Second-Order MFGC Systems Overview

• Second-order MFGC systems are defined by coupled backward–forward PDEs that incorporate diffusion in both the Hamilton–Jacobi–Bellman and Fokker–Planck equations with control-dependent interactions.
• They feature ergodic and time-dependent formulations on bounded domains or toroidal settings with unique boundary constraints and fixed-point conditions linking the state and control distributions.
• Variational and master-equation methods underpin the analysis, facilitating convex optimization frameworks and numerical schemes for understanding nonlinear dynamics and homogenization effects.

Searching arXiv for the cited papers to ground the article in current records. Second-order MFGC system denotes a class of coupled backward–forward equilibrium systems in which diffusion is present in both the Hamilton–Jacobi–Bellman and Fokker–Planck equations and the mean-field interaction involves controls either directly or through a control-dependent aggregate quantity. In the most literal recent usage, it refers to an ergodic second-order mean field game of controls with state constraints on a bounded domain, where the unknowns are a value function $u$, an ergodic constant $\rho$, a stationary density $m$, and a joint state–control distribution $\mu$, linked by an HJB equation, an FP equation, and a fixed-point condition $\mu=(I,-D_pH(D_xu,\mu))\# m$ (Graber et al., 8 Apr 2026). Closely related work develops time-dependent potential MFGC systems on the torus (Graber et al., 2020), master-equation formulations for potential or fully nonlinear second-order models (Liao et al., 2024, Bensoussan et al., 21 Mar 2025), and adjacent second-order MFG systems with local coupling, controlled diffusion, homogenization, or state constraints that clarify which features are specific to MFGC and which belong to second-order HJB–FP couplings more generally (Ignazio et al., 2024, Cesaroni et al., 2016, Porretta et al., 2023).

1. Canonical PDE realizations

The canonical ergodic second-order MFGC system with state constraints is

$\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x), &x \in \Omega, \[1mm] \sigma\Delta m + \nabla \cdot (mD_pH(D_x u,\mu)) = 0, &x \in \Omega, \[1mm] \mu = (I,-D_pH(D_xu,\mu)) \# m, & \[1mm] m \geq 0, \hspace{1cm} \displaystyle \int_\Omega m\, dx = 1, \hspace{1cm} \lim_{d(x) \to 0} u(x) = \infty. \end{cases}$

Here $u$ is the ergodic value function, $\rho$ is the long-run average cost per unit time, $m$ is the invariant state density, and $$\mu\in\mathcal P_{q'}(\overline\Omega\times\mathbb R^n)$$ is the joint distribution of states and controls (Graber et al., 8 Apr 2026). The fixed-point identity is the defining extra ingredient distinguishing an MFG of controls from a standard MFG: the coupling is not only through the first marginal $\rho$0, but through the full equilibrium law on $\rho$1.

A time-dependent second-order MFGC system on the flat torus appears in the potential framework

$\rho$2

This formulation is second order and possibly degenerate, since the diffusion matrix $\rho$3 is assumed symmetric nonnegative and need not be uniformly elliptic (Graber et al., 2020). It combines a local density coupling $\rho$4 with a nonlocal control coupling through the aggregate quantity $\rho$5.

A further potential formulation, developed through the master equation rather than through an explicit $\rho$6–$\rho$7 PDE pair, considers generalized mean field control and potential MFGC in which the Hamiltonian depends on the joint law of position and momentum. The finite-$\rho$8 HJB equation,

$\rho$9

and the limiting Wasserstein-space HJB/master equations make the second-order structure explicit through both idiosyncratic and common noise (Liao et al., 2024).

2. Second-order structure and the meaning of the coupling

In this literature, “second-order” has a precise PDE and probabilistic meaning: both the backward value equation and the forward law equation contain second-order diffusion terms generated by Brownian noise. In bounded-domain minimal-time models, the second-order character is visible in

$m$0

arising from the controlled diffusion $m$1 (Ducasse et al., 2020). In torus models with local coupling, the same role is played by

$m$2

within a standard finite-horizon HJB–FP system (Nakamura et al., 26 Jan 2026). In controlled-diffusion models, the second-order operator itself becomes endogenous because agents control both drift and diffusion, which yields a fully nonlinear HJB equation and a forward equation whose coefficients depend on $m$3 and $m$4 (Ignazio et al., 2024).

The coupling mechanism varies substantially across papers. In strict MFGC, the cost depends on the joint distribution of states and controls. This is the explicit viewpoint in the ergodic state-constrained system, where $m$5 and $m$6 depend on $m$7 (Graber et al., 8 Apr 2026). The same distinction from standard MFG is central in potential MFGC, where players interact through the aggregate control signal $m$8 and the flux $m$9 is part of the state of the system (Graber et al., 2020). In the generalized potential framework, the Hamiltonian depends on the joint empirical law of positions and momenta $\mu$0, and the limiting object is the joint law $\mu$1 (Liao et al., 2024).

Other second-order systems are adjacent rather than strictly MFGC. Local coupling models use $\mu$2 or $\mu$3, so the mean field acts pointwise through the density rather than through the control distribution (Nakamura et al., 26 Jan 2026, Cesaroni et al., 2016). In the minimal-time bounded-domain model, the coupling acts as a density-dependent control constraint,

$\mu$4

which the paper describes as close to a second-order control-constrained or congestion-type MFG, while also noting that it does not explicitly use the acronym MFGC (Ducasse et al., 2020). In the ergodic state-constraint paper, the coupling appears only in the running cost $\mu$5; that paper is explicitly not about congestion, and its relevance is strongest when the letter “C” is read as “constraints” rather than “controls” (Porretta et al., 2023). This suggests that the phrase “Second-Order MFGC System” is best understood as a family resemblance term whose exact meaning depends on where the mean-field dependence enters the coupled HJB–FP structure.

3. Boundary mechanisms, ergodic regimes, and constrained dynamics

A major branch of the theory concerns bounded domains and state constraints. In the ergodic MFGC system with state constraints, the admissible controls are those for which trajectories remain in $\mu$6 almost surely, with no imposed Dirichlet or Neumann condition on the boundary. The state constraint is instead encoded by the blow-up condition

$\mu$7

which forces the optimal drift to become singular and the stationary density to vanish at a commensurate rate (Graber et al., 8 Apr 2026). The paper proves the matched asymptotics

$\mu$8

so that the control blow-up remains integrable at the level of the joint measure $\mu$9.

The ergodic second-order MFG with state constraints develops the same boundary mechanism for the simpler Hamiltonian $\mu=(I,-D_pH(D_xu,\mu))\# m$0, $\mu=(I,-D_pH(D_xu,\mu))\# m$1. Its system is

$\mu=(I,-D_pH(D_xu,\mu))\# m$2

and the sharp boundary behavior is

$\mu=(I,-D_pH(D_xu,\mu))\# m$3

The value function blows up, the optimal feedback drift is singular inward, and the invariant density “flattens” near the boundary (Porretta et al., 2023).

A distinct bounded-domain mechanism appears in second-order local minimal-time MFG. There the boundary is absorbing rather than constraining, and both equations carry Dirichlet data: $\mu=(I,-D_pH(D_xu,\mu))\# m$4 Agents stop as soon as they reach $\mu=(I,-D_pH(D_xu,\mu))\# m$5, the forward density vanishes on the boundary because particles are removed, and the long-time behavior is

$\mu=(I,-D_pH(D_xu,\mu))\# m$6

uniformly as $\mu=(I,-D_pH(D_xu,\mu))\# m$7, where $\mu=(I,-D_pH(D_xu,\mu))\# m$8 solves

$\mu=(I,-D_pH(D_xu,\mu))\# m$9

The literature therefore treats two distinct second-order boundary regimes: exclusion by singular inward drift and removal by absorption (Ducasse et al., 2020).

4. Variational formulations, duality, and computational schemes

Potentiality is the main structural device that turns a second-order MFGC system into a convex optimization problem. In the weak theory for potential MFGC, the flux variable $\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x), &x \in \Omega, \[1mm] \sigma\Delta m + \nabla \cdot (mD_pH(D_x u,\mu)) = 0, &x \in \Omega, \[1mm] \mu = (I,-D_pH(D_xu,\mu)) \# m, & \[1mm] m \geq 0, \hspace{1cm} \displaystyle \int_\Omega m\, dx = 1, \hspace{1cm} \lim_{d(x) \to 0} u(x) = \infty. \end{cases}$0 is introduced so that the forward equation becomes linear in $\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x), &x \in \Omega, \[1mm] \sigma\Delta m + \nabla \cdot (mD_pH(D_x u,\mu)) = 0, &x \in \Omega, \[1mm] \mu = (I,-D_pH(D_xu,\mu)) \# m, & \[1mm] m \geq 0, \hspace{1cm} \displaystyle \int_\Omega m\, dx = 1, \hspace{1cm} \lim_{d(x) \to 0} u(x) = \infty. \end{cases}$1, and the coupled PDE system is shown to be the Euler–Lagrange system of a Fenchel–Rockafellar dual pair (Graber et al., 2020). The primal problem minimizes

$\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x), &x \in \Omega, \[1mm] \sigma\Delta m + \nabla \cdot (mD_pH(D_x u,\mu)) = 0, &x \in \Omega, \[1mm] \mu = (I,-D_pH(D_xu,\mu)) \# m, & \[1mm] m \geq 0, \hspace{1cm} \displaystyle \int_\Omega m\, dx = 1, \hspace{1cm} \lim_{d(x) \to 0} u(x) = \infty. \end{cases}$2

over the Fokker–Planck constraint, while the dual problem minimizes

$\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x), &x \in \Omega, \[1mm] \sigma\Delta m + \nabla \cdot (mD_pH(D_x u,\mu)) = 0, &x \in \Omega, \[1mm] \mu = (I,-D_pH(D_xu,\mu)) \# m, & \[1mm] m \geq 0, \hspace{1cm} \displaystyle \int_\Omega m\, dx = 1, \hspace{1cm} \lim_{d(x) \to 0} u(x) = \infty. \end{cases}$3

over a relaxed HJ admissible set. Existence of a weak solution and uniqueness of $\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x), &x \in \Omega, \[1mm] \sigma\Delta m + \nabla \cdot (mD_pH(D_x u,\mu)) = 0, &x \in \Omega, \[1mm] \mu = (I,-D_pH(D_xu,\mu)) \# m, & \[1mm] m \geq 0, \hspace{1cm} \displaystyle \int_\Omega m\, dx = 1, \hspace{1cm} \lim_{d(x) \to 0} u(x) = \infty. \end{cases}$4, with $\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x), &x \in \Omega, \[1mm] \sigma\Delta m + \nabla \cdot (mD_pH(D_x u,\mu)) = 0, &x \in \Omega, \[1mm] \mu = (I,-D_pH(D_xu,\mu)) \# m, & \[1mm] m \geq 0, \hspace{1cm} \displaystyle \int_\Omega m\, dx = 1, \hspace{1cm} \lim_{d(x) \to 0} u(x) = \infty. \end{cases}$5 unique on $\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x), &x \in \Omega, \[1mm] \sigma\Delta m + \nabla \cdot (mD_pH(D_x u,\mu)) = 0, &x \in \Omega, \[1mm] \mu = (I,-D_pH(D_xu,\mu)) \# m, & \[1mm] m \geq 0, \hspace{1cm} \displaystyle \int_\Omega m\, dx = 1, \hspace{1cm} \lim_{d(x) \to 0} u(x) = \infty. \end{cases}$6, follow from this duality framework.

A more recent potential theory works directly at the level of the master equation. The generalized MFC HJB on Wasserstein space,

$\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x), &x \in \Omega, \[1mm] \sigma\Delta m + \nabla \cdot (mD_pH(D_x u,\mu)) = 0, &x \in \Omega, \[1mm] \mu = (I,-D_pH(D_xu,\mu)) \# m, & \[1mm] m \geq 0, \hspace{1cm} \displaystyle \int_\Omega m\, dx = 1, \hspace{1cm} \lim_{d(x) \to 0} u(x) = \infty. \end{cases}$7

with $\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x), &x \in \Omega, \[1mm] \sigma\Delta m + \nabla \cdot (mD_pH(D_x u,\mu)) = 0, &x \in \Omega, \[1mm] \mu = (I,-D_pH(D_xu,\mu)) \# m, & \[1mm] m \geq 0, \hspace{1cm} \displaystyle \int_\Omega m\, dx = 1, \hspace{1cm} \lim_{d(x) \to 0} u(x) = \infty. \end{cases}$8, generates the potential MFGC master equation after differentiation and the potentiality identification $\begin{cases} -\sigma\Delta u + H(D_xu,\mu) + \rho = F(\mu,x), &x \in \Omega, \[1mm] \sigma\Delta m + \nabla \cdot (mD_pH(D_x u,\mu)) = 0, &x \in \Omega, \[1mm] \mu = (I,-D_pH(D_xu,\mu)) \# m, & \[1mm] m \geq 0, \hspace{1cm} \displaystyle \int_\Omega m\, dx = 1, \hspace{1cm} \lim_{d(x) \to 0} u(x) = \infty. \end{cases}$9 (Liao et al., 2024). That paper proves global classical well-posedness of the generalized MFC HJB equation and of the associated potential MFGC master equation, derives a Lipschitz particle approximation of the optimal feedback with rate $u$0, and constructs a closed-loop approximate Nash equilibrium with error $u$1.

For second-order systems with local coupling terms, potential structure also underlies algorithmic developments. The generalized conditional gradient method is formulated on the primal variables $u$2, with exploitability

$u$3

and optimality gap

$u$4

Under a quadratic Hamiltonian with convection, local coupling, and potential assumptions, the paper proves explicit convergence estimates in exploitability and optimality gap, with exponential convergence in one dimension for adaptive step-sizes and polynomial decay in higher dimensions (Nakamura et al., 26 Jan 2026).

5. Probabilistic formulations, master equations, and fully nonlinear generalizations

A different axis of generalization replaces convex duality by stochastic maximum principle and FBSDE methods. In second-order fully nonlinear MFGs with degenerate diffusions, the drift depends nonlinearly on state, distribution, and control, the diffusion depends on state and distribution but not on control, and the diffusion may be degenerate, unbounded in the state variable, and nonlinear in the distribution argument (Bensoussan et al., 21 Mar 2025). The central object is the master equation

$u$5

and the main result is global-in-time well-posedness of the corresponding FBSDEs and uniqueness of the classical solution to the MFG master equation under $u$6-monotonicity and additional structural conditions. The paper is explicit that this is not an MFG of controls in the strict sense, but it is a reference point for second-order systems with control-dependent nonlinear drift and state-law interaction.

Controlled diffusion changes the second-order structure in another way. In the finite-horizon whole-space model

$u$7

agents control both drift and diffusion, so the HJB equation becomes

$u$8

with forward equation

$u$9

This is a second-order MFG rather than an MFGC in the strict control-distribution sense, but it shows that once the diffusion coefficient is optimized, the HJB ceases to be semilinear and the FP equation acquires coefficients depending on second derivatives of the value function (Ignazio et al., 2024). The paper proves viscosity existence for the HJB equation, derives a $\rho$0 regularity result for $\rho$1 in the space variable by adapting and extending a result from Krylov, and then proves well-posedness for the full MFG system.

The probabilistic and master-equation viewpoints therefore isolate a broad lesson: in second-order problems, the forward law and backward value equations may remain formally similar while the regularity theory, monotonicity assumptions, and even the choice of basic unknowns change substantially once the coupling is moved from the density to the control distribution, the state law, or the diffusion coefficient.

6. Asymptotic, homogenized, and non-Markovian extensions

Second-order MFGC intuition is also shaped by asymptotic regimes in which the coupled HJB–FP structure survives microscopically but changes macroscopically. In the small-noise homogenization problem,

$\rho$2

the joint limit of vanishing noise and rapidly oscillating coefficients produces the effective first-order system

$\rho$3

where $\rho$4 are determined by a second-order ergodic MFG cell problem (Cesaroni et al., 2016). The paper proves continuity, coercivity, monotonicity, and local Lipschitz properties of the effective operators, but also shows that in general

$\rho$5

so the effective limit may lose the usual MFG structural relation between Hamiltonian and transport drift. Although this work is about MFG rather than a full MFGC system, it is directly relevant to second-order HJB–FP asymptotics because it shows that homogenization can destroy the original variational or MFG structure even when the microscopic cell problem remains second order.

A non-Markovian extension replaces standard diffusion by subdiffusion and yields a time-fractional second-order MFG system with local coupling: $\rho$6 Its variational formulation replaces the standard Benamou–Brenier density $\rho$7 by the memory-weighted quantity $\rho$8, so the kinetic action becomes

$\rho$9

This paper is not about MFGC in the strict sense, but it extends the variational architecture of second-order HJB–FP systems to a non-Markovian regime driven by inverse stable subordinators (Qing et al., 2018).

Across these extensions, the persistent invariant is the coupled backward–forward structure; what changes is the object through which equilibrium is closed. In strict second-order MFGC this object is the joint state–control law $m$0 or the control-induced aggregate signal $m$1. In adjacent second-order models it may be a pointwise density $m$2, an invariant measure under a singular inward drift, an optimized diffusion coefficient, a homogenized effective current, or a fractional memory term. The literature thus uses “Second-Order MFGC System” most precisely for mean field games of controls, but the broader second-order HJB–FP corpus supplies the structural comparisons needed to interpret that term rigorously.