Particle-Based Gradient Flow Methods

• Particle-based gradient flows are a variational approach that approximates infinite-dimensional gradient flows using interacting particles, preserving the dissipative structure of continuum models.
• The method discretizes the continuum via empirical measures and JKO schemes, ensuring energy dissipation and stability through non-overlapping partitions and finite-dimensional dynamics.
• Applications span variational inference, high-dimensional sampling, and aggregation-diffusion modeling, with convergence and propagation of chaos underpinning its theoretical guarantees.

A particle-based gradient flow is a variational, interacting-particle approach to approximate infinite-dimensional gradient flows of probability measures—especially those defined in Wasserstein geometry and information geometry—by systems of evolving particles or atomic measures. This methodology underlies high-dimensional sampling, variational inference, aggregation-diffusion models, and nonparametric optimization. Particle-based gradient flows preserve the dissipative variational structure of their continuum PDE counterparts while enabling mesh-free, Lagrangian discretization and empirical algorithmic realization.

1. Gradient Flows in Measure Spaces and Particle Discretization

Consider a functional $\mathcal{F}$ on the space $\mathcal{P}_2(\mathbb{R}^d)$ of probability measures with finite second moment, typically combining energy and entropic components. The gradient flow of $\mathcal{F}$ with respect to the $2$-Wasserstein metric is governed by the nonlinear Fokker–Planck equation: $$\partial_t \rho_t = \nabla\cdot[\rho_t D_\mu \mathcal{E}(\cdot, \rho_t)] + \sigma^2 \Delta \rho_t,$$ where $\mathcal{E}$ is the energy part, and $D_\mu \mathcal{E}$ denotes its Wasserstein-intrinsic derivative. For functional forms involving both interaction potentials and entropic regularization,

$$\mathcal{F}(\mu) = \mathcal{E}(\mu) + \sigma^2 \int \mu(x)\log\mu(x)\,dx,$$

these flows unify diffusion, drift, and interaction mechanisms (Monmarché, 18 Oct 2025).

To approximate the evolution by a finite number of particles, one replaces the continuum $\rho_t$ by an empirical measure $\mu_N = \frac{1}{N}\sum_{i=1}^N \delta_{x_i}$. The energy functional is discretized via non-overlapping balls (or local partitions),

$\mathcal{P}_2(\mathbb{R}^d)$0

where $\mathcal{P}_2(\mathbb{R}^d)$1 is the internal energy (e.g., entropy or porous medium), and $\mathcal{P}_2(\mathbb{R}^d)$2 is an interaction kernel (Carrillo et al., 2015, Lei, 7 Jan 2025).

The particle positions $\mathcal{P}_2(\mathbb{R}^d)$3 then evolve according to finite-dimensional ODEs or SDEs that serve as discrete gradient flows of $\mathcal{P}_2(\mathbb{R}^d)$4 in a space equipped with a suitable weighted norm. Time-discrete schemes employ minimizing movements (JKO steps) to ensure energy dissipation and stability (Carrillo et al., 2015, Lee et al., 2023).

2. Mean-Field Interacting Particle Systems and the Langevin Paradigm

For energies with entropic penalization, a prototypical particle-based gradient flow is realized as a stochastic interacting particle system: $\mathcal{P}_2(\mathbb{R}^d)$5 where $\mathcal{P}_2(\mathbb{R}^d)$6 is the total potential stemming from mean-field contributions and confining terms. The law of large numbers and propagation of chaos warrant that, as $\mathcal{P}_2(\mathbb{R}^d)$7, the empirical measure of the particle system converges to the solution of the Fokker–Planck PDE (Monmarché, 18 Oct 2025, Yang et al., 3 Feb 2026).

Crucially, in certain models (e.g., mean-field Curie–Weiss with entropic regularization), the barycenter of the system evolves as a reduced diffusion in low dimension with noise $\mathcal{P}_2(\mathbb{R}^d)$8, permitting direct importation of sharp finite-dimensional metastability and exit time results to the infinite-dimensional setting (Monmarché, 18 Oct 2025).

3. Functional Inequalities and Long-Time Behavior

The convergence rate and stability of particle-based gradient flows are controlled by functional inequalities:

• Polyak–Łojasiewicz (PL) inequalities: For suitable potentials, the free-energy exhibits a PL-type inequality directly analogous to finite-dimensional optimization,

$\mathcal{P}_2(\mathbb{R}^d)$9

where $\mathcal{F}$0 is the Fisher information and $\mathcal{F}$1 the local equilibrium (Monmarché, 18 Oct 2025).

• Log-Sobolev inequalities (LSI): The $\mathcal{F}$2-particle stationary measure satisfies, possibly in a degenerate form,

$\mathcal{F}$3

ensuring exponential or polynomial decay of the free energy and uniform propagation of chaos (Monmarché, 18 Oct 2025).

• Degenerate Łojasiewicz inequalities: At critical points with degenerate Hessians, functional inequalities with sublinear $\mathcal{F}$4 control only polynomial rates, necessitating refined analysis in such regimes.

4. Extensions: Generalized Metrics, Preconditioning, and Nonlinear Energies

Particle-based flows extend naturally to non-Euclidean metrics and generalized energies:

• Generalized Wasserstein flows: The minimizing movement scheme can be formulated with transport costs induced by convex "Young" functions $\mathcal{F}$5, yielding doubly nonlinear diffusion equations and transport maps involving the Legendre dual $\mathcal{F}$6 (Cheng et al., 2023, Lei, 7 Jan 2025).
• Preconditioned and functional flows: Preconditioners (e.g., Mahalanobis mass matrices or neural nets) can be incorporated at the particle level, yielding flows that adapt to anisotropy and ill-conditioning, with convergence rates governed by log-Sobolev constants and spectral properties of the preconditioner (Dong et al., 2022).
• Accelerated and Sobolev flows: Momentum-based (Nesterov-type) accelerations and Sobolev-metric preconditioning have been introduced for improved convergence in sampling and variational inference settings (Stein et al., 30 Mar 2025, Sbalzarini et al., 2014).

5. High-Dimensional and Efficient Algorithms

The scaling of particle-based flows in high dimension poses algorithmic challenges. The Radon–Wasserstein geometry restricts velocity fields to averages over one-dimensional projections, enabling FFT-based evaluation and $\mathcal{F}$7 per-step complexity for large $\mathcal{F}$8 (Hess-Childs et al., 5 Feb 2026). Regularized Radon–Wasserstein (RRW) flows retain well-posedness and convergence to equilibrium with particle-based approximations converging in the mean-field limit.

In time-implicit frameworks (Deep JKO), neural ODE representations of velocity fields parametrize gradient flows in high dimensions, ensuring variational energy dissipation and supporting complex nonlinear and nonlocal mobilities (Lee et al., 2023). These methods leverage continuous normalizing flows for tractable density computations along evolving particle trajectories.

6. Theoretical Guarantees and Convergence

Convergence analysis leverages the $\mathcal{F}$9-convergence of discrete energies, consistency results for curves of maximal slope in measure spaces, and propagation of chaos for interacting particle systems. Well-posedness and uniqueness are established under convexity and regularity assumptions on the energy functionals and interaction kernels (Carrillo et al., 2016, Monmarché, 18 Oct 2025, Amend et al., 26 Mar 2026).

Energy dissipation inequalities and Talagrand-type transport inequalities guarantee decay of entropy functionals and contraction of flows in Wasserstein distance, with explicit rates under strong convexity or under functional inequalities such as LSI (Monmarché, 18 Oct 2025, Yip, 2 Dec 2025).

7. Applications and Benchmarks

Particle-based gradient flows serve as the computational foundation for:

• Variational inference and Bayesian filtering via Fisher–Rao and Wasserstein geometries (Yi et al., 6 May 2025, Dong et al., 2022).
• Sampling in high dimensions (e.g., SVGD, accelerated SVGD, GWG, SIFG), generative models, and robust high-dimensional optimization (Stein et al., 30 Mar 2025, Cheng et al., 2023, Zhang et al., 2024).
• Aggregation-diffusion models, pressureless Euler-alignment systems, and nonlocal mean-field dynamics in mathematical biology and swarming (Yang et al., 3 Feb 2026, Galtung, 2024).
• Total variation and sparsity-regularized optimization in Banach spaces, via atomic flows and Wasserstein-lifted representations (Amend et al., 26 Mar 2026).

Experiments validate qualitative and quantitative convergence, energy decay, metastability predictions, noise-induced regularization effects, algorithmic efficiency, and scalability to thousands of particles and ambient dimensions beyond $2$0 (Carrillo et al., 2015, Hess-Childs et al., 5 Feb 2026).

---

In summary, particle-based gradient flows constitute a flexible and theoretically principled framework for approximating Wasserstein and related gradient flows in probability spaces. This approach preserves the variational dissipation structure, accommodates general energy functionals and geometries, and admits rigorous convergence analysis, making it a central algorithmic and conceptual tool across sampling, inference, PDE modeling, and optimization (Monmarché, 18 Oct 2025, Carrillo et al., 2015, Cheng et al., 2023, Hess-Childs et al., 5 Feb 2026).