Theorem 5 — needle lower bound
OpenOrderedExperts.Supplement.theorem_5bandit-algorithmsonline-learningp2o-gran-per-chapterp2o-plan-paperp2o-v1regret-bounds
There is an absolute constant such that for every and every randomized two-point algorithm for the BV problem, some oblivious BV instance satisfies
This is the search cost of finding the hidden zero-loss action when each round can inspect at most two candidates.
Formalization Note The learner observes only the two aggregate losses. Recommendations and losses are fixed before interaction and may depend on the learner.
Differences from the source:
- encoding: arbitrary randomized two-point algorithms and expected regret → conditional finite pair laws and an exact finite recursion; reason: this is the same nonanticipating learner class on finite active chains
- encoding: → one positive constant quantified before , , the learner, and the hard instance; reason: Appendix C.3 supplies a uniform constant (indeed )
- encoding: an unspecified ordered action space for the hard instance → the finite ordered action space
Fin n; reason: the needle construction uses only finitely many ordered candidates
Preamble
import Definitions.Def_OrderedExperts_Supplement_twoExpLoss import Definitions.Def_OrderedExperts_Supplement_bestExpertLoss import Definitions.Def_OrderedExperts_Supplement_IsBVLoss
Formal statement
namespace OrderedExperts.Supplement
/-- **Theorem 5 (Needle lower bound).** Zuo, *Online Learning with Ordered Experts (via
Two-point Feedback)*, STOC 2027 submission (held version), Theorem 5, p. 13 and proof in
Appendix C.3, p. 26.
One absolute positive constant works for every two-point BV learner: a fixed oblivious BV loss
sequence forces expected regret at least `c min(n,T)`.
**Formalization Note.** The finite action set is `Fin n`. The learner observes only the two
aggregate losses. The existential recommendations and loss sequence may depend on the learner
but are fixed before interaction. -/
theorem theorem_5 :
∃ c : ℝ, 0 < c ∧ ∀ (n T : ℕ) [NeZero n], 2 ≤ n → 2 ≤ T →
∀ P : TwoPointPolicy (Fin n) n (ℝ × ℝ),
∃ (rec : ℕ → Fin n → Fin n) (loss : ℕ → Fin n → ℝ),
(∀ t < T, IsBVLoss (rec t) (loss t)) ∧
c * (min n T : ℕ) ≤
twoExpLoss P rec loss
(fun t a b ↦ (loss t (chainPt rec t a), loss t (chainPt rec t b))) T 0 [] -
bestExpertLoss T rec loss := by sorry
end OrderedExperts.Supplement
Source
Zuo, Online Learning with Ordered Experts (via Two-point Feedback), STOC 2027 submission (held version), Theorem 5, p. 13; Appendix C.3, p. 26