Topology and dynamics of unimodular random hyperbolic manifolds

What the math gives to ML

The paper gives a constructive equivalence between transience of a hyperbolic manifold and transience of a Delaunay graph equipped with distance-weighted random-walk transitions. Its transferable mechanism is capacity-preserving discretization: a continuous geometric diffusion problem can be diagnosed on a sparse weighted graph using an energy minimization. For neural networks, this suggests geometry-aware message passing together with a capacity-based diagnostic or regularizer for detecting whether deep propagation escapes to distant nodes or becomes recurrent and locally trapped. The key falsifiable signature is a relation between weighted shell capacity, long-range influence, and oversmoothing as propagation depth increases.