Phase Transitions in Turnpike Theory For Mean-Field Games

Technical Summary: Phase Transitions in Turnpike Theory for Mean-Field Games

1. Problem Statement and Motivation
This paper investigates the long-time behavior of translation-invariant mean-field games (MFGs) on the flat torus $T^d$, specifically focusing on the "turnpike property." The turnpike phenomenon posits that optimal trajectories over a long time horizon $[0, T]$ spend the majority of their time near a stationary optimal state, with deviations confined to thin boundary layers near $t=0$ and $t=T$.

While existing turnpike results in MFG theory often rely on the Lasry–Lions monotonicity condition (which enforces a specific sign structure on the interaction coupling and excludes pattern formation), this work isolates a complementary mechanism. The authors study a system where the interaction is of potential type, defined by a smooth, even, mean-zero kernel $K$. The central question is how the Fourier spectrum of $K$ and the competition between diffusion ($\nu$) and coupling strength ($\gamma$) determine the stability of the uniform equilibrium and the nature of the turnpike rate. The mechanism is structurally analogous to Turing's diffusion-driven instability.

2. Mathematical Model and Setup
The study considers a finite-horizon MFG system $(MFG_T)$ on $T^d$:
$$\begin{cases} -\partial_t \phi - \nu \Delta \phi + \frac{1}{2}|\nabla \phi|^2 = \gamma (K * m)(x), \\ \partial_t m - \nu \Delta m - \text{div}(m \nabla \phi) = 0, \\ m(0, \cdot) = m_0, \quad \phi(T, \cdot) = g. \end{cases}$$
The interaction term is $F(x, m) = \gamma (K * m)(x)$, derived from the quadratic energy functional $F(m) = \frac{\gamma}{2} \iint K(x-y)m(x)m(y) dx dy$. The uniform equilibrium is $(\bar{\phi}, \bar{m}, \bar{\lambda}) = (0, 1, 0)$.

The analysis relies on linearizing the system around this uniform state. For each Fourier mode $\xi \in \mathbb{Z}^d \setminus \{0\}$, the linearized dynamics are governed by a $2 \times 2$ matrix $M_\xi$. The stability is determined by the dispersion relation:
$$\sigma_\xi(\gamma) = \nu^2 (2\pi|\xi|)^4 + \gamma (2\pi|\xi|)^2 \hat{K}(\xi).$$
A critical coupling strength $\gamma_c$ is defined as the threshold where the spectral gap closes for the most unstable mode:
$$\gamma_c := \min_{\xi: \hat{K}(\xi) < 0} \frac{\nu^2 (2\pi|\xi|)^2}{|\hat{K}(\xi)|}.$$
If $\hat{K}(\xi) \geq 0$ for all $\xi$, then $\gamma_c = +\infty$.

3. Methodology
The paper employs a multi-scale analytical approach combining:

• Fourier Spectral Analysis: Exact computation of the eigenvalues of the linearized mode matrices. The authors establish the identity $M_\xi^2 = \sigma_\xi(\gamma)I$, which simplifies the mode-wise dynamics significantly.
• Boundary Layer Estimates: For the subcritical regime ($\gamma < \gamma_c$), the authors derive explicit exponential decay estimates for the boundary layers using the hyperbolic nature of the linearized modes.
• Center Manifold and Lyapunov–Schmidt Reduction: At the critical threshold ($\gamma = \gamma_c$) and in the supercritical regime ($\gamma > \gamma_c$), the authors perform a reduction to the critical eigenspace. This involves fixing the translation phase (due to the torus symmetry) to handle the neutral direction and deriving reduced scalar amplitude equations.
• Perturbative Fixed-Point Arguments: Nonlinear terms are treated as quadratic perturbations to the linear turnpike estimates, ensuring local well-posedness and stability for small initial/terminal data.

4. Key Results

• Subcritical Regime ($0 \leq \gamma < \gamma_c$):
  The uniform state is exponentially stable in the turnpike sense. The solution $(m_T, \phi_T)$ satisfies:
  $$\|m_T(t) - 1\|_{H^{-1}}^2 + \|\nabla \phi_T(t)\|_{L^2}^2 \leq C \left( e^{-c\rho(\gamma)t} + e^{-c\rho(\gamma)(T-t)} \right),$$
  where the spectral gap $\rho(\gamma) = \min_{\xi \neq 0} \sqrt{\sigma_\xi(\gamma)}$ dictates the decay rate. As $\gamma \uparrow \gamma_c$, the rate degenerates as $\rho(\gamma) \sim C^* \sqrt{\gamma_c - \gamma}$.

• Critical Regime ($\gamma = \gamma_c$):
  The spectral gap closes. Under nondegeneracy assumptions (specifically that the critical set consists of a single pair $\{\pm \xi_0\}$ and the cubic reduction coefficient $\beta > 0$), the turnpike behavior transitions from exponential to algebraic. The critical center coordinate $a(t)$ obeys the reduced dynamics $\dot{a} = -\beta a^3 + O(a^5)$. Consequently, the midpoint amplitude decays algebraically:
  $$|a(T/2)| \leq C T^{-1/2}.$$

• Supercritical Regime ($\gamma > \gamma_c$):
  A pitchfork bifurcation occurs. A smooth branch of non-uniform stationary solutions emerges from the uniform state. In a phase-fixed slice, the density takes the form:
  $$m_\gamma(x) = 1 + A(\gamma) \cos(2\pi \xi_0 \cdot x) + R_\gamma(x),$$
  where the amplitude scales as $A(\gamma) \sim \sqrt{\gamma - \gamma_c}$. The full family of solutions is generated by translations of this profile.

• Propagation of Chaos:
  Under standard asymptotic consistency assumptions for symmetric $N$-player equilibria in the subcritical regime, the paper establishes qualitative propagation of chaos. Specifically, if the empirical measures of the $N$-player system converge in probability to the uniform measure, the joint laws of any fixed number of particles converge to the product of the uniform measure. The paper notes that quantitative Wasserstein rates are not claimed due to dimension dependence.

5. Significance and Claims
The paper claims to provide an explicit Fourier spectral control of the turnpike rate, contrasting with previous monotonicity-based proofs. The primary contributions are:

1. Explicit Thresholds: The critical coupling $\gamma_c$ and the rate constant $C^*$ are entirely computable from the Fourier coefficients of the kernel $K$ and the diffusion coefficient $\nu$.
2. Turing Analogy: The work precisely formalizes the analogy between MFG phase transitions and Turing's diffusion-driven instability, identifying the exact spectral condition where diffusion fails to suppress the interaction.
3. Three-Scale Coherence: The same spectral threshold governs three distinct phenomena: the exponential turnpike rate in finite-horizon problems, the bifurcation of stationary states, and the transition in the asymptotic regime relevant for propagation of chaos.
4. Limitation of Monotonicity: The paper demonstrates that the Lasry–Lions monotonicity condition is sufficient but not necessary for turnpike behavior; turnpike properties persist in the presence of attractive (negative Fourier) interactions as long as the coupling remains below the critical threshold $\gamma_c$.

The authors explicitly state that the results are local (near the uniform equilibrium) and that global classification of stationary solutions for $\gamma \gg \gamma_c$ remains an open problem. Similarly, quantitative rates for propagation of chaos in the $N$-player limit are not provided, as they depend on the dimension $d$.