Internals

For contributors. None of these names is public.

The two forms of DeltaApproximation, which the dependencies and rules dispatch on:

The routines the Gaussian rules share: the two methods differ only in these.

DeltaMessagePassingRules.approximate_normal — Function
approximate_normal(method, f, distributions::Tuple)

The normal approximation of f(x₁, …, xₙ) for independent normal inputs distributions, by method, Unscented or Linearization: a normal with the pushforward's mean and covariance, univariate or multivariate as the output of f is. The message rules towards out, and towards an input through a known inverse, are this call.

source
DeltaMessagePassingRules.forward_statistics — Function
forward_statistics(method, f, μs, Σs, μ_in, Σ_in) -> (μ, Σ, C)

The mean μ and covariance Σ of f of the inputs, and the cross-covariance C of the inputs with it, by method, for inputs with means μs and covariances Σs (tuples, one entry per input) whose joint has mean μ_in and covariance Σ_in. The rule for the joint over the inputs smooths these with the message from out, by smoothRTS.

source

The error of a method the node does not accept, and the projection families of CVIProjection: