Ergodic Mean Field Games of Controls with State Constraints

Abstract: In a mean field game of controls, players seek to minimize a cost that depends on the joint distribution of players' states and controls. We consider an ergodic problem for second-order mean field games of controls with state constraints, in which equilibria are characterized by solutions to a second-order MFGC system where the value function blows up at the boundary, the density of players vanishes at a commensurate rate, and the joint distribution of states and controls satisfies the appropriate fixed-point relation. We prove that such systems are well-posed in the case of monotone coupling and Hamiltonians with at most quadratic growth.

• The paper establishes existence and uniqueness of classical solutions for ergodic MFGC systems with state constraints using fixed-point and monotonicity methods.
• A sharp asymptotic analysis near the boundary is provided, ensuring the value function’s blow-up rate aligns with the vanishing density to enforce constraints.
• The study offers explicit a priori bounds and a stochastic control interpretation linking optimal feedback controls to the long-run ergodic cost.

Ergodic Mean Field Games of Controls with State Constraints: Analysis and Well-Posedness

Introduction and Problem Formulation

The paper "Ergodic Mean Field Games of Controls with State Constraints" (2604.07550) addresses the analysis of mean field games of controls (MFGCs) in the context of ergodic (long-time average) regimes with state constraints. Unlike classical mean field games (MFGs), where a player's cost depends only on the distribution of states, MFGCs introduce couplings through both states and controls, significantly elevating the analytical complexity. The consideration of state constraints, whereby the state variable is constrained to remain within a given domain $\Omega \subseteq \mathbb{R}^n$, reflects many applied contexts, especially in economics.

The focus of the paper is the well-posedness—existence, uniqueness, and regularity—of ergodic second-order MFGC systems under state constraints, described by the coupled nonlinear PDE system: $$\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x) & \text{in } \Omega \ \sigma\Delta m + \nabla \cdot (m D_pH(D_xu,\mu)) = 0 & \text{in } \Omega \ \mu = (I, -D_pH(D_xu,\mu)) \# m \ m \geq 0, \int_\Omega m dx = 1, \lim_{d(x) \to 0}u(x) = \infty \end{cases}$$ where $u$ is the value function, $m$ is the invariant player density, $\mu$ is the joint law of states and controls, $H$ is a Hamiltonian, and $F$ is a coupling functional.

A prominent feature of the analysis is the singularity structure: as the value function $u$ blows up at the boundary of $\Omega$, the density $m$ must converge to zero at a commensurate rate. This is necessary to enforce the state constraint and reveals the nontrivial coupling between the boundary behavior of the HJB and FP equations.

Technical Framework and Assumptions

The analysis is carried out under key structural assumptions on the Hamiltonian $$\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x) & \text{in } \Omega \ \sigma\Delta m + \nabla \cdot (m D_pH(D_xu,\mu)) = 0 & \text{in } \Omega \ \mu = (I, -D_pH(D_xu,\mu)) \# m \ m \geq 0, \int_\Omega m dx = 1, \lim_{d(x) \to 0}u(x) = \infty \end{cases}$$0 and the joint coupling $$\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x) & \text{in } \Omega \ \sigma\Delta m + \nabla \cdot (m D_pH(D_xu,\mu)) = 0 & \text{in } \Omega \ \mu = (I, -D_pH(D_xu,\mu)) \# m \ m \geq 0, \int_\Omega m dx = 1, \lim_{d(x) \to 0}u(x) = \infty \end{cases}$$1, including:

• Monotonicity: $$\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x) & \text{in } \Omega \ \sigma\Delta m + \nabla \cdot (m D_pH(D_xu,\mu)) = 0 & \text{in } \Omega \ \mu = (I, -D_pH(D_xu,\mu)) \# m \ m \geq 0, \int_\Omega m dx = 1, \lim_{d(x) \to 0}u(x) = \infty \end{cases}$$2 and the Lagrangian $$\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x) & \text{in } \Omega \ \sigma\Delta m + \nabla \cdot (m D_pH(D_xu,\mu)) = 0 & \text{in } \Omega \ \mu = (I, -D_pH(D_xu,\mu)) \# m \ m \geq 0, \int_\Omega m dx = 1, \lim_{d(x) \to 0}u(x) = \infty \end{cases}$$3 satisfy Lasry-Lions monotonicity, ensuring uniqueness by the classical monotonicity method in the space of probability measures.
• Growth Conditions: The Hamiltonian is strictly convex in the momentum variable with at most quadratic (or subquadratic) growth.
• Data Regularity: The mappings $$\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x) & \text{in } \Omega \ \sigma\Delta m + \nabla \cdot (m D_pH(D_xu,\mu)) = 0 & \text{in } \Omega \ \mu = (I, -D_pH(D_xu,\mu)) \# m \ m \geq 0, \int_\Omega m dx = 1, \lim_{d(x) \to 0}u(x) = \infty \end{cases}$$4, $$\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x) & \text{in } \Omega \ \sigma\Delta m + \nabla \cdot (m D_pH(D_xu,\mu)) = 0 & \text{in } \Omega \ \mu = (I, -D_pH(D_xu,\mu)) \# m \ m \geq 0, \int_\Omega m dx = 1, \lim_{d(x) \to 0}u(x) = \infty \end{cases}$$5, and $$\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x) & \text{in } \Omega \ \sigma\Delta m + \nabla \cdot (m D_pH(D_xu,\mu)) = 0 & \text{in } \Omega \ \mu = (I, -D_pH(D_xu,\mu)) \# m \ m \geq 0, \int_\Omega m dx = 1, \lim_{d(x) \to 0}u(x) = \infty \end{cases}$$6 are continuous in the appropriate weak-$$\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x) & \text{in } \Omega \ \sigma\Delta m + \nabla \cdot (m D_pH(D_xu,\mu)) = 0 & \text{in } \Omega \ \mu = (I, -D_pH(D_xu,\mu)) \# m \ m \geq 0, \int_\Omega m dx = 1, \lim_{d(x) \to 0}u(x) = \infty \end{cases}$$7 topologies on spaces of measures with polynomial integrability of controls.
• Asymptotic Structure: Near the boundary, the value function and its gradient exhibit precise blow-up asymptotics. The associated feedback optimal controls also diverge, necessitating careful boundary analysis.

The fixed-point formulation of $$\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x) & \text{in } \Omega \ \sigma\Delta m + \nabla \cdot (m D_pH(D_xu,\mu)) = 0 & \text{in } \Omega \ \mu = (I, -D_pH(D_xu,\mu)) \# m \ m \geq 0, \int_\Omega m dx = 1, \lim_{d(x) \to 0}u(x) = \infty \end{cases}$$8, the joint law of states and controls, is a central technical leap compared to classical MFGs. The feedback control $$\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x) & \text{in } \Omega \ \sigma\Delta m + \nabla \cdot (m D_pH(D_xu,\mu)) = 0 & \text{in } \Omega \ \mu = (I, -D_pH(D_xu,\mu)) \# m \ m \geq 0, \int_\Omega m dx = 1, \lim_{d(x) \to 0}u(x) = \infty \end{cases}$$9 induces a self-consistent coupling in law, requiring Schauder's fixed-point theorem due to the non-contractive nature of the map in the infinite-dimensional setting.

Analytical Results

Existence and Uniqueness

Existence is established via a careful compactness and a priori bounds argument. The authors construct a suitable compact convex subset of the space of joint probability measures and demonstrate continuity of the solution map using elliptic regularity theory. Solutions are obtained as classical (not mild) solutions, a significant strengthening over previous work on deterministic or degenerate systems.

Uniqueness is obtained through a monotonicity argument tailored for the MFGC setting. The joint Lasry-Lions monotonicity for both $u$0 and $u$1 is essential here. The authors systematically show that if two solutions exist, they must coincide by exploiting the convexity and monotonicity structure, both for the HJB and the FP components, and for the fixed-point consistency of $u$2.

Boundary Estimates and Asymptotic Behavior

The paper derives fine asymptotic expansions for both the value function and its gradient near the boundary. In particular, for a value function $u$3, the boundary blow-up rate is determined explicitly and matches the vanishing rate of the density $u$4. This ensures that the stationary distribution avoids leakage through the boundary, consistent with the state constraint.

A gradient blow-up estimate of the form $u$5 as $u$6 (where $u$7) is established. This is crucial for well-posedness of the associated Fokker-Planck operator with unbounded drift.

Stochastic Control Interpretation

The authors provide a stochastic control representation of the ergodic cost, showing that the critical ergodic constant $u$8 is the long-run average cost under optimal feedback controls restricted to be admissible for the state constraint. The admissibility of the feedback law is rigorously verified despite its singular nature near the boundary.

Joint Law of States and Controls

Analysis of the fixed-point relation for $u$9 is central: $m$0. Existence and uniqueness of such a joint law is established via monotonicity and regularity, together with tightness and stability arguments in the space $m$1.

Fokker-Planck Well-Posedness

The analysis leverages recent sharp results for singular Fokker-Planck equations on bounded domains, guaranteeing uniqueness and regularity of invariant measures even when the drift diverges at the boundary, provided the divergence matches the singularity of the value function.

Numerical and Contradictory Claims

A prominent numerical result is the derivation of explicit a priori bounds for the norms $m$2 of the joint law, independent of smallness parameters, and a demonstration that solutions exist in naturally weighted function spaces without recourse to regularization of the data. The paper claims, in contradiction with the previous literature, that classical (as opposed to mild) well-posedness can be achieved for this second-order ergodic MFGC system under state constraints, in both existence and uniqueness.

Implications and Future Directions

Theoretical implications: The framework opens the door for ergodic theory and long-time averaged dynamics in the context of MFGCs with state constraints and fully nonlinearly coupled PDEs, connecting variational analysis, stochastic control, and PDE theory. The results extend previous ergodic MFG literature from state couplings to joint control couplings, and from first-order to second-order systems.

Practical applications: Constrained ergodic MFGCs naturally model controlled systems with feasible region restrictions—examples include asset management with regulatory boundaries, or spatial models in economics where agents cannot exit allowed regions. The theory enables analysis of equilibrium behavior in such constrained large-population settings.

Potential developments:

• Extension of existence and uniqueness to non-monotone couplings, possibly via variational or viscosity solution frameworks.
• Analysis of time-dependent (parabolic) MFGC systems with state constraints and their ergodic limits.
• Numerical approximation schemes for ergodic MFGCs under state constraints, leveraging the explicit asymptotics and a priori estimates provided.
• Examination of constrained MFGC systems in higher regularity or degenerate scenarios (e.g., non-uniformly elliptic diffusions or irregular domains).

Conclusion

This paper establishes strong well-posedness results for ergodic second-order mean field games of controls with state constraints, resolving existence and uniqueness under monotonicity and quadratic growth assumptions on the coupling and Hamiltonian. The analysis leverages sharp boundary asymptotics and a precise treatment of the singular Fokker-Planck equation, and introduces a general framework for addressing ergodic MFGCs with joint control-state coupling. The implications span both theoretical advances in nonstandard MFG theory and potential practical models in constrained agent-based systems.