The Bethe-Hessian down to the Percolation Threshold

What the math gives to ML

The paper identifies a sparse-graph spectral operator whose negative eigenspace remains informative down to the percolation regime, where ordinary adjacency or Laplacian embeddings can be dominated by degree noise. The transferable asset is the Bethe correction D-I, which compensates for local tree-like degree fluctuations while preserving a symmetric matrix suitable for stable eigensolvers. A practical neural-network use is to compute a small number of negative Bethe-Hessian eigenvectors as structure-aware node positional features, or to use the corresponding operator as a graph propagation and preconditioning layer. The first implementation should test whether these features improve community-sensitive graph prediction on sparse stochastic block models and real sparse benchmarks at equal parameter count.