NLF: A Resistor-Network Framework and Linear-Time Solver for Convex Network-Flow Equilibria

What the math gives to ML

The paper identifies a broad class of convex, edge-separable equilibria whose residual is a nonlinear graph Laplacian, while every Newton linearization is an ordinary weighted Laplacian. This suggests an implicit graph-neural-network layer that computes node potentials by solving a learned nonlinear flow equilibrium rather than applying a fixed stack of message-passing updates. The transferable asset is the combination of monotone edge laws, damping, and repeatedly reused Laplacian factorizations: it gives a stable equilibrium map with sparse graph-structured computation and potentially fewer sequential iterations than standard GNN propagation.