Appendix 2.A Auxiliary result for the one-dimensional BSDE
3.3 FBSDEs with superquadratic growth
3.3.1 Proof of Theorem 3.2.1
h(x)−h(x0) ≤k7
x−x0
and |h(x)| ≤λ5
for allx, x0 ∈Rm.
The second main result of this work is the following:
Theorem 3.2.2. If (B1) - (B4) hold, then there exists a constantCk,λ depending only on the coefficients ki and λi such that if T ≤ Ck,λ, then there exist two constants C1 andC2 such that FBSDE (3.2.2)has a unique solution (X, Y, Z) such that(X, Y, Z·W) ∈ S2(Rm)× S∞(Rm
0)×BMO andkYkS∞(Rm0) ≤C1, andkZ·WkBMO≤C2.
3.3 FBSDEs with superquadratic growth
3.3.1 Proof of Theorem 3.2.1
Step 1: We first assume thath, b, gare continuously differentiable in all variables.
We will show that the sequence(Xn, Yn, Zn)given byX0 = 0,Y0 = 0,Z0 = 0 and
(Xtn+1 =x+Rt
0 b(Xun+1, Yun)du+Rt
0σudWu
Ytn+1 =h(XTn+1) +RT
t gu(Xun+1, Yun+1, Zun+1)du−RT
t Zun+1dWu, n≥1 is well defined and that there exists a constantC >0which does not depend onn such that|Zn|< Cfor alln.
By [68] and [61] the processX1is well defined, belongs toD1,2(Rm)and its Malliavin’s derivative satisfies
DtXr1= 0, 0≤r < t≤T, DtXr1=
r
Z
t
(∂xbDtXu1+∂ybDtYu0)du+Dt
r
Z
t
σudWu
, 0≤t≤r≤T,
withDt(Rr
t σudWu) = σ1[t,r], see [61, Theorem 2.2.1]. Hence, sincebis Lips-chitz continuous, by Gronwall’s inequality we have
DtXr1
≤eT k1λ2.
Moreover, by the chain rule, see [61, Proposition 1.2.4] it follows thath(XT1) ∈ D1,2(Rm
0) and D(h(XT1)) = ∂xh(XT1)DXT1. Therefore, h(XT1) has bounded Malliavin derivative since∂xhis bounded. We then deduce from Theorem 3.A.2 and its proof that(Y1, Z1)exists, (Y1, Z1) ∈ D1,2(Rm0)× D1,2(Rm0×d),DY1 is bounded and Zt1 = DtYt1. Now let n ∈ N, assume that (Xn, Yn, Zn) ∈ D1,2(Rm)×D1,2(Rm
0)×D1,2(Rm
0×d),DXn, DYnare bounded andZtn=DtYtn. The processXn+1is well defined, belongs toD1,2(Rm)and its Malliavin deriva-tive satisfies
DtXrn+1 = 0, 0≤r < t≤T, DtXrn+1 =σ1[t,r]+
r
Z
t
(∂xbDtXun+1+∂ybDtYun)du, 0≤t≤r ≤T.
Since∂xb,∂yb andσ are bounded byk1, k2 andλ2 respectively, it follows from Gronwall’s inequality that
DtXrn+1
≤eT k1
λ2+k2 T
Z
0
|DtYun|du
. Hence,
DtXn+1
S∞ ≤eT k1(λ2+k2TkDtYnkS∞)<∞. (3.3.1) By the chain rule,D(h(XTn+1))exists and is bounded. It then follows again from Theorem 3.A.2 and its proof that(Yn+1, Zn+1) exists andZn+1 is bounded. In addition,(Yn+1, Zn+1)are Malliavin differentiable and the derivatives satisfy, for j= 1, . . . , d,
DtjYrn+1 = 0, DtjZrn+1 = 0, 0≤r < t≤T, DtjYrn+1 =∂xh(XTn+1)DjtXTn+1+
T
Z
r
∂xgDjtXun+1+∂ygDjtYun+1
+∂zgDjtZun+1 du−
T
Z
r
DjtZun+1dWu, 0≤t≤r ≤T.
By (A3)-(A5) and the boundedness ofZn+1andDXn+1, it follows from the same
procedure of the proof of Lemma 3.A.1 that fori= 1, . . . , m0;j= 1, . . . , d. Plugging the above estimate in (3.3.1), we obtain
0×d). Indeed, using (A1) we can estimate the norm of the dif-ferenceXtn+1−Xtnas
Taking expectation on both sides and using Cauchy-Schwarz’ inequality, we have E
ChoosingT to be small enough so that2T2k12 ≤ 12, it follows
E
"
sup
0≤t≤T
|Xtn+1−Xtn|2
#
≤4T2k22E
"
sup
0≤t≤T
|Ytn−Ytn−1|2
#
. (3.3.2)
On the other hand, applying Itô’s formula toeβt|Ytn+1−Ytn|2,β ≥0, we have
eβt|Ytn+1−Ytn|2
=eβT|h(XTn+1)−h(XTn)|2−2
T
Z
t
eβs(Ysn+1−Ysn)(Zsn+1−Zsn)dWs
−
T
Z
t
eβs(Zsn+1−Zsn)2ds−
T
Z
t
βeβs(Ysn+1−Ysn)2ds
+ 2
T
Z
t
eβs(Ysn+1−Ysn)
gs(Xsn+1, Ysn+1, Zsn+1)−gs(Xsn, Ysn, Zsn) ds.
Hence, due to the condition (A3) and the boundedness of(Zn), it holds
eβt|Ytn+1−Ytn|2+
T
Z
t
eβs(Zsn+1−Zsn)2ds
≤eβT
h(XTn+1)−h(XTn)
2−2
T
Z
t
eβs(Ysn+1−Ysn)(Zsn+1−Zsn)dWs
−
T
Z
t
βeβs(Ysn+1−Ysn)2ds+ 2
T
Z
t
eβsρ(Q)
Ysn+1−Ysn
Zsn+1−Zsn ds
+ 2
T
Z
t
eβsk7
Ysn+1−Ysn
Xsn+1−Xsn ds+ 2
T
Z
t
eβsk4
Ysn+1−Ysn
2ds.
With some positive constantsα1,α2, it follows from (A5) and Young’s inequality
that
2 + 2k8 and taking expectation on both sides above, we have Next, taking conditional expectation with respect toFtin (3.3.3),
eβt|Ytn+1−Ytn|2+E
Thus, by Burkholder-Davis-Gundy’s inequality, with a positive constantc1 and It now follows from (3.3.2) that
E TakingT small enough so that
8(c1+ 1)eβT(k25+T)T2k22≤ 1 2,
we obtain that (Xn, Yn, Zn) is a Cauchy sequence in S2(Rm) × S2(Rm
0) × H2(Rm
0×d). By continuity of b, g and h we have the existence of a solution (X, Y, Z)inS2(Rm)× S2(Rm
0)× H2(Rm
0×d)of FBSDE (3.2.1) and it follows from the boundedness of(Zn)thatZsatisfies (3.2.3). The uniqueness follows from the boundedness ofZ and by repeating the above arguments on the difference of two solutions.
It is easy to check thatbnsatisfies (A1) with the constantsk1, k2 and2λ1and that gn andhn satisfy (A3) - (A4) and (A5), respectively, with the same constants.
From the above argument, there existsC¯k,λ,m0,dindependent ofnsuch that ifT ≤ C¯k,λ,m0,d, FBSDE (3.2.1) with parameters (bn, hn, gn) admits a unique solution (Xn, Yn, Zn)∈ S2(Rm)× S2(Rm
0)× S∞(Rm
0×d)and
|Ztij,n| ≤2λ2m0ek1T+m0k4T (k5+k3T) P⊗dt-a.e.
By the Lipschitz continuity conditions onbandhand the locally Lipschitz condi-tion ofg, the sequences(bn)and(hn)converge uniformly tobandh onRm+m
0
andRm, respectively, and (gn)converges toguniformly onRm+m
0 ×Λfor any compact subset Λ of Rm
0×d. Combining these uniform convergences with the boundedness of Zn, similar to above, we can show that there exists a constant C˜k,λ,m0,ddepending only onk1, k2, k3, k4, k5, λ2, m0, dsuch that ifT ≤C˜k,λ,m0,d, (Xn, Yn, Zn)is a Cauchy sequence inS2(Rm)× S2(Rm
0)× H2(Rm
0×d). Hence withCk,λ,m0,d = ¯Ck,λ,m0,d∧C˜k,λ,m0,d, for anyT ≤Ck,λ,m0,d, the FBSDE (3.2.2) admits a solution (X, Y, Z) ∈ S2(Rm)× S2(Rm
0)× S∞(Rm
0×d) and|Ztij| ≤ 2λ2m0ek1T+m0k4T(k5+k3T). The uniqueness follows from similar arguments.
Step 4: Now, assumeT > Ck,λ,m0,dand the additional growth conditions onb, gandhgiven by (3.2.4) hold. Let˜hQ : R → Rbe a continuously differentiable function whose derivative is bounded by1and such that˜h0Q(a) = 1for all−Q≤ a≤Qand
˜hQ(a) =
(Q+ 1) if a > Q+ 2 a if |a| ≤Q
−(Q+ 1) if a <−(Q+ 2).
An example of such a function is given by h˜Q(a) =
( −Q2+ 2Qa−a(a−4)
/4 if a∈[Q, Q+ 2]
Q2+ 2Qa+a(a+ 4)
/4 if [−(Q+ 2),−Q], see [43]. By the assumptions (A3) the functiong˜: [0, T]×Rm×Rm
0×Rm
0×d→R defined by
˜
gt(x, y, z) :=gt(x, y, hQ(z)) (3.3.4) withhQ(z) := (˜hQ(zij))ijis Lipschitz continuous in all variables. Thus, it follows from [20, Theorem 2.6] that the equation
(X˜t=x+Rt
0bu( ˜Xu,Y˜u)du+Rt
0 σudWu Y˜t=h( ˜XT) +RT
t ˜gu( ˜Xu,Y˜u,Z˜u)du−RT
t Z˜udWu, t∈[0, T] (3.3.5) admits a unique solution( ˜X,Y ,˜ Z)˜ ∈ S2(Rm)× S∞(Rm
0)× H∞(Rm
0×d). More-over, there exists a bounded functionθ : [0, T]×Rm → Rm
0 which is Lipschitz continuous inxsuch thatY˜t=θ(t,X˜t)for allt∈[0, T]. PutN = [T /Ck,A,q,m0,d],
where[a]denotes the integer part ofa, andti := iCk,A,q,m0,d,i = 0, . . . , N and
That is,(X, Y, Z)satisfies Equation (3.3.6). Uniqueness follows from [20,
Theo-rem 2.6]. This concludes the proof.
3.3.2 Fully coupled systems
In order to consider the fully coupled forward-backward system, i.e., we allow the dependence in(x, y, z)ofbandσ, we assume boundedness conditions on the
Malliavin derivatives of the generator and the terminal condition. Under this as-sumption, we can obtain solvability on any time interval[0, T],T ∈(0,∞)for the Markovian case. Now, consider the following conditions
(A1’) b: Ω×[0, T]×Rm×Rm
0×Rm
0×d→Rmis a continuous and measurable function such that there existk1, k2, k3, λ1≥0such that
0is a continuous and measurable function such thatgit(x, y, z) = gti(x, y, zi)fori = 1, . . . , m0 and there
(A5’) h: Ω×Rm →Rm
0 is continuous andFT-measurable such thath(XT)∈ D1,2(Rm
Theorem 3.3.1. If (A1’) - (A5’) hold, then there exist two constants Ck,A,q,m0,d andεk,A,q,m0,ddepending only onk1, k2, k3, k4, k5, k7, k8, k9, A, q, m0, dsuch that solution of the FBSDE
(Xtn+1 =x+Rt and|Zn+1−Zn|follows from exactly the same procedure as in the the proof of Theorem 3.2.1. Indeed, we have
|Xtn+1−Xtn|2 ≤6
Taking supremum with respect tot, then expectation to both sides and using Cauchy-Schwarz’ and Burkholder-Davis-Gundy’s inequalities, we have
E Hence the result follows directly from the arguments in the proof of Theorem
3.2.1.
0 is continuous and continu-ously differentiable inyandz, and is such thatgti(x, y, z) =git(x, y, zi)
i = 1, . . . , m0 and there existk7, k8, λ3 ≥0 as well as a nondecreasing
0 is continuously differentiable and such thath(XT) ∈ D1,2 for any XT ∈ L2(FT) and there exist constants k9, λ4, Aij ≥ 0
Theorem 3.3.2. If (A1”) - (A5”) hold, then the FBSDE (Xt=x+Rt
0×d)satisfying
|Zt| ≤Q:=
Proof. Consider the constant Ck,A,q,m0,d introduced in Theorem 3.3.1. If T ≤ Ck,A,q,m0,d, the result follows from Theorem 3.3.1.
In the rest of the proof let us assume that T > Ck,A,q,m0,d. The functiong˜ defined by (3.3.4) is Lipschitz continuous and differentiable in(y, z), and satisfies (A4”). Hence, by [20, Theorem 2.6] the FBSDE
(X˜t=x+Rt
0bu( ˜Xu,Y˜u,Z˜u)du+Rt
0σu( ˜Xu,Y˜u)dWu Y˜t=h( ˜XT) +RT
t ˜gu( ˜Xu,Y˜u,Z˜u)du−RT
t Z˜udWu, t∈[0, T] (3.3.7) admits a unique solution( ˜X,Y ,˜ Z)˜ ∈ S2(Rm)× S∞(Rm
0)× H∞(Rm
0×d). More-over, by [28] the processesY˜ andZ˜ are Malliavin differentiable and we have for j= 1, . . . , d,
DrjY˜t= 0, DrjZ˜t= 0, 0≤t < r < T, DrjY˜t=Drjh( ˜XT) +
T
Z
t
∂yg˜uDjrY˜u+∂zg˜uDjrZ˜u+Drjg˜u( ˜Xy,Y˜u,Z˜u)du
−
T
Z
t
DjrZ˜udWu, 0≤r ≤t≤T.
Since ˜g is Lipschitz inz, ∂z˜g( ˜Xy,Y˜u,Z˜u) is bounded. By (A4”) and (A5”), it follows from the same procedure as in the proof of Lemma 3.A.1 that
|DrjY˜ti| ≤
m0
X
i=1
Aij +
m0
X
i=1 T
Z
t
qij(s)e−m0k8(T−s)ds
em0k8(T−t), P⊗dt-a.e., i = 1, . . . , m0;j = 1, . . . , d. Let Ck,A,q,m˜ 0,d be the constant given by Theorem 3.3.1 replacingAij by A˜ij. One can easily check thatCk,A,q,m˜ 0,d ≤ Ck,A,q,m0,d sinceAij ≤ A˜ij. Considering a sequence0 = t0 ≤ t1 ≤ · · · ≤ tN = T with max1≤i≤N|ti−ti−1| ≤Ck,A,q,m˜ 0,dsimilar to the last part of the proof of Theorem 3.2.1. SinceDrY˜ti ∈L∞for allr∈[ti−1, ti]we can get that fori= 1, . . . , Nthat
(Xt= ˜Xti−1+Rt
ti−1bu(Xu, Yu, Zu)du+Rt
ti−1σu(Xu, Yu)dWu Yt= ˜Yti+Rti
t gu(Xu, Yu, Zu)du−Rti
t ZudWu, t∈[ti−1, ti] has a unique solution (Xi, Yi, Zi) ∈ S2(Rm) × S2(Rm
0) × S∞(Rm
0×d) and Zti
≤ Q for all t ∈ [ti−1, ti]. By the uniqueness of FBSDE (3.3.7), we have (Xi, Yi, Zi)1[ti−1,ti] = ( ˜X,Y ,˜ Z˜)1[ti−1,ti]. The result follows from a recursion
and pasting procedure as above.