A network of pairwise connections — each node linked to others through simple two-body interactions — should produce pairwise dynamics. The structure of the interactions should constrain the structure of the behavior. Kuehn and Murphy prove that it doesn't.
Their system is an adaptive network where coupling weights evolve on a fast timescale while node states evolve slowly. Each connection between two nodes adjusts its strength rapidly based on the states of those two nodes. Everything is pairwise: two nodes, one edge, one weight. No triplet interactions are written into the equations.
When the fast dynamics are eliminated through Fenichel reduction — a rigorous technique from geometric singular perturbation theory that replaces the fast variables with their equilibrium values on a slow manifold — the reduced equations for the slow variables contain genuine three-body terms. Triplet interactions that weren't in the microscopic description appear in the macroscopic one.
The result is not an approximation artifact. Kuehn and Murphy derive an explicit irreducibility criterion: a mixed second derivative condition that determines whether the emergent triplet terms can be decomposed back into pairwise interactions through a coordinate change. When the criterion is satisfied, the higher-order structure is genuine — no change of variables can make it pairwise again.
They verify this with the adaptive Kuramoto model, the standard model of coupled oscillators with evolving connections. The reduced dynamics after eliminating the fast coupling weights contain irreducible three-oscillator phase terms that don't exist in the original equations.
Pairwise systems are not closed under timescale reduction. Simplifying the fast dynamics complicates the slow structure. The interaction order isn't a property of the network — it's a property of the resolution at which you observe it.
Note (2026-11-11): checked against the full text of the paper (Kuehn and Murphy, arXiv 2603.19382). The main result is as the essay says: for adaptive phase-oscillator networks with fast coupling weights, Fenichel reduction of the strictly pairwise system gives explicit triplet terms in the slow phase dynamics, and the paper verifies an irreducibility criterion for the adaptive Kuramoto model. Four points differ. (1) The essay says that when the criterion holds 'no change of variables can make it pairwise again'. The paper says the opposite about general coordinate changes: multi-index expressions could be created or removed by general coordinate changes, so it defines pairwise and higher-order structure to be intrinsic to node coordinates and preserved only by 'node-respecting' changes of variables, and its irreducibility result is that the reduced vector field does not admit a pairwise decomposition in node coordinates. It also says the higher-order structure is not produced by coordinate mixing but by the slow-manifold reduction. (2) The essay calls the mixed-second-derivative condition a criterion that 'determines' whether terms can be decomposed. The paper states a sufficient criterion: a non-zero mixed derivative with respect to two other nodes rules out pairwise representability, and it does not claim the converse. (3) The triplet terms are the first-order, O(epsilon) correction to the slow manifold, so the result is for the fast-adaptation regime with small epsilon and the paper derives the first-order correction; the essay's 'macroscopic' description does not say that the terms are small corrections at that order. (4) The essay's closing lines (the interaction order is 'a property of the resolution at which you observe it') are the essay's; the paper's own conclusion is that the class of pairwise-coupled fast-slow adaptive network systems is not closed under slow-manifold reduction.