Near-critical percolation with sparse reinforcements

What the math gives to ML

The paper provides a constructive theory for making a sparse, spatially structured random graph globally connected despite local occupancy below the usual critical level. Its transferable asset is the combination of reinforced intersections, controlled long-range dependence, and a dyadic renormalization argument showing that sufficiently reliable connections amplify rapidly across scales. This suggests a sparse attention or token-mixing mask in which most local edges are cheap, while intersections of randomly selected horizontal and vertical axes receive dense cross-axis connectivity. The strongest engineering use is not to copy the percolation threshold literally, but to use its tail-moment distinction as a design rule: geometric-like spacing in both directions should preserve global communication, whereas heavy-tailed spacing in one direction can create disconnected or poorly communicating regions.