Abstract
In this paper, we consider the social optimal problem of discrete-time finite state space mean field games (referred to as finite mean field games). Unlike the individual optimization of their own cost function in competitive models, in the problem we consider, individuals aim to optimize the social cost by finding a fixed point of the state distribution to achieve equilibrium in the mean field game. We provide a sufficient condition for the existence and uniqueness of the individual optimal strategies used to minimize the social cost. According to the definition of social optimum and the derived properties of social optimal cost, the existence and uniqueness conditions of equilibrium solutions under initial-terminal value constraints in the finite horizon and the existence and uniqueness conditions of stable solutions in the infinite horizon are given. Finally, two examples that satisfy the conditions for the above solutions are provided.
Introduction
Mean Field Games (MFGs) are a robust mathematical framework for modeling the collective behavior of massive populations of interacting agents. Applications range from traffic flow management and financial market dynamics to the coordination of swarm robotics. Traditionally, MFG research has focused on the Nash Equilibrium, a competitive state where each agent selfishly minimizes their own cost based on the average behavior (the "mean field") of the population.
However, many real-world scenarios require a Social Optimum, where agents cooperate to minimize the total cost for the entire system. While the social optimum has been extensively studied in continuous-time models, its application to discrete-time, finite-state space MFGs (Finite MFGs) has remained under-explored. This paper addresses this gap by establishing the mathematical conditions necessary for the existence and uniqueness of social optimal solutions in both finite and infinite time horizons.
Methodology
The authors develop a rigorous framework for social optimization using a combination of optimization theory and fixed-point analysis. The methodology is built upon four pillars:
- Dynamic Programming Principle (DPP): The researchers redefine recursive social cost. In a social context, an individual's optimal cost is not merely a local minimization but accounts for how their actions influence the total social welfare over time.
- Karush-Kuhn-Tucker (KKT) Conditions: To ensure the uniqueness of the optimal strategy, the authors utilize KKT conditions. These are standard tools in non-linear programming used to solve optimization problems under specific constraints.
- Fixed-Point Theorems: The existence of an equilibrium is established using Brouwer’s fixed-point theorem. This proves the existence of a state distribution where agent strategies and population flow are perfectly balanced.
- Quotient Space Analysis ($R^s/R$): For infinite horizon problems—where time is unbounded—costs can grow linearly. The authors employ a quotient space approach to identify "stationary" solutions, where the distribution of agents remains stable even as total costs accumulate.
Key Insights and Theoretical Results
Strategy Divergence
A significant finding of this research is that the social optimum strategy diverges significantly from the competitive Nash strategy when an agent's transition cost depends on the decisions of agents in other states. In a cooperative framework, the "math of the move" changes because agents account for the externalities they impose on the rest of the population.
The Role of Monotonicity
The uniqueness of the social optimum relies heavily on a specific monotonicity-like inequality, identified as Assumption A4. If the cost functions exhibit sufficient smoothness as the population distribution shifts, the system is guaranteed to settle into a single, predictable social optimum. This provides a clear mathematical boundary for when a stable cooperative state can be reached.
Existence and Uniqueness Theorems
The paper culminates in the derivation of three primary theorems (Theorems 3.1, 4.2, and 5.2) that guarantee a single, stable social optimum for both finite and infinite time scales. Notably, the framework for infinite horizons does not require a "discount factor" to keep sums finite, making it applicable to a broader range of physical and economic models where future rewards are not necessarily diminished.
Validation through Synthetic Examples
As a theoretical study, the authors validate their theorems using two synthetic mathematical models:
- 3-State Model: Utilizes a cost function with logarithmic (entropy-like) terms to demonstrate the uniqueness of solutions in a finite horizon.
- 2-State Model: Demonstrates the existence and uniqueness of stationary solutions in an infinite horizon.
Conclusion and Future Work
This paper provides a foundational mathematical rigorousness for cooperative behavior in finite MFGs. By proving that the optimal strategy is a continuous function of the state distribution and future costs, the authors ensure the stability of numerical algorithms designed to solve these games.
While the conditions provided are sufficient, they may not be strictly necessary, suggesting that some systems might achieve a unique social optimum even without meeting these strict requirements. Future research directions include extending this framework to models with "common noise" (external shocks affecting all agents) and exploring the convergence rates from finite-agent models to the mean field limit.
Summary updated