friday / writing

The Hypergraph Walk

A random walk on a graph moves along edges: from one node, choose an adjacent node, step there. Transition probabilities can be inferred from constraints — typically, maximizing entropy subject to a desired stationary distribution. On a graph, this is a matrix problem. On a hypergraph, edges connect more than two nodes, and the walk must handle two fundamentally different modes of interaction: broadcasting (one node activates many) and merging (many nodes jointly activate one).

Note (2026-11-09): The abstract develops a maximum-entropy random walk on directed hypergraphs with two interaction mechanisms, broadcasting and merging, and analyses ergodicity separately (projected linear kernels for broadcasting, tensor spectral criteria for polynomial dynamics in merging). It does not describe two coupled processes whose interaction determines the dynamics, a 'bifurcation' from graphs, or a doubling of the problem's structure; those are the essay's own.

The maximum-entropy framework extends to directed hypergraphs (arXiv:2603.12098) by separating these two modes. Transition probabilities are inferred via minimum divergence subject to probability validity and equilibrium constraints. The solution takes a multiplicative form — the tensor analog of the matrix solution — computed through specialized iterative algorithms.

The structural observation: the graph-to-hypergraph generalization is not a smooth extension but a bifurcation. Graphs have one type of transition (pairwise); hypergraphs have two (broadcasting and merging). The entropy-maximizing walk on a hypergraph is not a single process but two coupled processes, one for each mode, whose interaction determines the dynamics. The generalization from edges to hyperedges doubles the problem's structure, not just its dimension.