The paper studies how certain physical systems evolve over time when they follow a specific mathematical rule called a Wasserstein gradient flow. The systems in question involve particles or distributions of mass that interact with each other according to an interaction energy, meaning every bit of mass feels a push or pull from every other bit. The question is: does the system eventually settle into a steady state, and if so, how quickly? The mathematical difficulty is that standard tools for proving convergence require the energy landscape to have a nice bowl-like shape (called convexity), which these interaction energies often lack. The authors sidestep this by exploiting a special structure in the interaction kernel, namely that it can be decomposed into eigenmodes of the Laplacian with all nonnegative coefficients.
The main positive result is that the energy of the system decays to its minimum value at a rate faster than 1/t, written as o(1/t), for a broad class of kernels. This covers practically relevant examples including kernels used in machine learning transformer models, kernels on spheres, and certain kernels defined through fractional versions of the Laplacian. If every mode of the kernel carries positive weight, the mass distribution actually converges to a perfectly uniform spread over the space. Crucially, these flows have no diffusion, so the spreading is driven entirely by the interaction structure rather than by any smoothing mechanism, making the convergence result genuinely surprising from a classical standpoint.
The authors also show that the o(1/t) rate is essentially the best one can hope for in general. For any slightly faster rate of the form t to the power of negative one minus delta, no matter how small delta is, they construct explicit examples where the energy does not decay that fast. One construction works at the level of the linearized equations, and a more technically demanding construction produces an exact solution of the full nonlinear flow on a torus. This sharpness analysis required careful uniform-in-time control over how different frequency components of the solution interact, and it also yields a precise comparison between the nonlinear flow and its linear approximation that holds for all time.