Finite policy rollouts preserve valid histories and pathwise bounds
ProvedPandoraBO.Private20261023.policy_rollout_structuraloptimal-searchpandoras-box
Every policy rollout yields a nonempty occupied prefix with no repeated box, and every occupied slot records that box’s sampled reward. Its terminal reward is bounded above by the maximum of all sampled box rewards, and its terminal cost is bounded above by the sum of all opening costs. These are pointwise assertions for every outcome and seed, without exceptional null sets.
Preamble
import Definitions.Def_PandoraBO_Model open MeasureTheory PandoraBO set_option autoImplicit false
Formal statement
theorem PandoraBO.Private20261023.policy_rollout_structural {n : ℕ} (M : PandoraBO.Instance n) (π : PandoraBO.Policy n)
(ω : PandoraBO.Outcome n) :
PandoraBO.validHistory (PandoraBO.terminalHistory π ω) ∧
(∀ k, PandoraBO.occupied (PandoraBO.terminalHistory π ω) k →
PandoraBO.observedReward (PandoraBO.terminalHistory π ω) k =
ω.1 (PandoraBO.observedBox (PandoraBO.terminalHistory π ω) k)) ∧
PandoraBO.terminalReward π ω ≤ PandoraBO.fullInformationPayoff ω.1 ∧
PandoraBO.terminalCost M π ω ≤ ∑ i : PandoraBO.Box n, M.cost i := by sorry
Source
Finite policy and capped-reward foundations supporting Xie et al., author revised manuscript reviewed 23 October 2026, Appendix B.5, Lemma 7 and equation (42). These pathwise auxiliary statements are isolated for the Lean proof rather than separately numbered in the paper.