Mean-field games with unbounded controls: a weak formulation approach to global solutions

This paper establishes the existence of equilibria for non-Markovian mean-field games with unbounded controls and quadratic running costs using a weak formulation approach grounded in new stability results for quadratic-growth generalized McKean-Vlasov BSDEs, thereby removing previous restrictions on model parameter boundedness and time horizons.

The Gist

Here is an explanation of the paper "Mean-Field Games with Unbounded Controls: A Weak Formulation Approach to Global Solutions," translated into simple language with creative analogies.

The Big Picture: The "Crowd Dance"

Imagine a massive dance floor with millions of people (agents). Everyone wants to dance in a way that makes them look good (maximize their payoff), but their moves affect the overall vibe of the room.

In the world of Mean-Field Games (MFG), we don't track every single person. Instead, we look at the "average" movement of the crowd. Each dancer adjusts their steps based on the general flow of the crowd, and the crowd's flow changes based on what everyone is doing. The goal is to find a Nash Equilibrium: a state where no single dancer wants to change their steps because, given how everyone else is moving, they are already doing the best they can.

The Problem: The "Unbounded" Dancers

Previous math models for these games had a major limitation: they assumed dancers could only move within a small, safe circle (bounded controls).

• The Old Rule: "You can only spin at speeds between 0 and 10 mph."
• The Reality: In real life (finance, traffic, energy), people can sometimes spin very fast. If the cost of spinning is quadratic (meaning spinning twice as fast costs four times as much), a dancer might theoretically want to spin infinitely fast if the reward is high enough.
• The Issue: When you allow for "unbounded" speeds (infinite possibilities), the math usually breaks. The equations explode, and no solution exists.

The Solution: The "Weak" Approach and the "Ghost" Map

The authors, Horst and Sato, found a new way to solve this. Instead of trying to force the dancers to stay in a small circle, they changed the rules of the game slightly.

1. The "Weak" Formulation (The Ghost Map)

Instead of saying, "You must pick a specific speed at this exact second," they say, "You pick a probability distribution of speeds."

• Analogy: Imagine a GPS navigation app. A "strong" approach tells you, "Turn left now." A "weak" approach says, "There is a 70% chance you should turn left, and a 30% chance you should go straight, depending on the traffic."
• By looking at the probability of actions rather than a single fixed action, the math becomes much more flexible. It allows the system to handle "wild" behaviors without breaking.

2. The "Young Measure" (The Crowd's Shadow)

To handle the chaos of millions of people moving unpredictably, the authors use a mathematical tool called Young Measures.

• Analogy: Imagine trying to describe a swirling cloud of smoke. You can't track every single molecule. Instead, you describe the "density" of the smoke at different points.
• In this paper, the "density" isn't just where people are, but how likely they are to move in a certain way. The authors lift the problem from tracking individual dancers to tracking the "shadow" or "density" of their potential moves. This allows them to prove that a stable pattern (equilibrium) exists, even if the dancers are moving wildly.

3. The "Quadratic" Cost (The Price of Speed)

The paper specifically tackles quadratic costs.

• Analogy: Imagine driving a car. If you drive at 10 mph, it costs $1. If you drive at 20 mph, it doesn't cost $2; it costs $4 (because of air resistance and fuel). If you drive at 100 mph, it costs $10,000.
• The authors prove that even with this steep price tag, there is still a "sweet spot" where the system stabilizes. They use a special mathematical "safety net" (called the BMO norm) to ensure that even if the drivers get crazy, the system doesn't crash.

The "Magic Trick": How They Proved It Exists

The authors didn't just guess the solution exists; they built a machine to find it.

1. The Feedback Loop: They created a loop.
   • Step A: Guess how the crowd moves.
   • Step B: Calculate the best move for an individual based on that guess.
   • Step C: Update the crowd's movement based on that new individual move.
2. The Fixed Point: They needed to prove that if you keep doing this loop, it eventually settles down to a single, unchanging pattern (a "fixed point").
3. The Compactness Trick: Usually, when you have infinite possibilities, the loop never settles; it just keeps bouncing around. The authors used the Young Measures to create a "container" (a compact set). Even though the dancers can move infinitely, the probability distribution of their moves stays inside a manageable box.
4. The Result: Because the loop stays inside the box and is continuous, a mathematical theorem (Schauder's Fixed Point Theorem) guarantees that the loop must hit a stopping point. That stopping point is the Global Solution.

Why This Matters

Before this paper, if you tried to model a financial market where traders could bet infinite amounts (unbounded control) with costs that skyrocketed (quadratic growth), the math said, "No solution exists."

This paper says: "Actually, a solution does exist. We just needed to look at the probabilities of the moves, not the moves themselves."

Real-world applications:

• High-Frequency Trading: Where algorithms can trade massive amounts instantly.
• Energy Grids: Where power plants can ramp up or down infinitely fast.
• Traffic Flow: Where cars can accelerate or brake with extreme force.

Summary in One Sentence

The authors invented a new mathematical lens (using "weak" probabilities and "density maps" instead of rigid rules) that proves a stable equilibrium exists for massive groups of agents, even when those agents are allowed to make wild, unbounded moves with expensive consequences.

Technical Summary

Here is a detailed technical summary of the paper "Mean-Field Games with Unbounded Controls: A Weak Formulation Approach to Global Solutions" by Ulrich Horst and Takashi Sato.

1. Problem Statement

The paper addresses the existence of Nash equilibria in Mean-Field Games (MFGs) under the following challenging conditions:

• Non-Markovian Dynamics: The state processes and coefficients depend on the entire history of the path, not just the current state.
• Unbounded Control Spaces: The set of admissible controls is not compact (e.g., $\mathbb{R}^k$), allowing for controls that can grow arbitrarily large.
• Quadratic Growth in Controls: The running cost function $f(t, x, \mu, \alpha)$ exhibits quadratic growth with respect to the control variable $\alpha$ (i.e., $f \sim |\alpha|^2$).
• Weak Formulation: The problem is formulated in a weak sense, where the probability measure is part of the solution, rather than fixing the measure a priori (strong formulation).

The Core Difficulty:
Standard approaches to MFGs often rely on:

1. Compactness of the control space: To apply fixed-point theorems (e.g., Kakutani) and ensure the existence of maximizers for the Hamiltonian.
2. Boundedness of parameters: To ensure Lipschitz continuity of the drivers in the associated Backward Stochastic Differential Equations (BSDEs).
3. Strong Formulation: Which often fails when running costs are path-dependent and discontinuous in the state variable.

The authors aim to prove existence without assuming compact control sets or bounded model parameters, specifically handling quadratic-growth drivers in a generalized McKean-Vlasov BSDE framework.

---

2. Methodology

The authors employ a probabilistic weak formulation approach, transforming the MFG equilibrium problem into a fixed-point problem involving Generalized McKean-Vlasov BSDEs (MV-BSDEs).

A. The Weak Formulation and Hamiltonian

Instead of solving a coupled Forward-Backward SDE (FBSDE) system directly, the authors characterize the equilibrium via a generalized MV-BSDE.

• State Dynamics: $dX_t = \sigma_t(X) dW_t$ (under a reference measure $P$).
• Change of Measure: The equilibrium measure $\bar{P}$ is defined via a Girsanov transformation involving the drift $b$ and the optimal control.
• The BSDE System: The equilibrium is linked to a solution $(X, Y, Z, \bar{P})$ of:
  $$\begin{cases} dX_t = \sigma_t(X) dW_t \\ dY_t = -H_t(X, Z_t, \bar{L}(X, Z_t)) dt + Z_t dW_t \\ Y_T = G(X, \bar{L}(X)) \\ \frac{d\bar{P}}{dP} = \mathcal{E}\left( \int_0^T B_s(X, Z_s, \bar{L}(X)) dW_s \right) \end{cases}$$
  Here, $H$ is the maximized Hamiltonian, and $\bar{L}$ denotes the law of the processes.

B. The Solution Mapping and Young Measures

To handle the lack of compactness in the control space and the potential discontinuity of the optimal control, the authors lift the solution mapping to the space of Integrable Young Measures.

1. Lifting: Instead of mapping a measure flow to another measure flow directly, they map a pair $(\mu, \nu)$ (where $\mu$ is the law of the state and $\nu$ is a Young measure representing the joint law of state and control) to a new pair.
2. Stable Topology: They utilize the stable topology on the space of Young measures. This topology is weaker than the standard Wasserstein topology but strong enough to preserve compactness for integrable measures, which is crucial when the control space is unbounded.
3. Fixed Point: The existence of an equilibrium is reduced to finding a fixed point of this solution mapping $\Phi: \mathcal{P}_1 \times \mathcal{Y}_1 \to \mathcal{P}_1 \times \mathcal{Y}_1$.

C. Key Technical Tools

• BMO Norms: The analysis relies heavily on the Bounded Mean Oscillation (BMO) norm for the $Z$-component of the BSDE solutions. This is essential for handling quadratic drivers where standard Lipschitz assumptions fail.
• Stability of Quadratic BSDEs: The authors prove a novel stability result for quadratic MV-BSDEs with respect to the driver and the underlying probability measure. This ensures the continuity of the solution mapping.
• Truncation Argument: For the case of unbounded parameters, they use a truncation argument (cutting off the coefficients) combined with uniform a priori bounds to pass to the limit.

---

3. Key Contributions

1. Existence with Unbounded Controls: The paper establishes the first existence result for MFG equilibria with non-compact action sets and quadratic running costs. Previous works (e.g., Carmona & Lacker, Possamaï & Tangpi) typically required compact control spaces or bounded coefficients.
2. Generalized MV-BSDEs with Quadratic Drivers: The authors provide existence and stability results for generalized MV-BSDEs with quadratic growth in the $Z$-variable (which corresponds to the control). This extends the literature on quadratic BSDEs to the McKean-Vlasov setting with unbounded coefficients.
3. Use of Integrable Young Measures: They successfully adapt the use of Young measures (previously used for relaxed controls in compact settings) to quadratic settings with unbounded controls. This allows them to bypass the need for compact action spaces by working with the convex hull of laws.
4. Relaxation of Boundedness Assumptions: The framework allows for:
   • Unbounded drifts (e.g., Geometric Brownian Motion with controlled drift).
   • Running costs that are unbounded in the law of the state-control process.
   • Path-dependent and potentially discontinuous cost functions.

---

4. Main Results

The paper presents two primary theorems regarding the existence of solutions to the generalized MV-BSDE (2.12), which imply the existence of MFG equilibria:

• Theorem 2.14 (Bounded Parameters): If the model parameters (drift, cost, terminal condition) are bounded in the mean-field terms, the generalized MV-BSDE admits a solution. The proof uses Schauder's fixed-point theorem on a compact, convex set of integrable Young measures, relying on the uniform BMO bound of the $Z$-component.
• Theorem 2.16 (Unbounded Parameters): If the parameters are unbounded, existence is still guaranteed provided:
  1. The driver satisfies a strictly quadratic growth condition (either bounded from above or below by a quadratic term).
  2. Linear growth conditions hold for the drift and cost in the mean-field terms.
  3. The solution is constructed via a truncation argument, utilizing uniform a priori estimates (Proposition 5.1) derived from the equivalence between MV-BSDEs and MV-FBSDEs.

Corollary: Under these conditions, there exists a mean-field equilibrium in weak formulation where the optimal control is an admissible process in $H^2_{BMO}$.

---

5. Significance and Impact

• Theoretical Advancement: This work bridges a significant gap between the theory of quadratic BSDEs and Mean-Field Games. It demonstrates that the "weak formulation" combined with Young measures is a robust tool for handling non-compactness and quadratic growth, which are common in realistic economic and engineering models but mathematically intractable under strong formulations.
• Applications: The results apply to a broader class of models, including:
  • Portfolio Optimization: Where transaction costs or risk penalties are quadratic in trading rates (unbounded controls).
  • Energy Transition & Resource Extraction: Models involving geometric Brownian motions with controlled drifts.
  • Decentralized Control: Systems where agents have unbounded control capabilities.
• Methodological Innovation: The technique of lifting the problem to the space of integrable Young measures and proving stability with respect to the BMO norm provides a new blueprint for solving MFGs with singular or unbounded controls, potentially influencing future research in stochastic control and game theory.

In summary, Horst and Sato provide a rigorous mathematical foundation for solving complex, non-Markovian MFGs with unbounded controls and quadratic costs, overcoming previous limitations regarding compactness and boundedness assumptions.