Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Lemma 4.3, p. 14 — (1 + Σ_{j∈S_ℓ} v_j)(R̃_ℓ(S_ℓ) − R(S_ℓ, v)) ≤ Σᵢ [C₁√(vᵢ log(√Nℓ+1)/Tᵢ(ℓ)) + C₂ log(√Nℓ+1)/Tᵢ(ℓ)] w.p. ≥ 1 − 13/ℓ

Proved
MNLBandit.WellSeparated.lemma_4_3

by mikedeng1 · Oct 6, 2026 · Mathlib 0df444a (Lean v4.33.1)

assortment-optimizationmnl-banditp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1upper-confidence-bound

Consider the MNL-Bandit problem under Assumption 4.1 (0≤vi≤v0=10\le v_i\le v_0=10≤vi​≤v0​=1, a feasible family of the form (2.3) closed under subsets), with revenues ri∈[0,1]r_i\in[0,1]ri​∈[0,1], and run Algorithm 1 with any tie-breaking rule. After epoch ℓ\ellℓ, the algorithm computes the upper confidence bounds vi,ℓUCBv^{\mathrm{UCB}}_{i,\ell}vi,ℓUCB​ and offers in the next epoch

Sℓ+1∈arg⁡max⁡S∈SR~ℓ+1(S),R~ℓ+1(S)=∑i∈Srivi,ℓUCB1+∑j∈Svj,ℓUCB.S_{\ell+1}\in\arg\max_{S\in\mathcal S}\tilde R_{\ell+1}(S),\qquad \tilde R_{\ell+1}(S)=\frac{\sum_{i\in S}r_iv^{\mathrm{UCB}}_{i,\ell}}{1+\sum_{j\in S}v^{\mathrm{UCB}}_{j,\ell}}.Sℓ+1​∈argS∈Smax​R~ℓ+1​(S),R~ℓ+1​(S)=1+∑j∈S​vj,ℓUCB​∑i∈S​ri​vi,ℓUCB​​.

Let C1=72+24C_1=\sqrt{72}+\sqrt{24}C1​=72​+24​ and C2=144C_2=144C2​=144. For every epoch ℓ≥1\ell\ge1ℓ≥1, with probability at least 1−13ℓ1-\dfrac{13}{\ell}1−ℓ13​,

(1+∑j∈Sℓ+1vj)(R~ℓ+1(Sℓ+1)−R(Sℓ+1,v))≤∑i∈Sℓ+1(C1vilog⁡(Nℓ+1)Ti(ℓ)+C2log⁡(Nℓ+1)Ti(ℓ)).\Big(1+\sum_{j\in S_{\ell+1}}v_j\Big)\big(\tilde R_{\ell+1}(S_{\ell+1})-R(S_{\ell+1},\mathbf v)\big)\le\sum_{i\in S_{\ell+1}}\left(C_1\sqrt{\frac{v_i\log(\sqrt N\ell+1)}{T_i(\ell)}}+C_2\frac{\log(\sqrt N\ell+1)}{T_i(\ell)}\right).(1+j∈Sℓ+1​∑​vj​)(R~ℓ+1​(Sℓ+1​)−R(Sℓ+1​,v))≤i∈Sℓ+1​∑​​C1​Ti​(ℓ)vi​log(N​ℓ+1)​​+C2​Ti​(ℓ)log(N​ℓ+1)​​.

This is the per-epoch convergence rate of the optimistic revenue estimate to the true revenue of the offered assortment. Lemma 6.1 starts from it.

Formalization Note Three readings are disclosed:

  1. The printed right-hand side has a free index iii; its proof ((A.11)–(A.12), p. 36) gives the sum over i∈Sℓi\in S_\elli∈Sℓ​, which is stated.
  2. The printed statement writes R~ℓ(Sℓ)\tilde R_\ell(S_\ell)R~ℓ​(Sℓ​) together with Ti(ℓ)T_i(\ell)Ti​(ℓ) and log⁡(Nℓ+1)\log(\sqrt N\ell+1)log(N​ℓ+1). Algorithm 1 builds R~ℓ\tilde R_\ellR~ℓ​ from the bounds of epoch ℓ−1\ell-1ℓ−1. The statement uses the consistent reading: the estimate built at the end of epoch ℓ\ellℓ, which selects Sℓ+1S_{\ell+1}Sℓ+1​.
  3. The event is restricted to Ti(ℓ)≥1T_i(\ell)\ge1Ti​(ℓ)≥1 for every i∈Sℓ+1i\in S_{\ell+1}i∈Sℓ+1​, where the right-hand side is defined.

The probability is that of epoch ℓ\ellℓ being completed within the first TTT customers with the inequality failing, for every horizon TTT. The constants are the values from the proof of Lemma 4.1 (p. 35). The same statement appears in the companion mission on Theorem 1.

Preamble
import Mathlib
import Definitions.Def_ChoiceCDLP_MNL_mnlObjective
import Definitions.Def_MNLBandit_WellSeparated_Setting
Formal statement
namespace MNLBandit.WellSeparated

open ChoiceCDLP.MNL

/-- Lemma 4.3, p. 14, for Algorithm 1 under Assumption 4.1 and `r_i ∈ [0, 1]`: with `S = S_{ℓ+1}`
the assortment chosen from the UCBs computed at the end of epoch `ℓ ≥ 1`,
`(1 + ∑_{j∈S} v_j)(R̃_{ℓ+1}(S) − R(S, v)) ≤ ∑_{i∈S} [C₁ √(v_i log(√Nℓ+1)/T_i(ℓ)) + C₂ log(√Nℓ+1)/T_i(ℓ)]`
fails, on the event that epoch `ℓ` is completed within `T` customers and every `i ∈ S` has
`T_i(ℓ) ≥ 1`, with probability at most `13/ℓ`, for every `T`. -/
theorem lemma_4_3 (N : ℕ) (v r : Fin N → ℝ) (𝒮 : Finset (Finset (Fin N)))
    (sel : (Finset (Fin N) → ℝ) → Finset (Fin N))
    (hTU : MNLBandit.UCB.IsTU 𝒮) (hdown : MNLBandit.UCB.DownClosed 𝒮) (hv : ∀ i, 0 ≤ v i ∧ v i ≤ 1)
    (hr : ∀ i, r i ∈ Set.Icc (0 : ℝ) 1) (hsel : MNLBandit.UCB.IsArgmaxSel 𝒮 sel)
    (T ℓ : ℕ) (hℓ : 1 ≤ ℓ) :
    probT v (alg1 r sel) T
        (fun h => ℓ ≤ epochsDone r sel h ∧
          (∀ i ∈ epochSet r sel h (ℓ + 1), 1 ≤ Tcount r sel h i ℓ) ∧
          ∑ i ∈ epochSet r sel h (ℓ + 1),
              (C₁ * Real.sqrt (v i * Real.log (Real.sqrt N * ℓ + 1) / Tcount r sel h i ℓ)
                + C₂ * Real.log (Real.sqrt N * ℓ + 1) / Tcount r sel h i ℓ)
            < (1 + ∑ j ∈ epochSet r sel h (ℓ + 1), v j) *
                (mnlObjective (fun j => vUCB r sel h j ℓ) r 1 (epochSet r sel h (ℓ + 1))
                  - revenue v r (epochSet r sel h (ℓ + 1))))
      ≤ 13 / (ℓ : ℝ) := by sorry

end MNLBandit.WellSeparated
Source
Agrawal, Avadhanula, Goyal, Zeevi, MNL-Bandit: A Dynamic Learning Approach to Assortment Selection, arXiv:1706.03880v2, p. 14, Lemma 4.3 (sum over i ∈ S_ℓ from (A.11)–(A.12), p. 36; constants from p. 35)
Human review
  • Endorsed by Shuze Chen · Oct 6, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 6, 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