HAL Id: hal-00196841
https://hal.archives-ouvertes.fr/hal-00196841v4
Submitted on 11 Nov 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
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
Representation theorems for backward doubly stochastic differential equations
Auguste Aman
To cite this version:
Auguste Aman. Representation theorems for backward doubly stochastic differential equations. 2008.
�hal-00196841v4�
doubly stohasti dierential equations
Auguste Aman
∗
UFR de Mathématiques et Informatique,
22 BP 582 Abidjan 22, Cte d'Ivoire
Abstrat
In this paperwe study the lass of bakward doubly stohasti dierential equations
(BDSDEs, for short) whose terminal value depends on the history of forward diusion.
We rst establisha probabilisti representation for the spatial gradient of the stohasti
visositysolution to a quasilinear paraboli SPDE inthe spirit of the Feynman-Ka for-
mula,withoutusingthederivativesoftheoeientsoftheorrespondingBDSDE.Then
suha representation leadsto a losed-formrepresentation ofthemartingaleintegrand of
BDSDE,underonly standard Lipshitz onditionon theoeients.
Key Words: Adapted solution, antiipatingstohasti alulus, bakward doubly SDEs,
stohasti partial dierential equation,stohasti visosity solutions.
MSC:60H15; 60H20
1 Introdution
Bakward stohasti dierential equations (BSDEs, for short) were rstly been onsidered in
itlinear formby Bismut[1,2℄inthe ontextof optimalstohasti ontrol. However, nonlinear
BSDEs and theirtheory have been introduedby Pardoux and Peng [12℄. Ithas been enjoying
a great interest in the last ten year beause of its onnetion with applied elds. We an ite
stohasti ontrol and stohasti games (see [8℄) and mathematial nane (see [6℄). BSDEs
alsoprovidea probabilistiinterpretationforsolutionstoelliptiorparabolinonlinearpartial
dierential equations generalizing the lassial Feynman-Ka formula [13, 14℄. A new lass
∗
augusteaman5yahoo.fr
onsideredby PardouxandPeng[15℄. ThisnewkindofBSDEspresenttwostohastiintegrals
driven by two independent Brownian motions
B
andW
and isof the formY s = ξ +
Z T s
f (r, Y r , Z r ) dr + Z T
s
g(r, Y r , Z r ) ↓ dB r
− Z T
s
Z r dW r , s ∈ [t, T ],
(1.1)where
ξ
is a square integrable variable. Let us remark that in the sens of Pardoux Peng, theintegral driven by
{B r } r≥0
is a bakward It integral and the other one driven by{W r } r≥0
is the standard forward It integral. Further, bakward doubly SDEs seem to be suitable
giving a probabilisti representation for a system of paraboli stohasti partial dierential
equations (SPDEs, in short). We refer to Pardoux and Peng [15℄ for the link between SPDEs
andBDSDEsinthepartiularasewheresolutionsofSPDEsareregular. Thegeneralsituation
ismuh deliatetotreat beauseof diulties ofextending the notionof visosity solutionsto
SPDEs. The stohasti visosity solution for semi-linear SPDEs was introdued for the rst
timeinLionsandSouganidis[10℄. Theyusedtheso-alled"stohastiharateristi"toremove
the stohasti integrals froman SPDE.Another way ofdening astohastivisosity solution
of SPDE is via an appealto the Doss-Sussman transformation. Bukdahn and Ma [3, 4℄ were
the rst to use this approah in order to onnet the stohasti visosity solutions of SPDEs
with BDSDEs.
In this paper we onsider the approah of dening stohasti visosity solution of SPDEs
given by Bukdahn and Ma
[
3,
4]
whih, in our mind is natural and oinide (ifg ≡ 0
) withthe well-knowvisosity solution ofPDEs introduedby Crandallet
al [
5]
. In this fat, we willworkinthesequel ofthis paperwith theversionofbakward doubly SDEsintroduedin[3,4℄,
whih is in fat a time reversal of that onsidered by Pardoux and Peng [15℄. Indeed, for
l, f
beLipshitz ontinuous funtionsintheirspatialvariablesand
g ∈ C b 0,2,3 ([0, T ] ×
IRd ×
IR;
IRd )
,they onsider a lass of bakward doubly SDEs is of this followingform:
Y s t,x = l(X 0 t,x ) + Z s
0
f (r, X r t,x , Y r t,x , Z r t,x ) dr + Z s
0
g(r, X r t,x , Y r t,x ) dB r
− Z s
0
Z r t,x ↓ dW r , s ∈ [0, t].
(1.2)The diusion proess
X t,x
is the unique solutionof the forward SDEX s t,x = x +
Z t s
b(r, X r t,x ) dr + Z t
s
σ(r, X r t,x ) ↓ dW r s ∈ [0, t],
(1.3)where
b
andσ
are some measurable funtions. Here the supersript(t, x)
indiatesthe depen-deneof the solutiononthe initialdate
(t, x)
,and itwillbeomittedwhen the ontext islear.Bukdahn and Ma proved in their two works [3, 4℄,among other things, that
u(t, x) = Y t t,x
isa stohasti visosity solutionof nonlinear paraboliSPDE:
du(t, x) = [Lu(t, x) + f (t, x, u(t, x), (∇uσ)(t, x))] dt +g(t, x, u(t, x)) dB t , (t, x) ∈ (0, T ) ×
IRd , u(0, x) = l(x), x ∈
IRd ,
(1.4)
where
L
dened byL = 1 2
n
X
i,j=1 k
X
l=1
σ il σ lj (x)∂ x 2 i x j +
n
X
j=1
b j (x)∂ x j ,
isthe innitesimaloperator generatedbythe diusion proess
X t,x
. More preisely,they showthank to the Blumenthal
0
-1
law thatu(t, x) =
IEl(X 0 t,x ) + Z t
0
f(r, X r t,x , Y r t,x , Z r t,x ) ds +
Z t 0
g (r, X r t,x , Y r t,x ) dB r | F t B
.
(1.5)Itiswellknowthat
u
is aF B
-measurableeld. However, tothebest ofour knowledge,todate there has been no disussion inthe literature onerning the path regularity of the proessZ
when
f
andl
are only Lipshitz ontinuous, even inthe speial ases wherethe oeientsareenoughregular.
Our goal in this paper is twofold. First we show that if the oeients
l
andf
are on-tinuously dierentiable, then the visosity solution
u
ofthe SPDE(1.4)
willhavea ontinuousspatial gradient
∂ x u
and, more important,the followingprobabilisti representation holds:∂ x u(t, x) =
IEl(X 0 t,x ) N 0 t + Z s
0
f (r, X r t,x , Y r t,x , Z r t,x )N r s dr +
Z t 0
g(r, X r t,x , Y r t,x )N r s dB r | F t B
(1.6)
where
N . s
issomeproess dened on[0, s]
,dependingonlyonthesolutionsofthe forward SDE(1.3)
and itsvariationalequationrespetively. This representation an be thought ofas a new typeofnonlinear Feynman-Kaformulaforthederivativeofu
,whih doesnot seemtoexistintheliterature. Themainsignianeofthe formula, however, liesinthatitdoesnotdependon
the derivativesof the oeients of the bakward doubly BSDE
(1.2)
. Beause of this speialfeature, we an then derive arepresentation
Z s t,x = ∂ x u(t, X s t,x )σ(t, X s t,x ), s ∈ [0, t],
(1.7)underonlyaLipshitzonditionon
l
andf
. Thislatterrepresentationthenenablesustoprove the path regularity of the proessZ
, the seond goal of this paper, even in the ase wherethe terminal value of
Y t,x
is of the forml(X t 0 , ..., X t n )
, whereπ : 0 = t 0 < .. < t n = t
is anypartitionof
[0, t]
, aresult that does not seem tobeamendable by any existing method.Letus reall that this two representations have already be given by Maand Zhang [11℄ in
the ase of a probabilistirepresentation for solutionsof PDEs via BSDEs. Consequently our
approah isbeinspired by their works. However, there are partiularities: rst, the derivative
notionistakeinthe owsense (independentof
ω 1
)beauseu
,the stohasti visosity solutionoftheSPDE
(1.4)
,isarandomeld. Seondly,the prooftotheontinuityoftherepresentation of the proessZ
need, sineF t
s = (F s B ⊗ F s,t W ) 0≤s≤t
is not a ltration,G t
s = (F s B ⊗ F 0,t W ) 0≤s≤t
whihis altration.
The rest of this paper is organized as follows. In setion 2 we give all the neessary pre-
liminaries. In setion 3 we establish the new Feynman-Ka formula between oupled forward
bakward doubly SDE
(1.1)
-(1.3)
and the SPDE(1.4)
, under theC 1
-assumption of the oef- ients. The setion4
is devoted to give the main representation theorem assuming only the Lipshitz ondition of the oeientsl
andf
. In setion5
we study the path regularity oftheproess
Z
.2 Preliminaries
Let
T > 0
a xed time horizon. Throughout this paper{W t , 0 ≤ t ≤ T }
and{B t , 0 ≤ t ≤ T }
willdenote two independentd
-dimensionalBrownian motionsdened onthe omplete probabilityspaes(Ω 1 , F 1 ,
IP1 )
and(Ω 2 , F 2 ,
IP2 )
respetively. Forany proess{U s , 0 ≤ s ≤ T }
dened on
(Ω i , F i ,
IPi ) (i = 1, 2)
,we writeF s,t U = σ(U r − U s , s ≤ r ≤ t)
andF t U = F 0,t U
. Unlessotherwise speied we onsider
Ω = Ω 1 × Ω 2 , F = F 1 ⊗ F 2
and IP=
IP1 ⊗
IP2 .
In addition,we put for eah
t ∈ [0, T ]
,F = {F s = F s B ⊗ F s,T W ∨ N , 0 ≤ s ≤ T }
where
N
is the olletion of IP -null sets. In other words, the olletionF
is IP -omplete but is neither inreasing nor dereasing so that, it is not a ltration. Let us tellalso that randomvariables
ξ(ω 1 ), ω 1 ∈ Ω 1
andζ(ω 2 ), ω 2 ∈ Ω 2
are onsidered as random variables onΩ
via thefollowingidentiation:
ξ(ω 1 , ω 2 ) = ξ(ω 1 ); ζ(ω 1 , ω 2 ) = ζ(ω 2 ).
Let
E
denote a generi Eulidean spae; and regardless of its dimension we denoteh; i
to bethe innerprodut and
|.|
thenorminE
. if anotherEulidean spaesareneeded, weshalllabelthem as
E 1 ; E 2 , ., .,
et. Furthermore, we use the notation∂ t = ∂t ∂ , ∂ x = ( ∂x ∂ 1 , ∂x ∂ 2 , .., ∂x ∂
d )
and∂ 2 = ∂ xx = (∂ x 2 i x j ) d i,j=1
, for(t, x) ∈ [0, T ] ×
IRd
. Note that ifψ = (ψ 1 , .., ψ d ) :
IRd →
IRd
, then∂ x ψ , (∂x j ψ i ) d i,j=1
is amatrix. The meaningof∂ xy , ∂ yy
,et. should belear fromthe ontext.The following spaes willbe used frequently in the sequel (let
X
denotea generi Banahspae):
1. For
t ∈ [0, T ], L 0 ([0, t]; X )
is the spae of allmeasurable funtionsϕ : [0, t] 7→ X
.2. For
0 ≤ t ≤ T, C([0, t]; X )
is the spae of allontinuous funtionsϕ : [0, t] 7→ X
; further,for any
p > 0
we denote|ϕ| ∗,p 0,t = sup
0≤s≤t
kϕ(s)k p X
when the ontextis lear.3. Forany
k, n ≥ 0, C k,n ([0, T ]×E; E 1 )
isthespaeofallE 1
-valuedfuntionsϕ(t, e), (t, e) ∈ [0, T ] × E
, suhthat they arek
-times ontinuously dierentiableint
andn
-times ontin-uously dierentiablein
e
.4.
C b 1 ([0, T ] × E; E 1 )
is the spae of thoseϕ ∈ C 1 ([0, T ] × E; E 1 )
suh that all the partialderivativesare uniformlybounded.
5.
W 1,∞ (E, E 1 )
is the spae of all measurable funtionsψ : E 7→ E 1
, suh that for someonstant
K > 0
it holds that|ψ(x) − ψ(y)| E 1 ≤ K |x − y| E , ∀x, y ∈ E
.6. Forany sub-
σ
-eldG ⊆ F T B
and0 ≤ p < ∞, L p (G; E)
denote allE
-valuedG
-measurable random variableξ
suh that IE|ξ| p < ∞
. Moreover,ξ ∈ L ∞ (G; E)
means it isG
-measurable and bounded.
7. For
0 ≤ p < ∞, L p ( F , [0, T ]; X )
is the spae of allX
-valued,F
-adapted proessesξ
satisfying IE
Z T 0
kξ t k p X dt
< ∞
;andalso,ξ ∈ L ∞ ( F , [0, T ];
IRd )
meansthattheproessξ
is uniformlyessentially bounded in(t, ω)
.8.
C( F , [0, T ] × E; E 1 )
isthespae ofE 1
-valued,ontinuousrandomeldϕ : Ω × [0, T ] × E
,suh thatfor xed
e ∈ E
,ϕ(., ., e)
is anF
-adapted proess.Tosimplifynotationweoftenwrite
C([0, T ]×E ; E 1 ) = C 0,0 ([0, T ]×E ; E 1 )
;andifE 1 =
IR,thenweoftensuppress
E 1
forsimpliity(e.g.,C k,n ([0, T ] ×E;
IR) = C k,n ([0, T ]× E), C k,n ( F , [0, T ]×
E;
IR) = C k,n ( F , [0, T ] × E), ...,
et.). Finally, unless otherwise speied (suh as proessZ
mentionedin Setion1), allvetors inthe paperwill be regarded asolumn vetors.
Throughoutthis paper we shall make use of the followingstanding assumptions:
( A1 )
The funtionsσ ∈ C b 0,1 ([0, T ] ×
IRd ;
IRd×d ), b ∈ C b 0,1 ([0, T ] ×
IRd ;
IRd )
;and allthe partialderivativesof
b
andσ
(with respet tox
) are uniformlybounded by a ommononstantK > 0
. Further, there exists onstantc > 0
, suh thatξ T σ(t, x)σ(t, x) T ξ ≥ c|ξ| 2 , ∀x, ξ ∈
IRd , t ∈ [0, T ].
(2.1)( A2 )
The funtionf ∈ C(F B , [0, T ] ×
IRd ×
IR×
IRd ) ∩ W 1,∞ ([0, T ] ×
IRd ×
IR×
IRd )
andl ∈ W 1,∞ (
IRd )
. Furthermore,we denote the Lipshitz onstants off
andl
by aommonone
K > 0
asin( A1 )
;and weassume thatsup
0≤t≤T
{|b(t, 0)| + |σ(t, 0)| + |f (t, 0, 0, 0)| + |g(0)|} ≤ K.
(2.2)( A3 )
The funtiong ∈ C b 0,2,3 ([0, T ] ×
IRd ×
IR;
IRd )
The following results are either standard or slight variations of the well-know results in
SDE and bakward doubly SDE literature;wegiveonly the statement forready referene.
Lemma 2.1 Suppose that
b ∈ C( F , [0, T ] ×
IRd ;
IRd ) ∩ L 0 ( F , [0, T ]; W 1,∞ (R d ;
IRd )),
σ ∈ C( F , [0, T ] ×
IRd ;
IRd×d ) ∩ L 0 ( F , [0, T ]; W 1,∞ (R d ;
IRd×d ))
, withaommonLipshitzonstantK > 0
. Suppose also thatb(t, 0) ∈ L 2 ( F , [0, T ];
IRd )
andσ(t, 0) ∈ L 2 ( F , [0, T ];
IRd×d )
. LetX
bethe unique solution of the followingforward SDE
X s = x + Z t
s
b(r, X r ) dr + Z t
s
σ(r, X r ) dW r .
(2.3)Then for any
p ≥ 2
, there exists a onstantC > 0
depending only onp, T
andK
, suh thatE(|X| ∗,p 0,t ) ≤ C
|x| p +
IEZ T
0
[|b(s, 0)| p + |σ(s, 0)| p ] ds
(2.4)
Lemma 2.2 Assume
f ∈ C( F , [0, T ] ×
IR×
IRd ) ∩ L 0 ( F , [0, T ]; W 1,∞ (
IR× R d ))
, witha uniformLipshitz onstant
K > 0
, suh thatf (s, 0, 0) ∈ L 2 ( F , [0, T ])
andg ∈ C( F , [0, T ] ×
IR×
IR
d ;
IRd ) ∩ L 0 ( F , [0, T ]; W 1,∞ (
IR× R d ;
IRl ))
with a ommon uniform Lipshitz onstantK > 0
with respet the rst variable and the Lipshitz onstant
0 < α < 1
whih respet the seondvariableandsuhthat
g(s, 0, 0) ∈ L 2 ( F , [0, T ])
. Foranyξ ∈ L 2 (F 0 ;
IR)
,let(Y, Z )
betheadaptedsolution to the BDSDE:
Y s = ξ + Z s
0
f(r, Y r , Z r ) dr + Z s
0
g(r, Y r , Z r ) dB r − Z s
0
Z r ↓ dW r .
(2.5)Thenthere existsa onstant
C > 0
depending only onT
and on theLipshitz onstantsK
andα
, suh thatIE
Z T 0
|Z s | 2 ds ≤ C
IE|ξ| 2 + Z T
0
[|f(s, 0, 0)| 2 + |g(s, 0, 0)| 2 ] ds
.
(2.6)Moreover, for all
p ≥ 2
, there existsa onstantC p > 0
, suh thatIE
(|Y | ∗,p 0,t ) ≤ C p
IE|ξ| p + Z T
0
[|f(s, 0, 0)| p + |g(s, 0, 0)| p ] ds
(2.7)
tothebakward doublySDEs(see Pardoux-Peng
[
15]
). Foranyrandomvariablesξ
oftheformξ = F
Z T 0
ϕ 1 dW t , .., Z T
0
ϕ n dW s ; Z T
0
ψ 1 dB s , ..., Z T
0
ψ p dB s
with
F ∈ C b ∞ (
IRn+p ), ϕ 1 , ..., ϕ n ∈ L 2 ([0, T ],
IRd ), ψ 1 , ..., ψ n ∈ L 2 ([0, T ],
IRd ),
we letD t ξ =
n
X
i=
∂F
∂x i
Z T 0
ϕ 1 dW t , .., Z T
0
ϕ n dW s ; Z T
0
ψ 1 dB s , ..., Z T
0
ψ p dB s
ϕ i (t).
For suh a
ξ
, we deneits1, 2
-norm as:kξk 2 1,2 =
IE|ξ| 2 +
IEZ T
0
|D r ξ| 2 dr
.
S
denoting the set of randomvariable of the above form,we dene the SobolevspaeID
1,2 , S k.k 1,2 .
The "derivation operator"
D .
extends as anoperatorfrom ID1,2
into
L 2 (Ω, L 2 ([0, T ],
IRd ))
.We shall apply the previous antiipativealulusto the oupledforward bakward doubly
SDEs
(1.3)
-(1.2)
. In this fat, letus onsider the following variational equation that will play a important role inthis paper: fori = 1, .., d
,∇ i X s t,x = e i + Z t
s
∂ x b(r, X r t,x )∇ i X r t,x dr +
d
X
j=1
Z t s
∂ x σ j (r, X r t,x )∇ i X r t,x ↓ dW r j ,
∇ i Y s t,x = ∂ x l(X 0 t,x )∇ i X 0 t,x +
Z s 0
[∂ x f(r, Ξ t,x (r))∇ i X r t,x + ∂ y f (r, Ξ t,x (r))∇ i Y r t,x + h∂ z f (r, Ξ t,x (r)), ∇ i Z r t,x i]dr +
Z s 0
[∂ x g(r, Θ t,x (r))∇ i X r t,x + ∂ y g(r, Θ t,x (r))∇ i Y r t,x ]dB r − Z s
0
∇ i Z r t,x ↓ dW r ,
(2.8)where
e i = (0, ..., 1, ..., i 0) T ∈
IRd , Ξ t,x = (Θ t,x , Z t,x ), Θ t,x = (X t,x , Y t,x )
andσ j (.)
is thej
-tholumnofthe matrix
σ(.)
. Wereallagainthat the supersriptiont,x
indiatesthe dependene
of thesolutiononthe initialdate
(t, x)
,and willbeomittedwhenthe ontextislear. We alsoremark that under the above assumptions,
∇X t,x , ∇Y t,x , ∇Z t,x
∈ L 2 ( F ; C([0, T ];
IRd×d ) × C([0, T ];
IRd ) × L 2 ([0, T ];
IRd×d )).
Further the
d × d
-matrix-valued proess∇X t,x
satises a linear SDE and∇X t t,x = I
, so that[∇X s t,x ] −1
exists fors ∈ [0, t],
IP-a.s. and wehave the following:Lemma 2.3 Assume that
( A1 )
holds; and suppose thatf ∈ C b 0,1 ([0, T ] ×
IR2d+1 )
andg ∈ C b 0,2,3 ([0, T ] ×
IRd+1 ;
IRd )
. Then(X, Y, Z) ∈ L 2 ([0, T ];
ID1,2 (
IR2d+1 ))
, and there exists aversion of
(D s X r , D s Y r , D s Z r )
that satises
D s X r = ∇X r (∇X s ) −1 σ(s, X s ) 1 {s≤r} , D s Y r = ∇Y r (∇X s ) −1 σ(s, X s ) 1 {s≤r} , D s Z r = ∇Z r (∇X s ) −1 σ(s, X s ) 1 {s≤r} ,
0 ≤ s, r ≤ t.
(2.9)Lemma 2.4 Suppose that
F ∈
ID1,2
. Then(i)
(Integration byparts formula): for anyu ∈ Dom(δ)
suh thatF u ∈ L 2 ([0, T ] × Ω;
IRd )
, onehas
F u ∈ Dom(δ)
, and it holds thatZ T
0
hF u t , dW t i = δ(F u) = F Z T
0
hu t , dW t i − Z T
0
D t F u t dt;
(ii)
(Clark-Hausman-Oone formula):F =
IE(F ) + Z T
0
IE
{D t F | F t }dW t .
3 Relations to stohasti PDE revisited
Inthissetionweprovetherelation(1.7)betweentheforwardbakward doublySDE
(1.2)
-(1.3)
andthequasi-linearSPDE
(1.4)
,undertheonditionthattheoeientsareonlyontinuously dierentiable. Indeed,sine Bukdahn andMa[
3,
4]
providethat, iff
andl
are onlyLipshitzontinuous, the quantity
u(t, x) = Y t t,x
is a stohasti visosity solution of the quasi-linear SPDE(1.4)
, relation in (1.7) beomes questionable. Our objetive is to ll this gap in the literature and toextend the results ofMa and Zhang[
11]
given inthe ase of the probabilisti interpretation of PDEs viathe BSDEs.Theorem 3.1 Assume
( A1 )
and( A3 )
and suppose thatf ∈ C b 0,1 ([0, T ] ×
IRd ×
IR×
IRd )
andl ∈ C b 1 (
IRd )
. Let(X t,x , Y t,x , Z t,x )
be the adapted solution to the FBDSDE(1.2)
-(1.3)
, and setu(t, x) = Y t t,x
the stohasti visosity of SPDE(1.4)
. Then,(i) ∂ x u(t, x)
exists for all(t, x) ∈ [0, T ] ×
IRd
; and for eah(t, x)
and i=1,...,d, the followingrepresentation holds:
∂ x i u(t, x) =
IE∂ x l(X 0 t,x )∇ i X 0 t,x +
Z t 0
[∂ x f (r, Ξ t,x (r))∇ i X r t,x + ∂ y f (r, Ξ t,x (r))∇ i Y r t,x + ∂ z f(r, Ξ t,x (r))∇ i Z r t,x ]dr +
Z t 0
[∂ x g (r, Θ t,x (r))∇ i X r t,x + ∂ y g(r, Θ t,x (r))∇ i Y r t,x ]dB r | F t B
(3.1)
where
Θ t,x = (X t,x , Y t,x ), Ξ t,x = (Θ t,x , Z t,x )
, and(∇X t,x , ∇Y t,x , ∇Z t,x )
the unique solution ofequation
(2.8)
;(ii) ∂ x u(t, x)
is ontinuous on[0, T ] ×
IRd
;(iii) Z s t,x = ∂ x u(s, X s t,x )σ(s, X s t,x ), ∀ s ∈ [0, t],
IP-a.s.Proof. Forthesimple presentationwetake
d = 1
. Thehigherdimensionalasean betreated inthesamewaywithoutsubstantialdiulty. Weusethesimplernotationsl x , (f x , f y , f z ), (g x , g y , g z )
respetively for the partialderivatives of
l, f
andg
.The proof is inspired by the approah of Ma and Zhang
[
11]
(see Theorem3.1
). Nevertheless, thereexists slightdierenedue inthefatthatthe solutionofSPDE'sisarandomeld; morepreisely willshow thatit is aonditional expetation with respet the ltration
(F t B ) 0≤t≤T
.We rst prove
(i)
. Let(t, x) ∈ [0, T ] ×
IR be xed. Forh 6= 0
,wedene:∇X s h = X s t,x+h − X s t,x
h ; ∇Y s h = Y s t,x+h − Y s t,x
h ; ∇Z s h = Z s t,x+h − Z s t,x
h s ∈ [0, t].
It follows analogouslyof the proof of Theorem 2.1in [15℄) that
IE
{|∆Y h | ∗,2 0,t =
IE{|∇Y h − ∇Y t,x | ∗,2 0,t } → 0 as h → 0.
(3.2)We know also that proesses
Y t,x , Y t,x+h , ∇Y h
and∆Y h
are all adapted to theσ
-algebraF t = (F s t ) 0≤s≤t
, whereF s t = F s B ⊗ F s,t W
. In partiular, sineW
is a Brownian motion on(Ω 2 , F 2 ,
IP2 )
, applying the Blumenthal0
-1
law (see, e.g, [9℄),Y t t,x = u(t, x), Y t t,x+h = u(t, x + h), ∇Y t h = h 1 [u(t, x + h) − u(t, x)]
and∆Y t h
are allindependent of (ora onstant with respet to)ω 2 ∈ Ω 2
. Therefore we onlude from the above that∂ x u
exist, as the random eld and∂ x u(t, x) = ∇Y t t,x
, for all(t, x)
. Finally, taking the onditional expetation on the both sidesof (2.8) at
s = t
,the representation (3.1) hold and nish the prove of(i)
.We nowprove(ii). Let
(t i , x i ) ∈ [0, T ] ×
IR, i = 1, 2
. Knowing thatt 1
andt 2
played inverserolesoneanother, weassumewithoutlosingageneralitythat
t 1 < t 2
. Sine∂ x u
isaonditionalexpetation with respet the ltration
(F s B ) 0≤s≤t
,we have|∂ x u(t 1 , x 1 ) − ∂ x u(t 2 , x 2 )| ≤
IE{A(t 1 , x 1 ) − A(t 2 , x 2 ) | F t B 1 } +
IE{A(t 2 , x 2 ) | F t B 1 } −
IE{A(t 2 , x 2 ) | F t B 2 }
,
(3.3)where
A(t, x) = l x (X 0 t,x )∇X 0 t,x +
Z t 0
[f x (r, Ξ t,x (r))∇X r t,x + f y (r, Ξ t,x (r))∇Y r t,x + f z (r, Ξ t,x (r))∇Z r t,x ] dr +
Z t 0
[g x (r, Θ t,x (r))∇X r t,x + g y (r, Θ t,x (r))∇Y r t,x ]dB r .
(3.4)Thanks tothe quasi-left-ontinuity of
(F s B ) 0≤s≤t
, we see thatt lim 1 ↓t 2
IE(A(t 2 , x 2 ) | F t B 1 ) −
IE(A(t 2 , x 2 ) | F t B 2 )
= 0,
(3.5)independently of
x 2
. In virtue of(3.3)
and(3.5)
),to prove(ii)
itremainto showthatt 1 ↓t lim 2 x 1 →x 2
IE
{A(t 1 , x 1 ) − A(t 2 , x 2 ) | F t B 1 } = 0.
(3.6)Tothis end,sine
A(t, x)
isastohasti proess andinvirtue ofKolmogorov-CentsovTheorem (see [9℄),it sues toshowthatIE
|A(t 1 , x 1 ) − A(t 2 , x 2 )| 2
≤ C(|t 1 − t 2 | 2 + |x 1 − x 2 | 2 ),
what we donow. Reallingthe denition of
A(t i , x i ), i = 1, 2
and denotingG t,x (r) = f x (r, Ξ t,x (r))∇X r t,x + f y (r, Ξ t,x (r))∇Y r t,x + f z (r, Ξ t,x (r))∇Z r t,x
and
H t,x (r) = g x (r, Θ t,x (r))∇X r t,x + g y (r, Θ t,x (r))∇Y r t,x ,
we get
|A(t 1 , x 1 ) − A(t 2 , x 2 )| ≤ |l x (X 0 t 1 ,x 1 )∇X 0 t 1 ,x 1 − l x (X 0 t 2 ,x 2 )∇X 0 t 2 ,x 2 | +
Z t 1
0
|G t 1 ,x 1 (r) − G t 2 ,x 2 (r)|dr +
Z t 1
0
(H t 1 ,x 1 (r) − H t 2 ,x 2 (r))dB r
+
Z t 2
t 1
|G t 2 ,x 2 (r)|dr +
Z t 2
t 1
H t 2 ,x 2 (r)dB r
.
Taking the expetation, it follows by Hölder's and Burkölder-Gundy Davis inequalitiesthat
IE
|A(t 1 , x 1 ) − A(t 2 , x 2 )| 2
≤ C
IE|l x (X 0 t 1 ,x 1 )∇X 0 t 1 ,x 1 − l x (X 0 t 2 ,x 2 )∇X 0 t 2 ,x 2 | 2 +
Z t 1
0
|G t 1 ,x 1 (r) − G t 2 ,x 2 (r)| 2 dr +
Z t 1
0
(H t 1 ,x 1 (r) − H t 2 ,x 2 (r))dB r
2
+ Z t 2
t 1
|G t 2 ,x 2 (r)| 2 dr +
Z t 2
t 1
H t 2 ,x 2 (r)dB r
2 )
≤ C
IE|l x (X 0 t 1 ,x 1 )∇X 0 t 1 ,x 1 − l x (X 0 t 2 ,x 2 )∇X 0 t 2 ,x 2 | 2 +
Z t 1
0
|G t 1 ,x 1 (r) − G t 2 ,x 2 (r)| 2 dr + Z t 1
0
|H t 1 ,x 1 (r) − H t 2 ,x 2 (r)| 2 dr + (t 2 − t 1 )
Z t 2
t 1
|G t 2 ,x 2 (r)| 2 dr + (t 2 − t 1 ) Z t 2
t 1
|H t 2 ,x 2 (r)| 2 dr
.
Bysimilarstandard omputationsinMaand Zhang[11℄ (see proof ofTheorem 3.1), weobtain
IE
|A(t 1 , x 1 ) − A(t 2 , x 2 )| 2
≤ C(|t 2 − t 1 | 2 + |x 2 − x 1 | 2 )
that providethe proof of
(ii)
.It remainstoprove
(iii)
. Foraontinuous funtionϕ
,letusonsider{ϕ ε } ε>0
afamilyofC 0,∞
funtionsthatonvergesto
ϕ
uniformly. Sineb, σ, l, f
arealluniformlyLipshitzontinuous,wemayassumethattherstorderpartialderivativesof
b ε , σ ε , l ε , f ε
arealluniformlybounded,by the orresponding Lipshitz onstants of
b, σ, l, f
uniformlyinε > 0
. Now we onsider thefamily of FBDSDEs parameterized by
ε > 0
:
X s t,x = x + R t
s b ε (r, X r t,x )dr + R t
s σ ε (r, X r t,x ) ↓ dW r ; Y s t,x = l ε (X 0 t,x ) + R s
0 f ε (r, X r t,x , Y r t,x , Z r t,x )dr + R s
0 g(r, X r t,x , Y r t,x )dB r − R s
0 Z r t,x ↓ dW r
(3.7)
and denote it solution by
(X t,x (ε), Y t,x (ε), Z t,x (ε))
. We deneu ε (t, x) = Y t t,x (ε)
. Theorem3.2
of
[
15]
provide thatu ε
is the lassialsolution of stohasti PDEdu ε (t, x) = [L ε u(t, x) + f ε (t, x, u ε (t, x), (∇u ε σ ε )(t, x))] dt +g(t, x, u ε (t, x)) dB t , (t, x) ∈ (0, T ) ×
IRd , u ε (0, x) = l ε (x), x ∈
IRd .
(3.8)
For any
{x ε } ⊂
IRn
suh thatx ε → x
asε → 0
, dene(X ε , Y ε , Z ε ) = (X t,x ε (ε), Y t,x ε (ε), Z t,x ε (ε))
. ThenitiswellknowaordingtheworkofPardouxand Peng [15℄ that
Y s ε = u ε (s, X s ε ); Z s ε = ∂ x u ε (s, X s ε )σ ε (s, X s ε ), ∀ s ∈ [0, t],
IP− a.s.
(3.9)Now by Lemma
2.1
and Lemma2.2
,for allp ≥ 2
ithold thatIE
|X ε − X| ∗,p 0,t + |Y ε − Y | ∗,p 0,t + Z t
0
|Z s ε − Z s | 2 ds
→ 0
(3.10)as
ε → 0
. Moreover let us reall(∇X ε , ∇Y ε , ∇Z ε )
the unique solution of the variational equation of(3.7)
. Using again Lemma2.1
and Lemma2.2
wegetIE
|∇X ε − ∇X| ∗,p 0,t + |∇Y ε − ∇Y | ∗,p 0,t + Z t
0
|∇Z s ε − ∇Z s | 2 ds
→ 0,
(3.11)as
ε → 0
. Thus itis readily seen thatIE
{l ε x (X 0 ε )∇X 0 ε |F t B } →
IE{l x (X 0 )∇X 0 |F t B },
IP -a.s., as
ε → 0
. Furthermore,by the analoguestep used in [11℄, one an show that IEZ t 0
[f x ε (r)∇X r ε + f y ε (r)∇Y r ε + f z ε (r)∇Z r ε ]dr + Z t
0
[g x (r)∇X r ε + g y (r)∇Y r ε ]dB r |F t B
IE
Z t 0
[f x (r)∇X r + f y (r)∇Y r + f z (r)∇Z r ]dr + Z t
0
[g x (r)∇X r + g y (r)∇Y r ]dB r |F t B
IP -a.s., as
ε → 0
. Therefore, we get∂ x u ε (t, x ε ) → ∂ x u(t, x), as ε → 0
IP− a.s.,
for eahxed
(t, x) ∈ [0, T ] ×
IR . Consequently, possibly alonga subsequene, we obtainZ s ε = lim
ε→0 ∂u ε (s, X s ε )σ ε (s, X ε ) = ∂u(s, X s )σ(s, X s ), ds × d
IP− a.e.
Sine forIP
− a.e. ω, ∂ x u(., .)
andX
are both ontinuous, theabove equalitiesatually holdsfor all
s ∈ [0, t]
, IP -a.s., proving(iii)
and end the proof.The following orollary is the diret onsequene of the Theorem
3.1
. The onvention onthe generi onstant
C > 0
stilltrue.Corollary 3.2 AssumethatthesameonditionsasinTheorem
3.1
hold,andlet(X t,x , Y t,x , Z t,x )
be the solution of FBDSDE
(1.2)
-(1.3)
. Then, there existsa onstantC > 0
depending only onK, T,
and for anyp ≥ 1,
a positiveL p (Ω, (F s t ) 0≤s≤t ,
IP)
-proessΓ t,x
, suh that|∂ x u(t, x)| ≤ CΓ t,x t , ∀ (t, x) ∈ [0, T ] ×
IRd ,
IP− a.s.
(3.12)Consequently, one has
|Z s t,x | ≤ CΓ t,x s (1 + |X s t,x |), ∀s ∈ [0, t],
IP− a.s.
(3.13)Furthermore,
∀ p > 1
, there exists a onstantC p > 0
, depending onK, T
, andp
suh thatIE
|X t,x | ∗,p 0,t + |Y t,x | ∗,p 0,t + |Z t,x | ∗,p 0,t ≤ C p (1 + |x| p ).
(3.14)Proof. We assume rst that
p ≥ 2
. The ase1 < p < 2
then follows easilyfrom Hölderinequality. ByLemma
2.1
and Lemma2.2
,we an nd onstantC > 0
suh thatIE
(
|∇X t,x | ∗,p 0,t + |∇Y t,x | ∗,p 0,t + Z T
0
|∇Z r t,x | 2 dr p/2 )
≤ C.
Then, from the identity
(3.1)
, we dedue immediatelythat|∂ x u(t, x)| ≤ CΓ t,x t
, for all(t, x) ∈ [0, T ] ×
IR,whereΓ t,x s =
IE|∇X 0 t,x | + Z s
0
[|∇X r t,x | + |∇Y r t,x | + |∇Z r t,x |]dr +
Z s 0
[∇X r t,x + ∇Y r t,x ]dB r
| F s t
.
Moreover we get for
s ∈ [0, t],
IE(|Γ s t,x | p ) ≤ C
. ThenTheorem3.1 (iii)
implies that|Z s t,x | ≤ CΓ t,x s (1 + |X s t,x |), ∀s ∈ [0, t],
IP -a.s.Now, applyingagain Lemma
2.1
and2.2
and realling(3.13)
we get(3.14)
,forp ≥ 2
.To onlude this setion, we would like to point out that in Theorem
3.1
, the funtionsf
andl
are assumed to be ontinuously dierentiable in all spatial variables with uniformly bounded partial derivatives, whih is muh stronger than standing assumption( A2 )
. Thefollowingtheorem redues the requirementon
f
andl
to only uniformlyLipshitz ontinuous,whihwillbeimportantin our future disussion.
Theorem 3.3 Assume
( A1 )
-( A4 )
, and let(X, Y, Z )
be the solution to the FBDSDE(1.2)
-(1.3)
. Then for allp > 0
, there exists a onstantC p > 0
suhthatIE
|X| ∗,p 0,t + |Y | ∗,p 0,t + ess sup
0≤s≤t
|Z s | p
≤ C p (1 + |x| p ).
(3.15)Proof. In the light of the orollary
3.2
, we need only onsiderp ≥ 2
. By Lemma2.1
andLemma
2.2
it follows that for anyp > 0
there existsC p > 0
suh thatIE
{|X| ∗,p 0,t + |Y | 0,t ∗,p } ≤ C p (1 + |x| p ).
(3.16)Next,bysimilarargumentof Theorem
3.1 (iii)
, weonsider two sequenes ofsmooth funtions{f ε } ε
and{l ε } ε
with theirrst order derivativesin(x, y, z)
uniformlybounded int
andε
suhthat
lim ε→0
( sup
(t,x,y,z)
|f ε (t, x, y, z ) − f (t, x, y, z)| + sup
x
|l ε (x) − l(x)|
)
= 0.
Denoting
(X ε , Y ε , Z ε )
the uniquesolutionofthe orrespondingFBDSDEsand applyingCorol- lary3.2
, we an nd for anyp ≥ 2
a onstantC p > 0
,independent ofε
, suhthatIE
|Z ε | ∗,p 0,t
≤ C p (1 + |x| p ).
(3.17)Furthermore, by
(3.10)
we know that IEZ t
0
|Z s ε − Z s | 2 ds → 0
asε → 0
. Thus, possibly alonga sequene say
(ε n ) n≥1
we havelim n→∞ Z ε n = Z ds × d
IP -a.s. Applying Fatou's lemma andrealling
(3.17)
we the obtainIE
ess sup
0≤s≤t
|Z s | p
≤ C p (1 + |x| p )
whihleads to
(3.15)
,as desired.In this setion we shall prove the rst main theorem of the paper. This theorem an be
regarded as an extension of the nonlinear Feynman-Ka formula obtained by Pardoux-Peng
[
15]
. It gives a probabilisti representation of the gradient (rather than the solution itself) of the stohasti visosity solution, whenever it exists, to a quasi-linear paraboli stohastiPDE. Unlike the ases studied in
(3.1)
, in this setion, our representation does not depend on the partial derivatives of the funtionsf, l
andg
. In this ontext suh representation is the best tool for us to study the path regularity of the proessZ
in the BDSDE with non-smoothoeients. Fornotationalsimpliity,weshalldropthesupersript
t,x
fromthesolution
(X, Y, Z)
of FBDSDE(1.2)
-(1.3)
.To begin with, letusintroduethe two important stohasti integralsthat willplay a key
role in the representation:
M r s = Z s
r
[σ −1 (τ, X τ )∇X τ ] T ↓ dW τ
and
N r s = 1
s − r (M r s ) T [∇X r ] −1 , 0 ≤ r < s ≤ t.
Let usreall that
IE
|M r s | 2p ≤ C p
IEZ s
r
|σ −1 (τ, X τ )∇X τ | 2 dτ p
(4.1)
≤ C p (s − r) p
IE|∇X τ | ∗,2p s,r
≤ C p (s − r) p ,
where
C p > 0
isa generi onstant.An other hand, let us dene the ltration
G t =
F s B ⊗ F t W , 0 ≤ s ≤ t
whih will play aimportantrole in the proof of the ontinuity of the proess
Z
in the BDSDE.Lemma 4.1 For any xed
t ∈ [0, T ]
and anyH ∈ L ∞ ( F t , [0, T ];
IR)
we have(i)
IE| R s
0 1
s−r H r M s r dB r | < +∞
(ii)
for IP.a.e., ω ∈ Ω
, the mappings 7→ R s 0
1
s−r H r (ω)M s r (ω)dB r (ω)
is ontinuous on[0, t]
(iii)
for IP.a.e. , ω ∈ Ω
, themappings 7→
IE{ R s 0
1
s−r H r M s r dB r /G s t }(ω)
isontinuous on[0, t]
Proof. First, for any
0 ≤ τ < s ≤ t
we denoteA s τ =
R s
τ 1
s−r H r M r s dr, 0 ≤ τ < s
0, if s = τ.
(4.2)
Tosimplifynotation,when
τ = 0
wedenoteA s 0 = A s
. Further,letβ
besuhthatα = 1−2β < 1 2
and
β < 1
. Consider the randomvariableM ∗ = sup
0≤t 1 <t 2 ≤t
|M t t 1 2 |
(t 2 − t 1 ) α ;
(4.3)then by
(4.2)
and Theorem2.1
of Revuz-Yor[
16]
,we see that IE[M ∗ ] 2 < +∞
.To prove (i) we note that for any
0 ≤ τ ≤ s ≤ t
by Burkhölder-Gundy- Davis's inequality one hasIE
|A s τ | ≤ C
IEZ s
τ
H r M r s s − r
2
dr
! 1/2
≤ C
IEZ s
τ
|H r | 2
(s − r) 2β . |M r s | 2 (s − r) 2α dr
1/2
≤ C
IEZ s
τ
|H r (s − r) β
2
dr
! 1/2
M ∗
≤ C
IEZ s
τ
dr (s − r) 2β
1/2
kHk ∞ M ∗ = C(s − τ) (1/2)−β
IE(kHk ∞ M ∗ ),
(4.4)where
k.k
denotes thenormofL ∞ ([0, T ])
. AgainlettingC > 0
beagenerionstantdependingonly on
β
andT
,we haveIE
|A s τ | ≤ C{
IEkHk ∞ 2 } 1/2 {
IE(M ∗ ) 2 } 1/2
≤ CkHk L ∞ ([0,T ]×Ω) kM ∗ k L 2 (Ω) < ∞.
(4.5)Setting
τ = 0
in(4.5)
we proved(i)
.To prove
(ii)
letτ = 0
and observe that, in view of(i), A s
is a stohasti integral for0 < s ≤ t
. Consequently, the mappings 7→ A s
is ontinuous on[0,t℄. It remain to prove(iii)
.In this fat, we remark that the right-hand side of the inequality
(4.4)
(withτ = 0
)is learlyin
L 1
; thus we hek easily that the proessA
is uniformly integrable. Therefore, by similar step in Ma and Zhang [11℄ (see proof for(iii)
of Theorem 4.1) itfollows that theG t
-optional projetion ofA
, denotingo A s =
IE(A s |G s t ), s ∈ [0, t]
, has ontinuous path. This prove(iii)
,whene the lemma.
Theorem 4.2 Assume that the assumptions
( A1 )
-( A4 )
hold, and let(X, Y, Z)
be the adaptedsolution to FBDSDE