MF-FBSDE: Theory, Methods, and Applications

• Mean-field FBSDEs are coupled stochastic systems where forward and backward dynamics interact through state distributions, modeling phenomena in control theory, economics, and physics.
• They employ decoupling fields and Riccati representations to manage complex couplings, ensuring well-posedness under Lipschitz and monotonicity conditions.
• Numerical methods for MF-FBSDEs range from particle approximations and cubature techniques to deep learning algorithms for high-dimensional problems.

Mean-field forward-backward stochastic differential equations (MF-FBSDEs) are coupled stochastic systems in which a forward SDE and a backward SDE interact through the state, adjoint variables, and the law of the solution, typically through marginal or joint distributions in Wasserstein spaces or through expectation-type aggregates. In the literature represented here, they appear as probabilistic objects for mean field games, McKean-Vlasov optimal control, stochastic differential games, linear-quadratic regulation, social optimization in large populations, and related problems in economics, finance, physics, and chemistry (Carmona et al., 2012, 0711.2162). The subject combines well-posedness theory, decoupling-field and Riccati methods, stochastic maximum principles, particle approximations, and increasingly high-dimensional numerical schemes (Tian et al., 2022, Han et al., 2022).

1. Canonical forms and modeling content

A standard fully coupled formulation is the McKean-Vlasov system

$dX_t=b\bigl(t,X_t,Y_t,Z_t,\law(X_t,Y_t,Z_t)\bigr)\,dt +\sigma\bigl(t,X_t,Y_t,Z_t,\law(X_t,Y_t,Z_t)\bigr)\,dW_t,\qquad X_0=x_0,$

$dY_t=-f\bigl(t,X_t,Y_t,Z_t,\law(X_t,Y_t,Z_t)\bigr)\,dt+Z_t\,dW_t,\qquad Y_T=g\bigl(X_T,\law(X_T)\bigr),$

with the dependence on the measure argument measured by the $2$-Wasserstein distance $W_2$ (Carmona et al., 2012). In this formulation, the forward state $X$ and the backward pair $(Y,Z)$ are jointly endogenous, and the mean-field term may involve the full law of $(X_t,Y_t,Z_t)$ rather than only a finite set of moments.

Other papers adopt structurally equivalent but more specialized forms. In expectation-coupled systems, the coefficients depend on aggregates such as $m(t)=\mathbb E[X_i(t)]$ and $m_u(t)=\mathbb E[u_i(t)]$, as in the linear social optimization model

$$dX_i(t)=\bigl(A\,X_i(t)+\bar A\,m(t)+B\,u_i(t)\bigr)\,dt +\bigl(C\,X_i(t)+\bar C\,m(t)+D\,u_i(t)\bigr)\,dW_i(t),$$

$dY_t=-f\bigl(t,X_t,Y_t,Z_t,\law(X_t,Y_t,Z_t)\bigr)\,dt+Z_t\,dW_t,\qquad Y_T=g\bigl(X_T,\law(X_T)\bigr),$0

with terminal condition $dY_t=-f\bigl(t,X_t,Y_t,Z_t,\law(X_t,Y_t,Z_t)\bigr)\,dt+Z_t\,dW_t,\qquad Y_T=g\bigl(X_T,\law(X_T)\bigr),$1 (Wang et al., 2024). In Agram and Choutri’s control framework, the forward law $dY_t=-f\bigl(t,X_t,Y_t,Z_t,\law(X_t,Y_t,Z_t)\bigr)\,dt+Z_t\,dW_t,\qquad Y_T=g\bigl(X_T,\law(X_T)\bigr),$2 and backward law $dY_t=-f\bigl(t,X_t,Y_t,Z_t,\law(X_t,Y_t,Z_t)\bigr)\,dt+Z_t\,dW_t,\qquad Y_T=g\bigl(X_T,\law(X_T)\bigr),$3 enter the coefficients through expectation functionals such as $dY_t=-f\bigl(t,X_t,Y_t,Z_t,\law(X_t,Y_t,Z_t)\bigr)\,dt+Z_t\,dW_t,\qquad Y_T=g\bigl(X_T,\law(X_T)\bigr),$4 and $dY_t=-f\bigl(t,X_t,Y_t,Z_t,\law(X_t,Y_t,Z_t)\bigr)\,dt+Z_t\,dW_t,\qquad Y_T=g\bigl(X_T,\law(X_T)\bigr),$5 (Agram et al., 2019).

The modeling scope is broader than the Markovian setting. A path-dependent formulation allows the forward and backward coefficients to depend on the law of the paths of $dY_t=-f\bigl(t,X_t,Y_t,Z_t,\law(X_t,Y_t,Z_t)\bigr)\,dt+Z_t\,dW_t,\qquad Y_T=g\bigl(X_T,\law(X_T)\bigr),$6 up to time $dY_t=-f\bigl(t,X_t,Y_t,Z_t,\law(X_t,Y_t,Z_t)\bigr)\,dt+Z_t\,dW_t,\qquad Y_T=g\bigl(X_T,\law(X_T)\bigr),$7, written as $dY_t=-f\bigl(t,X_t,Y_t,Z_t,\law(X_t,Y_t,Z_t)\bigr)\,dt+Z_t\,dW_t,\qquad Y_T=g\bigl(X_T,\law(X_T)\bigr),$8; in that regime the state variable is effectively measure-valued on a path space $dY_t=-f\bigl(t,X_t,Y_t,Z_t,\law(X_t,Y_t,Z_t)\bigr)\,dt+Z_t\,dW_t,\qquad Y_T=g\bigl(X_T,\law(X_T)\bigr),$9 (Buckdahn et al., 2023). Jumps can also be incorporated, either in the state equation or in the backward component, by adding Poisson random measures and compensated jump martingales (Tang et al., 2016, Li et al., 2018).

A recurring technical point is that notation is not uniform across the literature. In some optimality systems the forward component is an adjoint state and the original controlled state is backward, so the label “forward-backward” describes the temporal orientation of the equations rather than a fixed semantic role for $2$0 (Li et al., 2016, Chen et al., 2019).

2. Solvability theory and conditions for existence and uniqueness

A foundational existence result is due to Carmona and Delarue, who prove that under Lipschitz continuity of $2$1 in the state and measure arguments, linear growth of $2$2 and $2$3, boundedness of $2$4 and $2$5, and uniform ellipticity of $2$6, there exists an adapted solution

$2$7

to the fully coupled MF-FBSDE (Carmona et al., 2012). Their argument freezes a candidate flow of marginals, solves the resulting standard coupled FBSDE, constructs a decoupling field, and applies Schauder’s fixed-point theorem on a compact convex subset of measure flows and bounded decoupling functions. The same source emphasizes that uniqueness is not automatic; in general it may fail unless additional monotonicity or small-coupling assumptions are imposed (Carmona et al., 2012).

Several later works replace ellipticity-based fixed-point arguments by monotonicity structures tailored to the coupling. For fully coupled mean-field FBSDEs with jumps, Li and Min impose a weak-monotonicity condition on the coefficients and the terminal map, together with uniform Lipschitz and growth bounds, and obtain existence and uniqueness in

$2$8

They also prove continuity of the solution with respect to parameters (Li et al., 2018). Chen, Djehiche, and Hamadène study mean-field backward-forward systems under weak monotonicity assumptions and no non-degeneracy condition on the forward equation; their method is an implicit penalized approximation scheme that yields a contraction in $2$9 when the Lipschitz constants are sufficiently small relative to the monotonicity constants (Chen et al., 2019).

Tian and Yu introduce domination-monotonicity conditions for a class of mean-field type FBSDEs in which the coupling appears not only in running terms but also in initial and terminal constraints,

$W_2$0

Under their Lipschitz and domination-monotonicity hypotheses they obtain unique square-integrable solvability together with an a priori estimate and continuous dependence on coefficients (Tian et al., 2022). This framework is specifically designed to cover Hamiltonian systems arising from mean-field linear-quadratic problems with controlled initial or terminal data.

Strong solvability beyond the classical Lipschitz regime is addressed by Nam and Xu. For multi-population mean-field FBSDEs with measurable coefficients, possible discontinuity in the forward state, and non-Lipschitz continuity with respect to the time-sectional distribution, they prove existence of strong solutions via a fixed point on the flow of laws combined with Girsanov decoupling and classical FBSDE estimates (Nam et al., 2024). For path-dependent mean-field coupled systems, existence is obtained under continuity of the coefficients with respect to the path-law in $W_2$1, while uniqueness requires Lipschitz conditions and a non-anticipativity assumption on the forward-law dependence (Buckdahn et al., 2023).

Taken together, these results show that the central analytical difficulty is not merely the forward-backward coupling, but the simultaneous interaction of temporal coupling, distributional feedback, and terminal or initial constraints. A plausible implication is that no single existence theory dominates the subject; rather, different structural assumptions support different classes of MF-FBSDEs.

3. Decoupling fields, master fields, and Riccati representations

The most basic decoupling mechanism is the Markovian ansatz $W_2$2, which converts the joint law of $W_2$3 into the law of $W_2$4 together with a deterministic decoupling function. In the Carmona-Delarue framework, this ansatz is part of the fixed-point construction itself: one freezes a flow of marginals, solves a standard FBSDE, obtains the associated decoupling function $W_2$5, and searches for a fixed point in $W_2$6 (Carmona et al., 2012).

In weak formulations of stochastic differential mean-field games, the decoupling object becomes a master field. Under Lipschitz, nondegeneracy, and Lasry-Lions monotonicity conditions, and with additional smoothness and Lions differentiability assumptions, it is shown that there exists a deterministic function

$W_2$7

such that

$W_2$8

and $W_2$9 satisfies a master PDE involving both $X$0 and the Lions derivative $X$1 (Morgado et al., 2023). The same paper uses Malliavin calculus to derive classical differentiability in the initial datum, Malliavin differentiability of $X$2, and the identity $X$3.

In linear-quadratic settings, the decoupling is often finite-dimensional and Riccati-based. Li, Sun, and Xiong derive the representation

$X$4

which decouples the optimality system for a mean-field backward stochastic control problem into two deterministic Riccati ODEs and one residual linear MF-BSDE (Li et al., 2016). Tang and Meng obtain an analogous decomposition for a mean-field backward SDE with jumps, again through two coupled Riccati equations plus a residual MF-BSDE with jumps (Tang et al., 2016).

The linear social-optimization problem studied in “Social Optima of Linear Forward-Backward Stochastic System” uses the ansatz

$X$5

where $X$6 and $X$7 solve coupled Riccati ODEs with terminal values $X$8 and $X$9 (Wang et al., 2024). This yields a decentralized linear feedback law

$(Y,Z)$0

with deterministic gain matrices determined by $(Y,Z)$1 (Wang et al., 2024).

The Riccati program is not universal. In the deterministic-coefficient LQ case it produces explicit feedback laws and often complete characterization of the optimal control, but the literature explicitly notes that for nonlinear mean-field FBSDEs or random-coefficient LQ problems the neat ansatz may fail and one is led instead to backward stochastic Riccati equations, continuation methods, monotonicity arguments, or infinite-dimensional master equations (Li et al., 2016).

4. Optimal control, social optima, and mean-field games

MF-FBSDEs frequently appear as Hamiltonian systems associated with optimal control. In Agram and Choutri’s setting, the controlled MF-FBSDE consists of a forward mean-field SDE for $(Y,Z)$2 and a backward mean-field BSDE for $(Y,Z)$3, together with a payoff

$(Y,Z)$4

They derive sufficient and necessary optimality conditions in terms of a stochastic maximum principle with adjoint processes for both state variables and law derivatives, and the first-order condition takes the form

$(Y,Z)$5

under the control filtration $(Y,Z)$6 (Agram et al., 2019). The same framework is used to solve an optimal portfolio problem with mean-field risk minimization.

The path-dependent extension of Pontryagin theory replaces the classical Hamiltonian by a vector-valued object that incorporates the forward equation, the backward equation, and the terminal cost simultaneously. The law dependence on the paths of $(Y,Z)$7 yields adjoint equations with delay and anticipation features, and the corresponding stationarity condition is necessary and, under convexity, sufficient for optimality (Buckdahn et al., 2023). For jump-diffusion systems, Li and Min derive a stochastic maximum principle in which the Hamiltonian includes the Brownian and jump coefficients, and a first-order inequality in the control variable characterizes optimality; convexity again yields sufficiency (Li et al., 2018).

A distinct but closely related theme is the social planner’s problem. In the large-population LQ system of Xie, Huang, and Wang, agents cooperate to minimize the social cost

$(Y,Z)$8

which the paper explicitly distinguishes from a mean field game (Wang et al., 2024). Using forward-backward person-by-person optimality, an MF-type consistency condition, and Riccati decoupling, the authors obtain a decentralized strategy and prove asymptotic social optimality:

$(Y,Z)$9

Thus the decentralized law is $(X_t,Y_t,Z_t)$0-optimal with $(X_t,Y_t,Z_t)$1 as $(X_t,Y_t,Z_t)$2 (Wang et al., 2024).

MF-FBSDEs also encode equilibrium conditions in games. Chen, Djehiche, and Hamadène derive explicit open-loop Nash equilibrium strategies for nonzero-sum mean-field linear-quadratic stochastic differential games with random coefficients, where the equilibrium controls are expressed through backward adjoint processes solving a fully coupled mean-field backward-forward system (Chen et al., 2019). Nam and Xu use a Pontryagin stochastic maximum principle and a martingale approach to connect strong solutions of MF-FBSDEs with Nash equilibria in multi-population mean-field games, where each population shares an objective function and population sizes are part of the modeling data (Nam et al., 2024).

A common misconception is to identify all such systems with competitive mean field games. The cooperative social-optimization setting, the single-controller mean-field control problem, and the nonzero-sum game setting all lead to MF-FBSDEs, but the optimization criteria and equilibrium concepts are different.

5. Particle approximations and numerical methods

A basic approximation principle replaces law dependence by empirical measures of interacting or decoupled particle systems. Buckdahn, Djehiche, Li, and Peng study an MF-BSDE driven by a McKean-Vlasov forward SDE and approximate it by a decoupled $(X_t,Y_t,Z_t)$3-particle forward-backward system governed by $(X_t,Y_t,Z_t)$4 independent copies. They prove the estimate

$(X_t,Y_t,Z_t)$5

that is, mean-square order $(X_t,Y_t,Z_t)$6 and equivalently $(X_t,Y_t,Z_t)$7-order $(X_t,Y_t,Z_t)$8 (0711.2162). They further show that the fluctuation process

$(X_t,Y_t,Z_t)$9

converges in law to the solution of a linear mean-field FBSDE driven by both the original Brownian motion and an independent zero-mean Gaussian field (0711.2162).

For decoupled McKean-Vlasov FBSDEs, Crisan and McMurray propose a deterministic cubature-on-Wiener-space algorithm. Their forward approximation builds a cubature tree, while the backward pass uses first-order or second-order schemes. Under smooth-boundary assumptions, the first-order backward scheme has error $m(t)=\mathbb E[X_i(t)]$0 and the second-order scheme has error $m(t)=\mathbb E[X_i(t)]$1; analogous final-time rates are recovered under a Lipschitz terminal condition and uniform ellipticity with a suitable nonuniform grid (Raynal et al., 2013).

High-dimensional problems with full distribution dependence motivate learning-based methods. Han, Hu, and Long develop a fictitious-play deep learning algorithm for McKean-Vlasov FBSDEs in which the mean-field interaction may depend on the full law rather than only on expectations or moments. Each iteration freezes law-dependent coefficients, learns deterministic surrogates for the law terms, and solves the resulting standard FBSDE by Deep BSDE or Deep Backward Dynamic Programming (Han et al., 2022). Their convergence analysis uses integral probability metrics $m(t)=\mathbb E[X_i(t)]$2, for which

$m(t)=\mathbb E[X_i(t)]$3

independently of the ambient dimension, leading to a theorem that the total error can be made arbitrarily small without dependence on the ambient dimension (Han et al., 2022). Numerically, the paper reports absolute error in $m(t)=\mathbb E[X_i(t)]$4 below $m(t)=\mathbb E[X_i(t)]$5 for a benchmark example in dimensions $m(t)=\mathbb E[X_i(t)]$6, and final errors still below $m(t)=\mathbb E[X_i(t)]$7 when increasing $m(t)=\mathbb E[X_i(t)]$8 from $m(t)=\mathbb E[X_i(t)]$9 to $m_u(t)=\mathbb E[u_i(t)]$0 with corresponding scaling of sample size and network width; it also reports $m_u(t)=\mathbb E[u_i(t)]$1-error of order $m_u(t)=\mathbb E[u_i(t)]$2 in a mean-field Cucker-Smale flocking game benchmark (Han et al., 2022).

These developments show that numerical treatment of MF-FBSDEs has split into three broad paradigms already present in the literature: particle approximations with propagation-of-chaos control, deterministic weak approximation via cubature, and high-dimensional learning methods based on repeated standard FBSDE solves.

6. Extensions, variants, and current scope

Several extensions clarify the breadth of the field. Mean-field backward stochastic control with jumps leads to stochastic Hamiltonian systems that are fully coupled MF-FBSDEs with jumps, decoupled by two Riccati equations and a residual MF-BSDE with jumps (Tang et al., 2016). Fully coupled mean-field FBSDEs with jumps under monotonicity support stochastic maximum principles and applications to mean-variance portfolio problems and linear-quadratic control (Li et al., 2018). Controlled systems under partial observation can be formulated as mean-field forward-backward equations with a noisy observation process; in that setting, a backward separation method and decomposition technique yield optimality conditions, optimal filters, and closed-form LQ controls (Wang et al., 2015).

The dependence on the law need not be moment-based. One explicit motivation of the deep-learning literature is that many existing methods are tailored to cases where the mean-field interaction only depends on expectation or other moments and are therefore inadequate when the interaction has full distribution dependence (Han et al., 2022). Nor need the law be the law of a current state only: the path-dependent theory allows dependence on the law of the entire past path of $m_u(t)=\mathbb E[u_i(t)]$3, and the weak-formulation mean-field game theory studies regularity of the decoupling or master field by Malliavin methods (Buckdahn et al., 2023, Morgado et al., 2023).

Recent control-theoretic work also uses MF-FBSDEs as synthesis tools in robust regulation. In the continuous-time mean-field stochastic $m_u(t)=\mathbb E[u_i(t)]$4 problem with affine terms, MF-FBSDEs provide equivalence conditions for open-loop $m_u(t)=\mathbb E[u_i(t)]$5 control strategies, while closed-loop solvability is reduced to four coupled Difference Riccati Equations (CDREs), two sets of backward stochastic differential equations and ordinary equations, from which state-feedback gains are recovered (Fang et al., 26 Jul 2025).

Two clarifications follow from these extensions. First, MF-FBSDE is a family of structures rather than a single canonical equation: the coupling may be through marginal laws, joint laws, empirical means, path-laws, jumps, observation filtrations, or robust game variables. Second, explicit feedback formulas are abundant in LQ models, but the general nonlinear theory still relies on fixed points, monotonicity, and measure derivatives rather than universal closed-form decoupling.