Nonparametric maximum likelihood estimation (NPMLE) for Gaussian mixtures finds the mixing distribution that best explains the data without assuming a fixed number of components. It's a natural approach but raises a stability question: does a small change in the data produce a small change in the estimated mixture?
The authors (arXiv:2603.23196) prove stability bounds using tools from statistical mechanics. The KL divergence between NPMLE estimates from different samples is bounded by min{(log n)^(d+2)/n, log n/√n}, covering all regimes of sample size n and dimension d. The bounds also apply to approximate NPMLE solutions — important because exact optimization is often intractable.
The statistical mechanics connection is not metaphorical. The NPMLE optimization landscape has the structure of a random energy model, where the optimal mixing distribution sits in a rugged landscape. The stability results exploit the same techniques used to analyze spin glasses.
The through-claim: the stability of a statistical estimator can be understood through the geometry of its optimization landscape. NPMLE is stable not because the landscape is smooth but because the basins of attraction around good solutions are deep enough relative to the barriers between them. Statistical mechanics provides the natural language for this — the answer comes from the landscape, not from the estimator.