HAL Id: hal-00199024
https://hal.archives-ouvertes.fr/hal-00199024v5
Preprint submitted on 10 May 2008
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.
Stein’s method on Wiener chaos
Ivan Nourdin, Giovanni Peccati
To cite this version:
Ivan Nourdin, Giovanni Peccati. Stein’s method on Wiener chaos. 2008. �hal-00199024v5�
hal-00199024, version 5 - 10 May 2008
byIvanNourdin
∗
and GiovanniPeati†
University of Paris VI
Revisedversion: May10,2008
Abstrat: We ombine Malliavin alulus with Stein's method, in order to derive expliit bounds
in theGaussian and Gamma approximationsof random variables in axed Wiener haos of a gen-
eral Gaussianproess. Ourapproahgeneralizes,renesanduniestheentral andnon-entrallimit
theorems formultiple Wiener-Itintegralsreentlyproved(inseveralpapers,from 2005to 2007)by
Nourdin,Nualart,Ortiz-Latorre,Peati andTudor. Weapplyourtehniques toproveBerry-Esséen
bounds in the Breuer-Major CLT for subordinated funtionals of frational Brownian motion. By
usingthewell-knownMehler'sformulaforOrnstein-Uhlenbeksemigroups,wealsoreoveratehnial
resultreentlyprovedbyChatterjee,onerningtheGaussian approximationoffuntionals ofnite-
dimensional Gaussianvetors.
Key words: Berry-Esséenbounds;Breuer-MajorCLT;FrationalBrownianmotion;Gamma
approximation;Malliavinalulus;Multiplestohastiintegrals;Normalapproximation;Stein's
method.
2000 Mathematis SubjetClassiation: 60F05;60G15;60H05; 60H07.
1 Introdution and overview
1.1 Motivations
Let
Z
be arandom variable whose lawisabsolutely ontinuouswith respetto theLebesguemeasure(for instane,
Z
is a standard Gaussian random variable or a Gamma random vari-able). Supposethat
{ Z n : n > 1 }
isasequeneofrandomvariablesonvergingindistribution towardsZ
,that is:for all
z ∈ R
,P (Z n 6 z) −→ P(Z 6 z)
asn → ∞
. (1.1)It is sometimes possible to assoiate an expliit uniform bound with the onvergene (1.1),
providingaglobaldesriptionoftheerroronemakeswhenreplaing
P (Z n 6 z)
withP (Z 6 z)
for axed
n > 1
. Oneof the most elebrated results inthis diretion isthe following Berry-Esséen Theorem (see e.g. Feller [17℄for aproof), thatwe reordhere for futurereferene:
Theorem 1.1 (Berry-Esséen) Let
(U j ) j>1
beasequene ofindependentandidentiallydis- tributed random variables, suh thatE( | U j | 3 ) = ρ < ∞
,E(U j ) = 0
andE(U j 2 ) = σ 2
. Then,∗
LaboratoiredeProbabilitésetModèlesAléatoires,UniversitéPierreetMarieCurie,Boîteourrier188,4
PlaeJussieu,75252Paris Cedex5,Frane,ivan.nourdinupm.fr
†
LaboratoiredeStatistiqueThéoriqueetAppliquée,UniversitéPierreetMarieCurie,8èmeétage,bâtiment
A,175rueduChevaleret,75013Paris,Frane,giovanni.peatigmail.om
by setting
Z n = σ √ 1 n
P n
j=1 U j , n > 1,
one has thatZ n −→ Law Z ∼ N (0, 1)
, asn → ∞
, andmoreover:
sup
z ∈ R | P (Z n 6 z) − P(Z 6 z) | 6 3ρ
σ 3 √ n .
(1.2)The aim of this paper is to show that one an ombine Malliavin alulus (see e.g. [35 ℄)
andStein'smethod (seee.g. [11℄),inordertoobtainboundsanalogous to(1.2),wheneverthe
randomvariables
Z n
in(1.1)an be representedasfuntionals ofa given Gaussian eld. Our results are general, in the sense that (i) they do not rely on any spei assumption on theunderlyingGaussianeld,(ii)theydonot requirethatthevariables
Z n
havethespeiformofpartialsums,and (iii)they allow todeal(atleastintheaseofGaussian approximations)
withseveraldierent notions of distane between probabilitymeasures. Assuggested bythe
title, a prominent role will be played by random variables belonging to a Wiener haos of
order
q
(q > 2
), that is, random variables having the form of a multiple stohasti Wiener-It integral of order
q
(see Setion 2 below for preise denitions). It will be shown that our results provide substantial renements of the entral and non-entral limit theorems formultiple stohastiintegrals, reently proved in [33 ℄ and [37℄. Among other appliations and
examples, wewill provideexpliitBerry-EsséenboundsintheBreuer-Major CLT(see[5℄) for
elds subordinatedto a frational Brownian motion.
Conerning point (iii), we shall note that, as a by-produt of the exibility of Stein's
method,we will indeedestablishboundsfor Gaussian approximations relatedtoa numberof
distanes of the type
d H (X, Y ) = sup {| E(h(X)) − E(h(Y )) | : h ∈ H } ,
(1.3)where
H
is some suitable lass of funtions. For instane: by takingH = { h : k h k L 6 1 } ,
where
k·k L
denotestheusualLipshitzseminorm,oneobtainstheWasserstein(orKantorovih- Wasserstein) distane; by takingH = { h : k h k BL 6 1 } ,
wherek·k BL = k·k L + k·k ∞
, oneobtains the Fortet-Mourier (or bounded Wasserstein) distane; by taking
H
equal to theolletionofallindiators
1 B
ofBorelsets,oneobtains thetotalvariationdistane;bytakingH
equal to the lass of all indiators funtions1 ( −∞ ,z]
,z ∈ R
, one has the Kolmogorov distane, whih is the one taken into aount in the Berry-Esséen bound (1.2). In whatfollows, we shall sometimes denote by
d W (., .)
,d FM (., .)
,d TV (., .)
andd Kol (., .)
, respetively, the Wasserstein, Fortet-Mourier, total variation and Kolmogorov distanes. Observe thatd W (., .) > d FM (., .)
andd TV (., .) > d Kol (., .)
. Also, the topologies indued byd W
,d TV
andd Kol
are stronger than the topology of onvergene in distribution, while one an show thatd FM
metrizestheonvergeneindistribution(seee.g. [16,Ch. 11℄fortheseandfurtherresults involvingdistanes onspaes of probabilitymeasures).1.2 Stein's method
We shall now give a short aount of Stein's method, whih is basially a set of tehniques
allowingtoevaluatedistanesofthetype(1.3)bymeansofdierentialoperators. Thistheory
has been initiated by Stein in the path-breaking paper [48℄, and then further developed in
the monograph [49 ℄. The readerisreferredto [11 ℄,[43 ℄ and [44℄for detailedsurveys ofreent
resultsandappliations. ThepaperbyChatterjee[8 ℄providesfurtherinsightsintotheexisting
literature. In what follows, we will apply Stein's method to two types of approximations,
namely Gaussian and (entered) Gamma. We shall denote by
N (0, 1)
a standard Gaussianrandom variable. Theentered Gammarandom variableswe areinterested inhavethe form
F (ν ) Law = 2G(ν/2) − ν, ν > 0,
(1.4)where
G(ν/2)
has a Gamma law with parameterν/2
. This means thatG(ν/2)
is a (a.s.stritly positive)random variablewithdensity
g(x) = x
ν 2 −1 e −x
Γ(ν/2) 1 (0, ∞ ) (x),
whereΓ
istheusualGamma funtion. We hoose this parametrization inorder to failitate the onnetion with
our previous paper [33℄ (observe in partiular that, if
ν > 1
is an integer, thenF (ν)
has aentered
χ 2
distribution withν
degrees of freedom).StandardGaussian distribution. Let
Z ∼ N (0, 1)
. Consider a real-valued funtionh : R → R
suhthattheexpetationE(h(Z))
iswell-dened. TheSteinequationassoiatedwithh
andZ
islassially given byh(x) − E(h(Z)) = f ′ (x) − xf (x), x ∈ R .
(1.5)A solution to (1.5) is a funtion
f
whih is Lebesgue a.e.-dierentiable, and suh that there exists a version off ′
verifying (1.5) for everyx ∈ R
. The following result is basially due to Stein [48, 49℄. Theproof of point (i) (whose ontent is usually referredas Stein's lemma)involves astandard use ofthe Fubini theorem(see e.g. [47 ℄ or [11 ,Lemma 2.1℄). Point (ii)is
proved e.g. in [11 , Lemma 2.2℄; point (iii) an be obtained by ombining e.g. thearguments
in[49, p. 25℄and [9, Lemma 5.1℄; a proofof point (iv)is ontained in[49 , Lemma 3,p. 25℄;
point (v) isproved in[8 , Lemma 4.3℄.
Lemma 1.2 (i) Let
W
be a random variable. Then,W Law = Z ∼ N (0, 1)
if, andonly if,E[f ′ (W ) − W f (W )] = 0,
(1.6)for every ontinuous and pieewise ontinuously dierentiable funtion
f
verifying therelation
E | f ′ (Z) | < ∞
.(ii) If
h(x) = 1 ( −∞ ,z] (x)
,z ∈ R
, then (1.5)admitsa solutionf
whih isbounded by√ 2π/4
,pieewise ontinuously dierentiable andsuh that
k f ′ k ∞ 6 1
.(iii) If
h
is bounded by 1/2, then (1.5) admits a solutionf
whih is bounded byp π/2
,Lebesgue a.e. dierentiable and suh that
k f ′ k ∞ 6 2
.(iv) If
h
is bounded and absolutely ontinuous (then, in partiular, Lebesgue-a.e. dieren-tiable), then (1.5) has a solution
f
whih is bounded and twie dierentiable, and suh thatk f k ∞ 6 p
π/2 k h − E(h(Z)) k ∞
,k f ′ k ∞ 6 2 k h − E(h(Z )) k ∞
andk f ′′ k ∞ 6 2 k h ′ k ∞
.(v) If
h
isabsolutely ontinuous withbounded derivative, then (1.5) has a solutionf
whihis twie dierentiable andsuh that
k f ′ k ∞ 6 k h ′ k ∞
andk f ′′ k ∞ 6 2 k h ′ k ∞
.We alsoreall the relation:
2d TV (X, Y ) = sup {| E(u(X)) − E(u(Y )) | : k u k ∞ 6 1 } .
(1.7)Notethat point (ii) and (iii) (via(1.7)) imply the following boundson theKolmogorov and
totalvariation distanebetween
Z
andan arbitraryrandom variableY
:d Kol (Y, Z) 6 sup
f ∈F Kol
| E(f ′ (Y ) − Y f (Y )) |
(1.8)d TV (Y, Z) 6 sup
f ∈ F TV
| E(f ′ (Y ) − Y f (Y )) |
(1.9)where
F Kol
andF TV
are,respetively,thelassof pieewiseontinuouslydierentiablefun- tions that are bounded by√ 2π/4
and suh that their derivative is bounded by 1, and thelassof pieewise ontinuously dierentiable funtions that are bounded by
p π/2
and suhthattheir derivative is bounded by 2.
Analogously,byusing (iv)and(v) alongwith therelation
k h k L = k h ′ k ∞
,oneobtainsd FM (Y, Z) 6 sup
f ∈F FM
| E(f ′ (Y ) − Y f (Y )) | ,
(1.10)d W (Y, Z) 6 sup
f ∈ F W
| E(f ′ (Y ) − Y f (Y )) | ,
(1.11)where:
F FM
is the lass of twie dierentiable funtions that are bounded by√ 2π
, whoserst derivative is bounded by 4, and whose seond derivative is bounded by 2;
F W
is thelassoftwiedierentiablefuntions,whoserstderivativeisboundedby1andwhoseseond
derivative isbounded by2.
Centered Gamma distribution. Let
F (ν )
be as in (1.4). Consider a real-valued funtionh : R → R
suhthattheexpetationE[h(F (ν))]
exists. TheSteinequationassoiatedwithh
and
F (ν)
is:h(x) − E[h(F (ν))] = 2(x + ν)f ′ (x) − xf (x), x ∈ ( − ν, + ∞ ).
(1.12)The following statement ollets some slight variations around results proved by Stein [49 ℄,
Diaonis and Zabell [15 ℄, Luk [26℄, Shoutens [46℄ and Pikett [41 ℄. It is theGamma oun-
terpart of Lemma 1.2. Theproofis detailedinSetion 7.
Lemma 1.3 (i) Let
W
be a real-valued random variable (not neessarily with values in( − ν, + ∞ )
) whose law admits a density with respet to the Lebesgue measure. Then,W Law = F(ν)
if, andonly if,E[2(W + ν) + f ′ (W ) − W f(W )] = 0,
(1.13)where
a + := max(a, 0)
, for every smooth funtionf
suh that the mappingx 7→ 2(x + ν ) + f ′ (x) − xf (x)
isbounded.(ii) If
| h(x) | 6 c exp(ax)
foreveryx > − ν
andforsomec > 0
anda < 1/2
,andifh
istwiedierentiable, then(1.12) hasa solution
f
whihisbounded on( − ν, + ∞ )
, dierentiable and suh thatk f k ∞ 6 2 k h ′ k ∞
andk f ′ k ∞ 6 k h ′′ k ∞
.(iii) Suppose that
ν > 1
is an integer. If| h(x) | 6 c exp(ax)
for everyx > − ν
and forsome
c > 0
anda < 1/2
, and ifh
is twie dierentiable with bounded derivatives, then (1.12) hasa solutionf
whihisbounded on( − ν, + ∞ )
, dierentiable andsuh thatk f k ∞ 6 p
2π/ν k h k ∞
andk f ′ k ∞ 6 p
2π/ν k h ′ k ∞
.Now dene
H 1 = { h ∈ C b 2 : k h k ∞ 6 1, k h ′ k ∞ 6 1 } ,
(1.14)H 2 = { h ∈ C b 2 : k h k ∞ 6 1, k h ′ k ∞ 6 1, k h ′′ k ∞ 6 1 } ,
(1.15)H 1,ν = H 1 ∩ C b 2 (ν)
(1.16)H 2,ν = H 2 ∩ C b 2 (ν)
(1.17)where
C b 2
denotes the lass of twie dierentiable funtions (with support inR
) and with boundedderivatives, andC b 2 (ν)
denotes thesubsetofC b 2
omposedoffuntionswithsupportin
( − ν, + ∞ )
. Note that point (ii) in the previous statement implies that, by adopting thenotation (1.3) and for every
ν > 0
and every real random variableY
(not neessarily withsupportin
( − ν, + ∞ )
),d H 2,ν (Y, F (ν)) 6 sup
f ∈F 2,ν
| E[2(Y + ν)f ′ (Y ) − Y f (Y )] |
(1.18)where
F 2,ν
is the lass of dierentiable funtions with support in( − ν, + ∞ )
, bounded by 2andwhose rstderivatives arebounded by 1. Analogously,point (iii) impliesthat, for every
integer
ν > 1
,d H 1,ν (Y, F (ν)) 6 sup
f ∈ F 1,ν
| E[2(Y + ν)f ′ (Y ) − Y f (Y )] | ,
(1.19)where
F 1,ν
is the lass of dierentiable funtions with support in( − ν, + ∞ )
, bounded byp 2π/ν
and whose rst derivatives are also bounded byp 2π/ν
. A little inspetion showsthatthefollowingestimates also hold: for every
ν > 0
andevery random variableY
,d H 2 (Y, F (ν)) 6 sup
f ∈ F 2
| E[2(Y + ν) + f ′ (Y ) − Y f (Y )] |
(1.20)where
F 2
is the lass of funtions (dened onR
) that are ontinuous and dierentiable onR \{ ν }
, bounded bymax { 2, 2/ν }
, and whose rst derivatives are bounded bymax { 1, 1/ν + 2/ν 2 }
. Analogously,for everyintegerν > 1
,d H 1 (Y, F (ν)) 6 sup
f ∈ F 1
| E[2(Y + ν) + f ′ (Y ) − Y f (Y )] | ,
(1.21)where
F 1
is the lass of funtions (onR
) that are ontinuous and dierentiable onR \{ ν }
,boundedby
max { p
2π/ν, 2/ν }
,andwhoserstderivativesareboundedbymax { p
2π/ν, 1/ν + 2/ν 2 }
.Now, the ruial issue is how to estimatethe right-hand side of (1.8)(1.11) and (1.18)
(1.21) for a given hoie of
Y
. Sine Stein's initial ontribution [48 ℄, an impressive panoply of tehniques has been developed in this diretion (see again [10 ℄ or [43 ℄ for a survey; here,we shallquote e.g.: exhangeable pairs [49℄,diusion generators [3, 19℄, size-bias transforms
[20 ℄, zero-bias transforms [21 ℄, loal dependeny graphs [10℄ and graphial-geometri rules
[8℄). Starting from the next setion, we will show that, when working within theframework
of funtionals of Gaussian elds, one an veryeetively estimate expressions suh as(1.8)
(1.11),(1.18) and(1.19) byusingtehniquesofMalliavinalulus. Interestingly,aentralrole
isplayed byan innite dimensional version of the same integration by parts formula that is
atthevery heart ofStein's haraterization oftheGaussian distribution.
1.3 The basi approah (with some examples)
Let
H
be a realseparable Hilbertspaeand, forq > 1
, letH ⊗ q
(resp.H ⊙ q
) be theq
thtensorprodut (resp.
q
thsymmetri tensorprodut) ofH
. We writeX = { X(h) : h ∈ H }
(1.22)to indiate a entered isonormal Gaussian proess on
H
. For everyq > 1
, we denote byI q
the isometry between
H ⊙ q
(equipped with the norm√
q! k · k H ⊗q
) and theq
th Wiener haosof
X
. Note that, ifH
is aσ
-nite measure spae withno atoms, theneah random variableI q (h)
,h ∈ H ⊙ q
, has the form of a multiple Wiener-It integral of orderq
. We denote byL 2 (X) = L 2 (Ω, σ(X), P )
the spae of square integrable funtionals ofX
, and byD 1,2
the domain of the Malliavin derivative operatorD
(see the forthoming Setion 2 for preisedenitions). Reall that, for every
F ∈ D 1,2
, the derivativeDF
is a random element withvalues in
H
.We start by observing that, thanks to (1.6), for every
h ∈ H
suh thatk h k H = 1
and foreverysmooth funtion
f
,wehaveE[X(h)f (X(h))] = E[f ′ (X(h))]
. Ourpointisthatthis lastrelation is a very partiular ase of the following orollary of the elebrated integration by
partsformulaofMalliavin alulus: for every
Y ∈ D 1,2
withzero mean,E[Y f (Y )] = E[ h DY, − DL − 1 Y i H f ′ (Y )],
(1.23)where thelinear operator
L − 1
is the inverse of the generator of theOrnstein-Uhlebek semi- group,notedL
. The readerisreferredtoSetion 2andSetion 3for denitionsandfor afulldisussion of this point; here, we shall note that
L
is an innite-dimensional version of the generator assoiated with Ornstein-Uhlenbek diusions (see [35 , Setion 1.4℄ for a proof ofthisfat),anobjetwhihisalsoruialintheBarbour-Götzegeneratorapproah toStein's
method [3,19℄.
Itfollowsthat,forevery
Y ∈ D 1,2
withzeromean,theexpressionsappearingontheright- hand side of (1.8)(1.11) (or (1.18)(1.21)) an be assessed by rst replaingY f (Y )
withh DY, − DL − 1 Y i H f ′ (Y )
insidetheexpetation,andthenbyevaluatingtheL 2
distanebetween1
(resp.2Y + 2ν
)andtheinnerproduth DY, − DL − 1 Y i H
. Ingeneral,theseomputationsare arriedout byrstresorting to therepresentation ofh DY, − DL − 1 Y i H
asa(possiblyinnite)series of multiple stohasti integrals. We will see that, when
Y = I q (g)
, forq > 2
andsome
g ∈ H ⊙ q
,thenh DY, − DL − 1 Y i H = q − 1 k DY k 2 H
. Inpartiular, byusing this last relationone an dedue bounds involving quantities that are intimately related to the entral and
non-entral limit theorems reently proved in[37℄,[36 ℄ and[33 ℄.
Remark 1.4 1. Theruialequality
E[I q (g)f (I q (g))] = E[q − 1 k DI q (g) k 2 H f ′ (I q (g))]
,inthease where
f
is a omplex exponential, has been rst used in [36℄, in order to give renements (as well asalternate proofs) of the main CLTs in [37℄ and [39 ℄. The samerelation has been later applied in [33℄, where a haraterization of non-entral limit
theoremsformultipleintegralsisprovided. Notethatneither[33℄nor[36 ℄areonerned
with Stein's method or, more generally, with bounds on distanes between probability
measures.
2. We will see that formula (1.23) ontains as a speial ase a result reently proved by
Chatterjee [9 ,Lemma 5.3℄,intheontextof limittheoremsfor linearstatistisofeigen-
values of random matries. Theonnetion between thetwo results an be established
by means of the well-known Mehler's formula (see e.g. [28, Setion 8.5, Ch. I℄ or
[35 , Setion 1.4℄), providing a mixture-type representation of the innite-dimensional
Ornstein-Uhlenbek semigroup. See Remarks 3.6 and 3.12 below for a preise disus-
sion of this point. See e.g. [31℄ for a detailed presentation of the innite-dimensional
Ornstein-Uhlebek semigroup.
3. We stress that therandom variable
h DY, − DL − 1 Y i H
appearing in (1.23) is ingeneralnot measurable with respet to
σ(Y )
. For instane, ifX
is taken to be the Gaussianspae generated by a standard Brownian motion
{ W t : t > 0 }
andY = I 2 (h)
withh ∈ L 2 s ([0, 1] 2 )
,thenD t Y = 2 R 1
0 h(u, t)dW u
,t ∈ [0, 1]
,andh DY, − DL − 1 Y i L 2 ([0,1]) = 2 I 2 (h ⊗ 1 h) + 2 k h k 2 L 2 ([0,1] 2 )
whih is, ingeneral, not measurable withrespetto
σ(Y )
(thesymbolh ⊗ 1 h
indiatesa ontrationkernel, anobjet thatwill be dened inSetion 2).
4. Note that(1.23) also impliestherelation
E[Y f (Y )] = E[τ (Y )f ′ (Y )],
(1.24)where
τ (Y ) = E[ h DY, − DL − 1 Y i H | Y ]
. Some general results for the existeneof a real-valued funtion
τ
satisfying (1.24) are ontained e.g. in [6℄. Note that, in general, itis very hard to nd an analyti expression for
τ (Y )
, espeially whenY
is a randomvariablewitha very omplex struture,suh ase.g. a multipleWiener-It integral. On
the other hand,we will seethat, in manyases, the random variable
h DY, − DL − 1 Y i H
is remarkably tratable and expliit. See the forthoming Setion 6,whih is based on
[49 , Leture VI℄, for a general disussions of equations of the type (1.24). See Remark
3.10 belowfor aonnetion withGoldsteinand Reinert'szerobias transform[20℄.
5. The reader is referredto [42 ℄ for appliations of integration by parts tehniques to the
Stein-typeestimationofdriftsofGaussianproesses. See[23℄foraSteinharaterization
ofBrownianmotionsonmanifoldsbymeansofintegrationbypartsformulae. See[13 ℄for
aonnetionbetween Stein'smethodandalgebrasofoperatorsonongurationspaes.
Before proeeding to a formal disussion, and in order to motivate the reader, we shall
provide two examples of the kind of results that we will obtain in the subsequent setions.
The rst statement involves double Wiener-It integrals, that is, random variables living in
the seondhaosof
X
. The proof isgiven inSetion 7.Theorem 1.5 Let
(Z n ) n>1
be a sequene belonging to the seond Wiener haos ofX
.1. Assume that
E(Z n 2 ) → 1
andE(Z n 4 ) → 3
asn → ∞
. ThenZ n −→ Law Z ∼ N (0, 1)
asn → ∞
. Moreover, we have:d TV (Z n , Z) 6 2 r 1
6
E(Z n 4 ) − 3
+ 3 + E(Z n 2 ) 2
E(Z n 2 ) − 1 .
2. Fix
ν > 0
andassume thatE(Z n 2 ) → 2ν
andE(Z n 4 ) − 12E(Z n 3 ) → 12ν 2 − 48ν
asn → ∞
.Then,as
n → ∞
,Z n −→ Law F(ν)
, whereF (ν)
hasa entered Gammadistributionofparameterν
. Moreover, we have:d H 2 (Z n , F (ν)) 6 max { 1, 1/ν, 2/ν 2 }
s 1 6
E(Z n 4 ) − 12E(Z n 3 ) − 12ν 2 +48ν +
8 − 6ν + E(Z n 2 ) 2
E(Z n 2 ) − 2ν ,
where
H 2
isdened by (1.15).ever, these operators will appear in the general statements presented in Setion 3). For
instane, whenappliedto theasewhere
X
istheisonormalproessgenerated byafrationalBrownianmotion,therstpointofTheorem1.5anbeusedtoderivethefollowingboundfor
theKolmogorovdistaneintheBreuer-MajorCLTassoiatedwithquadratitransformations:
Theorem 1.6 Let
B
be a frational Brownian motionwithHurst indexH ∈ (0, 3 / 4 )
. We setσ 2 H = 1
2 X
t ∈ Z
| t + 1 | 2H + | t − 1 | 2H − 2 | t | 2H 2
< ∞ ,
and
Z n = 1 σ H √ n
n − 1
X
k=0
n 2H (B (k+1)/n − B k/n ) 2 − 1
, n > 1.
Then, as
n → ∞
,Z n −→ Law Z ∼ N (0, 1)
. Moreover, there exists a onstantc H
(dependinguniquely on
H
)suh that, foranyn > 1
:d Kol (Z n , Z ) 6 c H
n 1 2 ∧ ( 3 2 − 2H) .
(1.25)Notethatboth Theorem1.5and1.6will be signiantly generalizedinSetion 3and Setion
4(see,inpartiular, theforthoming Theorems 3.1, 3.11and4.1).
Remark1.7 1. When
H = 1 / 2
,thenB
isastandardBrownian motion(and thereforehasindependent inrements), and we reover from theprevious resultthe rate
n − 1/2
,thatouldbealsoobtainedbyapplyingtheBerry-EsséenTheorem1.1. Thisrateisstillvalid
for
H < 1 / 2
. But,forH > 1 / 2
,therateweanproveintheBreuer-Major CLTisn 2H − 2 3
.2. Totheauthorsknowledge,Theorem1.6anditsgeneralizationsaretherstBerry-Esséen
boundseverestablished for the Breuer-Major CLT.
3. Tokeepthelengthofthispaperwithinlimits,wedonotderivetheexpliitexpressionof
someoftheonstants(suhasthequantity
c H
informula(1.25))omposingourbounds.As will beome lear later on, the exat value of these quantities an be dedued by a
areful bookkeeping of the bounding onstants appearing at the dierent stagesof the
proofs.
1.4 Plan
Therestofthe paperisorganizedasfollows. InSetion 2wereallsome notionsof Malliavin
alulus. In Setion3 westate anddisuss our mainboundsinStein-typeestimates forfun-
tionalsofGaussianelds. Setion4ontainsanappliationtotheBreuer-MajorCLT. Setion
5 deals withGamma-type approximations. Setion 6 providesa unieddisussionof approx-
imations bymeans ofabsolutelyontinuous distributions. Proofsandfurther renementsare
olleted inSetion 7.
The reader isreferred to [25℄ or [35℄ for any unexplainednotion disussed inthis setion. As
in (1.22), we denote by
X = { X(h) : h ∈ H }
an isonormal Gaussian proess overH
. Bydenition,
X
is a entered Gaussian family indexed by the elements ofH
and suh that, forevery
h, g ∈ H
,E
X(h)X(g)
= h h, g i H .
(2.26)As before, we use the notation
L 2 (X) = L 2 (Ω, σ(X), P )
. It is well-known (see again [35 ,Ch. 1℄or [25℄)thatanyrandom variable
F
belonging toL 2 (X)
admits thefollowing haotiexpansion:
F = X ∞
q=0
I q (f q ),
(2.27)where
I 0 (f 0 ) := E[F]
,theseriesonvergesinL 2
andthesymmetrikernelsf q ∈ H ⊙ q
,q > 1
,areuniquelydeterminedby
F
. Asalreadydisussed,inthepartiularasewhereH = L 2 (A, A , µ)
,where
(A, A )
isa measurablespaeandµ
isaσ
-nite andnon-atomimeasure, one hasthatH ⊙ q = L 2 s (A q , A ⊗ q , µ ⊗ q )
is the spae of symmetri and square integrable funtions onA q
.Moreover, for every
f ∈ H ⊙ q
, the random variableI q (f )
oinides withthemultiple Wiener-It integral (of order
q
) off
with respet toX
(see [35, Ch. 1℄). Observe that a randomvariable of the type
I q (f )
, withf ∈ H ⊙ q
, has nite moments of all orders (see e.g. [25, Ch.VI℄). See again [35 , Ch. 1℄or [45 ℄ for a onnetionbetween multiple Wiener-Itand Hermite
polynomials. For every
q > 0
, we writeJ q
to indiate the orthogonalprojetion operator onthe
q
th Wiener haos assoiated withX
, so that, ifF ∈ L 2 (Ω, F , P )
is as in (2.27), thenJ q F = I q (f q )
for everyq > 0
.Let
{ e k , k ≥ 1 }
be aomplete orthonormal systeminH
. Givenf ∈ H ⊙ p
andg ∈ H ⊙ q
,forevery
r = 0, . . . , p ∧ q
,ther
thontration off
andg
is theelement ofH ⊗ (p+q − 2r)
dened asf ⊗ r g =
X ∞
i 1 ,...,i r =1
h f, e i 1 ⊗ . . . ⊗ e i r i H ⊗r ⊗ h g, e i 1 ⊗ . . . ⊗ e i r i H ⊗r .
(2.28)Notethat, inthe partiular asewhere
H = L 2 (A, A , µ)
(withµ
non-atomi), one hasthatf ⊗ r g = Z
A r
f (t 1 , . . . , t p − r , s 1 , . . . , s r ) g(t p − r+1 , . . . , t p+q − 2r , s 1 , . . . , s r )dµ(s 1 ) . . . dµ(s r ).
Moreover,
f ⊗ 0 g = f ⊗ g
equals the tensor produt off
andg
while, forp = q
,f ⊗ p g = h f, g i H ⊗p
. Notethat, ingeneral (and exeptfor trivial ases), theontrationf ⊗ r g
is not asymmetrielement of
H ⊗ (p+q − 2r)
. The anonial symmetrization off ⊗ r g
is writtenf ⊗ e r g
.We alsohave the usefulmultipliation formula: if
f ∈ H ⊙ p
andg ∈ H ⊙ q
,thenI p (f )I q (g) =
p ∧ q
X
r=0
r!
p r
q r
I p+q − 2r (f ⊗ e r g).
(2.29)Let
S
be the setof all smooth ylindrial random variables of theformF = g X(φ 1 ), . . . , X(φ n )
where
n > 1
,g : R n → R
is a smooth funtion with ompat support andφ i ∈ H
. TheMalliavinderivative of
F
withrespettoX
isthe element ofL 2 (Ω, H )
dened asDF =
X n
i=1
∂g
∂x i X(φ 1 ), . . . , X(φ n ) φ i .
In partiular,
DX(h) = h
for everyh ∈ H
. By iteration, one an dene them
th derivativeD m F
(whihisanelementofL 2 (Ω, H ⊗ m )
)foreverym > 2
. Asusual,form > 1
,D m,2
denotes the losureofS
withrespetto the normk · k m,2
,dened bytherelationk F k 2 m,2 = E F 2
+ X m
i=1
E
k D i F k 2 H ⊗i
.
Note that, if
F 6 = 0
andF
is equal to a nite sum of multiple Wiener-It integrals, thenF ∈ D m,2
for everym > 1
and the law ofF
admits a density with respet to the Lebesguemeasure. TheMalliavin derivative
D
satisesthe followinghainrule: ifϕ : R n → R
isinC b 1
(thatis,theolletionofontinuouslydierentiablefuntionswitha boundedderivative) and
if
{ F i } i=1,...,n
isavetor of elementsofD 1,2
,thenϕ(F 1 , . . . , F n ) ∈ D 1,2
andDϕ(F 1 , . . . , F n ) = X n
i=1
∂ϕ
∂x i (F 1 , . . . , F n )DF i .
Observe that the previous formula still holds when
ϕ
is a Lipshitz funtion and the law of(F 1 , . . . , F n )
hasa densitywithrespetto theLebesgue measureonR n
(see e.g. Proposition 1.2.3 in[35 ℄).We denote by
δ
the adjoint of the operatorD
, also alled the divergene operator. Arandom element
u ∈ L 2 (Ω, H )
belongs to the domain ofδ
, notedDomδ
, if, and only if, itsatises
| E h DF, u i H | 6 c u k F k L 2
for anyF ∈ S ,
where
c u
isaonstant depending uniquelyonu
. Ifu ∈ Domδ
,thentherandomvariableδ(u)
isdened by the duality relationship(ustomarily alledintegration bypartsformula):
E(F δ(u)) = E h DF, u i H ,
(2.30)whih holds for every
F ∈ D 1,2
. One sometimes needs the following property: for everyF ∈ D 1,2
and everyu ∈ Domδ
suh thatF u
andF δ(u) + h DF, u i H
aresquareintegrable, one hasthatF u ∈ Domδ
andδ(F u) = F δ(u) − h DF, u i H .
(2.31)Theoperator
L
,atingonsquareintegrablerandomvariablesofthetype(2.27),isdenedthrough theprojetion operators
{ J q } q > 0
asL = P ∞
q=0 − qJ q ,
and is alled the innitesimal generator of the Ornstein-Uhlenbek semigroup. It veries the following ruial property: arandom variable
F
isanelementofDomL (= D 2,2 )
if,andonlyif,F ∈ DomδD
(i.e.F ∈ D 1,2
and
DF ∈ Domδ
),and inthisase:δDF = − LF.
NotethatarandomvariableF
asin(2.27)is in
D 1,2
(resp.D 2,2
) if,and only if,X ∞
q=1
q k f q k 2 H ⊙q < ∞ (
resp.X ∞
q=1
q 2 k f q k 2 H ⊙q < ∞ ),
and also
E
k DF k 2 H
= P
q > 1 q k f q k 2 H ⊙q
. IfH = L 2 (A, A , µ)
(withµ
non-atomi), then the derivativeofarandomvariableF
asin(2.27) anbeidentiedwiththeelementofL 2 (A × Ω)
given by
D a F = X ∞
q=1
qI q − 1 f q ( · , a)
, a ∈ A.
(2.32)We also dene the operator
L − 1
,whih is the inverse ofL
,as follows: for everyF ∈ L 2 (X)
withzero mean, we set
L − 1 F = P
q>1 − 1 q J q (F)
. Note thatL − 1
is an operator with valuesin
D 2,2
. The following Lemma ontains two statements: therst one (formula (2.33)) is an immediateonsequeneofthedenitionofL
andoftherelationδD = − L
,whereastheseond(formula (2.34))orrespondsto Lemma 2.1in[33℄.
Lemma 2.1 Fix an integer
q > 2
and setF = I q (f )
, withf ∈ H ⊙ q
. Then,δDF = qF.
(2.33)Moreover, for every integer
s > 0
,E F s k DF k 2 H
= q
s + 1 E F s+2
.
(2.34)3 Stein's method and integration by parts on Wiener spae
3.1 Gaussian approximations
Our rst result provides expliit bounds for the normal approximation of random variables
thatareMalliavinderivable. Althoughitsproofisquiteeasytoobtain,thefollowingstatement
will beentral for the rest ofthe paper.
Theorem 3.1 Let
Z ∼ N (0, 1)
,andletF ∈ D 1,2
be suhthatE(F ) = 0
. Then,thefollowingbounds are inorder:
d W (F, Z) 6 E[(1 − h DF, − DL − 1 F i H ) 2 ] 1/2 ,
(3.35)d FM (F, Z) 6 4E[(1 − h DF, − DL − 1 F i H ) 2 ] 1/2 .
(3.36)If,inaddition,the lawof
F
isabsolutelyontinuouswithrespettotheLebesgue measure, onehasthat
d Kol (F, Z) 6 E[(1 − h DF, − DL − 1 F i H ) 2 ] 1/2 ,
(3.37)d TV (F, Z) 6 2E[(1 − h DF, − DL − 1 F i H ) 2 ] 1/2 .
(3.38)Proof. Start by observing that one an write
F = LL − 1 F = − δDL − 1 F
. Now letf
bea real dierentiable funtion. By using the integration by parts formula and the fat that
Df(F ) = f ′ (F )DF
(notethat, forthis formulato holdwhenf
isonlya.e. dierentiable, one needsF
to have anabsolutelyontinuous law, seeProposition 1.2.3 in[35 ℄),we dedueE(F f (F )) = E[f ′ (F) h DF, − DL − 1 F i H ].
Itfollowsthat
E[f ′ (F ) − F f (F )] = E(f ′ (F )(1 − h DF, − DL − 1 F i H ))
sothatrelations (3.35)(3.38) an bededued from(1.8)(1.11) andthe Cauhy-Shwarz inequality.
2
We shall now prove that the bounds appearing in the statement of Theorem 3.1 an be
expliitly omputed, whenever
F
belongs to axed Wienerhaos.Proposition 3.2 Let
q > 2
be an integer, and letF = I q (f )
, wheref ∈ H ⊙ q
. Then,h DF, − DL − 1 F i H = q − 1 k DF k 2 H
, andE[(1 − h DF, − DL − 1 F i H ) 2 ] = E[(1 − q − 1 k DF k 2 H ) 2 ]
(3.39)= (1 − q! k f k 2 H ⊗q ) 2 + q 2
q − 1
X
r=1
(2q − 2r)!(r − 1)! 2 q − 1
r − 1 4
k f ⊗ e r f k 2 H ⊗2(q−r)
(3.40)6 (1 − q! k f k 2 H ⊗q ) 2 + q 2
q − 1
X
r=1
(2q − 2r)!(r − 1)! 2 q − 1
r − 1 4
k f ⊗ r f k 2 H ⊗2(q−r) .
(3.41)Proof. The equality
h DF, − DL − 1 F i H = q − 1 k DF k 2 H
is an immediate onsequene of therelation
L − 1 I q (f ) = − q − 1 I q (f )
. Fromthemultipliationformulaebetweenmultiplestohasti integrals, see(2.29),one dedues thatk D[I q (f)] k 2 H = qq! k f k 2 H ⊗q + q 2
q − 1
X
r=1
(r − 1)!
q − 1 r − 1
2
I 2(q − r) f ⊗ e r f
(3.42)
(seealso[36,Lemma 2℄). Wethereforeobtain (3.40)byusingtheorthogonalityandisometri
properties of multiple stohasti integrals. The inequality in (3.41) is just a onsequene of
the relation
k f ⊗ e r f k H ⊗2(q−r) 6 k f ⊗ r f k H ⊗2(q−r)
.2
The previous result should be ompared with the forthoming Theorem 3.3, where we
ollet the main ndings of [36 ℄ and [37℄. In partiular, the ombination of Proposition 3.2
and Theorem 3.3 shows that, for every (normalized) sequene
{ F n : n > 1 }
living in a xedWiener haos, the bounds given in (3.35)(3.36) are tight with respet to theonvergene
in distributiontowards
Z ∼ N (0, 1)
,inthe sense thatthese boundsonvergeto zeroif, andonly if,
F n
onverges indistribution toZ
.Theorem 3.3 ([36, 37℄) Fix
q > 2
, and onsider a sequene{ F n : n > 1 }
suh thatF n = I q (f n )
,n > 1
, wheref n ∈ H ⊙ q
. Assume moreover thatE[F n 2 ] = q! k f n k 2 H ⊗q → 1
. Then, thefollowingfour onditions are equivalent, as
n → ∞
:(i)
F n
onverges in distributiontoZ ∼ N (0, 1)
;(ii)
E[F n 4 ] → 3
;(iii) for every
r = 1, ..., q − 1
,k f n ⊗ r f n k H ⊗2(q−r) → 0
;(iv)
k DF n k 2 H → q
inL 2
.The impliations (i)
↔
(ii)↔
(iii) have been rst proved in[
37]
by means of stohastialulus tehniques. The fat that (iv) is equivalent to either one of onditions (i)(iii) is
proved in[36℄. Note thatTheorem 3.1 and Proposition 3.2above provide an alternate proof
of the impliations (iii)
→
(iv)→
(i). The impliation (ii)→
(i) an be seen as a drastisimpliation of the method of moments and umulants, that is a ustomary toolin order
to prove limit theorems for funtionals of Gaussian elds (see e.g. [5, 7,18, 27 , 50 ℄). In [39 ℄
onean nd amultidimensional versionofTheorem 3.3.
works,suh as:
p
-variationsof stohastiintegrals withrespetto Gaussian proesses[2, 12 ℄, quadrati funtionalsofbivariateGaussian proesses[14℄,self-intersetionloaltimesoffra-tionalBrownianmotion[24℄,approximationshemesforsalarfrationaldierentialequations
[32 ℄,high-frequenyCLTsforrandomeldsonhomogeneousspaes[29,30,38 ℄,needletsanal-
ysis on the sphere [1 ℄, estimation of self-similarity orders [55℄, power variations of iterated
Brownian motions [34℄. We expetthat the new bounds proved in Theorem 3.1 and Propo-
sition 3.2 will lead to further renements of these results. See Setion 4 and Setion 5 for
appliationsand examples.
As shown in the following statement, the ombination of Proposition 3.2 and Theorem
3.3implies that,onanyxedWienerhaos,theKolmogorov,total variationandWasserstein
distanes metrize theonvergene indistribution towards Gaussian random variables. Other
topologial haraterizations of the set of laws of random variables belonging to a xed sum
of Wiener haoses aredisussed in[25 , Ch. VI℄.
Corollary 3.5 Let the assumptions and notation of Theorem 3.3 prevail. Then, the fat
that
F n
onverges in distribution toZ ∼ N (0, 1)
is equivalent to either one of the followingonditions:
(a)
d Kol (F n , Z) → 0
;(b)
d TV (F n , Z) → 0
;()
d W (F n , Z ) → 0
.Proof. If
F n Law → Z
then,byTheorem 3.3, we have neessarilythatk DF n k 2 H → q
inL 2
. Thedesired onlusionfollows immediately fromrelations (3.35), (3.37),(3.38) and (3.39).
2
Notethatthe previous resultisnot trivial, sine thetopologiesindued by
d Kol
,d TV
andd W
are strongerthan onvergene indistribution.Remark3.6 (Mehler's formula and Stein's method, I). In [9, Lemma 5.3℄, Chatterjee has
proved the following result (we use a notation whih is slightly dierent from the original
statement). Let
Y = g(V )
, whereV = (V 1 , ..., V n )
is a vetor of entered i.i.d. standardGaussian random variables, and
g : R n → R
is a smooth funtion suh that: (i)g
and itsderivatives have subexponential growth at innity, (ii)
E(g(V )) = 0
,and (iii)E(g(V ) 2 ) = 1
.Then,for anyLipshitz funtion
f
,one hasthatE[Y f (Y )] = E[S(V )f ′ (Y )],
(3.43)where,for every
v = (v 1 , ..., v n ) ∈ R n
,S(v) = Z 1
0
1 2 √
t E
" n X
i=1
∂g
∂v i (v) ∂g
∂v i ( √ tv + √
1 − tV )
#
dt,
(3.44)sothat, for instane, for
Z ∼ N (0, 1)
andbyusing(1.9),Lemma 1.2(iii),(1.7) and Cauhy-Shwarzinequality,
d TV (Y, Z) 6 2E[(S(V ) − 1) 2 ] 1/2 .
(3.45)of generality, we an assume that
V i = X(h i )
, whereX
is an isonormal proess over someHilbert spae of the type
H = L 2 (A, A , µ)
and{ h 1 , ..., h n }
is an orthonormal system inH
. SineY = g(V 1 , . . . , V n )
, we haveD a Y = P n i=1
∂g
∂x i (V )h i (a)
. On the other hand, sineY
is entered and square integrable, it admits a haoti representation of the formY = P
q>1 I q (ψ q )
. ThisimpliesinpartiularthatD a Y = P ∞
q=1 qI q − 1 (ψ q (a, · ))
. Moreover, onehasthat
− L − 1 Y = P
q>1 1
q I q (ψ q )
, so that− D a L − 1 Y = P
q>1 I q − 1 (ψ q (a, · ))
. Now, letT z
,z > 0
,denote the (innite dimensional) Ornstein-Uhlenbek semigroup, whose ation on random
variables
F ∈ L 2 (X)
isgiven byT z (F) = P
q>0 e − qz J q (F )
. Wean writeZ 1
0
1 2 √
t T ln(1/ √ t) (D a Y )dt = Z ∞
0
e − z T z (D a Y )dz = X
q> 1
1
q J q − 1 (D a Y )
= X
q>1
I q − 1 (ψ q (a, · )) = − D a L − 1 Y .
(3.46)NowreallthatMehler'sformula(seee.g. [35 ,formula(1.54)℄)impliesthat,foreveryfuntion
f
withsubexponential growth,T z (f(V )) = E
f (e − z v + p
1 − e − 2z V ) v=V , z > 0.
In partiular, by applying this last relation to the partial derivatives
∂g
∂v i
,
i = 1, ..., n
, wededue from(3.46) that
Z 1
0
1 2 √
t T ln(1/ √ t) (D a Y )dt = X n
i=1
h i (a) Z 1
0
1 2 √
t E ∂g
∂v i ( √ t v + √
1 − t V ) dt
v=V .
Consequently, (3.43) follows,sine
h DY, − DL − 1 Y i H =
* n X
i=1
∂g
∂v i
(V )h i , X n
i=1
Z 1
0
1 2 √
t E ∂g
∂v i
( √
t v + √
1 − t V ) dt
v=V h i +
H
= S(V ).
2
See also Houdré and Pérez-Abreu [22℄ for related omputations in an innite-dimensional
setting.
The following resultonerns nite sumsof multiple integrals.
Proposition 3.7 For
s > 2
, xs
integers2 6 q 1 < . . . < q s
. Consider a sequene of theform
Z n = X s
i=1
I q i (f n i ), n > 1,
where