Limit Theory for $N$-Player $α$-Potential Games

Source: arXiv:2606.09815 · Published 2026-06-08 · By Xin Guo, Meng Wang, Yufei Zhang

TL;DR

This paper develops a rigorous theoretical connection between finite-player $\alpha$-potential games and potential mean field games (MFGs) as the number of players $N$ tends to infinity. $\alpha$-potential games provide an approximate potential structure for analyzing complex stochastic differential games with finitely many players, by reducing equilibrium computation to minimizing an $\alpha$-potential function that approximates individual costs up to error $\alpha$. The authors construct a novel finite-dimensional $\alpha_N$-potential function for $N$-player games driven by controlled diffusions with both idiosyncratic and common noise, characterizing the approximation error $\alpha_N$ analytically. They then study the large population limit and prove that normalized $\alpha_N$-potential functions converge in value and minimizers to a mean field control (MFC) problem with measure-valued controls—thus deriving potential MFGs as natural asymptotic limits of $\alpha_N$-potential games as $N \to \infty$ with $\alpha_N \to 0$.

The authors provide a rigorous equivalence between the vanishing $\lim_{N\to\infty} \alpha_N=0$ condition and classical conditions for potential MFGs expressed via differential geometric tools on Wasserstein space. They prove that the objective functional of the limiting MFC problem serves as a potential for the MFG, extending classical results for finite-player potential games. The paper also demonstrates propagation of chaos results: approximate equilibria of the finite $N$-player $\alpha_N$-potential games converge to the MFG equilibria for general controlled diffusions with common noise and non-separable control-state interactions. Concrete examples illustrate the necessity of lifting controls to measure-valued spaces to handle nonlinear cost interactions. This work thus unifies and extends finite-player approximate potential game theory with infinite-player MFG potential frameworks through a limit theory grounded in stochastic analysis and optimal control.

Key findings

• Constructed a finite-dimensional N-player $\alpha_N$-potential function (Theorem 2.1) for stochastic differential games with controlled diffusions, quantifying $\alpha_N$ in terms of second derivatives of cost differences (Equation 2.4 and Corollary 2.1).
• Proved that the normalized $\alpha_N$-potential functions (scaled by 1/N) converge in optimal values and approximate minimizers to a mean field control problem (Theorem 3.1), even with common noise and nonlinear non-separable interactions.
• Showed necessity of lifting control spaces to measure-valued controls to capture limit points of approximate equilibria, illustrated via explicit counterexamples with and without common noise (Examples 3.1 and 3.2).
• Established equivalence of the asymptotic condition $\lim_{N\to\infty} \alpha_N = 0$ to classical potential MFG conditions using a Poincaré lemma on Wasserstein space for closed differential forms (Theorems 4.1 and 4.2).
• Proved that the limit MFC objective functional acts as a potential function for the corresponding MFG with measure-valued controls (Theorem 4.3).
• Demonstrated propagation of chaos: $\alpha_N$-Nash equilibria minimizing the $\alpha_N$-potential functions converge to mean field equilibria of the MFG (Corollary 4.1).
• The finite-player $\alpha_N$-potential construction avoids infinite-dimensional derivative computations used in prior work by exploiting the decoupled player dynamics and path integrals in an enlarged state-control space.
• The results cover general controlled diffusions with idiosyncratic and common noise, and cost coupling nonlinear in states and controls.

Threat model

n/a - The paper develops theoretical results about stochastic differential games and their limits; it does not address adversarial threat models or security settings.

Methodology — deep read

The paper proceeds through several rigorous methodological steps:

1. Threat model and assumptions: The adversary here is not a traditional attacker but rather the challenge is analyzing equilibrium in stochastic differential games with many interacting agents. The model considers N-player games where each player's state evolves as a controlled diffusion influenced by idiosyncratic Brownian motion and a common noise component. Players’ controls influence their own dynamics and cost functions that may depend on all players’ states and controls. Key assumptions include Lipschitz continuity and boundedness of dynamics and costs, and a non-degeneracy condition on diffusion to ensure regularity.

2. Data and problem setup: The setup uses a filtered probability space supporting independent initial states, idiosyncratic Brownian motions (one per player), and a common Brownian motion. Each player’s dynamics and cost functions are defined explicitly, with the controls taken from suitable progressive measurable processes valued in compact action sets. The costs depend on the joint states and controls. The finite-player game data include N, the function coefficients, and initial laws.

3. Construction of the $\alpha_N$-potential function: The authors innovate a novel $\alpha_N$-potential function defined on the product space of all players’ controls (Equation 2.4). Unlike previous constructions requiring differentiability of controlled states w.r.t. controls, this approach exploits the decoupled structure of players’ dynamics and constructs the potential via path integrals over an enlarged state-control space, yielding a finite-dimensional control problem. This enables characterizing the approximation error $\alpha_N$ analytically through norms of symmetric second derivatives of differences in running and terminal costs across players (Theorem 2.1).

4. Large population limit and lifting controls: To analyze the limit as $N \to \infty$, the authors impose homogeneity and weak interaction assumptions allowing permutation symmetries. The normalized $\alpha_N$-potential functions are expressed in terms of functions $F^N$ and $G^N$ depending on empirical joint distributions of states and controls. Passing to the limit formally leads to a mean field control problem with a cost depending on the conditional law of the state and control processes given the common noise.

Due to nonlinear dependence on control distributions, the limit control space is lifted from strict controls to measure-valued controls defined on a canonical space with flows of probability measures. This lifting handles oscillation effects in control sequences and discontinuities, ensuring stability and convergence of equilibria.

5. Convergence analysis: The main theorem (3.1) proves that approximate minimizers of the normalized $\alpha_N$-potential functions converge (in Wasserstein-2 distance) to optimal measure-valued controls of the limiting MFC problem. The key technical challenge addressed is that the finite-N costs $F^N$, $G^N$ differ slightly from the limiting costs $F^{\infty}$, $G^{\infty}$, with discrepancies that vanish as $O(1/N)$ uniformly over admissible controls.

6. Equivalence and geometric characterization: The authors prove that the asymptotic condition $\lim_{N\to \infty} \alpha_N=0$ is equivalent to classical conditions for potential MFGs via a Poincaré lemma on Wasserstein space, showing that suitable closed differential forms correspond to exact differentials represented by the potential function. The proof extends Green's theorem to this infinite-dimensional setting.

7. Reproducibility: The paper is theoretical with no code release or empirical dataset. The stochastic processes are defined rigorously on filtered probability spaces with canonical constructions. All definitions and proofs are detailed in the paper and supplemental materials.

Example walk-through: Consider a N-player game with identical dynamics and cost functions depending on empirical distributions. The $\alpha_N$-potential function is constructed as in (2.4). Minimizing this function yields an $\alpha_N$-Nash equilibrium. As $N$ grows, the normalized potential and minimizers converge to those of the MFC problem with measure-valued controls, which solves the limiting MFG. Counterexamples (Section 3.4) demonstrate that without measure-valued controls the limit fails to hold, illustrating the necessity of the lifting.

Throughout, the methodology blends stochastic analysis, control theory, and geometry on Wasserstein space to build a comprehensive limit theory connecting approximate finite-player games and infinite-player MFGs.

Technical innovations

• Novel finite-dimensional construction of $\alpha_N$-potential functions for stochastic N-player differential games using path integrals in an enlarged state-control space, avoiding infinite-dimensional derivative calculations of controlled state processes.
• Use of measure-valued controls in the limiting mean field control problem to capture limits of approximate equilibria when running costs depend nonlinearly on control distributions.
• Establishment of equivalence between vanishing $\alpha_N$-approximation errors and classical potential MFG conditions via a Poincaré lemma on Wasserstein space, extending Green's theorem to less regular forms.
• Proof of propagation of chaos results for general controlled diffusions including common noise, non-separable control interactions, and nonlinear dependence on the population distribution.

Limitations

• The analysis assumes Lipschitz continuity and boundedness of coefficients; extensions to unbounded coefficients or discontinuous controls are not addressed.
• The necessity and sufficiency of measure-valued controls stems from examples but practical algorithms for computing such controls are not discussed.
• The convergence results rely on non-degeneracy assumptions for idiosyncratic noise to handle nonlinear running costs; removing this assumption remains open.
• The setting assumes full knowledge of model coefficients and does not consider model uncertainty or adversarial perturbations.
• The theory focuses on approximation and limit characterization; numerical methods and empirical validations are not provided.
• The interaction among players is restricted to symmetric, weakly interacting setups; heterogeneous or networked interaction topologies are unaddressed.

Open questions / follow-ons

• How to devise practical algorithms to compute or approximate measure-valued controls for the limiting MFC problem accounting for nonlinear control interactions?
• Can the non-degeneracy assumption on idiosyncratic noise be relaxed while preserving convergence and propagation of chaos results?
• How do finite-sample behaviors of $\alpha_N$-potential games and their approximate equilibria compare with the theoretical asymptotic limits?
• Can these limit theories be extended to incorporate heterogeneity among players or more general network structures in interactions?

Why it matters for bot defense

Though this work is theoretical and focused on stochastic differential games and mean field limits, the convergence of approximate Nash equilibria via potential functions has conceptual connections to bot-defense scenarios where many weakly interacting agents (bots) may be modeled as a large population game. The lifting to measure-valued controls and handling of nonlinear interaction terms could inspire advanced modeling of collective bot behaviors interacting through shared system states or detection mechanisms.

Practitioners designing CAPTCHA or bot-detection strategies might look to these results as a theoretical foundation for scaling solutions from a few interacting entities to a large population, ensuring that approximate equilibria and optimal defense strategies converge in the mean field limit. The propagation of chaos arguments may inform approaches to decouple complex attacker populations into simpler representative models. However, direct application requires translating the stochastic game framework here into actionable bot interaction models and control objectives relevant to security contexts.

Cite

@article{arxiv2606_09815,
  title={ Limit Theory for $N$-Player $α$-Potential Games },
  author={ Xin Guo and Meng Wang and Yufei Zhang },
  journal={arXiv preprint arXiv:2606.09815},
  year={ 2026 },
  url={https://arxiv.org/abs/2606.09815}
}

Read the full paper

• arXiv:2606.09815 (abstract)
• arXiv:2606.09815 (PDF)