Section 2, Eq. (16) — the loss of tends to 0 for any two populations with finite means
ProvedRobbinsSeqDesign.ForcedSampling.forced_sampling_loss_tendsto_zeroLet and be any two populations whose expectations and exist (are finite), and let
be disjoint sequences of positive integers of density . Then Robbins' rule (forced draws from at the 's and from at the 's, otherwise from the population with the larger running mean, ties to ) satisfies
where is the loss of display (9).
This answers affirmatively Robbins' question (12) — whether some rule has for every — not only for coins but for any two populations with finite means, and it does so with one explicit rule, uniformly in the choice of the density-zero schedule. In later terminology: a forced-exploration greedy policy has regret on every two-armed bandit with integrable rewards.
Formalization Note. "For any such that exist" is: for every two-armed StochasticBandit with both reward laws integrable; no other restriction on the class of populations is imposed, and is allowed. The statement is for the specific rule and every admissible schedule, not merely the existence of some rule. The limit is along in .
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. (16), p. 532: for any two populations A, B such that
`α`, `β` exist (are finite) and any pair of sequences (13), the loss (9) of the rule `R̄`
tends to zero: `lim_{n→∞} L_n(A, B, R̄) = 0`. -/
theorem forced_sampling_loss_tendsto_zero (ν : StochasticBandit 2)
(hint : ∀ i, Integrable id (ν.P i)) (σ : ForcedSchedule) :
Tendsto (fun n : ℕ => loss ν (rule σ) n) atTop (𝓝 0) := by sorry
end RobbinsSeqDesign.ForcedSampling
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.