Nonnegative Bakry--Émery Curvature on Bounded-Degree Graphs Implies Volume Doubling and Poincaré Inequalities

What the math gives to ML

The paper supplies a local second-order curvature condition for graphs: nonnegative Bakry–Émery curvature is expressed through the carré du champ operator and its iterate, and implies global volume doubling and scale-invariant Poincaré control. The transferable asset is a computable certificate that local graph diffusion does not create excessively thin bottlenecks, which are a major source of oversquashing and unstable message passing. A practical adaptation is to regularize a learned graph or graph-convolution layer using sampled violations of \(\Gamma_2\geq 0\), while monitoring whether this improves long-range prediction without simply densifying the graph.