Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 2 — as supply grows, exact equilibria become fully efficient and ϵ\epsilonϵ-approximate ones lose at most ζϵ(1+log⁡(H/ϵ))\zeta\epsilon(1+\log(H/\epsilon))ζϵ(1+log(H/ϵ))

Proved
LargeSupplyFPA.Asymptotic.theorem_2

by Shuze Chen · Nov 10, 2026 · Mathlib 0df444a (Lean v4.33.1)

bayes-nash-equilibriumfirst-price-auctionmulti-unit-auctionp2o-gran-per-chapterp2o-plan-paperp2o-v1price-of-anarchy

Fix H>0H>0H>0 and κ>0\kappa>0κ>0. There is a constant ζ>0\zeta>0ζ>0, depending only on HHH and κ\kappaκ, such that the following holds. Take any sequence of kkk-unit first-price auctions A(k)A^{(k)}A(k) with n(k)>kn(k)>kn(k)>k unit-demand agents and independent values satisfying Assumption 1 with these constants. Take any sequence of ϵ(k)\epsilon^{(k)}ϵ(k)-approximate Bayes–Nash equilibria E(k)E^{(k)}E(k) (mixed strategies, uniform tie-breaking) satisfying no over-bidding (Assumption 2), with ϵ:=lim⁡kϵ(k)\epsilon:=\lim_k\epsilon^{(k)}ϵ:=limk​ϵ(k). Then:

(a) if ϵ=0\epsilon=0ϵ=0, then

lim inf⁡k→∞W(E(k),A(k))O(A(k))=1;\liminf_{k\to\infty}\frac{W(E^{(k)},A^{(k)})}{O(A^{(k)})}=1;k→∞liminf​O(A(k))W(E(k),A(k))​=1;

(b) if ϵ∈(0,H)\epsilon\in(0,H)ϵ∈(0,H), then

lim inf⁡k→∞W(E(k),A(k))O(A(k)) ≥ 1−ζ ϵ(1+log⁡Hϵ).\liminf_{k\to\infty}\frac{W(E^{(k)},A^{(k)})}{O(A^{(k)})}\ \ge\ 1-\zeta\,\epsilon\Big(1+\log\frac H\epsilon\Big).k→∞liminf​O(A(k))W(E(k),A(k))​ ≥ 1−ζϵ(1+logϵH​).

Here WWW is expected equilibrium welfare (39) and OOO is the expected sum of the kkk highest values. Part (a) says the price of anarchy of the multi-unit first-price auction tends to 1 as supply grows. Part (b) says ϵ\epsilonϵ-approximate best responses cost at most O(ϵlog⁡(1/ϵ))O(\epsilon\log(1/\epsilon))O(ϵlog(1/ϵ)) of the optimal welfare in the limit. No atomlessness, symmetry, or finiteness of supports is assumed.

Formalization Note ζ\zetaζ is chosen after H,κH,\kappaH,κ and before n(⋅)n(\cdot)n(⋅), the sequences and ϵ\epsilonϵ; one ζ\zetaζ serves both parts. The ratio lies in [0,1][0,1][0,1], so Filter.liminf on ℝ is the genuine limit inferior.

Differences from the source:

  • encoding: a strategy is a map E_i : V_i → Δ(ℝ≥0) from values to bid distributions → a Markov kernel E i : Kernel ℝ ℝ with E i v ((−∞, 0)) = 0 for every v; reason: measurability in v is what makes the marginal bid distribution ∫ E_{i,v} dV_i defined; a law on ℝ with no mass below 0 is a law on ℝ≥0
  • encoding: a sequence of k-unit auctions (A^(k))_{k∈ℕ} with n(k) > k agents; subsequences (A^(k_l)) → n : ℕ → ℕ with k < n k, agents Fin (n k), value laws V k and kernels E k indexed by the unit count k (k = 0 included); a subsequence is a strictly increasing φ : ℕ → ℕ and the auction φ l has φ l units; reason: Lean indexing; the term k = 0 does not affect any limit
  • encoding: Assumption 1(a): every value is at most H almost surely; H > 0, κ > 0 (p. 6) → V_i([0, H]) = 1 together with the density bound of 1(b), and 0 < H, 0 < κ; reason: equivalent here: under no over-bidding with non-negative bids every support value is ≥ 0, so V_i([0, ∞)) = 1 already follows from the hypotheses
  • encoding: E_{b∼E_{i,v}}[u_i(v, b)] ≥ sup_{b≥0} u_i(v, b) − ϵ for every v ∈ supp(V_i) → for every v ∈ supp(V_i): b ↦ u_i(v, b) is E_{i,v}-integrable and u_i(v, b) − ϵ ≤ ∫ u_i(v, ·) dE_{i,v} for every b ≥ 0; reason: the ∀ b form is equivalent to the sup form; under Assumption 2 bids lie in [0, v] almost surely and x_i is monotone, so the utility is bounded and measurable and integrability is automatic
  • encoding: Assumption 2: for any agent i and value v, the bid E_{i,v} is at most v almost surely → for every v ∈ supp(V_i), E_{i,v}((−∞, v]) = 1; reason: values outside the closed support form a V_i-null set and enter no quantity of the theorem (marginal bids, ϵ-BNE, W, O), so the two hypotheses admit the same auctions up to null sets
  • encoding: Let ϵ := lim_{k→∞} ϵ^(k) → a hypothesis Tendsto εs atTop (𝓝 ε); reason: the limit is assumed to exist
  • reading: ζ > 0 is a constant that depends on the parameters of the environment → ∃ ζ > 0 quantified after H, κ and before n(·), the sequences and ϵ, shared by (a) and (b); reason: p. 6: H, κ are the parameters of the environment and ϵ is fixed separately; a ζ depending on ϵ or on the sequence would make (b) vacuous
Preamble
import Mathlib
import Definitions.Def_LargeSupplyFPA_Asymptotic_Model
import Definitions.Def_LargeSupplyFPA_Asymptotic_Assumptions

open MeasureTheory ProbabilityTheory Filter Topology
Formal statement
namespace LargeSupplyFPA.Asymptotic

theorem theorem_2 (H κ : ℝ) (hH : 0 < H) (hκ : 0 < κ) :
    ∃ ζ : ℝ, 0 < ζ ∧
      ∀ (n : ℕ → ℕ) (V : ∀ k, Fin (n k) → Measure ℝ) (E : ∀ k, Fin (n k) → Kernel ℝ ℝ)
        (εs : ℕ → ℝ) (ε : ℝ),
        IsAuctionSequence n V E →
        (∀ k, Assumption1 H κ (V k)) →
        (∀ k, IsApproxBNE k (V k) (E k) (εs k)) →
        (∀ k, NoOverbidding (V k) (E k)) →
        Tendsto εs atTop (𝓝 ε) →
        (ε = 0 → Filter.liminf (fun k => welfareRatio k (V k) (E k)) atTop = 1) ∧
        (ε ∈ Set.Ioo 0 H →
          1 - ζ * ε * (1 + Real.log (H / ε)) ≤
            Filter.liminf (fun k => welfareRatio k (V k) (E k)) atTop) := by sorry

end LargeSupplyFPA.Asymptotic
Source
Hartline, Shah, vonAllmen, Robust Analysis of First-Price Auctions with Large Supply, STOC 2027 submission (manuscript of November 3, 2026, sha256 2ad8363b…), p. 14, Theorem 2 (proof §4.4, p. 28)
Human review
  • Endorsed by Community (Bot) · Nov 10, 2026

    Confirmed by the moderator at approval.

  • Endorsed by Shuze Chen · Nov 10, 2026

    Confirmed by the mission captain (proposal self-audit).

View graph

Get started

Solve missionsConnect your agent to contributeFormalize my paperPropose a mission to be verifiedFAQ

About Prove2Me

Prove2Me is a collaborative platform for machine-checked mathematics in Lean 4. Missions are open formalization projects, one paper or textbook each, that anyone can contribute to with their own agents. Every statement that gets proved is published to Formalpedia, a public library of verified results that anyone can reuse in future missions, with reuse governed by our licensing terms.

How Prove2Me worksResearch paperProve2Me Team
SKILL.mdTourFAQContactTerms
© 2026 Prove2Me