Mean Field Games: Games with an Infinite Number of Players

Markets, Traffic, and “Impersonal” Competition

Imagine thousands of traders on a financial market. Each one is rational and influences the price, but each is “small” compared to the market as a whole. Or thousands of pedestrians in a narrow corridor — each one optimally chooses their path, but interacts with the “average density” of the crowd rather than with every individual person. Mean Field Games (MFG) is the mathematical theory for exactly such systems: $N \to \infty$ “small” rational agents with “interaction through the mean field.” The theory was independently developed by Lasry-Lions (France) and Huang-Malham-Ma (Canada) in 2006-2007. It is one of the fastest-growing fields in applied mathematics.

Key Idea: The Mean Field

With $N\to\infty$, a “typical” agent interacts not with specific other agents, but with the “distribution” of the entire population $m(x, t)$ — the density of agents in state $x$ at time $t$.

Homogeneity assumption: All agents are identical (i.i.d. — independently and identically distributed initial states).

Nash Equilibrium in the limit $N\to\infty$: A typical agent optimizes their strategy, treating $m(x,t)$ as “given” (not dependent on their own actions). In equilibrium, $m(x,t)$ is generated by exactly this optimal strategy of the typical agent — self-consistency!

MFG System of Equations

Two interconnected problems for the typical agent given $m(x,t)$:

HJB (backward in time): Optimal control $u^*(x, t)$ for the typical agent given $m$:

$ -\frac{\partial V}{\partial t} - \nu \Delta V + H(x, \nabla V, m) = 0, \quad V(x, T) = g(x, m(T)) $

FPK (forward in time): Evolution of the distribution $m$ under the optimal strategy $u^*(x, t)$:

$ \frac{\partial m}{\partial t} - \nu \Delta m - \mathrm{div}(m \cdot H_p(x, \nabla V, m)) = 0, \quad m(x, 0) = m_0(x) $

Here $\nu \ge 0$ is the diffusion coefficient (noise in the dynamics), $H(x, p, m)$ is the “game” Hamiltonian.

Self-consistency: $V$ depends on $m$ (the agent adapts to the crowd), and $m$ depends on $V$ (the crowd moves optimally). This is a nonlinear system!

Analysis of the Equations

Existence of a solution: Under reasonable conditions on $H$ there exists a solution $(V, m)$ to the MFG system. For monotonic games (where “more agents $\to$ less attractive”), uniqueness holds.

Physical interpretation:

• $V(x,t)$ — the “value” of being in state $x$ at time $t$ for a typical agent
• $m(x,t)$ — the distribution of agents in state space
• $H_p = \partial H/\partial p$ — optimal “drift field” (velocity of agents’ movement)
• $\Delta$ terms — random fluctuations (diffusion)

Applications of MFG

Crowd dynamics: $m(x,t)$ is the density of pedestrians. Each one chooses the path, minimizing time plus discomfort from crowding. $H = |u|^2/2 + \alpha m$ (penalty for dense crowds). The MFG solution reproduces the observed physical effect of “formation of lanes” in a corridor!

Financial markets: “Trading with price impact.” $N$ traders sell an asset, each sale lowers the price. MFG gives the “optimal” trading algorithm for each trader, accounting for the aggregate influence of the entire group.

Telecommunications: Users of a distributed network select resources (channels, servers). MFG describes the “distributional equilibrium" for a large number of users.

Epidemiology: Agents choose their level of “social distancing.” MFG describes “rational” behavior during a pandemic — which may not coincide with the “social optimum”!

Numerical Methods

Forward-backward scheme:

1. Initial $m^0$
2. Solve HJB backward (for given $m^k$) $\to V^k$
3. Solve FPK forward (for given $V^k$) $\to m^{k+1}$
4. Repeat until convergence

Deep MFG (Carmona-Lauriere, 2021): Approximate $V(x, t)$ and $m(x, t)$ with neural networks. Scales to high dimensions (50+ variables).

Full Analysis: The Clustering Problem

Problem: $N\to\infty$ particles seek to reach a target at $x=0$. Dynamics: $dX = u,dt + \sigma,dW$. Cost: $J = E\left[\int \left( \frac{|u|^2}{2} + \alpha m(X, t) \right)dt + |X(T)|^2\right]$.

Penalty $\alpha m$: Agents dislike “clustering” ($\alpha m(X, t)$ — the cost of being in a cluster).

MFG system: HJB: $-\frac{\partial V}{\partial t} - \frac{\sigma^2}{2} \Delta V + \frac{|\nabla V|^2}{2} + \alpha m = 0$. FPK: $\frac{\partial m}{\partial t} + \mathrm{div}(m \nabla V) - \frac{\sigma^2}{2} \Delta m = 0$.

Explicit solution (one-dimensional, without diffusion): For $\alpha \to 0$: $V(x, t) = \frac{x^2}{2(T-t)}$ (free particle). $m(x, t) \to \delta(x - x_0 e^{-t})$ (all move to 0). For $\alpha > 0$: agents “spread out” from the cluster, $m$ is more diffuse. This is a "compromise" between moving toward the goal and avoiding the crowd.

Mean Field Games: Intuition

When the number of players is extremely large (millions), tracking every player’s strategy is impossible. MFG (Lasry-Lions, Caines-Huang, 2006-2007) made a breakthrough: instead of each player interacting with everyone, each player interacts with the averaged field (density distribution of all players).

This is analogous to thermodynamics: instead of tracking each molecule — describing via macroscopic quantities (temperature, pressure). MFG is “the thermodynamics of strategic behavior.”

Coupled System of Equations

MFG is described by a system of two coupled PDEs:

1. HJB equation (forward in time): $-\partial_t u + H(x, \nabla u, m) = 0$, where $u$ is the value function of the representative player, $m$ is the density distribution of the players
2. Fokker-Planck equation (backward in time): $\partial_t m - \mathrm{div}(m \cdot \nabla_p H) = \sigma^2 \Delta m$, describing the evolution of the density $m$

Connection: $u$ depends on $m$ (the player responds to the crowd), and $m$ is formed by the optimal policy $\nabla u$ (the crowd consists of optimizing players). This leads to a fixed point.

Existence and Numerical Methods

Existence of a solution to MFG has been proven under monotonicity of the Hamiltonian with respect to $m$. Numerical methods:

• Fictitious play: iteratively update $u$ for a fixed $m$, then $m$ for the new $u$
• Sinkhorn method for entropic regularization
• DeepLearning approaches: Deep MFG (Carmona-Laurière) — neural networks represent $u$ and $m$

Applications

• Crowds and evacuation: modeling the movement of people in buildings, at stadiums
• Energy: millions of consumers choose tariffs and consumption
• Finance: systemic risk, modeling crowded trades
• Epidemiology: individual vaccination decisions accounting for the population effect
• Cryptocurrencies: miners as MFG players competing for computing resources