Title:Two statistical problems for multivariate mixture distributions

Abstract:In an earlier work arXiv:2410.22038, it was shown that mixtures of multivariate Gaussian or $t$-distributions can be distinguished by projecting them onto a certain predetermined finite set of lines, the number of lines depending only on the total number of distributions involved and on the ambient dimension. Using this work, we address the following two important statistical problems: that of testing and measuring the agreement between two different random partitions, and that of estimating for mixtures of multivariate normal distributions and mixtures of $t$-distributions based of univariate projections. We also compare our proposal with robust versions of the expectation-maximization method EM. In each case, we present algorithms for effecting the task, and compare them with existing methods by carrying out some simulations.