Direct Message Approximation Improves Probabilistic Inference on Factor Graphs

Direct Message Approximation (DMA) approximates factor-to-variable messages directly to avoid the iterative round-robin scheduling required by expectation propagation. The method eliminates negative-precision messages and provides a formal bound on marginal KL divergence via a new master theorem.