Sources: code:
interop.jl,ext/LenticulumCoreExponentialFamilyExt.jl; Bagaev, Podusenko & de Vries, RxInfer: A Julia package for reactive real-time Bayesian inference, JOSS 2023Bibliography: Bagaev et al. 2023
Theory: RxInfer as a Backend · Belief Algebra · Beliefs
Implements: step 1 of RxInfer as a Backend §4. Every belief converts to the distribution type RxInfer.jl passes as a message, and back.
1. The functions
as_distribution(b) and as_belief(d). The stubs live in interop.jl; the methods are in a
package extension that loads when ExponentialFamily.jl, BayesBase.jl and Distributions.jl are
loaded, which using RxInfer does. LenticulumCore itself gains no dependency.
| belief | distribution | exact? |
|---|---|---|
GaussianBelief(η, Λ) | MvNormalWeightedMeanPrecision(η, Λ) (univariate = true: NormalWeightedMeanPrecision) | yes: both canonical |
DiracBelief | PointMass | yes |
CategoricalBelief | Categorical | probabilities yes; labels are dropped |
MixtureBelief | MixtureModel | yes, componentwise |
SampleBelief | SampleList | yes; weights normalised |
TrivialBelief | — | throws |
In the other direction, any normal (Distributions’ Normal and MvNormal, every
parametrisation in ExponentialFamily) goes through BayesBase’s weightedmean_precision, so it
lands in canonical form without inverting a covariance. Bernoulli becomes a two-state
CategoricalBelief.
2. Design decisions
- Canonical to canonical. RxInfer’s
MvNormalWeightedMeanPrecisionis exactly . Converting through the moment form would fail for the improper Gaussians that likelihood messages are, and would cost an inversion each way. - An extension on ExponentialFamily, not on RxInfer. The message types live in ExponentialFamily and BayesBase; RxInfer is the engine. Depending on the smaller packages means the conversion works for anyone using those types, RxInfer or not.
- Functions, not
convertmethods. The conversions are not always lossless (labels,TrivialBelief), andconvertis called implicitly.
3. How it is tested
In the top-level Lenticulum test suite (“interop”), because the checks need Mycelium’s
combine:
- round trips of every type, an improper Gaussian included;
- Lenticulum’s pooling (
combine) against BayesBase’sprodfor Gaussians and categoricals; - the posterior of a one-step conjugate model, , ,
, computed by RxInfer’s
inferand by aGaussianFactormessage pooled with the prior: identical.
4. What it does not do
BayesBase’s prod of a mixture and a Gaussian is a lazy ProductOf, while combine computes
the product mixture in closed form; the conversion makes the two interchangeable but does not
add rules to RxInfer. Step 3 of RxInfer as a Backend §4, a learned relation as an RxInfer
node, builds on this and is not done.
Related: gaussian_belief, open_model, messages
LP-018 finite particle weights (2026-10-29)
The SampleList conversion divides weights by their maximum before summing. Finite equal
weights at 1e308 retain their mean rather than becoming zero through an overflowing
normaliser. All-zero particle weights are refused. Evidence: the top-level
test/lp018_interop.jl, “LP-018.9 finite particle weights across interop and AD”, compares
against the two-particle mean 2. The optional extension still has no core dependency cost.
Scalar particles (2026-10-29, LP-019 item 8)
as_belief(::SampleList) collected every sample, which turned scalars into zero-dimensional
arrays: SampleBelief([1, 3]) came back as [fill(1), fill(3)], and pooling it with a
categorical belief then threw. Scalars are now unwrapped and only array samples collected.
Tested in the top-level test/lp019_interop.jl.