Approximate Message Passing for Not So Large niid Generalized Linear Models
Zilu Zhao, Fangqing Xiao, Dirk Slock
Abstract
Many signal processing problems involve a Generalized Linear Model (GLM), which is a linear model in which the unknowns may be non-identically independently distributed (n.i.i.d.). Vector Approximate Message Passing (VAMP) is a computationally efficient belief propagation technique used for Bayesian inference. However, the posterior variances obtained from (limited complexity) VAMP are only exact when an independent and identically distributed (i.i.d.) prior is assumed, due to the averaging operations involved. In many problems, it is desirable to not only get estimates of the unknowns but also correct posterior distributions. Whereas VAMP and esp. AMP is applicable to problems of high dimensions, in many applications the dimensions are not very high, allowing for more complex operations. Also, in finite dimensions, the asymptotic regime leading to correct variances under certain measurement matrix model assumptions does not hold. To address these challenges, we propose a revisited version of VAMP, called reVAMP, which provides both a multivariate Gaussian posterior approximation (including inter-parameter correlations) and accurate posterior marginals which only require the extrinsic distributions to become Gaussian.