Randomized mixtures of Gittins policies interpolate reward and cost
ProvedPandoraBO.Private20261023.gittins_policy_mixtureexpected-budgetmeasure-theorypandoras-box
Two Gittins policies at the same positive multiplier can be mixed with any weight a in [0,1]. There exists a Gittins policy whose expected cost and expected reward are the corresponding weighted averages. No continuity or atomlessness assumption on reward laws is needed.
Preamble
import Definitions.Def_PandoraBO_Model import Mathlib.MeasureTheory.Integral.Prod open MeasureTheory PandoraBO set_option autoImplicit false
Formal statement
theorem PandoraBO.Private20261023.gittins_policy_mixture {n : ℕ} (M : Instance n) (multiplier : ℝ)
(hpos : 0 < multiplier) (π ρ : Policy n)
(hπ : isGittinsPolicy M multiplier π) (hρ : isGittinsPolicy M multiplier ρ)
(a : ℝ) (ha0 : 0 ≤ a) (ha1 : a ≤ 1) :
∃ σ : Policy n, isGittinsPolicy M multiplier σ ∧
expectedCost M σ = a * expectedCost M π + (1 - a) * expectedCost M ρ ∧
expectedReward M σ = a * expectedReward M π + (1 - a) * expectedReward M ρ := by sorry
Source
Auxiliary construction for Xie et al., Cost-aware Bayesian Optimization via the Pandora’s Box Gittins Index, author revised manuscript reviewed 23 October 2026, Appendix B.5, Lemmas 7–8 and Theorem 2. This result isolates the measurable-policy existence or policy-randomization step; it is not a verbatim numbered claim.