comparison derivation

Three precise connections to algebraic geometry: singular statistical evidence has an RLCT correction, dual-number points encode Jacobian kernels, and linear relations admit a cellular sheaf of consistent states. None makes a general probabilistic factor graph a scheme.

Sources: Watanabe, A Widely Applicable Bayesian Information Criterion (2013), §3 Fundamental Conditions (1)–(4), Theorems 1–2, eqs. (15)–(16), Lemma 3; The Stacks Project, Definition 26.9.1 (tag 01II), Definitions 33.16.1 and 33.16.3, Lemma 33.16.4 (tag 0B28; numbering accessed 2026-10-27); Hansen & Ghrist, Toward a Spectral Theory of Cellular Sheaves (2019 author PDF / arXiv:1808.01513v2), Definitions 2.4–2.5, §2.2.2, Theorem 3.1. The constructions and elementary proofs below are original to this vault; code: lib/AlgebraicFactors.jl/src/toric.jl, lib/AlgebraicFactors.jl/src/plucker.jl, lib/AlgebraicFactors.jl/src/polynomial.jl, lib/Mycelium.jl/src/free_energy.jl (connections, not a scheme or sheaf implementation).

Bibliography: Watanabe 2013 · The Stacks Project Authors 2026 · Hansen & Ghrist 2019

Theory (CT-ML wiki): Information Geometry of Probabilistic Programs · Variational Free Energy

Scope

LP-012 Factor graphs and schemes asked what schemes add to Algebraic Statistics Bridge. A scheme is a locally ringed space locally isomorphic to an affine scheme (Stacks, Definition 26.9.1). Attaching vector spaces to a graph supplies neither these local rings nor affine open charts. The three connections below concern different structures; they do not identify the objects.

Toric Models as Factors, Grassmann Mixing and Plücker Coordinates and L-SPEC-004 Algebraic Factors Toric and Grassmannian already use polynomial equations. Positive probabilities, real roots, a noise metric and a probability normaliser are additional data: the equations alone do not supply them. See Varieties Ideals and Real Nullstellensatz and Branches and the Discriminant.

1. Singular evidence: what replaces parameter counting

Let be independent with true distribution , model and prior density . Write

Watanabe’s Theorem 2, under §3 Fundamental Conditions (1)–(4), gives

Here minimises population cross-entropy, is the real log canonical threshold (RLCT) and its multiplicity. This is not an unrestricted formula for any neural model: the source assumes a compact semianalytic parameter domain, an analytic times smooth positive prior, analytic log-density ratios in for , and its local moment inequality (12). The theorem states a stronger convergence property for the remainder; the displayed is sufficient here.

Theorem 1 resolves the excess population loss to and the prior volume to . Equations (15)–(16) define by the smallest chartwise ratio and by the maximum number of simultaneous minimisers on the essential charts. They depend on the loss and prior near the true distribution, not just the zero set of the model equations. For a regular, realisable model of parameter dimension , with an interior optimum and positive prior density there, Lemma 3 gives , : the leading BIC penalty is recovered.

Elementary singular example (proof). Take and , with uniform prior on . The population KL loss is . The Fisher matrix for is

because the score is . It has rank at most one, and is zero at the origin. Products identify distributions, so different parameter pairs represent the same law. For the deterministic loss integral, symmetry and give

The last order follows by : the coefficient of tends to a positive Gaussian integral, and the integral of is finite. Hence , (also directly from , ). This integral illustrates the local exponents; it is not a replacement for Theorem 2’s random-data statement.

Consequence for Lenticulum. Finding a Factor Graph fits parameters and approximates an integral over states. Watanabe’s integrates parameters. BP over states does not perform this second integral, and a Bethe entropy correction is not an RLCT estimator. Applying singular model selection to learned factors is an expectation, with no RLCT/WBIC implementation or measurement here. A singular equation locus, redundant parameters, and degenerate Fisher information should not be treated as equivalent tests: the example is injective but has Fisher information , which vanishes at zero (differentiate its mean to verify).

2. Dual-number points and forward-mode AD

Write (Stacks Definition 33.16.1). At a -rational point of , tangent vectors are -morphisms reducing to (Definition 33.16.3). For

such a morphism sends to . Substitution proves

This follows by multiplying monomials and discarding all terms containing ; therefore it is a proof for polynomial equations over any field. Stacks Lemma 33.16.4 gives the coordinate-free description by differentials/derivations. Over , forward-mode AD propagates the same multiplication rule, , and the chain rule; one seed computes , rather than the entire Jacobian.

What schemes retain (proof). and have the same underlying point. At zero, the first accepts for every , while the second requires . Passing to the reduced point loses that infinitesimal distinction. Tangent vectors do not promise finite motions: over , the first example requires , which forces over a field. First-order feasibility can be obstructed at higher order.

polynomial uses Jacobians of residual equations, and plucker uses tangent directions on a Grassmannian. These are useful first-order calculations; the code does not construct dual-number points of schemes. In particular, replacing a residual by changes its Jacobian and infinitesimal scheme even if the real zero set is unchanged (the preceding example proves this at zero). A residual presentation and the radical ideal of its zero set must be distinguished before interpreting an AD kernel as geometric tangent data.

3. A cellular sheaf for linear hard relations

Work over and use the bipartite incidence graph as a one-dimensional cell complex. A variable vertex carries . Suppose a factor imposes a homogeneous linear relation ; its vertex carries . The incidence edge carries . The two vertex-to-edge restriction maps are

This is a cellular sheaf of vector spaces in Hansen–Ghrist Definition 2.4: on a graph there are no higher-dimensional incidences requiring additional composition checks. A vertex assignment is a global section (Definition 2.5) precisely when on every incidence. Thus sections correspond bijectively to variable assignments satisfying every relation: each is necessarily the tuple of its neighbouring variables, and that tuple must belong to . This proves the correspondence, including factors with more than two variables.

Choose finite-dimensional inner products and orient each incidence. Section §2.2.2’s coboundary becomes . Define

Taking the inner product proves ; this is the degree-zero case of Hansen–Ghrist Theorem 3.1. The quadratic form measures inconsistency of these state assignments in the chosen metrics. It is not a Bethe free energy and has no entropy or probability normaliser. Affine constraints require translated affine spaces or homogenisation; nonlinear algebraic relations do not automatically give the vector-space sheaf above.

This supplies a precise limited reading of Composition is Elimination: eliminating factor-vertex states leaves exactly the simultaneous linear constraints on the variables (proved by the section correspondence). A Laplacian solver could seek consistency in this fragment. Identifying general Mycelium BP with a global-section algorithm, or identifying belief marginal consistency with this state consistency, remains a conjectural extension. No sheaf solver is implemented here. A sheaf of vector spaces on a cell complex and a scheme’s structure sheaf of rings share the local-to-global language, but Definition 26.9.1 requires much more than this construction provides.

Related: LP-012 Factor graphs and schemes, Finding a Factor Graph, Algebraic Statistics Bridge, Toric Models as Factors, Branches and the Discriminant, L-SPEC-004 Algebraic Factors Toric and Grassmannian.