Distributed Power Mean Fusion

The problem of distributed information (data) fusion is considered in this paper. Using a new class of functionals to perform data fusion – the power mean – information fusion can be performed in a dependency/correlation-agnostic manner while achieving better fusion results than many classical fusion techniques (i.e. linear or log-linear pooling). In addition, power mean techniques converge much more rapidly (even in finite time) in distributed network settings. Computing the power mean on generic probability distribution functions is specifically addressed in this paper, including a distributed protocol and specific steps for ensuring distributed convergence to the global power mean. Results demonstrate: 1). convergence to the global power mean even in a distributed setting, 2). the ability for certain power means to converge in finite time, and 3). the improved fusion results achieved by using the power mean over more traditional fusion techniques, even with relatively complex distributions.