Detection of first homology via random geometric graphs in the thermodynamic regime

What the math gives to ML

The paper gives a constructive way to recover global first-homology information from only a sparse random geometric graph, without building a dense Cech or Vietoris-Rips complex. Its key asset is the separation between short, local cycles and long cycles: in the supercritical thermodynamic regime, quotienting the graph cycle space by cycles below scale O(r|log r|) recovers H_1(M) with high probability, while any smaller scale fails. This suggests a topology-aware graph neural network preprocessing or pooling module that preserves global cycles while discarding redundant local loops, using only O(n) edges and sparse linear algebra. The most direct experiment is to compare this quotient-derived homology feature or graph coarsening rule against ordinary kNN graphs and persistent-homology features on manifold-valued graph classification tasks.