Quantile MDP Theorem 5 (corrected hypothesis) — conditional quantile order passes to the common mixture
ProvedQuantileMDP.theorem_5_correctedLet and be probability laws of total rewards conditional on starting state , with a common starting-state distribution . If
then the unconditional reward laws satisfy
The Lean statement allows arbitrary nonnegative mixture weights. For probability weights the two mixtures are the unconditional laws.
This is the mixture argument used in Appendix B.2.5 of Li, Zhong and Brandeau. The original theorem is valid for T ≥ 2 under its stated filtration: it includes times 0 through k for both chains. Applying (12) at k = T − 1 for T ≥ 2 makes both total rewards measurable with respect to the conditioning information, so their conditional quantiles are the totals themselves. The example with (B,B) and (1−B,B) only concerns the weaker hypothesis that compares future rewards after time 0. It does not satisfy (12).
import Mathlib.MeasureTheory.Measure.Typeclasses.Probability import Definitions.Def_QuantileMDP_quantile open MeasureTheory Set QuantileMDP open scoped ENNReal
theorem QuantileMDP.theorem_5_corrected {S : Type*} [Fintype S] (p : S → ℝ≥0∞)
(μ ν : S → Measure ℝ) (hμ : ∀ i, IsProbabilityMeasure (μ i))
(hν : ∀ i, IsProbabilityMeasure (ν i))
(h : ∀ i τ, τ ∈ Ioo (0 : ℝ) 1 → quantile (ν i) τ ≤ quantile (μ i) τ)
(τ : ℝ) (hτ : τ ∈ Ioo (0 : ℝ) 1) :
quantile (∑ i, p i • ν i) τ ≤ quantile (∑ i, p i • μ i) τ := by sorry