Almost Linear 3-Spanners of Temporal Cliques

What the math gives to ML

The paper gives a constructive way to replace a dense temporally ordered complete communication graph by a nearly linear number of edges while preserving every source-target temporal route within a factor of three in hop count. The transferable asset is not temporal graph theory itself, but the recursive decomposition: certify a large set of ordered source-target pairs using only a linear-size routing skeleton, then recurse only on uncovered pairs. This suggests sparse causal or time-respecting attention masks in which multi-layer message passing emulates dense attention with substantially fewer pairwise interactions. The guarantee is especially relevant when tokens, events, or agents have an intrinsic ordering and communication must respect that order.