Gaussian Mixture Reduction via Clustering

Recursive processing of Gaussian mixture functions inevitably leads to a large number of mixture components. In order to keep the computational complexity at a feasible level, the number of their components has to be reduced periodically. There already exists a variety of algorithms for this purpose, bottom-up and top-down approaches, methods that take the global structure of the mixture into account or that work locally and consider few mixture components at the same time. The mixture reduction algorithm presented in this paper can be categorized as global top-down approach. It takes a clustering algorithm originating from the field of theoretical computer science and adapts it for the problem of Gaussian mixture reduction. The achieved results are on the same scale as the results of the current “state-of-the-art” algorithm PGMR, but, depending on the input size, the whole procedure performs significantly faster.