverthestandard deviation ofcode observationsσ
ρ, aswell asthe square-root ofFig. 1. Skyplot of the simulated GNSS receiver-satellite geometry.
the corresponding (total)CRBs,witha zoomofthelow noise re-giongiveninFig.2(right).First,noticethatthestandardLS (equiv-alenttotheMLE)RMSE,asexpected,coincideswiththeCRBRealfor therangeoftested
σ
ρ values,whichgivestheultimateachievable performance withboth code andphase observablesif nointeger constraint isimposed forambiguitiesz. Secondly,the ILS perfor-mance, considering thatthe floatpositionestimate isalways cor-rectedbytheoutputambiguitiesoftheIAR,clearlydependsonthe noise level. Three regions of performance can identified: i) large noiseregime:theILScoincideswithbothstandardLSandCRBReal,which is clear from the ILS success rate shown in Fig. 3, where wecanseethatfor
σ
ρ >5[m]acorrectintegersolutionisnever found,then,onaverage,isasifnointegerconstraintwasimposed;ii) low noise regime:the IAR obtains the correct ambiguity solu-tionwithhighprobability,thentheILScoincideswiththeso-called
CorrectILS(whichonlyconsidersthesuccessfuloutputsoftheIAR)
and theCRBReal|Integer, whichshowsthat theILS is asymptotically
efficient; and iii) threshold region: below the so-called threshold point (inthiscase,
σ
ρ >0.1[m]),theILSRMSEdeparts fromtheCRBReal|Integerandrises towards theCRBReal,witheven asmall
re-gionwheretheRMSEoverpassestheperformanceofastandardLS
Fig. 3. Success rate for the integer parameter estimation of the RTK positioning problem.
(inthiscase,for1<
σ
ρ < 10[m]).Thisregiondescribesthe be-haviour ofthe ILS, which abandons its asymptotic efficiencyand ambiguouserrorsoccurduetothe(partially)wrongestimationof theintegerambiguities.Thethresholdpointvarieswiththe satel-litegeometry,numberofobservations(i.e.,numberoffrequencies tracked) andobservationnoises. Therefore, theprecise prediction forthe transitionpoint remains an open challenge. Finally,if we considered onlythesuccessfullyfixed ambiguities,theCorrect ILSwouldcoincide withtheCRBReal|Integer.However, the correct
solu-tiontotheILSproblemcannotbe guaranteedoutsidethe asymp-toticregion.
Regarding theCRBReal|Integer andCRBReal comparison,it isclear
that considering the integer nature ofa part of the vector to be estimatedhasastrongimpactontheachievableperformance,and therefore, highlights the interest of the estimating the so-called integer ambiguities. As a byproduct, this highlights the impor-tanceof theLB proposed inthiscontribution.Obviously, restrict-ing the set ofpossible values (integer instead ofreal) leads to a
LB such that CRBReal|Integer ≤ CRBReal. For this bound there exists
alsoa noisethresholdregion fromwherethereal/integer
parame-Fig. 2. Positioning RMSE and square-root of CRBs as a function of the standard deviation of observation noise σρ.