Two problems for threshold cascades of interacting diffusions on unimodular random trees: front propagation with a Bramson correction, and the continuous-type limit theory

What the math gives to ML

The paper supplies a multitype branching-process view of cascades in which stability and propagation are governed by the Perron root of a tilted mean operator, rather than by an average branching factor alone. This suggests a controller for dynamically sparse neural networks that treats active tokens, experts, or recurrent states as particles with continuous difficulty or strength types and regulates the dominant reproduction eigenvalue. The transferable asset is the combination of type-conditioned offspring statistics, Perron-eigenvector reweighting, and spectral-gap monitoring, which can prevent both dead routes and exponential activation growth. The most practical first target is a mixture-of-experts or adaptive-depth model with a router whose expected active-child operator is explicitly constrained near a chosen critical radius.