Large deviation principle for enhanced Gaussian processes
Peter Friz
∗, Nicolas Victoir
Department of Pure Mathematics and Mathematical Statistics, University of Cambridge, Cambridge, UK Received 10 January 2006; received in revised form 2 November 2006; accepted 22 November 2006
Available online 27 February 2007
Abstract
We study large deviation principles for Gaussian processes lifted to the free nilpotent group of stepN. We apply this to a large class of Gaussian processes lifted to geometric rough paths. A large deviation principle for enhanced (fractional) Brownian motion, in Hölder- or modulus topology, appears as special case.
©2007 Elsevier Masson SAS. All rights reserved.
Résumé
Nous etudions les principes de grandes deviations pour les « rough paths Gaussiens », pour une large classe de processus Gaus- siens. Cette classe contient le mouvement brownien fractionnaire.
©2007 Elsevier Masson SAS. All rights reserved.
MSC:60F10; 60G15; 60H99
Keywords:Rough paths; Large deviation principle; Gaussian processes; Fractional Brownian motion
1. Introduction
We prove large deviation principles for d-dimensional Gaussian processes lifted to the free nilpotent group of stepN. The example we have in mind is a class of zero mean Gaussian processes subject to certain conditions on the covariance, [5,17].
After recalls on nilpotent groups and Wiener Itô chaos, we give simple conditions under which a large devia- tion principle holds. To illustrate the method, we quickly check these conditions for enhanced Brownian motion (see [15,10] for earlier approaches). We then apply our methodology to the class of enhanced Gaussian processes con- sidered in [5,17]. Enhanced fractional Brownian motion appears as a special case and we strengthen the results in [18,4].
* Corresponding author.
E-mail address:P.K.Friz@statslab.cam.ac.uk (P. Friz).
0246-0203/$ – see front matter ©2007 Elsevier Masson SAS. All rights reserved.
doi:10.1016/j.anihpb.2006.11.002
1.1. The free nilpotent group and rough paths
LetN0. The truncated tensor algebra of degreeN is given by the direct sum TN
Rd
=R⊕Rd⊕ · · · ⊕ Rd⊗N
.
With the usual scalar product, vector addition, and tensor product⊗the spaceTN(Rd)is an algebra. Letπi denote the canonical projection fromTN(Rd)onto(Rd)⊗i.
Letq∈ [1,2)andx∈C0q-var([0,1],Rd), the space of continuousRd-valued paths of boundedq-variation started at the origin. The liftSN(x):[0,1] →TN(Rd)is defined as
SN(x)t=1+ N
i=1
0<s1<···<si<t
dxs1⊗ · · · ⊗dxsi,
the iterated integrals are Young integrals. Then GN
Rd
=
g∈TN Rd
: ∃x∈C1-var 0
[0,1],Rd
: g=SN(x)1
is a submanifold ofTN(Rd)and, in fact, a Lie group with product⊗, called the free nilpotent group of stepN. The dilation operatorδ:R×GN(Rd)→GN(Rd)is defined by
πi δλ(g)
=λiπi(g), i=0, . . . , N,
and a continuous norm onGN(Rd), homogeneous with respect toδ, is given g =inf
length(x): x∈C1-var
0
[0,1],Rd
, SN(x)1=g . By equivalence of such norms there exists a constantKNsuch that
1 KN
i=max1,...,N
πi(g)1/ igKN max
i=1,...,N
πi(g)1/ i. (1.1)
The norm · induces a metric onGN(Rd)known as Carnot–Caratheodory metric, d(g, h)= g−1⊗h .
Letx, y∈C0([0,1], GN(Rd)), the space of continuousGN(Rd)-valued paths started at the neutral element of (GN(Rd),⊗). We define a supremum distance,
d∞(x, y)= sup
t∈[0,1]d(xt, yt), a Hölder distance,α∈ [0,1],
dα-Hölder(x, y)= sup
0s<t1
d(xs−1⊗xt, ys−1⊗yt)
|t−s|α , write
xα-Hölder= sup
0s<t1
xs−1⊗xt
|t−s|α = sup
0s<t1
d(xs, xt)
|t−s|α ,
andLipinstead of 1-Hölder. Similarly, one definesp-variation regularity and distance forGN(Rd)-valued paths. We refer to [11] for a more detailed discussion of these topics.
The interest for all this comes from T. Lyons’ rough path theory, [16,17], a deterministic theory of control differ- ential equations (“Rough Differential Equation”), driven by a continuousG[p](Rd)-valued path of finitep-variation.
Brownian motion and Lévy’s area can be viewed as aG[p](Rd)-valued path of finitep-variation withp >2 and the corresponding RDE solution turns out to be a classical Stratonovich SDE solution. This gives a general recipe how to make sense of differential equations driven by anarbitrarystochastic processXwith continuous sample path of finite p-variation: find an appropriate lift ofXto aG[p](Rd)-valued pathXwith finitep-variation use Lyons’ machinery.
1.2. Gaussian process
The Wiener spaceC0([0,1],Rd)is the space of continuous function started at 0; the sup norm induces a topology and hence a Borelσ-algebra. We assume there is a probability measurePsuch that the coordinate processX, defined by X(ω)t =ω(t ), is a centered Gaussian process with covariancec(s, t ). We also defineHto be the reproducing kernel Hilbert space associated toP, i.e. the closure of the set of the linear span of the functionss→c(ti, s), under the Hilbert product
c(ti,·), c(tj,·)
=c(ti, tj).
Letidenote the canonical injection from the reproducing kernel Hilbert space to the Wiener space. We defineCn(B) the B-valued homogeneous Wiener chaos of degreen, where(B,| · |B)is a given Banach space. We also define Cn(B)the sum of the firstnhomogeneous Wiener chaos. We refer to [14,1–3] for more details on Gaussian spaces and Wiener chaos.
1.3. Enhanced Gaussian process
We make the assumption that there exists a lift of the Gaussian processXto a processXwith values inGN(Rd) for some fixedN 1. Such lifts have been constructed first by Coutin and Qian [5] as a.s. limits inp-variation of (canonically lifted) dyadic piecewise linear approximations ofX. We shall be less specific here and extract those properties of the approximations that we need in the sequel. We make the following
Assumption.There exists a sequence of continuous linear maps{Φm}fromC([0,1],Rd)(with uniform topology) onto the space of bounded variation path from[0,1]intoRd(with 1-variation topology) such that
(1) limm→∞Φm(x)=xfor all continuous pathsxwith respect to uniform topology;
(2) the uniform distance betweenSN◦Φm(X)andXconverges to 0 in probability.
In fact, Condition (2) is equivalent to the seemingly stronger condition ofLp-convergence.
Lemma 1.Convergence of d∞(SN◦Φm(X),X)to zero in probability is equivalent to convergence inLp for all p∈ [1,∞).
Proof. Theorem III.2 in [20] and Eq. (1.1). 2
Remark 1.Assumption 2 is always satisfied whenN=1. In the applications discussed in later sections,N will be related to the regularity ofX.
Example 1.Let{Φm}be the piecewise linear approximations based on a sequence of dissections{Dm}with mesh
|Dm| →0. IfXdenotes fractional Brownian motion of parameterH >14, then the liftXforN= [1/H]and was first constructed in [5], see also [4].
Remark 2.As is well known (e.g. Theorem 1 in [13]) any continuous Gaussian processX on[0,1]is the uniform limit of
X(m)t ≡ m
j=1
ξj(ω)ψj(t )
where{ψj}is an orthonormal basis for the reproducing Kernel Hilbert spaceHand the{ξj}are i.i.d. standard Gaussian given by the image ofψj under the isometric isomorphismθ:H→L2(Xt: t∈ [0,1]), the closure of{Xt: [0,1]}in L2(P). The Hilbert structure ofHimplies that for allh∈H,
h(m)≡ m
j=1
ψj, hψj→h inH
and hence uniformly. The construction ofXcan be based on such (or similar) approximations,1see [6,11,19,8], but they do not satisfy our assumption. It may be possible to adapt the subsequent proofs, aimed to establish a large deviation principle forX, to such approximations. As this requires further work without improving the results we shall not pursue this further here.
2. Wiener chaos andGN(Rd)-valued paths
The hypercontractivity of the Ornstein–Uhlenbeck semigroup leads to a useful Lemma 2.LetZnbe a random variable inCn(B). Assume1< p < q <∞. Then
ZnLq(P;B) q−1
p−1 n/2
ZnLp(P;B).
Proof. LetPt denote the Ornstein–Uhlenbeck semigroup. Whenever 1< p < q <∞andt >0 satisfies et
q−1 p−1
1/2
, (2.1)
then, for allf ∈Lp(P;B), PtfLq fLp. We also recall that
PtZn=e−ntZn.
See [14] for this two results. We now chooset such that equality holds in (2.1) and find q−1
p−1 −n/2
ZnLq =e−ntZnLq = PtZnLq ZnLp. 2
It is well known thatLp(P;B)- andLq(P;B)-norms are equivalent on Cn(B)=C1(B)+ · · · +Cn(B).
The next lemma quantifies this equivalence.
Lemma 3.Letn∈NandZbe a random variable inCn(B). Assume2pq <∞. Then there exists a constant Mnsuch that
ZLp(P;B)ZLq(P;B)Mn(q−1)n/2ZLp(P;B). Proof. Only the second inequality requires a proof. WriteZ=n
i=0Zi withZi∈Ci. From Lemma 2, for allin, ZiLq(P;B)
q−1 p−1
i/2
ZiLp(P;B). (2.2)
The sumCn(B)=C1(B)+ · · · +Cn(B)is topological direct inL0(P;B), see [3, p. 6] for instance, which implies that the projectionZ→Zi is continuous inL2(P;B). It follows thatZiL2(P;B)cZL2(P;B)for some constant c=c(n). Then, forp2,
ZiLp(P;B)(p−1)i/2ZiL2(P;B)c(p−1)i/2ZL2(P;B)c(p−1)i/2ZLp(P;B). (2.3) By Hölder’s inequality for finite sums,
|Z|qBnq−1 n
i=0
|Zi|qB,
1 At least for,N2 it is shown in [6] that reproducing kernel – and piecewise linear approximations yield the same lifted process.
and after taking expectations, ZLq(P;B)n1−1/q
n
i=0
ZiqLq(P;B) 1/q
. Hence, from (2.2) and (2.3),
ZLq(P;B)cn1−1/q n
i=0
q−1 p−1
qi/2
ZiqLp(P;B) 1/q
cn1−1/q n
i=0
(q−1)qn/2ZqLp(P;B)
1/q
cn(q−1)n/2ZLp(P;B). 2
Corollary 1.LetX, Y be two continuousGN(Rd)-valued processes such that, for eachi=1, . . . , N, the projection t→πi(Xt), πi(Yt)∈
Rd⊗i
belongs toCi(Bi)=C0(Bi)⊕C1(Bi)⊕ · · · ⊕Cn(Bi), where Bi=C0
[0,1], Rd⊗i
andCj(Bi)is theBi-valued homogeneous Wiener chaos of degreej. Assume2< q <∞. Then there exists a constant MNsuch that
d∞(X, Y )
L2N(P) d∞(X, Y )
LqN(P)√
q MN d∞(X, Y )
L2N(P). Proof. Again, only the second inequality requires a proof. Fori=1, . . . , Ndefine
Zi:t→πi
Xt−1⊗Yt
∈ Rd⊗i
.
Observe thatZi∈Ci(Bi)⊂CN(Bi).Let| · |∞denote the supremum norm on(Rd)⊗i-valued paths. From Eq. (1.1), 1
KN
i=max1,...,N|Zi|1/ i∞ d∞(X, Y )KN max
i=1,...,N|Zi|1/ i∞. (2.4)
Therefore, for allq2, 1
KN
i=max1,...,NZi1/ iLqN/ i(P) d∞(X, Y )
LqN(P)KN max
i=1,...,NZi1/ iLqN/ i(P). (2.5)
From Lemma 3
Zi1/ iLqN/ i(P)MN1/ i
qN/ i−1Zi1/ iL2N/ i(P)
√
qMN Zi1/ iL2N/ i(P)
√
qMN d∞(X, Y )
L2N(P)
where we used (2.5) withq=2 in the last line. Another look at (2.5) finishes the proof. 2 3. Large deviation results for(δεX)ε>0
From general principles,(εX)ε>0satisfies a large deviation principle with good rate functionIin uniform topology, whereI is given by (see [14] for example):
I (y)= 1
2x2H wheny=i(x)for somex∈H, +∞ otherwise.
It is clear that SN◦Φm:
C
[0,1],Rd ,| · |∞
→C
[0,1], GN Rd
, d∞
is continuous. By the contraction principle [7],SN(εΦm(X))satisfies a large deviation principle with good rate func- tion
Jm(y)=inf
I (x), xsuch thatSN Φm(x)
=y ,
the infimum of the empty set being+∞. Essentially, a large deviation principle forδεXis obtained by sendingmto infinity. To this end we need
Lemma 4.Letδ >0fixed. Then
mlim→∞lim
ε→0ε2logP d∞
SN
Φm(εX) , δεX
> δ
= −∞.
Proof. First observe that d∞
SN
Φm(εX) , δεX
=εd∞ SN
Φm(X) ,X
. By standing assumption onXand Lemma 1 we have
αm:= d∞ SN
Φm(X) ,X
L2N −→
m→∞0.
Then, by Corollary 1 , d∞
SN Φm(X)
,X
LqMN√
q αm ∀q2N. (3.1)
We then estimate P
d∞ SN
Φm(εX) , δεX
> δ
=P
d∞ SN
Φm(X) ,X
>δ ε
δ
ε −q√
qqαqm
exp
qlog
ε δαm√
q
, and after choosingq=1/ε2we obtain, forεsmall enough,
ε2logP
t∈D
d SN
Φm(εX)
t, δεXt
> δ
log αm
δ
.
Now take the limits limε→0and limm→∞to finish the proof. 2
The following theorem is a straight forward application of the extended contraction principle [7, Theorem 4.2.23]
and Lemma 4. The point is that althoughSN is a only a measurable map, defined as a.s. limit, it has exponentially good approximations given by{SN◦Φm: m1}.
Theorem 1.Assume thatSN(h),defined as the pointwise limit of(SN◦Φm)(h).whenm→ ∞exists forhsuch that I (h) <∞(i.e. allhin the Cameron–Martin space), and that for allΛ >0,
mlim→∞ sup
{h:I (h)Λ}d∞
(SN◦Φm)(h), SN(h)
=0. (3.2)
Then the family(δεX)ε>0satisfies a large deviation principle in uniform topology with good rate function J (y)=inf
I (x), xsuch thatSN(x)=y .
The following corollary allows us to strengthen the topology.
Corollary 2.Under the above assumptions and the additional condition
∃α>0, c >0:E
expcX2α-Hölder
<∞,
the family (δεX)ε>0 satisfies a large deviation principle in α-Hölder topology for all α∈ [0, α)with good rate functionJ (y).
Proof. By the inverse contraction principle [7, Theorem 4.2.4.] it suffices to show that the laws of(δεX)are expo- nentially tight inα-Hölder topology. But this follows from the compact embedding of
Cα-Hölder
[0,1], GN Rd
⊂⊂Cα-Hölder
[0,1], GN Rd and Gauss tails ofXα-Hölder. Indeed, for everyM >0,
KM=
x: xα-HölderM
defines a precompact set w.r.t.α-Hölder topology. Then ε2log
(δεX)∗P KMc
=ε2logP
δεXα-Hölder > M
=ε2logP
Xα-Hölder> M/ε
−M2. 2
Remark 3.The same proof applies to the (weaker)p-variation topology and the (stronger) modulus topology consid- ered in [10].
We emphasise the generality of this approach: a large deviation result for an enhanced Gaussian process holds provided (a) the enhanced Gaussian process exists and (b) a condition on the reproducing Kernel of the Gaussian process.
As a warm-up we show how to apply this to enhanced Brownian motion. We will then discuss a wider class of Gaussian process, which contains fractional Brownian motion with Hurst parameterH >14.
4. LDP for enhanced Brownian motion
LetX=Bis bed-dimensional Brownian motion. We takeΦmbe the piecewise linear approximation over at the dyadic dissection of[0,1]with mesh 2−m. Then S2(Φm(B))converges uniformly on[0,1]and in probability to a processB, see [17,9,10]). Large deviations for(δεB)εwere first obtained inp-variation topology [15], then in Hölder and modulus topology [10]. The following proof is considerably quicker. There are only two things to check.
(1) Gauss tails ofBα-Hölder for anyα∈ [0,1/2). There are many different ways to see this. In [10] we use the scalingd(Bs,Bt)= |D t−s|1/2B0,1, noting thatB0,1has Gauss tails, followed by an application of the Garsia–
Rodemich–Rumsey lemma. One could prove the result directly by extending Lemma 1 toα-Hölder metric.
(2) Condition (3.2). But this is a simple consequence of the fact that the 1-variation norm ofhis bounded by√ 2I (h) and that the 1-variation ofΦm(h)is bounded by the 1-variation ofh. From Theorem 1 and Corollary 2 we deduce Theorem 2.LetB=B(ω)∈C([0,1], G2(Rd))denote enhanced Brownian Motion. Letα∈ [0,1/2). Then(δεB)ε>0
satisfies a large deviation principle inα-Hölder topology with good rate function J (y)=inf
I (x), xsuch thatSN(x)=y .
When restrictingαtoα∈(1/3,1/2)the universal limit theorem gives the Freidlin–Wentzell theorem (inα-Hölder topology) as corollary. See [15,10,4] for details on this by now classical application of Lyons’ limit theorem.
Remark 4.The case of enhanced fractional Brownian motion,H > 14, is similar. First, Gauss tails of theα-Hölder norm, for anyα∈ [0, H ), follow from scaling just as above. (This has already been used in [11,4].) Secondly, condition (3.2) follows from the fact that the Cameron–Martin space for fBM is embedded in the space of finiteq-variation for
anyq less than 1/(1/2+H ), see [12]. In any case, enhanced fractional Brownian motion is a special case of the processes considered in the next section.
5. LDP for a class of Gaussian processes
We fix a parameterH >0. We now specialise to zero-meanRd-valued Gaussian processes{Xt: t∈ [0,1]} with independent componentsx=Xi, i=1, . . . , d, satisfying
E
|xs,t|2
c|t−s|2H, for alls < t, (5.1)
E(xs,s+hxt,t+h)c|t−s|2H h
t−s
2, for alls, t, hwithh < t−s (5.2)
for some constantc >0. Note that such a process has a.s. 1/p-Hölder continuous sample paths when p >1/H.
A particular example of such a Gaussian process is fractional Brownian motion of Hurst parameterH. This condition is the one which allowed Coutin and Qian [5] to construct the lift ofX to a rough path. As beforeHdenotes the Cameron–Martin space. We takeΦmto be the piecewise linear approximation over at the dyadic dissections of[0,1]
with mesh 2−m.
Definition 1.ForH >1/4,we define, according to [5], the enhanced Gaussian processXto be the (a.s. and inLqfor allq) limit in 1/p-Hölder norm ofS[p](Φm(X)), where 1/p < H.2
To obtain the large deviation principle for(δεX)ε>0there are, as in the case of EBM, only two things to check.
(1) Gauss tails ofX1/p-Hölder forp >1/H,that is Eexp
αX21/p-Hölder
<∞. (5.3)
First note that for allq1, sup
0s<t1
E
Xs,t
|t−s|H q
<∞. Similar to Corollary 1, for alls, t,andk1,
E Xs,t
|t−s|H
2k1/2k
C√
k E Xs,t
|t−s|H q1/q
,
whereq is any fixed real greater than 2[1/H]. Now expand the exponential in a power series to see that there existsα >0 s.t.Eexp(αXs,t2/|t−s|2H) <∞for alls, t. The proof of 5.3 is then finished using the Garsia–
Rodemich–Rumsey lemma, just as in [10].
(2) Condition (3.2). We show (a) that for some q ∈ [1,2) the q-variation norm of a Cameron–Martin path is bounded by a constant times its Cameron–Martin norm and (b) a uniform modulus of continuity for elements in{h: I (h)Λ}. A soft argument, spelled out in Theorem 3 below, will then imply (3.2).
We start by proving that the quadratic variation of Cameron–Martin paths converges exponentially fast to 0.
Lemma 5.LetH∈(1/4,1). Then there exists a constantCH,csuch that for alln0and allh∈H
2n−1 k=0
|hk
2n,k2+n1|2CH,c|h|2H2n(1−4H )/2.
2 The convergence in was proven inp-variation topology, but it is obvious when reading the proof that it also holds in Hölder norm.
Proof. Without loss of generalities, we assumed=1. WriteR(s, t )=E(XsXt)for the covariance. The Cameron–
Martin space is the closure of functions of formh=
iaiR(ti,·)with respect to the norm
|h|H=
i
aiR(ti,·) 2
H=
i,j
aiajR(ti, tj)=E Z2 whereZ is the zero-mean Gaussian r.v. given byZ=
iaiXti. Observe thaths,t=E(ZXs,t). We keepnfixed for the rest of the proof.
Qn:=
2n−1 k=0
|hk
2n,k2+n1|2=
2n−1 k=0
E(Xk
2n,k2+n1Z)2.
Note that ifU, V are two zero-mean Gaussian random variables, then E
U2V2
=2E(U V )2+E U2
E V2
or Cov U2, V2
=2E(U V )2. (5.4)
This allows us to write Qn=1
2
2n−1 k=0
E Z2
X2k
2n,k2+n1 −E X2k
2n,k2+n1
=1 2E
Z2
2n−1 k=0
X2k
2n,k2+n1 −E X2k
2n,k2+n1
1
2E Z41/2
2n−1 k,l=0
Cov X2k
2n,k2+n1, X2l
2n,l2+n1
1/2
= 3
2E
Z22n−1 k,l=0
E(Xk 2n,k2+n1Xl
2n,l2+n1)2 1/2
where we used (5.4) again. The double sum in line just above splits naturally in a diagonal part, where k=l, and twice the sum overk < l. With (5.1) we control the diagonal part,
2n−1 k=0
E X2k
2n,k+1
2n
2
c22n(1−4H ).
The other part is estimated by using (5.2). Indeed,
0k<l2n−1
E(Xk 2n,k+12n X l
2n,l+12n )2=
2n−1 k=0
2n−k−1 m=1
E(Xk
2n,k+12n Xk+m
2n ,k+m+12n )2 c2
2n−1 k=0
2n−k−1 m=1
m 2n
4H−42−4n
∞
m=1
m4H−4
c22n(1−4H ). Putting everything together yields a constantC=C(H, c)such that
QnCHE Z2
2n(1−4H )/2=CH,c|h|H2n(1−4H )/2. This finishes the proof. 2
Corollary 3.LetH∈(1/4,1/2]andq∈(1∨1/(H+1/4),2). ThenHis continuously embedded inCq-var0 ([0,1],Rd).
More precisely, there exists a constantCH,q,csuch that for allh∈H,
|h|q-varCH,q,c|h|H.
Proof. We use the inequality
|h|qq-varCγ ,q
∞ n=1
nγ
2n−1 k=0
|hk 2n,k+1
2n |q,
see [17] for example. From the previous lemma and Hölder’s inequality,
2n−1 k=0
|hk
2n,k2+n1|q 2n−1
k=0
|hk 2n,k2+n1|2
q/2
2n(1−q/2) C|h|qH2n[(1−4H )q4+1−q2]. Observe
(1−4H )q
4 +1−q
2<0⇔q
H+1 4
>1.
Hence, for everyq1 such thatq >1/(H+1/4), we have
|h|q-varCH,q,c|h|H. 2
This result is not optimal. Cameron–Martin paths associated to usual Brownian motion are of bounded variation (≡finite 1-variation) whereas the above corollary applied toH=1/2 only gives finiteq-variation forq >4/3. In [12], we prove that Cameron–Martin paths associated to fractional Brownian motion are of finiteq-variation for any qgreater than 1∨1/(H+1/2).
Lemma 6. Let H∈(0,1). Then His continuously embedded in C0H-Hölder([0,1]). More precisely, there exists a constantCH,csuch that for allh∈H,
|h|H-HölderCH,c|h|H.
Proof. We use the notation of the above proofs. In particular,hs,t=E(ZXs,t). By Cauchy–Schwartz and (5.1),
|hs,t| E
Z2 E X2s,t
C|h|H|t−s|H. 2 Theorem 3.For any fixedN1,for allΛ >0,
mlim→∞ sup
{h:I (h)Λ}d∞
(SN◦Φm)(h), SN(h)
=0.
Proof. Observe first that forhsuch thatI (h)Λ, we have Φm(h)−h
∞ sup
|t−s|2−m
|hs,t| Λ1/2CH2−mH. Observe also that forq1/(1/4+H ),
Φm(h)
q-var3|h|q-var3Cq,hΛ1/2.
By interpolation, forq> q,we therefore obtain that
mlim→∞ sup
{h:I (h)Λ}
Φm(h)−h
q-var=0.
Then using Theorem 2.2.2 in [16] (or just the continuity of the Young integral), we obtain our theorem. 2 We have therefore the following:
Corollary 4.Assume that for someH∈(1/4,1), the Gaussian process satisfies condition(5.1)and(5.2)and letX be as in Definition1. Then,(δεX)satisfies a large deviation principle in1/p-Hölder topology wherep >1/H with good rate function
J (y)=inf 1
2|x|H, xsuch thatS[p](x)=y
.
Once again, applying the Itô map to δεX,we can obtain a generalisation of the Freidlin–Wentzell theorem, i.e.
a large deviation principle for solution of stochastic differential equation driven by the enhanced Gaussian processX.
Acknowledgements
The authors would like to thank Christer Borell, Ben Garling for helpful discussions and an anonymous referee for his/her careful reading. The first author acknowledges partial support from a Leverhulme Research Fellowship.
References
[1] C. Borell, Tail probabilities in Gauss space, in: Vector Space Measures and Applications, Dublin, 1977, in: Lecture Notes in Math., vol. 644, Springer-Verlag, 1978, pp. 73–82.
[2] C. Borell, On polynomials chaos and integrability, Probab. Math. Statist. 3 (1984) 191–203.
[3] C. Borell, On the Taylor series of a Wiener polynomial, in: Seminar Notes on Multiple Stochastic Integration, Polynomial Chaos and their Integration, Case Western Reserve University, Cleveland, 1984.
[4] L. Coutin, P. Friz, N. Victoir, Good rough path sequences and applications to anticipating & fractional stochastic calculus, ArXiv-preprint, 2005.
[5] L. Coutin, Z. Qian, Stochastic analysis, rough path analysis and fractional Brownian motions, Probab. Theory Relat. Fields 122 (2002) 108–
140.
[6] L. Coutin, N. Victoir, Lévy area for Gaussian processes, preprint, 2005.
[7] A. Dembo, O. Zeitouni, Large Deviations Techniques and Applications, second ed., Applications of Mathematics, vol. 38, Springer-Verlag, New York, ISBN 0-387-98406-2, 1998, xvi+396 pp.
[8] D. Feyel, A. de La Pradelle, Curvilinear integrals along enriched paths, Electronic J. Probab. 11 (2006) 860–892.
[9] P. Friz, Continuity of the Itô-map for Hölder rough path with applications to the Support Theorem in Hölder norm, in: Probability and Partial Differential Equations in Modern Applied Mathematics, in: IMA Volumes in Mathematics and its Applications, vol. 140, 2003, pp. 117–136.
[10] P. Friz, N. Victoir, Approximations of the Brownian rough path with applications to stochastic analysis, Ann. Inst. H. Poincaré Probab.
Statist. 41 (4) (2005) 703–724.
[11] P. Friz, N. Victoir, On the notion of geometric rough paths, Probab. Theory Relat. Fields 136 (3) (2006) 395–416.
[12] P. Friz, N. Victoir, A variation embedding theorem and applications, J. Funct. Anal. 239 (2) (2006) 631–637.
[13] N.C. Jain, G. Kallianpur, A note on uniform convergence of stochastic processes, Ann. Math. Statist. 41 (1970) 1360–1362.
[14] M. Ledoux, Isoperimetry and Gaussian analysis, in: Lectures on Probability Theory and Statistics, Saint-Flour, 1994, in: Lecture Notes in Math., vol. 1648, Springer, Berlin, 1996, pp. 165–294.
[15] M. Ledoux, Z. Qian, T. Zhang, Large deviations and support theorem for diffusion processes via rough paths, Stochastic Process. Appl. 102 (2) (2002) 265–283.
[16] T.J. Lyons, Differential equations driven by rough signals, Rev. Mat. Iberoamericana 14 (2) (1998) 215–310.
[17] T. Lyons, Z. Qian, System Control and Rough Paths, Oxford University Press, 2002.
[18] A. Millet, M. Sanz-Sole, Large deviations for rough paths of the fractional Brownian motion, Ann. Inst. H. Poincaré Probab. Statist. 42 (2) (2006) 245–271.
[19] A. Millet, M. Sanz-Sole, Approximation of rough paths of fractional Brownian motion, ArXiv-preprint, 2005.
[20] M. Schreiber, Fermeture en probabilité de certains sous-espaces d’un espaceL2. Application aux chaos de Wiener, Z. Wahrsch. Verw. Gebi- ete 14 (1969/70) 36–48.
Further reading
[1] L. Decreusefond, A.S. Üstünel, Stochastic analysis of the fractional Brownian motion, Potential Anal. 10 (1997) 177–214.