On the Geometry of Separation in Finite Gaussian Mixtures

A new geometric framework clarifies how the minimum separation between components governs parameter estimation in finite Gaussian mixtures, a long-standing open problem in statistical learning [1]. The work, submitted in 2026, develops novel Hellinger lower bounds that link discrepancies between mixture densities directly to Wasserstein distances between their underlying mixing measures [1]. These bounds explicitly account for both the minimum separation and the minimum component weight [2]. The approach constructs specialized moment-extraction test functions by combining interpolation polynomials with confluent divided difference techniques [1]. Gaussian mixture models represent subpopulations within an overall population without requiring that individual observations be labeled by subpopulation [6]. Each component follows a normal distribution, characterized by a mean and variance [7]. When the number of components is known, the framework reveals a localization phenomenon: the complexity of separation is determined strictly by the spatial arrangement of the components—whether they cluster together, split into groups separated by a macroscopic gap, or lack structural constraints [1]. This finding contrasts with earlier work that often treated minimum separation as a single worst-case threshold. Prior research had established that a separation of order Ω(√log k) suffices for learning mixtures of spherical Gaussians with polynomial samples, while smaller separations can force super-polynomial lower bounds [3]. More recently, a structural measure called the Pair Correlation Factor was introduced to capture geometric difficulty beyond the minimum gap, reflecting interactions among all component means [4]. The new framework also addresses the case where the number of components is unknown and over-specified. Under these conditions, the separation complexity decreases slightly and the minimum mixture weight vanishes from the convergence rates [1]. This occurs because the geometry shifts from first-order to second-order Wasserstein distances [2]. The resulting separation-dependent convergence rates interpolate continuously between point-wise and uniform estimation regimes, establishing fundamental limits for parameter recovery [1]. The study of mixture separation has also been examined through the lens of computational hardness. When mixing weights are allowed to be exponentially small, distinguishing well-separated mixtures from a pure Gaussian can be exponentially hard, but for polynomially lower-bounded weights, quasi-polynomial time algorithms become possible [5].