Hamilton-Jacobi-Bellman Equations in Random Geometries: Homogenization on Continuum Percolation Clusters

Abstract

We develop a quenched homogenization theory for optimal control problems related to Hamilton--Jacobi--Bellman equations on random geometries arising from continuum percolation. The underlying state space is the infinite connected component of a continuum percolation model conditioned to contain the origin. The relevant law of the environment is no longer translation invariant, and the geometry of the state space becomes part of the homogenization problem. The associated controlled diffusion is allowed to degenerate according to the distance to the random boundary of the cluster. The degeneracy regime is determined by a balance between a negative-moment threshold for the distance-to-boundary function of the cluster and the coercivity of the Hamiltonian. We prove that the rescaled value functions converge, locally in $L^p$ on the rescaled random domains, almost surely to a deterministic limit governed by an effective Hamiltonian. The effective Hamiltonian admits dual variational characterizations involving a class of curl-free gradients satisfying an induced mean-zero condition. The resulting effective theory retains information about the continuum percolation geometry, the degeneracy of the diffusion, and the nonstationarity induced by conditioning on the infinite component. The proof introduces a variational framework for homogenization nonstationary conditioned laws. Its main ingredients are random shifts adapted to the geometry of the cluster, a two-step min--max construction for admissible gradients, and a novel relative entropy structure intrinsic to the stochastic control representation. The latter reveals a new connection between relative entropy and effective theories for nonlinear stochastic control problems and applies equally well in the general setting of stationary ergodic random media on ${\mathbb{R}}^d$ and is therefore of independent interest.

Completely revised manuscript, some earlier assumptions corrected; the revised assumptions are now sharp and apply to a substantially broader class of models, large parts of the proofs have been rewritten and are new