I. Introduction
After the discovery of Turbo codes and their powerful iterative decoders, the systematic design of iterative receiver structures which include other components beyond the decoder has attracted a lot of research interest. One of the most prominent techniques is Belief Propagation (BP), the theoretical foundation of Turbo and LDPC decoders [1], which is typically presented as a message passing scheme over a graphical model of the underlying probability network [2]. However, receiver tasks like synchronization and channel estimation mainly involve continuous parameters, for which BP is not well suited due to analytically intractable integrals. While this problem can be circumvented by discretizing the continuous random variables [3], this approach is quite ad hoc and, more importantly, results in a rather high complexity for sufficiently fine quantizations.