Infinite-Time Mean Field FBSDEs Overview

• Infinite-time mean field FBSDEs are coupled stochastic systems on [0,∞) that replace terminal conditions with decay requirements to analyze long-horizon control and game problems.
• They employ exponentially weighted L² spaces and strong monotonicity conditions to ensure existence, uniqueness, and stability of solutions.
• The framework bridges dynamic programming and master equations, underpinning value-function regularity and long-time asymptotic analysis in mean field games.

Infinite-time mean field forward-backward stochastic differential equations are coupled stochastic systems on the horizon $[0,\infty)$ in which the coefficients depend on the law of the solution and the backward component is controlled not by a finite terminal condition but by discounting, dissipativity, or transversality at infinity. They provide a probabilistic formulation for infinite-horizon mean field control and mean field game problems, especially in McKean–Vlasov settings where the representative agent interacts with the evolving population distribution. In recent work, they have been used to characterize discounted infinite-time equilibria, representative-player value functions, and elliptic master equations; the regularity theory now includes strong well-posedness, uniqueness in law, and Lions-differentiability of the value function with respect to the measure argument (Song et al., 16 Mar 2026).

1. Mathematical setting and canonical formulations

The finite-horizon precursor is the fully coupled mean field FBSDE studied by Carmona and Delarue, in which the coefficients depend on the joint law of $(X_t,Y_t)$ and the backward component is specified by a terminal condition: $$\begin{cases} dX_t = B\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + \Sigma\big(t, X_t, Y_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dW_t, \ dY_t = -F\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + Z_t\,dW_t, \ X_0 = x_0,\qquad Y_T = G\big(X_T, \mathbb{P}_{X_T}\big). \end{cases}$$ This formulation is inherently finite-horizon: the arguments used there rely on a fixed $[0,T]$, terminal data at time $T$, and compactness and regularity mechanisms adapted to that setting (Carmona et al., 2012).

Infinite-time theories replace terminal conditions by integrability and decay requirements in exponentially weighted spaces. One infinite-horizon McKean–Vlasov form is

$$\begin{cases} dX_t = B(t, X_t, Y_t, L(X_t, Y_t))\, dt + \sigma\, dW_t, \ dY_t = -F(t, X_t, Y_t, L(X_t, Y_t))\, dt + Z_t\, dW_t, \ X_0 = \xi, \end{cases}$$

with solutions sought in $L^2_K(0,\infty,\mathbb{R}^3)$, where

$$\| (v_t) \|^2_K = \mathbb{E} \left[ \int_0^{\infty} e^{-Kt} |v_t|^2 dt \right] < +\infty .$$

This weighted formulation is central because it encodes the long-time decay needed for solvability on $[0,\infty)$ (Bayraktar et al., 2021).

In discounted infinite-time mean field games, a canonical system is

$$\left\{ \begin{aligned} dX_t &= \partial_y H(X_t, \mathcal{L}_{X_t, Y_t})\,dt + dB_t, \ dY_t &= -\big[\partial_x H(X_t, \mathcal{L}_{X_t, Y_t}) - r Y_t\big]\,dt + Z_t\,dB_t, \ X_0 &= \xi, \end{aligned} \right.$$

where the Hamiltonian $(X_t,Y_t)$0 aggregates optimal controls and mean-field interactions, and $(X_t,Y_t)$1 is the discount rate (Song et al., 16 Mar 2026). In the discounted mean field game formulation of Yang and Song, related social-equilibrium and representative-player systems are written with the measure input $(X_t,Y_t)$2 and a representative state initialized at $(X_t,Y_t)$3, reflecting the distinction between the population flow and an individual best response (Yang et al., 4 Oct 2025).

The transition from finite to infinite time is not formal. The finite-horizon framework does not directly extend because one must replace terminal conditions by decay or stationary requirements, obtain uniform a priori estimates independent of the horizon, and work in weighted norms or discounted formulations adapted to $(X_t,Y_t)$4 (Carmona et al., 2012).

2. Well-posedness on the infinite horizon

A central problem is existence and uniqueness of adapted solutions for fully coupled McKean–Vlasov systems on $(X_t,Y_t)$5. Under Lipschitz continuity and monotonicity conditions, one strand of the theory establishes unique solvability via two distinct methods: a continuity method interpolating between a linear base equation and the nonlinear FBSDE, and a fixed-point/monotonicity approach exploiting strong monotonicity of $(X_t,Y_t)$6 in $(X_t,Y_t)$7 and $(X_t,Y_t)$8 in $(X_t,Y_t)$9 together with a balance between monotonicity and Lipschitz constants (Bayraktar et al., 2021). In both cases the solution lies in an exponentially weighted $$\begin{cases} dX_t = B\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + \Sigma\big(t, X_t, Y_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dW_t, \ dY_t = -F\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + Z_t\,dW_t, \ X_0 = x_0,\qquad Y_T = G\big(X_T, \mathbb{P}_{X_T}\big). \end{cases}$$0 space.

The 2026 regularity paper strengthens this solvability theory for an extended class of infinite-time FBSDEs. Under monotonicity, Lipschitz, and regularity conditions, it proves strong well-posedness—existence and pathwise uniqueness—and also uniqueness in law for the infinite-time mean field FBSDE above. The strong solution $$\begin{cases} dX_t = B\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + \Sigma\big(t, X_t, Y_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dW_t, \ dY_t = -F\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + Z_t\,dW_t, \ X_0 = x_0,\qquad Y_T = G\big(X_T, \mathbb{P}_{X_T}\big). \end{cases}$$1 is obtained in an exponentially weighted $$\begin{cases} dX_t = B\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + \Sigma\big(t, X_t, Y_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dW_t, \ dY_t = -F\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + Z_t\,dW_t, \ X_0 = x_0,\qquad Y_T = G\big(X_T, \mathbb{P}_{X_T}\big). \end{cases}$$2 space, and the proof combines a continuity method with strongly monotonic couplings and contraction mapping arguments adapted to the infinite horizon. The uniqueness-in-law statement is formulated as a Yamada–Watanabe type result and uses canonical Wiener space together with regular conditional probabilities to control the distributional structure of solutions (Song et al., 16 Mar 2026).

A further extension replaces fixed quadratic monotonicity structures by a generalized monotonicity condition involving two flexible nonnegative functions $$\begin{cases} dX_t = B\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + \Sigma\big(t, X_t, Y_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dW_t, \ dY_t = -F\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + Z_t\,dW_t, \ X_0 = x_0,\qquad Y_T = G\big(X_T, \mathbb{P}_{X_T}\big). \end{cases}$$3 and $$\begin{cases} dX_t = B\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + \Sigma\big(t, X_t, Y_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dW_t, \ dY_t = -F\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + Z_t\,dW_t, \ X_0 = x_0,\qquad Y_T = G\big(X_T, \mathbb{P}_{X_T}\big). \end{cases}$$4. In that framework, the infinite-horizon McKean–Vlasov FBSDE

$$\begin{cases} dX_t = B\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + \Sigma\big(t, X_t, Y_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dW_t, \ dY_t = -F\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + Z_t\,dW_t, \ X_0 = x_0,\qquad Y_T = G\big(X_T, \mathbb{P}_{X_T}\big). \end{cases}$$5

admits a unique solution in $$\begin{cases} dX_t = B\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + \Sigma\big(t, X_t, Y_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dW_t, \ dY_t = -F\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + Z_t\,dW_t, \ X_0 = x_0,\qquad Y_T = G\big(X_T, \mathbb{P}_{X_T}\big). \end{cases}$$6, together with an a priori estimate and continuous dependence on coefficients and initial data (Hua et al., 2024).

Stability with respect to initial conditions is also structurally important. For discounted mean field games, continuous dependence estimates of the form

$$\begin{cases} dX_t = B\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + \Sigma\big(t, X_t, Y_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dW_t, \ dY_t = -F\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + Z_t\,dW_t, \ X_0 = x_0,\qquad Y_T = G\big(X_T, \mathbb{P}_{X_T}\big). \end{cases}$$7

lead to a flow property: $$\begin{cases} dX_t = B\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + \Sigma\big(t, X_t, Y_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dW_t, \ dY_t = -F\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + Z_t\,dW_t, \ X_0 = x_0,\qquad Y_T = G\big(X_T, \mathbb{P}_{X_T}\big). \end{cases}$$8 This identifies the representative-player system with the social-equilibrium system when the individual initial state is sampled from the equilibrium law (Yang et al., 4 Oct 2025).

3. Value functions and elliptic master equations

The discounted infinite-time mean field game framework associates to the representative player the value function

$$\begin{cases} dX_t = B\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + \Sigma\big(t, X_t, Y_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dW_t, \ dY_t = -F\big(t, X_t, Y_t, Z_t, \mathbb{P}_{(X_t, Y_t)}\big)\,dt + Z_t\,dW_t, \ X_0 = x_0,\qquad Y_T = G\big(X_T, \mathbb{P}_{X_T}\big). \end{cases}$$9

where $[0,T]$0 is initialized at $[0,T]$1 while the population flow is generated by $[0,T]$2 with law $[0,T]$3 (Song et al., 16 Mar 2026). This function plays the same role as in finite-horizon mean field games, but the associated PDE is elliptic rather than parabolic because the horizon is stationary and discounted.

The corresponding infinite-horizon master equation is

$[0,T]$4

with classical derivatives in $[0,T]$5 and Lions derivatives in $[0,T]$6 (Song et al., 16 Mar 2026). In the formulation of Yang and Song, the same equation arises from the FBSDE system at Nash equilibrium, and the value function is shown to be a viscosity solution under the stated assumptions on $[0,T]$7 (Yang et al., 4 Oct 2025).

The probabilistic representation is not merely formal. The infinite-time FBSDE furnishes the dynamic programming principle, the compatibility between representative and social equilibria via the flow property, and the stochastic calculus needed to identify the master equation. In the discounted setting this produces an elliptic master equation, in contrast with the time-dependent parabolic master equations of finite-horizon mean field games (Yang et al., 4 Oct 2025).

This connection also clarifies the role of infinite-time mean field FBSDEs in the analytical structure of mean field games. They are not only equilibrium characterizations; they also control the smooth dependence of the value function on the initial law and thereby provide a probabilistic route to master-equation well-posedness (Song et al., 16 Mar 2026).

4. Lions differentiability and value-function regularity

The main technical advance of the 2026 regularity theory is the construction of the Lions derivative $[0,T]$8 in the infinite-time setting. The key difficulty is that finite-horizon differentiability arguments do not transfer directly: there is no terminal condition, and uniform estimates in perturbation parameters are more delicate on $[0,T]$9 (Song et al., 16 Mar 2026).

The approach starts from directional perturbations of the initial law. For a direction $T$0, linearized directional derivative FBSDEs produce processes $T$1 such that

$T$2

The derivative is then explicitly characterized as the time-zero value of a specifically constructed FBSDE with an initial perturbation at $T$3 (Song et al., 16 Mar 2026).

For discrete laws, $T$4 is realized by a coupled FBSDE indexed by the atoms of the distribution. For absolutely continuous laws, the construction proceeds by approximating the initial law with discrete measures and passing to the limit. The resulting map

$T$5

is bounded and jointly continuous, yielding a full Lions-differentiable structure for the value function in the measure variable (Song et al., 16 Mar 2026).

Related global-in-time analyses of mean field games use sensitivity estimates of backward solutions with respect to initial conditions through Jacobian flows. In the control-theoretic approach of 2024, such Jacobian estimates lead to linear functional differentiability of value functions and classical well-posedness of master equations under a small mean field effect (Bensoussan et al., 2024). In a degenerate finite-horizon setting, Jacobian and Hessian flow estimates yield a value functional that is $T$6 in time and $T$7 in spatial and distribution variables, again via probabilistic FBSDE methods (Bensoussan et al., 2024). These developments show that the infinite-time Lions-differentiability result fits into a broader probabilistic regularity program, but the discounted infinite-horizon case requires its own stability and approximation arguments.

5. Variants, control/game interpretations, and special structures

Infinite-time mean field FBSDEs arise directly from the stochastic maximum principle for mean field type control problems and mean field games. In one formulation, solvability of the infinite-horizon McKean–Vlasov FBSDE is used to establish existence and uniqueness for both control and game problems, with the corresponding optimal control or Nash equilibrium obtained from the Hamiltonian minimizer (Bayraktar et al., 2021). This identification is standard in the infinite-horizon literature: the FBSDE is the adjoint-state system that encodes optimality and the fixed-point consistency of the mean field.

In linear-quadratic models, the relation between infinite-horizon equilibria and FBSDE solvability becomes especially explicit. For two-person mean-field LQ stochastic differential games on an infinite horizon, existence of an open-loop Nash equilibrium is characterized by solvability of a system of mean-field FBSDEs together with convexity of the cost functionals, while closed-loop representations are expressed through coupled algebraic Riccati equations; the closed-loop Nash equilibrium itself is characterized by coupled symmetric algebraic Riccati equations (Li et al., 2020).

The framework also extends to conditional McKean–Vlasov systems with common noise and regime switching. An infinite-horizon fully coupled conditional McKean–Vlasov FBSDE with Markovian switching and common noise admits a unique solution in an exponentially weighted $T$8 space under Lipschitz, coercivity, and generalized domination-monotonicity conditions, and this result supports both infinite-horizon LQ control and mean-field game formulations (Wei et al., 21 Nov 2025). A closely related discounted LQ theory with mean-field switching diffusions establishes well-posedness of the state equation and the adjoint equation—formulated as infinite-horizon mean-field forward and backward SDEs with Markov chains—and derives an optimal feedback law from two algebraic Riccati equations (Ding et al., 19 Jun 2025).

These model classes indicate that infinite-time mean field FBSDEs are not confined to a single PDE regime. They support discounted and stationary mean field games, Pontryagin-based control problems, common-noise models, and Markov-switching LQ systems, provided one can impose a suitable combination of monotonicity, convexity, or dissipativity.

6. Numerical approximation and long-time asymptotics

The numerical analysis of mean field FBSDEs remains largely finite-horizon, but it is closely connected to infinite-time problems. A numerical study of fully coupled McKean–Vlasov FBSDEs proposes two Picard-based schemes—a tree structure algorithm and a grid discretization algorithm—and combines each with a continuation in time method. The continuation method divides $T$9 into short subintervals where Picard iteration converges, thereby extending the range of coupling strengths and time horizons that can be handled in practice (Angiuli et al., 2018).

That work is explicit about its limitation: it treats only finite, though arbitrarily large, $$\begin{cases} dX_t = B(t, X_t, Y_t, L(X_t, Y_t))\, dt + \sigma\, dW_t, \ dY_t = -F(t, X_t, Y_t, L(X_t, Y_t))\, dt + Z_t\, dW_t, \ X_0 = \xi, \end{cases}$$0. Truly infinite time would require separate analysis involving ergodicity or invariant measures and lies outside the scope of those algorithms (Angiuli et al., 2018). This distinction is important because infinite-time mean field FBSDEs are not simply large-$$\begin{cases} dX_t = B(t, X_t, Y_t, L(X_t, Y_t))\, dt + \sigma\, dW_t, \ dY_t = -F(t, X_t, Y_t, L(X_t, Y_t))\, dt + Z_t\, dW_t, \ X_0 = \xi, \end{cases}$$1 limits of numerical solvers; they require a separate well-posedness theory in weighted spaces.

Long-time asymptotics provides the conceptual bridge. In displacement monotone mean field games, the $$\begin{cases} dX_t = B(t, X_t, Y_t, L(X_t, Y_t))\, dt + \sigma\, dW_t, \ dY_t = -F(t, X_t, Y_t, L(X_t, Y_t))\, dt + Z_t\, dW_t, \ X_0 = \xi, \end{cases}$$2 distance between solutions of the associated Pontryagin FBSDEs yields quantitative stability of Nash equilibria as $$\begin{cases} dX_t = B(t, X_t, Y_t, L(X_t, Y_t))\, dt + \sigma\, dW_t, \ dY_t = -F(t, X_t, Y_t, L(X_t, Y_t))\, dt + Z_t\, dW_t, \ X_0 = \xi, \end{cases}$$3. The value function converges, after subtraction of a linear drift determined by an ergodic constant, to a limit described by an infinite-horizon mean field game system, and the convergence is exponential (Cirant et al., 2024). This shows that infinite-time mean field FBSDEs can arise either directly, through discounted or stationary formulations, or as asymptotic objects governing the long-time stabilization of finite-horizon equilibria.

Taken together, these results place infinite-time mean field FBSDEs at the intersection of probabilistic well-posedness, value-function regularity, and long-run equilibrium analysis. Their distinctive features are the use of exponentially weighted solution spaces, the need for monotonicity and stability at infinity, and the explicit role they play in constructing and differentiating value functions on Wasserstein space (Song et al., 16 Mar 2026).