• Aucun résultat trouvé

Mean square rate of convergence for random walk approximation of forward-backward SDEs

N/A
N/A
Protected

Academic year: 2021

Partager "Mean square rate of convergence for random walk approximation of forward-backward SDEs"

Copied!
41
0
0

Texte intégral

(1)

HAL Id: hal-01838449

https://hal.archives-ouvertes.fr/hal-01838449v2

Preprint submitted on 5 Mar 2020

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.

approximation of forward-backward SDEs

Christel Geiss, Céline Labart, Antti Luoto

To cite this version:

Christel Geiss, Céline Labart, Antti Luoto. Mean square rate of convergence for random walk approx-

imation of forward-backward SDEs. 2020. �hal-01838449v2�

(2)

MEAN SQUARE RATE OF CONVERGENCE FOR RANDOM WALK AP- PROXIMATION OF FORWARD-BACKWARD SDES

CHRISTEL GEISS,Department of Mathematics and Statistics, University of Jyvaskyla C´ELINE LABART,∗∗ Univ. Grenoble Alpes, Univ. Savoie Mont Blanc, CNRS, LAMA ANTTI LUOTO,Department of Mathematics and Statistics, University of Jyvaskyla

Abstract

Let (Y, Z) denote the solution to a forward-backward SDE. If one constructs a random walk Bn from the underlying Brownian motion B by Skorohod embedding, one can show L2-convergence of the corresponding solutions (Yn, Zn) to (Y, Z). We estimate the rate of convergence in dependence of smoothness properties, especially for a terminal condition function inC2,α. The proof relies on an approximative representation ofZnand uses the concept of discretized Malliavin calculus. Moreover, we use growth and smoothness properties of the PDE associated to the FBSDE as well as of the finite difference equations associated to the approximating stochastic equations. We derive these properties by probabilistic methods.

Keywords: Backward stochastic differential equations; approximation scheme;

finite difference equation; convergence rate; random walk approximation 2010 Mathematics Subject Classification: Primary 60H10; 60H35; 60G50

Secondary 60H30

1. Introduction

Let (Ω,F,P) be a complete probability space carrying the standard Brownian motion B = (Bt)t0 and assume that (Ft)t0 is the augmented natural filtration. Let (Y, Z) be the solution of the forward-backward SDE (FBSDE)

Xs=x+ Z s

0

b(r, Xr)dr+ Z s

0

σ(r, Xr)dBr, Ys=g(XT) +

Z T s

f(r, Xr, Yr, Zr)dr− Z T

s

ZrdBr, 0≤s≤T. (1)

Postal address: P.O.Box 35 (MaD) FI-40014 University of Jyvaskyla, Finland

∗∗Postal address: Univ. Grenoble Alpes, Univ. Savoie Mont Blanc, CNRS, LAMA, 73000 Chamb´ery, France

1

(3)

Let (Yn, Zn) be the solution of the FBSDE if the Brownian motion B is replaced by a scaled random walkBn given by

Bnt =√ h

[t/h]

X

i=1

εi, 0≤t≤T, (2)

whereh=Tn and (εi)i=1,2,... is a sequence of i.i.d. Rademacher random variables. Then (Yn, Zn) solves the discretized FBSDE

Xsn=x+ Z

(0,s]

b(r, Xrn)d[Bn]r+ Z

(0,s]

σ(r, Xrn)dBrn, Ysn =g(XTn) +

Z

(s,T]

f(r, XrnYrn, Zrn)d[Bn]r− Z

(s,T]

ZrndBnr, 0≤s≤T. (3) The approximation of BSDEs using random walk has been investigated by many authors, also numerically (see, for example, [5], [25], [29], [31], [32], [33], [16]). In 2001, Briand et al. [5]

have shown weak convergence of (Yn, Zn) to (Y, Z) for a Lipschitz continuous generatorf and a terminal condition in L2. The rate of convergence of this method remained an open problem.

Bouchard and Touzi in [7] and Zhang in [41] proposed instead of random walk an approach based on the dynamic programming equation, for which they established a rate of convergence. But this approach involves conditional expectations. Various methods to approximate these conditional expectations have been developed ([23], [17], [14]). Also forward methods have been introduced to approximate (1): a branching diffusion method ([26]), a multilevel Picard approximation ([40]) and Wiener chaos expansion ([6]). Many extensions of (1) have been considered, among them schemes for reflected BSDEs ([3], [13]), high order schemes ([10], [9]), fully-coupled BSDEs ([18], [8]), quadratic BSDEs ([12]), BSDEs with jumps ([22]) and McKean-Vlasov BSDEs ([1], [15], [11]).

The aim of this paper is to study the rate of the L2-approximation of (Ytn, Ztn) to (Yt, Zt) when X satisfies (1). For this, we generate the random walk Bn by Skorohod embedding from the Brownian motionB. In this case theLp-convergence ofBn toBis of orderh14 for anyp >0.

The special caseX=B has already been studied in [21], assuming a locallyα-H¨older continuous terminal functiongand a Lipschitz continuous generator. An estimate for the rate of convergence was obtained which is of orderhα4 for theL2-norm ofYtn−Yt,and of order h

α

4

Tt for theL2-norm ofZtn−Zt.

In the present paper, where we assume that X is a solution of the SDE in (1), rather strong conditions on the smoothness and boundedness onf and g and also on b and σ are needed. In Theorem 3.1, the main result of the paper, we show that the convergence rate for (Ytn, Ztn) to

(4)

(Yt, Zt) inL2is of orderh14α2 provided thatg′′is locallyα-H¨older continuous. To the best of our knowledge, these are the first cases a convergence rate for the approximation of forward-backward SDEs using random walk has been obtained.

Remark 1.1. For the diffusion setting – in contrast to the case X = B – we can derive the convergence rate for (Ytn, Ztn) to (Yt, Zt) in L2 only under strong smoothness conditions on the coefficients which include also thatg′′is locallyα-H¨older continuous (see Assumption2.3below).

These requirements appear to be necessary. This becomes visible in Subsection 2.2.2 where we introduce a discretized Malliavin weight to obtain a representation ˆZnforZn.While it holds that Zˆn =Zn when X =B, in our case ˆZn does not coincide with Zn.However, one can show that the difference ˆZtn−Ztn converges to 0 inL2 as n → ∞ using a H¨older continuity property (see (62) in Remark 4.1) for the space derivative of the generator in (3). For this H¨older continuity property to hold one needs enough smoothness in space from the solutionunto the finite difference equation associated to the discretized FBSDE (3). Provided that Assumption 2.3holds we show the smoothness properties for un in Proposition 4.2 applying methods known for L´evy driven BSDEs.

The paper is organized as follows: Section 2 contains the setting, main assumptions and the approximative representation ˆZn of Zn. Our main results about the approximation rate for the case of no generator (i.e. f = 0) and for the general case are in Section 3. One can see that in contrast to what is known for time discretization schemes, for random walk schemes the Lipschitz generator seems to cause more difficulties than the terminal condition: while in the casef = 0 we need thatgis locallyα-H¨older continuous, in the casef 6= 0 is this property is required forg′′.In Section 4 we recall some needed facts about Malliavin weights, about the regularity of solutions to BSDEs and properties of the associated PDEs. Finally, we sketch how to prove growth and smoothness properties of solutions to the finite difference equation associated to the discretized FBSDE. Section5contains technical results which mainly arise from the fact that the construction of the random walk by Skorohod embedding forces us to compare our processes on different ’time lines’, one coming from the stopping times of the Skorohod embedding, and the other one is ruled by the equidistant deterministic times due to the quadratic variation process [Bn].

(5)

2. Preliminaries 2.1. The SDE and its approximation scheme

We introduce

Xt=x+ Z t

0

b(s, Xs)ds+ Z t

0

σ(s, Xs)dBs, 0≤t≤T and its discretized counterpart

Xtnk =x+h Xk j=1

b(tj, Xtnj−1) +√ h

Xk j=1

σ(tj, Xtnj−1j, tj:=jTn, j= 0, ..., n, (4) where (εi)i=1,2,... is a sequence of i.i.d. Rademacher random variables. Letting Gk :=σ(εi : 1 ≤ i ≤ k) with G0 := {∅,Ω}, it follows that the associated discrete-time random walk (Btnk)nk=0 is (Gk)nk=0-adapted. Recall (2) and h = Tn. If we extend the sequence (Xtnk)k0 to a process in continuous time by definingXtn:=Xtnkfort∈[tk, tk+1),it is the solution of the forward SDE (3).

We formulate our first assumptions. Assumption 2.1 (ii) will be not used explicitely for our estimates but it is required for Theorem4.1below.

Assumption 2.1.

(i) b, σ∈Cb0,2([0, T]×R), in the sense that the derivatives of orderk = 0,1,2 w.r.t. the space variable are continuous and bounded on[0, T]×R,

(ii) the first and second derivatives ofbandσw.r.t. the space variable are assumed to beγ-H¨older continuous (for someγ∈(0,1],w.r.t. the parabolic metricd((t, x),(¯t,x)) = (¯ |t−¯t|+|x−x¯|2)12) on all compact subsets of[0, T]×R.

(iii) b, σare 12-H¨older continuous in time, uniformly in space, (iv) σ(t, x)≥δ >0for all(t, x).

Assumption 2.2.

(i) g is locally H¨older continuous with orderα∈(0,1]and polynomially bounded (p0≥0, Cg>

0) in the following sense

∀(x,x)¯ ∈R2, |g(x)−g(¯x)| ≤Cg(1 +|x|p0+|x¯|p0)|x−x¯|α. (5)

(ii) The function [0, T]×R3: (t, x, y, z)7→f(t, x, y, z)satisfies

|f(t, x, y, z)−f(¯t,x,¯ y,¯ ¯z)| ≤Lf(p

t−¯t+|x−x¯|+|y−y¯|+|z−z¯|). (6)

(6)

Notice that (5) implies

|g(x)| ≤K(1 +|x|p0+1) =: Ψ(x), x∈R, (7) for someK >0.From the continuity off we conclude that

Kf := sup

0tT|f(t,0,0,0)|<∞. Notation:

• k · kp:=k · kLp(P) forp≥1 and forp= 2 simplyk · k.

• If ais a function,C(a) represents a generic constant which depends ona and possibly also on its derivatives.

• E0,x:=E(·|X0=x).

• Letφbe aC0,1([0, T]×R) function. φxdenotes∂xφ, the partial derivative ofφw.r.t. x.

2.2. The FBSDE and its approximation scheme

Recall the FBSDE (1) and its approximation (3). The backward equation in (3) can equivalently be written in the form

Ytnk=g(XTn) +h

n1

X

m=k

f(tm+1, Xtnm, Ytnm, Ztnm)−√ h

nX1 m=k

Ztnmεm+1, 0≤k≤n, (8) if one putsXrn:=Xtnm, Yrn :=Ytnm andZrn:=Ztnm forr∈[tm, tm+1).

Remark 2.1. Equations (3) and (8) do not contain any orthogonal part to the random walkBn since we are in a special case where the orthogonal part is zero. Indeed, for (εk)k=1,···,nfollowing the Rademacher law assume (Gk:=σ(εi, i= 1,· · ·, k)) as filtration, and let for theGn-measurable random variableF(ε1, ..., εn) hold the representation

F(ε1, ..., εn) =c+ Xn m=1

hmεm+Nn,

where (hm)nm=1is predictable and (Nm)nm=1a martingale orthogonal to (Bntm)nm=1given byBntm =

√h(ε1+...+εm). By definition, orthogonality of the martingales N and Bn means that their product is a martingale, i.e. we have

E[Nk+1Btnk+1|Gk] =NkBtnk,

and sinceNkBtnk =E[Nk+1Btnk|Gk], this implies especially thatENk+1εk+1 = 0. Assume Nk+1 is given byNk+1=H(ε1, ..., εk+1).Then 0 =EH(ε1, ..., εk+1k+1=12[H(ε1, ..., εk,1)−H(ε1, ..., εk,−1)]

(7)

implying that Nk+1 is Gk-measurable since H(ε1, ..., εk,1) =H(ε1, ..., εk,−1), and therefore the martingale (Nm)nm=1 is identically zero. (See also [5, page 3] or [34, Proposition1.7.5].)

One can derive an equation for Zn = (Ztnk)nk=01 if one multiplies (8) by εk+1 and takes the conditional expectation w.r.t. Gk, so that

Ztnk = EGk(g(XTnk+1)

√h +EGk √ h

n1

X

m=k+1

f(tm+1, Xtnm, Ytnm, Ztnmk+1

!

, 0≤k≤n−1,(9)

whereEGk :=E(·|Gk).

Remark 2.2. Fornlarge enough, the BSDE (3) has a unique solution (Yn, Zn) (see [36, Propo- sition 1.2]), and (Ytnk, Ztnk)nk=01 is adapted to the filtration (Gk)nk=01.

2.2.1. Representation forZ We will use the following representation forZ, due to Ma and Zhang (see [30, Theorem 4.2])

Zt=Et g(XT)NTt + Z T

t

f(s, Xs, Ys, Zs)Nstds

!

σ(t, Xt), 0≤t≤T (10) whereEt:=E(·|Ft), and for alls∈(t, T], we have (cf. Lemma4.1)

Nst= 1 s−t

Z s t

∇Xr

σ(r, Xr)∇Xt

dBr, (11)

where∇X= (∇Xs)s[0,T] is the variational process i.e. it solves

∇Xs= 1 + Z s

0

bx(r, Xr)∇Xrdr+ Z s

0

σx(r, Xr)∇XrdBr, (12) with (Xs)s[0,T] given in (1).

Remark 2.3. In the following we will assume thatg′′exists. In such a case we have the following representation forZ:

Zt=Et g(XT)∇XT + Z T

t

f(s, Xs, Ys, Zs)Nstds

!

σ(t, Xt), 0≤t≤T. (13) 2.2.2. Approximation for Zn In this section we state the discrete counterpart to (10), which, in the general case of a forward processX, does not coincide withZn (given by (9)). In contrast to the continuous-time case, where the variational process and the Malliavin derivative are connected by XXst = σ(s,XDsXts) (s≤t), we can not expect equality for the corresponding expressions if we use the discretized version of the processes (∇Xt)tand (DsXt)stintroduced in (15). This counterpart Zˆn toZ is a key tool in the proof of the convergence ofZn to Z. As we will see in the proof of

(8)

Theorem3.1, the study ofkZtnk−Ztkkgoes through the study ofkZtnk−Zˆtnkk andkZˆtnk−Ztkk.

Before defining the discretized version of (∇Xt)t and (DsXt)st, we shortly introduce the discretized Malliavin derivative and refer the reader to [4] for more information on this topic.

Definition 2.1. (Definition of Tm,+, Tm, and Dnm.) For any function F : {−1,1}n → R, the mappingsTm,+ andTm, are defined by

Tm,±F(ε1, . . . , εn) :=F(ε1, . . . , εm1,±1, εm+1, . . . , εn), 1≤m≤n.

For anyξ=F(ε1, . . . , εn), the discretized Malliavin derivative is defined by Dmnξ:= E[ξεm|σ((εl)l∈{1,...,n}\{m})]

√h = Tm,+ξ−Tm,−ξ 2√

h , 1≤m≤n. (14)

Definition 2.2. (Definition of φ(k,l)x .) Letφbe aC0,1([0, T]×R)function. We denote φ(k,l)x := Dnkφ(tl, Xtnl−1)

DknXtnl−1 :=

Z 1 0

φx(tl, ϑTk,+Xtnl−1+ (1−ϑ)Tk,Xtnl−1)dϑ.

IfDknXtnℓ−1 6= 0the second := holds as an identity.

We are now able to define the discretized version of (∇Xt)tand (DsXt)st.

Definition 2.3. (Discretized processes (∇Xtn,tmk,x)m∈{k,...,n} and (DknXtnm)m∈{k,...,n}.) For allm in{k, . . . , n} we define

∇Xtn,tmk,x= 1 +h Xm l=k+1

bx(tl, Xtn,tl−1k,x)∇Xtn,tl−1k,x+√ h

Xm l=k+1

σx(tl, Xtn,tl−1k,x)∇Xtn,tl−1k,xεl, 0≤k≤n,

DnkXtnm=σ(tk, Xtnk−1) +h Xm l=k+1

b(k,l)x DknXtnl−1+√ h

Xm l=k+1

σx(k,l)(DknXtnl−1l, 0< k≤n. (15)

Remark 2.4. (i) Although∇Xn,tk,X

n tk

tm is not equal to σ(tDnk+1Xtmn

k+1,Xtkn), we can show that the differ- ence of these terms converges inLp (see Lemma5.4).

(ii) With the notation introduced above, (9) rewrites to Ztnk = EGk Dnk+1g(XTn)

+EGk h

nX1 m=k+1

Dk+1n f(tm+1, Xtnm, Ytnm, Ztnm)

!

. (16) In order to define the discrete counterpart to (10), we first define the discrete counterpart to (Nst)s[t,T] given in (11):

Ntn,t k:=√ h

X m=k+1

∇Xn,tk,X

n tk

tm−1

σ(tm, Xtnm−1) εm

t−tk

, k < ℓ≤n. (17)

(9)

Notice that there is some constantbκ2>0 depending onb, σ, T, δ such that

EGk|Ntn,t k|212

≤ bκ2

(t−tk)12, 0≤k < ℓ≤n. (18) Definition 2.4. (Discrete counterpart to (13).) Let the processZˆn= ( ˆZtnk)nk=01 be defined by

tnk:=EGk Dnk+1g(XTn)

+EGk h

n1

X

m=k+1

f(tm+1, Xtnm, Ytnm, Ztnm)Ntn,tmk

!

σ(tk+1, Xtnk), (19)

Remark 2.5. In (19) We could have used also the approximate expressionEGk(g(XTn)Ntn,tn kσ(tk+1, Xtnk)), but since we will assume thatg′′ exists, we work with the correct term.

The study of the convergenceEG0,x|Ztnk−Zˆtnk|2requires stronger assumptions on the coefficients b,σ, f andg.

Assumption 2.3. Assumptions 2.1 and 2.2 hold. Additionally, we assume that all first and second derivatives w.r.t. the variablesx, y, zofb(t, x), σ(t, x)andf(t, x, y, z)exist and are bounded Lipschitz functions w.r.t. these variables, uniformly in time. Moreover,g′′ satisfies (5).

Proposition 2.1. If Assumption 2.3holds, then

EG0,x|Ztnk−Zˆtnk|2≤C2.1Ψˆ2(x)hα,

where EG0,x := EG(·|X0 = x), the function Ψˆ is defined in (61) below, and C2.1 depends on b, σ, f, g, T, p0 andδ.

Proof. According to [5, Proposition 5.1] one has the representations

Ytnm =un(tm, Xtnm), and Ztnm =Dnm+1un(tm+1, Xtnm+1), (20) whereun is the solution of the finite difference equation (43) with terminal conditionun(tn, x) = g(x).Notice that by the definition ofDm+1n in (14) the expressionDm+1n un(tm+1, Xtnm+1) depends in fact onXtnm.Hence we can put

f(tm+1, Xtnm, Ytnm, Ztnm) = f(tm+1, Xtnm, un(tm, Xtnm),Dm+1n un(tm+1, Xtnm+1))

=: Fn(tm+1, Xtnm).

(10)

From (19) and (16) we conclude that (we useE:=EG0,x fork · k) kZtnk−Zˆtnkk

= EGk h

n1

X

m=k+1

Dnk+1f(tm+1, Xtnm, Ytnm, Ztnm)

!

−EGk h

nX1 m=k+1

f(tm+1, Xtnm, Ytnm, Ztnm)Ntn,tmkσ(tk+1, Xtnk)!

n1

X

m=k+1

h m−k

Xm ℓ=k+1

EGk

"

Dnk+1Fn(tm+1, Xtnm)− DnFn(tm+1, Xtnm)σ(tk+1, Xtnk)∇Xn,tk,X

n tk

t−1

σ(t, Xtnℓ−1)

#.

With the notation introduced in Definition2.2applied to Fn,

Dnk+1Fn(tm+1, Xtnm)− DnFn(tm+1, Xtnm)σ(tk+1, Xtnk)∇Xn,tk,X

n tk

t−1

σ(t, Xtnℓ−1)

≤ k(Dk+1n Xtnm)(Fn,(k+1,m+1)

x −Fxn,(ℓ,m+1))k +

Fxn,(ℓ,m+1)

(Dk+1n Xtnm)−(DnXtnm)σ(tk+1, Xtnk)∇Xn,tk,X

n tk

tℓ−1

σ(t, Xtnℓ−1)

=: A1+A2.

ForA1 we use Definition2.2again and exploit the fact that

x7→Fxn(tm+1, x) :=∂xf(tm+1, x, un(tm, x),Dnm+1un(tm+1, Xtn,tm+1m,x))

is locally α-H¨older continuous according to (62). By Hlder’s inequality and Lemma 5.4 (i) and (iii),

A1≤ kDk+1n Xtnmk4 Z 1

0 kFxn(tm+1, ϑTk+1,+Xtnm+ (1−ϑ)Tk+1,−Xtnm)

−Fxn(tm+1, ϑTℓ,+Xtnm+ (1−ϑ)Tℓ,Xtnm)k4dϑ≤C(b, σ, f, g, T, p0) ˆΨ(x)hα2. For the estimate of A2 we notice that by our assumptions theL4-norm ofFxn,(ℓ,m+1) is bounded byCΨ2(x),so that it suffices to estimate

(Dk+1n Xtnm)−(DnXtnm)σ(tk+1, Xtnk)∇Xn,tk,X

n tk

t−1

σ(t, Xtn−1)

4

(Dk+1n Xtnm)−σ(tk+1, Xtnk)DnXtnm σ(t, Xtnℓ−1)

Dnk+1Xtn−1 σ(tk+1, Xtnk)

4

+

σ(tk+1, Xtnk)DnXtnm

σ(t, Xtn−1) ∇Xn,tk,X

n

t−1 tk − Dnk+1Xtnℓ−1 σ(tk+1, Xtnk)

!

4

. (21)

(11)

The second expression on the r.h.s. of (21) is bounded by C(b, σ, T, δ)h12 as a consequence of Lemma 5.4 (ii)-(iii). To show that also the first expression is bounded by C(b, σ, T, δ)h12, we rewrite it using (15) and get

DnXtnm

σ(t, Xtnℓ−1)Dnk+1Xtn−1− Dk+1n Xtnm

= 1 +

Xm l=ℓ+1

DnXtnl−1

σ(t, Xtnℓ−1)(b(ℓ,l)x h+σx(ℓ,l)√ hεl)

!

× σ(tk+1, Xtnk) +

1

X

l=k+2

Dk+1n Xtnl−1(b(k+1,l)x h+σ(k+1,l)x √ hεl)

!

− σ(tk+1, Xtnk) + X1

l=k+2

+ Xm

l=ℓ

Dnk+1Xtnl−1(b(k+1,l)x h+σx(k+1,l)√ hεl)

!

≤Dk+1n Xtnℓ−1(b(k+1,ℓ)x h+σ(k+1,ℓ)x √ hε) +

Xm l=ℓ+1

DnXtnl−1

σ(t, Xtn−1)Dnk+1Xtn−1− Dnk+1Xtnl−1

b(ℓ,l)x h+σ(ℓ,l)x √ hεl

+

Xm l=ℓ+1

Dk+1n Xtnl−1

b(ℓ,l)x h+σx(ℓ,l)√ hεl

b(k+1,l)x h+σ(k+1,l)x √ hεl

. (22)

We take theL4-norm of (22) and apply the BDG inequality and H¨older’s inequality. The second term on the r.h.s. of (22) will be used for Gronwall’s lemma, while the first and the last one can be bounded byC(b, σ, T)h12,by using Lemma5.4-(iii). For the last term we also use the Lipschitz

continuity ofbx andσx in space and Lemma5.4-(i).

3. Main results

In order to compute the mean square distance between the solution to (1) and the solution to (3) we construct the random walk Bnfrom the Brownian motionB by Skorohod embedding. Let

τ0:= 0 and τk:= inf{t > τk1 :|Bt−Bτk−1|=√

h}, k≥1. (23)

Then (Bτk−Bτk−1)k=1 is a sequence of i.i.d. random variables with P(Bτk−Bτk−1 =±√

h) = 12,

which means that √ hεk

=d Bτk −Bτk−1. We will denote by Eτk the conditional expectation w.r.t.Fτk:=Gk.In this case we also use the notationXτk :=Xtnk for allk= 0, . . . , n,so that (4)

(12)

turns into

Xτk=x+ Xk j=1

b(tj,Xτj−1)h+ Xk j=1

σ(tj,Xτj−1)(Bτj −Bτj−1), 0≤k≤n.

Assumption 3.1. We assume that the random walk Bn in(3)is given by Btn=

[t/h]

X

k=1

(Bτk−Bτk−1), 0≤t≤T, where theτk, k= 1, ..., nare taken from(23).

Remark 3.1. Note that forp >0 there exists a C(p)>0 such that for allk= 1, . . . , n it holds

1

C(p)(tkh)14 ≤(E|Bτk−Btk|p)1p ≤C(p)(tkh)14.

The upper estimate is given in Lemma 5.1. For p ∈ [4,∞) the lower estimate follows from [2, Proposition 5.3]. We get the lower estimate forp∈(0,4) by choosing 0< θ <1 and 0< p < p1

such that 14 = 1pθ+ pθ

1. Then it holds by the log-convexity of Lp norms (see, for example [35, Lemma 1.11.5]) that

kBτk−Btkk1pθ≥ kBτk−Btkk4 kBτk−Btkkθp1

≥ C(4)1(tkh)14 C(p1)(tkh)14θ

C(p)(tkh)141θ

.

Since fort∈[tk, tk+1) it holdsBtn=Bτk andkBt−Btkkp≤C(p)h12,we have for anyp >0 that sup

0tTkBtn−Btkp=O(h14). (24) Proposition 3.1 states the convergence rate of (Yv, Zv) to (Yvn, Zvn) in L2 when f = 0 and Theorem3.1generalizes this result for anyf which satisfies Assumption2.3.

Proposition 3.1. Let Assumptions 2.1 and3.1 hold. If f = 0and g ∈ C1 is such that g is a locally α-H¨older continuous function in the sense of (5), then for all 0 ≤ v < T, we have (for sufficiently largen) that

E0,x|Yv−Yvn|2≤C3.1y Ψ(x)2h12, and E0,x|Zv−Zvn|2≤C3.1z Ψ(x)2hα2, whereC3.1y =C(Cg, b, σ, T, p0, δ)andC3.1z =C(Cg, b, σ, T, p0, δ).

Theorem 3.1. Let Assumptions2.3and3.1be satisfied. Then for allv∈[0, T)and large enough n, we have

E0,x|Yv−Yvn|2+E0,x|Zv−Zvn|2≤C3.1Ψ(x)ˆ 2h12α withC3.1=C(b, σ, f, g, T, p0, δ)andΨˆ is given in (61).

(13)

Remark 3.2. As noticed above, the filtrationGk coincides with Fτk, for all k = 0, . . . , n. The expectationE0,x appearing in Proposition 3.1 and in Theorem 3.1is defined on the probability space (Ω,F,P).

Remark 3.3. In order to avoid too much notation for the dependencies of the constants, if for example onlyg is mentioned and notCg,this means that the estimate might depend also on the bounds of the derivatives ofg.

From (24) one can see that the convergence rates stated in Proposition3.1 and Theorem3.1 are the natural ones for this approach. The results are proved in the next two sections. In both proofs, we will use the following remark.

Remark 3.4. Since the process (Xt)t0is strong Markov we can express conditional expectations with the help of an independent copy ofB denoted by ˜B, for exampleEτkg(XTn) = ˜Eg( ˜Xττnk,Xτk) for 0≤k≤n, where

ττnk,Xτk =Xτk+ Xn j=k+1

b(tj,X˜ττjk−1,Xτk)h+ Xn j=k+1

σ(tj,X˜ττjk−1,Xτk)( ˜B˜τj−k−B˜τ˜j−k−1), (25) (we define ˜τk := 0 and ˜τj := inf{t >τ˜j1 :|B˜t−B˜τ˜j−1|=√

h}forj ≥1 andτn:=τk+ ˜τnk for n≥k). In fact, to represent the conditional expectationsEtk andEτk we work here with ˜Eand the Brownian motionsB andB′′,respectively, given by

Bt=Bttk+ ˜B(ttk)+ and B′′t =Btτk+ ˜B(tτk)+, t≥0. (26) 3.1. Proof of Proposition3.1: the approximation rates for the zero generator case

To shorten the notation, we useE:=E0,x.Let us first deal with the error ofY. Ifv belongs to [tk, tk+1) we haveYvn=Ytnk. Then

E|Yv−Yvn|2≤2(E|Yv−Ytk|2+E|Ytk−Ytnk|2).

Using Theorem4.1we boundkYv−Ytkk by

C4.1y Ψ(x)(v−tk)12 =C(Cg, b, σ, T, p0, δ)Ψ(x)(v−tk)12

(sinceα= 1 can be chosen wheng is locally Lipschitz continuous). It remains to bound E|Ytk−Ytnk|2 = E|Etkg(XT)−Eτkg(XTn)|2=E|Eg( ˜˜ Xttnk,Xtk)−Eg( ˜˜ Xττnk,Xτk)|2. By (5) and the Cauchy-Schwarz inequality (Ψ1:=Cg(1 +|X˜ttnk,Xtk|p0+|X˜ττnk,Xτk|p0)),

|Eg( ˜˜ Xttnk,Xtk)−Eg( ˜˜ Xττnk,Xτk)|2 ≤ (˜E(Ψ1|X˜ttnk,Xtk−X˜ττkn,Xτk|))2≤E(Ψ˜ 21)˜E|X˜ttnk,Xtk−X˜ττnk,Xτk|2.

(14)

Finally, we get by Lemma5.2-(v) that

E|Ytk−Ytnk|2

EE(Ψ˜ 41)12

EE˜|X˜ttnk,Xtk −X˜ττnk,Xτk|412

≤C(Cg, b, σ, T, p0)Ψ(x)2h12.

Let us now deal with the error ofZ. We usekZv−Zvnk ≤ kZv−Ztkk+kZtk−Ztnkkand the representation

Zt=σ(t, Xt)˜E(g( ˜XTt,Xt)∇X˜Tt,Xt) (see Theorem4.2), where

st,x = x+ Z s

t

b(r,X˜rt,x)dr+ Z s

t

σ(r,X˜rt,x)dB˜rt, (27)

∇X˜st,x = 1 + Z s

t

bx(r,X˜rt,x)∇X˜rt,xdr+ Z s

t

σx(r,X˜rt,x)∇X˜rt,xdB˜rt, 0≤t≤s≤T.

For the first term we get by the assumption ong and Lemma5.2-(i) and (iii)

kZv−Ztkk = kσ(v, Xv)˜E(g( ˜XTv,Xv)∇X˜Tv,Xv)−σ(tk, Xtk)˜E(g( ˜XTtk,Xtk)∇X˜Ttk,Xtk)k

≤ kσ(v, Xv)−σ(tk, Xtk)k4kE(g˜ ( ˜XTv,Xv)∇X˜Tv,Xv)k4 +kσkkE(g˜ ( ˜XTv,Xv)∇X˜Tv,Xv)−E(g˜ ( ˜XTtk,Xtk)∇X˜Tv,Xv)k +kσkkE(g˜ ( ˜XTtk,Xtk)∇X˜Tv,Xv)−E(g˜ ( ˜XTtk,Xtk)∇X˜Ttk,Xtk)k

≤ C(Cg, b, σ, T, p0)Ψ(x)h

h12+kXv−Xtkk4+

EE˜|X˜Tv,Xv−X˜Ttk,Xtk|14 +

EE˜|∇X˜Tv,Xv− ∇X˜Ttk,Xtk|414i

≤ C(Cg, b, σ, T, p0)Ψ(x)hα2.

We compute the second term usingZtnkas given in (16). Hence, with the notation from Definition 2.2,

kZtk−Ztnkk2 = Eσ(tk, Xtk)˜Eg( ˜Xttnk,Xtk)∇X˜ttnk,Xtk−ED˜ nk+1g( ˜Xττnk,Xτk)2

≤ kσk2E

E(g˜ ( ˜Xttnk,Xtk)∇X˜ttnk,Xtk)−E˜Dk+1n g( ˜Xττnk,Xτk) σ(tk, Xtk)

2

= kσk2E

E(g˜ ( ˜Xttnk,Xtk)∇X˜ttnk,Xtk)−E˜

g(k+1,n+1)x Dk+1nττnk,Xτk

σ(tk, Xtk)

2

.

Références

Documents relatifs