[Paper Review] Linear Quadratic Mean Field Game with Control Input Constraint

Ying Hu, Huang Jianhui+1 more|arXiv (Cornell University)|Oct 19, 2016

Stochastic processes and financial applications10 references3 citations

TL;DR

This paper studies linear-quadratic mean-field games with control input constraints using a projection-based Hamiltonian system to model decentralized strategies. It establishes well-posedness of the consistency condition via monotonicity methods and proves an $\epsilon$-Nash equilibrium with $\epsilon = O(1/\sqrt{N})$, where $N$ is the number of agents, under general closed convex control constraints.

ABSTRACT

In this paper, we study a class of linear-quadratic (LQ) mean-field games in which the individual control process is constrained in a closed convex subset $Γ$ of full space $\mathbb{R}^m$. The decentralized strategies and consistency condition are represented by a class of mean-field forward-backward stochastic differential equation (MF-FBSDE) with projection operators on $Γ$. The wellposedness of consistency condition system is obtained using the monotonicity condition method. The related $ε$-Nash equilibrium property is also verified.

Motivation & Objective

• To analyze large-population stochastic differential games where individual controls are constrained in a closed convex set $\Gamma \subset \mathbb{R}^m$.
• To derive decentralized optimal strategies based on local information and the mean field limit.
• To establish the well-posedness of the consistency condition system arising from the mean-field forward-backward stochastic differential equation (MF-FBSDE) with projection operators.
• To verify the $\epsilon$-Nash equilibrium property of the derived strategies under the control constraint.
• To generalize existing LQ-MFG frameworks by incorporating general closed convex control constraints, including no-shorting and market access constraints.

Proposed method

• Model the problem using a linear-quadratic mean-field game framework with control constraints in $\Gamma$.
• Apply the stochastic maximum principle to derive the Hamiltonian system with projection operators onto $\Gamma$.
• Formulate the consistency condition as a mean-field forward-backward stochastic differential equation (MF-FBSDE) involving projection operators.
• Use the monotonicity condition method to prove the well-posedness of the MF-FBSDE system.
• Employ projection operator properties (e.g., non-expansiveness and variational inequality characterization) to analyze the system dynamics.
• Establish convergence of the $N$-player game to the mean-field limit by bounding the difference in cost functionals using $L^2$-estimates and $O(1/\sqrt{N})$ error terms.

Experimental results

Research questions

• RQ1 How can decentralized strategies be characterized in a linear-quadratic mean-field game when individual controls are constrained to a closed convex set $\Gamma$?
• RQ2 What is the structure of the consistency condition system in the presence of control constraints, and is it well-posed?
• RQ3 Can the derived strategies form an $\epsilon$-Nash equilibrium, and what is the rate of convergence as $N \to \infty$?
• RQ4 How do projection operators onto $\Gamma$ affect the Hamiltonian system and the resulting MF-FBSDE?
• RQ5 What conditions ensure the existence and uniqueness of solutions to the MF-FBSDE with projection operators under control constraints?

Key findings

• The consistency condition system, represented as a mean-field forward-backward stochastic differential equation (MF-FBSDE) with projection operators, is well-posed under the monotonicity condition.
• The optimal decentralized strategies are characterized via a Hamiltonian system involving projection onto the constraint set $\Gamma$.
• The $\epsilon$-Nash equilibrium property holds with $\epsilon = O(1/\sqrt{N})$, indicating convergence of the $N$-player game to the mean-field limit.
• The error in cost functional between the $N$-player and mean-field solutions is bounded by $O(1/\sqrt{N})$, derived through $L^2$-estimates and projection operator inequalities.
• The projection operator $\mathbf{P}_\Gamma$ satisfies key properties: non-expansiveness ($|\mathbf{P}_\Gamma[x] - \mathbf{P}_\Gamma[y]| \leq |x - y|$) and variational inequality ($\langle \mathbf{P}_\Gamma[x] - x, \mathbf{P}_\Gamma[x] - y \rangle \leq 0$ for all $y \in \Gamma$).
• The analysis confirms that the system remains stable and consistent under general closed convex control constraints, including positive orthant and general cones.