Convergence in Wasserstein Distance for Empirical Measures of Dirichlet Diffusion Processes on Manifolds

Feng-Yu Wang
DOI: https://doi.org/10.4171/jems/1269
2022-01-01
Journal of the European Mathematical Society
Abstract:Let M be a d-dimensional connected compact Riemannian manifold withboundary ∂ M, let V∈ C^2(M) such that μ( d x):=e^V(x) d x is a probability measure, and let X_t be the diffusionprocess generated by L:=Δ+∇ V with τ:=inf{t≥ 0:X_t∈∂ M}. Consider the empirical measure μ_t:=1/t∫_0^tδ_X_s d s under the condition t<τ for the diffusion process.If d≤ 3, then for any initial distribution not fully supported on ∂M, c∑_m=1^∞2/(λ_m-λ_0)^2≤lim inf_t→∞inf_T≥ t{t 𝔼[𝕎_2(μ_t, μ_0)^2|T<τ]} ≤lim sup_t→∞sup_T≥ t{ t 𝔼[𝕎_2(μ_t, μ_0)^2|T<τ] }≤∑_m=1^∞2/(λ_m-λ_0)^2 holds for someconstant c∈ (0,1] with c=1 when ∂ M is convex, where μ_0:=ϕ_0^2μ for the first Dirichet eigenfunction ϕ_0 of L,{λ_m}_m≥ 0 are the Dirichlet eigenvalues of -L listed in theincreasing order counting multiplicities, and the upper bound is finite if andonly if d≤ 3. When d=4, sup_T≥ t𝔼[𝕎_2(μ_t,μ_0)^2|T<τ] decays in the order t^-1log t, while for d≥5 it behaves like t^-2/d-2, as t→∞.
What problem does this paper attempt to address?