Ultrametric organization of energy landscapes on random Erdős--Rényi graphs: topological origin of barrier hierarchy

What the math gives to ML

The paper identifies a concrete mechanism by which large energy heterogeneity converts a kinetic landscape metric into a single-linkage ultrametric: paths are dominated by their largest barrier, while suboptimal paths and thermal corrections are exponentially suppressed. This suggests representing neural-network parameter or activation states by a barrier hierarchy rather than a flat Euclidean geometry. A practical transfer is to build an online minimum-spanning-tree hierarchy over low-loss checkpoints or latent states, then use it for hierarchical trust regions, basin-aware replay, or coarse-to-fine optimization. The key falsifiable prediction is that the fraction of triples satisfying the strong triangle inequality should increase monotonically with the imposed loss or barrier spread and approach one as the spread grows.