Processing math: 100%
Research article Special Issues

A sufficient maximum principle for backward stochastic systems with mixed delays


  • In this paper, we study the problem of optimal control of backward stochastic differential equations with three delays (discrete delay, moving-average delay and noisy memory). We establish the sufficient optimality condition for the stochastic system. We introduce two kinds of time-advanced stochastic differential equations as the adjoint equations, which involve the partial derivatives of the function f and its Malliavin derivatives. We also show that these two kinds of adjoint equations are equivalent. Finally, as applications, we discuss a linear-quadratic backward stochastic system and give an explicit optimal control. In particular, the stochastic differential equations with time delay are simulated by means of discretization techniques, and the effect of time delay on the optimal control result is explained.

    Citation: Heping Ma, Hui Jian, Yu Shi. A sufficient maximum principle for backward stochastic systems with mixed delays[J]. Mathematical Biosciences and Engineering, 2023, 20(12): 21211-21228. doi: 10.3934/mbe.2023938

    Related Papers:

    [1] Edward J. Allen . Derivation and computation of discrete-delayand continuous-delay SDEs in mathematical biology. Mathematical Biosciences and Engineering, 2014, 11(3): 403-425. doi: 10.3934/mbe.2014.11.403
    [2] H. J. Alsakaji, F. A. Rihan, K. Udhayakumar, F. El Ktaibi . Stochastic tumor-immune interaction model with external treatments and time delays: An optimal control problem. Mathematical Biosciences and Engineering, 2023, 20(11): 19270-19299. doi: 10.3934/mbe.2023852
    [3] Yuanpei Xia, Weisong Zhou, Zhichun Yang . Global analysis and optimal harvesting for a hybrid stochastic phytoplankton-zooplankton-fish model with distributed delays. Mathematical Biosciences and Engineering, 2020, 17(5): 6149-6180. doi: 10.3934/mbe.2020326
    [4] Chaofeng Zhao, Zhibo Zhai, Qinghui Du . Optimal control of stochastic system with Fractional Brownian Motion. Mathematical Biosciences and Engineering, 2021, 18(5): 5625-5634. doi: 10.3934/mbe.2021284
    [5] An Ma, Shuting Lyu, Qimin Zhang . Stationary distribution and optimal control of a stochastic population model in a polluted environment. Mathematical Biosciences and Engineering, 2022, 19(11): 11260-11280. doi: 10.3934/mbe.2022525
    [6] Tingting Yu, Sanling Yuan . Dynamics of a stochastic turbidostat model with sampled and delayed measurements. Mathematical Biosciences and Engineering, 2023, 20(4): 6215-6236. doi: 10.3934/mbe.2023268
    [7] Sheng Wang, Lijuan Dong, Zeyan Yue . Optimal harvesting strategy for stochastic hybrid delay Lotka-Volterra systems with Lévy noise in a polluted environment. Mathematical Biosciences and Engineering, 2023, 20(4): 6084-6109. doi: 10.3934/mbe.2023263
    [8] Yan Wang, Tingting Zhao, Jun Liu . Viral dynamics of an HIV stochastic model with cell-to-cell infection, CTL immune response and distributed delays. Mathematical Biosciences and Engineering, 2019, 16(6): 7126-7154. doi: 10.3934/mbe.2019358
    [9] ZongWang, Qimin Zhang, Xining Li . Markovian switching for near-optimal control of a stochastic SIV epidemic model. Mathematical Biosciences and Engineering, 2019, 16(3): 1348-1375. doi: 10.3934/mbe.2019066
    [10] Zeyan Yue, Sheng Wang . Dynamics of a stochastic hybrid delay food chain model with jumps in an impulsive polluted environment. Mathematical Biosciences and Engineering, 2024, 21(1): 186-213. doi: 10.3934/mbe.2024009
  • In this paper, we study the problem of optimal control of backward stochastic differential equations with three delays (discrete delay, moving-average delay and noisy memory). We establish the sufficient optimality condition for the stochastic system. We introduce two kinds of time-advanced stochastic differential equations as the adjoint equations, which involve the partial derivatives of the function f and its Malliavin derivatives. We also show that these two kinds of adjoint equations are equivalent. Finally, as applications, we discuss a linear-quadratic backward stochastic system and give an explicit optimal control. In particular, the stochastic differential equations with time delay are simulated by means of discretization techniques, and the effect of time delay on the optimal control result is explained.



    The forward stochastic differential equation characterizes the dynamic changes of state processes with given initial state trajectories. A backward stochastic differential equation (BSDE) can be seen as an Itô stochastic differential equation in which the terminal rather than the initial condition is given. BSDE will admit a pair of adapted process under some conditions on the coefficient and the terminal value of the BSDE. A stochastic differential equation coupled with a BSDE that has both initial and terminal conditions are known as forward-backward stochastic differential equations. In contrast to the development of forward stochastic differential equations, the study of backward stochastic differential equations (BSDEs) have been started in the last three decades. Pardoux and Peng [1] first proposed BSDEs. Since then, BSDEs have been systematically studied due to their wide applications in social fields such as applied mathematics, engineering and finance [2,3,4,5,6,7]. A BSDE can be viewed as an Itô's stochastic differential equation, where the endpoints are given except for the initial conditions. It is worth pointing out that the solution of BSDE includes two processes, which is the essential difference from forward stochastic differential equation. Due to the wide-range of applications and the interesting properties and structure of BSDE, it is valuable to study the optimal control problem of such equations.

    There is a lot of research focused on stochastic optimal control problems. The stochastic optimal control for BSDE was studied by Dokuchaer and Zhou [8]. Subsequently, Lim and Zhou [9] solved the linear-quadratic control problem for BSDE, and two alternative expressions for optimal control are obtained. Huang et al. [10] obtained a stochastic maximum principle for partial information control problem of backward stochastic systems. Zhen et al. [11] discussed the optimal control for fully coupled forward-backward stochastic control systems. They obtained the necessary and sufficient stochastic maximum principles. [12] focused on nonzero sum differential games for BSDEs. They established necessary and sufficient conditions for Nash equilibrium points (see [13] and [14] for the game problems of BSDE). For the control problems of BSDEs, see also [15,16,17,18,19,20].

    In the above problems, the BSDEs depend only on the value of the current time. Control systems with time-delay have existed for several decades, which are widely applied in financial areas. For example, investment portfolio management, pricing, hedging problems and automobile traffic model [21,22,23]. It will be more realistic for an investor to better strategize his investment when considering a delay factor. [24] introduced the time-advanced stochastic differential equations. [25] extended for the first time the applicability of the optimal homotopy asymptotic method to find the algorithm of approximate analytic solution of delay differential equations. However, the control system contained only discrete delay. [26] generalized the results to partial information. Apart from the discrete delays, it is more reasonable and practical to include moving average delays, which indicate that the current situation depends on a past period of history in the form of an integral. As an example of an optimal consumption problem, at time t, let x(t) and u(t) be wealth and consumption, respectively. It is reasonable to assume that the incremental wealth is a combination of the present value x(t) plus some sliding average of previous value 0δeλsx(t+s)ds and negative consumption amount u(t). Therefore, it is meaningful to incorporate moving-average delays, which indicates that the current situation of the system depends on a past period of history in integral form. Moreover, in [27], the authors introduced another delay, the noisy memory or Browinan delay, which is an Itô's integral over a past time periods. Indeed, noisy memories arise naturally in economics and finance, e.g., in modeling stochastic volatility (see, e.g., [28]). [29] obtained necessary and sufficient conditions for mean-field BSDEs with noisy memory over an infinite time span using Malliavin calculus technique. They also applied the theoretical results to pension fund problem. Zhang [31] established the sufficient maximum principle for optimal control problems with three delays and introduced a unified adjoint equation.

    In this paper, we consider a BSDE with three delays: The discrete delay, the moving-average delay and the noisy memory. Our work differs from the above in several ways.

    1) Compared with [24] and [29], three kinds of delays are incorporated in the optimal control problem: the discrete delay and the moving-average delay and noisy memory, resulting in two kinds of adjoint equations, which will increase the difficulty of analyzing the maximum condition. Thus, our optimal control system can describe more stochastic delay phenomena and cover more backward stochastic delay optimization control problems.

    2) We give one kind of a 2-dimensional time-advanced stochastic differential equation as our adjoint equation in a straightforward form. More importantly, we show that the two kinds of adjoint equations are equivalent. It is also emphasized that only the solution P1(t) of (3.2) is needed in the theory of the sufficient maximum principle, while P2(t) is an auxiliary process.

    3) As we know from [30], the four-step scheme is used to investigate the adapted solutions for forward-backward stochastic differential equations. Unfortunately, we study the optimal control problem for the state governed by BSDE with delay, that is to say, the state is past-dependence, so the system is not Markovian, and we cannot derive a certain system to get the explicit solution of the state Y and Z. In order to obtain the effect of delay, we divide the interval according to the delay, and then continue to subdivide each subinterval to get the numerical solution using Euler method or improved Euler method.

    This paper is organized as follows. Section 2 presents the problem statement and some preliminaries. In Section 3, we introduce two kinds of adjoint equations, obtain the sufficient maximum principle in Section 4. In Section 5, we apply the theoretical results to linear-quadratic systems and derive explicit expressions for the optimal control. In addition, we study an optimal consumption problem. Finally, we conclude the paper with concluding remarks.

    Assume that T is a finite time horizon, λ is a constant, and δ is a positive constant time delay such that δ<T. Let (Ω,F,(Ft)0tT,P) be a complete filtered probability space on which one-dimensional standard Brownian motion is defined. Moveover, it is also assumed that {Ft}0tT is a natural filtration generated by W(t) and FT=F. Let E[] be the expectation under P. Denote Et[]=E[|Ft] for each t[0,T].

    Throughout this paper, we denote by L2(Ft;R) the space of R-valued Ft-measurable random variables η satisfying E|η|2<+, by L2F(r,s;R) the space of R-valued Ft-adapted process ψ() satisfying Esr|ψ(t)|2dt<+. Let ϕ():[δ,0]R be given deterministic functions such that supδt0|ϕ(t)|2<+. Let U be a nonempty convex control subset of R. We introduce the admissible control set U={u()L2F(0,T;R)|u(t)U,t[0,T]}.

    Consider the following BSDE with delay:

    {dy(t)=f(t,y(t),z(t),u(t))dtz(t)dW(t),y(T)=ξ,y(t)=ϕ(t),t[δ,0), (2.1)

    where y(t):=(y(t),y1(t),y2(t),y3(t)), y1(t):=ttδeλ(st)y(s)ds, y2(t):=ttδy(s)dW(s) and y3(t):=y(tδ). y1(t),y2(t),y3(t) are called the moving-average delayed state, the noisy memory process and the discrete delayed state, respectively. Here f:Ω×R×R×R×R×R×RR is a given continuous function in (t,y,y1,y2,y3,z,u), ξL2(FT,R), ϕ(t) is a deterministic function.

    The following assumption will be inforce.

    (H1) f is continuously differentiable in (y,y1,y2,y3,z,u). Moreover, the partial derivatives fy,fy1,fy2,fy3,fz,fu of f are uniformly bounded.

    We give the corresponding functional

    J(u(t))=E{T0l(t,y(t),z(t),u(t))dt+g(y(0))}, (2.2)

    where l and g satisfy the following condition:

    (H2) The functions l and g are differentiable in (y,y1,y2,y3,z,u) and y, respectively, and all the derivatives are bounded.

    The following is a stochastic optimal control problem.

    Problem (BSO) Find an admissible control that maximizes J(u()) over uU.

    Any admissible control that solves Problem (BSO) is called an optimal control. Assume that ˉu()U is an admissible control. For convenience, we use the following notations throughout the paper. f(t,ˉu)=f(t,ˉy,z,ˉu), f(t,u)=f(t,y,z,u), l(t,ˉu)=f(t,y,z,ˉu), l(t,u)=l(t,y,z,u), and for k=y,y1,y2,y3,z,u, we use fk(t)=fk(ˉy,ˉy1,ˉy2,ˉy3,ˉz,ˉu) and for j=y,z,u, we use lj(t)=lj(ˉy,ˉz,ˉu).

    Remark 2.1. In contrast to [29], the present optimal control problem can also include moving-average delay and noisy memory. Thus, the model can cover more stochastic phenomena and stochastic optimization problems.

    The adjoint equation plays a key role in the theory of maximum principle. In this section, we introduce some adjoint equations. Then, we decompose the two-dimensional adjoint equation (3.1) into two one-dimensional systems of equations and give the equivalent adjoint equation (3.6), which is called a unified one since it connects the two kinds of adjoint equations (3.2), (3.3) and (3.6).

    A general reference to Malliavin derivative Dt for Lévy processes is the book (Di Nunno et al. [32]). See also Nualart [33] and Di Nunno et al. [34].

    For all FL2(FT,P), we have DtF(S) and (t,ω)E[DtF|Ft] belongs to L2(FT,λ×P), where (S)L2(P) denotes the Hida space of stochastic distributions and λ denotes the Lebesgue measure on [0,T]. Moreover, the following generalized Clark-Ocone theorem holds:

    F=E[F]+T0E[DtF|Ft]dW(t).

    As also noted in Dahl et al. [27], we can also apply the Clark-Ocone theorem to show that:

    Proposition 3.1. (The generalized duality formula). Let FL2(FT,P) and let φ(t)L2(λ×P) be adapted. Then

    E[FT0φ(t)dB(t)]=E[T0E[DtF|Ft]φ(t)dt].

    Proof. By Clark-Ocone theorem and the Itô isometry we have

    E[FT0φ(t)dB(t)]=E[(E[F]+T0E[DtF|Ft]dW(t))(T0φ(t)dW(t))]=E[(T0E[DtF|Ft]dW(t))(T0φ(t)dW(t))]=E[T0E[DtF|Ft]φ(t)dt].

    In this section, we will give two adjoint equations and prove that they are equivalent.

    The adjoint equations for Problem (BSO) is given by

    {dP(t)={A1(t)P(t)+Et[A2(s)P(s)]|s=t+δχ[0,Tδ](t)+Ett+δtDt(B1(s)P(s))ds+˜b1(t)}dt+(B2(t)P(t)+˜b2(t))dW(t),P(0)=ζ,P(t)=0,t(T,T+δ], (3.1)

    where

    A1(t)=(fy(t)1fy1(t)λ),A2(s)=(fy3(s)eλδ00),B1(s)=(fy2(s)000),
    B2(t)=(fz(t)000),˜b1(t)=(ly(t)0),˜b2(t)=(lz(t)q2(t)),ζ=(gy(ˉy(0))0).

    Here and in what follows, χ denotes the indicator function.

    In this section, we decompose the two-dimensional adjoint equation (3.1) into two one-dimensional systems of equations.

    Define the solution P() of the adjoint equation (3.1) by

    P(t)=(P1(t)P2(t)).

    Then by (3.1), it is easy to check that

    {dP1(t)={fy(t)P1(t)+Et[fy3(t+δ)P1(t+δ)]χ[0,Tδ](t)ly(t)+t+δtE[Dtfy2(s)P1(s)|Ft]ds+P2(t)eλδEt[P2(t+δ)]χ[0,Tδ](t)}dt+{fz(t)P1(t)lz(t))}dW(t)P1(0)=gy(ˉy(0)), (3.2)
    {dP2(t)={fy1(t)P1(t)+λP2(t)}dtq2(t)dW(t)P2(0)=0, (3.3)

    Remark 3.2. Note that, (3.2) and (3.3) is a system of coupled equations, where P1() depends on P2(). We will show that P1() satisfies an independent time-advanced BSDE involving Malliavin derivatives (see (3.6)).

    Note that, it is difficult to derive the adjoint equations P1() and P2() by focusing on the coupled equations (3.2) and (3.3). We will give the equivalent adjoint equation to show that P1() satisfies an independent time-advanced BSDE involving Malliavin derivatives (see (3.6)). We call it a unified one since it connects the two kinds of adjoint equations (3.2), (3.3) and (3.6).

    Lemma 3.3. Under (H1), it holds that

    P2(t)eλδEt[P2(t+δ)]χ[0,Tδ](t)=(t+δ)(Tδ)tEt[eλ(ts)(ly1(s)P1(s)fy1(s))]ds. (3.4)

    Proof. Given (3.3), applying Itô's formula to eλsP2(s),0st yields

    P2(t)=t0Et[eλ(ts)(ly1(s)P1(s)fy1(s))]ds. (3.5)

    Set Σ(t)=P2(t)eλδEt[P2(t+δ)]χ[0,Tδ](t). Then by (3.5), we have

    Σ(t)=t0Et[eλ(ts)(ly1(s)P1(s)fy1(s))]ds+t+δ0Et[eλ(ts)(ly1(s)P1(s)fy1(s))]dsχ[0,Tδ](t).

    (i) We first consider the case: t+δ<Tδ, Σ(t)=t+δtEt[eλ(ts)(ly1(s)P1(s)fy1(s))]ds.

    (ii) Otherwise, Σ(t)=TδtEt[eλ(ts)(ly1(s)P1(s)fy1(s))]ds.

    Therefore, we have Σ(t)=(t+δ)(Tδ)tEt[eλ(ts)(ly1(s)P1(s)fy1(s))]ds.

    Thus, the proof is complete.

    Substituting (3.4) into (3.2), we get the following equivalent adjoint equation

    {dP1(t)={fy(t)P1(t)+Et[fy3(t+δ)P1(t+δ)]χ[0,Tδ](t)ly(t)+t+δtE[Dtfy2(s)P1(s)|Ft]ds+(t+δ)(Tδ)tEt[eλ(ts)(ly1(s)P1(s)fy1(s))]ds}dt+{fz(t)P1(t)lz(t))}dW(t)P1(0)=gy(ˉy(0)). (3.6)

    Remark 3.4. Three types of delays of the state are incorporated in Problem (BSO). Let us consider some special cases.

    (a) If the moving-average delayed state y1() vanishes, then it holds that P2(t)=0, (3.6) degenerates to

    {dP1(t)={fy(t)P1(t)+Et[fy3(t+δ)P1(t+δ)]χ[0,Tδ](t)ly(t)+t+δtE[Dtfy2(s)P1(s)|Ft]ds}dt+{fz(t)P1(t)lz(t))}dW(t)P1(0)=gy(ˉy(0)).

    This is the case studied in [35] without mean-field term and on finite time horizon.

    (b) If the noisy memory y2() vanishes, then (3.6) degenerates to

    {dP1(t)={fy(t)P1(t)+Et[fy3(t+δ)P1(t+δ)]χ[0,Tδ](t)ly(t)+(t+δ)(Tδ)tEt[eλ(ts)(ly1(s)P1(s)fy1(s))]ds}dt+{fz(t)P1(t)lz(t))}dW(t)P1(0)=gy(ˉy(0)).

    This is the case studied in [36] without mean-field term and on finite time horizon. Note that, in [36], the discrete delay is not considered.

    (c) If the discrete state y3() vanishes, then (3.6) degenerates to

    {dP1(t)={fy(t)P1(t)ly(t)+t+δtE[Dtfy2(s)P1(s)|Ft]ds+(t+δ)(Tδ)tEt[eλ(ts)(ly1(s)P1(s)fy1(s))]ds}dt+{fz(t)P1(t)lz(t))}dW(t)P1(0)=gy(ˉy(0)).

    This case has not been studied before.

    (d) If both y1() and y2() vanish, then (3.6) degenerates to

    {dP1(t)={fy(t)P1(t)+Et[fy3(t+δ)P1(t+δ)]χ[0,Tδ](t)ly(t)}dt+{fz(t)P1(t)lz(t))}dW(t)P1(0)=gy(ˉy(0)).

    This is the case studied in [37] with full information.

    Define the Hamiltonian function H by

    H(t,y,y1,y2,y3,z,u,P1,P2)=l(t,y,z,u)f(t,y,y1,y2,y3,z,u)P1(yλy1eλδy3)P2.

    In this section, we form sufficient condition of optimality based on some concavity assumptions on H and g.

    The following assumption will also be needed.

    (H3) The functions H(t,y,y1,y2,y3,z,u,P1,P2) and g(y) are concave with respect to the corresponding variables, for t[0,T] and given P1,P2.

    For simplicity, we use (y(t),u(t)) and (ˉy(t),ˉu(t)) to denote the state trajectories of system (2.1) correspongding to u(t) and ˉu(t). We denote by ˉH(t)=H(t,ˉy,ˉy1,ˉy2,ˉy3,ˉz,ˉu,P1,P2), H(t)=H(t,y,y1,y2,y3,z,u,P1,P2) and the similar definitions for ˉf(t) and f(t). We give the sufficient maximum principle now.

    Theorem 4.1. Let (H1)(H3) hold. If an admissible control ˉu() satisfies the following condition:

    (ˉuu)Hu(t)0,

    then ˉu() is an optimal control.

    Proof. We write J(ˉu())J(u())=I1+I2 with

    I1=ET0{l(t,ˉy(t),ˉz(t),ˉu(t))l(t,y(t),z(t),u(t))}dt, (4.1)
    I2=E[g(ˉy(0))g(y(0))].

    Since g is concave on y, we have

    I2E[gy(ˉy(0))(ˉy(0)y(0))]. (4.2)

    Applying Itô's formula to P1(t)(ˉy(t)y(t)) and taking expectation, we get

    E(gy(ˉy(0))(ˉy(0)y(0))=ET0{(ˉy(t)y(t))(fy(t)P1(t)+Et[fy3(t+δ)P1(t+δ)]χ[0,Tδ](t)ly(t)+t+δtE[Dtfy2(s)P1(s)|Ft]ds+P2(t)eλδEt[P2(t+δ)]χ[0,Tδ](t))}dt+ET0{P1(t)((f(t,ˉy(t),ˉz(t),ˉu(t))f(t,y(t),z(t),u(t))))}dt+ET0{(ˉz(t)z(t))(fz(t)P1(t)lz(t))}dt. (4.3)

    By the definition H, it holds that

    I1=ET0{ˉH(t)H(t)+(ˉf(t)f(t))P1(t)(ˉy(t)y(t)λ(ˉy1(t)y1(t))eλδ(ˉy3(t)y3(t)))P2(t)}dt. (4.4)

    Noting that

    dy1(t)=[y(t)λy1(t)eλδy3(t)],t[0,T].

    Using Itô's formula to P2(t)(ˉy1(t)y1(t)) and integrating from 0 to T, and noticing P2(0)=0 and ˉy1(T)y1(T)=0, we get

    EP2(T)(ˉy1(T)y1(T))]E[P2(0)(ˉy1(0)y1(0))]=ET0{ly1(t)+P1(t)fy1(t)+λP2(t))(ˉy1(t)y1(t)}dt+ET0{P2(t)(ˉy(t)y(t)λ(ˉy1(t)y1(t))eλδ(ˉy3(t)y3(t)))}dt=0. (4.5)

    Inserting (4.5) into (4.4), we have

    I1=ET0{ˉH(t)H(t)+(ˉf(t)f(t))P1(t)+(ly1(t)+P1(t)fy1(t)+λP2(t))(ˉy1(t)y1(t))}dt. (4.6)

    Since H is concave, and from (4.2), (4.3) and (4.6), we can obtain

    J(ˉu)J(u)ET0{fy3(t)(ˉy3(t)y3(t))P1(t)+Et[fy3(t+δ)P1(t+δ)]χ[0,Tδ](t)(ˉy(t)y(t))}dt+ET0{eλδEt[P2(t+δ)]χ[0,Tδ](t)(ˉy(t)y(t))+eλδ(ˉy3(t)y3(t))P2(t)}dt+ET0{fy2(t)(ˉy2(t)y2(t))P1(t)+t+δtE[Dtfy2(s)P1(s)|Ft]ds(ˉy(t)y(t))+lu(t)(ˉu(t)u(t))fu(t)(ˉu(t)u(t))P1(t)}dt=J1+J2+J3+J5.

    Based on the definition of y3 and through transformation of variables, we have

    ET0fy3(t)(ˉy3(t)y3(t))P1(t)dt=ET0fy3(t)(ˉy(tδ)y(tδ))P1(t)dt=ETδδfy3(t+δ)(ˉy(t)y(t))P1(t+δ)dt.

    On the other hand,

    ET0Et[fy3(t+δ)P1(t+δ)]χ[0,Tδ](t)(ˉy(t)y(t))dt=ET+δδfy3(t)(ˉy(tδ)y(tδ))P1(t)dt.

    Therefore,

    J1=ET0fy3(t)(ˉy(tδ)y(tδ))P1(t)dt+ET+δδfy3(t)(ˉy(tδ)y(tδ))P1(t)dt=Eδ0fy3(t)(ˉy(tδ)y(tδ))P1(t)dtETδfy3(t)(ˉy(tδ)y(tδ))P1(t)dt+ET+δδfy3(t)(ˉy(tδ)y(tδ))P1(t)dt=Eδ0fy3(t)(ˉy(tδ)y(tδ))P1(t)dt+ET+δTfy3(t)(ˉy(tδ)y(tδ))P1(t)dt.

    By (2.1), we have ˉy(tδ)y(tδ)=0 when t[0,δ). So δ0fy3(t)(ˉy(tδ)y(tδ))P1(t)dt = 0. By (3.1), P1(t)=0,t(T,T+δ], we get ET+δTfy3(t)(ˉy(tδ)y(tδ))P1(t)dt. Then, we obtain J1=0.

    J3=ET0{fy2(t)(ˉy2(t)y2(t))P1(t)+t+δtE[Dtfy2(s)P1(s)|Ft]ds(ˉy(t)y(t))}dt.

    Using Fubini's theorem and the definition of y2, as well as the duality formula for Malliavin derivatives, we get

    ET0fy2(t)(ˉy2(t)y2(t))P1(t)dt=ET0fy2(t)(ttδˉy(s)y(s)dW(s))P1(t)dt=T0E[ttδE[Ds(fy2(t)P1(t))|Fs](ˉy(s)y(s))ds]dt=ET0s+δsE[Ds(fy2(t)P1(t))|Fs](ˉy(s)y(s))dtds.

    So, J3=0. In the similar ways, we can show that J2=0. This implies

    J(ˉu)J(u)T0E[Hu(t)(ˉu(t)u(t))]dt0.

    From this, it can be seen that ˉu(t) is an optimal control. We conclude.

    In this section, applying the results obtained in the previous section, we derive an explicit optimal control of the linear-quadratic optimal control with delayed BSDE.

    We give the following linear BSDE:

    {dy(t)={a(t)y(t)+b(t)y3(t)+c(t)y1(t)+d(t)y2(t)+e(t)u(t)}dtz(t)dW(t),y(T)=ξ,y(t)=φ(t),t[δ,0)

    and the performance functional

    J(u(t))=12E[T0a0(t)y2(t)+b0(t)z2(t)+r(t)u2(t)]dt12E[y2(0)],

    where a(t),b(t),c(t),d(t),e(t) are deterministic functions, with y1(t):=ttδeλ(st)y(s)ds, y2(t):=ttδy(s)dW(s) and y3(t):=y(tδ).

    We give the following assumption:

    (H4) a0(t) and b0(t) are non-negative bounded deterministic functions. Moreover, r(t),r1(t) are deterministic positive bounded functions.

    The Hamiltonian function is

    H(t,y,y1,y2,y3,z,u,P1,P2)=12(a0(t)y2(t)+b0(t)z2(t)+r(t)u2(t))(a(t)y(t)+b(t)y3(t)+c(t)y1(t)+d(t)y2(t)+e(t)u(t))P1(t)(y(t)λy1(t)eλδy3(t))P2(t)

    and the adjoint equation satisfies

    {dP(t)={A1(t)P(t)+Et[A2(s)P(s)]|s=t+δχ[0,Tδ](t)+Ett+δtDt(B1(s)P(s))dt+˜b1(t)}dt+(B2(t)P(t)+˜b2(t))dW(t),P(0)=ζ,

    where

    A1(t)=(a(t)1c(t)λ),A2(s)=(b(s)eλδ00),B1(s)=(d(s)000),
    B2(t)=(0000),˜b1(t)=(a0(t)0),˜b2(t)=(b0(t)q2(t)),ζ=(E[y(0)]0).

    Our objective is to maximize the performance functional over U.

    According to Theorem 4.1, we can derive the following result.

    Theorem 5.1. Under (H4), if the admissible control ˉu(t) satisfies

    ˉu(t)=r1(t)e(t)P1(t), (5.1)

    then it is an optimal control.

    Proof. Let ˉu(t) be a control given by (5.1), ˉy(t) and ˉz(t) be the corresponding state equation. By (5.1), (ˉy(t),ˉz(t),ˉu(t)) satisfies the optimality condition in Theorem 4.1, which means E[Hu(t)(ˉu(t)u(t))]0. From (H4), it is easy to check that all other conditions in Theorem 5.1 are satisfied, so ˉu is optimal.

    We consider a financial market with two investment possibilities. One is a risk-free asset with price S0(t) and the other is a risky asset with price S1(t). They satisfies:

    {dS0(t)=μ(t)S0(t)dt,S0(0)=0,

    where μ(t) is a bonded deterministic function.

    {dS1(t)=S1(t)(b(t)dt+σ(t)dW(t)),S1(0)=0, (5.2)

    where b(t),σ(t) are bounded deterministic functions.

    Suppose someone invests in bonds and stocks. They want to achieve their wealth goal ξ at time T. Moreover, the investor wants to maximize his utility function J.

    We denote by π(t) the proportion of fund that the investor invests in the stock. c(t) is the instantaneous consumption rate of investor. In real life, usually there are different levels of consumption. For simplicity, there is only one level. For example, c(t)U=[0,a1], that is we only consider the ultra-low consumption. Here a1 is a constant such that a1>(E[eγtP1(t)]L)1R. y(t) is the wealth process of the investor at time t and k(y(t)y(tδ)ttδy(s)dW(s))ttδeλ(st)y(s)ds is the surplus premium.

    We propose a wealth process with delay, which is caused by the instantaneous capital inflow or outflow from the investor's current wealth.

    Now the investor's wealth process is governed by the following BSDE:

    {dy(t)=[(μk)y(t)+ky(tδ)+kttδy(s)dW(s)+k1ttδeλ(st)y(s)dsc(t)+λz(t)]dtz(t)dW(t),y(T)=ξ,y(t)=0,t[δ,0). (5.3)

    We give the corresponding cost functional:

    J(c(t))=E{T0Leγt(c(t))1R1RdtMy(0)}, (5.4)

    where L>0,K>0,M>0,γ>0, and 0<R<1.

    Here is the Hamiltonian function

    H(t,y,y1,y2,y3,z,c,P1,P2)=Leγt(c(t))1R1R((μk)y(t)+ky3(t)+ky2(t)+k1y1(t)c(t)+λz(t))P1(t)(y(t)λy1(t)eλδy2(t))P2(t),

    where y1(t):=ttδeλ(st)y(s)ds, y2(t):=ttδy(s)dW(s) and y3(t):=y(tδ).

    The corresponding adjoint equation satisfies the following:

    {dP(t)={A1(t)P(t)+Et[A2(s)P(s)]|s=t+δχ[0,Tδ](t)+Ett+δtDt(B1(s)P(s))ds+˜b1(t)dt}+(B2(t)P(t)+˜b2(t))dW(t),P(0)=ζ,

    where

    A1(t)=(μk1k1λ),A2(s)=(keλδ00),B1(s)=(k000),
    B2(t)=(λ000),˜b1(t)=(00),˜b2(t)=(0q2(t)),ζ=(M0).

    Proposition 5.2. Suppose that (5.2), (5.3) and the adjoint process P(t) satisfy the conditions in Theorem 4.1. Then the optimal consumption c(t) is given by (E[eγtP1(t)]L)1R.

    Finally, let's look at a numerical example. Take δ1=0.1,δ2=0.15,δ3=0.2,δ4=0.23,δ5=0.26, λ=0.4,μ=0.6,k=0,k1=1,γ=0.3,M=1,R=0.5,T=3 and L=3.

    In order to obtain the effect of delay, we can divide the time horizon according to the delay, and then continue to subdivide each subinterval to get the numerical solution using the Euler method or improved Euler method. The corresponding parameter values are: δ1=0.1,δ2=0.15,δ3=0.2,δ4=0.23,δ5=0.26. Figure 1 plots the solution P1(t) for different time delays. We observe that the solution is very sensitive to δ, i.e., the smaller δ is, the larger P1(t) is.

    Figure 1.  The solution P1(t) with different time delays.

    Figure 2 shows the optimal consumption c(t) for different time delays. We find that the smaller δ is, the lower the consumption rate is. Furthermore, it can be seen from Figure 2 that no matter what value δ, the optimal consumption strategy is close to almost the same number. This is because, for convenience, we consider only one level of consumption: Ultra-low consumption.

    Figure 2.  The optimal consumption c(t) with different time-delays.

    For a more comprehensive analysis of the factors affecting the optimal control, we will show how the terminal condition y(T)=ξ affects the optimal control result. Here k=k1=0,μ=1. The state process satisfies the following BSDE

    {dy(t)=(y(t)c(t)+z(t))dtz(t)dW(t),y(T)=ξ.

    The cost function is given by

    J(c(t))=E{T0Leγt(c(t))1R1Rdty2(0)}.

    and the adjoint equation (3.2) is reduced to the following form

    {dP1(t)=P1(t)dt,P1(0)=2ˉy(0).

    By Proposition 5.2, the optimal consumption c(t) is given by (E[eγtP1(t)]L)1R.

    The following is the numerical result. Figure 3 plots the optimal consumption c(t) with different terminal conditions. The parameter values used in the calculations are: γ=0.3, R=0.5, T=3, L=3, ξ1=1, ξ2=3, ξ3=5. We observe that the smaller terminal condition will lead to a smaller consumption.

    Figure 3.  The optimal consumption c(t) with different terminal conditions.

    Motivated by some interesting financial and economic phenomena, we study the optimal control problem for BSDE with three delays. Some new-type adjoint equations are introduced for the first time in the literature. More especially, we obtain a sufficient condition for BSDE with three delays. The major features of this work include the introduction of a unified adjoint equation as well as the new type of the adjoint equations. The theoretical results established here are applied to solve a linear-quadratic optimal control problem and a financial problem. Then, a linear-quadratic optimal control and an optimal investment strategy are derived, respectively. There are several interesting issues that deserve further study.

    One could consider more complicated optimal control problems, such as problems with mean-field term and partial information rather than full information.

    On the one hand, the mean-field stochastic differential equation (SDE) was introduced to study physical systems with a large number of interacting particles. The mean-field models have many applications in economics and finance. Stochastic maximum principle for the optimal control system of mean-field type has become a popular topic. When considering mean-field terms, some new mean-field stochastic differential equations will be introduced as the adjoint equations and the maximum principle is generalized to the mean-field framework.

    On the other hand, in general, controllers can get only partial information in most cases. For example, we consider a cash flow process for an insurance firm, where due to the discreteness of account information, it is impossible for the firm to fully observe the cash flow. Unlike the system with full information, we will give a noisy process Y for partially observed optimal control problem. In addition, an extra adjoint equation will be introduced to obtain stochastic maximum principle, which does not appear in the full information case. The optimal control will contain the filtering of adjoint process which will increase the difficulty of analyzing the maximum condition and improve the skills of pure mathematical calculation. We plan to discuss these problems in the future.

    The author declares she has not used Artificial Intelligence (AI) tools in the creation of this article.

    Heping Ma is partially supported by National Natural Science Foundation of China (Grant No. 11801154). Hui Jian is partially supported by Jiangxi Provincial Natural Science Foundation (Grant No. 20212BAB211006) and National Natural Science Foundation of China (Grant No. 11761032). Thanks the anonymous reviewers and editors for their valuable comments and suggestions which led to the improvement of the original manuscript.

    The authors declare there is no conflict of interest.



    [1] E. Pardoux, S. G. Peng, Adapted solutions of backward differential equations, Syst. Control Lett., 14 (1990), 55–61. https://doi.org/10.1016/0304-4149(95)00024-2 doi: 10.1016/0304-4149(95)00024-2
    [2] N. El Karoui, S. G. Peng, M. Quenez, Backward stochastic differential equations in finance, Math. Finance, 7 (1997), 1–71. https://doi.org/10.1111/1467-9965.00022 doi: 10.1111/1467-9965.00022
    [3] D. Duffie, P. Geoffard, C. Skiadas, Efficient and equilibrium allocations with stochastic differential utility, J. Math. Econ., 23 (1994), 133–146. https://doi.org/10.1016/0304-4068(94)90002-7 doi: 10.1016/0304-4068(94)90002-7
    [4] G. Wang, H. Zhang, Mean-field backward stochastic differential equation with non-Lipschitz coefficient, Asian J. Control, 22 (2019), 1986–1994. https://doi.org/10.1002/asjc.2087 doi: 10.1002/asjc.2087
    [5] P. Huang, G. Wang, H. Zhang, An asymmetric information non-zero sum differential game of mean-field backward stochastic differential equation with applications, Adv. Differ. Equ., 236 (2019), 1–25. https://doi.org/10.1186/s13662-019-2166-5 doi: 10.1186/s13662-019-2166-5
    [6] J. Yong, X. Zhou, Stochastic Controls: Hamiltonian Systems and HJB Equations, Springer-Verlag, New York, NY, USA, 1999.
    [7] A. Bressan, B. Piccoli, Introduction to the Mathematical Theory of Control, Springfield: American Institute of Mathematical Sciences, 1 (2007).
    [8] N. Dokuchaev, X. Zhou, Stochastic controls with terminal contingent conditions, J. Math. Anal. Appl., 238 (1999), 143–165. https://doi.org/10.1006/JMAA.1999.6515 doi: 10.1006/JMAA.1999.6515
    [9] A. E. B. Lim, X. Zhou, Linear-quadratic control of backward stochastic differential equations, SIAM J. Control Optim., 40 (2001), 450–474. https://doi.org/10.1137/S0363012900374737 doi: 10.1137/S0363012900374737
    [10] J. Huang, G. Wang, J. Xiong, A maximum principle for partial information backward stochastic control problems with applications, SIAM J. Control Optim., 48 (2009), 2106–2117. https://doi.org/10.1137/080738465 doi: 10.1137/080738465
    [11] Z. Huang, Y. Wang, X. R. Wang, A mean-field optimal control for fully coupled forward-backward stochastic control systems with Lévy processes, J. Syst. Sci. Complex, 35 (2022), 205–220. https://doi.org/10.1007/s11424-021-0077-5 doi: 10.1007/s11424-021-0077-5
    [12] G. Wang, Z. Yu, A Pontryagin's maximum principle for non-zero sum differential games of BSDEs with applications, IEEE Trans. Autom. Control, 55 (2010), 1742–1747. https://doi.org/10.1109/TAC.2010.2048052 doi: 10.1109/TAC.2010.2048052
    [13] G. Wang, Z. Yu, A partial information non-zero sum differential game of backward stochastic differential equations with applications, Automatica, 48 (2012), 342–352. https://doi.org/10.1016/j.automatica.2011.11.010 doi: 10.1016/j.automatica.2011.11.010
    [14] G. Wang, H. Xiao, J. Xiong, A kind of linear quadratic non-zero differential game of backward stochastic differential equation with asymmetric information, Automatica, 63 (2018), 346–352. https://doi.org/10.1016/j.automatica.2018.08.019 doi: 10.1016/j.automatica.2018.08.019
    [15] S. Peng, Backward stochastic differential equations and applications to optimal control, Appl. Math. Optim., 27 (1993), 125–144. https://doi.org/10.1007/BF01195978 doi: 10.1007/BF01195978
    [16] W. Xu, Stochastic maximum principle for optimal control problem of forward and backward system, J. Aust. Math. Soc. Ser. B., 37 (1995), 172–185. https://doi.org/10.1017/S0334270000007645 doi: 10.1017/S0334270000007645
    [17] L. Chen, P. Zhou, H. Xiao, Backward Stackelberg games with delay and related forward-backward stochastic differential equaitons, Mathematics, 11 (2023), 2898. https://doi.org/10.3390/math11132898 doi: 10.3390/math11132898
    [18] Y. Zheng, J. Shi, A stackelberg game of backward stochastic differential equations with applications, J. Optim. Theory Appl., 10 (2019), 968–992. https://doi.org/10.1007/s13235-019-00341-z doi: 10.1007/s13235-019-00341-z
    [19] T. Wang, Backward stochastic volterra integro-differential equations and applications, SIAM J. Control Optim., 60 (2022), 2393–2419. https://doi.org/10.1137/20m1371464 doi: 10.1137/20m1371464
    [20] X. Li, J. Sun, J. Xiong, Linear quadratic optimal control problems for mean-field backward stochastic differential equations, Appl. Math. Optim., 80 (2019), 223–250. https://doi.org/10.1007/s00245-017-9464-7 doi: 10.1007/s00245-017-9464-7
    [21] L. Delong, P. Imleller, Backward stochastic differential equations with time delayed generators-results and counter examples, Ann. Appl. Probab., 20 (2010), 1512–1536. https://doi.org/10.1214/09-AAP663 doi: 10.1214/09-AAP663
    [22] S. Tomasiello, L. Rarità, An approximation technique and a possible application for a class of delay differential equations, in Proceedings of the 34th European Modeling & Simulation Symposium (EMSS 2022), (2022), 45. https://doi.org/10.46354/i3m.2022.emss.045
    [23] L. Rarità, C. D'Apice, B. Piccoli, D. Helbing, Sensitivity analysis of permeability parameters for flows on Barcelona networks, J. Differ. Equations, 249 (2010), 3110–3131. https://doi.org/10.1016/J.JDE.2010.09.006 doi: 10.1016/J.JDE.2010.09.006
    [24] L. Chen, J. H. Huang, Stochastic maximum principle for controlled backward delayed system via advanced stochastic differential equation, J. Optim. Theroy Appl., 167 (2015), 1112–1135. https://doi.org/10.1007/s10957-013-0386-5 doi: 10.1007/s10957-013-0386-5
    [25] N. Anakira, A. Alomari, I. Hashim, Optimal homotopy asymptotic method for solving delay differential equations, Math. Probl. Eng., 2013 (2013), 498902. https://doi.org/10.1155/2013/498902 doi: 10.1155/2013/498902
    [26] S. Wu, G. Wang, Optimal control problem of backward stochstic differential delay equation under partial information, Syst. Control Lett., 82 (2015), 71–78. https://doi.org/10.1016/j.sysconle.2015.05.008 doi: 10.1016/j.sysconle.2015.05.008
    [27] K. Dahl, S. E. A. Mohammed, B. Øksendal, E. Røse, Optimal control of systems with noisy memory and BSDEs with Malliavin derivatives, J. Funct. Anal., 271 (2016), 289–329. https://doi.org/10.1016/J.JFA.2016.04.031 doi: 10.1016/J.JFA.2016.04.031
    [28] Y. Kazmerchuk, A. Swishchuk, J. Wu, A continuous-time GARCH model for stochastic volatility with delay, Can. Appl. Math. Q., 13 (2005), 123–149.
    [29] A. Delavarkhalafi, A. S. Fatemion Aghda, M. Tahmasebi, Maximum principle for infinite horizon optimal control of mean-field backward stochastic systems with delay and noisy memory, Int. J. Control, 95 (2020), 535–543. https://doi.org/10.1080/00207179.2020.1800822 doi: 10.1080/00207179.2020.1800822
    [30] J. Ma, P. Protter, J. M. Yong, Solving forward-backward stochastic differential equations explicitly–-a four step scheme, Probab. Theory Relat. Fields, 98 (1994), 339–359. https://doi.org/10.1007/BF01192258 doi: 10.1007/BF01192258
    [31] F. Zhang, Sufficient maximum prinicple for stochastic optimal control problems with general delays, J. Optim. Theory Appl., 192 (2022), 678–701. https://doi.org/10.1007/s10957-021-01987-9 doi: 10.1007/s10957-021-01987-9
    [32] G. Di Nunno, B. Øksendal, F. Proske, Malliavin Calculus for Lévy Processes with Applications to Finance, Berlin: Springer, 2009.
    [33] D. Nualart, The Malliavin Calculus and Related Topics, Berlin: Springer, 2006.
    [34] G. Di Nunno, T. Meyer-Brandis, B. Øksendal, F. Proske, Malliavin calculus and anticipative Itô formula for Lévy processes, Infin. Dimension. Anal. Quantum Probab. Relat. Top., 08 (2005), 235–258. https://doi.org/10.1142/S0219025705001950 doi: 10.1142/S0219025705001950
    [35] A. Delavarkhalafi, A. S. Fatemion Aghda, M. Tahmasebi, Maximum principle for infinite horizon optimal control of mean-field backward stochastic systems with delay and noisy memory, Int. J. Control, 95 (2022), 535–543. https://doi.org/10.1080/00207179.2020.1800822 doi: 10.1080/00207179.2020.1800822
    [36] H. Ma, L. Bin, Infinite horizon optimal control problem of mean-field backward stochastic delay differential equation under partial information, Eur. J. Control, 36 (2017), 43–50. https://doi.org/10.1016/j.ejcon.2017.04.001 doi: 10.1016/j.ejcon.2017.04.001
    [37] S. Wu, Partially-observed maximum principle for backward stochastic differential delay equations, IEEE/CAA J. Autom. Sin., (2017), 1–6. https://doi.org/10.1109/JAS.2017.7510472
  • Reader Comments
  • © 2023 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Metrics

Article views(1486) PDF downloads(69) Cited by(0)

Figures and Tables

Figures(3)

Other Articles By Authors

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog