definition theorem

GaussianBelief(η, Λ) is a Gaussian in canonical (information) form, , with precision and information vector . In this form pooling is addition, and a precision may be singular, so likelihood messages are representable too.

Sources: code: beliefs.jl (in the top-level Lenticulum package), messages.jl; Koller & Friedman, Probabilistic Graphical Models (MIT Press 2009), §14.2 (canonical forms)

Bibliography: Koller & Friedman 2009

Theory (CT-ML wiki): Gaussian Relations · Markov Category · Bayesian Inversion

Why the canonical form

Pooling is addition. The product of two Gaussian densities over the same variable is

so combine is exact, associative, commutative and never fails. In moment form the same operation needs two matrix inversions.

Improper beliefs are allowed. A factor → variable message is a likelihood, and a likelihood that constrains only some directions has a rank-deficient precision (a measurement of says nothing about ). The moment form cannot represent it at all; the canonical form just has zeros. Such a belief is valid as a message and becomes a distribution once combined with a prior that covers the remaining directions. isproper(b) checks , and belief_mean, belief_cov, belief_logdensity and variable_entropy refuse improper beliefs rather than return a silent pseudo-answer. A factor that knows its message is rank-deficient builds it as GaussianBelief(η, Λ, true): rounding often makes such a factorise, and the flag keeps isproper false all the same (LenticulumCore gaussian_belief.md, “Declared improper”).

The other beliefs are its limits

limitbelief
the unit: Trivial Belief (uninformative(n))
, fixeda point mass: Dirac Belief, which is why a Dirac dominates combine
a precision on some coordinatesa soft clamp (Channels and Polarity); the diffusion factor’s is exactly this

So the precision is a single dial from “no information” to “certainty”, and the polarity precisions of Implicit Diffusion Learners §2 are Gaussian beliefs on the clamped coordinates.

Operations

  • Gaussian(μ, Σ) builds one from moments; belief_mean, belief_cov go back.
  • logpartition(b) .
  • variable_entropy(b) . It can be negative, and the sign matters for the Bethe counting correction (Bethe Free Energy).
  • kl_divergence(q, p) for proper beliefs; belief_distance compares canonical parameters directly, so it works for improper messages too.
  • Gaussian messages can be damped (mixed) during loopy message passing (Loopy Message Passing).

Where it lives, and who emits it

The type and its accessors are in LenticulumCore (gaussian_belief.jl), its message rules in Mycelium (gaussian.jl), the projections in the top-level package (beliefs.jl). It used to live entirely in the top-level package, out of reach of the factors in lib/.

The implicit diffusion learner now emits it: the Laplace approximation at an answer (laplace_belief), a mixture of them over all answers (implicit_mixture), and as a factor’s message with the target’s anchor divided out (ImplicitProx(…; message = :gaussian), implicit_factor §5). The equilibrium factors could do the same through their sensitivity and do not yet.

Related: Beliefs, Dirac Belief, Trivial Belief, The Linear Gaussian Chain, Messages are Inversions, Bethe Free Energy