Laksaci
Laksaci
Laksaci
In this paper, we study a kernel estimator of the conditional hazard function when
the covariates take values in some abstract function space. The almost completely
convergence (with rate) of this estimate is obtained when the sample considered
is collected in spatial order with mixing structure. These results are extensions
to spatial dependent data of the results given by Ferraty et al. (Rev. Roumaine
Math. Pures Appl. 53 (2008), 1–18).
AMS 2010 Subject Classification: 62G05, 62G20, 62N02.
1. INTRODUCTION
hx , est
conditionnelle, noté b
fbx (y)
hx (y) =
b ∀y ∈ R.
1 − Fbx (y)
Le but de ce travail est d’étudier la convergence presque complète de
hx vers hx , lorsque le champ aléatoire fonctionnel (Zi )i∈NN vérifie
l’estimateur b
la condition de mélange suivante.
Il existe une fonction ϕ (t) ↓ 0 quand t → ∞, telle que
∀ E, E 0 sous ensemble de NN de cardinal fini
α (B(E), B(E 0 )) = sup |P (B ∩ C) − P (B) P (C)| ≤
B∈B(E), C∈B(E 0 )
≤ ψ (Card (E) , Card(E 0 )) ϕ (dist (E, E 0 )) ,
où B(E) (resp. B(E 0 )) la tribu borélienne engendrée par (Zi , i ∈ E) resp.
(Zi , i ∈ E 0 ) , Card(E) resp. Card(E 0 ) est le cardinal de E resp. E 0 ,
3. PROPRIÉTÉS ASYMPTOTIQUES
(H5) K est de support [0, 1], vérifie C1I[0,1] (·) ≤ K(·) ≤ C 0 1I[0,1] (·), où
1I[0,1] est la fonction indicatrice de [0, 1].
(H6) H est une fonction de classe C 2 et à support compact.
(H7) Il existe 0 < α < (δ − 5N )/3N et η0 > 0, tels que:
(5+3α)N −δ
b α hH = ∞
lim n et Cn
b δ
+η0
≤ hH φx (h),
n→∞
où n
b = n1 . . . n N .
Nos hypothèses sont relativement standard, puisque les conditions (H1),
(H3)–(H7) sont très similaires à celles utilisées par Ferraty et al. [9], tandis
que l’hypothèse (H2) est la même que dans Tran [21].
1 h i hx (y) h bx i
hx (y) − hx (y) =
b x x
f (y) − f (y) +
b x
F (y) − F (y) .
1 − Fbx (y) 1 − Fbx (y)
1 X
FbNx (y) = −1
K(h−1 −1
K d(x, Xi ))H(hH (y − Yi ),
b E K(hK d(x, X1 )) i∈I
n
n
et
1 X
x
fbN (y) = K(h−1 0 −1
K d(x, Xi ))H (hH (y − Yi ),
b hH E K(h−1
n K d(x, X 1 )) i∈I n
En général, cet estimateur n’est pas unique. Ainsi, dans toute la suite de cet
article θ(x)
b désignera toute variable aléatoire vérifiant (4).
Afin d’étudier la convergence presque complète de l’estimateur θ(x),
b on
garde les hypothèses de la section précédente et on suppose que la fonction hx
est de classe C 2 par rapport à y et telle que
0 00
(5) hx (θ(x)) = 0 et hx (θ(x)) > 0.
Finalement, le théorème précédent nous permet de conclure le corollaire
suivant:
Corollaire 9. Sous les conditions (1), (2), (H1)–(H7) et (5), on a,
1 !
b /2
b − θ(x) = O h 1 + h 2 b /2
log n
b 4
θ(x) K H + O , p.co.
n
b hH φx (hK )
42 Ali Laksaci et Boubakeur Mechab 8
5. APPENDICE
et
Hi0 (y) = H 0 (h−1
H (y − Yi )).
où ∆i (x)Ki (x) − E [Ki (x)]. On considère la décomposition spatiale de Tran [21]
sur les variables ∆i (x), définie, pour un entier pn fixé, de la manière suivante
2jkX
pn +pn
U (1, n, x, j) = ∆i (x),
ik =2jk pn +1
k=1,...,N
2jkX
pn +pn 2(jN +1)pn
X
U (2, n, x, j) = ∆i (x),
ik =2jk pn +1 iN =2jN pn +pn +1
k=1,...,N −1
2jkX
pn +pn 2(jN −1 +1)pn 2jNX
pn +pn
X
U (3, n, x, j) = ∆i (x),
ik =2jk pn +1 iN −1 =2jN −1 pn +pn +1 iN =2jN pn +1
k=1,...,N −2
2(jk +1)pn
X
U (2N , n, x, j) = ∆i (x).
ik =2jk pn +pn +1
k=1,...,N
9 Fonction de hasard spatiale et variable fonctionnelle 43
parce que même si ni n’est pas exactement égal à 2ri pn , on regroupe les
variables restantes dans un bloc T (n, x, 2N + 1) (ceci ne changera pas la
démonstration voir Biau et Cadre [1]).
Maintenant, en vertu de la dernière équation (6), pour chaque η > 0 on a
P |FbDx (x) − E[FbDx (x)]| ≥ η ≤ 2N max P (T (n, x, i) ≥ ηb
nE [K1 (x)]) .
i=1,...
éléments.
Le Lemme 4.5 de Carbon et al. [3] nous permet d’approximer Z1 ,Z2 , . . .,
ZM par des variables aléatoires indépendantes Z1∗ , . . . , ZM ∗ de même loi que
où
XM
M ηb
nE [K1 (x)]
Zj∗ ≥
B1 = P ≤
2M
j=1
nE [K1 (x)])2
(ηb
≤ 2 exp − ,
M Var [Z1∗ ] + CpN
n ηb
nE [K1 (x)]
M M
X ηb
n E [K 1 (x)] C X
B2 = P |Zj − Zj∗ | ≥ ≤ E|Zj − Zj∗ | ≤
2 ηb
nE [K1 (x)]
j=1 j=1
B2 ≤ n b pN b )−1/2 (b
n (log n nφx (hK ))−1/2 ϕ(pn ).
b x (hK ) 1/2N
On prend pn C nφlog n
b , alors
(7) b )−1 n
B2 ≤ (log n b ϕ(pn ).
P
D’après (H7), on montre que b ϕ(pn ) < ∞.
n
n∈In
On va traiter, maintenant, B1 . Pour cela, il suffit d’évaluer Var [Z1∗ ].
En effet,
" #
X X
∗
Var [Z1 ] = Var ∆i (x) = |Cov(∆i (x), ∆j (x))| .
i∈I(1,n,x,1) i,j∈I(1,n,x,1)
P P
Posons Qn = Var [∆i (x)] et Rn = |Cov(∆i (x), ∆j (x))|.
i∈I(1,n,x,1) i6=j∈I(1,n,x,1)
En vertu de (H1), on a Var[∆i (x)] ≤ C(φx (hK ) + (φx (hK ))2 ), donc Qn =
O pNn φx (hK ) . En ce qui concerne Rn , on utilise les techniques de Masry [18]
et on considère les ensembles
S1 = {i, j ∈ I(1, n, x, 1) : 0 < ki − jk ≤ cn },
S2 = {i, j ∈ I(1, n, x, 1) : ki − jk > cn },
où cn est une suite réelle tendant vers +∞. Ainsi,
X X
Rn = |Cov (∆i (x), ∆j (x))| + |Cov (∆i (x), ∆j (x))| = Rn1 + Rn2 .
(i,j)∈S1 (i,j)∈S2
11 Fonction de hasard spatiale et variable fonctionnelle 45
D’ une part, on a
X
Rn1 = |E [Ki (x)Kj (x)] − E [Ki (x)] E [Kj (x)]| ≤
(i,j)∈S1
1/a
≤ CpN cN
φ
n n x K(h ) (φ (h
x K )) + φ (h
x K ) ≤ CpN N
n cn φx (hK )
(a+1)/a
.
D’après (H2) on peut écrire Rn2 ≤ CpN n φx (hK ). De plus, avec ce choix de cn
on a
Rn1 ≤ CpNn φx (hK ).
D’où
Var [Z1∗ ] = O pN
n φx (hK ) .
En utilisant ce dernier résultat, avec les définitions de pn , M et η, on montre
B1 ≤ exp (−Cη0 log n
b) .
Par conséquent, un choix approprié de η0 permet de déduire le résultat du
Lemme.
Démonstration du Corollaire 3. On a, E FbDx (x) = 1, donc
P FbDx (x) ≤ 1/2 ≤ P FbDx (x) − E FbDx (x) > 1/2
donc
Z
|E(H1 (y)/X) − F (y)| ≤ x
H 0 (y)|F X (y − hH t) − F x (y)|dt.
R
Sous (H4), on a
Z
x
H 0 (y) hbK1 + |t|b2 hbH2 dt.
∀y ∈ S, 1IB(x,hK ) (X)|E(H1 (y)/X) − F (y)| ≤
R
−α− 12 α+ 12
avec ln = n
b et zn ≤ n
b . En posant, j(y) = arg minj∈{1,2,...,zn } |y − tj |,
on obtient la décomposition
bx
FN (y) − EFbNx (y) ≤
≤ FbNx (y) − FbNx (tj(y) ) + FbNx (tj(y) ) − EFbNx (tj(y) ) + EFbNx (tj(y) ) − EFbNx (y) .
| {z } | {z } | {z }
T1 T2 T3
• En ce qui concerne (T1 ) et (T3 ) on utilise le fait que H 0 est bornée pour
démonter
bx
1 X
FN (y) − FbNx (tj(y) ) ≤
Ki (x) Hi (y) − Hi (tj(y) ) ≤
n
b E [K1 (x)]
i∈In
1 X ln b x
≤C Ki (x) Hi (y) − Hi (tj(y) ) ≤ C F .
n
b E [K1 (x)] hH D
i∈In
q
log n
De la définition de ln et sous l’hypothèse (H7), on a hlnH = o b φ(hK ) .
n
b
et
s !
bx x
log n
EFN (tj(y) ) − EFN (y) = o
b
(9) .
b
n
b φ(hK )
• Pour (T3 ), ∀ η > 0, on a
s !
log n
P sup FbNx (tj(y) ) − EFbNx (tj(y) ) > η
b
=
y∈S n
b φ(hK )
s !
bxk x
log n
FN (tj ) − EFbN (tj ) > η ≤
k
b
=P max
j∈{1,2,...,zn } n
b φ(hK )
s !
log n
P FN (tj ) − EFbNxk (tj ) > η
b xk
≤ zn
b
max .
j∈{1,2,...,zn } n
b φ(hK )
Posons
∆0i = Ki (x)Hi (tj ) − E [Ki (x)Hi (tj )] .
Il est clair que
0
|∆i | ≤ C |∆i (x)| ,
Var [∆0i ] ≤ CVar [∆i (x)] et
Cov(∆0i , ∆0j ) ≤ CCov(∆i (x), ∆j (x)) ∀ i 6= j.
et on en déduit
Z
|h−1 0 x
H E(H1 (y)/X) − f (y)| ≤ H 0 (y)|f X (y − hH t) − f x (y)|dt.
R
h−1 X
Ki (x) Hi0 (y) − Hi0 (tj(y) ) ≤
bx x H
fN (y) − fbN (tj(y) ) ≤
n
b E [K1 (x)]
i∈In
h−1 X ln
H
Ki (x) Hi0 (y) − Hi0 (tj(y) ) ≤ C 2 FbDx .
≤C
n
b E [K1 (x)] hH
i∈In
q
ln log n
La définition de ln et l’hypothèse (H7) nous permet d’écrire h2H
o n
b
b hH φ(hK ) .
La convergence presque complète de Fbx implique D
s !
bx x
log n
fN (y) − fbN
b
(10) (tj(y) ) = o , p.co.
n
b hH φ(hK )
s !
bx x
log n
EfN (tj(y) ) − EfbN
b
(11) (y) = o .
n
b hH φ(hK )
15 Fonction de hasard spatiale et variable fonctionnelle 49
Posons
∆00i = Ki (x)Hi0 (tj ) − E Ki (x)Hi0 (tj ) .
Alors
|∆00i | ≤ C |∆i (x)| ,
Var [∆00 ] = O(hH φ(hK )) et
EH 0 Hi 0 |(X , X ) = O(h2 ) ∀ i 6= j.
i j i j H
En utilisant les mêmes étapes du Lemme 2, avec une légère modification dans
le choix de cn , on montre que
Cov(∆0i , ∆0j ) = O(hH φ(hK )) ∀ i 6= j.
Ainsi,
s !
x log n
≤ zn B10 + B20
x
(z) − E fbN
b
P max fbN (z) > η0
z∈Gn n
b hH φx (hK )
B20 ≤ n b pN
n (log nb )−1/2 (bnhH φx (hK ))−1/2 ϕ(pn ).
1/2N
Prenons pn C nhHlogφx (hK )
, la définition de zn sous l’hypothèse (H7), un
b
n
b
choix convenable de η0 nous permettent d’écrire
X 3 1
nb 2 α+ 2 (B20 + B10 ) < ∞.
n∈In
D’où
b − θ(x)|2 ≤ C
|θ(x) x00
hx (y) − hx (y)|.
sup |b
min h (y) y∈S
y∈S
En utilisant le résultat du Théorème 1, on montre que
1 !
2
b b
log n 2
|θ(x)
b − θ(x)| = O h + h 1 2
b
K H +O .
n
b hH φx (hK )
REFERENCES
[1] G. Biau and B. Cadre, Nonparametric spatial prediction. Stat. Inference Stoch. Process
7 (2004), 327–349.
[2] D. Bosq, Linear Processes in Function Spaces: Theory and Applications. Lecture Notes
in Statistics 149, Springer, 2000.
17 Fonction de hasard spatiale et variable fonctionnelle 51
[3] M. Carbon, L.T. Tran and B. Wu, Kernel density estimation for random fields. Statist.
Probab. Lett. 36 (1997), 115–125.
[4] M. Carbon, C. Francq and L.T. Tran, Kernel regression estimation for random fields.
J. Statist. Plann. Inference 137 (2007), 778–798.
[5] N.A. Cressie, Statistics for Spatial Data. Wiley, New York, 1991.
[6] S. Dabo-Niang and A.F. Yao, Kernel regression estimation for continuous spatial pro-
cesses. Math. Methods Statist 16 (2007), 298–317.
[7] G. Estévez-Pérez, A. Quintela-del-Ro and Ph. Vieu, Convergence rate for cross-
validatory bandwidth in kernel hazard estimation from dependent samples. J. Statist.
Plann. Inference 104 (2002), 1–30.
[8] F. Ferraty, A. Laksaci and Ph. Vieu, Estimating some characteristics of the conditional
distribution in nonparametric functional models. Stat. Inference Stoch. Process 9 (2006),
47–76.
[9] F. Ferraty, A. Rabhi and Ph. Vieu, Estimation non-paramétrique de la fonction de
hasard avec variable explicative fonctionnelle. Rev. Roumaine Math. Pures Appl. 53
(2008), 1–18.
[10] F. Ferraty, A. Tadj, A. Laksaci and Ph. Vieu, Rate of uniform consistency for nonpara-
metric estimates with functional variables. J. Statist. Plann. Inference. (accepté)
[11] F. Ferraty and Ph. Vieu, Nonparametric Functional Data Analysis. Springer Series in
Statistics. Springer, New York, 2006.
[12] X, Guyon, Estimation d’un champ par pseudo-vraisemblance conditionnelle: Etude
asymptotique et application au cas Markovien. Proceedings of the Sixth Franco-Belgian
Meeting of Statisticians, 1987.
[13] A. Quintela-del-Ro, Nonparametric estimation of the maximum hazard under depen-
dence conditions. Statist. Probab. Lett. 76 (2006), 1117–1124.
[14] A. Quintela-del-Ro, Hazard function given a functional variable: Non-parametric esti-
mation under strong mixing conditions. J. Nonparametr. Stat. 20 (2008), 413–430.
[15] J.P. Lecoutre and E. Ould-Said, Estimation de la densité et de la fonction de hasard
conditionnelle pour un processus fortement mélangeant avec censure. C.R. Math. Acad.
Sci. Paris 314 (1992), 295–300.
[16] J. Li and L.T. Tran, Hazard rate estimation on random fields. J. Multivariate Anal. 98
(2007), 1337–1355.
[17] J.Li and L.T. Tran, Nonparametric estimation of conditional expectation. J. Statist.
Plann. Inference 139 (2009), 164–175.
[18] E. Masry, Recursive probability density estimation for weakly dependent stationary pro-
cesses. IEEE Trans. Inform. Theory 32 (1986), 254–267.
[19] J.O. Ramsay and B.W. Silverman, Applied Functional Data Analysis; Methods and Case
Studies. Springer-Verlag, New York, 2002.
[20] G. Roussas, Hazard rate estimation under dependence conditions. J. Statist. Plann.
Inference 22 (1989), 81–93.
[21] L.T. Tran, Kernel density estimation on random fields. J. Multivariate Anal. 34 (1990),
37–53.
[22] G.S. Watson and M.R. Leadbetter, Hazard analysis. Sankhyia 26 (1964), 101–116.