model derivation

Assembling everything into AutoBayes Definition 20 and Definition 27. Every abstract slot gets a concrete, proposed mathematical inhabitant. The implemented subset is the polynomial residual, Sampson scalarisation, complete-data fitting and single-branch Gaussian messages. Branch-index inference and the full statistical game remain planned.

Sources: original design and analysis; code: lib/AlgebraicFactors.jl/src/polynomial.jl, polyfit.jl, veronese.jl, lib/AlgebraicFactors.jl/test/polynomial.jl (residual, fitting, local messages, rank and objective checks; no branch game).

Theory (CT-ML wiki): Lax Functor · Bayesian Inversion · Statistical Game · Variational Free Energy

The quadruple

A parameterized statistical game is with . For an algebraic factor of degree on channels with generators:

slotAutoBayesalgebraic realisation
parameter space any space, — The Parameter is a Grassmannian
(update space)tangent bundle (Rem. 30), via
forward kernel root-solve in the chosen polarity
latent space scratchthe branch index — Branches and the Discriminant
inversion root-solve in the opposite polarity, weighted by
vector energy
energy space ours — the residual noise space, Algebraic versus Geometric Distance
scalarisation ours; (algebraic) or (Sampson)
vector entropy over branches, times a direction
gradient couplingDef. 29’s laxnessExactCoupling — the IFT gives the exact Jacobian

The table is a design sketch, not an implemented game contract. PolynomialFactor declares every single-target observed polarity and supplies a local pure PolynomialModel lens with frozen-metric Gaussian/GN messages. It has no branch beliefs, game entropy/free energy, IFT derivatives or ExactCoupling declaration. Non-default couplings, including ExactCoupling, are refused at construction by connect!, Edge and ComposedFactor (LP-011 option 2); per-edge gradient semantics remain planned with L-SPEC-002 Stop Gradient. Its invert is the final site pooled with the prior once. The tests check the linear oracle, circle, improper fibre and a three-channel graph, not this proposed branch game. Its packed pointwise convention is x=nothing, y=z, wrt=:y, differentiating all coordinates; polynomial explains named channels.

The vector energy is forced, not chosen

is -valued by construction. There is no scalar residual replacing all components without losing their Jacobian: k denotes the chosen number of equations, not necessarily the codimension. A variety need not be a complete intersection. So the algebraic family is the existence proof for Scalar and Multivariate Energy: it is a family in which the paper’s -valued cannot be used without first throwing away the model, because

  • the IFT condition for a square system is rank Jᵤ=q=k, a statement about the vector residual’s Jacobian, unstatable for a scalar;
  • the adjoint solves , requiring the Jacobian;
  • the correct scalarisation is derived from the Jacobian.

The energy/entropy split is genuinely two things here

Proposition 18 insists energy and entropy behave differently. In this family the difference is visible:

For the displayed proposed objective the residual term vanishes at exact roots, provided the metric is defined. This alone does not establish the correct free energy of a branch model: priors, branch Jacobians and normalisation may add terms. An overdetermined system can still have an exact root, so k>q alone does not imply nonzero energy. No branch free-energy claim is tested or implemented here.

Composition: local only

By Composition is Elimination the factors must not be fused. The composite game is formed by the Definition 22 laws applied to the local data:

The energy direct sum here is concrete: the composite residual is the stacked residual vector, graded by factor. That is exactly the of Scalar and Multivariate Energy, and in this family it is literally how you would write the joint system down — which is a good sign the grading is the right abstraction rather than an imposition.

The expectation in the entropy law is over branches, so it is a finite sum, not a Monte Carlo estimate. In this family the chain rule is exact.

Proposed alternating learning

From Fitting is a Nullspace Problem §“missing channels”:

  • E-step the inversion : root-find the unobserved channels; obtain the branch belief .
  • M-step descent on : with branches fixed, the energy is quadratic in , so the update is the bottom- eigenspace of — a globally optimal algebraic-distance parameter step.

This spectral step holds for fixed branch weights/completions and the algebraic objective only. Root solving is not generally closed form, and a hard completion is not an expectation under the posterior. The full Sampson fit differentiates its varying covariance and uses an iterative non-convex solver (polyfit). This proposed alternating procedure is not implemented or validated as EM.

Proposed exact gradient coupling away from singularities

Definition 29 is lax in general because moves the pushforward prior and the sampling distribution. Here:

  • the “sampling distribution” is a finite branch set, and its dependence on is differentiable away from the discriminant, with derivative given by the IFT;
  • so the terms Definition 29 drops are computable, and the coupling can be ExactCoupling rather than DiagonalCoupling.

The exception is at the discriminant, where the branch count changes — a discrete jump that no derivative captures. So the honest statement is: exact away from , discontinuous on it. The GradientCoupling annotation should be ExactCoupling with singular-value/rank checks as well as conditioning, if that proposed branch solve and its derivatives are implemented. A scalar condition number alone misses tangency (Branches and the Discriminant).

What this factor cannot do

To be explicit, since the table above is otherwise flattering:

  • , (parameter count and sample complexity);
  • unobserved channels per message (Bézout);
  • coefficients must be fitted, not composed (Composition is Elimination);
  • no guarantee the learned real locus is nonempty or of the intended dimension (Varieties Ideals and Real Nullstellensatz);
  • gradients meaningless near the discriminant.

Related: Factors are Parameterized Statistical Games, Scalar and Multivariate Energy, Branches and the Discriminant, Open Problems in Algebraic Implicit Learning