Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 8 — for any γ, Λ₂(T) ≤ (max{b,h}ȳ/T){⌈(m/ρ′)²⌉ + E[K(ρ′, 2√2m²/ρ′)]} + 2hm E[|J₁(ρ′)|²]/√T

Proved
AIMInventory.UnknownRho.theorem_8

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

asymptotic-analysiscensored-demandinventoryp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1queueingstochastic-gradient

Consider a lost-sales inventory system with i.i.d. nonnegative demands D1,D2,…D_1,D_2,\dotsD1​,D2​,… of finite positive mean, holding cost h>0h>0h>0 and penalty b>0b>0b>0, and expected one-period cost Q(y)=h E[y−D1]++b E[D1−y]+Q(y)=h\,E[y-D_1]^+ + b\,E[D_1-y]^+Q(y)=hE[y−D1​]++bE[D1​−y]+. Excess inventory is non-perishable. The manager knows an upper bound yˉ>0\bar y>0yˉ​>0 on the newsvendor quantity, yNV=inf⁡{y≥0:P(D1≤y)≥b/(b+h)}≤yˉy^{NV}=\inf\{y\ge0: P(D_1\le y)\ge b/(b+h)\}\le\bar yyNV=inf{y≥0:P(D1​≤y)≥b/(b+h)}≤yˉ​, and runs the AIM algorithm with an arbitrary step scale γ>0\gamma>0γ>0 and any initial target y^1∈[0,yˉ]\hat y_1\in[0,\bar y]y^​1​∈[0,yˉ​], producing targets y^t\hat y_ty^​t​ and implemented levels yt=max⁡{y^t,xt}y_t=\max\{\hat y_t,x_t\}yt​=max{y^​t​,xt​}. Let

m=hγyˉmax⁡{b,h},Λ2(T)=E[1T∑t=1T(Q(yt)−Q(y^t))].m=\frac{h\gamma\bar y}{\max\{b,h\}},\qquad \Lambda_2(T)=E\Big[\frac1T\sum_{t=1}^T\big(Q(y_t)-Q(\hat y_t)\big)\Big].m=max{b,h}hγyˉ​​,Λ2​(T)=E[T1​t=1∑T​(Q(yt​)−Q(y^​t​))].

Let K(θ,z)=min⁡{t≥0:z+∑s=1t(θ−Ds)≤0}K(\theta,z)=\min\{t\ge0: z+\sum_{s=1}^t(\theta-D_s)\le0\}K(θ,z)=min{t≥0:z+∑s=1t​(θ−Ds​)≤0} be the hitting time of (−∞,0](-\infty,0](−∞,0] by the random walk started at zzz with increments θ−Ds\theta-D_sθ−Ds​, and let ∣J1(θ)∣|J_1(\theta)|∣J1​(θ)∣ be the length of the first busy period of the queue W1(θ)=0W_1(\theta)=0W1​(θ)=0, Wt+1(θ)=[Wt(θ)+θ−Dt]+W_{t+1}(\theta)=[W_t(\theta)+\theta-D_t]^+Wt+1​(θ)=[Wt​(θ)+θ−Dt​]+.

Then for every T≥1T\ge1T≥1 and every ρ′\rho'ρ′ with 0<ρ′<min⁡{m,E[D1]}0<\rho'<\min\{m,E[D_1]\}0<ρ′<min{m,E[D1​]},

Λ2(T)≤max⁡{b,h} yˉT{⌈(m/ρ′)2⌉+E[K(ρ′,22 m2ρ′)]}+2hm E[∣J1(ρ′)∣2]T.\Lambda_2(T)\le\frac{\max\{b,h\}\,\bar y}{T}\Big\{\big\lceil(m/\rho')^2\big\rceil+E\Big[K\Big(\rho',\frac{2\sqrt2\,m^2}{\rho'}\Big)\Big]\Big\}+\frac{2hm\,E\big[|J_1(\rho')|^2\big]}{\sqrt T}.Λ2​(T)≤Tmax{b,h}yˉ​​{⌈(m/ρ′)2⌉+E[K(ρ′,ρ′22​m2​)]}+T​2hmE[∣J1​(ρ′)∣2]​.

The cost of AIM splits as Λ1(T)+Λ2(T)\Lambda_1(T)+\Lambda_2(T)Λ1​(T)+Λ2​(T), where Λ1(T)\Lambda_1(T)Λ1​(T) is controlled by stochastic subgradient descent for every γ\gammaγ. Theorem 8 controls the remaining term without the restriction γ≤ρmax⁡{b,h}/(hyˉ)\gamma\le\rho\max\{b,h\}/(h\bar y)γ≤ρmax{b,h}/(hyˉ​) of Theorem 2, so the manager does not need a lower bound on mean demand: whenever E[K(⋅,⋅)]E[K(\cdot,\cdot)]E[K(⋅,⋅)] and E[∣J1(⋅)∣2]E[|J_1(\cdot)|^2]E[∣J1​(⋅)∣2] are finite, the bound is O(1/T)O(1/\sqrt T)O(1/T​).

Formalization Note. The right side is computed in [0,∞][0,\infty][0,∞]: E[K]E[K]E[K] and E[∣J1∣2]E[|J_1|^2]E[∣J1​∣2] are Lebesgue integrals that may be ∞\infty∞, in which case the bound is trivially true rather than vacuous through a junk value; the left side Λ2(T)\Lambda_2(T)Λ2​(T) is a real Bochner integral of a bounded integrand, embedded with ENNReal.ofReal (a negative Λ2(T)\Lambda_2(T)Λ2​(T) embeds as 000, consistent with the real inequality). The hypotheses ρ′>0\rho'>0ρ′>0, h,b,yˉ,γ>0h,b,\bar y,\gamma>0h,b,yˉ​,γ>0 are implicit in the paper ("for any choice of the parameter γ\gammaγ" means any γ>0\gamma>0γ>0, as in (4)) and are added explicitly. The newsvendor bound yNV≤yˉy^{NV}\le\bar yyNV≤yˉ​ is the theorem's standing hypothesis and is kept although the bound does not use it.

Preamble
import Mathlib
import Definitions.Def_AIMInventory_UnknownRho_Model

noncomputable section
open MeasureTheory
open scoped ENNReal NNReal
Formal statement
namespace AIMInventory.UnknownRho

/-- Theorem 8, p. 13: the non-perishable AIM excess-cost bound for any
positive step scale `γ`. The right side is in `ℝ≥0∞`, so an infinite
hitting-time or busy-period expectation makes the bound infinite rather than
junk. `0 < ρ′` is added (the printed bound divides by `ρ′`). -/
theorem theorem_8 (P : DemandLaw) (c : Parameters) (ρ' : ℝ) (T : ℕ)
    (hNV : LostSalesOrderUpTo.Asymptotic.newsvendorLevel P.law c.h c.b ≤
      Real.toNNReal c.barY)
    (hT : 1 ≤ T) (hρ : 0 < ρ') (hsmall : ρ' < m c)
    (hmean : ρ' < CappedBaseStock.mean P) :
    ENNReal.ofReal (lambdaTwo P c T) ≤
      ENNReal.ofReal (max c.b c.h * c.barY / (T : ℝ)) *
        ((Nat.ceil ((m c / ρ') ^ 2) : ℝ≥0∞) + expectedK P ρ' (initialHeight c ρ')) +
      ENNReal.ofReal (2 * c.h * m c / Real.sqrt (T : ℝ)) *
        busySecondMoment P ρ' := by sorry

end AIMInventory.UnknownRho
Source
Huh & Rusmevichientong, A Non-Parametric Asymptotic Analysis of Inventory Planning with Censored Demand, Math. Oper. Res. (2009), author's manuscript, p. 13, Theorem 8
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