The most satisfying concrete grounding in this note set: for an algebraic factor, the abstract latent space of Definition 1 is the set of solution branches in the proposed multi-branch model. Target-rank loss obstructs the root IFT; it need not make the full Sampson covariance singular.
Sources: original design and analysis; code:
lib/AlgebraicFactors.jl/src/polynomial.jl,lib/AlgebraicFactors.jl/test/polynomial.jl(circle tangency and scalar conditioning checks),lib/AlgebraicFactors.jl/ext/AlgebraicFactorsHomotopyContinuationExt.jl,lib/AlgebraicFactors.jl/test/homotopy.jl(square isolated roots, mixtures and entropy bounds); Breiding & Timme, HomotopyContinuation.jl: A Package for Homotopy Continuation in Julia (2018); Breiding, Rose & Timme, Certifying Zeros of Polynomial Systems Using Interval Arithmetic (2023), Theorem 3.6 and Lemma 7.1.Bibliography: Breiding & Timme 2018 · Breiding et al. 2023
Theory (CT-ML wiki): Bayesian Inversion · Statistical Game · Open Model
The projection is a branched cover
Fix a polarity. The projection
can be read locally as a covering map of degree D when the fibres are finite, the total locus is smooth, the projection has no critical points and roots do not escape to infinity. These assumptions are needed; an arbitrary polynomial relation need not be a finite proper cover. Under them the real root count is constant on each component away from critical values.
So an algebraic factor’s forward kernel really does have the type of Definition 1,
and the latent variable has a meaning: which solution you are on. Elbow up or elbow down. Which of the ten essential matrices. This is not a bookkeeping device; it is the physically meaningful hidden state.
Everything the vault says abstractly about now has a concrete reading:
| abstract (Open Models and Latent Channels, Composition of Open Models) | algebraic reading |
|---|---|
| latent space | the set of branches over |
reveal — a free retyping | tell the consumer which root was taken |
| marginalising | sum/average over all real roots |
| composition files into the latent | the intermediate branch choice is remembered |
| a pure model, | : the relation is a function in this polarity |
That last row is worth pausing on. A factor is a function exactly when its branch count is one. Explicit learning is the special case , and the entire “implicit” story is about what happens when . Implicit Learners’s table row “may be multi-valued” is the statement ; “or have no solution” is .
The discriminant
is not constant. It changes across the discriminant locus
— the set of observed values over which two branches collide. For suitable finite square systems elimination/discriminant polynomials describe critical values (often a hypersurface); this is not a universal assertion that one resultant describes every real projection’s bifurcations. Roots escaping to infinity can change the count without belonging to this finite critical locus.
What target-rank loss does and does not imply
The following effects require different assumptions; they do not all coincide:
- The implicit function theorem fails. is singular, so is undefined and the adjoint solve of Backpropagation by the Implicit Function Theorem is ill-conditioned. For the circle, |dy/dx|=|x|/√(1−x²) grows like the inverse square root of the distance to x=±1, not inverse distance. Other degeneracies have other rates.
- Inference is discontinuous. Two real roots merge and become a complex-conjugate pair. Smooth separate branch labellings fail at collision. A root set can nevertheless approach the merged root continuously from the side with real solutions; set-valued continuity needs a specified topology.
- The entropy jumps. With a uniform belief over branches, , which is a step function of . The statistical game’s regulariser is genuinely discontinuous here.
- The target site can flatten while the full Sampson metric stays finite. Full J and target Jᵤ are different matrices. On the unit circle, with Σ_f=δI, JΣ_fJᵀ=4δ on the entire circle, including (±1,0). The y site precision is Jᵤ²/A=y²/δ and tends to zero, not infinity. It is loss of full residual rank (or covariance support), not target rank alone, that makes the undamped residual covariance singular. The implemented site is finite and flat at tangency; singular A still throws (polynomial).
This is intrinsic, not a defect
Take the circle: is genuinely undefined for and genuinely two-valued for , and at . No parametrisation, solver or regulariser removes this, because it is a property of the relation, not of the representation. Any implicit learner that permits multi-valued relations inherits it.
Detect target-rank loss and inspect singular values, rather than assume the condition number alone detects proximity. For a nonzero scalar Jᵤ, κ(Jᵤ)=1 even as it tends to zero. The circle tests check that fact and the square-root derivative rate. No IFT gradient is returned by PolynomialFactor.
The optional §4 solver now returns Gaussian mixtures over regular real isolated
roots of square systems (homotopy). The default PolynomialModel remains a
pure local branch approximation. Neither exposes a joint latent branch-index
API; mixture components and weights can be inspected explicitly.
The branch belief
For a square system with regular isolated real roots aⱼ, the residual-density convention of L-SPEC-004 §4 yields a Gaussian mixture of local branch posteriors. With π=N(μ,P), Sⱼ=Jᵤⱼ⁻¹AⱼJᵤⱼ⁻ᵀ,
For a trivial prior the density is flat in target Lebesgue measure and masses are proportional to 1/|det Jᵤⱼ|. Prior-only root weighting would specify a different discrete selection measure. In particular, maximising the branch mass generally includes a Jacobian term and is not merely projection onto the root set by maximising π. One may separately define that constrained prox, but it is not the implemented branch weighting rule.
invert returns MixtureBelief of the branch posteriors. Likelihood messages
exclude the prior: their normalised branch-site masses are proportional to
1/|det Jᵤⱼ|; pooling supplies the Gaussian overlap. Singular roots, infinity
and failed paths are diagnostic, never weighted as regular roots. No regular
real root makes inference throw; polynomial_branches remains inspectable.
The label entropy H(w)=−Σwⱼlog wⱼ is discrete. For the returned Gaussian mixture, ΣwⱼHⱼ ≤ H ≤ ΣwⱼHⱼ+H(w); H(w) alone is not its differential entropy or an implemented game regulariser. homotopy records the quadrature and Monte Carlo checks and the distinction between normalised residual density and varying-A unnormalised soft energy.
Monodromy — the branches are not independently labelled
Transporting around a loop in can permute the branches. The resulting monodromy group measures how globally inconsistent any branch labelling is: if the monodromy is transitive, there is no continuous global choice of branch, and the “elbow up” branch is only locally well defined.
Two consequences:
- Any implementation that caches “which branch we took last time” is making a local choice that cannot be made global. Message passing around a cycle in the factor graph can return to a different branch than it started on — a genuine and under-appreciated failure mode for equilibrium-style iteration.
- Monodromy is also a tool: monodromy loops are the cheapest known way to find many solutions of a polynomial system from one, and are how modern solvers bootstrap.
Related: Inference as Root Finding, Open Models and Latent Channels, Factors are Parameterized Statistical Games, The Algebraic Factor as a Statistical Game