optimal stopping of expected profit and cost yields in an investment under uncertainty
TRANSCRIPT
This article was downloaded by: [Moskow State Univ Bibliote]On: 31 August 2013, At: 08:47Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registeredoffice: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK
Stochastics An International Journal ofProbability and Stochastic Processes:formerly Stochastics and StochasticsReportsPublication details, including instructions for authors andsubscription information:http://www.tandfonline.com/loi/gssr20
Optimal stopping of expected profitand cost yields in an investment underuncertaintyBoualem Djehiche a , Said Hamadène b & Marie-Amélie Morlais ba Department of Mathematics, The Royal Institute of Technology,SE-100 44, Stockholm, Swedenb Département de Mathématiques, Equipe Statistique et Processus,Université du Maine, Avenue Olivier Messiaen, 72085, Le Mans,Cedex 9, FrancePublished online: 08 Jun 2011.
To cite this article: Boualem Djehiche , Said Hamadne & Marie-Amlie Morlais (2011) Optimalstopping of expected profit and cost yields in an investment under uncertainty, Stochastics AnInternational Journal of Probability and Stochastic Processes: formerly Stochastics and StochasticsReports, 83:4-6, 431-448, DOI: 10.1080/17442508.2010.516828
To link to this article: http://dx.doi.org/10.1080/17442508.2010.516828
PLEASE SCROLL DOWN FOR ARTICLE
Taylor & Francis makes every effort to ensure the accuracy of all the information (the“Content”) contained in the publications on our platform. However, Taylor & Francis,our agents, and our licensors make no representations or warranties whatsoever as tothe accuracy, completeness, or suitability for any purpose of the Content. Any opinionsand views expressed in this publication are the opinions and views of the authors,and are not the views of or endorsed by Taylor & Francis. The accuracy of the Contentshould not be relied upon and should be independently verified with primary sourcesof information. Taylor and Francis shall not be liable for any losses, actions, claims,proceedings, demands, costs, expenses, damages, and other liabilities whatsoever orhowsoever caused arising directly or indirectly in connection with, in relation to or arisingout of the use of the Content.
This article may be used for research, teaching, and private study purposes. Anysubstantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,systematic supply, or distribution in any form to anyone is expressly forbidden. Terms &Conditions of access and use can be found at http://www.tandfonline.com/page/terms-and-conditions
Dow
nloa
ded
by [
Mos
kow
Sta
te U
niv
Bib
liote
] at
08:
47 3
1 A
ugus
t 201
3
Optimal stopping of expected profit and cost yields in an investmentunder uncertainty
Boualem Djehichea*, Said Hamadeneb1 and Marie-Amelie Morlaisb2
aDepartment of Mathematics, The Royal Institute of Technology, SE-100 44 Stockholm, Sweden;bDepartement de Mathematiques, Equipe Statistique et Processus, Universite du Maine,
Avenue Olivier Messiaen, 72085 Le Mans, Cedex 9, France
(Received 29 October 2009; final version received 16 August 2010)
We consider a finite horizon optimal stopping problem related to trade-off strategiesbetween expected profit and cost cash flows of an investment under uncertainty.The optimal problem is first formulated in terms of a system of Snell envelopes for theprofit and cost yields which act as obstacles to each other. We then construct both aminimal solution and a maximal solution using an approximation scheme of theassociated system of reflected backward stochastic differential equations (SDEs).We also address the question of uniqueness of solutions of this system of SDEs. Whenthe dependence of the cash flows on the sources of uncertainty, such as fluctuationmarket prices, assumed to evolve according to a diffusion process, is made explicit, weobtain a connection between these solutions and viscosity solutions of a system ofvariational inequalities with interconnected obstacles.
Keywords: optimal stopping; Snell envelop; backward stochastic differentialequations; merger and acquisition
AMS Subject Classification: 60G40; 93E20; 62P20; 91B99
1. Introduction
The trade-off between the expected profit and cost yields is a central theme in the cash flow
analysis of any investment project or any industry which produces a commodity or provides
services that are subject to uncertainties such as fluctuating market prices or demand and
supply flows (see [1,3,9] and the references therein). The project is profitable when the
expected profit yield is larger than the expected cost yield, a relationship that cannot always
be sustained, due to many sources of uncertainty. Timing of exit from the project based on
optimal trade-off between expected profit and cost yields is thus a crucial decision.
An approach to this problem, which is widely used in portfolio choice with transaction
costs (see [6] and the references therein), is to impose a predetermined form of the cost
yield, to formulate an optimal switching or impulse control problem for the expected profit
yield and to determine exit and re-entry strategies. But in many investment projects subject
to uncertain demand and supply flows, such as merger and acquisition operations, it is
often impossible to fully capture the expected cost yield with a given predetermined
model. It is this situation that we focus on in this paper. Our hope is to give some insight
into the foundations of the complex structure of the type of uncertain cash flows related to,
ISSN 1744-2508 print/ISSN 1744-2516 online
q 2011 Taylor & Francis
http://dx.doi.org/10.1080/17442508.2010.516828
http://www.tandfonline.com
*Corresponding author. Email: [email protected]
Stochastics: An International Journal of Probability and Stochastic Processes
Vol. 83, Nos. 4–6, August–December 2011, 431–448
Dow
nloa
ded
by [
Mos
kow
Sta
te U
niv
Bib
liote
] at
08:
47 3
1 A
ugus
t 201
3
e.g. merger and acquisition operations, where there is almost nothing written on this
topical issue.
In this work, we do not assume any predetermined model for the cost yield. We rather
approach the problem by formulating a finite horizon optimal stopping problem that
involves both the expected profit and cost yields which will act as obstacles to each other.
More precisely, given the profit (resp. cost) c1ðtÞdt (resp. c2ðtÞdt) per unit time dt and the
cost a(t) (resp. profit b(t)) incurred when exiting/abandoning the project, if we let Y 1 and
Y 2 denote the expected profit and the cost yields, respectively, the decision to exit the
project at time t depends on whether Y1t $ Y2
t 2 aðtÞ or Y2t # Y1
t þ bðtÞ. If Ft denotes the
history of the project up to time t, the expected profit yield at time t is expressed in terms of
a Snell envelope as follows:
Y1t ¼ ess supt$tE
ðtt
c1ðsÞ ds þ ðY2t 2 aðtÞÞ1½t,T� þ j11½t¼T�jF t
� �; ð1:1Þ
where the supremum is taken over all exit times t from the project. Moreover, for any
t # T , the random time
t*t ¼ inf s $ t; Y1
s ¼ Y2s 2 aðsÞ
� �^ T ; ð1:2Þ
related to the cost Y 2 2 a incurred when exiting the project should be an optimal time to
abandon the project after t, in which case, we should also get
Y1t ¼ E
ðt*t
t
c1ðsÞ ds þ ðY2t*
t2 aðt*
t ÞÞ1½t*t ,T� þ j11½t*
t ¼T�jF t
" #: ð1:3Þ
In a similar fashion, the expected cost yield at time t reads
Y2t ¼ ess infs$tE
ðst
c2ðsÞ ds þ ðY1s þ bðs ÞÞ1½s,T� þ j21½s¼T�jF t
� �; ð1:4Þ
where the infimum is taken over all exit times s from the project. The random time
s *t ¼ inf s $ t; Y2
s ¼ Y1s þ bðsÞ
� �^ T ð1:5Þ
related to the profit Y 1 þ b incurred when exiting the project should be optimal after t as
well. In this case, we should get
Y2t ¼ E
ðs *t
t
c2ðsÞ ds þ Y1s *
tþ bðs *
t Þ� �
1½s *t ,T� þ j21½s *
t ¼T�jF t
" #: ð1:6Þ
In other words, the cost Y 2 2 a and the profit Y 1 þ b act as obstacles that define the
exit strategy.
The main result of this paper is to show existence of the pair ðY 1; Y 2Þ that solves the
system of Equations (1.1) and (1.4) and also to prove that t* and s * given, respectively,
by (1.2) and (1.5) are optimal strategies for our problem. Using the relation between Snell
envelopes, reflected backward stochastic differential equations (RBSDEs) and variational
inequalities (VI, see [2] for more details), it then follows that solving the system of
Equations (1.1) and (1.4) is equivalent to finding a solution to the following RBSDEs with
B. Djehiche et al.432
Dow
nloa
ded
by [
Mos
kow
Sta
te U
niv
Bib
liote
] at
08:
47 3
1 A
ugus
t 201
3
interconnected obstacles: for all t # T ,
ðSÞ
Y1t ¼ j1 þ
Ð T
tc1ðsÞ ds þ ðK1
T 2 K1t Þ2
Ð T
tZ1
s dBs;
Y2t ¼ j2 þ
Ð T
tc2ðsÞ ds 2 ðK2
T 2 K2t Þ2
Ð T
tZ2
s dBs;
Y1t $ Y2
t 2 aðtÞ and Y2t # Y1
t þ bðtÞ;Ð T
0Y1
s 2 ðY2s 2 aðsÞÞ
� �dK1
s ¼ 0 andÐ T
0ðY1
s þ bðsÞ2 Y2s Þ dK2
s ¼ 0:
8>>>>>><>>>>>>:Using an approximation scheme for systems of RBSDEs, we establish existence of both a
maximal solution and a minimal solution of (S). When the dependence of the cash flows
ðY 1; Y 2Þ on the sources of uncertainty, such as fluctuation market prices, which are
assumed to evolve according to a diffusion process X is made explicit, we also obtain a
connection between the solutions of system (S) and the viscosity solutions of the following
system of VI with interconnected obstacles:
ðVIÞ
min{u1ðt; xÞ2 u2ðt; xÞ þ aðt; xÞ;2›tu1ðt; xÞ2 Lu1ðt; xÞ2 c1ðt; xÞ} ¼ 0;
max{u2ðt; xÞ2 bðt; xÞ2 u1ðt; xÞ;2›tu2ðt; xÞ2 Lu2ðt; xÞ2 c2ðt; xÞ} ¼ 0;
u1ðT ; xÞ ¼ g1ðxÞ; u2ðT ; xÞ ¼ g2ðxÞ:
8>><>>:This paper is organized as follows: Section 2 is devoted to the formulation of the
optimal stopping problem under consideration. In Section 3, we construct a minimal
solution and a maximal solution of (S), using an approximation scheme, where the
minimal solution is obtained as a limit of an increasing sequence of solutions of a system
of RBSDEs, while the maximal one is obtained as a limit of a decreasing sequence of
solutions of another system of RBSDEs. Next, we address the question of uniqueness of
the solution of (S). In general, uniqueness does not hold as it is shown through two
counter-examples. However, we give some sufficient conditions on c1;c2; a and b, for
which a uniqueness result is derived. Finally, in Section 4, we establish a connection
between the solutions of system (S) and viscosity solutions of the system of VI with
interconnected obstacles (VI). We actually show that (VI) admit a solution. Uniqueness
and finer regularity properties of the solutions of (VI) require heavy PDE techniques which
we prefer not to include in this paper but will appear elsewhere.
2. Preliminaries and the main result
In this section, we introduce some basic notions and results concerning RBDSEs, which
will be needed in the subsequent sections.
Throughout this paper, T . 0 denotes an arbitrarily fixed time horizon, and ðV;F ;PÞis a given probability space on which is defined a d-dimensional Brownian motion
B ¼ ðBtÞ0#t#T . We also denote by F ¼ ðF tÞ0#t#T the filtration generated by B and
completed by the P-null sets of F. Throughout the sequel, we always denote the process
restricted to ½0; T� by B and assume that all processes are defined on ½0; T�.
We shall also introduce the following spaces of processes which will be frequently
used in the sequel:
. S2 is the set of all continuous F-adapted processes Y ¼ ðYtÞ such that E½supt[½0;T�
jYtj2� , 1;
. A2 is the subset of S2 of increasing processes ðKtÞt#T with K0 ¼ 0;
. Md;2 denotes the set of F-adapted and d-dimensional processes Z such that
EðÐ T
0jZsj
2dsÞ , 1.
Stochastics: An International Journal of Probability and Stochastic Processes 433
Dow
nloa
ded
by [
Mos
kow
Sta
te U
niv
Bib
liote
] at
08:
47 3
1 A
ugus
t 201
3
The following results on RBSDEs are by now well known. For a proof, the reader is
referred to [5]. A solution for the RBSDE associated with a triple (f ; j; S), where f :
ðt;v; y; zÞ 7! f ðt;v; y; zÞ (R-valued) is the generator, j is the terminal condition and
S U ðStÞt#T is the lower barrier, is a triple ðYt; Zt;KtÞ0#t#T of F-adapted stochastic
processes that satisfies
Y [ S2; K [ A2 and Z [ Md;2;
Yt ¼ jþÐ T
tf ðs;v; Ys; ZsÞds þ ðKT 2 KtÞ2
Ð T
tZs dBs;
Yt $ St; 0 # t # T ;Ð T
0ðSt 2 YtÞ dKt ¼ 0:
8>>>>><>>>>>:ð2:7Þ
The RBSDE(f ; j; S) is said standard if the following conditions are satisfied:
(A1): The generator f is Lipschitz with respect to ðy; zÞ uniformly in ðt;vÞ.(A2): The process ðf ðt;v; 0; 0; 0ÞÞ0#t#T is F-progressively measurable and dt ^ dP-
square integrable.
(A3): The random variable j is in L2 V;F T ;P� �
.
(A4): The barrier S is continuous F-adapted and satisfies E½sup0#s#T jS
þs j
2� , 1 and
ST # j, P-a.s.
Theorem 2.1. (See [5]) Let the coefficients ðf ; j; SÞ satisfy Assumptions (A1)–(A4). Then
the RBSDE (2.7) associated with ðf ; j; SÞ has a unique F-progressively measurable
solution (Y ; Z;K) which belongs to S2 £Md;2 £A2. Moreover, process Y enjoys the
following representation property as a Snell envelope: for all t # T ,
Y1t ¼ ess supt$tE
ðtt
f ðs; Y1s ; Z1
s Þds þ St1½t,T� þ j11½t¼T�jF t
� �: ð2:8Þ
The proof of Theorem 2.1 is related to the following, by now standard, estimates and
comparison results for RBSDEs. For the proof, see Proposition 3.5 and Theorem 4.1 in [5].
Lemma 2.1. Let ðY ; Z;KÞ be a solution of the RBSDE ðf ; j; SÞ. Then there exists a constant
C depending only on the time horizon T and on the Lipschitz constant of f such that
E sup0#t#T
jYtj2þ
ðT
0
jZsj2
ds þ jKT j2
# CE
ðT
0
j f ðs; 0; 0Þj2
ds þ jjj2þ sup
0#t#T
jSþt j
2
:
ð2:9Þ
Lemma 2.2. (Comparison of solutions) Assume that ðY ; Z;KÞ and ðY 0; Z 0;K 0Þ are solutions
of the RBSDEs associated with (f ; j; S) and ðf 0; j0; S0Þ, respectively, where only one of the
two generators f or f 0 is assumed to be Lipschitz continuous. If
. j # j 0, P-a.s.,
. f ðt; y; zÞ # f 0ðt; y; zÞ; dP^ dt-a.s. and for all (y; z),
. P-a.s., for all t # T , St # S0t,
then
P-a:s: ;t # T ; Yt # Y 0t: ð2:10Þ
The previous results are also valid for RBSDE with upper barriers. Indeed, if ðY; Z;KÞ
solves the RBSDE associated with ðf ; j;UÞ with upper barrier equal to U, then
B. Djehiche et al.434
Dow
nloa
ded
by [
Mos
kow
Sta
te U
niv
Bib
liote
] at
08:
47 3
1 A
ugus
t 201
3
(2Y ;2Z;K) solves the RBSDE associated with ð~f;2j; SÞ with parameters given by
S ¼ 2U and ~fðs; y; zÞ ¼ 2f ðs;2y;2zÞ.
The first objective in this paper is to study existence of solutions of the coupled system
of RBSDEs (S). Let us introduce the following assumptions:
(B1): For each i ¼ 1; 2, the mappings ðt;v; y; zÞ 7! ciðt;v; y; zÞ are Lipschitz in ðy; zÞ
uniformly in ðt;vÞ meaning that there exists C . 0 such that
jciðt;v; y; zÞ2 ciðt;v; y0; z0Þj # Cðjy 2 y0j þ jz 2 z0jÞ; for all t; y; z; y0; z0:
Moreover, the processes ðc iðt; 0; 0; 0ÞÞ0#t#T are F-progressively measurable and
dt^ dP -square integrable.
(B2): The obstacles ðaðt;vÞÞ0#t#T and ðbðt;vÞÞ0#t#T belong to S2.
(B3): The random variables j 1 and j 2 are FT -measurable and square integrable.
Moreover, we assume that P-a:s:; j1 2 j2 $ max{ 2 aðTÞ;2bðTÞ}.
From now on, we will also make use of either one of the two following assumptions:
(B4): The process ðbðtÞÞ0#t#T is of Ito type, i.e. for any t # T .
bðtÞ ¼ bð0Þ þ
ðt
0
U2s ds þ
ðt
0
V2s dBs; ð2:11Þ
for some F-progressively measurable processes U 2 and V 2 which are dt^ dP-
integrable and square integrable, respectively.
(B40): The process ðaðtÞÞ0#t#T is of Ito type, i.e. for any t # T ,
aðtÞ ¼ að0Þ þ
ðt
0
U1s ds þ
ðt
0
V1s dBs; ð2:12Þ
for some F-progressively measurable processes U 1 and V 1 which are dt^ dP-
integrable and square integrable, respectively
Remark 2.1. Assumption (B4) is required to prove the continuity of the minimal solution,
which is obtained by using an increasing approximation scheme, whereas Assumption
(B40) is required to get the continuity of the maximal solution.
Let us now make precise: on the one hand, the notion of a solution and, on the other
hand, the notions of minimal and maximal solutions of system (S).
Definition 2.1. A 6-tuple of processes ðY 1; Z 1; K 1; Y 2; Z 2; K 2Þ is a solution of system
(S) if the two triples ðY 1; Z 1; K 1Þ and ðY 2; Z 2; K 2Þ belong to S2 £Md;2 £A2 and if it
satisfies (S).
The process ðY 1; Z 1; K 1; Y 2; Z 2; K 2Þ is a minimal solution of the system (S) if it is a
solution of (S) and if whenever another 6-tuple of processes ð ~Y1; ~Z1; ~K1; ~Y2; ~Z2; ~K2Þ is a
solution of (S), then
P-a:s: ;t # T ; ~Y1
t $ Y1t and Y2
t $ Y2t ;
whereas it is a maximal solution (S) if
P-a:s: ;t # T; ~Y1
t # Y1t and ~Y
2
t # Y2t :
The following theorems, related to existence of minimal and maximal solutions of S,
are the main results of this paper.
Stochastics: An International Journal of Probability and Stochastic Processes 435
Dow
nloa
ded
by [
Mos
kow
Sta
te U
niv
Bib
liote
] at
08:
47 3
1 A
ugus
t 201
3
Theorem 2.2. Assume that the data ðc1;c2; j1; j2; a; bÞ satisfy Assumptions (B1)–(B4).
Then system (S) of RBSDEs associated with ðc1;c2; j1; j2; a; bÞ admits a minimal
solution ðY 1; Z 1;K 1; Y 2; Z 2;K 2Þ.
Theorem 2.3. Suppose that the data ðc1;c2; j1; j2; a; bÞ satisfy Assumptions (B1)–(B3)
and, in addition, Assumption (B40) on the process ðaðtÞÞt#T . Then, system (S) of RBSDEs
associated with ðc1;c2; j1; j2; a; bÞ admits a maximal solution ðY 1; Z 1;K 1; Y 2; Z 2;K 2Þ.
The proof of Theorem 2.3 can be obtained from Theorem 2.2 by considering the
minimal solution of the system associated with ð2c1ðt;v;2y;2zÞ;2c2ðt;v;2y;2zÞ;2j1;2j2;2a;2bÞ.
Section 3 is devoted to the proof of Theorem 2.2.
3. Proof of Theorem 2.2
Step 1: Construction of the sequences and properties.
We first introduce two increasing approximation schemes ðY 1;n; Z 1;n;K 1;nÞ and
ðY 2;n; Z 2;n;K 2;nÞ that converge to the minimal solution of (S).
Consider the following BSDEs defined recursively, for any n $ 1, by
ðY 1;0; Z 1;0Þ [ S2 £Md;2;
Y1;0t ¼ j1 þ
Ð T
tc1ðs; Y1;0
s ; Z1;0s Þds 2
Ð T
tZ1;0
s dBs; t # T;
8<: ð3:13Þ
and for n $ 0 and any t # T ,
ðSnÞ
Y2;nþ1t ¼ j2 þ
Ð T
tc2ðs;Y
2;nþ1s ;Z2;nþ1
s Þds 2 ðK2;nþ1T 2 K
2;nþ1t Þ2
Ð T
tZ2;nþ1
s dBs;
Y2;nþ1t # Y
1;nt þ bðtÞ;
Y1;nþ1t ¼ j1 þ
Ð T
tc1ðs;Y
1;nþ1s ;Z1;nþ1
s Þds þ ðK1;nþ1T 2 K
1;nþ1t Þ2
Ð T
tZ1;nþ1
s dBs;
Y1;nþ1t $ Y
2;nþ1t 2 aðtÞ;Ð T
0ðY1;nþ1
s 2 ðY2;nþ1s 2 aðsÞÞdK1;nþ1
s ¼ 0 andÐ T
0ðY1;n
s þ bðsÞ2 Y2;nþ1s ÞdK2;nþ1
s ¼ 0:
8>>>>>>>>><>>>>>>>>>:In view of Assumptions (B1)–(B4), it is easily shown by induction that for any n $ 1,
the triples ðY 1;n; Z 1;n;K 1;nÞ and ðY 2;n; Z 2;n;K 2;nÞ are well defined and belong to the space
S2 £Md;2 £A2, because the pair of processes ðY 1;0; Z 1;0Þ solution of (3.13) exists.
Additionally, by the Comparison Lemma 2.2, we have P-a.s., for all t # T , Y1;0t # Y
1;1t
because the process K 1;1 is increasing and then K1;1T 2 K
1;1t $ 0;;t # T . Next, using
once more the comparison result, we obtain Y 2;1 # Y 2;2: Finally, an induction argument
leads to
;n $ 0; P-a:s:; for all t # T ; Y1;nt # Y1;nþ1
t and Y2;nþ1t # Y2;nþ2
t :
Next, let us consider the following standard BSDE:
�Y2 [ S2 and �Z [ Md;2;
�Y2t ¼ j2 þ
Ð T
tc2ðs; �Y
2s ; �Z
2s Þds 2
Ð T
t�Z
2s dBs; t # T :
8<:
B. Djehiche et al.436
Dow
nloa
ded
by [
Mos
kow
Sta
te U
niv
Bib
liote
] at
08:
47 3
1 A
ugus
t 201
3
The solution of this equation exists (see, e.g. [7]). Furthermore, since the process K 2;n
is non-decreasing then using standard comparison theorem for BSDEs (see, e.g. [7])
we obtain
P-a:s:; ;t # T; ;n Y2;nt # �Yt: ð3:14Þ
Finally, let ð ~Y; ~Z; ~KÞ be the solution of the following RBSDE associated with ðc1; j1; �YÞ,i.e. for any t # T ,
ð ~Y; ~Z; ~KÞ [ S2 £Md;2 £ S2;
~Yt ¼ j1 þÐ T
tc1ðs; ~Ys; ~ZsÞds þ ð ~KT 2 ~KtÞ2
Ð T
t~ZsdBs;
~Yt $ �Yt 2 aðtÞ;Ð T
0ð ~Ys 2 ð �Ys 2 aðsÞÞd ~Ks ¼ 0:
8>>>>><>>>>>:Again using the Comparison Lemma 2.2 and relying on (3.14), we have
P-a:s:; ;t # T ; ;n; Y1;nt # ~Yt: ð3:15Þ
Therefore, from (3.14) and (3.15), it follows that
E supn$0
sup0#t#T
ðjY1;nt j þ jY2;n
t jÞ2� �
, 1: ð3:16Þ
Moreover, using the estimates given in Lemma 2.1 for standard RBSDEs, there exists a
real constant C $ 0 such that for all n $ 0,
E
ðT
0
jZ1;ns j þ jZ2;n
s j� �2
ds
� �þ E ðK
1;nT Þ2 þ ðK
2;nT Þ2
h i# C: ð3:17Þ
Let Y 1 and Y 2 be two optional processes defined, for all t # T , by
Y1t ¼ lim
n!1Y1;n
t and Y2t ¼ lim
n!1Y2;n
t :
Step 2: Existence of a solution for (S).
Since the processes b and Y 1;n are of Ito type, then thanks to a result by El Karoui et al.
([5], Proposition 4.2, p. 713), the process K 2;n is absolutely continuous w.r.t. t. Moreover,
we have, for all t # T ,
dK2;nt # 1
½Y2;nþ1t ¼Y
1;nt þbt�
c2ðt; Y1;nt þ bt; Z1;n
t þ V2t Þ þ U2
t þ c1ðt; Y1;nt ; Z1;n
t Þ� �þ
dt:
Hence, by (B1) and (B4), there exists a constant C $ 0 such that, for all n $ 1,
E
ðT
0
dK2;nt
dt
!2
dt
24 35 # C:
In view of this estimate together with (3.17), there exists a subsequence along which
ððdK2;nt =dtÞ0#t#T Þn$1, ððc2ðt; Y
2;nþ1t ; Z
2;nþ1t ÞÞ0#t#T Þn$1 and ððZ
2;nþ1t Þ0#t#T Þn$1 converge
weakly in their respective spaces to the processes ðk2t Þt#T , ðw2ðtÞÞt#T and ðZ2
t Þt#T which
also belong to M1;2, M1;2 and Md;2, respectively.
Next, for any n $ 0 and any stopping time t, we have
Y2;nþ1t ¼ Y
2;nþ10 2
ðt0
c2ðs; Y2;nþ1s ; Z2;nþ1
s Þds þ K2;nþ1t þ
ðt0
Z2;nþ1s dBs:
Stochastics: An International Journal of Probability and Stochastic Processes 437
Dow
nloa
ded
by [
Mos
kow
Sta
te U
niv
Bib
liote
] at
08:
47 3
1 A
ugus
t 201
3
Taking the weak limits on each side and along this subsequence yields
Y2t ¼ Y2
0 2
ðt0
w2ðsÞds þ
ðt0
k2s ds þ
ðt0
Z2s dBs; P-a:s:
Since the processes appearing in each side are optional, using the optional section theorem
(see, e.g. [2], Chapter IV, p. 220), it follows that
P-a:s:; ;t # T ; Y2t ¼ Y2
0 2
ðt
0
w2ðsÞds þ
ðt
0
k2s ds þ
ðt
0
Z2s dBs: ð3:18Þ
Therefore, the process Y 2 is continuous. Relying on both Dini’s Theorem and Lebesgue’s
dominated convergence one, we also get that
limn!1
E supt#T
jY2;nt 2 Y2
t j2
� �¼ 0:
We will now focus on the convergence of ðY 1;nÞn$0. Using estimates (3.16) and (3.17) and
then applying Peng’s monotone limit theorem (see [8]) to the sequence (Y 1;n;Z 1;n;K 1;n),
we get that Y 1 is cadlag. Moreover, there exist an F-adapted cadlag non-decreasing
process K1 and a process Z1 of Md;2 such that ðZ 1;nÞn$0 converges to Z 1 in Lpðdt^ dPÞ for
any p [ ½1; 2Þ. Moreover, for any stopping time t, the sequence ðK1;nt Þn$1 converges
weakly to K1t in L2ðV;F t; dPÞ. Relying now on the Snell envelope representation (see [5],
Proposition 2.3, p. 705), we have, for any n $ 1 and t # T ,
Y1;nþ1t ¼ ess supt$tE
ðtt
c1ðs; Y1;nþ1s ; Z1;nþ1
s Þds þ ðY2;nþ1t 2 aðtÞÞ1½t,T� þ j11½t¼T�jF t
� �:
ð3:19Þ
But for any n $ 0 and t # T ,
E ess supt$tE
ðtt
c1ðs; Y1;nþ1s ; Z1;nþ1
s Þds þ ðY2;nþ1t 2 aðtÞÞ1½t,T� þ j11½t¼T�jF t
� ������2ess supt$tE
ðtt
c1ðs; Y1s ; Z1
s Þds þ ðY2t 2 aðtÞÞ1½t,T� þ j11½t¼T�jF t
� ����� # E ess supt$t E
ðtt
ðc1ðs; Y1;nþ1s ; Z1;nþ1
s Þ2 c1ðs; Y1s ; Z1
s ÞÞds
������þðY2;nþ1
t 2 Y2tÞÞ1½t,T�jF t
����# E E
ðT
0
c1ðs; Y1;nþ1s ; Z1;nþ1
s Þ2 c1ðs; Y1s ; Z1
s Þ�� ��ds þ supt#T jY
2;nþ1t 2 Y2
t kF t
� �� :
As ðY 2;nÞn$1 converges uniformly (with respect to t) in L2ðdPÞ to Y 2 (actually the
convergence holds in S2), ðY 1;nÞn$0 converges to Y 1 in L2ðdt^ dPÞ, thanks to (3.16) and
ðZ 1;nÞn$0 converges to Z 1 in Lpðdt^ dPÞ ð1 # p , 2Þ, then the right-hand side of the
previous inequality converges to 0 as n !1. But for any t # T , once more thanks to
(3.16), ðY1;nt Þn$0 converges to Y1
t in L2ðdPÞ, therefore, from equality (3.19), we deduce
that for any t # T:
Y1t ¼ ess supt$tE
ðtt
c1ðs;Y1s ;Z
1s Þds þ ðY2
t 2 aðtÞÞ1½t,T� þ j11½t¼T�jF t
� �; P-a:s: ð3:20Þ
B. Djehiche et al.438
Dow
nloa
ded
by [
Mos
kow
Sta
te U
niv
Bib
liote
] at
08:
47 3
1 A
ugus
t 201
3
As Y 1 and the right-hand side of the latter equality are cadlag, then those processes are
indistinguishable, i.e. P-a.s., for all t # T , equality (3.20) holds.
Next, the process ðY2t 2 aðtÞÞt#T is continuous and Y2
T 2 aðTÞ ¼ j2 2 aðTÞ # j1
through assumption (B3); therefore, using Theorem 2.1, we get that the right-hand side of
(3.20) is continuous.
Henceforth, Y 1 is continuous and, using Dini’s theorem, the convergence of ðY 1;nÞn$1
to Y 1 holds in S2. Relying next on the Doob–Meyer decomposition of the supermartingale
ðY1t þ
Ð t
0c1ðs;Y
1s ;Z
1s ÞdsÞt#T (see also Theorem 2.1), there exist Z 1 and K 1 such that, for all
t # T ,
Y1t ¼ j1 þ
Ð T
tc1ðs; Y1
s ; Z1s Þds þ ðK1
T 2 K1t Þ2
Ð T
tZ1
s dBs;
Y1t $ Y2
t 2 aðtÞ andÐ T
0Y1
s 2 ðY2s 2 aðsÞÞ
� �dK1
s ¼ 0:
8<:Since the convergence of ðY 1;nÞn$1 to Y 1 holds in S2, we can now rely on standard
arguments and, in particular, on Ito’s formula applied to ðY 2;n 2 Y 2;mÞ2 (m; n $ 0) to
claim that ðZ 2;nÞn$1 is a Cauchy sequence, and, therefore, that converges to Z 2 in Md;2.
Using this and taking into account the decomposition obtained in (3.18), we finally get, for
any t # T ,
Y2t ¼ j2 þ
Ð T
tc2ðs; Y2
s ; Z2s Þds 2
Ð T
tk2
s ds 2Ð T
tZ2
s dBs;
Y2t # Y1
t þ bðtÞ:
8<:Due to the weak convergence of ððdK
2;nt =dtÞt#T Þn$1 to the process k 2 and the strong
convergence of (Y 1;n) and ðY 2;nÞ in S2, it follows that
0 ¼
ðT
0
ðY1;ns þ bðsÞ2 Y2;nþ1
s ÞdK2;nþ1s !
ðT
0
ðY1s þ bðsÞ2 Y2
s Þk2s ds ¼ 0;
which implies that ðY 2; Z 2;K 2 UÐ :
0k2
s dsÞ is solution for the second part of ðSÞ, and
henceforth, the 6-tuple ðY 1; Z 1;K 1; Y 2; Z 2;K 2Þ is a solution of (S).
This solution is actually a minimal one. Indeed, if there is another one
ðY1; Z1;K1; Y2; Z2;K2Þ then, by comparison, we obviously get Y1 $ Y 1;0 and then
Y2 $ Y 2;1. Finally, by induction we have, for any n $ 1, Y1 $ Y 1;n and then Y2 $ Y 2;n,
which implies the desired result after taking the limit as n goes to 1.
3.1 On the uniqueness of the solution of system (S)As shown in Section 3.2, in general, we do not have uniqueness of the solution of (S).
However, in some specific cases, such as in the following result, uniqueness holds.
Theorem 3.1. Assume that
(i) the mappings c1 and c2 do not depend on ðy; zÞ, i.e. ci U ðciðt;vÞÞ, i ¼ 1; 2
(ii) the barriers a and b satisfy
P-a:s:
ðT
0
1½aðsÞ¼bðsÞ�ds ¼ 0: ð3:21Þ
Then, the solution of (S) is unique.
Stochastics: An International Journal of Probability and Stochastic Processes 439
Dow
nloa
ded
by [
Mos
kow
Sta
te U
niv
Bib
liote
] at
08:
47 3
1 A
ugus
t 201
3
Proof. The proof relies on the uniqueness of the solution of a RBSDE with one lower
barrier. Indeed, let ðY 1; Z 1;K 1; Y 2; Z 2;K 2Þ be a solution of (S) and, for t # T , let us set
Yt ¼ Y1t 2 Y2
t , Zt ¼ Z1t 2 Z2
t and Kt ¼ K1t þ K2
t . Therefore, the triple ðY ; Z;KÞ belongs to
S2 £Md;2 £A2. Moreover, for any t # T , it satisfies
Yt ¼ j1 2 j2 þÐ T
t{c1ðu;vÞ2 c2ðu;vÞ}du þ ðKT 2 KtÞ2
Ð T
tZudBu;
Yt $ max{ 2 aðtÞ;2bðtÞ} andÐ T
0Ys þ min{aðsÞ; bðsÞ}� �
dKs ¼ 0:
8<: ð3:22Þ
Indeed, the two first relations being obvious, it only remains to show the third one. ButðT
0
Ys þ min{aðsÞ; bðsÞ}� �
dKs ¼
ðT
0
Y1s 2 Y2
s þ aðsÞ� �
1½aðsÞ#bðsÞ�dðK1s þ K2
s Þ
þ
ðT
0
Y1s 2 Y2
s þ bðsÞ� �
1½aðsÞ.bðsÞ�dðK1s þ K2
s Þ: ð3:23Þ
However, ðT
0
Y1s 2 Y2
s þ aðsÞ� �
1½aðsÞ#bðsÞ�dK1s ¼ 0;
because, for any t # T , dK1t ¼ 1½Y1
t 2Y2t þaðtÞ¼0�dK1
t . On the other hand,
0 #
ðT
0
Y1s 2 Y2
s þ aðsÞ� �
1½aðsÞ#bðsÞ�dK2s #
ðT
0
Y1s 2 Y2
s þ bðsÞ� �
1½aðsÞ#bðsÞ�dK2s ¼ 0;
because for any t # T , it holds that dK2t ¼ 1½Y2
t 2Y1t þbðtÞ¼0�dK2
t : In the same way, one can
show that the second term in (3.23) is null, and therefore, the third relation in (3.22) holds
true. It follows that ðY; Z;KÞ is a solution for the one lower barrier RBSDE associated with
ðc1 2 c2; j1 2 j2;max{ 2 aðtÞ;2bðtÞ}Þ. As the solution of the latter equation is unique by
Theorem 2.1, then for any solution ðY 1; Z 1;K 1; Y 2; Z 2;K 2Þ of (S), the differences Y 1 2
Y 2 and Z 1 2 Z 2 and the increasing process K 1 þ K 2 are unique.
Next, let us express K 1 and K 2 in terms of K. For any t # T , we claim that
K1t ¼
ðt
0
1½Y1s2Y2
sþaðsÞ¼0�dK1s
¼
ðt
0
1½Y1s2Y2
sþaðsÞ¼0�1½aðsÞ,bðsÞ� dK1s þ
ðt
0
1½Y1s2Y2
sþaðsÞ¼0�1½aðsÞ.bðsÞ� dK1s ; ð3:24Þ
where to get this second equality, we make use of the both absolute continuity of dK 1, the
increasing property of K 1, and condition (ii) on the barriers to argue thatðT
0
1½Y1s2Y2
sþaðsÞ�1½aðsÞ¼bðsÞ� dK1s ¼
ðT
0
ds1½Y1s2Y2
sþaðsÞ�1½aðsÞ¼bðsÞ�
dK1s
ds
¼ 0:
On the other hand,ðT
0
1½Y1s2Y2
sþaðsÞ¼0�1½aðsÞ,bðsÞ� dK2s ¼
ðT
0
1½Y1s2Y2
sþaðsÞ¼0�1½aðsÞ,bðsÞ�1½Y1s2Y2
sþbðsÞ¼0� dK2s ¼ 0:
In a similar fashion, we haveðT
0
1½Y1s2Y2
sþaðsÞ¼0�1½aðsÞ.bðsÞ� dK2s ¼ 0:
B. Djehiche et al.440
Dow
nloa
ded
by [
Mos
kow
Sta
te U
niv
Bib
liote
] at
08:
47 3
1 A
ugus
t 201
3
Therefore, going back to (3.24), we obtain, for all t # T ,
K1t ¼
ðt
0
1½Y1s2Y2
sþaðsÞ¼0�dKs;
which implies that K 1 is unique and then so is K 2. Next, writing the equations satisfied by
Y 1 and Y 2 and taking the conditional expectation (w.r.t. F t), we obtain their uniqueness.
From standard arguments, uniqueness of Z 1 and Z 2 follows immediately. A
Remark 3.1. This uniqueness result can be slightly generalized to generators of the
following forms:
c1ðt;v; y; zÞ ¼ ~c1ðt;vÞ þ aty þ btz and c2ðt;v; y; zÞ ¼ ~c2ðt;vÞ þ aty þ btz;
where a and b are P-progressively measurable bounded processes taking their values in R
and Rd, respectively. The proof is the same as the previous one, noting that the process
Y ¼ Y 1 2 Y 2 solves a linear RBSDE, which is explicitly solvable.
For the sake of completeness, we now consider the more general case, i.e. when
condition (3.21) on the barriers is no more satisfied.
Lemma 3.1. Under (B1)–(B3) together with (B4) or (B40), if condition (3.21) on the
barriers is no more satisfied, then on any interval ½a;b� of ½0; T�, where a ; b, uniqueness
of a solution for system (S) holds only in the trivial cases where neither the first component
Y 1 nor the second component Y 2 is reflected processes.
Proof. To prove this, let us consider a non-trivial interval ½a;b� with 0 , a , b , T and
where aðtÞ ¼ bðtÞ; ;t [ ½a;b� and assume that uniqueness of the solution of (S) holds. So
let ðY1; Z1;K1; Y2; Z2;K2Þ be the minimal solution of (S) which is then equal to the
maximal one by uniqueness. Next, let us consider the following system of RBSDEs on the
time interval ½a;b�: for any s [ ½a;b�,
ðSminÞ
dY1s ¼ 2c1ðs; Y1
s ; Z1s Þds þ Z1
s dBs; Y1b ¼ Y1
b;
dY2s ¼ 2c2ðs; Y2
s ; Z2s Þds þ Z2
s dBs þ dK2s ; Y2
b ¼ Y2b;
Y2s # Y1
s þ bðsÞ;Ð ba
Y1s þ bðsÞ2 Y2
s
� �dK2
s ¼ 0:
8>>><>>>:Note also that both existence and uniqueness for ðSminÞ on ½a;b� result from standard
results for BSDEs (or RBSDEs with one barrier).
On the other hand, let us consider the following system of RBSDEs which is similar to
(S) but on the time interval ½0;a� and which actually has a solution. For any t [ ½0;a�,
g1t ¼ Y1
a þÐ a
tc1ðs; g
1s ; u
1s Þds þ ðz1
a 2 z1t Þ2
Ð atu1
s dBs;
g2t ¼ Y2
a þÐ a
tc2ðs; g
2s ; u
2s Þds þ ðz2
a 2 z2t Þ2
Ð atu2
s dBs;
g1t $ g2
t 2 aðtÞ and g2t # g1
t þ bðtÞ;Ð a0g1
s 2 ðg2s 2 aðsÞÞ
� �dz1
s ¼ 0 andÐ a
0ðg1
s þ bðsÞ2 g2s Þdz
2s ¼ 0;
8>>>>><>>>>>:where both Y1
a and Y2a stand for the value at time t ¼ a of the first component Y 1 (and the
second component Y 2) of the unique solution to system (Smin) defined on ½a;b�. Therefore,
Stochastics: An International Journal of Probability and Stochastic Processes 441
Dow
nloa
ded
by [
Mos
kow
Sta
te U
niv
Bib
liote
] at
08:
47 3
1 A
ugus
t 201
3
the following process
ðg1t ; u
1t ; z
1t ; g
2t ; u
2t ; z
2t Þ1½t#a� þ ðY1
t ; Z1t ; z
1a; Y2
t ; Z2t ;K2
t 2 K2a þ z2
aÞ1½a,t#b�
�þðY1
t ; Z1t ;K1
t 2 K1b þ z1
a;Y2t ;Z
2t ;K2
t 2 K2b þ K2
b 2 K2a þ z2
aÞ1½b,t#T�
ot#T
;
obtained by concatenation, is also a solution for (S). Using once more uniqueness for (S)
yields for any t [ ½a;b�
Y1t ¼ Y1
t and Z1t ¼ Z1
t :
It implies that process Y1 is not reflected on the time interval ½a;b�. Now going back to
system ðSminÞ and making the reflection on Y 2 and not on Y 1, we define a new system
denoted by (Smax), and, therefore, proceeding as above, we obtain that the solution
ð �Y1; �Y2Þ on ½a;b� is such that the second component is not reflected. By uniqueness, we
can claim that ðY1; Y2Þ ; ð �Y1; �Y2Þ on ½a;b�, and hence, Y2 is not reflected on ½a;b�. A
Remark 3.2. On any time interval where the solution of (Smin) provides a solution of the
system (S), it is straightforward to check, by using standard comparison result for BSDEs
or RBSDEs (see Theorem 2.2), that solution (Y1; Y2) is the minimal solution of (S).
Similarly, whenever the solution ( �Y1; �Y2) associated with system (Smax) is a solution of
(S), it can be proved that it is the maximal one.
3.2 Non-uniqueness: two counter-examples
In this section, we provide two explicit counter-examples of system (S) where uniqueness
does not hold.
First example. In this first example, we study the case when the two penalties a and b are
equal to zero and the terminal conditions j1 and j2 are equal. In addition, the two generators
ðs;vÞ! ciðs;v; y; zÞ, for i ¼ 1; 2, are assumed to be independent of ðy; zÞ and satisfy
P-a:s:; ;t # T ; c1ðt;vÞ2 c2ðt;vÞ , 0: ð3:25Þ
Therefore, this leads to the following system of RBSDEs:
ðS1Þ
Y1t ¼ j1 þ
Ð T
tc1ðsÞds 2
Ð T
tZ1
s dBs þ ðK1T 2 K1
t Þ;
Y2t ¼ j1 þ
Ð T
tc2ðsÞds 2
Ð T
tZ2
s dBs 2 ðK2T 2 K2
t Þ;
Y1t $ Y2
t ; t # T ;Ð T
0Y1
s 2 Y2s
� �dðK1
s þ K2s Þ ¼ 0:
8>>>>>><>>>>>>:Introduce ðY 1;min; Y 2;minÞ (resp. ðY 1;max; Y 2;maxÞ) as being the minimal (resp. the
maximal) solution of the system (S1) as constructed in Section 2.
Our objective is to establish that, even in this simple example, uniqueness for system
(S1) fails to hold. To do this, we prove that the minimal and maximal solutions for that
system do not coincide. So let us consider an arbitrary solution ðY 1; Y 2Þ of the system ðS1Þ.
If we set �Y ¼ Y 1 2 Y 2, then we have, for all t # T ,
�Yt ¼Ð T
tðc1ðsÞ2 c2ðsÞÞds þ ð �KT 2 �KtÞ þ
Ð T
t�Zs dBs;
�Yt $ 0; t # T ;Ð T
0�Yt d �Kt ¼ 0:
8>>><>>>: ð3:26Þ
B. Djehiche et al.442
Dow
nloa
ded
by [
Mos
kow
Sta
te U
niv
Bib
liote
] at
08:
47 3
1 A
ugus
t 201
3
Hence, standard results imply that the solution of this RBSDE is unique. Then taking
conditional expectation w.r.t. F t in (3.26), it yields
Y1t 2 Y2
t ¼ E
ðT
t
ðc1ðsÞ2 c2ðsÞÞds þ ð �KT 2 �KtÞjF t
:
Therefore, thanks to both (3.25) and the constraint condition on Y 1 2 Y 2, we obtain the
strict increasing property of �K. Next, for i ¼ 1; 2 denoting by K i;min and by K i;max, the pair
of increasing processes associated with each component of both the minimal solution and
the maximal solution, we show below
dK1;mint ; 0 and dK2;max
t ; 0; for any t # T :
To this end, let us consider the following system: for any s # T ,
dY1s ¼ 2c1ðsÞds þ Z1
s dBs; Y1T ¼ j1;
dY2s ¼ 2c2ðsÞds þ Z2
s dBs þ dK2s ; Y2
T ¼ j1;
Y2s # Y1
s ;Ð T
0Y1
s 2 Y2s
� �dK2
s ¼ 0:
8>>><>>>: ð3:27Þ
Then, for any t # T , we have
Y1t 2 Y1;min
t ¼ 2
ðT
t
ðZ1s 2 Z1;min
s ÞdBs 2 K1;minT 2 K1;min
t
� �:
As Y1t 2 Y
1;mint $ 0, it follows from standard arguments that K
1;mint ¼ K
1;minT for any t # T
and then K 1;min ; 0. Using now uniqueness of BSDEs, reflected or not, we obtain that
Y1 ; Y 1;min and then Y2 ; Y 2;min. In the same way, by considering the reflection on the
other equation in (3.27), we obtain that K 2;max ; 0.
Next, using again the uniqueness of the triple ð �Y ¼ Y 1 2 Y 2; �Z ¼ Z 1 2 Z 2; �K ¼
K 1 þ K 2Þ and solving the one lower barrier RBSDE (3.26), it follows that
�K ¼ K 1;min þ K 2;min ¼ K 2;min ¼ K 1;max þ K 2;max ¼ K 1;max:
The process �K uniquely defined by (3.26) being strictly increasing then, in view of the
second line of the previous equality, K 1;max and K 2;min are also strictly increasing.
Consequently, both Y 1;min and Y 1;max solve BSDEs with the same generator c1 and the
same terminal condition j1. However, Y 1;min solves a standard BSDE without any
reflection, whereas Y 1;max solves a RBSDE with the associated process K 1;max which is
strictly increasing, which yields that Y 1;min – Y 1;max and achieves the proof. A
Second example. Let us assume the following structure of the generators:
c1ðt;v; yÞ ¼ y and c2ðt;v; yÞ ¼ 2y:
We also assume that for any t # T , at ¼ bt ¼ 0. Then, we are led to consider the following
system, for all t [ ½0; T�,
ðS2Þ
Y1t ¼ 1 þ
Ð T
tY1
s ds 2Ð T
tZ1
s dBs þ ðK1T 2 K1
t Þ;
Y1t $ Y2
t andÐ T
tðY1
s 2 Y2s ÞdK1
s ¼ 0;
Y2t ¼ 1 þ
Ð T
t2Y2
s ds 2Ð T
tZ2
s dBs 2 ðK2T 2 K2
t Þ;
Y1t $ Y2
t andÐ T
tðY2
s 2 Y1s ÞdK2
s ¼ 0:
8>>>>>>><>>>>>>>:
Stochastics: An International Journal of Probability and Stochastic Processes 443
Dow
nloa
ded
by [
Mos
kow
Sta
te U
niv
Bib
liote
] at
08:
47 3
1 A
ugus
t 201
3
In this second example, we will show that the minimal and maximal solutions are not
equal, and, therefore, uniqueness does not hold. We note that, considering an arbitrary
solution of system (S2), the difference Y 1 2 Y 2 does not solve any more RBSDEs, which
was the crucial fact, therefore we rely on in the previous example.
To prove that uniqueness does not hold for (S2), let us consider the minimal
(resp. maximal) solution (Y 1;min; Y 2;min) (resp. (Y 1;max; Y 2;max)) of (S2) which is given,
for all t [ ½0; T�, by
dY1;mint ¼ 2Y
1;mint dt 2 Z
1;mint dBt and Y
1;minT ¼ 1;
dY2;mint ¼ 22Y
2;mint dt 2 Z
2;mint dBt þ dK
2;mint and Y
2;minT ¼ 1;
Y1;mint $ Y
2;mint and
Ð T
0ðY
1;mint 2 Y
2;mint ÞdK
2;mint ¼ 0;
8>>><>>>:and
dY1;maxt ¼ 2Y
1;maxt dt 2 Z
1;maxt dBt 2 dK
1;maxt and Y
1;maxT ¼ 1;
dY2;maxt ¼ 22Y
2;maxt dt 2 Z
2;maxt dBt and Y
2;maxT ¼ 1;
Y1;maxt $ Y
2;maxt and
Ð T
0ðY
1;maxt 2 Y
2;maxt ÞdK
1;maxt ¼ 0;
8>>><>>>:meaning that K 1;min ; 0 and K 2;max ; 0. But, the solution of the first system is given, for
all t # T , by
Y1;mint ¼ Y2;min
t ¼ eT2t; Z1;mint ¼ Z2;min
t ¼ 0 and K2;mint ¼ eT ð1 2 e2tÞ:
On the other hand, one of the second systems is given, for all t # T , by
Y1;maxt ¼ Y2;max
t ¼ e2ðT2tÞ; Z1;maxt ¼ Z2;max
t ¼ 0 and K1;maxt ¼
1
2e2T ð1 2 e22tÞ:
Therefore, as we can see, uniqueness for (S2) does not hold in this case. Finally, let us
point out that in order to exhibit the solutions of the previous systems, we have kept in
mind two facts: (1) the solutions of those systems are deterministic and (2) we have used
properties of the Snell envelope of processes. A
Remark 3.3. As a slight generalization of the previous counter-example, let us consider
here the more general case when the generators take the following form:
c1ðs; yÞ ¼ a1s y and c2ðs; yÞ ¼ a2
s y;
where a1 and a2 are deterministic functions, integrable on ½0; T� and satisfy
a1t , a2
t ;;t # T .
We are led to study the following system given, for all t [ ½0; T�, by
ðS02Þ
dY1t ¼ 2a1
t Y1t dt þ Z1
t dBt 2 dK1t ; Y1
T ¼ 1;
dY2t ¼ 2a2
t Y2t dt þ Z2
t dBt þ dK2t ; Y2
T ¼ 1;
Y1t $ Y2
t andÐ T
0ðY2
t 2 Y1t ÞdK1
t ¼Ð T
0ðY2
t 2 Y1t ÞdK2
t ¼ 0:
8>>><>>>:Similarly as in the previous proof, we consider here the minimal solution (resp. the
maximal solution) of the previous system which is denoted by (Y 1;min; Y 2;min) (resp. by
B. Djehiche et al.444
Dow
nloa
ded
by [
Mos
kow
Sta
te U
niv
Bib
liote
] at
08:
47 3
1 A
ugus
t 201
3
(Y 1;max; Y 2;max)). In addition and as already mentioned in Remark 3.2, the increasing
processes associated with these solutions are such that
K 1;min ; 0 and K 2;max ; 0:
Thus, Y 1;min and Y 2;max solve a standard BSDE which, by uniqueness, implies that, for all
t [ ½0; T�,
Y1;mint ¼ exp
ðT
t
a1s ds
; Y2;max
t ¼ exp
ðT
t
a2s ds
and Z1;min
t ¼ Z2;maxt ¼ 0:
Next, since the data of the system are deterministic and using the characterization of the
Snell envelope of a process as the smallest supermartingale which dominates it (see, e.g.
[4]), we get that
;t [ ½0;T�; Y2;mint ¼ Y1;min
t ; Z2;mint ¼ 0 and K2;min
t ¼
ðt
0
a2s 2a1
s
� �Y1;min
s ds: ð3:28Þ
Thus, we have obtained the minimal solution of ðS02Þ.
Concerning the maximal solution, we proceed in the same way and then we can check
that Y 1;max, Z 1;max and K 1;max are given by
;t [ ½0; T�; Y1;maxt ¼ Y2;max
t ; Z1;maxt ¼ 0 and K1;max
t ¼
ðt
0
a2s 2 a1
s
� �Y2;max
s ds:
Therefore, since a1 , a2, we obtain, for i ¼ 1; 2, Y i;min , Y i;max, which entails that, in
this case, uniqueness does not hold for ðS02Þ as well.
4. Study of the related system of VI
4.1 The system of VI
In this section, we briefly describe the connection between solutions of (S) and existence
viscosity solutions of the following system of VI with interconnected obstacles:
ðVIÞ
min u1ðt; xÞ2 u2ðt; xÞ þ aðt; xÞ;2›tu1ðt; xÞ2 Lu1ðt; xÞ2 c1ðt; xÞ
� �¼ 0;
max u2ðt; xÞ2 u1ðt; xÞ2 bðt; xÞ;2›tu2ðt; xÞ2 Lu2ðt; xÞ2 c2ðt; xÞ
� �¼ 0;
u1ðT ; xÞ ¼ g1ðxÞ and u2ðT ; xÞ ¼ g2ðxÞ;
8>><>>:when the dependence of dynamics of the cash flows Y 1; Y 2 of, e.g. the fluctuations of the
market prices, X, assumed to be a diffusion process, is made explicit. Note that the
obstacles may depend on the diffusion process X. In (VI), L denotes the infinitesimal
generator of the diffusion X.
For ðt; xÞ [ ½0; T� £ Rk, let X t;x U ðXt;xs Þs#T be the solution of the following standard
differential equation:
dXt;xs ¼ mðs;Xt;x
s Þds þ s ðs;Xt;xs ÞdBs; t # s # T; Xt;x
s ¼ x; s # t; ð4:29Þ
where the two functions m U ðmðt; xÞÞ and s U ðs ðt; xÞÞ defined on ½0; T� £ Rk and taking
their respective values in Rk and Rk£d are uniformly Lipschitz w.r.t. x and have linear
growth. This means that, for all t, x, y,
ðC1Þjmðt; xÞ2 mðt; yÞj þ js ðt; xÞ2 s ðt; yÞj # Cjx 2 yj;
jmðt; xÞj þ js ðt; xÞj # Cð1 þ jxjÞ:
(
Stochastics: An International Journal of Probability and Stochastic Processes 445
Dow
nloa
ded
by [
Mos
kow
Sta
te U
niv
Bib
liote
] at
08:
47 3
1 A
ugus
t 201
3
These properties ensure both existence and uniqueness of a solution for (4.29).
Additionally and for any u, u $ 2, there exists a constant C such that for any x in Rk
E sup0#s#T
Xt;xs
�� ��u� �# Cð1 þ jxj
uÞ: ð4:30Þ
Next, in this setting, the infinitesimal generator L of the diffusion X :¼bX t;x is defined,
for any function F in C1;2ð½0; T� £ RÞ, as follows:
LðFÞðt; xÞ ¼ kmðt; xÞ;Fðt; xÞlþ1
2Trace ss TD2Fðt; xÞ
� �;
where T stands for the transpose operation. Let us now make the following assumption on
the functions ci, i ¼ 1; 2:
ðC2Þ
The functions c1 and c2 do not depend on ðy; zÞ; i:e: c1 ¼ c1ðt; xÞ and
c2 ¼ c2ðt; xÞ; they are jointly continuous in ðt; xÞ and have polynomial
growth; i:e: there exist two positive constants q and C such that;
for all ðt; xÞ; jciðt; xÞj # Cð1 þ jxjqÞ; i ¼ 1; 2:
8>>>>><>>>>>:The assumption that each generator ci does not depend on ðy; zÞ but only on x is quite
natural especially for applications in economics. It means that the payoffs are not of
recursive type, and the utilities c1 and c2 depend only on the process X t;x which stands,
e.g. for the price of a commodity such as electricity or oil price in the market.
Next, let gi U ðgiðxÞÞ, i ¼ 1; 2, a U ðaðt; xÞÞ and b U ðbðt; xÞÞ be the given functions
defined, respectively, on Rk and ½0; T� £ Rk, with values in R and satisfy
ðC3Þ
ðiÞ a and b are of polynomial growth; belong to C1;2ð½0;T�£RdÞ and satisfy
;ðt;xÞ[ ½0;T�£Rk;Ð T
01½aðs;Xt;x
s Þ¼bðs;Xt;xs Þ�ds ¼ 0;
ðiiÞ g1 and g2 are continuous and of polynomial growth;
ðiiiÞ ;x [ Rk; g1ðxÞ2 g2ðxÞ$ max{2 aðT ;xÞ;2bðT ;xÞ}:
8>>>>><>>>>>:With the help of the solutions of the system of RBSDEs (S), we show below that (VI)
has a solution in viscosity sense whose definition is the following.
Definition 1. Let ðu1; u2Þ be a pair of continuous functions on ½0; T� £ Rk. It is called
(i) a viscosity supersolution (resp. subsolution) of the system (VI) if for any ðt0; x0Þ [½0; T� £ Rk and any pair of functions ðw1;w2Þ [ ðC 1;2ð½0; T� £ RkÞÞ2 such that
ðw1;w2Þðt0; x0Þ ¼ ðu1; u2Þðt0; x0Þ and for any i ¼ 1; 2, ðt0; x0Þ is a maximum
(resp. minimum) of wi 2 ui, then we have
min u1ðt0; x0Þ2 u2ðt0; x0Þ þ aðt0; x0Þ;2›tw1ðt0; x0Þ2 Lw1ðt0; x0Þ2 c1ðt0; x0Þf g $ 0
ðresp: # 0Þ;
and
max u2ðt0; x0Þ2 u1ðt0; x0Þ2 bðt0; x0Þ;2›tw2ðt0; x0Þ2 Lw2ðt0; x0Þf
2c2ðt0; x0Þg $ 0 ðresp: # 0Þ:
(ii) a viscosity solution of system ðVIÞ if it is both a viscosity supersolution and
subsolution.
B. Djehiche et al.446
Dow
nloa
ded
by [
Mos
kow
Sta
te U
niv
Bib
liote
] at
08:
47 3
1 A
ugus
t 201
3
4.2 Construction and regularity of viscosity solution of system
Our objective is to construct and identify a continuous viscosity solution for the system (VI)
by relying on both standard results for the representation of viscosity solutions by BSDEs
(see, for example, [5], Theorem 8.5) and on the results obtained in the previous sections.
For ðt; xÞ [ ½0; T� £ Rk, let ðY 1;ðt;xÞ; Y 2;ðt;xÞ; Z 1;ðt;xÞ; Z 2;ðt;xÞ;K 1;ðt;xÞ;K 2;ðt;xÞÞ be a solution
of the following system of RBSDEs: for any s [ ½t;T�,
ð ~SÞ
Y1;ðt;xÞs ¼ g1ðX
ðt;xÞT Þ þ
Ð T
sc1ðu;Xt;x
u Þ du þ ðK1;ðt;xÞT 2 K1;ðt;xÞ
s Þ2Ð T
sZ1;ðt;xÞ
u dBu;
Y2;ðt;xÞs ¼ g2ðX
ðt;xÞT Þ þ
Ð T
sc2ðu;Xt;x
u Þ du 2 ðK2;ðt;xÞT 2 K2;ðt;xÞ
s Þ2Ð T
sZ2;ðt;xÞ
u dBu;
Y1;ðt;xÞs $ Y2;ðt;xÞ
s 2 aðs;Xt;xs Þ and Y2;ðt;xÞ
s # Y1;ðt;xÞs þ bðs;Xt;x
s Þ;Ð T
tY1;ðt;xÞ
s 2 ðY2;ðt;xÞs 2 aðs;Xt;x
s ÞÞ� �
dK1;ðt;xÞs ¼ 0;Ð T
tðY1;ðt;xÞ
s þ bðs;Xt;xs Þ2 Y2;ðt;xÞ
s ÞdK2;ðt;xÞs ¼ 0:
8>>>>>>>>><>>>>>>>>>:Note that, thanks to Theorems 2.2 and 3.1 and Assumptions (C1)–(C3), especially the fact
that the processes ðaðs;Xt;xs ÞÞs[½t;T� and ðbðs;Xt;x
s ÞÞs[½t;T� verify (B4) and (B40), respectively,
both the minimal and maximal solutions exist and, furthermore, these two solutions
coincide. Indeed, uniqueness holds because the functions ci, i ¼ 1; 2, do not depend on
ðy; zÞ and the barriers satisfy the condition (ii) of Theorem 3.1. Moreover, this unique
solution is obtained as a limit of the increasing and decreasing schemes because the
functions a U ðaðt; xÞÞ and b U ðbðt; xÞÞ belong to C1;2ð½0; T� £ RkÞ and are of polynomial
growth.
According to the construction of the minimal solution of ð ~SÞ we have, for any
s [ ½t; T�,
Y1;ðt;xÞs ; Y2;ðt;xÞ
s
� �¼ lim
n!1Y1;ðt;xÞ;n
s ; Y2;ðt;xÞ;ns
� �;
where ðY 1;ðt;xÞ;n; Y 2;ðt;xÞ;nÞ are defined in the same way as in (Sn) but with the specific data
above. Thanks to Theorem 8.5 in [5], and by an induction argument there exist
deterministic functions uin U ðui
nðt; xÞÞ, i ¼ 1; 2, continuous on ½0; T� £ Rk such that, for
any s [ ½t; T�,
Y1;ðt;xÞ;ns ; Y2;ðt;xÞ;n
s
� �¼ u1
nðs;Xt;xs Þ; u2
nðs;Xt;xs Þ
� �:
Moreover, there exist two positive constants a1 and a2 such that, for any
ðt; xÞ [ ½0; T� £ Rk,
juinðt; xÞj þ jui
nðt; xÞj # a1ð1 þ jxja2Þ:
Finally, the sequences ðuinÞn$0, i ¼ 1; 2, are increasing because Y i;ðt;xÞ;n # Y i;ðt;xÞ;nþ1,
i ¼ 1; 2. Therefore, there exist two deterministic lower semi-continuous functions ui,
i ¼ 1; 2, with polynomial growth such that, for i ¼ 1; 2 and for any ðt; xÞ [ ½0; T� £ Rk,
Y1;ðt;xÞs ; Y2;ðt;xÞ
s
� �¼ u1ðs;Xt;x
s Þ; u2ðs;Xt;xs Þ
� �; ;s [ ½t; T�: ð4:31Þ
Now, and in the same fashion, when considering the decreasing approximating
scheme, it follows from the same result in [5] and from the uniqueness that u1 and u2 are
also upper semi-continuous. Therefore, u1 and u2 are continuous with polynomial growth.
Finally, relying on Theorem 8.5 in [5], we directly obtain the following result.
Stochastics: An International Journal of Probability and Stochastic Processes 447
Dow
nloa
ded
by [
Mos
kow
Sta
te U
niv
Bib
liote
] at
08:
47 3
1 A
ugus
t 201
3
Theorem 4.1. The pair ðu1; u2Þ defined in (4.31) is a continuous viscosity solution for the
system (VI).
Acknowledgements
The authors are grateful to Paavo Salminen and the two anonymous reviewers for their insightfulremarks and suggestions that improved the content of the work. This work was completed while thefirst author was visiting the Department of Mathematics of Universite du Maine. Financial supportfrom MATPYL (RPL) is gratefully acknowledged.
Notes
1. Email: [email protected]. Email: [email protected]
References
[1] M.J. Brennan and E.S. Schwartz, Evaluating natural resource investments, J. Bus. 58 (1985),pp. 135–137.
[2] C. Dellacherie and P.A. Meyer, Probabilites et potentiels, Chapter I–IV Hermann, Paris, 1975.[3] A. Dixit and R.S. Pindyck, Investment Under Uncertainty, Princeton University Press,
Princeton, NJ, 1994.[4] N. El Karoui, Les aspects probabilistes du controle stochastique, in Ecole d’ete de probabilites
de Saint-Flour, Lecture Notes in Mathematics No 876, Springer Verlag, Berlin, 1980.[5] N. El Karoui, C. Kapoudjan, E. Pardoux, S. Peng, and M.-C. Quenez, Reflected solutions of
backward SDE’s and related problems for PDE’s, Ann. Prob. 25(2) (1997), pp. 702–737.[6] R. Korn, Some applications of impulse control in mathematical finance, Math. Methods Oper.
Res. 50 (1999), pp. 493–518.[7] E. Pardoux and S. Peng, Adapted solution of BSDE, Syst. Control Lett. 14 (1990), pp. 55–61.[8] S. Peng, Monotonic limit theorem of BSDE and nonlinear decomposition theorem of
Doob-Meyer’s type, Prob. Theory Relat. Fields 113 (1999), pp. 473–499.[9] L. Trigeorgis, Real Options: Managerial Flexibility and Strategy in Resource Allocation,
MIT Press, Cambridge, MA, 1996.
B. Djehiche et al.448
Dow
nloa
ded
by [
Mos
kow
Sta
te U
niv
Bib
liote
] at
08:
47 3
1 A
ugus
t 201
3