• No results found

Nonparametric recursive density estimation for spatial data

N/A
N/A
Protected

Academic year: 2021

Share "Nonparametric recursive density estimation for spatial data"

Copied!
6
0
0

Loading.... (view fulltext now)

Full text

(1)

Contents lists available atScienceDirect

C. R.

Acad.

Sci.

Paris,

Ser. I

www.sciencedirect.com

Statistics

Nonparametric

recursive

density

estimation

for

spatial

data

Estimation nonparamétrique récursive de la densité pour données spatiales

Aboubacar

Amiri

a

,

Sophie

Dabo-Niang

a

,

b

,

Mohamed

Yahaya

a

,

c

aUniversitéLille3,LaboratoireLEMCNRS9221,France bINRIA-MODALTeam,France

cFST,UniversitédesComores,Comoros

a

r

t

i

c

l

e

i

n

f

o

a

b

s

t

r

a

c

t

Articlehistory:

Received26February2014

Acceptedafterrevision18October2015 Availableonline23December2015 PresentedbytheEditorialBoard

Thispaper dealswith non-parametricdensityestimationfor spatialdata. Westudythe asymptoticpropertiesofanewrecursiveversionoftheParzen–Rozenblattestimator.The meansquareerrorandanalmostsureconvergenceresultwithrateofsuchestimatorare derived.

©2015PublishedbyElsevierMassonSASonbehalfofAcadémiedessciences.Thisisan openaccessarticleundertheCCBY-NC-NDlicense (http://creativecommons.org/licenses/by-nc-nd/4.0/).

r

é

s

u

m

é

Cepapiertraitedel’estimationdeladensitéspatialedanslecasrécursif.Nousétudionsles propiétésasymptotiquesd’unenouvelleversiondel’estimateurdeParzen–Rozenblatt.Nous établissonslesconvergencesenmoyennequadratiqueetpresquesûredecetestimateur ; desvitessesdeconvergencesontdonnées.

©2015PublishedbyElsevierMassonSASonbehalfofAcadémiedessciences.Thisisan openaccessarticleundertheCCBY-NC-NDlicense (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction

Weconsiderrecursivekerneldensityestimationforspatiallydependentdata.Spatialdensityestimationisaninteresting andcrucialprobleminstatisticalinferenceforanumberofapplications,wheretheconsideredvariableshavespatial depen-dence.Spatialdataaremodeledasfiniterealizationsofrandomfieldscollectedfromdifferentspatiallocationsontheearth andtheyare widelyusedina varietyoffields,includingsoilscience,geology,oceanography, econometrics,epidemiology, environmentalscience, forestry, andmanyothers,seeCressie [9],Chiles andDelfiner [7].Althoughpotential applications ofnonparametric spatialmodels areabundant, littletheoretical workhas beendevotedtononparametric spacemodeling comparedtotheparametriccase.

Thiswork concernsa nonparametricestimationof theprobability densityfunctionfromdependentspatial data,using a recursive kernel approach. For an overview on results and applications considering independent ordependent spatial data fordensityestimation by classical kernelmethod, we highlight theworks by Biau andCadre[3],Carbon et al.[5], Dabo-NiangandYao[10],Tran[19].Themethodandtheresultsobtainedheregeneralizethepreviousworksintherecursive framework.

E-mailaddresses:[email protected](A. Amiri),[email protected](S. Dabo-Niang),[email protected](M. Yahaya). http://dx.doi.org/10.1016/j.crma.2015.10.010

1631-073X/©2015PublishedbyElsevierMassonSASonbehalfofAcadémiedessciences.ThisisanopenaccessarticleundertheCCBY-NC-NDlicense (http://creativecommons.org/licenses/by-nc-nd/4.0/).

(2)

Withtheadvancesinhardwaretechnologyandthesophisticationofmoderncomputinganddataacquisitiontechniques, voluminousspatialdataarecollectedinvariousappliedareas.Newchallengesarisesinceinmanyapplicationsthevolumeof thedataissolargethatitmaybeimpossibletostorethemondeskandtheirmodelingrequiresmanytime.Nonparametric estimatorssuchasthekerneldensityestimatorofParzen–Rozenblattcanbeusedtoestimatethedensityfunction.However, theysufferfromaseriouscomputationaldrawbackinbigdatacontext.Infact,ifthedensityisestimatedbyanon-recursive estimator,thislastmustbecompletelyrecalculatedwhennewdataitemsarereceived.

In situations wheredata itemsarrivesequentially anda largeamount ofdata canbe generated ata rapidrate, a re-cursive updating algorithms providegreat help for the difficulties arising from the large volume of data. Then they are muchpreferredovernon-recursiveones,sincerecursiveestimatorscanbeupdatedfromtheirimmediatepastandthenew observation. Therefore,theupdatecanbecomputedinstantlyanddoesnotrequireextensivestorageofdata.This arrange-ment isparticularto recursivemethodsandiscalledtheonlineorreal-timeupdatingproperty. Thisrecursive propertyis particularlyinterestingwhenthesampledataarecontinuallycapturedovertime.

Recursivekernelapproacheswere introducedandabundantlystudiedinthenon-spatialcase.Werefer,forinstance,to the pioneerworksby WolvertonandWagner [20],Deheuvels [12],Wegman andDavies [11], whohavepaid attentionto nonparametricrecursivedensityestimationfortimeprocesses.Therearesomerecentcontributionsonasymptoticproperties of recursive kernel densityfor dependent data. Amiri [1]generalized all the above-cited versions to a generalfamily of recursive estimators, usingdifferent valuesof a parameter

∈ [

0,1

]

associatedwitheach estimator. Forthe fixed values

=

1/2 and

=

1, the asymptotic normality under negative association was previously treated by Liang and Baek [2], while themean-squareconvergenceandtheasymptoticnormality werestudiedby Masry[14] under

α

-mixingcondition. Mezhoudetal.[15]haveextendedtheresultsobtainedbyAmiri[1]to

η

-dependenttimeprocess.

We extend the feature of time-varying sample recursive density estimate to the spatial case. Namely, we presenta recursive versionoftheclassical spatialkerneldensityestimatorandestablishsomeasymptoticresultsofsuchestimator. Asfarasweknow,theinvestigationofrecursivekernelmethodsforspatialdataremainanopenproblem.

The restof the paper isorganized asfollows. In section 2, we present the spatial recursive density estimator, while Section 3 gives general assumptions. In section 4, we establishasymptotic convergence, namely mean square error and almost-sure convergence with the rate of our estimator. Finally Section 5 is devoted to some indications to prove the theoreticalresultspresentedinthisnote.

2. Therecursivespatialkerneldensityestimate

Let

(

Xi

)

i∈NN

,

N

1 be a measurable strictly stationary spatial process defined on a probability space

(,

A

,

P

)

and

valued in

R

d,whered

1.We assume that the X

i’shave thesame distributionasa random variable X,withunknown

probability densityfunction p.Suppose that weobservethe process onaspatial set ofsites

I

n ofcardinal n,whichis a finite subsetofapotentially observableregion

D

R

N,where

R

N isendowedwiththeuniformmetric.Forconvenience, we treat the observationssites asan array that is

I

n

=

sj

,

j

=

1, . . . ,n

N

N. The estimator ofthe probability density functionbasedon

(

Xi

,

i

I

n

)

tobestudiedinthepresentworkisoftheform:

pn

(

x0

)

:=

1 Sn,

iIn 1 hdi K

Xi

x0 hi

,

x0

R

d

(

∈ [

0

,

1

]

),

(1) whereSn,

:=

iIn

hd(i 1−),hiisabandwidthcorrespondingtotheobservationXi.Byenumeratingthesites,onemayrewrite

p n

(

x0

)

as pn

(

x0

)

:=

1 Sn, n

j=1 1 hdsj K

Xsj

x0 hsj

,

x0

R

d

(

∈ [

0

,

1

]

),

(2) whereSn,

:=

n

j=1

hd(sj1−).Thisrepresentationoftheestimatorpresentsagreatinterestfromthecomputationalpointofview.

Infact,ifXsn+1 isanewobservationoftheprocessatasitesn+1 addedto

I

n,theestimatorcanbeupdatedrecursivelyby thefollowingformula:

pn+1

(

x0

)

=

Sn, Sn+1, pn

(

x0

)

+

Kn+1

Xsn+1

x0

with Ki

(.)

:=

1 hdsiSi, K

.

hsi

.

(3)

In(3),pn+1

(

x0

)

isthedensityestimatorbasedonthedomain

I

n

∪ {

sn+1

}

.Thisrecursiveformulaisusefulwhenthenumber

ofthespatialsitesincreasessequentiallyonspace.

Tosimplifythepresentation,theparameter

isthroughoutsupposedtobe equalto0 andthecorrespondingestimator willbelaternoted pn.

Basically,twoasymptoticmethodsoccurinthespatialliterature:increasingdomainandinfillasymptotics,see[9],p. 480. The growthof thesample inincreasing domainasymptotics isa consequenceofan unbounded expansionof thesample

(3)

region

I

n,whereasunderinfillasymptoticsthesampleregionisfixedandthegrowthofthesamplesizeisduetosampling thatisdenseintheregion

D

.Inwhatfollowsweconsidertheincreasingdomainasymptoticsandforsimplicitythebivariate regularlattice(N

=

2),describedinthefollowingassumption.

H2.1.D isaregularlatticeand

I

n

= {

i

=

(

i1

,

i2

),

1

ij

nj

,

j

=

1,2

}

isrectangular.Inthiscase,n

=

n1

×

n2goestoinfinitymeans thatminnj j=1,2goestoinfinity.Therefore,weusethelexicographicorderandrenumbertheobservationsasatriangulararrayinthe followingway.Theobservationsite

(

i

,

j

)

canbeindexedbyt

=

n2

(

i

1)

+

j inthetriangulararraysetting,see[17].Inthiscase,the newobservationsiteis

(

n1

+

1,1),whichisindexedbyn

+

1inthearraysetting.

Noticethattoobtainourasymptoticresultsintheirregularlatticecontext,onecanconsideradditionalassumptionson thenumberofspatialunitonaclosedballof

D

withthefollowing:

H2.2.Thepotentialobservableregion

D

isanirregularlatticeandinfinitecountable.Allelementsof

D

arelocatedatdistancesofat least

δ >

1fromeachother.

Inthiscase(asin[8]),weconsiderasequenceoffiniteclosed,convexregions

{

Ru

}

ofincreasingareaasu

→ +∞

andletthe samplesetconsistsoftheintersectionofoneoftheconvexregionsandD:

I

n

=

D

Ru.Wedonotspecifyhereanyorderandany otherassumptionontheconfigurationandgrowthofsamplesizeexcepttheassumptionofauniformlyincreasingareaof

{

Ru

}

inat leasttwonon-opposingdirectionstoincreasethesamplesizen (asu

→ +∞

),see[8].Thus,anewobservationcorrespondstoanynew sitesn+1of

D

notalreadyincludedin

I

n.

3. Generalassumptions

Inordertoestablishtheasymptoticresults,thefollowingassumptionswillbeconsidered.Letusconsiderthe represen-tationoftherecursivedensityestimatorgivenin(2),with

=

0.

H3.1.(a)hsn

0andnh d+2 sn

→ ∞

asn

→ ∞

;(b)Forr

∈ ]−∞

,

2

+

d

]

,Bn,r

:=

1 n n

i=1

hsi hsn

r

Br

>

0.

H3.2.K isaboundedsymmetricdensitysuchthat

Rd u K

(

u

)d

u

=

0and

Rd

u

2K

(

u

)d

u

<

. H3.3.Thedensityp isboundedandtwicecontinuouslydifferentiableon

R

d.

H3.4.Foranyi

,

j

∈ {

1,

. . . ,

n

}

suchthatsi

=

sj,therandomvector

(

Xsi

,

Xsj

)

hasadensity fsi,sjsuchthat:

sup

si=sj

gs

i,sj

<

,

where gsi,sj

=

fsi,sj

p

p

,

i

,

j

=

1

, . . . ,

n

.

H3.5.

(i) Thefield

(

Xsi

)

1≤inis

α

-mixing:thereexistsafunction

φ

:

R

+

R

+with

φ (

t

)

0ast

→ ∞

,suchthatwheneverE

,

E

R

2 withfinitecardinalsCard

(

E

)

,Card

(

E

)

α

σ

(

E

) ,

σ

E

:=

sup

Aσ(E),Bσ(E)

|

P

(

A

B

)

P

(

A

)

P

(

B

)

| ≤

ψ

Card

(

E

),

Card

(

E

)

φ

dist

(

E

,

E

)

,

where

σ

(

E

)

= {

Xi

,

i

E

}

and

σ

E

=

Xi

,

i

E

,dist

(

E

,

E

)

istheEuclideandistancebetweenE andEand

ψ(

·

)

isapositive symmetricfunctionnondecreasingineachvariable.

Thefunctions

φ

and

ψ

aresuchthat

φ (

i

)

C iθand

ψ(

n

,

m

)

Cmin(n

,

m

)

. (ii) For

>

0,letun

=

Ni=1

(log

ni

)(log log

ni

)

1+andassumethat

nlog4−N−2h θ−4N n u 2N 4Nθ n

→ ∞

,

where

θ >

max

4N

,

Nd

+

2 2

.

ThefirstconditionofH3.1isusualinnonparametricrecursiveestimation,whilethesecondoneiscrucialinourcalculus. This last is particularly used in the recursive kernel estimation, for instance by Masry [14], Wegman and Davies [11], Yamato[21],SamantaandMugisha[16]andIsogai[13].Ifhsn

=

Ann

q whereA

n

A

>

0, 0

<

q

<

1,thentheassumption

H 3.1 is satisfied for all q such that rq

<

1 and Br

=

1

1

rq. Also the choice hsn

=

An

lnn n

q

,

0

<

q

<

1 2

+

d satisfies

assumption H 3.1. Assumptions H 3.2 and H 3.3 are classical in nonparametric estimation and easily verified by many kernelssuchasEpanechnikov,Gaussiankernels.

(4)

AbouttheassumptionH3.5(i),basically,inthespatialliteratureitissupposedthat

φ (

i

)

tendstozerowithpolynomial rate, or

φ (

i

)

Cexp(

si

),

for someC

,

s

>

0,i.e.

φ (

i

)

tendstozerowithexponentialrate.Ourresultscanbeprovedunder theabovecondition.Finallyletusmentionthatthedefinitionofun inH3.5leadsto

n∈NNn1u

n

<

(recallthatn

=

n). 4. Consistencyresults

Theorem 1belowestablishestheasymptoticmeansquareerroroftheestimatorpn viaitsvarianceandbias.

Theorem1.UndertheassumptionsH2.2 andH3.1–3.5,wehave: nhdsnVar

(

pn

(

x0

))

p

(

x0

)

Bd1

Rd K2

(

u

)

du (4) and hs2 n

(

E

(

pn

(

x0

))

p

(

x0

))

B −1 d Bd+2

Rd K

(

u

)

2 d

j1,j2=1 uj1uj2

2p

(

x0

)

xj1

xj2 du (5)

asn

→ ∞

,forallx0suchthatp

(

x0

)

>

0.Inparticular,ifhsn

=

n −1

4+d,wehavetheasymptoticbehavioraccordingtothemeansquare error: n4+4d

E

(

pn

(

x0

)

p

(

x0

))

2

Rd K

(

u

)

d

j1,j2=1 uj1uj2

2p

(

x0

)

xj1

xj2 du

2

+

4 4

+

dp

(

x0

)

Rd K2

(

u

)

du (6)

asn

→ ∞

forallx0suchthatp

(

x0

)

>

0.

Theorem 1isanextensionofthepreviousresultsobtainedinthetemporalcasebyAmiri[1]with

α

-mixingcondition. Also, asimilarresultwereprovedbyMezhoudetal.[15] fortemporal

η

-dependenceprocess.Now,westatethefollowing theoremthatestablishesthealmostconvergenceofpn.

Theorem2.UndertheassumptionsH2.1,H3.1–3.5,wehave:

|

pn

(

x0

)

p

(

x0

)

| =

O

logn nhdn

,

a

.

s

.

(7)

5. Briefoutlineofproofs 5.1. Theorem 1 Forx0

R

d,let I1

=

1 B2 n,dn2h2snd n

i=1 Var

Zsi

andI2

=

1 B2 n,dn2h2snd

si=sj Cov

Zsi

,

Zsj

,

with Zsi

=

K

Xsi

x0 hsi

. Ontheonehand,underH3.2withthehelpoftheToeplitzLemma(see[18]),wereadilyhave

nhds n

|

I1

| →

p

(

x0

)

B −1 d

Rd K2

(

u

)

du asn

→ ∞

.

Ontheotherhand,forx0

R

d,letussplit:I2

=

J1

+

J2,where J1

=

1 B2n,dn2h2d sn

sisjcn Cov

(

Zsi

,

Zsj

)

and J2

=

1 B2n,dn2h2d sn

cn<sisj Cov

(

Zsi

,

Zsj

),

cn beingasequencetendingtoinfinityasntendstoinfinity.Ifd

3,then2d

>

d

+

2.Inthiscase,weconsider

ζ

R

such that N

θ

1

< ζ

2

d,andusingthedecreaseof

(

hsn

)

n,onemaywrite:

Cov

(

Zsi

,

Zsj

)

h d sihdsj sup si=sj

g

si,sj

h2sd i

+

h 2d sj 2 ssupi=sj

g

si,sj

hd(ζsi +1)h d(1−ζ ) s1

+

h d(ζ+1) sj h d(1−ζ ) s1 2 ssupi=sj

gsi,sj

.

(5)

If1

d

3,take

ζ

=

1 andmakeaspreviously.As

I

n isaregulardesign(seeforinstance[19]) Card

(

u

,

v

)

I

n2

:

u

v

cn

=

O

ncnN

.

Itfollowsthatnhdsn

|

J1

| ≤

Bn,d(ζ+1) B2 n,d hd(s11−ζ )h snc N n sup si=sj

gsi,sj

=

O

hdζsnc N n

.

Turningto J2,bytheBillingsleyinequality(see[4]),wehave

Cov(Zsi

,

Zsj

)

4φ (

si

sj

)

K

2 ∞.Consequently,weget nhds n

|

J2

| ≤

4

K

2 B2n,dhdsn

1

)

c1nθ

=

O

c1−θ n hdsn

.

The choice ofcn

=

hd(ζ+1) N+θ−1 sn

leads to nhdsn

|

I2

| =

O

h d

(ζ (θ

1)

N

)

N

+

θ

1 sn

0 asn

→ ∞

, aswell as

ζ >

N

θ

1,andthe result(4)follows.Aboutthebiasterm,weapply theTaylorformulaandthedominatedconvergence,whichtogether with Toeplitz’lemma,leadstotheresult(5).

2

5.2. Theorem 2 LetussetGn

(

x0

)

=

pn

(

x0

)

E

pn

(

x0

)

=

n

i=1

iwith

i

=

1 Sn,0

Zsi

E

Zsi

.Next,leta

=

an bean integerandsplitthe randomvariables

i intoblocksasfollows:

U

(

1

,

n

,

j

,

x0

)

=

(2j

k+1)an ik=2jkan+1 k=1,...,N

i

;

U

(

2

,

n

,

j

,

x0

)

=

(2jk

+1)an ik=2jkan+1 k=1,...,N−1 2(jN

+1)an iN=(2jN+1)an+1

i

;

U

(

3

,

n

,

j

,

x0

)

=

(2j

k+1)an ik=2jkan+1 k=1,...,N−2 2(jN

−1+1)an iN−1=(2jN−1+1)an+1 (2jN

+1)an iN=2jNan+1

i

;

U

(

4

,

n

,

j

,

x0

)

=

(2j

k+1)an ik=2jkan+1 k=1,...,N−2 2(jN

−1+1)an iN−1=(2jN−1+1)an+1 2(jN

+1)an iN=(2jN+1)an+1

i

,

(8)

andsoon.FinallynotethatU

(2

N−1

,

n

,

j

,

x0

)

=

(2j

k+1)an ik=2jkan+1 k=1,...,N−1 2(jN

+1)an iN=2jNan+1

i

;

U

(2

N

,

n

,

j

,

x0

)

=

(2jk

+1)an ik=2jkan+1 k=1,...,N

i.

Foreachinteger1

i

2N,define E

(

n

,

i

,

x 0

)

=

r

k−1 jk=0 k=1,...,N

U

(

i

,

n

,

j

,

x0

),

thenonemaywriteGn

(

x0

)

=

2N+1

i=1 E

(

n

,

i

,

x0

).

Without

lossofgenerality,wewillshowtheresultfori

=

1. Because K is bounded, we get U

(1,

n

,

j

,

x0

)

a

N nK

Bn,d 1 nhdn

. Consider the sequences

λ

n

=

nhdnlogn

1/2

!

, and ri

=

ni

/(2

an

),

i

=

1,

. . . ,

N.Wededucethat,fornlargeenough,

λ

n

|

U

(

1

,

n

,

j

,

x0

)

| =

O

"

logn nhdn

.

(9)

Enumerate the randomvariables U

(1,

n

,

j

,

x

)

in E

(

n

,

1,x

)

inan arbitrary mannerand refer tothem as U1

ˆ

,

. . . ,

U

ˆ

M, M

=

r1

...

rN.ByMarkov’sinequalityandLemma4.5of[6],thereexistr.v.’sU1

˜

,

. . . ,

U

˜

M independentofU1

ˆ

,

. . . ,

U

ˆ

M andsuchthat

ˆ

Ui

= ˆ

Ui

(

x0

)

hasthesamedistributionasU

˜

i

= ˜

Ui

(

x0

)

and:

P

U

ˆ

i

− ˜

Ui

>

18

U

ˆ

i

ψ (

n

,

aNn

)φ (

an

).

(10)
(6)

Let

n

=

η

logn nhd

n

,where

η

isapositivenumber.Obviously

P

(

|

E

(

n

,

1

,

x0

)

|

>

n

)

P

M j=1

Ui

˜

>

n 2

+

P

M j=1

Ui

ˆ

− ˜

Ui

>

n 2

.

(11)

UsingtheindependenceoftheU

˜

i’sandapplyingBernsteininequality,weget

P

M j=1

Ui

˜

>

n 2

2 exp

λ

n

n 2

+

λ

n2 4 M

i=1

E

U

ˆ

2i

(

x0

)

2 exp

η

logn 2

+

λ

2n 4 M

i=1

E

U

ˆ

i2

(

x0

)

.

Clearly,

λ

2 n 4 M

i=1

E

U

ˆ

i2

(

x0

)

λ

2n 4Var(pn

(

x0

))

log(n

)

4 nh d nVar(pn

(

x0

)).

From(4),wededucethatfornlargeenough,nhdnVar(pn

(

x0

))

p

(

x0

)

Bd−1

Rd K2

(

u

)d

u.Itfollowsthat

λ

2n 4 M

i=1

E

U

ˆ

2i

(

x0

)

log

(

n

)

p

(

x0

)

Bd1 4

Rd K2

(

u

)

du

.

Hence

P

M j=1

U

˜

i

>

n 2

nδ,where

δ >

1,for

η

largeenough.Concerningthesecondtermintheright-hand-sideof(11), from(10)onehas

U

ˆ

i

(

x0

)

aN n

K

nhd n

.Weget,byMarkov’sinequality:

P

M

i=1

U

ˆ

i

− ˜

Ui

>

n 2

C M

anN

K

nhd n

n−1

ψ (

n

,

anN

)φ (

an

),

then

P

M

i=1

U

ˆ

i

− ˜

Ui

>

n 2

C

1 hdn

n−1

ψ(

n

,

anN

)

(φ (

an

))

C

1 hnd

n−1anNa. Letan

=

logn nhd n

−1/(2N) ,then

P

M

i=1

U

ˆ

i

− ˜

Ui

>

n 2

Cn 2Nθ 2N logn θ−2N 2N h 2N n

=

C

β

n. By hypothesis, we have nun

β

n

=

nlog4−N2−h θ−4N n u 2N 4Nθ n

4Nθ 2N

0, then Theorem 2 follows by the Borel–Cantelli lemma.

2

References

[1]A.Amiri,Surunefamilleparamétriqued’estimateursséquentielsdeladensitépourunprocessusfortementmélangeant,C.R.Acad.Sci.Paris,Sér.I 347(2009)309–314.

[2]J.Baek,H.-Y.Liang,Asymptoticnormalityofrecursivedensityestimatesundersomedependenceassumptions,Metrika60(2004)155–166. [3]G.Biau,B.Cadre,Nonparametricspatialprediction,Stat.InferenceStoch.Process.7(2004)327–349.

[4]D.Blanke,D.Bosq,InferenceandPredictioninLargeDimensions,WileySeriesinProbabilityandStatistics,Dunod,2007. [5]M.Carbon,L.-T.Tran,B.Wu,Kerneldensityestimationforrandomfields:theL1theory,J.Nonparametr.Stat.6(1997)157–170. [6]M.Carbon,L.-T.Tran,B.Wu,Kerneldensityestimationforrandomfields,Stat.Probab.Lett.36(1997)115–125.

[7]J.-P.Chiles,P.Delfiner,StatisticsforSpatialData,Wiley,NewYork,1999.

[8]T.Conley,GMMestimationwithcrosssectionaldependence,J.Econom.92(1999)1–45.

[9]N.-A.-C.Cressie,StatisticsforSpatialData,WileySeriesinProbabilityandMathematicalStatistics,Wiley-Interscience,NewYork,1991. [10]S.Dabo-Niang,A.-F.Yao,Kernelregressionestimationforcontinuousspatialprocesses,Math.MethodsStat.16(2007)298–317. [11]H.-I.Davies,E.-J.Wegman,Remarksonsomerecursiveestimatorsofaprobabilitydensity,Ann.Stat.7(1979)316–327. [12]P.Deheuvels,Surunefamilled’estimateursdeladensitéd’unevariablealéatoire,C.R.Acad.Sci.Paris276(1974)1013–1015. [13]E.Isogai,ABerry–Esseen-typeboundforrecursiveestimatorsofadensityanditsderivatives,J.Stat.Plan.Inference40(1994)1–4. [14]E.Masry,Recursiveprobabilitydensityestimationforweaklydependentstationaryprocesses,IEEETrans.Inf.Theory32(1986)254–267. [15]K.-A.Mezhoud,etal.,Recursivekernelestimationofthedensityunderη-dependence,J.KoreanStat.Soc.(2014)1–12.

[16]R.-X.Mugisha,M.Samanta,Ontheclassofestimatesoftheprobabilityfunctionandmodebasedonrandomnumberofobservations,CalcuttaStat. Assoc.Bull.117(1981)23–40.

[17]P.-M.Robinson,Asymptotictheoryfornonparametricregressionwithspatialdata,J.Econom.165(2011)5–19. [18]W.-F.Stout,AlmostSureConvergence,AcademicPress,NewYork,1974,pp. 120–121.

[19]L.-T.Tran,Kerneldensityestimationonrandomfields,J.Multivar.Anal.34(1990)37–53.

[20]T.Wagner,C.Wolverton,Recursiveestimatesofprobabilitydensity,IEEETrans.Syst.Sci.Cybern.5(1969)307–308. [21]H.Yamato,Sequentialestimationofacontinuousprobabilitydensityfunctionandmode,Bull.Math.Stat.14(1972)1–12.

C. R. Acad. Sci. Paris, Ser. I 354 (2016) 205–210 ScienceDirect www.sciencedirect.com (http://creativecommons.org/licenses/by-nc-nd/4.0/ 309–314. 155–166. 327–349. 2007. 157–170. 115–125. 1999. 92 1991. 298–317. 316–327. 1013–1015. 40 254–267. K.-A. 23–40. 5–19. 1974, 37–53. 307–308. 14

References

Related documents

A different approach is to adopt physiologically informed models of BOLD response, such as the Balloon model ( Buxton et al., 1998; Friston et al., 2000b ).. In this model, a set

“How Individuals Smooth Spending: Evidence from the 2013 Government Shutdown Using Account Data.” NBER WP (Revised Oct 2015)... Work

(2013) reported no difference in CH 4 emissions between high- and low- RFI beef cattle. The extent to which CH 4 can be reduced by selection for RFI depends on the heritability

to estimate the significant contribution of support price and fertilizer offtake towards rice production and wheat acreage, to work out short and long run elasticities of price

Life Long learning measures at the university include, a.o., streamlining curricula (shorter duration of study programmes), Fachhochschule programmes, the utilization of new

Insect Pests and Weeds (P. DeBach, editor). Chapman and Hall Ltd, London. A history of biological control. Smith, editors). Inc., Palo Alto, California. A history of

In that case the worsening of the state of the fossil-fuel based technology may decrease the supply of energy enough so that fossil fuel will be reallocated from both the present

However, the DWS technique has an important limitation: it is specifically thought to typical data warehouses organized in an ideal star schema consisting of a large fact