Phase transitions in turnpike theory for mean-field games

[Siddharth Karuturi] · 2026

Read Phase transitions in turnpike theory for mean-field games by [Siddharth Karuturi] (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-20213.txt.

Introduction

Siddharth Karuturi studies phase shifts in turnpike theory for mean-field games. The game uses a smooth, even, mean-zero kernel, linearized around the uniform state into mode systems. When the kernel transform turns negative, a finite threshold marks loss of stability, else infinite. Near it, the gap scales like the square root of the distance to it, and below it the uniform state is exponentially stable. At the threshold the gap closes and decay turns algebraic, while above it nonuniform states split off through a pitchfork equation.

Catalog record

CREATOR[Siddharth Karuturi]
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE43499
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.20213

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Phase transitions in turnpike theory for mean-field games [Siddharth Karuturi] Abstract . We study a translation-invariant mean-field game on the flat torus with interaction F ⁡ ( x , m ) = γ ⁡ ( K ∗ m ) ​ ( x ) F(x,m)=\gamma(K*m)(x) , where K K is smooth, even, and mean-zero. The interaction is of potential type, arising as the first variation of a quadratic energy, though the stationary system is not treated variationally. Linearizing around the uniform equilibrium yields mode-wise 2 × 2 2\times 2 systems with dispersion σ ξ ​ ( γ ) = ν 2 ​ ( 2 ​ π ​ | ξ | ) 4 + γ ​ ( 2 ​ π ​ | ξ | ) 2 ​ K ^ ​ ( ξ ) \sigma_{\xi}(\gamma)=\nu^{2}(2\pi|\xi|)^{4}+\gamma(2\pi|\xi|)^{2}\hat{K}(\xi) . If K ^ \hat{K} is negative for some mode, a finite threshold γ c = min K ^ ​ ( ξ ) < 0 ⁡ ν 2 ​ ( 2 ​ π ​ | ξ | ) 2 | K ^ ​ ( ξ ) | \gamma_{c}=\min_{\hat{K}(\xi)<0}\frac{\nu^{2}(2\pi|\xi|)^{2}}{|\hat{K}(\xi)|} marks loss of stability; otherwise γ c = + ∞ \gamma_{c}=+\infty . Near criticality, the spectral gap scales as ρ ⁡ ( γ ) ∼ C ∗ ​ γ c − γ \rho(\gamma)\sim C_{*}\sqrt{\gamma_{c}-\gamma} . For γ < γ c \gamma<\gamma_{c} , the uniform state is exponentially stable in the turnpike sense for finite-horizon problems, with rate ρ ⁡ ( γ ) \rho(\gamma) . At γ = γ c \gamma=\gamma_{c} , the gap closes and, after phase fixing and center-manifold reduction, one obtains algebraic midpoint decay of order T − 1 / 2 T^{-1/2} . For γ > γ c \gamma>\gamma_{c} , a branch of nonuniform stationary solutions bifurcates via a pitchfork-type amplitude equation, with translations generating the full family. Finally, under standard asymptotic-consistency assumptions on symmetric N N -player equilibria in the subcritical regime, we obtain qualitative propagation of chaos, without quantitative rates. 2020 Mathematics Subject Classification 49N80, 35Q89, 91A16, 35B32, 35B36 1. Introduction 1.1. Background and motivation The turnpike phenomenon asserts that, on a long time horizon [0,T] , an optimal trajectory spends most of its time near a stationary optimal state, approaching it from both endpoints through thin boundary layers. This idea originated in economics [ 7 ] and has since been developed for finite-dimensional optimal control [ 14 , 18 ] , infinite-dimensional systems [ 17 ] , and mean-field games [ 2 , 3 ] . In mean-field game theory, turnpike results are often proved under the Lasry–Lions monotonicity condition, introduced independently by Lasry–Lions [ 13 ] and Huang, Malhamé, and Caines [ 10 ] . That condition enforces a sign structure on the coupling and excludes many pattern-forming mechanisms. The present paper isolates a complementary mechanism in a translation-invariant setting: the long-time picture is determined by the Fourier spectrum of the interaction kernel K . The mechanism is analogous to Turing’s diffusion-driven instability [ 16 ] . In the present MFG setting, the diffusion coefficient \nu and the coupling strength \gamma compete mode by mode. As long as the spectral gap stays positive, the uniform equilibrium remains the dominant long-time state. When the first Fourier mode reaches the threshold, the linear gap closes, and a pattern-forming bifurcation becomes possible. 1.2. Relation to existing work Cardaliaguet and Porretta [ 2 ] proved exponential turnpike for MFG under strong monotonicity using the master equation. The monograph [ 3 ] establishes convergence of N -player equilibria under similar structural assumptions. Trélat [ 17 ] gave a linear turnpike theorem for optimal control systems in which the long-time rate is controlled by a spectral gap; the earlier work of Porretta–Zuazua [ 14 ] studied long-time versus steady-state optimal control in a related setting. The present paper connects that spectral-gap viewpoint to the Fourier structure of the MFG interaction kernel. The bifurcation of nonuniform stationary states in mean-field models has been studied in the context of McKean–Vlasov equations [ 8 ] and aggregation-diffusion systems [ 5 ] . The Lyapunov–Schmidt and Crandall–Rabinowitz approach [ 6 , 11 ] used here is standard; the point here is to identify the exact spectral condition that drives the transition and to connect it to the turnpike rate. Propagation of chaos for mean-field games was established by Sznitman [ 15 ] and further developed by Lacker [ 12 ] and Cecchin–Delarue [ 4 ] . Quantitative Wasserstein rates for empirical measures depend on dimension [ 9 ] , which is why we restrict to a qualitative convergence statement. 1.3. Standing assumptions and regimes We work on the flat torus \mathbb{T}^{d}=(\mathbb{R}/\mathbb{Z})^{d} , fix \nu>0 , and consider a smooth even kernel K\in C^{\infty}(\mathbb{T}^{d}) with zero mean. The general model allows the possibility that \widehat{K}(\xi)\geq 0\qquad\text{for all }\xi\neq 0, in which case no phase transition occurs. If some negative Fourier coefficient exists, then the critical coupling strength is defined by (1) \gamma_{c}:=\min_{\begin{subarray}{c}\xi\in\mathbb{Z}^{d}\setminus\{0\}\\ \widehat{K}(\xi)<0\end{subarray}}\frac{\nu^{2}(2\pi|\xi|)^{2}}{|\widehat{K}(\xi)|}, with the convention \gamma_{c}=+\infty if no \xi\in\mathbb{Z}^{d}\setminus\{0\} satisfies \widehat{K}(\xi)<0 . The results in this paper fall into three regimes: \displaystyle 0\leq\gamma<\gamma_{c} \displaystyle\text{(subcritical turnpike)}, \displaystyle\gamma=\gamma_{c} \displaystyle\text{(critical regime)}, \displaystyle\gamma>\gamma_{c} \displaystyle\text{(supercritical bifurcation)}. For the critical and supercritical statements, we impose the following nondegeneracy assumptions. Assumption 1.1 (Critical-mode hypothesis) . Assume that \gamma_{c}<\infty and that the critical set \mathcal{C}:=\Bigl\{\xi\in\mathbb{Z}^{d}\setminus\{0\}:\widehat{K}(\xi)<0,\;\frac{\nu^{2}(2\pi|\xi|)^{2}}{|\widehat{K}(\xi)|}=\gamma_{c}\Bigr\} consists of a single pair \{\pm\xi_{0}\} for some \xi_{0}\in\mathbb{Z}^{d}\setminus\{0\} . Assumption 1.2 (Small boundary data) . The terminal cost g\in C^{\infty}(\mathbb{T}^{d}) satisfies \|g-\bar{g}\|_{C^{2,\alpha}(\mathbb{T}^{d})}\leq\varepsilon for some small \varepsilon>0 , where \bar{g}:=\int_{\mathbb{T}^{d}}g\,dx . The initial density m_{0}\in C^{\infty}(\mathbb{T}^{d}) is strictly positive, satisfies \int_{\mathbb{T}^{d}}m_{0}\,dx=1,\qquad\|m_{0}-1\|_{C^{2,\alpha}(\mathbb{T}^{d})}\leq\varepsilon. Assumption 1.3 (Critical cubic reduction) . At \gamma=\gamma_{c} , after fixing the translation phase so that the critical center space is one-dimensional, the reduced critical dynamics admit an odd scalar normal form \dot{a}=-\beta a^{3}+O(a^{5}) for some \beta>0 . Assumption 1.4 (Bifurcation nondegeneracy) . In the phase-fixed stationary problem at \gamma=\gamma_{c} , the reduced scalar bifurcation equation has the form \alpha(\gamma-\gamma_{c})A-\beta A^{3}+R(A,\gamma)=0, where \alpha>0 , \beta>0 , and the remainder satisfies |R(A,\gamma)|\leq C\Bigl(|\gamma-\gamma_{c}|^{2}|A|+|\gamma-\gamma_{c}|\,|A|^{3}+|A|^{5}\Bigr) for (A,\gamma) near (0,\gamma_{c}) . 1.4. Main results The finite-horizon mean-field game system is ( \mathrm{MFG}_{T} ) \begin{cases}-\partial_{t}\phi-\nu\Delta\phi+\dfrac{1}{2}|\nabla\phi|^{2}=\gamma(K*m)(x),\\[4.0pt] \partial_{t}m-\nu\Delta m-\operatorname{div}(m\,\nabla\phi)=0,\\[4.0pt] m(0,\cdot)=m_{0},\qquad\phi(T,\cdot)=g.\end{cases} The stationary (ergodic) system is ( \mathrm{SMFG} ) \begin{cases}-\nu\Delta\bar{\phi}+\dfrac{1}{2}|\nabla\bar{\phi}|^{2}=\gamma(K*\bar{m})(x)-\bar{\lambda},\\[4.0pt] -\nu\Delta\bar{m}-\operatorname{div}(\bar{m}\,\nabla\bar{\phi})=0,\\[4.0pt] \displaystyle\int_{\mathbb{T}^{d}}\bar{m}\,dx=1,\qquad\displaystyle\int_{\mathbb{T}^{d}}\bar{\phi}\,dx=0.\end{cases} The uniform equilibrium is (\bar{\phi},\bar{m},\bar{\lambda})=(0,1,0) . For \xi\in\mathbb{Z}^{d}\setminus\{0\} , set k_{\xi}:=(2\pi|\xi|)^{2} and define (2) \sigma_{\xi}(\gamma):=\nu^{2}k_{\xi}^{2}+\gamma k_{\xi}\widehat{K}(\xi). For \gamma<\gamma_{c} (or for all \gamma\geq 0 if \gamma_{c}=+\infty ), one has \sigma_{\xi}(\gamma)>0 for every \xi\neq 0 , and we define the spectral gap (3) \rho(\gamma):=\min_{\xi\neq 0}\sqrt{\sigma_{\xi}(\gamma)}. Theorem 1.5 (Linear spectral threshold) . Assume the model above. For every \xi\neq 0 , the Fourier mode matrix M_{\xi} satisfies M_{\xi}^{2}=\sigma_{\xi}(\gamma)I. Hence: (i) if \sigma_{\xi}(\gamma)>0 , the eigenvalues of M_{\xi} are \pm\sqrt{\sigma_{\xi}(\gamma)} ; (ii) if \sigma_{\xi}(\gamma)=0 , then 0 is a double eigenvalue and M_{\xi} is nilpotent of index 2 ; (iii) if \sigma_{\xi}(\gamma)<0 , the eigenvalues are purely imaginary \pm i\sqrt{|\sigma_{\xi}(\gamma)|} . The following are equivalent: (i) \sigma_{\xi}(\gamma)>0 for all \xi\neq 0 ; (ii) \gamma<\gamma_{c} ; (iii) the linearization around the uniform equilibrium has no neutral Fourier mode. Under Assumption 1.1 , as \gamma\uparrow\gamma_{c} , \rho(\gamma)=C_{\ast}\sqrt{\gamma_{c}-\gamma}+o\!\bigl(\sqrt{\gamma_{c}-\gamma}\bigr),\qquad C_{\ast}=\sqrt{k_{\xi_{0}}\,|\widehat{K}(\xi_{0})|}. Proposition 1.6 (No phase transition under monotonicity) . If \widehat{K}(\xi)\geq 0 for all \xi\neq 0 , then \gamma_{c}=+\infty and the linearized finite-horizon and stationary systems have no neutral Fourier mode for any \gamma\geq 0 . In particular, there is no

Read the complete text