open-problem

Status: open (2026-10-24). Area: information geometry, message passing · Prerequisites: Fisher metric, belief propagation · Related: Scalar and Multivariate Energy §6, LP-002 The Jacobian side is not implemented, the CT-ML wiki’s Information Geometry of Probabilistic Programs and Natural Gradient.

Sources: Karwowski, Nielsen & Staton, Compositional Information Geometry of Probabilistic Programs, arXiv:2610.17886, 2026 (Proposition 8, Theorem 2, Example 1); Martens, New Insights and Perspectives on the Natural Gradient Method, JMLR 21, 2020 (§9.2); Welling & Teh, Linear Response Algorithms for Approximate Inference in Graphical Models, Neural Computation 16, 2004; the question is original to this vault.

Bibliography: Karwowski et al. 2026 · Martens 2020 · Welling & Teh 2004

Question

For probabilistic programs, Karwowski, Nielsen & Staton prove a chain rule for the Fisher metric of a composite: , the local metric plus the downstream metric pulled back (arXiv:2610.17886, Proposition 8(7)). Is there a local formula of this kind for the Fisher metric in the parameters of a factor graph with loops, computed from the messages, and does natural-gradient descent with it converge to the same points as with the exact metric?

Setting

A factor graph with parameterized factors , data clamped on some variables, and the model evidence . The exact Fisher metric is . A local metric is one assembled from per-factor quantities evaluated at the current messages, e.g. per-factor Gauss–Newton blocks plus coupling blocks.

What is known

  • For first-order probabilistic programs, composed sequentially and in parallel, the chain rule above holds exactly for the presentation-level metric of Karwowski et al., which is lax monoidal and depends on how the model is presented (their Theorem 2 and Example 1). Whether it carries over to an undirected factor graph, even a tree, is part of the question.
  • For a single Gaussian factor with mean the Fisher metric is the Gauss–Newton matrix (Martens, §9.2); Lenticulum keeps available for this (energy_jacobian, by finite differences for now).
  • On loopy graphs nothing is known here.

What would count as an answer

A formula for a local metric on loopy graphs with its error relative to the exact metric (for example in terms of the Bethe approximation error), or a proof that no local metric can be exact; and an experiment on a small loopy linear-Gaussian graph where the exact metric is computable.

Why it matters here

AutoBayes intends parameterized games to be trained by Fisher-metric descent. A local metric is what would make that tractable on graphs, and it is the main payoff of keeping energies as vectors.

Leads

The Bethe Hessian; linear response for belief propagation (Welling & Teh); the presentation dependence found by Karwowski et al., which suggests the answer depends on how factors are grouped.