Section 2, Eq. (14) — under , with probability 1
ProvedRobbinsSeqDesign.ForcedSampling.average_reward_tendsto_max_aeLet and be two populations whose expectations and exist (are finite), let and be disjoint sequences of density , and sample according to Robbins' rule : forced draws from at the times and from at the times , and otherwise a draw from the population whose running mean is larger (ties to ). With , with probability 1,
This is the almost-sure form of the statement that asymptotically does as well, per draw, as an experimenter who knows which population is better. The paper attributes it to the strong law of large numbers.
Formalization Note. "The expectations exist" is the hypothesis that both reward laws are integrable; no further condition (variance, boundedness) is assumed, and is allowed. "With probability 1" is an almost-everywhere statement under the trajectory law banditTrajMeasure ν (rule σ).
import Mathlib import Definitions.Def_RobbinsSeqDesign_ForcedSampling_Rule import Definitions.Def_RobbinsSeqDesign_ForcedSampling_Loss
namespace RobbinsSeqDesign.ForcedSampling
open MeasureTheory ProbabilityTheory Filter Topology BanditAlgorithm
/-- Robbins (1952), Section 2, Eq. (14), p. 532: for any two populations A, B whose
expectations `α`, `β` exist (are finite) and any pair of sequences (13), under the rule `R̄`,
with probability 1, `S_n / n → max(α, β)`. -/
theorem average_reward_tendsto_max_ae (ν : StochasticBandit 2)
(hint : ∀ i, Integrable id (ν.P i)) (σ : ForcedSchedule) :
∀ᵐ ω ∂(banditTrajMeasure ν (rule σ)),
Tendsto (fun n : ℕ => partialSum n ω / (n : ℝ)) atTop
(𝓝 (max (banditArmMean ν 0) (banditArmMean ν 1))) := by sorry
end RobbinsSeqDesign.ForcedSampling
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.