Abstract
We study optimal control for mean-field forward–backward stochastic differential equations with payoff functionals of mean-field type. Sufficient and necessary optimality conditions in terms of a stochastic maximum principle are derived. As an illustration, we solve an optimal portfolio with mean-field risk minimization problem.
1. Introduction
After the seminal work by Lasry and Lions [Citation1], where they introduced mean-field game theory that is devoted to the analysis of differential games with infinitely many players. Mean-field games attracted a lot of attention and forward/backward stochastic differential equations of mean-field type are used, extensively, as dynamics (see e.g., Huang et al. [Citation2], Xu and Zhang [Citation3] and Xu and Shi [Citation4]). In Huang [Citation5], the author studies a linear–quadratic game with a major player and a large number of minor players. The dynamics of the major player is influenced by an aggregation of all minor players (mean-field coupling) whereas the minor players’ dynamics depend on the control of the major player in addition to their individual controls as well as the mean-field coupling, i.e., a system of partially control-coupled forward stochastic differential equations (SDEs). This work (Ref. [Citation5]) was generalized to the non-linear case in Nourian and Caines [Citation6]. In all previously mentioned works, the authors find -Nash equilibrium for mean-field games, where each player play a game with the aggregation of the other players (the mass). In the present paper, the setting is different. We consider a mean-field type control problem where the goal is to find an optimal control via stochastic maximum principle. The mass or the laws of state processes are not freezed, they vary with the change of the control. Thus, finding an optimal control will yield optimal laws. Furthermore, in our control problem we consider a controlled partially coupled forward–backward SDE of mean-field type (MF-FBSDE) as dynamics, which is a novel contribution. We have also used the Sobolov space of random measures, introduced in Agram et al. [Citation7–9], in which, the Fréchet derivative with respect to the measure can be taken directly. This is a new approach compared to what is standard in the literature, where the Wasserstein metric space for measures and the lifting technique, introduced by Lions [Citation10], is used to differentiate a function of a measure.
Existence of a fully coupled MF-FBSDE is studied by Carmona and Delarue [Citation11] under Lipschitz assumption on the coefficients but no uniqueness result was proven. Bensoussan et al. [Citation12] prove existence and uniqueness of a fully coupled MF-FBSDE by assuming Lipschitz and monotonicity conditions. Recently, Djehiche and Hamadene [Citation13] prove the same results but under weak monotonicity assumptions and without the non-degeneracy condition on the forward equation.
The purpose of our work is to derive necessary and sufficient optimality conditions in terms of a stochastic maximum principle for a set of admissible controls which maximize a cost functional of the form with respect to admissible controls u, for some functions under dynamics governed by MF-FBSDEs. More specifically, we consider the coupled system for some functions and a Brownian motion M(t) and N(t) denote the marginal laws of X and Y, respectively. As an application, we will consider a risk minimization control problem. More precisely, we want to minimize the risk given by such that is the convex risk measure by means of backward stochastic differential equations of mean-field type (MF-BSDEs). Let us recall what we mean by the convex risk measure:
Definition 1.1.
A convex risk measure is a map that satisfies the following properties:
(Convexity) for all and all
(Monotonicity) If then
(Translation invariance) for all and all constants a.
The construction of risk measures from solutions of BSDEs is given as follows: Assume that in the driver of the above MF-BSDE and that is convex for all t. Then defines a convex risk measure. This shows how crucial is the choice of the functional g. Through this connection, the problem of risk minimization is equivalent to stochastic optimal control of MF-FBSDEs, as shown in Øksendal and Sulem [Citation14], for the non-mean-field case. The rest of the paper is organized as follows. In Section 2, we give some mathematical background. In Section 3, we study a stochastic optimal control of MF-FBSDE where sufficient and necessary optimality conditions are derived. In the last section, we construct a dynamic risk measure by means of MF-BSDE and then we solve an associated risk minimization problem.
2. Generalities
Let be a one-dimensional Brownian motion defined in a complete filtered probability space The filtration is assumed to be the P-augmented filtration generated by B.
Definition 2.1.
Let be the set of integers.
• Let be the space of random measures μ on equipped with the norm (2.1) (2.1) where is the Fourier transform of the measure μ, i.e., We endow with the inner product and are the Fourier transform of the measures μ and η, respectively. Then is a pre-Hilbert space, for each k. Let be the union (inductive limit) of
• We denote by the set of all deterministic elements of
We give some examples:
Example 2.2
(Measures). Let us give some examples of measures in and :
Suppose that , the unit point mass at . Then and
and hence
Suppose , where . Then and by Riemann–Lebesque lemma, , i.e., is continuous and when . In particular, is bounded on and hence
Suppose that μ is any finite positive measure on . Then and
and hence
Next, suppose is random. Then is a random measure in . Similarly, if is random, then is a random measure in
We denote by U a nonempty convex subset of and we denote by the set of U-valued -progressively measurable processes where with for all we consider them as the admissible control processes.
We will also use the following spaces:
is the set of -valued -adapted càdlàg processes such that
is the set of -valued -adapted processes such that
denotes the set of absolutely continuous functions
is the set of bounded linear functionals equipped with the operator norm
is the set of -adapted stochastic processes such that
is the set of -adapted stochastic processes such that
We recall now the notion of differentiability which will be used in the sequel.
Let be two Banach spaces with norms respectively, and let
We say that F has a directional derivative (or Gateaux derivative) at in the direction if
We say that F is Fréchet differentiable at if there exists a continuous linear map such that where is the action of the liner operator A on h. In this case we call A the gradient (or Fréchet derivative) of F at v and we write
If F is Fréchet differentiable at v with Fréchet derivative then F has a directional derivative in all directions and
In particular, note that if F is a linear operator, then for all v.
3. Optimal control problem
Here we denote by the law of X(t) at time t and by the law of Y(t) at time t. We assume that our system is governed by a coupled system of MF-FBSDE as follows:
The MF-SDE is given by (3.1) (3.1) for functions which are supposed to be -measurable and the initial value
The couple MF-BSDE satisfies (3.2) (3.2) where is -adapted and is -measurable.
It follows from the definition of the norm (Equation2.1(2.1) (2.1) ) that where and are random variables that follow the distributions and respectively.
Assume that (C is a constant that may change from line to line)
(A1) there exists C > 0, such that
for all for all fixed
for all for all fixed
(A2) there exists C > 0, such that, for all fixed and all knowing of Equation (Equation3.1(3.1) (3.1) ) and we have
for all
for all
Proposition 3.1.
Under Assumptions (ACitation1) and (ACitation2), the MF-FBSDE (3.Citation1)–(3.Citation2) admits a unique solution
Since the system is partially coupled i.e., the forward equation does not depend on the solution of the backward one, we can solve the system separately as follows: we first find a solution X(t) of the MF-SDE (Equation3.1(3.1) (3.1) ) and then we plug it into the backward EquationEquation (3.2)(3.2) (3.2) , then we solve it.
Our aim is to maximize the performance functional of the form over all admissible controls, for functions and
Now, we can define the Hamiltonian by (3.3) (3.3)
Remark 3.2.
For ease of notation we drop the dependence of all variables except for the time we write Moreover, we will use
We assume that
(A3) are continuously differentiable with bounded partial derivatives w.r.t all the variables.
For with corresponding solution define, whenever solutions exist, and and by the adjoint equations:
The BSDE for the unknown processes (3.4) (3.4)
The MF-BSDE for the unknown processes (3.5) (3.5)
The forward SDE (3.6) (3.6) and (3.7) (3.7)
Remark 3.3.
The real-valued linear system of FBSDE (Equation3.4(3.4) (3.4) ) and (Equation3.6(3.6) (3.6) ) have a unique solution by Proposition 3.1 since the coefficients satisfy condition (ACitation3). However, EquationEquation (3.5)(3.5) (3.5) is equivalent to the degenerate BSDE
We take conditional expectation to obtain
Similarly, a solution for (Equation3.7(3.7) (3.7) ) is given by
Before stating and proving sufficient and necessary conditions of optimality, we need the following result, which is Lemma 2.3 in Agram and Øksendal [Citation7].
Lemma 3.4.
Suppose that X(t) is an Itô process of the form where are adapted processes.
Then the map is absolutely continuous.
It follows that is differentiable for t-a.e. We will in the following use the notation
In fact, it is proven in [Citation7] that if then
3.1. Sufficient optimality conditions
We state and prove a type of a verification theorem.
Theorem 3.5.
Suppose that with corresponding solutions to Equations (3.1), (Equation3.2(3.2) (3.2) ), (Equation3.4(3.4) (3.4) ), (Equation3.5(3.5) (3.5) ), (Equation3.6(3.6) (3.6) ) and (Equation3.7(3.7) (3.7) ), respectively. Suppose that are concave functions P-a.s for each Moreover,
P-a.s for all t Then is an optimal control.
Proof.
We show that for an arbitrary u and a fixed optimal
We introduce first the following notation and and
From the definition of the Hamiltonian (Equation3.3(3.3) (3.3) ), we have and (3.8) (3.8)
We use the concavity of h and as well as the boundary values of EquationEquations (3.4)(3.4) (3.4) , (3.5), (3.Citation6) and (Equation3.7(3.7) (3.7) ) (3.9) (3.9)
Applying It formula to and yields the following duality relations: (3.10) (3.10) (3.11) (3.11) (3.12) (3.12)
Concavity of ψ gives (3.13) (3.13)
By the concavity of H, we obtain (3.14) (3.14)
Finally, by substituting the derived duality relations (Equation3.10(3.10) (3.10) ), (Equation3.11(3.11) (3.11) ), (Equation3.12(3.12) (3.12) ) and (Equation3.13(3.13) (3.13) ) in (Equation3.8(3.8) (3.8) ) and using the estimates (Equation3.9(3.9) (3.9) ), (Equation3.14(3.14) (3.14) ), we obtain
Using the tower property and the fact that u(t) is -adapted the desired result follows and thus, is optimal.□
3.2. Necessary optimality conditions
Given an arbitrary but fixed control we define (3.15) (3.15)
Note that, the convexity of U and guarantees that We denote by and by the solution processes corresponding to and respectively.
For each and all bounded -measurable random variables the process belongs to
In general, if is a process depending on we define the operator D on K by (3.16) (3.16) whenever the derivative exists.
Define the following derivative processes such that (3.17) (3.17) and (3.18) (3.18)
Remark 3.6.
Equations (Equation3.17(3.17) (3.17) ), (Equation3.18(3.18) (3.18) ) are linear FBSDE with bounded coefficients, then by Proposition 3.1 they have a unique solution.
Theorem 3.7.
Let be the optimal control and be the corresponding solutions to the EquationEquations (Equation3.17(3.17) (3.17) )(3.17) (3.17) , (Equation3.18(3.18) (3.18) ), (Equation3.4(3.4) (3.4) ), (Equation3.5(3.5) (3.5) ), (Equation3.6(3.6) (3.6) ), (Equation3.7(3.7) (3.7) ), respectively. Then, the following statements are equivalent
for all bounded
for all
Proof
We first prove Theorem 3.7 by assuming (i) and aiming to show (ii)
{we substitute f(t) from Equation } by using the chain rule, we obtain and
We apply It formula to and then we take the expectation, we obtain the following important duality relations:
By substituting the derived duality relations and the partial derivatives of f(t) the desired result follows. This proof can be reversed to prove We omit the details.□
4. Mean-field risk minimization
4.1. Mean-field dynamic risk measure
In this section, we are interested in a particular class of MF-BSDE of the following form (4.1) (4.1) where
We assume that the generator is -adapted, uniformly Lipschitz and concave, and the terminal condition
Definition 4.1.
Define by where is a component of the solution of the MF-BSDE (Equation4.1(4.1) (4.1) ) with terminal horizon T, terminal condition ξ and driver f. Then is a dynamic risk measure induced by the MF-BSDE (Equation4.1(4.1) (4.1) ).
We may remark that the driver f depends linearly on Y and its expected value and nonlinearly on Z. This is interpreted as a market with interest rates We can reformulated this as a problem with a driver independent of Y and by discounting the financial position ξ. We assume that the instantaneous interest rates r(t) and are deterministic. We denote by the corresponding discounted risk measure.
Define the discounted process
Then Yr with driver and the terminal value is a part of the solution of the associated BSDE. We obtain also a discounted risk measure accordingly
This discounted risk measure is translation-invariant because Fr does not depend on Y, we have for and
Similarly, we can get for each that is translation-invariant.
4.2. Optimal portfolio with mean-field risk minimization
Consider a financial market with two investment possibilities:
Safe, or risk-free asset with unit price
Risky asset with unit price
Let be a self-financing portfolio invested in the risky asset at time t. We want to minimize the risk of the terminal value of the wealth process corresponding to a portfolio π which satisfies the linear SDE (4.2) (4.2) such that where satisfies a MF-BSDE (4.3) (4.3)
Here we assume that are given deterministic functions and is some given concave function. We want to find such that
Define the Hamiltonian H that corresponds to our problem by
The couple solution of the following BSDE and satisfies
The equation for is given by the forward SDE (4.4) (4.4) and satisfies
The first order necessary optimality condition gives
where we denoted by and so on. Since for all t P-a.s., we obtain (4.5) (4.5) which implies this together with EquationEquation (Equation4.4(4.4) (4.4) )(4.4) (4.4) , yields
From (Equation4.5(4.5) (4.5) ), we get
For example, if we choose (4.6) (4.6)
That is
Substituting the expression of above into the MF-BSDE (Equation4.3(4.3) (4.3) ), we obtain (4.7) (4.7)
Consequently thus (4.8) (4.8)
Define to be the solution of the linear SDE
or explicitly (4.9) (4.9)
By the Girsanov theorem of change of measures, we know that there exists an equivalent local martingale measure such that with is the Radon–Nikodym derivative of Q with respect to P on
Substituting (Equation4.8(4.8) (4.8) ), (Equation4.9(4.9) (4.9) ) into (Equation4.7(4.7) (4.7) ) we have
Taking the expectation but now with respect to the new measure Q, we get (4.10) (4.10) where is the entropy of Q with respect to P.
Since we obtained the optimal value of we can get the corresponding optimal terminal wealth
Summarizing, we have the following conclusion:
Theorem 4.2.
Suppose that (Equation4.6(4.6) (4.6) ) holds. Then the minimal risk of our problem is given by (Equation4.10(4.10) (4.10) ).
Additional information
Funding
References
- Lasry, J.-M., Lions, P.-L. (2007). Mean field games. Jpn. J. Math. 2(1):229–260. DOI: 10.1007/s11537-007-0657-8.
- Huang, J., Wang, S., Wu, Z. (2016). Backward mean-field linear–quadratic–Gaussian (LQG) games: Full and partial information. IEEE Trans. Automat. Contr. 61(12):3784–3796. DOI: 10.1109/TAC.2016.2519501.
- Xu, R., Zhang, F. (2020). ε-Nash mean-field games for general linear–quadratic systems with applications. Automatica. 114:108835. DOI: 10.1016/j.automatica.2020.108835.
- Xu, R., Shi, J. (2019). ε-Nash mean-field games for linear–quadratic systems with random jumps and applications. Int. J. Control. DOI: 10.1080/00207179.2019.1651940.
- Huang, M. (2010). Large-population LQG games involving a major player: the Nash certainty equivalence principle. SIAM J. Control Optim. 48(5):3318–3353. DOI: 10.1137/080735370.
- Nourian, M., Caines, P. E. (2013). ε-Nash mean field game theory for nonlinear stochastic dynamical systems with major and minor agents. SIAM J. Control Optim. 51(4):3302–3331. DOI: 10.1137/120889496.
- Agram, N., Øksendal, B. (2019). Model uncertainty stochastic mean-field control. Stoch. Anal. Appl. 37(1):36–21. DOI: 10.1080/07362994.2018.1499036.
- Agram, N., Øksendal, B. (2019). Stochastic control of memory mean-field processes. Appl. Math. Optim. 79(1):181–204. DOI: 10.1007/s00245-017-9425-1.
- Agram, N., Bachouch, A., Øksendal, B., Proske, F. (2019). Singular control optimal stopping of memory mean-field processes. SIAM J. Math. Anal. 51(1):450–468. DOI: 10.1137/18M1174787.
- Lions, P. Cours au college de france: Th´eorie des jeux ´a champs moyens (2014).
- Carmona, R., Delarue, F. (2015). Forward–backward stochastic differential equations and controlled McKean–Vlasov dynamics. Ann. Probab. 43(5):2647–2700. DOI: 10.1214/14-AOP946.
- Bensoussan, A., Yam, S. C. P., Zhang, Z. (2015). Well-posedness of mean-field type forward–backward stochastic differential equations. Stoch. Process. Appl. 125(9):3327–3354. DOI: 10.1016/j.spa.2015.04.006.
- Djehiche, B., Hamadene, S. (2019). Mean-field backward–forward stochastic differential equations and nonzero sum stochastic differential games. arXiv preprint arXiv:1904.06193
- Øksendal, B., Sulem, A. (2015). Risk minimization in financial markets modeled by Ito⁁-L´evy processes. Afr. Math. 26(5–6):939–979. DOI: 10.1007/s13370-014-0248-9.