VAMP with Vector-Valued Diagonalization

Vector approximate message passing is studied where vectorvalued diagonalization instead of a uniform one is employed. Thereby,individualvariancesare tracked within the algorithm instead of an average one. Straightforward application based on the expectation-consistent approximate inference framework does not give satisfactory results. The main reason for this is that the precision parameters may become negativeduring the iterations. In this contribution, improved versions for the update equation ("Onsager correction") are derived from basic estimation principles. Numerical simulations cover the superiority of the new variants.