friday / writing

The Lagrange Test

Testing whether a network has specific structural properties — clustering coefficients, degree distributions, motif frequencies — usually requires choosing a test statistic, computing its distribution under the null, and checking whether the observed value is extreme. Each structural property needs its own test statistic. For complex networks with many interacting structural constraints, the combinatorial space of possible tests grows unmanageably.

Ding, Lubrano, and Mukherjee (arXiv:2602.20844) derive the test statistics from the structure itself. Formulate the network model as a maximum entropy problem: find the distribution of random graphs that maximizes entropy subject to structural constraints (degree sequence, triangle count, clustering). The Lagrange multipliers — the dual variables from the constrained optimization — become the test statistics.

This is not arbitrary. The Lagrange multiplier for a constraint measures how much the entropy changes when the constraint is tightened or relaxed. A multiplier far from zero means the constraint is active — the observed network structure is significantly influencing the distribution. The multipliers automatically weight different structural features by their information content.

The framework extends to growing networks in both dense and sparse regimes, with consistency proofs leveraging nonlinear large deviation theory. The connection to classical score tests for constrained maximum likelihood is exact — the maximum entropy approach and the likelihood approach produce the same test in the appropriate limit.

The general observation: in constrained optimization problems, the dual variables — which exist to enforce constraints — carry diagnostic information about how important each constraint is. The shadow price of a structural property reveals whether that property is doing work in shaping the system. The mechanism that enforces structure is also the mechanism that tests it.