Emission posterior: inferring a partially observed joint action #
For a kernel κ : α → PMF (β × γ) emitting a pair of which only the first
component is observed, PMF.emissionPosterior κ μ b is the Bayesian posterior
over α × γ — state together with unobserved component — given observation
b: the normalization of (a, g) ↦ μ a · κ a (b, g).
This is the dual of JointPosterior.lean: there the state is a product and
the observation is total; here the emission is a product and the observation
is partial, which no PMF-valued kernel into β can express (the slice
(a, g) ↦ κ a (b, g) sums to the γ-marginal, not to 1). At γ := Unit
the construction collapses to PMF.posterior.
Main definitions #
PMF.emissionPosterior— posterior overα × γfrom observing the first component of a jointly emitted pair.
Main results #
emissionPosterior_fst_lt_iff/emissionPosterior_snd_lt_iff— marginal comparisons reduce to prior-weighted emission sums.emissionPosterior_uniform_fst_lt_iff/_uniform_snd_lt_iff— at a uniform prior, to bare emission sums.
The total score of an emission posterior is the observation marginal of
the fst-projected kernel.
The conditional distribution over state and unobserved emission component
(a, g), given that the kernel κ emitted a pair with observed first
component b.
Equations
- PMF.emissionPosterior κ μ b h = PMF.normalize (fun (x : α × γ) => μ x.1 * (κ x.1) (b, x.2)) ⋯ ⋯
Instances For
A single witness (a, g) with μ a ≠ 0 and κ a (b, g) ≠ 0 makes the
observation marginal non-zero.
Comparing state marginals of the emission posterior reduces to comparing prior-weighted emission sums; the observation marginal cancels.
Companion of emissionPosterior_fst_lt_iff for the unobserved component.
At a uniform prior the prior cancels: state comparison reduces to bare emission sums over the unobserved component.
At a uniform prior the prior cancels: emission-component comparison reduces to bare emission sums over states.