Lemma 5.4.9 — MTP2 likelihood ratio order
ProvedMDPFinance.BayesianModels.lemma_5_4_9Lemma 5.4.9, under the "Monotonicity Results" subsection's standing simplifications (; independent of with density ; has a -density ; the order on is ):
a) If is MTP2 in for every , then is increasing for every (and every ).
b) Fixing : for all if and only if is MTP2 in .
Part b) gives a practically checkable sufficient condition (MTP2 of a joint density) for the likelihood-ratio monotonicity hypothesis of the goal Theorem 5.4.10; part a) is the key propagation step used in that theorem's own proof, showing the information-state update itself inherits the monotonicity.
Formalization Note. MTP2 is a joint condition on as a function of the
pair , not a property of or separately — getting the two-variable
quantification right (IsMTP2 on ℝ × ℝ) is the entire content of part b).
Moderation note. The order on is the likelihood ratio order of the measures (LRMeasure, Definition B.3.5). Part a) needs the step "by definition " of the book's proof, which is now the hypothesis hPhihat_bayes (the information-state update is the Bayes update wherever the latter is defined; it holds at every reachable information state by sufficiency, sequentiality and the filter recursion), and it is stated at pairs where both Bayes updates are defined. The draft asserted monotonicity of an unconstrained on all of , with no link between and the sufficient statistic, which is refutable.
import Mathlib import Definitions.Def_MDPFinance_BayesianModels_Model import Definitions.Def_MDPFinance_BayesianModels_Posterior import Definitions.Def_MDPFinance_BayesianModels_SuffStat import Definitions.Def_MDPFinance_BayesianModels_Orders open MeasureTheory ProbabilityTheory
namespace MDPFinance.BayesianModels
/-- Lemma 5.4.9 (Bäuerle–Rieder, p. 163, PDF 177), under the "Monotonicity Results" standing
simplifications of that subsection (p. 164, PDF 177: `Z ⊂ ℝ`, `E_X ⊂ ℝ^d`, `Θ ⊂ ℝ`; `Q^Z(\cdot|
x,θ,a)` independent of `x` with density `q_Z(\cdot|θ,a)` w.r.t. a σ-finite measure — here `qZbar`,
with `hqZ` recording the `x`-independence of `M.qZ`; `μ̂` has a `Q_0`-density `\hat p(θ|i)` —
here `phat`, with `hphat`; the order on `I`, `i ≤ i' :⇔ \hat μ(\cdot|i) ≤_{lr} \hat μ(\cdot|i')`,
i.e. the likelihood ratio order of Definition B.3.5 on the measures `\hat μ(\cdot|i)`, `\hat
μ(\cdot|i')` — here `hI_order` via `LRMeasure`). The proof's step "by definition `\hat μ(\cdot|
\hat Φ(i,a,z)) = Φ(\cdot|\hat μ(\cdot|i),a,z)`" is the hypothesis `hPhihat_bayes`: the
information-state update is the Bayes update of `\hat μ(\cdot|i)` wherever the latter is defined
(normalizing constant positive and finite); it holds at every reachable information state by
`SuffStat.ht_suff`, `IsSequential` and `Posterior.hmu_rec`. a) If `q_Z(z|θ,a)` is an `MTP2`
function in `(z,θ)` for all `a`, then `(i,z) ↦ Φ̂(x,i,a,z)` is increasing for all `a ∈ A` (and
every `x`), at pairs where both Bayes updates are defined. b) Fix `a ∈ A`. We have
`q_Z(\cdot|θ,a) ≤_{lr} q_Z(\cdot|θ',a)` for all `θ ≤ θ'` if and only if `q_Z(z|θ,a)` is an `MTP2`
function in `(z,θ)`. -/
theorem lemma_5_4_9 {EX A I : Type*} [MeasurableSpace EX] [MeasurableSpace A] [MeasurableSpace I]
[Preorder I] (M : BayesModel EX ℝ A ℝ) (Pf : M.Posterior) (S : SuffStat M Pf I)
(Phihat : EX → I → A → ℝ → I) (qZbar : ℝ → A → ℝ → ℝ)
(hqZ : ∀ x θ a z, M.qZ x θ a z = qZbar θ a z) (phat : I → ℝ → ℝ)
(hphat : ∀ i, S.muHat i = M.Q0.withDensity fun θ => ENNReal.ofReal (phat i θ))
(hI_order : ∀ i i' : I, i ≤ i' ↔ LRMeasure M.Q0 (S.muHat i) (S.muHat i'))
(hSeq : S.IsSequential Phihat)
(hPhihat_bayes : ∀ x i a z,
0 < (∫⁻ θ, ENNReal.ofReal (qZbar θ a z) ∂(S.muHat i)) →
(∫⁻ θ, ENNReal.ofReal (qZbar θ a z) ∂(S.muHat i)) < ⊤ →
∀ C : Set ℝ, MeasurableSet C →
S.muHat (Phihat x i a z) C =
(∫⁻ θ in C, ENNReal.ofReal (qZbar θ a z) ∂(S.muHat i)) /
(∫⁻ θ, ENNReal.ofReal (qZbar θ a z) ∂(S.muHat i))) :
((∀ a, IsMTP2 fun q : ℝ × ℝ => qZbar q.2 a q.1) →
∀ x a (i i' : I) (z z' : ℝ), i ≤ i' → z ≤ z' →
0 < (∫⁻ θ, ENNReal.ofReal (qZbar θ a z) ∂(S.muHat i)) →
(∫⁻ θ, ENNReal.ofReal (qZbar θ a z) ∂(S.muHat i)) < ⊤ →
0 < (∫⁻ θ, ENNReal.ofReal (qZbar θ a z') ∂(S.muHat i')) →
(∫⁻ θ, ENNReal.ofReal (qZbar θ a z') ∂(S.muHat i')) < ⊤ →
Phihat x i a z ≤ Phihat x i' a z') ∧
(∀ a, (∀ θ θ' : ℝ, θ ≤ θ' → LikelihoodRatioOrder (qZbar θ a) (qZbar θ' a)) ↔
IsMTP2 fun q : ℝ × ℝ => qZbar q.2 a q.1) := by sorry
end MDPFinance.BayesianModels
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.