open-problem

Status: resolved within the agreed theory-note scope (2026-10-27; maintainer decision 2026-10-26). Original question (2026-10-22, Daniel: “Also is there a connection to schemes from algebraic geometry? I do not know much about that.”). Related: Factor Graphs and Schemes, Algebraic Statistics Bridge, Varieties Ideals and Real Nullstellensatz, Branches and the Discriminant, Composition is Elimination, Bethe Free Energy.

Question. Is there a connection between Lenticulum’s factor graphs and schemes?

What is already in the vault. Algebraic Statistics Bridge shows that conditional independence is a polynomial (determinantal) constraint, so a discrete or Gaussian graphical model is an algebraic variety, and marginalisation is elimination. Schemes generalise varieties, so the question is what the generalisation adds. Three candidate links, none worked out here:

  1. Singular models. Where a model’s variety is singular, parameters are not identifiable and the Fisher metric degenerates; that is where model selection gets hard. Watanabe’s singular learning theory (Algebraic Geometry and Statistical Learning Theory, 2009) handles this with resolution of singularities, and the real log canonical threshold replaces the parameter count in the evidence. Lenticulum’s evidence and Bethe-criterion comparisons (Dynamic Graphs §4) would meet exactly this when learned factors are overparametrised.
  2. Tangent vectors are dual-number points. In scheme language the tangent space of at a point is the set of maps : solutions of the defining equations over the dual numbers. That is forward-mode automatic differentiation, and the “jets” of the CT-ML wiki’s Information Geometry of Probabilistic Programs. Schemes keep this first-order information even where the variety is singular, which is where the Fisher metric of Scalar and Multivariate Energy §6 is degenerate.
  3. Sheaves. A scheme is built by gluing local pieces with a sheaf of functions. A factor graph can be read as a cellular sheaf: every variable and factor carries a space of local states, every edge a map saying how they must agree; global sections are the consistent assignments, and the sheaf Laplacian measures inconsistency (Hansen & Ghrist, Toward a spectral theory of cellular sheaves, 2019). Message passing would then be a way to compute (approximate) global sections. Sheaves are the ingredient factor graphs and schemes share.

Leaning. Link 3 is the closest to the code and would give the CT-ML wiki a Cellular Sheaf note; link 1 is the one with consequences for model selection. Both need reading before a vault note; the references above are to be checked and added to refs.bib first.

Factor Graphs and Schemes works out the three links with exact source definition/theorem numbers and elementary proofs: Watanabe’s parameter-evidence expansion and an RLCT example; dual-number points and the Jacobian kernel (including a nonreduced point); and a cellular sheaf for homogeneous linear factor relations, whose Laplacian kernel is exactly the space of consistent state assignments. It links the toric/Plücker implementation notes, L-SPEC-004 Algebraic Factors Toric and Grassmannian and Finding a Factor Graph.

Qualifications to the original questions above: singular equation loci, nonidentifiability and Fisher degeneracy are not equivalent conditions; the note gives an injective model with a degenerate Fisher matrix. Watanabe’s evidence integrates parameters, whereas the worked Bethe example integrates states at fitted parameters. A cellular vector-space sheaf does not supply a scheme’s sheaf of local rings. The proved state-section correspondence applies to linear hard constraints, not general probabilistic beliefs. Identifying general BP with section computation remains a conjectural extension.

The theory note fulfils the agreed scope; no scheme objects, RLCT estimator, sheaf solver or general nonlinear sheaf construction is implemented. The CT-ML wiki is not edited.