Theorem 6.1 — one-bit feedback achieves regret for independent regular valuations
ProvedBilateralProfitRegular.theorem_6_1_regret_boundbilateral-tradeonline-learningp2o-gran-per-chapterp2o-plan-paperp2o-v1posted-pricesregret-boundsregular-distributions
There exist a learning algorithm with one-bit feedback and a constant such that, for every horizon and every pair of independent regular seller and buyer distributions on with strictly positive densities,
The expectation is over the i.i.d. valuations and the algorithm's randomness. The learner sees only the trade bit of each round. The comparator ranges over all of , including . The paper's proof gives the sharper . Together with Theorem 7.1's lower bound, the rate for regular distributions is .
Formalization Note The algorithm is a family of policies indexed by the horizon, on one seed space, chosen with before and . Logarithms are natural.
Differences from the source:
- reading: "a learning algorithm … for every " → a family of one-bit policies on one seed space, chosen (with ) before and the distributions; reason: the paper's algorithms take the horizon as input (Algorithms 7–8; "we choose as a function of ", p.37)
- encoding: → with the bound for every ; reason: at the right side is ; is asymptotic in
- encoding: of the expected regret; the learner's randomness → the bound for every comparator in ; a seed drawn once from a probability space, valuations i.i.d. independent of it; rounds ; natural log; reason: a supremum is at most iff every term is; standard model of a randomized policy
- encoding: independent regular seller and buyer distributions on with strictly positive densities (§2, §6) →
RegularUnit Efor a publishedEnvironmentE(supports ,E.Regular); , from the joint lawE.prior, =E.cdfS,E.cdfB; reason: the same class: Definitions 2.1–2.2 with the standing positivity of the densities; the valuation laws are the environment's
Preamble
import Mathlib import Definitions.Def_BilateralProfitRegular_Stochastic
Formal statement
namespace BilateralProfitRegular
open MeasureTheory MechanismDesign.BilateralTrade
/-- **Theorem 6.1 (Regret for regular distributions)**, Brüstle, di Gregorio, Fusco, Leonardi
(manuscript, sha256 e91e8127…), p.36 (PDF 38); proof pp.46–49 (PDF 48–51). There exists a learning
algorithm `𝒜` with one-bit feedback such that, for every `T` and every pair `𝒟` of independent
regular buyer and seller distributions on `[0, 1]` with strictly positive densities,
`R_T(𝒜, 𝒟) ≤ O(T^{3/5} log⁵(T))`.
**Formalization Note.**
* The algorithm is a horizon-indexed family `π T` of randomized one-bit policies on one seed
space `(Ω, P)`, chosen (with the constant `C`) before `T` and the distributions: the paper's
algorithms take the horizon `T` as input (Algorithms 7–8; `η` is chosen as a function of `T`).
* `O(·)` is a universal constant `C` for all `T ≥ 2` (at `T = 1` the right side is `0`).
* `R_T = sup_{(p,q) ∈ [0,1]²} ∑_t (E[Profit_t(p, q)] − E[Profit_t(p_t, q_t)])`; the bound is
stated for every comparator in `[0,1]²`, which is equivalent.
* The distributions are a published `Environment` with supports `[0, 1]` and `E.Regular`
(Definitions 2.1–2.2); the valuations are i.i.d. with law `E.prior`, independent of the seed.
Natural logarithm; rounds `Fin T`. -/
theorem theorem_6_1_regret_bound :
∃ C : ℝ, 0 < C ∧
∃ (Ω : Type) (_ : MeasurableSpace Ω) (P : Measure Ω) (_ : IsProbabilityMeasure P)
(π : (T : ℕ) → Policy T Ω),
(∀ T, (π T).OneBit) ∧
∀ T : ℕ, 2 ≤ T → ∀ E : Environment, RegularUnit E →
∀ a ∈ Set.Icc (0 : ℝ) 1 ×ˢ Set.Icc (0 : ℝ) 1,
(π T).regret P E.prior a ≤ C * (T : ℝ) ^ ((3 : ℝ) / 5) * Real.log T ^ 5 := by sorry
end BilateralProfitRegular
Source
Brüstle, di Gregorio, Fusco, Leonardi, Regret Bounds for Profit Maximization in Bilateral Trade: Learning the Best Posted-Price Mechanism (manuscript, STOC 2027 submission, sha256 e91e81279505b765c7c96c219d674c0fbbbd6ca47a00da8fe228d498ca25ac87), Theorem 6.1, p.36 (PDF 38); proof §6.4, pp.46–49 (PDF 48–51)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.