Formation of clusters and coarsening in weakly interacting diffusions

1. Problem Statement

The paper investigates the dynamical behavior of weakly interacting diffusions on a one-dimensional torus ($T \simeq [0,1)$) subject to localized attractive interaction potentials. The system is modeled by a set of $N$ stochastic differential equations (SDEs) where particles interact via a potential $W_{\gamma, \ell}$ and are driven by Brownian motion.

The core problem addresses the discrepancy between the microscopic particle system and its mean-field limit (the McKean–Vlasov PDE) regarding the long-term evolution of clustered states. Specifically:

• Phenomenon: Particles initially spread out quickly form clusters due to attraction.
• The Gap: While the mean-field PDE predicts that the system converges to a stationary single-cluster state (or uniform state), it fails to capture the transient dynamics of how multiple clusters merge (coalesce) or exchange mass over time.
• Key Question: How do clusters evolve, merge, or dissolve in the particle system versus the mean-field limit, and what are the associated timescales?

2. Methodology

The authors employ a multi-faceted approach combining rigorous analysis, heuristic derivations, and numerical simulations:

• Linear Stability Analysis: They analyze the linearization of the McKean–Vlasov PDE around the uniform state to determine the onset of clustering (instability) and the number of initial clusters formed.
• Variational Analysis & Riesz Rearrangement: To characterize the stationary states of the free energy, they prove that global minimizers are either uniform or single-cluster states (symmetrically decreasing). This relies on a new variant of the strict Riesz rearrangement inequality applicable to compactly supported kernels (unlike previous results requiring globally strictly decreasing kernels).
• Asymptotic Analysis (Laplace's Method): They derive the timescales for particle escape from clusters using the Eyring–Kramers law, approximating the effective potential near cluster centers as quadratic wells.
• Model Reduction: They derive a coarse-grained model (a system of ODEs and SDEs) describing the dynamics of cluster centers and masses. This model separates the dynamics into:
  1. Coalescence: Brownian motion of cluster centers leading to collisions.
  2. Mass Exchange: Diffusive leakage of particles between clusters driven by thermal fluctuations.
• Numerical Experiments: They validate their theoretical models and timescale predictions using Scharfetter–Gummel schemes for the PDE and Euler–Maruyama schemes for the particle system.

3. Key Contributions

A. Characterization of Stationary States

• Phase Transition: They prove the existence of a discontinuous phase transition at a critical interaction strength $\gamma_c$. For $\gamma > \gamma_c$, the free energy has non-trivial global minimizers (single clusters) distinct from the uniform state.
• Symmetry of Minimizers: They prove (Theorem 3.4) that all global minimizers of the free energy are symmetrically decreasing (single-cluster states). This is achieved by extending the strict Riesz rearrangement inequality to kernels that are only locally strictly decreasing.
• Gaussian Approximation: For large $\gamma$, they show that single-cluster stationary states are approximately Gaussian with variance $\sigma^2 \approx \frac{\ell}{\gamma w''(0)}$.

B. Timescale Analysis and Dynamical Regimes

The paper identifies four distinct dynamical phenomena with vastly different timescales:

1. Initial Clustering ($t_{clustering}$):
   • Driven by linear instability of the uniform state.
   • Timescale: $O(1)$ for generic initial data; $O(\log N)$ if starting from a perfectly uniform distribution (particle system).
2. Coalescence ($t_{coalescence}$):
   • Clusters move as "heavy Brownian motions" and merge upon contact.
   • Timescale: $O(N)$.
   • Crucial Finding: This mechanism is absent in the mean-field limit ($N \to \infty$) because the timescale diverges.
3. Mass Exchange / Leakage ($t_{metastable}$):
   • Particles escape clusters due to Brownian excursions over the potential barrier.
   • Timescale: $O(e^{\gamma \ell \Delta m})$, where $m$ is the cluster mass and $\Delta$ is the depth of the potential well.
   • Crucial Finding: This mechanism dominates in the mean-field limit. The PDE exhibits dynamical metastability: multi-cluster states persist for exponentially long times before one cluster absorbs the others via mass transfer, rather than physical merging.
4. Microscopic Reversibility ($t_{reversibility}$):
   • The spontaneous dissolution of a single cluster back to a uniform state (large deviation event).
   • Timescale: $O(e^{N \Delta F})$. This is absent in the mean-field limit.

C. The Reduced Model

The authors propose a reduced model (Equation 4.3) for the cluster dynamics:

• Centers: $dX_t^{(j)} = \sqrt{\frac{2}{N m^{(j)}}} dW_t^{(j)}$ (Brownian motion).
• Masses: An ODE system describing mass exchange rates $\phi$ between neighbors, derived from Eyring–Kramers escape rates.
• Limit Behavior: As $N \to \infty$, the center motion freezes ($dX \to 0$), leaving only the mass exchange ODE. This explains why the PDE cannot describe cluster merging (coalescence) but only mass transfer.

4. Key Results

• Discontinuous Phase Transition: The system undergoes a first-order phase transition where the uniform state loses stability, and stable single-cluster states emerge.
• Failure of Mean-Field for Coarsening: The mean-field PDE fails to capture the coalescence of clusters (merging of distinct peaks). In the PDE, clusters are stationary in position; coarsening occurs solely through mass exchange (one cluster grows while the other shrinks).
• Dynamical Metastability: The PDE exhibits metastability where a multi-cluster configuration remains stable for timescales exponential in the cluster mass ($e^{\gamma \ell \Delta m}$) before collapsing to a single cluster.
• Scaling Regimes:
  • If $\log N \ll \gamma \ell \Delta$: Coalescence (merging) dominates.
  • If $\log N \gg \gamma \ell \Delta$: Mass exchange dominates.
• Conjecture on Massive Arratia Flow: In a specific scaling limit where $N \to \infty$, $\ell \to 0$, and $\gamma \ell \gg \log N$, the particle system is conjectured to converge to the Modified Massive Arratia Flow, a process of coalescing Brownian motions with masses.

5. Significance

• Theoretical Rigor: The paper provides a rigorous mathematical justification for the "failure" of mean-field descriptions to capture certain transient phenomena (coalescence) in interacting particle systems, highlighting the importance of noise-induced metastability.
• New Inequality: The development of a strict Riesz rearrangement inequality for locally monotone kernels is a significant contribution to variational analysis and statistical mechanics.
• Modeling Insight: The derived reduced model bridges the gap between microscopic particle dynamics and macroscopic PDEs, offering a framework to understand Ostwald ripening and metastability in systems with localized interactions.
• Applications: The findings are relevant to understanding collective behavior in swarming, opinion dynamics, and the sampling properties of algorithms (e.g., in machine learning and optimization) where metastable states can trap dynamics.

In summary, the paper demonstrates that while the mean-field limit correctly predicts the final state of the system, it fundamentally misrepresents the pathway to that state in the presence of localized attractive potentials, missing the crucial mechanism of cluster coalescence which is only visible at the particle level or in finite-$N$ regimes.