HAL Id: hal-01060985
https://hal.inria.fr/hal-01060985
Preprint submitted on 4 Sep 2014
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
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
Asymptotic behaviour of extreme geometric quantiles and their estimation under moment conditions
Stéphane Girard, Gilles Stupfler
To cite this version:
Stéphane Girard, Gilles Stupfler. Asymptotic behaviour of extreme geometric quantiles and their
estimation under moment conditions. 2014. �hal-01060985�
and their estimation under moment onditions
Stéphane Girard
(1)
& GillesStuper
(2)
(1)
TeamMistis, InriaGrenobleRhne-Alpes&LJK,Inovallée,655,av. del'Europe,
Montbonnot,38334Saint-Ismieredex,Frane
(2)
AixMarseilleUniversité,CNRS,EHESS,CentraleMarseille,GREQAMUMR7316,
13002 Marseille,Frane
Abstrat. A popular way to study the tail of a distribution is to onsider its extreme quantiles.
While this is a standard proedure for univariate distributions, it is harder for multivariate ones, pri-
marilybeausethere is nouniversally aepteddenition of what amultivariatequantileshould be. In
thispaper,wefousonextremegeometriquantiles. Theirasymptotisareestablished,bothindiretion
andmagnitude, under suitablemoment onditions,when thenorm of theassoiated index vetortends
to one. Inpartiular, it appears that if arandom vetorhas anite ovarianematrix, then the mag-
nitude of its extreme geometri quantilesgrows at a xed rate. We take advantageof these results to
deneanestimatorofextremegeometriquantilesofsuharandomvetor. Theonsistenyandasymp-
totinormalityoftheestimatorareestablishedandourresultsareillustratedonsomenumerialexamples.
AMS Subjet Classiations: 62H05,62G20,62G32.
Keywords: Extremequantile,geometriquantile,onsisteny,asymptotinormality.
1 Introdution
Let
X
bearandomvetorinR d
. Uptonow,severaldenitionsofmultivariatequantilesofX
havebeenproposedinthestatistialliterature. WerefertoSering(2002)forareviewofvariouspossibilitiesforthis
notion. Here,wefousonthenotionofspatial orgeometriquantiles,introduedbyChaudhuri(1996),
whihgeneralisestheharaterisationof aunivariatequantileshowninKoenkerandBassett(1978). For
agivenvetor
u
belongingto theunitopenballB d
ofR d
, whered ≥ 2
,ageometriquantilewithindexvetor
u
isanysolutionoftheoptimisationproblemdenedbyarg min
q∈R
dE(φ(u, X − q) − φ(u, X)),
(1)withthelossfuntion
φ : R d × R d → R, (u, t) 7→ k t k + h u, t i ,
whereh· , ·i
istheusualsalarprodutonR d
and
k · k
istheassoiatedEulideannorm. Notethatq(u) ∈ R d
possessesbothadiretionandmagnitude.Itanbeseenthatgeometriquantilesareinfatspeialasesof
M
quantilesintroduedbyBreklingandChambers(1988)whihwerefurtheranalysedbyKolthinskii(1997). Besides,suhquantileshavevarious
strongproperties. First,thequantilewithindex vetor
u ∈ B d
isuniquewheneverthedistribution ofX
isnotonentratedonasinglestraightlinein
R d
(seeChaudhuri,1996,orTheorem2.17inKemperman,1987). Seond,although theyarenotfullyaneequivariant,theyareequivariantunder any orthogonal
transformation (Chaudhuri, 1996). Third, geometri quantiles haraterise the assoiated distribution.
Namely, if two random variables
X
andY
yield the same quantile funtionq
, thenX
andY
havethesame distribution(Kolthinskii, 1997). Finally, for
u = 0
, the well-knownL 2 −
geometri median isobtained,whihis thesimplestexampleofaentral quantile(seeSmall,1990). Wepointoutthat one
mayomputeanestimationofthegeometrimedian inaneientway,seeCardotetal. (2013).
Thesepropertiesmakegeometriquantilesreasonableandidateswhentryingtodenemultivariatequan-
tiles, whih is why their estimation was studied in several papers. We refer for instane to Chaud-
huri (1996),who establishedaBahadurexpansion forthe estimatorof geometri quantilesobtained by
solvingthe sample ounterpart of problem (1). Chakraborty(2001) then introdued atransformation-
retransformationproedure toobtain aneequivariantestimates of multivariatequantiles. This notion
wasextendedtoamultiresponselinearmodelbyChakraborty(2003). Reently,Dharetal. (2014)dened
amultivariatequantile-quantileplotusinggeometri quantiles. Conditionalgeometriquantilesanalso
be dened by substituting a onditional expetation to the expetation in (1). We referto Cadre and
Gannoun (2000) for the estimation of the onditional geometri median and to Cheng and de Gooijer
(2007)fortheestimationof anarbitraryonditional geometriquantile. Theestimationofaonditional
medianwhenthereisaninnite-dimensionalovariateisonsideredinChaouhandLaïb(2013).
Ourfousinthispaperisratheronextremegeometriquantiles,obtainedwhen
k u k → 1
. Thetheoryofunivariateextremequantilesiswellestablished,seeforinstanethemonographbydeHaanandFerreira
(2006). Ontheontrary,thefewworksonextrememultivariatequantilesrelyonthestudyofextremelevel
setsoftheprobabilitydensityfuntionof
X
whenitisabsolutelyontinuouswithrespettotheLebesguemeasure. We referforinstane to Cai et al. (2011)foran appliation tothe estimationof extremerisk
regions for nanial data orto Einmahl et al. (2013)who fous on the ase of bivariate distributions
withan appliation to insuranedata. Oneanalso analyse extremequantilesof multivariatedatasets
by seleting a univariate variable and onsidering the other variables as ovariates. This amounts to
estimating onditional univariate extreme quantiles: for a nite-dimensional ovariate, this problem is
onsidered in Daouia et al. (2013), the ase of a funtional ovariate being addressed in Gardes and
Girard(2012).
Inthisstudy,weprovideanequivalentofthediretionandmagnitudeoftheextremegeometriquantile
q(u)
,k u k → 1
undersuitablemomentonditions. Apartiularorollaryofourresultsisthatthemagnitudeof the extreme geometri quantilesof a random vetor
X
having a nite ovarianematrix growsat axed rate. Moreover,in this ase, themagnitude of the extreme geometri quantiles is asymptotially
haraterisedbytheovarianematrixof
X
. Thispropertyopensthedoortothedenitionofanextremequantileestimator,whoseasymptotipropertiesarestudiedin thiswork.
Theoutlineofthepaperisasfollows. AsymptotipropertiesofgeometriquantilesarestatedinSetion2.
An appliation to the estimation of extreme geometri quantilesis given in Setion 3. Some examples
andillustrationsofourresultsarepresentedinSetion4. Setion5oersaoupleofonludingremarks.
ProofsaredeferredtoSetion 6.
2 Asymptoti behaviour of extreme geometri quantiles
Fromnowon,weassumethatthedistributionof
X
isnotonentratedonasinglestraightlineinR d
andnon-atomi. Weshallreformulatetheoptimisationproblem(1)as
arg min
q∈ R
dψ(u, q)
where
ψ : R d × R d → R, (u, q) 7→ E(φ(u, X − q) − φ(u, X))
anberewrittenasψ(u, q) = E( k X − q k − k X k ) − h u, q i .
(2)Chaudhuri(1996)provedthatinthisontext,thesolution
q(u)
of(1),namelythegeometriquantilewithindexvetor
u
,existsand isuniqueforeveryu ∈ B d
. Dene furtherthatt/ k t k = 0
ift = 0
;ifu ∈ R d
issuh thatthere isasolution
q(u) ∈ R d
to problem(1), thenthegradientofq 7→ ψ(u, q)
mustbezeroatq(u)
,thatisu + E
X − q(u) k X − q(u) k
= 0.
(3)This ondition immediately entails that if
u ∈ R d
is suh that problem (1) has a solutionq(u)
, thenk u k ≤ 1
. Infat,weanproveastrongerresult:Proposition1. The optimisationproblem (1)has asolutionif andonly if
u ∈ B d
.Moreover,remarking that the funtion
ψ(u, · )
isstritly onvex,Chaudhuri (1996)proved thefollowingharaterisationofageometri quantile: forevery
u ∈ B d
,q(u)
isthesolutionofproblem(1)ifandonlyifitsatisesequation(3). Inpartiular,thisentailsthatthefuntion
G : R d → B d
denedby∀ q ∈ R d , G(q) = − E
X − q k X − q k
isaontinuousbijetion. Proposition2.6(iii)inKolthinskii(1997)showsthattheinverseofthefuntion
G
,i.ethegeometriquantilefuntionu 7→ q(u)
,is alsoontinuousonB d
.Inmostaseshowever,omputingexpliitlythefuntion
G
isahopelesstask,whihmakesitimpossibleto obtain a losed-form expression for the geometri quantile funtion. It is thus of interest to prove
generalresultsaboutthegeometri quantile
q(u)
,espeiallyregardingitsdiretionand magnitude. Ourrstmainresultfousesonthespeialaseofspheriallysymmetridistributions.
Proposition2. If
X
has aspheriallysymmetri distributionthen:(i) Themap
u 7→ q(u)
ommutes witheverylinearisometryofR d
. Espeially, the normofageometriquantile
q(u)
only dependson the normofu
.(ii) Forall
u ∈ B d
,the geometri quantileq(u)
has diretionu
ifu 6 = 0
andq(0) = 0
otherwise.(iii) The funtion
k u k 7→ k q(u) k
isaontinuousinreasingfuntion on[0, 1)
.(iv) Itholdsthat
k q(u) k → ∞
ask u k → 1
.Although therst and third statement of Proposition 2annot be expeted to hold true for a random
variablewhihis notspheriallysymmetri,onemaywonderiftheseond andfourthstatement,namely
that a geometri quantile shares the diretion of its index vetorand that the norm of the geometri
quantile funtion tends to innity on the unit sphere, an be extended to the general ase. The next
result,whih examinesthebehaviourof thegeometri quantilefuntion nearthe boundaryof theopen
ball
B d
,providesananswertothis question.Theorem1. Let
S d−1
bethe unitsphereofR d
.(i) Itholdsthat
k q(v) k → ∞
ask v k → 1
.(ii) Moreover,if
v → u
withu ∈ S d−1
andv ∈ B d
thenq(v)/ k q(v) k → u.
Theorem 1 shows twoproperties of geometri quantiles: rst, the norm of the geometri quantile
q(v)
with index vetor
v
diverges to innity ask v k ↑ 1
. In other words, Proposition 2(iv) still holds for any distribution. This is arather intriguing property of geometri quantiles, sine it holds even ifthedistributionof
X
hasaompatsupport. Arelatedpointis thefat that samplegeometriquantilesdonotneessarilyliewithin theonvexhullofthesample,seeBreklingetal. (2001)foraounter-example.
Seond,if
v → u ∈ S d−1
thenthe geometriquantileq(v)
has asymptotidiretionu
. Proposition 2(ii) thusremainstrueasymptotiallyforanydistribution.ItispossibletospeifytheonvergenesobtainedinTheorem1undermomentassumptions. Theorem2
providesarst-order expansionofthe diretion andof themagnitude ofan extremegeometri quantile
q(αu)
inthediretionu
,whereu
isaunit vetorandα
tendsto1.Theorem2. Let
u ∈ S d−1
.(i) If
E k X k < ∞
thenq(αu) − {k q(αu) k u + E(X − h X, u i u) } → 0
asα ↑ 1.
(ii) If
E k X k 2 < ∞
andΣ
denotes the ovarianematrixofX
thenk q(αu) k 2 (1 − α) → 1
2 (tr Σ − u ′ Σu) > 0
asα ↑ 1.
AsaonsequeneofTheorem2,itappearsthatif
X
hasaniteovarianematrixΣ
thenthemagnitudeofanextremegeometriquantileinthediretion
u
isdetermined(intheasymptotisense)byΣ
. Inotherwords,sinetheasymptotidiretion ofanextremegeometriquantileinthediretion
u
isexatlyu
byTheorem1,it followsthat theextremegeometri quantilesof twoprobabilitydistributions whihadmit
thesameniteovarianematrixareasymptotiallyequivalent. Furthermore,weobservethat
k q(βu) k k q(αu) k =
1 − α 1 − β
1/2
(1 + o(1))
when
α → 1
andβ → 1
. Inotherwords,givenanarbitraryextremegeometriquantile,oneandeduetheasymptotibehaviourofeveryotherextreme geometriquantilesharingits diretion,independently
ofthedistribution. Thisisfundamentallydierentfromtheunivariateasewhendeduingthevalueofan
extremequantilefromanotheronerequirestheknowledgeoftheextreme-valueindexofthedistribution,
seedeHaanandFerreira(2006),Chapter 4. Ourresultsanatuallybeused todene aonsistentand
asymptotiallyGaussianestimatorofextremegeometriquantiles,asshowninSetion 3below.
3 An estimator of extreme geometri quantiles
Let
X 1 , . . . , X n
beindependentrandomopiesofarandomvetorX
havinganiteovarianematrixΣ
.ItfollowsfromTheorem2thatanyextremegeometriquantile
q(αu)
ofX
,withα ↑ 1
andu ∈ S d−1
anbeapproximatedby:
q eq (αu) := (1 − α) −1/2 1
2 (tr Σ − u ′ Σu) 1/2
u.
(4)Thisanbeusedtodeneanestimatoroftheextremegeometriquantilesof
X
: letX n = n −1 P n k=1 X k
bethesamplemeanand
Σ b n = 1 n
X n k=1
(X k − X n )(X k − X n ) ′
bethe empirialestimatoroftheovarianematrix
Σ
ofX
. Let further(α n )
beaninreasingsequeneofpositiverealnumberstendingto 1. Ourestimator
q b n (α n u)
ofq(α n u)
isthenb
q n (α n u) = (1 − α n ) −1/2 1
2
tr Σ b n − u ′ Σ b n u 1/2 u.
Theonsistenyof
q b n (α n u)
isexaminedin thenextresult.Theorem3. Let
u ∈ S d−1
andassumethatα n ↑ 1
. IfE k X k 2 < ∞
then√ 1 − α n ( b q n (α n u) − q(α n u)) → 0
almost surely asn → ∞ .
This resultatually means that theextreme geometri quantileestimator is relatively onsistentin the
sensethat
b
q n (α n u) − q(α n u)
k q(α n u) k → 0
almostsurelyasn → ∞ ,
sine
k q(α n u) k −1 = O( √
1 − α n )
, see Theorem 2(ii). This normalisation ould be expeted sine the quantity to be estimated diverges in magnitude. Under the additional assumption thatX
has a nitefourthmoment,anasymptotinormalityresultanbeestablishedforthisestimator:
Theorem4. Let
u ∈ S d−1
andassumethatα n ↑ 1
issuhthatn(1 − α n ) → 0
. IfE k X k 4 < ∞
thenp n(1 − α n ) ( q b n (α n u) − q(α n u)) −→ d Z
asn → ∞
where
Z
isaGaussian entredrandom vetor.Letus highlight that the ovarianematrix of the Gaussian limit in Theorem 4 essentially depends on
theovarianematrix
M
oftheGaussian limitof√ n( Σ b n − Σ)
, seetheproofin Setion6. Althoughthematrix
M
has aheavy and ompliated expression (see e.g. Neudeker and Wesselman, 1990), it anbeestimated when
E k X k 4 < ∞
, whihmakesitpossibleto onstrutasymptotiondene regionsforextremegeometri quantiles.
Extremegeometriquantilesanthusbeonsistentlyestimatedby
q b n (α n u)
,whatevertheorderα n
,andanasymptotinormalityresultisobtainedwhen
α n ↑ 1
quiklyenough. Theproposedestimatoristhusabletoextrapolatearbitrarilyfarfromtheoriginalsample. Thisisverydierentfromtheunivariatease,
where the empirial quantile
b q n (α n ) = inf { t ∈ R | F b (t) ≥ α n }
, dedued from the empirial umulativedistributionfuntion
F b
, estimatesthe true quantileq(α n )
onsistently onlyifα n
onverges to 1slowlyenough. Theextrapolationwithfasterrates
α n
isthenhandledassumingthattheunderlyingdistribution funtion isheavy-tailed andbyusing adapted estimators,seee.g. Weissman(1978)and themonographbydeHaanandFerreira(2006).
4 Numerial illustrations
Inthis setion, our main results are illustrated, partiularly Theorems 2, 3and 4in the bivariate ase
d = 2
to makethedisplayeasier. Inthis framework,u ∈ S 1
anberepresentedbyanangle andwemay writeu = u θ = (cos θ, sin θ)
,θ ∈ [0, 2π)
. Theiso-quantileurvesC q(α) = { q(αu θ ), θ ∈ [0, 2π) }
and theirestimates
Cb q n (α) = {b q n (αu θ ), θ ∈ [0, 2π) }
anthenbeonsideredinordertogetagraspofthebehaviourof extreme quantiles in everydiretion. The following two distributions are onsidered forthe random
vetor
X
:•
theentredGaussianmultivariatedistributionN (0, v X , v Y , v XY )
,withprobabilitydensityfuntion:∀ x, y ∈ R, f(x, y) = 1 2π √
det Σ exp
− 1 2
x y
′
Σ −1
x y
withΣ =
v X v XY
v XY v Y
.
•
adouble exponentialdistributionE (λ − , µ − , λ + , µ + )
, withλ −
,µ −
,λ +
,µ + > 0
, whose probability∀ x, y ∈ R, f(x, y) =
λ + µ +
4 e −λ
+|x|−µ
+|y|
ifxy > 0, λ − µ −
4 e −λ
−|x|−µ
−|y|
ifxy ≤ 0.
Inthisase,
X
isentredandhasovarianematrixΣ =
1 λ 2 − + 1
λ 2 +
1 2
1
λ + µ + − 1 λ − µ −
1 2
1
λ + µ + − 1 λ − µ −
1 µ 2 − + 1
µ 2 +
.
Inourstudy, threedierentsetsof parameterswereused foreah distribution,inorder thattherelated
ovarianematriesoinide:
• N (0, 1/2, 1/2, 0)
andE (2, 2, 2, 2)
withspherialovarianematries;• N (0, 1/8, 3/4, 0)
andE (4, 2 p
2/3, 4, 2 p
2/3)
withdiagonalovarianematries;• N (0, 1/2, 1/2, 1/6)
andE (2 √ 3, 2 √
3, 2 p
3/5, 2 p
3/5)
withfullovarianematries.Ineahase,wearryoutthefollowingomputations:
•
foreahα ∈ { 0.99, 0.995, 0.999 }
,thetruequantileurvesC q(α)
obtainedbysolvingproblem(1)nu-merially,aswellastheiranalogues
C q eq (α)
usingapproximation(4)areomputed. Thenormalised squaredapproximationerrore(α) = (1 − α) Z 2π
0 k q eq (αu θ ) − q(αu θ ) k 2 dθ
isthenreorded.
•
for eah value ofα
, wedrawN = 1000
repliationsof ann −
sample(X 1 , . . . , X n )
of independent opiesofX
,withn ∈ { 100, 200, 500 }
. Theestimated quantileurvesC q b (j) n (α)
orrespondingtothej −
threpliationandtheassoiatednormalisedsquarederrorE n (j) (α) = (1 − α)
Z 2π 0
b q (j) n (αu θ ) − q(αu θ )
2
dθ
areomputedaswellas themeansquarederror
E n (α) = N −1 P N
j=1 E (j) n (α)
.Thetruequantileurves,aswellastheapproximatedandtheestimatedonesaredisplayedonFigures16
in the ase
n = 200
andα = 0.995
. The true quantile urves look very similar on Figures 1 and 4,on Figures 2 and 5 and Figures 3 and 6. This is in aordane with Theorem 2: eventually, extreme
geometriquantilesonlydepend ontheovarianematrixoftheunderlying distribution. Moreover,the
approximatedquantilesurvesare lose tothe trueones in all ases,and the estimated quantile urves
aresatisfyingin allsituationswithamoderatevariability. Similarresultswereobservedfor
n = 100, 500
and
α = 0.99, 0.999
. We do not report the graphshere for the sakeof brevity; wedo howeverdisplaytheapproximationandestimationerrorsinTable1. Unsurprisingly,theestimationerror
E n (α)
dereasesasthesample size
n
inreases. Both approximationandestimation errorse(α)
andE n (α)
haveastablebehaviourwithrespetto
α
.5 Conluding remarks
In this paper, we established the asymptotis, both in diretion and magnitude, of extreme geometri
quantiles. A partiularonsequeneof ourresultsis thatiftheunderlyingdistribution possessesanite
ovariane matrix
Σ
, then an extreme geometri quantile may be estimated aurately, no matter howextremeitis,withthehelpofthestandardempirialestimatorof
Σ
. Thisissupportedbyournumerialresults.
This work, however, was arried out under moment onditions suh asthe existene of nite rst and
seond-ordermomentsfor
k X k
. It would denitelybeinterestingto seeifouronlusions arryover,to someextent,totheasewhentheseassumptionsareviolated. Furthermore,althoughgeometriquantilesmake an appealing andidate for multivariate quantiles, they lak a ouple of nie properties suh as
aneequivariane, forinstane. Totaklethis issue, onemayapply atransformation-retransformation
proedure,seeSering(2010);suhproedures admitsampleanalogues,see forinstaneChakrabortyet
al. (1998)and Chakraborty(2001). Futureworkon extremegeometri quantilesthus inludes building
andstudyingananalogueofourestimatorfortransformed-retransformeddata.
6 Proofs
SomepreliminaryresultsareolletedinParagraph6.1,theirproofsarepostponedtoParagraph6.3. The
proofsofthemain resultsareprovidedinParagraph6.2.
6.1 Preliminary results
TherstlemmaprovidessometehnialtoolsneessarytoshowTheorem2(ii).
Lemma1. Let
ϕ : R d × R + × S d−1 → R
bethe funtion denedbyϕ(x, r, v) = r 2
1 + h x − rv, v i k x − rv k
.
Then forall
v ∈ S d−1
,ϕ( · , · , v)
isnonnegativeandwehave that∀ x ∈ R d , ∀ r ≤ k x k , ϕ(x, r, v) ≤ 2r 2
and∀ r > k x k , ϕ(x, r, v) ≤ k x k 2 .
Inpartiular,
ϕ(x, r, v) ≤ 2 k x k 2
for every(x, r, v) ∈ R d × R + × S d−1
.Thenextlemmaistherststepto proveTheorem2(i).