Phase transitions in Doi-Onsager, Noisy Transformer, and other multimodal models

Abstract: We study phase transitions for repulsive-attractive mean-field free energies on the circle. For a $\frac{1}{n+1}$-periodic interaction whose Fourier coefficients satisfy a certain decay condition, we prove that the critical coupling strength $K_c$ coincides with the linear stability threshold $K_#$ of the uniform distribution and that the phase transition is continuous, in the sense that the uniform distribution is the unique global minimizer at criticality. The proof is based on a sharp coercivity estimate for the free energy obtained from the constrained Lebedev--Milin inequality. We apply this result to three motivating models for which the exact value of the phase transition and its (dis)continuity in terms of the model parameters was not fully known. For the two-dimensional Doi--Onsager model $W(θ)=-|\sin(2πθ)|$, we prove that the phase transition is continuous at $K_c=K_#=3π/4$. For the noisy transformer model $W_β(θ)=(e^{β\cos(2πθ)}-1)/β$, we identify the sharp threshold $β*$ such that $K_c(β) = K#(β)$ and the phase transition is continuous for $β\leq β*$, while $K_c(β)<K#(β)$ and the phase transition is discontinuous for $β> β*$. We also obtain the corresponding sharp dichotomy for the noisy Hegselmann--Krause model $W{R}(θ) = (R-2π|θ|)_{+}^2$.

• The paper establishes that the critical phase transition point coincides with the linear stability threshold, showing a continuous transition under specific multimodal interactions.
• The study applies its theoretical framework to canonical models—Doi–Onsager, noisy transformer, and Hegselmann–Krause—to derive precise threshold values and transition continuity conditions.
• The analysis leverages a sharp constrained Lebedev–Milin inequality to control the entropy-functional trade-offs, enabling exact characterization of phase transition dynamics.

Phase Transitions in Multimodal Mean-Field Free Energies: Doi-Onsager, Noisy Transformer, and Related Models

Introduction and Scope

The paper "Phase transitions in Doi-Onsager, Noisy Transformer, and other multimodal models" (2604.16288) provides a rigorous analysis of phase transitions in a broad class of mean-field systems governed by repulsive-attractive free energies on the unit circle. The work develops and leverages a sharp coercivity estimate for the mean-field free energy, rooted in the constrained Lebedev–Milin inequality, to exactly characterize the existence and nature (continuity vs. discontinuity) of phase transitions for multimodal interaction potentials.

Three archetypal models serve as motivating and substantive applications: the Doi–Onsager model (classical in soft-matter/statistical physics), the noisy Transformer (a mean-field surrogate for self-attention), and the noisy Hegselmann-Krause (opinion dynamics/bounded-confidence interactions). The paper resolves several previously open or incompletely characterized aspects of these models, most notably the precise values and continuity properties of their phase transition thresholds.

Theoretical Framework

The central object of study is the mean-field free energy functional

$$\mathcal{F}_K(q) = \int_T \log q \, dq(\theta) - K\iint_{T \times T} W(\theta - \theta')\, dq(\theta) dq(\theta'),$$

defined over probability measures $q \in \mathcal{P}(T)$ on the circle $T$. Here, $K \ge 0$ is the coupling strength and $W$ is a real-valued, even, $L^2$ kernel (with further regularity/decay, as made precise in the main theorem). The central dynamical interpretation proceeds from the McKean–Vlasov (MV) equation: $$\partial_t q_t = \frac{1}{2} \partial_\theta^2 q_t - K\partial_\theta\big(q_t \partial_\theta(W * q_t)\big),$$ viewed as the 2-Wasserstein gradient flow of $\mathcal{F}_K$.

A phase transition is said to occur at an interaction strength $K_c$ if, as $K$ increases past $q \in \mathcal{P}(T)$0, the set of global minimizers of $q \in \mathcal{P}(T)$1 bifurcates from uniqueness (the uniform state $q \in \mathcal{P}(T)$2) to nonuniqueness (emergence of non-uniform, symmetry-broken states). Of particular interest is the relationship between $q \in \mathcal{P}(T)$3 and the linear stability threshold $q \in \mathcal{P}(T)$4—the value at which $q \in \mathcal{P}(T)$5 loses local stability.

Main Results

General Theorem: Characterization of Continuous Phase Transitions

For a wide class of $q \in \mathcal{P}(T)$6-periodic interaction potentials $q \in \mathcal{P}(T)$7 with Fourier coefficients satisfying

$q \in \mathcal{P}(T)$8

and normalized so that $q \in \mathcal{P}(T)$9, it is proven that:

• The critical phase transition point $T$0 coincides with the linear stability threshold $T$1, both exactly equal to 1.
• The phase transition is continuous: at $T$2, the uniform distribution $T$3 remains the unique global minimizer.
• For $T$4, $T$5 is the unique critical point of $T$6.

The proof exploits a sharp dual form of the Lebedev–Milin inequality, furnishing an optimal relationship between the entropy and the $T$7 seminorm of $T$8. This functional analytic input enables coercivity of the free energy and strong control on the set of critical points.

Applications to Canonical Models

Doi–Onsager Model

For the two-dimensional setting with $T$9, the analysis yields:

• $K \ge 0$0 exactly.
• The phase transition is continuous.

This strengthens previous lower and upper bounds in the literature, settling the precise value and the nature of the transition in $K \ge 0$1.

Noisy Transformer Model

For the model $K \ge 0$2, the paper establishes a sharp bifurcation at a distinguished value $K \ge 0$3, defined as the solution to $K \ge 0$4:

• For $K \ge 0$5, $K \ge 0$6 and the transition is continuous.
• For $K \ge 0$7, $K \ge 0$8 and the transition becomes discontinuous.

This closes conjectural gaps from empirical and analytical studies in prior art, identifying the exact point of tricriticality.

Noisy Hegselmann–Krause Model

For $K \ge 0$9 and $W$0 (the solution to $W$1):

• For $W$2, $W$3 and the transition is continuous.
• For $W$4, $W$5 and the transition is discontinuous.

This generalizes known results from the small $W$6 regime to the full range of bounded confidence.

Technical Insights

A distinctive technical achievement of the paper is the application of the constrained Lebedev–Milin inequality and its dual, quantifying entropy-function regularity tradeoffs in the mean-field context. This allows for an exact minimax characterization of phase transition thresholds in the presence of multimodal interactions—which typically elude reduction to the unimodal (classic Kuramoto/HMF) case.

Additionally, the analysis makes minimalistic or no quantitative assumptions on the repulsive part of the potential. The methodology therefore captures essentially all multimodal (periodic) interactions with dominating attractive Fourier coefficients, without requiring smallness or decay on multimodal repulsive contributions.

In the limit case of singular interactions (e.g., circular log gas), the boundedness-from-below constraint on the free energy breaks down for $W$7, providing a regime where no minimizer exists but nontrivial critical points may still be fully characterized.

Implications and Future Directions

The established relationship between linear stability analysis and global minimizer dynamics provides a powerful framework for interpreting “order–disorder” transitions in complex, high-dimensional mean-field systems. The sharp demarcation of continuous versus discontinuous transitions has direct implications for nonequilibrium behavior and the long-time selection of macroscopic states under the associated gradient flows.

The results for the noisy transformer model illuminate the limits of mean-field synchronization in self-attention dynamics—indicating parameter regimes where symmetry-broken states emerge either smoothly or in a genuinely first-order manner. In the Hegselmann–Krause model, the complete phase diagram for $W$8 equips the analysis of opinion polarization and stability with precise thresholds.

From a technical perspective, a gap remains regarding uniqueness of critical points in the interval $W$9: the approach provides uniqueness for $L^2$0 and nonuniqueness for $L^2$1, but the intermediate regime awaits further structural or variational insight.

Regarding future research, two directions are natural:

• Extension to Higher Dimensions and Singular Potentials: The approach should be adapted and extended to handle non-circular geometry and more singular, possibly non-smooth, interactions.
• Detailed Long-Time Asymptotics for the McKean–Vlasov Dynamics: While the paper addresses steady-state selection, refined rates of convergence to equilibrium in both continuous and discontinuous cases, especially beyond unimodal (purely attractive) potentials, remain partially open and would benefit from the tools introduced here.

Conclusion

This work rigorously determines the thresholds and character of phase transitions for a class of multimodal, repulsive-attractive mean-field systems on the circle. Through sharp coercivity and entropy-functional inequalities, it provides complete solutions to the stationary and variational problems associated with the Doi–Onsager, noisy transformer, and noisy Hegselmann–Krause models. These results clarify the demarcation between continuous (second-order) and discontinuous (first-order) transitions in complex mean-field systems, setting benchmarks for both mathematical analysis and the theoretical understanding of synchronization, alignment, and consensus in statistical physics, artificial intelligence, and social dynamics.