Lemma 4.5 — inexpensive detection (corrected: )
ProvedBilateralProfit.lemma_4_5bilateral-tradeonline-learningp2o-correctedp2o-gran-per-chapterp2o-plan-paperp2o-v1posted-pricesregret-bounds
There is a constant such that for every , every distribution on and every there are non-negative weights , , depending only on and , such that
where is optimal according to ; and 2. for every random valuation on with probability function of trade such that ,
with the trade function induced by .
The same weights work for every such once and are fixed. The lemma drives the proof by contradiction of the information inequality (Theorem 4.2).
Formalization Note is the maximum of the marginal expected profits over . A random valuation is a probability measure carried by , with . Logarithms are natural.
Differences from the source:
- correction: every → ; reason: at the right side of 1 is ; with concentrated on valuation and , every gap is , so 1 forces , but has profit at while the sum in 2 is ; every use in the paper has
- reading: "for some constant " after "for every distribution … and " → chosen before , and ; reason: the proof of Theorem 4.2 uses one for every and ("when ")
- encoding: a random valuation with probability function of trade ; optimal according to → a probability measure on carried by ; and through the maximum value
optVal; reason: same instances; the gap does not depend on the maximizer
Preamble
import Mathlib import Definitions.Def_BilateralProfit_Analysis
Formal statement
namespace BilateralProfit
open MeasureTheory
/-- **Lemma 4.5 (Inexpensive Detection)**, Brüstle, di Gregorio, Fusco, Leonardi (manuscript,
sha256 e91e8127…), p.20 (PDF 22), restated p.23 (PDF 25); corrected: `K ≥ 2`.
There is a constant `C` such that for every `K ≥ 2`, every distribution `ρ` on `G_K × 𝒜` and every
`x ∈ (0, 1]` there are non-negative weights `n_x(a)`, depending only on `ρ` and `x`, with
1. `Σ_{a ∈ 𝒜} n_x(a)(Δ_ρ(a⋆, a) + x) ≤ C log³K (K/x + log³K/x²)`, and
2. for every random valuation `Ṽ` on `[0,1]²` with trade function `P` such that
`max_{a ∈ 𝒜} E[Profit(Ṽ, a)] ≥ E_ρ[Profit(V, a⋆)] + x`:
`Σ_{a ∈ 𝒜} n_x(a)(P(a) − P0(a))² ≥ 1`.
**Formalization Note.** `a⋆` is optimal for the valuation marginal of `ρ`; `Δ_ρ(a⋆, ·)` and
`E_ρ[Profit(V, a⋆)]` do not depend on the choice (`gapStar`, `optVal`). `C` is chosen before
`K`, `ρ`, `x`; the weights before the law of `Ṽ`. `log` is natural. The printed lemma has no
lower bound on `K`; at `K = 1` it is false (see the item's differences). -/
theorem lemma_4_5 :
∃ C : ℝ, ∀ K : ℕ, 2 ≤ K → ∀ ρ : Grid K × Act K → ℝ, IsDist ρ →
∀ x : ℝ, 0 < x → x ≤ 1 →
∃ n : Act K → ℝ, (∀ a, 0 ≤ n a) ∧
∑ a, n a * (gapStar ρ a + x) ≤
C * Real.log K ^ 3 * ((K : ℝ) / x + Real.log K ^ 3 / x ^ 2) ∧
∀ μ : Measure (ℝ × ℝ), IsValLaw μ →
(∃ a : Act K, optVal ρ + x ≤ muP μ (pt K a.1)) →
1 ≤ ∑ a, n a * (Ptrade μ (pt K a.1) - P0 ρ (pt K a.1)) ^ 2 := by sorry
end BilateralProfit
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), Lemma 4.5, p.20 (PDF 22), restated p.23 (PDF 25); proof pp.28–29 (PDF 30–31)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.