Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 1, p. 18 — DIP has E(R_T) ≤ C T^{2/3} log T + 4p_max L Σ_k 2^{k−2}ℓ₂ E‖θ̂_{k−1} − θ₀‖₁ (Õ pinned from the proof, pp. 54–55)

Proved
LuoSunLiu.DIP.theorem_1

by mikedeng1 · Oct 9, 2026 · Mathlib 0df444a (Lean v4.33.1)

contextual-pricingdynamic-pricinglinear-banditsp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1regret

This is the main regret bound of the DIP policy.

Let the standing assumptions of the contextual pricing model hold (∥θ0∥1≤W\|\theta_0\|_1 \le W∥θ0​∥1​≤W, pmax⁡>0p_{\max} > 0pmax​>0, ∥x∥∞≤1\|x\|_\infty \le 1∥x∥∞​≤1 on X\mathcal XX, and p∗(x)∈(0,pmax⁡)p^*(x) \in (0, p_{\max})p∗(x)∈(0,pmax​) maximizes the expected revenue p(1−F(p−x⊤θ0))p(1 - F(p - x^\top\theta_0))p(1−F(p−x⊤θ0​)) over p>0p > 0p>0), let FFF be LLL-Lipschitz (Assumption 1), and let Assumption 2 hold. Fix DIP inputs α1,α2≥1\alpha_1, \alpha_2 \ge 1α1​,α2​≥1 (so ℓ1=α1\ell_1 = \alpha_1ℓ1​=α1​, ℓ2=α2\ell_2 = \alpha_2ℓ2​=α2​), a discretization constant CdC_dCd​ with pmax⁡+2W<Cd pmax⁡p_{\max} + 2W < C_d\,p_{\max}pmax​+2W<Cd​pmax​, and λ>0\lambda > 0λ>0.

Then there is a constant C>0C > 0C>0 with the following property. On any probability space with a filtration, for any adapted covariates, i.i.d. noise with CDF FFF, any admissible Inner Algorithm A (a measurable estimator of the previous episode's data with values in {∥θ∥1≤W}\{\|\theta\|_1 \le W\}{∥θ∥1​≤W}) and any run of the DIP policy, for every horizon T≥2T \ge 2T≥2,

E(RT)≤C T2/3log⁡T+4pmax⁡L∑k=2n2k−2ℓ2 E∥θ^k−1−θ0∥1,\mathbb E(R_T) \le C\,T^{2/3}\log T + 4p_{\max}L\sum_{k=2}^{n}2^{k-2}\ell_2\,\mathbb E\|\hat\theta_{k-1} - \theta_0\|_1,E(RT​)≤CT2/3logT+4pmax​Lk=2∑n​2k−2ℓ2​E∥θ^k−1​−θ0​∥1​,

where RT=∑t=1TrtR_T = \sum_{t=1}^T r_tRT​=∑t=1T​rt​ is the cumulative regret (1) and n=n(T,α1,α2)n = n(T, \alpha_1, \alpha_2)n=n(T,α1​,α2​) is the number of episodes up to TTT.

The first term is the price of learning the unknown noise distribution by discretized UCB pricing; the second isolates the contribution of the estimation errors of θ0\theta_0θ0​, so that any estimator with a known ℓ1\ell_1ℓ1​ error rate yields an explicit regret rate for DIP.

Formalization Note The printed O~(T2/3)\tilde O(T^{2/3})O~(T2/3) is pinned to C T2/3log⁡TC\,T^{2/3}\log TCT2/3logT for T≥2T \ge 2T≥2: the proof (p. 55) gives ℓ1pmax⁡+C3′C4log⁡(2C2′T)T2/3\ell_1p_{\max} + C'_3C_4\log(2C'_2T)T^{2/3}ℓ1​pmax​+C3′​C4​log(2C2′​T)T2/3, which this absorbs. The constant CCC depends only on the model, LLL, the Assumption 2 constant, α1,α2,Cd\alpha_1, \alpha_2, C_dα1​,α2​,Cd​ and λ\lambdaλ, and is chosen before the probability space, the processes, the estimator and TTT; this is what the proof supports and is stronger than a constant chosen per run. Inner Algorithm A is any admissible estimator, with the paper's Algorithm 2 (logistic regression, then projection onto {∥θ∥1≤W}\{\|\theta\|_1 \le W\}{∥θ∥1​≤W}) as one instance. Algorithm 3 is read with forced exploration of unpulled arms (Lemma 2's convention). The grid hypothesis pmax⁡+2W<Cd pmax⁡p_{\max} + 2W < C_d\,p_{\max}pmax​+2W<Cd​pmax​ makes every candidate price set nonempty, without which the policy is undefined. The noise is independent of the past and the covariates are predictable for a general filtration, which needs a standard Borel sample space. For the last, truncated episode the weight is 2n−2ℓ22^{n-2}\ell_22n−2ℓ2​, as printed. Expectations are lower Lebesgue integrals of nonnegative quantities.

Preamble
import Mathlib
import Definitions.Def_LuoSunLiu_DIP_Model
import Definitions.Def_LuoSunLiu_DIP_InnerB
import Definitions.Def_LuoSunLiu_DIP_Policy

open MeasureTheory ProbabilityTheory NNReal ENNReal
Formal statement
namespace LuoSunLiu.DIP

/-- Theorem 1 (Luo, Sun and Liu, arXiv:2109.07340v2, p. 18; proof pp. 54–55), with the `Õ(T^{2/3})`
pinned to `C T^{2/3} log T`. Under the standing assumptions of §2.1 and Assumptions 1–2, for DIP
inputs `α₁, α₂ ≥ 1`, `C` (here `Cdisc`) with `p_max + 2W < C p_max`, and `λ > 0`, there is a
constant `C₀ > 0` such that for every probability space, filtration, adapted covariates, i.i.d.
noise, admissible estimator (Inner Algorithm A) and every run of the DIP policy, and every
`T ≥ 2`,
`E(R_T) ≤ C₀ T^{2/3} log T + 4p_max L ∑_{k=2}^{n} 2^{k-2}ℓ₂ E‖θ̂_{k-1} - θ₀‖₁`,
where `n = n(T, α₁, α₂)` is the number of episodes up to `T` and `ℓ₂ = α₂`. -/
theorem theorem_1 {d0 : ℕ} (M : PricingModel d0) (hM : M.Standing) (L : ℝ≥0)
    (hL : LipschitzWith L M.F) (CA : ℝ) (hA2 : Assumption2 M CA)
    (α1 α2 Cdisc : ℕ) (hα1 : 1 ≤ α1) (hα2 : 1 ≤ α2)
    (hCdisc : M.pmax + 2 * M.W < Cdisc * M.pmax) (lam : ℝ) (hlam : 0 < lam) :
    ∃ C : ℝ, 0 < C ∧
      ∀ (Ω : Type*) (mΩ : MeasurableSpace Ω) [StandardBorelSpace Ω]
        (P : Measure Ω) [IsProbabilityMeasure P] (𝒢 : Filtration ℕ mΩ)
        (x : ℕ → Ω → Fin d0 → ℝ) (z p : ℕ → Ω → ℝ)
        (est : (k : ℕ) → (Fin (epLen α1 α2 k) → (Fin d0 → ℝ) × ℝ × ℝ) → (Fin d0 → ℝ)),
        IsPricingEnv M P 𝒢 x z p → IsAdmissibleEstimator M α1 α2 est →
        (∀ ω, IsDIPRun M α1 α2 Cdisc lam est x z p ω) →
        ∀ T : ℕ, 2 ≤ T →
          ∫⁻ ω, ENNReal.ofReal (cumRegret M x p T ω) ∂P ≤
            ENNReal.ofReal (C * (T : ℝ) ^ ((2 : ℝ) / 3) * Real.log T) +
              ENNReal.ofReal (4 * M.pmax * L) *
                ∑ k ∈ Finset.Icc 2 (nEp α1 α2 T),
                  ((2 ^ (k - 2) * α2 : ℕ) : ℝ≥0∞) *
                    ∫⁻ ω, ENNReal.ofReal
                      (l1 (thetaHat M α1 α2 est x z p (k - 1) ω - M.θ0)) ∂P := by sorry

end LuoSunLiu.DIP
Source
Luo, Sun and Liu, arXiv:2109.07340v2, p. 18, Theorem 1; proof pp. 54–55
Human review
  • Endorsed by Shuze Chen · Oct 10, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 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