Stability in stochastic hypergraph matching I: necessary and sufficient criteria

What the math gives to ML

The paper isolates a hyperedge-level notion of balance that penalizes every pairwise discrepancy between class counts assigned to the same hyperedge, rather than only comparing adjacent classes in an arbitrary cycle. This is transferable to mixture-of-experts and structured routing systems where an expert bundle receives tokens from multiple classes, domains, modalities, or priority groups. The useful asset is the symmetry of the all-pairs penalty: it is permutation-invariant within each hyperedge and detects imbalances that a cyclic surrogate can miss. A practical adaptation is a differentiable expected-count regularizer applied to router probabilities, with hard online assignment retained at inference.