Appendix 2.A Auxiliary result for the one-dimensional BSDE
3.4 FBSDEs with quadratic growth
3.4.2 Regularity of solutions
8k22T2e2k1T k42+k72
+ 8k52T2
kδyk2S∞
+ 24√
3k26L24T1−ε
Tε+ 2 +εL24kz·W˜ik2
BMO( ˜Pi)
+εL24k¯z·W˜ik2
BMO( ˜Pi)
kδz·W˜ik2
BMO( ˜Pi).
With the strictly positive constantsc1, c2depending only onk3andC2from Lemma A.1.3,
kδYk2S∞+c1kδZ·Wk2BMO(P)
≤m0
8k22T2e2k1T k42+k27
+ 8k52T2
kδyk2S∞
+ 24
√
3k26L24T1−εc2m0 Tε+ 2 + 2εL24c2C22
kδz·Wk2BMO(P). LettingT be small enough so that
m0
1 +c1
1
8k22T2e2k1T k42+k27
+ 8k52T2
≤ 12 24√
3k26L24T1−εc2m0 1 +c1
1
Tε+ 2 + 2εL24c2C22
≤ 12, (3.4.3) it follows thatΨdefines a contraction mapping. Then, there exists a fixed point (Y, Z)∈ B. Hence there exists a constantCk,λwhich depends only onki, λisuch that whenT ≤Ck,λ, FBSDE (3.2.2) admits a unique solution(X, Y, Z)such that (X, Y, Z ·W) belongs toS2(Rm)× S∞(Rm
0)×BMO andkYkS∞(Rm0) ≤ C1, kZ·WkBMO ≤C2.
3.4.2 Regularity of solutions
For any initial valuex ∈ Rm, we denote by(Xx, Yx, Zx) the unique solution of the FBSDE (3.2.2). The following two results provide regularity of the solution upon the parameterx.
Theorem 3.4.1 (Continuity). Assume (B1) - (B4). With the same constant Ck,λ as in Theorem 3.2.2, ifT ≤Ck,λ, the functionx7→(Xx, Yx, Zx)is continuous.
Proof. LetT ≤ Ck,λ and(Xx, Yx, Zx)be the solution of the FBSDE (3.2.2) for anyx∈R. Notice thatXx−Xx0is bounded. In fact, using the Lipschitz continuity condition onb, we have
|Xtx−Xtx0| ≤ |x−x0|+k1 t
Z
0
|Xux−Xux0|du+k2 t
Z
0
|Yux−Yux0|du
≤ |x−x0|+k2TkYx−Yx0kS∞+k1
t
Z
0
|Xux−Xux0|du
≤
|x−x0|+k2TkYx−Yx0kS∞ ek1t, by Gronwall’s lemma. Thus
kXx−Xx0kS∞ ≤
|x−x0|+k2TkYx−Yx0kS∞
ek1T. (3.4.4) On the other hand, arguing such as in the proof of Theorem 3.2.2, we have, for eachi= 1, . . . , m0,
Yti,x−Yti,x0 +
T
Z
t
Zui,x−Zui,x0dW˜ui
=hi(XTx)−hi(XTx0) +
T
Z
t
liu(Xux, Yux, Zux)−liu(Xux0, Yux0, Zux0)du
whereW˜i=W−R.
0ηisdswith|ηsi| ≤k3(1+|Zsi,x|+|Zsi,x0)is a Brownian motion under the equivalent measureP˜i = E(ηi·W)T ·P. Hence, similar to Theorem 3.2.2, with the same constantsc1, c2andC2,
kYx−Yx0k2∞+c1k(Zx−Zx0)·Wk2BMO
≤m0
16k22T2e2k1T k24+k27
+ 8k25T2
kYx−Yx0k2∞ + 16m0e2k1T k24+k27
|x−x0|2 + 24
√
3k62L24T1−εc2m0 Tε+ 2 + 2εL24c2C22
k(Zx−Zx0)·Wk2BMO. Therefore, it follows from (3.4.3) that
kYx−Yx0k2S∞ ≤ 16m0e2k1T k42+k72 1−m0 16k22T2e2k1T k24+k27
+ 8k25T2|x−x0|2. (3.4.5) and
c1kZx−Zx0k2BMO ≤32m0e2k1T k42+k27
|x−x0|2. (3.4.6) Combining with (3.4.4)
kXx−Xx0kS∞
≤ 1 +k2T s
16m0e2k1T k42+k72 1−m0 16k22T2e2k1T k24+k27
+ 8k25T2
!
ek1T|x−x0|.
(3.4.7)
This proves continuity of the solution.
(B5) The functionsb;h;f andlare continuously differentiable.
(B6) The functionsh0;∂xb;∂yb;∂xl;∂yl;∂zlandf0are Lipschitz continuous in all variables with Lipschitz constantK.
Theorem 3.4.2 (Differentiability). Assume (B1) - (B6). With the same constant Ck,λas in Theorem 3.2.2, ifT ≤Ck,λ, the functionx7→(Xx, Yx, Zx)is differen-tiable.
Proof. Let T ≤ Ck,λ, x, x0 ∈ Rm andλ, λ0 > 0. Let ej = (0, . . . ,1, . . . ,0) be the unit vector in Rm the jth component of which is 1 and all the others 0.
Given (Xx+λei, Yx+λej, Zx+λej), (Xx0+λ0ej, Yx0+λ0ej, Zx0+λ0ej), (Xx, Yx, Zx) and(Xx0, Yx0, Zx0)solutions of the FBSDE (3.2.2), we define the processesNx,λ :=
(Xx+λej −Xx)/λ;Nx0,λ0 := (Xx0+λ0ej −Xx0)/λ0;Ux,λ:= (Yx+λej −Yx)/λ;
Ux0,λ0 := (Yx0+λ0ej−Yx0)/λ0andVx,λ:= (Zx+λej−Zx)/λ;Vx0,λ0 := (Zx0+λ0ej− Zx0)/λ0. Furthermore, for θ ∈ [0,1], λ > 0, x ∈ Rm, we define the processes Λx,λ := Xx+θλNx,λ,Γx,λ := Yx +θλUx,λ and∆x,λ := Zx+θλVx,λ. Let Ni,x,λ,Ui,x,λ, Vi,x,λ,Λi,x,λ,Γi,x,λand∆i,x,λbe theith component ofNx,λ, Ux,λ, Vx,λ,Λx,λ,Γx,λ and∆x,λ, respectively for each i= 1, . . . , m0. Let us first show that there exists a constantCindependent ofxandλsuch that
kNλk2S∞+kUλk2S∞+kVλ·Wk2BMO≤C. (3.4.8) Since
Ntx,λ=ei+
t
Z
0 1
Z
0
∂xbu(Xux+θ(Xux+λei−Xux), Yux+θ(Yux+λei−Yux))Nux,λdθ du
+
t
Z
0 1
Z
0
∂ybu(Xux+θ(Xux+λei−Xux), Yux+θ(Yux+λei−Yux))Uux,λdθ du,
and∂xband∂ybare bounded, it follows from Gronwall’s inequality that
|Ntx,λ| ≤ek1t
1 +k2TkUx,λkS∞
. (3.4.9)
We have Uti,x,λ=
1
Z
0
∂xhi(Λx,λT )NTx,λdθ+
T
Z
t 1
Z
0
∂zfui(∆i,x,λu )Vui,x,λ
+∂xliu(Λx,λu ,Γx,λu ,∆x,λu )Nux,λ+∂yliu(Λx,λu ,Γx,λu ,∆x,λu )Uux,λ +∂zliu(Λx,λu ,Γx,λu ,∆x,λu )Vux,λdθ du−
T
Z
t
Vui,x,λdWu.
Hence, similar to the proof of Theorem 3.2.2, we have Uti,x,λ+
T
Z
t
Vui,x,λdW˜ui
=
1
Z
0
∂xhi(Λx,λT )NTx,λdθ+
T
Z
t 1
Z
0
∂xliu(Λx,λu ,Γx,λu ,∆x,λu )Nux,λ
+∂yliu(Λx,λu ,Γx,λu ,∆x,λu )Uux,λ+∂zlui(Λx,λu ,Γx,λu ,∆x,λu )Vux,λdθ du, whereW˜i =W −R.
0ζsidswith|ζsi| ≤k3(1 + 2|(1−θ¯s)Zsi,x+ ¯θsZsi,x+λej|)for some predictable process θ¯s ∈ [0,1] is a Brownian motion under the equivalent measureP˜i =E(ζi·W)T ·P. Therefore similar to Theorem 3.2.2, with the same constantsc1, c2andC2,
kUx,λk2∞+c1kVx,λ·Wk2BMO
≤m0
16k22T2e2k1T k42+k27
+ 8k52T2
kUx,λk2∞+ 16m0e2k1T k42+k72 + 24√
3k26L24T1−εc2m0 Tε+ 2 + 2εL24c2C22
kVx,λ·Wk2BMO. Therefore, it follows from (3.4.3) that
kUx,λk2S∞ ≤ 16m0e2k1T k42+k27 1−m0 16k22T2e2k1T k24+k72
+ 8k25T2. (3.4.10) and
c1kVx,λk2BMO≤32m0e2k1T k24+k27
. (3.4.11)
Combining with (3.4.9), kNx,λkS∞ ≤ 1 +k2T
s 16m0e2k1T k24+k27 1−m0 16k22T2e2k1T k42+k72
+ 8k52T2
! ek1T. (3.4.12) Now, estimating the difference gives
|Ntx,λ−Ntx0,λ0|=
t
Z
0 1
Z
0
∂xbu(Λx,λu ,Γx,λu )Nux,λ+∂ybu(Λx,λu ,Γx,λu )Uux,λ
−∂xbu(Λxu0,λ0,Γxu0,λ0)Nux0,λ0 −∂ybu(Λxu0,λ0,Γxu0,λ0)Uux0,λ0dθdu
≤ Zt
0
Z1
0
|∂xbu(Λx,λu ,Γx,λu )||Nux,λ−Nux0,λ0|
+|∂xbu(Λx,λu ,Γx,λu )−∂xbu(Λxu0,λ0,Γxu0,λ0)||Nux0,λ0| +|∂ybu(Λx,λu ,Γx,λu )||Uux,λ−Uux0,λ0|
+|∂ybu(Λx,λu ,Γx,λu )−∂ybu(Λxu0,λ0,Γxu0,λ0)||Uux0,λ0|dθdu.
(3.4.13)
Then, using (B1) and (B6) and applying Gronwall’s lemma, we have kNx,λ−Nx0,λ0kS∞
≤ek1T k2TkUx,λ−Ux0,λ0kS∞+K(kNx0,λ0kS∞+kUx0,λ0kS∞)
2 kXx−Xx0kS∞
+K(kNx0,λ0kS∞+kUx0,λ0kS∞)
2 kXx+λej −Xx0+λ0ejkS∞ +K(kNx0,λ0kS∞+kUx0,λ0kS∞)
2 kYx−Yx0kS∞ +K(kNx0,λ0kS∞ +kUx0,λ0kS∞)
2 kYx+λej−Yx0+λ0ejkS∞
! . On the other hand,
Uti,x,λ−Uti,x0,λ0+
T
Z
t
(Vui,x,λ−Vui,x0,λ0)dW˜ui
=
1
Z
0
∂xhi(Λx,λT )NTx,λ−∂xhi(XΛxT0,λ0)NTx0,λ0dθ
+
T
Z
t 1
Z
0
(∂zfui(∆i,x,λu )−∂zfui(∆i,xu 0,λ0))Vui,x0,λ0dθ du
+
T
Z
t 1
Z
0
∂xlui(Λx,λu ,Γx,λu ,∆x,λu )Nux,λ+∂ylui(Λx,λu ,Γx,λu ,∆x,λu )Uux,λ +∂zlui(Λx,λu ,Γx,λu ,∆x,λu )Vux,λ−∂xliu(Λxu0,λ0,Γxu0,λ0,∆xu0,λ0)Nux0,λ0
−∂yliu(Λxu0,λ0,Γxu0,λ0,∆xu0,λ0)Uux0,λ0−∂zliu(Λxu0,λ0,Γxu0,λ0,∆xu0,λ0)Vux0,λ0dθ du, where W˜i = W −R.
0ζsids is defined as above. Rearranging the terms on the right hand side such as in (3.4.13) using successively (B3), (B4), (B6) and using Cauchy-Schwarz’ inequality, similar to Theorem 3.2.2, with the same constants c1, c2, C2, we have
kUx,λ−Ux0,λ0k2∞+c1k(Vx,λ−Vx0,λ0)·Wk2BMO
≤m0
16k22T2e2k1T k42+k72
+ 16k52T2
kUx,λ−Ux0,λ0k2∞+I1+I2 + 24√
3k26L24T1−εc2m0 Tε+ 2 + 2εL24c2C22
k(Vx,λ−Vx0,λ0)·Wk2BMO, where
I1 = 4K2m0e2k1T(k42+k27)
kNx0,λ0kS∞+kUx0,λ0kS∞
2
kXx−Xx0kS∞
+kXx+λej−Xx0+λ0ejkS∞+kYx−Yx0kS∞+kYx+λej−Yx0+λ0ejkS∞2
,
I2 = 24m0K2
kNTx0,λ0k2S∞
kXTx−XTx0kS∞+kXTx+λej−Xx
0+λ0ej
T kS∞2
+
4c22L24kVx0,λ0·Wk2BMO+ 2T c2
kNTx0,λ0kS∞+kUTx0,λ0kS∞2
·
k(Zx−Zx0)·Wk2BMO+k(Zx+λej−Zx0+λ0ej)·Wk2BMO +
T c2kVx0,λ0·Wk2BMO+T2
kNTx0,λ0kS∞+kUTx0,λ0kS∞2
·
kXx−Xx0kS∞+kXx+λej−Xx0+λ0ejkS∞ +kYx−Yx0kS∞+kYx+λej−Yx0+λ0ejkS∞
2 .
Hence, it follows from the Equations (3.4.3), (3.4.5), (3.4.6), (3.4.7) and (3.4.8) that there exists a constantC >˜ 0which does not depend onx, x0 andλ, λ0 such that
kNx,λ−Nx0,λ0k2S∞+kUx,λ−Ux0,λ0k2S∞+k(Vx,λ−Vx0,λ0)·Wk2BMO
≤C˜ |x−x0|+|λ−λ0| .
This proves the differentiability ofx7→(Xx, Yx, Zx).
3.A Multidimensional BSDEs with terminal condition of bounded Malliavin derivative
In this section, we extend the existence result of Cheridito and Nam [16] to the multidimensional case where theithcomponent of the generator depends only on (y, zi). For simplicity, we prove the crucial boundedness ofZ in this setting and leave out the existence since it follows as in [16, Theorem 2.2]. We consider the BSDE
Yt=ξ+
T
Z
t
gu(Yu, Zu)du−
T
Z
t
ZudWu. (3.A.1)
We make the following assumptions:
(D1) g : Ω×[0, T]×Rm
0 ×Rm
0×d → Rm
0 is a continuous and measurable function such thatgti(y, z) = git(y, zi),i = 1, . . . , m0 and there exists a constantB ∈R+and a nondecreasing functionρ:R+→R+such that
gt(y, z)−gt(y0, z0) ≤B
y−y0
+ρ |z| ∨ z0
z−z0 for allt∈[0, T],y, y0 ∈Rm
0 andz, z0∈Rm
0×d.
(D2) g·(0,0)∈ H4 and there exist Borel-measurable functionsqij : [0, T]→
We first prove a useful lemma under the following stronger conditions:
(D1’) gis continuously differentiable in(y, z)is such thatgti(y, z) =gti(y, zi),
Lemma 3.A.1. If (D1’), (D2’) and (D3) hold, then the BSDE (3.A.1)admits a unique solution(Y, Z)∈ S4(Rm [1,∞). It follows from [28, Theorem 5.1 and Proposition 5.3] that the BSDE (3.A.1) has a unique solution(Y, Z)∈ S4(Rm
and for each fixedr, denoting(Utj,r, Vtj,r) = (DjrYt, DrjZt), then(Uj,r, Vj,r) is
Using the conditions (D1’), we have Utij,r=Drjξi+
It is easy to see that the unique solution of the above ODE is given by ujt =
Theorem 3.A.2. If (D1) - (D3) hold, then the BSDE(3.A.1)has a unique solution as in the proof of [16, Theorem 2.2] and in combination with [28, Proposition 5.1]
the result follows.
3.B Multidimensional BSDEs with superquadratic growth
In this section, we will drop the assumption that theithcomponent of the generator depends only on(y, zi). We obtain solvability for small time horizon. Under an ad-ditional condition on the growth functionρ, the existence result hold for arbitrarily large time horizon. We consider the BSDE
Yt=ξ+
We make the following assumptions:
(H1) g : Ω×[0, T]×Rm
0 ×Rm
0×d → Rm
0 is a continuous and measurable function such that there exists a constantB ∈ R+ and a nondecreasing functionρ:R+→R+such that
(H3) The terminal conditionξ ∈ D1,2(Rm0)and there exist constantsAij ≥0 such that
Dtjξi
≤Aij for alli= 1, . . . , m0;j= 1, . . . , d.
We first prove a useful lemma under the following stronger conditions:
(H1’) gis continuously differentiable in(y, z)is such that there exist constants B ∈R+, ρ∈R+such that
|∂ygt(y, z)| ≤B, |∂zgt(y, z)| ≤ρ, for allt∈[0, T],y, y0 ∈Rm
0 andz, z0∈Rm
0×d. (H2’) Condition (D2) holds for all(y, z)∈Rm
0 ×Rm
0×d.
Lemma 3.B.1. If (H1’), (H2’) and (H3) hold, then the BSDE (3.B.1) admits a unique solution(Y, Z)∈ S4(Rm
0)× H4(Rm
0×d), and
|Ztj|2≤
m0
X
i=1
A2ij +
T
Z
t
q2ij(s)e−(2B+ρ2+1)(T−s)ds
e(2B+ρ2+1)(T−t), P⊗dt-a.e.
Proof. By [16, Lemma 2.5], condition (H3) implies E|ξ|p < +∞, for all p ∈ [1,∞). It follows from [28, Theorem 5.1 and Proposition 5.3] that the BSDE (3.B.1) has a unique solution(Y, Z)∈ S4(Rm
0)×H4(Rm
0×d). Moreover,(Y, Z)∈ L1,2a (Rm0+m0×d)fori= 1, . . . , m0; j= 1, . . . , d,
(DjrYti, DjrZti) = (Utij,r, Vtij,r) P⊗dt⊗dr-a.e. andZtij =Utij,t P⊗dt-a.e., where
Utij,r= 0, Vtij,r= 0, for 0≤t < r ≤T,
and for each fixedr, denoting(Utj,r, Vtj,r) = (DjrYt, DrjZt), then(Uj,r, Vj,r) is the unique solution inS2(Rm
0)× H2(Rm
0×d)of the BSDE
Utj,r =Drjξ+
T
Z
t
∂ygs(Ys, Zs)Usj,r+∂zgs(Ys, Zs)Vsj,r+Drjgs(Ys, Zs)ds
−
T
Z
t
Vsj,rdWs.
Applying Itô’s formula to|Utj,r|2yields
Taking conditional expectation with respect toFtandP, using condition (H3)
|Utj,r|2≤E Gronwall’s inequality implies that
|Utj,r|2 ≤
where the constant λ ∈ R+ is chosen such that R
Rm0+m0×dβ(x)dx = 1. Set βn(x) :=nm0+m0×dβ(nx),n∈N\ {0}, and define
gtn(ω, x) :=
Z
Rm0+m0×d
˜
gt(ω, x0)βn(x−x0)dx0.
Then allgnsatisfy (H1’) and (H2’). Therefore by Lemma 3.B.1 there exist unique solutions (Yn, Zn) ∈ S4(Rm
0) × H4(Rm
0×d) to the BSDEs corresponding to (gn, ξ), and
|Ztn,j|2 ≤
m0
X
i=1
A2ij+
T
Z
t
qij2(s)e−(2B+ρ2(Q)+1)(T−s)ds
e(2B+ρ2(Q)+1)(T−t)
≤
m0
X
i=1
A2ij+
T
Z
0
qij2(s)ds
e(2B+ρ2(Q)+1)T. SinceT ≤ 2B+ρlog 22(Q)+1, we obtain
|Ztn,j|2 ≤2
m0
X
i=1
A2ij +
m0
X
i=1 T
Z
0
q2ij(s)ds
.
Thus, following the same procedure as in the proof of [16, Theorem 2.2] and in combination with [28, Proposition 5.1] the result follows.
Theorem 3.B.3. If (H1), (H3) hold andρis such thatP∞ n=0
log 2
2B+ρ2(2nQ)+1 > T, (H2) holds for all(y, z) ∈ Rm
0 ×Rm
0×d such that|z| ≤ 2NQwhere N is the smallest integer such thatPN
n=0
log 2
2B+ρ2(2nQ)+1 ≥T. Then the BSDE(3.B.1)has a unique solution inS4(Rm0)× H∞(Rm0×d)and
|Zt| ≤2NQ, P⊗dt-a.e.
Proof. From Theorem 3.B.2, the BSDE (3.B.1) has a unique solution inS4(Rm
0)×
H∞(Rm0×d)and|Zt| ≤Qon[T −2B+ρlog 22(Q)+1, T]. By Lemma 3.B.1, we have
|DrjYT− log 2 2B+ρ2(Q)+1
|2≤
m0
X
i=1
2|Aij|2+
m0
X
i=1 T
Z
0
2|qij(t)|2dt.
By similar arguments as in Lemma 3.B.1 and Theorem 3.B.2, the BSDE (3.B.1) has a unique solution inS4(Rm0)×H∞(Rm0×d)on[T−2B+ρlog 22(Q)+1−2B+ρlog 22(2Q)+1, T−
log 2
2B+ρ2(Q)+1]with terminal conditionYT− log 2
2B+ρ2(Q)+1, and
|DjrYT− log 2
2B+ρ2(Q)+1−2B+ρ2(2Q)+1log 2 |2≤
m0
X
i=1
22|Aij|2+
m0
X
i=1 T
Z
0
(22+ 2)|qij(t)|2dt,
|Zt| ≤2Q, t∈[T − log 2
2B+ρ2(Q) + 1− log 2
2B+ρ2(2Q) + 1, T − log 2
2B+ρ2(Q) + 1].
By recurrence, form≥2, the BSDE (3.B.1) has a unique solution inS4(Rm
0)× H∞(Rm0×d)on[T−Pm
n=0
log 2
2B+ρ2(2nQ)+1, T−Pm−1 n=0
log 2
2B+ρ2(2nQ)+1]with termi-nal conditionYT−Pm−1
n=0
log 2
2B+ρ2(2nQ)+1, and
|DrjYT−Pm n=0
log 2 2B+ρ2(2nQ)+1
|2≤
m0
X
i=1
2m+1|Aij|2+
m0
X
i=1 T
Z
0
(
m+1
X
k=1
2k)|qij(t)|2dt,
|Zt| ≤2mQ, t∈[T−
m
X
n=0
log 2
2B+ρ2(2nQ) + 1, T −
m−1
X
n=0
log 2
2B+ρ2(2nQ) + 1].
Hence the existence follows from a pasting argument. The uniqueness follows from a similar argument as in the proof of [16, Theorem 2.2].
Remark 3.B.4. Ifρ(x) = p
log(1 +x)for x ≥ 0, then P∞ n=0
log 2
2B+ρ2(2nQ)+1 =
∞. Indeed, sinceρ(2nQ)≤p
log(2n(1 +Q)), we have
∞
X
n=0
log 2
2B+ρ2(2nQ) + 1 ≥
∞
X
n=0
log 2
2B+ log(2n(1 +Q)) + 1
=
∞
X
n=0
log 2
2B+ log(1 +Q) +nlog 2 + 1 =∞.
BSDEs on Finite and Infinite Horizon with Time-delayed Generators
4.1 Introduction
In Delong and Imkeller [24, 25], the theory of backward stochastic differential equations (BSDEs) was extended to BSDEs with time delay generators (delay BS-DEs). These are non-Markovian BSDEs in which the generator at each positive time t may depend on the past values of the solutions. This class of equations turned out to have natural applications in pricing and hedging of insurance con-tracts, see Delong [23].
The existence result of Delong and Imkeller [24], proved for standard Lips-chitz generators and small time horizon, has been refined by dos Reis et al. [26]
who derived additional properties of delay BSDEs such as path regularity and ex-istence of decoupled systems. Furthermore, exex-istence of delay BSDE constrained above a given continuous barrier has been established by Zhou and Ren [77] in a similar setup. More recently, Briand and Elie [13] proposed a framework in which quadratic BSDEs with sufficiently small time delay in the value process can be solved.
In addition to the inherent non-Markovian structure of delay BSDEs, the dif-ficulty in studying these equations comes from that the inter-temporal changes of the value and control processes always depend on their entire past, hence making it hard to obtain boundedness of solutions or even BMO-martingale properties of the stochastic integral of the control process. This suggests that delay BSDEs can ac-tually be solved forward and backward in time and in this regard, share similarities with forward backward stochastic differential equations (FBSDEs), see Section 4.4 for a more detailed discussion.
The object of the present chapter is to study delay BSDEs in the case where the past values of the solutions are weighted with respect to some scaling function.
In economic applications, these weighting functions can be viewed as representing the perception of the past of an agent. For multidimentional BSDEs with possibly infinite time horizon, we derive existence, uniqueness and stability of delay BSDE in this weighting-function setting. In particular, we show that when the delay van-ishes, the solutions of the delay BSDEs converge to the solution of the BSDE with no delay, hence recovering a result obtained by Briand and Elie [13] for different types of delay. Moreover, we prove that in our setting existence and uniqueness also hold in the case of reflexion on a càdlàg barrier. We observe a link between delay BSDEs and coupled FBSDE and, based on the findings in chapter 3, we de-rive existence of delay quadratic BSDEs in the case where only the value process is subjected to delay. We refer to Briand and Elie [13] for a similar result, again for a different type of delay and in the one-dimensional case.
In the next section, we specify our probabilistic structure and the form of the equation, then present existence, uniqueness and stability results. Sections 4.3 and 4.4 are dedicated to the study of reflected delay BSDEs and quadratic and superquadratic BSDEs with delay in value process, respectively.
4.2 BSDEs with time delayed generators
We work on a filtered probability space(Ω,F,(Ft)t∈[0,T], P) with T ∈ (0,∞].
We assume that the filtration is generated by ad-dimensional Brownian motion W and it is complete and right continuous. Let us also assume that F = FT. We endow Ω×[0, T] with the predictable σ-algebra and Rk with its Borel σ-algebra. Unless otherwise stated, all equalities and inequalities between random variables and stochastic processes will be understood in theP-a.s. and P ⊗ dt-a.e. sense, respectively. Forp ∈ [1,∞) andm ∈ N, we denote by Sp(Rm) the space of predictable and continuous processesXvalued inRmsuch thatkXkpSp:=
E[(supt∈[0,T]|Xt|)p] < ∞and byHp(Rm)the space of predictable processesZ valued inRm×d such thatkZkpHp := E[(RT
0 |Zu|2 du)p/2] < ∞. For a suitable integrandZ, we denote byZ ·W the stochastic integral(Rt
0ZudWu)t∈[0,T]ofZ with respect toW. From Protter [68],Z·W defines a continuous martingale for everyZ ∈ Hp(Rm). Processes(φt)t∈[0,T]will always be extended to[−T,0)by settingφt= 0fort∈[−T,0). We equipRwith theσ-algebraB(R)consisting of Borel sets of the usual real line with possible addition of the points−∞,+∞, see Bogachev [11].
Letξ be anFT-measurable terminal condition and ganRm-valued function.
Given two measuresα1 andα2 on[−∞,∞], and two weighting functionsu, v : [0, T]→R, we study the existence of the BSDE
Yt=ξ+
T
Z
t
g(s,Γ(s))ds−
T
Z
t
ZsdWs, t∈[0, T], (4.2.1)
where
Γ(s) :=
0
Z
−T
u(s+r)Ys+rα1(dr),
0
Z
−T
v(s+r)Zs+rα2(dr)
. (4.2.2) Example 4.2.1. 1. BSDE with infinite horizon: Ifu =v = 1andα1 =α2 =δ0 the Dirac measure at0, then Equation (4.2.1) reduces to the classical BSDE with infinite time horizon and stantard Lipschitz generator.
2.Pricing of insurance contracts:Let us consider the pricing problem of an insur-ance contractξwritten on a weather derivative. It is well know, see for instance [3]
that such contracts can be priced by investing in a highly correlated, but tradable derivative. In the Merton model, assuming that the latter asset has dynamics
dSt=St(µtdt+σtdWt),
then the insurer chooses a numberπtof shares ofSto buy at timetand fixes a cost ctto be paid by the client. Hence, he seeks to find the priceV0such that
dVt=ctdt+πtσt(dWt+θtdt)
withθt=σ0t(σtσt)−1µt. It is natural to demand the costctat timetto depend on the past values of the insurance premiumVt, for instance to account for historical weather data. A possible cost criteria is
ct:=Mt Z0
−T
cos(2π
P (t+s))Vt+sds
whereP accounts for the weather periodicity andM is a scaling parameter. Thus, the insurance premium satisfies the delay BSDE
Vt=ξ+
T
Z
t
0
Z
−T
Mucos(2π
P (u+s))Vu+sds+Zuσuθu
du−
T
Z
t
ZudWu. ♦
4.2.1 Existence
Our existence result for the BSDE (4.2.1) is obtained under the following assump-tions:
(A1) α1, α2are two deterministic, finite valued measures supported on[−T,0].
(A2) u, v : [0, T]→ Rare Borel measurable functions such thatu ∈ L1(dt) andv∈L2(dt).
(A3) g: Ω×[0, T]×Rm×Rm×d→Rmis measurable, such thatRT
0 g(s,0,0)ds∈ L2(Rm)and satisfies the standard Lipschitz condition: there exists a con-stantK >0such that
|g(t, y, z)−g(t, y0, z0)| ≤K(|y−y0|+|z−z0|) for everyy, y0 ∈Rmandz, z0∈Rm×d.
(A4) ξ ∈L2(Rm)and isFT-measurable.
Theorem 4.2.2. Assume (A1)-(A4). If
(K2α21([−T,0])kuk2L1(dt)≤ 251 ,
K2α22([−T,0])kvk2L2(dt)≤ 251, (4.2.3) then BSDE(4.2.1)admits a unique solution(Y, Z)∈ S2(Rm)× H2(Rm×d).
For the proof we need the following lemma ona prioriestimates of solutions of (4.2.1).
Lemma 4.2.3 (A priori estimation). Assume (A1)-(A3). For everyξ,ξ¯∈L2(Rm), (y, z),(¯y,z)¯ ∈ S2(Rm)× H2(Rm×d)and(Y, Z),( ¯Y ,Z)¯ ∈ S2(Rm)× H2(Rm×d) satisfying
(Yt=ξ+RT
t g(s, γ(s))ds−RT
t ZsdWs
Y¯t= ¯ξ+RT
t g(s,γ(s))ds¯ −RT
t Z¯sdWs, t∈[0, T] with
γ(s) = R0
−T u(s+r)ys+rα1(dr),R0
−Tv(s+r)zs+rα2(dr)
¯ γ(s) =
R0
−T u(s+r)¯ys+rα1(dr),R0
−Tv(s+r)¯zs+rα2(dr)
. Then, one has
kY −Y¯k2S2(Rm)+kZ−Zk¯ 2H2(Rm×d)
≤20K2α21([−T,0])kuk2L1(dt)ky−yk¯ 2S2(Rm)
+ 10 ξ−ξ¯
2
L2(Rm)+ 20K2α22([−T,0])kvk2L2(dt)kz−zk¯ 2H2(Rm×d). Proof. Let(y, z)∈ S2(Rm)× H2(Rm×d), by assumptions (A1) and (A3), using
2ab≤a2+b2 and [26, Lemma 1.1], we have and taking conditional expectation with respect toFtyields
Yt−Y¯t=E
By Doob’s maximal inequality and2ab≤a2+b2, we obtain side, taking square and expectation to both sides and2ab≤a2+b2, we have E
≤2K2α21([−T,0])kuk2L1(dt)ky−yk¯ 2S2+ 2K2α22([−T,0])kvk2L2(dt)kz−zk¯ 2H2. Hence,
kY −Y¯k2S2(Rm)+kZ−Zk¯ 2H2(Rm×d)≤20K2α21([−T,0])kuk2L1(dt)ky−yk¯ 2S2(Rm)
10E
|ξ−ξ|¯2
+ 20K2α22([−T,0])kvk2L2(dt)kz−zk¯ 2H2(Rm×d).
This concludes the proof.
Proof ( of Theorem 4.2.2). Let(y, z)∈ S2(Rm)× H2(Rm×d)and define the pro-cess γ(s) :=
R0
−Tu(s+r)ys+rα1(dr),R0
−T v(s+r)zs+rα2(dr)
. Similar to Lemma 4.2.3, it follows from (A1)-(A4) that
E
ξ+
T
Z
0
g(s, γ(s))ds
2
<∞.
According to the martingale representation theorem, there exists a unique Z ∈ H2(Rm×d)such that for allt∈[0, T],
E
ξ+
T
Z
0
g(s, γ(s))ds
Ft
=E
ξ+
T
Z
0
g(s, γ(s))ds
+
t
Z
0
ZsdWs. Putting
Yt:=E
ξ+
T
Z
t
g(s, γ(s))ds Ft
, 0≤t≤T, the pair(Y, Z)belongs toS2(Rm)× H2(Rm×d)and satisfies
Yt=ξ+
T
Z
t
g(s, γ(s))ds−
T
Z
t
ZsdWs, 0≤t≤T.
Thus we have constructed a mappingΦfromS2(Rm)× H2(Rm×d)to itself such thatΦ(y, z) = (Y, Z). Let(y, z),(¯y,z)¯ ∈ S2(Rm)× H2(Rm×d), and(Y, Z) = Φ(y, z),( ¯Y ,Z) = Φ(¯¯ y,z). By Lemma 4.2.3, we have¯
kY −Y¯k2S2(Rm)+kZ−Zk¯ 2H2(Rm×d)≤10K2α21([−T,0])kuk2L1(dt)ky−yk¯ 2S2(Rm)
+ 10K2α22([−T,0])kvk2L2(dt)kz−zk¯ 2H2(
Rm×d)
so that if condition (4.2.3) is satisfied,Φis a contraction mapping which therefore admits a unique fixed point on the Banach space S2(Rm) × H2(Rm×d). This
completes the proof.
4.2.2 Stability
In this subsection, we study stability of the BSDE (4.2.1) with respect to the delay measures. In particular, in Corollary 4.2.5 below we give conditions under which a sequence of solutions of BSDEs with time delayed generator converges to the solution of a standard BSDE with no delay. Given two measuresαandβ, we write α≤βifα(A)≤β(A)for every measurable setA.
Theorem 4.2.4. Assume (A2)-(A4). Fori = 1,2andn∈ N, letαni, αi be mea-sures satisfying (A1); with αni satisfying(4.2.3)in Theorem 4.2.2 and such that αni([−T,0]) converges toαi([−T,0]). Ifαn1 ≤ α1 (orα1 ≤ αn1) and αn2 ≤ α2 (orα2 ≤ αn2), thenkYn−YkS2(Rm) → 0andkZn−ZkH2(Rm×d) → 0, where (Yn, Zn) and(Y, Z) are solutions of the BSDE (4.2.1)with delay given by the measures(αn1, αn2)and(α2, α2), respectively.
Proof. From Theorem 4.2.2, for everyn, there exists a unique solution(Yn, Zn) to the BSDE (4.2.1) with delay given by the measures(αn1, αn2). Sinceαni, i = 1,2satisfy (4.2.3) in Theorem 4.2.2 andαin([−T,0])converges toαi([−T,0]), it follows thatαisatisfy (4.2.3) and by Theorem 4.2.2 there exists a unique solution (Y, Z)to the BSDE with delay given by(α1, α2). Using it follows similar to the proof of Lemma 4.2.3 that
E
+ 2K2E i= 1,2, define positive measures satisfying (A1). Therefore,
E Using [26, Lemma 1.1], we obtain
E Similarly, for the control processes we have
E
Hence
kYn−Yk2S2(Rm)+kZn−Zk2H2(Rm×d)
≤20K2(αn1([−T,0]))2kuk2L1(dt)kYn−Yk2S2(Rm)
+ 20K2((αn1 −α1)([−T,0]))2kuk2L1(dt)kYk2S2(Rm)
+ 20K2(αn2([−T,0]))2kvk2L2(dt)kZn−Zk2H2(Rm×d)
+ 20K2((αn2 −α2)([−T,0]))2kvk2L2(dt)kZk2H2(Rm×d)
≤ 4
5kYn−Yk2S2(Rm)+4
5kZn−Zk2H2(Rm×d)
+ 20K2((αn1 −α1)([−T,0]))2kuk2L1(dt)kYk2S2(Rm)
+ 20K2((αn2 −α2)([−T,0]))2kvk2L2(dt)kZk2H2(
Rm×d).
Therefore, the result follows from the convergence ofαni([−T,0]),i= 1,2.
The following is a direct consequence of the above stability result. We denote by δ0the Dirac measure at0.
Corollary 4.2.5. Assume (A2)-(A4). Fori= 1,2andn∈Nletαni be measures satisfying (A1) and(4.2.3)in Theorem 4.2.2 and such thatαni([−T,0])converges to1. Ifδ0 ≤α1n(orαn1 ≤δ0) andδ0 ≤αn2 (orαn2 ≤δ0), thenkYn−YkS2(Rm)→ 0andkZn−ZkH2(Rm×d)→0, where(Yn, Zn)is the solution of the BSDE(4.2.1) with delay given by(αn1, αn2)and(Y, Z)is the solution of BSDE without delay.
We conclude this section with the following counterexample which shows that the conditionα1 ≤ αn1 (orαn1 ≤ α1) andα2 ≤ α2n(or α2n ≤ α2) is needed in the above theorem.
Example 4.2.6. Assume thatm = d = 1. We denote by δ0 andδ−1 the Dirac measures at 0 and−1, respectively. It is clear that δ0([−1,0]) = δ−1([−1,0]).
Consider the delay BSDEs Yt= 1 +
1
Z
t
1/5
0
Z
−1
Ys+r+Zs+r
δ0(dr)ds−
1
Z
0
ZsdWs (4.2.5) and
Y¯t= 1 +
1
Z
t
1/5
0
Z
−1
Y¯s+r+ ¯Zs+r
δ−1(dr)ds−
1
Z
0
Z¯sdWs. (4.2.6)
Since BSDE (4.2.6) takes the formY¯t= 1−R1
t Z¯udWs, it follows thatY¯t= 1for allt ∈ [0,1]. On the other hand, (4.2.5) is the standard BSDE without delay, its solution can be written asYt=E[H1t| Ft], where the deflator(Hst)s≥tat timetis given bydHst=−H5st(ds+dWs). Thus,Yt=exp(−1/5(1−t))and fort∈[0,1),
Yt<Y¯t. ♦
4.3 Reflected BSDEs with time-delayed generators
The probabilistic setting and the notation of the previous section carries over to the present one. In particular, we fix a time horizonT ∈(0,∞]and we assumem= 1.
Forp∈[1,∞), we further introduce the spaceMp(R)of adapted càdlàg processes Xvalued inRsuch thatkXkpMp :=E[(supt∈[0,T]|Xt|)p]<∞and byAp(R), we denote the subspace of elements ofMp(R)which are increasing processes starting at0. Let (St)t∈[0,T] be a càdlàg adapted real-valued process. In this section, we study existence of solutions(Y, Z, K)of BSDEs reflected on the càdlàg barrierS and with time-delayed generators. That is, processes satisfying
Yt =ξ+ ZT
t
g(s,Γ(s))ds+KT −Kt− ZT
t
ZsdWs, t∈[0, T](4.3.1)
Y ≥S (4.3.2)
RT
0 (Yt−−St−)dKt= 0 (4.3.3)
withΓdefined by (4.2.2). Consider the condition (A5) E
sup0≤t≤T(St+)2
<∞andST ≤ξ.
Theorem 4.3.1. Assume (A1)-(A5). If
(K2α21([−T,0])kuk2L1(dt)≤ 361,
K2α22([−T,0])kvk2L2(dt) ≤ 361, (4.3.4) then RBSDE (4.3.1)admits a unique solution (Y, Z, K) ∈ M2(R)× H2(Rd)× A2(R)satisfying
Yt= ess sup
τ∈Tt
E
τ
Z
t
g(s,Γ(s))ds+Sτ1{τ <T}+ξ1{τ=T}
Ft
,
whereT is the set of all stopping times taking values in[0, T]andTt={τ ∈ T : τ ≥t}.
Proof. For any given(y, z) ∈ M2(R)× H2(Rd), similar to the proof of Lemma 4.2.3, we have
E
ξ+
T
Z
0
g(s, γ(s))ds
2
<∞
withγdefined as in Lemma 4.2.3. Hence, from [50, Theorem 3.3] forT <∞and [1, Theorem 3.1] forT =∞the reflected BSDE
Yt=ξ+
T
Z
t
g(s, γ(s))ds+KT −Kt−
T
Z
t
ZsdWs
with barrierSadmits a unique solution(Y, Z, K)such that(Y, Z)∈ B, the space of processes (Y, Z) ∈ M2(R)× H2(Rd) such that Y ≥ S, andK ∈ A2(R).
Moreover,Y admits the representation Yt= ess sup Doob’s maximal inequality implies that
E
Since(Y, K)and( ¯Y ,K)¯ satisfy (4.3.2) and (4.3.3), we have
|Yt−Y¯t|2+
T
Z
t
|Zs−Z¯s|2ds≤2
T
Z
t
(Ys−Y¯s)(g(s, γ(s))−g(s,γ(s)))ds¯
−2
T
Z
t
(Ys−Y¯s)(Zs−Z¯s)dWs. Hence
E
T
Z
0
|Zs−Z¯s|2ds
≤E
"
sup
0≤t≤T
|Yt−Y¯t|2
# +E
T
Z
0
|g(s, γ(s))−g(s,γ¯(s))|ds
2
. In view of the proof of Lemma 4.2.3, we deduce
kY −Y¯k2M2(R)+kZ−Z¯k2H2(Rd)≤9E
T
Z
0
|g(s, γ(s))−g(s,¯γ(s))|ds
2
≤18K2α21([−T,0])kuk2L1(dt)ky−yk¯ 2M2(R)
+ 18K2α22([−T,0])kvk2L2(dt)kz−zk¯ 2H2(Rd).
By condition (4.3.4),Φis a contraction mapping and therefore it admits a unique fixed point which combined with the associated processK is the unique solution
of the RBSDE (4.3.1).
4.4 Quadratic and superquadratic BSDEs with delay in value process
In this section, we study quadratic and superquadratic BSDEs with delay in value process through the connection between BSDEs with time-delayed generators and FBSDEs. We work in the probabilistic setting and with the notation of Section 4.2.
Standard methods to solve BSDEs with quadratic growth in the control variable often rely either on boundedness of the control process, see for instance [69] and [16], or on BMO estimates for the stochastic integral of the control process, see for instance [73]. However, as shown in [24], solutions of BSDEs with time-delayed generators do not, in general, satisfy boundedness and BMO properties so that new methods are required to solve quadratic BSDE with time-delayed generators.
Recently, [13] obtained existence and uniqueness of solution for a quadratic BSDE
with delay only in the value process. We show below that using FBSDE theory, it is possible to generalize their results to multidimension and considering a different kind of delay. Moreover, our argument allows to solve equations with generators of superquadratic growth.
Let α1 be the uniform measure on [−T,0], α2 the Dirac measure at 0. Put u(s) = v(s) = 1, fors ∈ [0, T]. We are considering the following BSDE with time delay only in the value process:
Yt=ξ+
T
Z
t
g(s,
s
Z
0
Yrdr, Zs)ds−
T
Z
t
ZsdWs, t∈[0, T]. (4.4.1)
We denote byD1,2 the space of all Malliavin differentiable random variables and forξ ∈ D1,2 denote byDtξ its Malliavin derivative. We refer to Nualart [61] for a thorough treatment of the theory of Malliavin calculus, whereas the definition and properties of the BMO-space and norm can be found in [47]. We make the following assumptions:
(B1) g : [0, T]×Rm ×Rm×d → Rm is a continuous function such that gi(y, z) =gi(y, zi)and there exists a constantK >0as well as a nonde-creasing functionρ:R+→R+such that
|g(s, y, z)−g(s, y0, z0)| ≤K|y−y0|+ρ(|z| ∨ |z0|)|z−z0|,
|g(s, y, z)−g(s, y0, z)−g(s, y, z0) +g(s, y0, z0)| ≤K(|y−y0|+|z−z0|) for alls∈[0, T],y, y0∈Rmandz, z0 ∈Rm×d.
(B2) ξ is FT-measurable such that ξ ∈ D1,2(Rm) and there exist constants Aij ≥0such that
|Djtξi| ≤Aij, i= 1, . . . , m; j = 1, . . . , d, for allt∈[0, T].
(B3) g: Ω×[0, T]×Rm×Rm×dis measurable,g(s, y, z) =f(s, z)+l(s, y, z) where f and lare measurable functions with fi(s, z) = fi(s, zi), i = 1, . . . , mand there exists a constantK ≥0such that
|f(s, z)−f(s, z0)| ≤K(1 +|z|+|z0|)|z−z0|,
|l(s, y, z)−l(s, y0, z0)| ≤K|y−y0|+K(1 +|z|+|z0|)|z−z0|,
|f(s, z)| ≤K(1 +|z|2),
|l(s, y, z)| ≤K(1 +|z|1+),
for some0≤ <1and for alls∈[0, T],y, y0 ∈Rmandz, z0∈Rm×d.
(B4) ξ is FT-measurable such that there exist a constant K ≥ 0 such that
|ξ| ≤K.
(B5) g : Ω×[0, T]×R×Rd → Ris progressively measurable, continuous process for any choice of the spatial variables and for each fixed(s, ω)∈ [0, T]×Ω, g(s, ω,·) is continuous. g is increasing in y and for some constantK ≥0such that
|g(s, y, z)| ≤K(1 +|z|), for alls∈[0, T],y∈Randz∈Rd.
(B6) ξisFT-measurable such thatξ ∈L2.
(B7) g : Ω×[0, T]×R×Rd → Ris progressively measurable, continuous process for any choice of the spatial variables and for each fixed(s, ω)∈ [0, T]×Ω, g(s, ω,·) is continuous. g is increasing in y and for some constantK ≥0such that
|g(s, y, z)| ≤K(1 +|z|2), for alls∈[0, T],y∈Randz∈Rd.
Proposition 4.4.1. AssumeT ∈(0,∞).
1. If (B1)-(B2) are satisfied, then there exists a constant C ≥ 0 such that for sufficiently small T, BSDE (4.4.1) admits a unique solution (Y, Z) ∈ S2(Rm)× H2(Rm×d)such that|Z| ≤C.
2. If (B3)-(B4) are satisfied, then there exist constantsC1, C2 ≥ 0such that for sufficiently small T, BSDE (4.4.1) admits a unique solution (Y, Z) ∈ S2(Rm)× H2(Rm×d)such that|Y| ≤C1andkZ·dWkBMO ≤C2. 3. Ifm=d= 1and (B5)-(B6) are satisfied, then BSDE(4.4.1)admits at least
a solution(Y, Z)∈ S2(R)× H2(Rd).
4. If m = d = 1 and (B4) and (B7) are satisfied, then BSDE(4.4.1)admits at least a solution(Y, Z) ∈ S2(R)× H2(Rd)such thatY is bounded and Z·W is a BMO martingale.
Proof. Define the function b : Rm → Rm by setting for y ∈ Rm, bi(y) = yi, i= 1, . . . , m. Fort∈[0, T], put
Xt=
t
Z
0
b(Ys)ds.
Thus BSDE (4.4.1) can be written as the coupled FBSDE (Xt=Rt
0b(Ys)ds, Yt=ξ+RT
t g(s, Xs, Zs)ds−RT
t ZsdWs
(4.4.2) so that 1. and 2. follow from chapter 3, and 3. and 4. from [5].
The above theorem provides an explanation why it is not enough to solve a time-delayed BSDE backward in time, one actually needs to consider both the forward and backward parts of the solution due to the delay.
Appendix
A.1 BMO martingales
We recall some results and properties of BMO martingales, for a thorough treat-ment, we refer to Kazamaki [47]. For any uniformly integrable martingaleM with M0= 0andp∈[1,∞), define
kMkBM Op := sup
τ∈T
kE[hMiT − hMiτ|p2Fτ]1pk∞.
We will useBM Op(P)when it is necessary to indicate the underlying probability measure, and just writeBM O whenp = 2. We recall the following results from the literature.
Lemma A.1.1. LetM be a BMO martingale. Then we have:
(1) The stochastic exponentialE(M)is uniformly integrable.
(2) There exists a numberr >1such thatE(M)T ∈Lr. This property follows from the Reverse Hölder inequality. The maximalr with this property can be expressed explicitly in terms of theBM O norm of M. There exists as well an upper bound forkE(M)TkrLr depending only onT, rand theBM O norm ofM.
(3) For probability measuresP andQ satisfyingdQ = E(M)TdP for M ∈ BM O(P), the processMˆ =M − hMiis aBM O(Q)martingale.
(4) Energy inequalities imply the inclusionBM O ⊂ Hp for allp ≥ 1. More precisely, for M = R
αdW with BM O norm C, the following estimate
αdW with BM O norm C, the following estimate