Partial differential equations on the space of probability measures arise naturally in mean-field theory. They are typically of transport type or, in the presence of common noise, of second order but highly degenerate. Their well-posedness usually requires regular data. In finite dimension, irregular data can be handled by ``adding a Laplacian''. From a probabilistic point of view, the resulting equation is the Kolmogorov equation of a diffusion process driven by Brownian motion, and is solved by the associated semigroup. It is then natural to look for a Laplacian on the space of probability measures, or at least a weak version of it, and for the diffusion it generates. To this end, a reference measure is needed and, unlike in finite dimension with the Lebesgue measure, there is no canonical choice.
Working on the flat torus, we present one possible choice, given by the so-called Dirichlet-Ferguson measure. Combined with the Wasserstein distance, this measure defines a Wasserstein-Sobolev space (Fornasier, Savaré and Sodini, JFA'23) and an associated Laplacian, which generates a probability-measure-valued stochastic process, the Dirichlet-Ferguson diffusion (Dello Schiavo, AOP'22). We first discuss the link between these objects and then, within this framework, we prove well-posedness of linear transport-diffusion equations driven by this Laplacian, with bounded measurable velocity fields and square-integrable terminal data.
Finally, we introduce an interacting particle system associated with these equations, which involves countably many particles with random masses. This system provides a probabilistic representation of the solution, in the sense of Feynman-Kac.
This is joint work with François Delarue and Giacomo Enrico Sodini.