Lemma 5.4.2 — the filter process is a martingale
ProvedMDPFinance.BayesianModels.lemma_5_4_2Lemma 5.4.2. For a fixed event , the process is a martingale: for every policy and , .
Since depends on the history only through once the history up to stage and the action are fixed, this conditional expectation reduces to an ordinary integral of against the predictive law of given the history — the mixture of the disturbance density over the current posterior, .
Formalization Note. No canonical path measure/filtration is constructed anywhere in this
series (matching MDPFinance.POMDP.lemma_5_2_2, chunk 05a); the martingale property is stated
directly as this predictive-marginal integral identity, which is definitionally what a
conditional expectation given a -algebra generated by the listed random variables would
mean here.
import Mathlib import Definitions.Def_MDPFinance_BayesianModels_Model import Definitions.Def_MDPFinance_BayesianModels_Posterior open MeasureTheory ProbabilityTheory
namespace MDPFinance.BayesianModels
/-- Lemma 5.4.2 (Bäuerle–Rieder, p. 160, PDF 173). For fixed `C ∈ B(Θ)` the process
`M_n := μ_n(C|X_0,A_0,Z_1,X_1,\dots,X_n)` is a martingale, i.e. for all `π` and `x ∈ E_X`:
`𝔼^π_x[M_{n+1} | X_0,A_0,Z_1,X_1,\dots,X_n] = M_n`, `n ∈ ℕ_0`. Since `M_{n+1}` is a function of
`Z_{n+1}` alone (through the Bayes update) once `(x_n,a_n)` and the history up to `n` are fixed,
the conditional expectation reduces to integrating `M_{n+1}` against the *predictive* law of
`Z_{n+1}` given the history — the mixture `(μ_n(\cdot|h̃_n)).bind (q_Z(x_n,\cdot,a_n,\cdot)ν)` of
the disturbance density over the current posterior — matching
`MDPFinance.POMDP.lemma_5_2_2` (chunk `05a`)'s use of a predictive-marginal integral identity in
place of a literal conditional-expectation-with-filtration statement (no canonical path measure
is constructed in this series; see `MODERATION_NOTES.md`). -/
theorem lemma_5_4_2 {EX Θ A Z : Type*} [MeasurableSpace EX] [MeasurableSpace Θ]
[MeasurableSpace A] [MeasurableSpace Z] (M : BayesModel EX Θ A Z) (Pf : M.Posterior)
(n : ℕ) (xs : ℕ → EX) (as : ℕ → A) (zs : ℕ → Z) (han : as n ∈ M.Dx (xs n)) (C : Set Θ)
(hC : MeasurableSet C) :
∫ z, (Pf.mu (n + 1) xs as (Function.update zs (n + 1) z) C).toReal
∂((Pf.mu n xs as zs).bind fun θ => M.nu.withDensity fun z =>
ENNReal.ofReal (M.qZ (xs n) θ (as n) z)) =
(Pf.mu n xs as zs C).toReal := by sorry
end MDPFinance.BayesianModels
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.