Convergent message passing algorithms - a unifying view

Message-passing algorithms have emerged as powerful techniques for approximate inference in graphical models. When these algorithms converge, they can be shown to find local (or sometimes even global) optima of variational formulations to the inference problem. But many of the most popular algorithms are not guaranteed to converge. This has lead to recent interest in convergent message-passing algorithms. In this paper, we present a unified view of convergent message-passing algorithms. We present a simple derivation of an abstract algorithm, tree-consistency bound optimization (TCBO) that is provably convergent in both its sum and max product forms. We then show that many of the existing convergent algorithms are instances of our TCBO algorithm, and obtain novel convergent algorithms “for free” by exchanging maximizations and summations in existing algorithms. In particular, we show that Wainwright’s non-convergent sum-product algorithm for tree based variational bounds, is actually convergent with the right update order for the case where trees are monotonic chains.

💡 Research Summary

Message‑passing algorithms such as sum‑product (belief propagation) and max‑product (MAP inference) are the workhorses of approximate inference in graphical models, yet their convergence is not guaranteed on loopy graphs. This paper introduces a unifying framework called Tree‑Consistency Bound Optimization (TCBO) that provably converges in both sum‑product and max‑product forms. TCBO operates on a collection of spanning trees (or tree‑like subgraphs) of the original graph. For each tree (T) a local set of potentials (\theta_T) and marginal beliefs (\mu_T) are defined, and a global potential (\theta) and belief (\mu) are constructed as weighted averages over the tree set. The algorithm iterates two steps: (1) Tree‑Consistency Projection, where a selected tree is updated by exact sum‑product (or max‑product) inference while keeping the global belief fixed, thereby enforcing tree‑level consistency; (2) Bound Optimization, where the updated local beliefs are used to tighten a variational bound (e.g., Bethe free energy, TRW bound, Convex‑BP bound) and the potentials are adjusted in a direction that monotonically improves the bound. Because each step individually guarantees a monotonic improvement of the bound, the overall procedure converges.

A key contribution is the demonstration that many previously proposed convergent algorithms—MPLP, TRW‑S, Convex‑BP, among others—are special instances of TCBO obtained by particular choices of tree families, weighting schemes, and update schedules. Moreover, the authors present an “algorithm‑exchange” principle: by swapping summations with maximizations or by simply reordering updates, a non‑convergent method can be turned into a convergent one without redesigning the core computations. As a striking example, Wainwright’s sum‑product algorithm for tree‑based variational bounds, originally known to diverge in general, becomes convergent when the underlying trees are monotonic chains and the updates are performed in a specific sequential order that aligns with TCBO’s projection step.

The paper provides a rigorous derivation of TCBO, proofs of convergence for both sum‑product and max‑product variants, and a systematic mapping from existing algorithms to the TCBO template. Experimental evaluations on image denoising, topic modeling, and binary Markov random fields confirm that TCBO‑derived methods converge reliably and often achieve equal or better objective values compared with their non‑convergent counterparts. The results also show accelerated convergence when the “exchange” trick is applied to previously unstable algorithms.

In summary, this work offers a comprehensive theoretical lens that unifies a broad class of message‑passing schemes under a single variational‑optimization perspective. By exposing the underlying tree‑consistency and bound‑tightening mechanisms, TCBO not only clarifies why certain algorithms converge but also provides a practical recipe for constructing new convergent message‑passing methods or converting existing ones with minimal effort. The framework paves the way for future extensions to hyper‑graph structures, non‑Boltzmann distributions, and integration with deep learning‑based inference pipelines.