HAL Id: hal-00488676
https://hal.archives-ouvertes.fr/hal-00488676
Preprint submitted on 3 Jun 2010
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.
Berry-Esseen’s central limit theorem for non-causal linear processes in Hilbert space
Mohamed El Machkouri
To cite this version:
Mohamed El Machkouri. Berry-Esseen’s central limit theorem for non-causal linear processes in Hilbert space. 2010. �hal-00488676�
linear proesses in Hilbert spae
Mohamed EL MACHKOURI
June 2, 2010
Abstrat
LetHbearealseparableHilbertspaeand(ak)k∈Zasequeneofboundedlinear
operatorsfromH to H. WeonsiderthelinearproessX dened foranykinZby Xk=P
j∈Zaj(εk−j)where(εk)k∈Zisasequeneofi.i.d. enteredH-valuedrandom
variables. WeinvestigatetherateofonvergeneintheCLTforXandinpartiular
we obtain the usual Berry-Esseen's bound provided that
P
j∈Z|j|kajkL(H) < +∞
andε0 belongs to L∞H.
Shorttitle: Berry-Esseen'sCLT for Hilbertian linear proesses.
Key words: Central limit theorem, Berry-Esseen bound, linear proess, Hilbert
spae.
1 Introdution and notations
Let(H,k.kH) be a separable real Hilbert spae and (L,k.kL(H)) be the lass of bounded
linear operators fromH to H with its usual uniform norm. Consider a sequene (εk)k∈Z
of i.i.d. entered random variables, dened on a probability spae (Ω,A,P), with values
inH. If(ak)k∈Z isasequene inL,wedenethe(non-ausal)linearproessX = (Xk)k∈Z
inH by
Xk=X
j∈Z
aj(εk−j), k ∈Z. (1)
If
P
j∈ZkajkL(H) <∞ and Ekε0kH < +∞ then the series in (1) onverges almost surely
andinL1H(Ω,A,P)(see Bosq[2℄). Theondition P
j∈ZkajkL(H) <∞isknowtobesharp
for the
√n-normalized partial sums of X to satises a CLT provided that (εk)k∈Z are
i.i.d. entered having nite seond moments(see Merlevede et al. [6℄). In this work, we
investigate the rate of onvergene in the CLT for X under the ondition
X
j∈Z
|j|τkajkL(H) <∞ (2)
withτ = 1when(εk)k∈Z areassumed tobei.i.d. enteredandsuhthatε0 belongstoL∞H
andτ = 1/2when(εk)k∈Zarei.i.d. enteredandsuhthatε0belongstosomeOrlizspae LH,ψ (see setion 2). This problem was previously studied(with τ = 1 in Condition (2))
by Bosq [3℄ for (ausal) Hilbert linear proesses but a mistake in his proof was pointed
out by V. Paulauskas [7℄. However, in the partiular ase of Hilbertian autoregressive
proesses of order 1, Bosq [4℄ obtained the usual Berry-Esseen inequality provided that
(εk)k∈Z are i.i.d. entered with ε0 in L∞H.
2 Main result
Inthe sequel,Cε0 istheautoovarianeoperatorofε0,A:=P
j∈Zaj andA∗ istheadjoint
of A. For any sequene Z = (Zk)k∈Z of randomvariableswith values in H wedenote
∆n(Z) = sup
t∈R
P
√1 n
n
X
k=1
Zk
H
≤t
!
−P(kNkH ≤t)
whereN ∼ N(0, ACε0A∗).
Forany j ∈Z, denote cj,n=Pn
i=1bi−j where bi =ai for any i6= 0and b0 =a0−A.
Lemma 1 For any positive integer n,
n
X
k=1
Xk =A
n
X
k=1
εk
!
+Qn+Rn
where Qn=Pn k=1
P
|j|>nak−j(εj) and Rn=P
|j|≤ncj,n(εj).
Reall that a Young funtion ψ is a real onvex nondereasing funtion dened on R+
whih satises limt→+∞ψ(t) = +∞ and ψ(0) = 0. We dene the Orliz spae LH,ψ as
thespaeof H-valuedrandomvariablesZ dened onthe probabilityspae(Ω,F,P)suh
that E[ψ(kZkH/c)] < +∞ for some c > 0. The Orliz spae LH,ψ equipped with the
so-alledLuxemburg norm k.kψ dened for any H-valued randomvariableZ by kZkψ = inf{c >0 ;E[ψ(kZkH/c)]≤1}
is a Banah spae. In the sequel, c(N) denotes a bound of the density of N(0, ACε0A∗)
(see Davydov et al. [5℄). Our main result isthe following.
Theorem 1 Let (εk)k∈Z be a sequene of i.i.d. entered H-valued random variables and
let X be the Hilbertian linear proess dened by (1).
i) If ε0 belongsto L∞H and P
j∈Z|j|kajkL(H) <∞ then
∆n(X)≤ c1
√n (3)
where c1 = c2 + 14c(N)kε0k∞P
j∈Z|j|kajkL(H) and c2 is a positive onstant whih
depend only on the distribution of ε0.
ii) If ψ is a Young funtion then
∆n(X)≤∆n(A(ε)) +ϕ
c(N)kQn+Rnkψ
√n
(4)
where ϕ(x) =xh−1(1/x) and h(x) =xψ(x) for any real x >0.
The inequality (4) ensures a rate of onvergene to zero for ∆n(X) as n goes to innity
provided that ∆n(A(ε0)) goes to zero as n goes to innity and a bound for kQn+Rnkψ
exists. As just anillustration, wehave the followingorollary.
Corollary 1 Assume that(εk)k∈Z arei.i.d. enteredH-valued random variablesandthat
the ondition (2) holds with τ = 1/2.
i) Ifε0belongstoLH,ψ1 then∆n(X) =O
log√n n
whereψ1 istheYoungfuntiondened
byψ1(x) = exp(x)−1.
ii) If ε0 belongsto LrH for r≥3 then ∆n(X) = O
n−2(r+1)r
.
Proof of Lemma 1. For any positiveinteger n, we have Rn =
n
X
j=−n
cj,n(εj) =
n
X
k=1 n
X
j=−n
bk−j(εj)
=
n
X
k=1
X
j∈[−n,n]\{k}
ak−j(εj) + (a0−A)
n
X
k=1
εk
!
=
n
X
k=1 n
X
j=−n
ak−j(εj)−A
n
X
k=1
εk
!
=−Qn+
n
X
k=1
Xk−A
n
X
k=1
εk
! .
The proof of Lemma1 isomplete.
Proof of Theorem 1. Let λ > 0 and t > 0 be xed and denote U = A(Pn
k=1εk/√ n)
and V = (Qn+Rn)/√
n. SoU +V =Pn
k=1Xk/√ n and
P(kU+VkH ≤t)≤P(kUkH ≤t+λ) +P(kVkH ≥λ) (5)
Forλ0 = 2kVk∞, we obtain
P(kU +VkH ≤t)−P(kNkH ≤t)≤P(kUkH ≤t+λ0)−P(kNkH ≤t).
Ifc(N) denotes abound for the density of kNkH (see Davydov et al. [5℄)then
∆n(X)≤∆n(A(ε)) + 2c(N)kQ√n+Rnk∞
n .
Noting that
Qn = X
j≥n+2
aj
−n−1
X
k=1−j
εk
! +X
j<0
aj n−j
X
k=n+1
εk
!
(6)
and
Rn=R′n+R′′n (7)
where
Rn′ =−
−1
X
j=−n
aj
−j
X
k=1
εk
!
− X
j<−n
aj
n
X
k=1
εk
!
−X
j>0
aj
n
X
k=n−j+1
εk
!
R′′n =
n
X
j=1
aj 0
X
k=−j+1
εk
! +
2n
X
j=n+1
aj n−j
X
k=−n
εk
! ,
wederive that kQn+Rnk∞≤7kε0k∞P
j∈Z|j|kajkL(H) and onsequently
∆n(X)≤∆n(A(ε)) + 14c(N)kε0k∞P
j∈Z|j|kajkL(H)
√n .
Combining the last inequality with the Berry-Esseen inequality for i.i.d. entered H-
valued randomvariables (see Yurinski [9℄or Bosq [2℄, Theorem 2.9) we obtain (3).
In the other part, if ψ is a Young funtion we have P(kVkH ≥ λ) ≤ ψ(λ/k1Vkψ) and
keeping in mindinequality (5), we derive
∆n(X)≤∆n(A(ε)) +c(N)λ+ 1 ψ(λ/kVkψ).
Noting that c(N)λ = 1
ψ(λ/kVkψ) if and only if λ = ϕ(c(N)kVkψ)
c(N) where ϕ is dened by ϕ(x) =xh−1(1/x)and h by h(x) =xψ(x), we onlude
∆n(X)≤∆n(A(ε)) +ϕ
c(N)kQn+Rnkψ
√n
.
The proof of Theorem 1is omplete.
Proof of Corollary 1. Assume that kε0kψ1 < ∞ where ψ1 is the Young funtion
dened by ψ1(x) = exp(x)−1. There exists a > 0 suh that E(exp(akε0kH)) ≤ 2. So,
thereexist (see Arak and Zaizsev [1℄) onstants B and L suh that Ekε0kmH ≤ m!
2 B2Lm−2, m= 2,3,4, ...
Applying Pinelis-Sakhanenkoinequality (see Pinelis and Sakhanenko [?℄ or Bosq [2℄), we
obtain
P
q
X
k=p
εk
H
≥x
!
≤exp
− x2
2(q−p+ 1)B2+ 2xL
, x >0
andusingLemma2.2.10inVanDerVaartandWellner[?℄,thereexistsauniversalonstant
K suh that
q
X
k=p
εk ψ1
≤K
L+Bp
q−p+ 1
(8)
Combining (6), (7) and (8), we derive kQn+Rnkψ1 ≤ CP
j∈Z
p|j|kajkL(H) where the
onstantC doesnot depend on n. Keeping inmind the Berry-Esseen's entrallimitthe- oremfori.i.d. entered H-valuedrandomvariables(seeYurinski [9℄orBosq [2℄, Theorem
2.9), we apply Theorem 1 with the Young funtion ψ1. Sine the funtion ϕ dened by ϕ(x) =xh−1(1/x)with h(x) =xψ1(x) satises
limx→0
ϕ(x)
xlog(1 + x1) = 0,
wederive ∆n(X) = O
logn
√n
.
Now, assume that kε0kr < ∞ for some r ≥ 3. Applying Pinelis inequality (see Pinelis
[8℄), thereexists a universal onstant K suh that
q
X
k=p
εk
r
≤K
r
q
X
k=p
EkεkkrH
!1/r
+√ r
q
X
k=p
Ekεkk2H
!1/2
and onsequently
q
X
k=p
εk
r
≤2Krkε0kr
pq−p+ 1. (9)
Combining (6), (7) and (9), we derive kQn +Rnkr ≤ CP
j∈Z
p|j|kajkL(H) where the
onstant C does not depend on n. Again, applying Berry-Esseen's CLT (see Yurinski [9℄ or Bosq [2℄, Theorem 2.9) and Theorem 1 with the Young funtion ψ(x) = xr and
the funtion ϕ given by ϕ(x) = xr/(r+1), we obtain∆n(X) = O
n−2(r+1)r
. The proof of
Corollary1 isomplete.
Referenes
[1℄ T. V. Arak and A. Yu. Zatsev. Uniform limit theorems for sums of independent
random variables. Pro. Steklov Inst. Math., (1(174)), 1988. A translation of Trudy
Mat.Inst. Steklov. 174(1986).
[2℄ Denis Bosq. Linear proesses in funtion spaes, volume 149 of Leture Notes in
Statistis. Springer-Verlag, New York,2000. Theory and appliations.
[3℄ Denis Bosq. Berry-Esseen inequality for linear proesses in Hilbert spaes. Statist.
Probab. Lett., 63(3):243247,2003.
essesinHilbertspaes [Statist.Probab.Lett.63(2003),no.3,243247;mr1986323℄.
Statist. Probab. Lett., 70(2):171174,2004.
[5℄ Yu.A.Davydov,M.A.Lifshits,andN.V.Smorodina.Loalpropertiesofdistributions
of stohasti funtionals, volume 173 of Translations of Mathematial Monographs.
Amerian Mathematial Soiety, Providene, RI, 1998. Translated from the 1995
Russian originalby V. E. Nazakinskiand M. A. Shishkova.
[6℄ FloreneMerlevède, MagdaPeligrad,andSergey Utev. Sharponditions forthe CLT
of linear proesses ina Hilbert spae. J. Theoret. Probab., 10(3):681693,1997.
[7℄ V. Paulauskas. Personal ommuniation, 2004.
[8℄ Iosif Pinelis. Optimum bounds for the distributions of martingales inBanah spaes.
Ann. Probab., 22(4):16791706,1994.
[9℄ V.V.Yurinski. Ontheaurayofnormalapproximationoftheprobabilityofhitting
a ball. Teor. Veroyatnost. i Primenen., 27(2):270278,1982.
MohamedEL MACHKOURI
Laboratoirede Mathématiques Raphaël Salem
UMRCNRS 6085, Université de Rouen
Avenue de l'université
76801Saint-Etienne du Rouvray
mohamed.elmahkouriuniv-rouen.fr