Lemma 15 — tracking a cumulated sum of proportions
ProvedOptimalBAI.TrackStop.cumulative_tracking_boundbest-arm-identificationdeterministicp2o-batch-p200bp2o-gran-per-chapterp2o-plan-paperp2o-v1tracking
Let be a positive integer, the simplex of dimension , and the vertex of with a on coordinate . For a positive integer , let and for . Let and, for ,
Then
The lemma is the deterministic core of C-Tracking: greedily drawing the arm whose count lags most behind the cumulated targets keeps every count within of its target.
Formalization Note Arms are indexed ; the sequence keeps the paper's indices , and its other values are ignored.
Preamble
import Mathlib import Definitions.Def_OptimalBAI_TrackStop_OptimalProportions
Formal statement
namespace OptimalBAI.TrackStop
/-- Lemma 15 (Garivier–Kaufmann, arXiv:1602.04589v2, p. 20). Proportions `p(1), …, p(n) ∈ Σ_K`
with cumulated sums `P(k) = p(1) + ⋯ + p(k)`; `N(0) = 0`, `I_{k+1}` maximizes
`P_i(k+1) - N_i(k)` over `i`, and `N(k+1) = N(k) + δ_{I_{k+1}}`. Then
`max_i |N_i(n) - P_i(n)| ≤ K - 1`. Arms are `Fin K` (the paper's arm `i` is index `i - 1`); the
sequence `p` keeps the paper's 1-based index, and its values outside `1, …, n` are ignored. -/
theorem cumulative_tracking_bound {K : ℕ} (hK : 0 < K) (n : ℕ) (hn : 0 < n)
(p : ℕ → Fin K → ℝ) (hp : ∀ k, 1 ≤ k → k ≤ n → p k ∈ simplex K)
(N : ℕ → Fin K → ℝ) (I : ℕ → Fin K)
(hN0 : N 0 = 0)
(hI : ∀ k, k < n → ∀ i : Fin K,
(∑ j ∈ Finset.Icc 1 (k + 1), p j i) - N k i ≤
(∑ j ∈ Finset.Icc 1 (k + 1), p j (I (k + 1))) - N k (I (k + 1)))
(hN : ∀ k, k < n → N (k + 1) = N k + Pi.single (I (k + 1)) (1 : ℝ)) :
∀ i : Fin K, |N n i - ∑ j ∈ Finset.Icc 1 n, p j i| ≤ (K : ℝ) - 1 := by sorry
end OptimalBAI.TrackStop
Source
Garivier, Kaufmann, Optimal Best Arm Identification with Fixed Confidence, arXiv:1602.04589v2, p. 20, Lemma 15
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.