Quantile MDP Theorem 5 — conditional order given the joint history
ProvedQuantileMDP.theorem_5For two finite-state processes on a common probability space, let each total reward be the sum of the same state reward function over the horizon. Suppose the initial state marginals agree and the comparison of conditional quantiles in Theorem 5 holds given the joint histories from time 0 through time k. Then the unconditional quantiles are ordered. It is sufficient, for a horizon of at least two, to compare at a single interior quantile level on each positive-probability atom of the final joint history (k = T − 1). This gives the result at every level, including the endpoints.
Preamble
import Mathlib.Probability.ConditionalProbability import Definitions.Def_QuantileMDP_quantile open MeasureTheory ProbabilityTheory Set QuantileMDP open scoped ENNReal
Formal statement
theorem QuantileMDP.theorem_5 {Ω S : Type*} [MeasurableSpace Ω] [Fintype S]
[MeasurableSpace S] [MeasurableSingletonClass S] (P : Measure Ω) [IsProbabilityMeasure P]
(T : ℕ) (X Y : Fin T → Ω → S) (hX : ∀ t, Measurable (X t)) (hY : ∀ t, Measurable (Y t))
(r : S → ℝ) (hstart : ∀ hT : 0 < T, P.map (X ⟨0, hT⟩) = P.map (Y ⟨0, hT⟩))
(u : ℝ) (hu : u ∈ Ioo (0 : ℝ) 1)
(h : 1 < T → ∀ j : (Fin T → S × S), P {w | (fun t => (X t w, Y t w)) = j} ≠ 0 →
quantile ((cond P {w | (fun t => (X t w, Y t w)) = j}).map
(fun w => ∑ t, r (Y t w))) u ≤
quantile ((cond P {w | (fun t => (X t w, Y t w)) = j}).map
(fun w => ∑ t, r (X t w))) u)
(τ : ℝ) :
quantile (P.map (fun w => ∑ t, r (Y t w))) τ ≤
quantile (P.map (fun w => ∑ t, r (X t w))) τ := by sorrySource
Li, Zhong and Brandeau, Quantile Markov Decision Processes, arXiv:1711.05788v5, §5.2, Theorem 5, p. 22. https://arxiv.org/abs/1711.05788v5