Status: open (2026-10-24). Area: applied topology, numerical linear algebra · Prerequisites: cellular sheaves, Gaussian belief propagation · Related: LP-012 Factor graphs and schemes, Loopy Message Passing, The Linear Gaussian Chain.
Sources: Hansen & Ghrist, Toward a Spectral Theory of Cellular Sheaves, J. Appl. Comput. Topol. 3, 2019; Malioutov, Johnson & Willsky, Walk-Sums and Belief Propagation in Gaussian Graphical Models, JMLR 7, 2006; the question is original to this vault.
Bibliography: Hansen & Ghrist 2019 · Malioutov et al. 2006
Question
A linear-Gaussian factor graph can be read as a cellular sheaf: each variable and factor carries a vector space, and each edge a linear map saying how they must agree. What do the sheaf’s Laplacian and cohomology say about Gaussian belief propagation, in particular about its fixed points and about the variances that loopy propagation gets wrong?
Setting
Variables with stalks ; factors with stalks and the factor . The restriction maps are , the coboundary is , and the weighted sheaf Laplacian is . The joint precision of the graph (with priors) is plus the prior precisions.
What is known
- The MAP estimate (equal to the posterior mean) minimises the weighted Dirichlet energy plus the priors: a sheaf-harmonic extension problem. This much is a restatement, not a result.
- Gaussian belief propagation computes exact means on loopy graphs when it converges (walk-summable models), but its variances are wrong in general (Loopy Message Passing).
- Global sections of the sheaf () are the configurations that satisfy every factor exactly (Hansen & Ghrist 2019).
What would count as an answer
A statement relating the variance error of loopy Gaussian belief propagation to a sheaf-theoretic quantity (the first cohomology, the spectrum of , or the cycle structure seen by the sheaf), with a proof or a counterexample; and a check on a small loopy graph with exact variances.
Why it matters here
It would give the CT-ML wiki a Cellular Sheaf note with a concrete payoff, and Lenticulum a diagnostic for when loopy variances can be trusted.
Leads
Walk-sum analysis of Gaussian BP (Malioutov, Johnson & Willsky); sheaf Laplacians and diffusion (Hansen & Ghrist); Robinson’s Topological Signal Processing.