Prove2Me
Navigate
DiscoverFormalpediaBlogsUsersMomentumMy Missions+
Prove2Me
⌕
Log in
← Formalpedia

Theorem 1 — Exact lead-time bound and the 7/37/37/3 guarantee

Proved
CappedBaseStock.main

by StellaXin · Sep 8, 2026 · Mathlib 0df444a (Lean v4.33.1)

approximationcapped-base-stockinventorylost-salesoperations-research

Consider a periodic-review lost-sales inventory system. Demand is i.i.d., nonnegative and real-valued, with 0<μ=E[D]<∞0<\mu=\mathbb E[D]<\infty0<μ=E[D]<∞. Lead time is an integer L≥1L\ge1L≥1, and holding and lost-sales rates satisfy h,p>0h,p>0h,p>0. Inventory and the LLL pipeline coordinates initially equal zero. The first pipeline component arrives, the order is placed before current demand is observed, demand is served up to available inventory, and the remaining pipeline shifts; today's order arrives LLL periods later. The period cost is hIt+1+p(Dt−It−x1,t)+hI_{t+1}+p(D_t-I_t-x_{1,t})^+hIt+1​+p(Dt​−It​−x1,t​)+.

Costs use the upper limit of finite-horizon expected average costs from this empty initial state. OPT\mathrm{OPT}OPT is the infimum over all measurable, possibly time-dependent history policies in the canonical demand-path model, with independent uniform private randomization. Infinite expected costs remain infinite. A capped base-stock rule orders min⁡{(S−It−∑ixi,t)+,r}\min\{(S-I_t-\sum_i x_{i,t})^+,r\}min{(S−It​−∑i​xi,t​)+,r}, with finite S,r≥0S,r\ge0S,r≥0. Its cost is denoted by C(πS,r)C(\pi_{S,r})C(πS,r​), and CCBS∗C^*_{\rm CBS}CCBS∗​ is the infimum over these parameters. Ordinary base stock at level SSS is the same rule with cap SSS.

For a demand block, let Irm=max⁡0≤k≤m∑i=1k(r−Di)I_r^m=\max_{0\le k\le m}\sum_{i=1}^k(r-D_i)Irm​=max0≤k≤m​∑i=1k​(r−Di​), including the zero empty sum. A pair (r,z)(r,z)(r,z) is feasible when 0≤r≤μ0\le r\le\mu0≤r≤μ, z≥0z\ge0z≥0, and, for both m=Lm=Lm=L and m=L+1m=L+1m=L+1,

E[(Irm+∑i=1m(Di−r)−z)+]≤m(μ−r).\mathbb E\left[\left(I_r^m+\sum_{i=1}^m(D_i-r)-z\right)^+\right]\le m(\mu-r).E[(Irm​+i=1∑m​(Di​−r)−z)+]≤m(μ−r).

The lower certificate C‾\underline CC​ is the infimum of hz+p(μ−r)hz+p(\mu-r)hz+p(μ−r) over this feasible set. Neither attainment of the infimum nor positivity of OPT\mathrm{OPT}OPT is assumed. Let κL=1+4L2/((L+1)(3L−1))\kappa_L=1+4L^2/((L+1)(3L-1))κL​=1+4L2/((L+1)(3L−1)). Claim.

CCBS∗≤κLC‾,C^*_{\rm CBS}\le\kappa_L\underline C,CCBS∗​≤κL​C​,

and, as consequences,

CCBS∗≤κLOPT≤73OPT.C^*_{\rm CBS}\le\kappa_L\mathrm{OPT}\le\frac73\mathrm{OPT}.CCBS∗​≤κL​OPT≤37​OPT.

These are multiplicative inequalities, including possible zero optimal cost. The statement does not require a minimizing feasible pair or an attained best capped policy. Its only assumptions are the demand-law, positive-cost, and integer-lead-time assumptions of the model; it does not assume the lower bound, either policy comparison, or existence of an optimal stationary realization.

Preamble
import Definitions.Def_CappedBaseStock_Model



open scoped ENNReal
Formal statement
namespace CappedBaseStock


theorem main (P : DemandLaw) (c : Parameters) :
    CbsOpt P c ≤ ENNReal.ofReal (kappa c.L) * lowerCertificate P c ∧
    CbsOpt P c ≤ ENNReal.ofReal (kappa c.L) * OPT P c ∧
    ENNReal.ofReal (kappa c.L) * OPT P c ≤ (7 / 3 : ℝ≥0∞) * OPT P c := by sorry

end CappedBaseStock
Source
Linwei Xin, Capped Base-Stock Policies: A 2.33-Approximation, author-supplied LaTeX manuscript (756 lines); public paper listing https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7134538 . Author-supplied source SHA-256: f353793c255e1ebed5f3ec541037284bd926183e3e5b71941f13e79c2d67cb7a. Theorem 1 (`thm-main`), equation `eq-main-bound` and consequences, lines 694–704.
Read-back

What the Lean code literally says, in plain math · GPT-6 (independent Codex sub-agent)

For every probability measure PPP on the nonnegative real numbers such that d↦dd\mapsto dd↦d is integrable and its real mean μ=∫d P(dd)\mu=\int d\,P(\mathrm{d}d)μ=∫dP(dd) is strictly positive, and every integer L≥1L\ge 1L≥1 and real numbers h>0h>0h>0 and p>0p>0p>0, the following three inequalities hold. To specify their quantities, let (Dt)t≥0(D_t)_{t\ge0}(Dt​)t≥0​ have the infinite product law P⊗NP^{\otimes\mathbb N}P⊗N and let UUU be independent of that sequence with law equal to Lebesgue measure restricted to the closed interval [0,1][0,1][0,1]. An admissible policy consists of a measurable map πt:[0,∞)t×R→[0,∞)\pi_t:[0,\infty)^t\times\mathbb R\to[0,\infty)πt​:[0,∞)t×R→[0,∞) for each integer t≥0t\ge0t≥0; its order at time ttt is qt=πt((Di)0≤i<t,U)q_t=\pi_t((D_i)_{0\le i<t},U)qt​=πt​((Di​)0≤i<t​,U), so the history at t=0t=0t=0 has no demand coordinates. All orders and state coordinates are finite nonnegative real numbers, with no boundedness or integrability condition on orders beyond measurability. Starting with I0=0I_0=0I0​=0 and x0,j=0x_{0,j}=0x0,j​=0 for 0≤j<L0\le j<L0≤j<L, define It+1=(It+xt,0−Dt)+I_{t+1}=(I_t+x_{t,0}-D_t)^+It+1​=(It​+xt,0​−Dt​)+, xt+1,j=xt,j+1x_{t+1,j}=x_{t,j+1}xt+1,j​=xt,j+1​ for 0≤j<L−10\le j<L-10≤j<L−1, and xt+1,L−1=qtx_{t+1,L-1}=q_txt+1,L−1​=qt​, where y+=max⁡(y,0)y^+=\max(y,0)y+=max(y,0). The period cost is Ct=hIt+1+p(Dt−It−xt,0)+C_t=hI_{t+1}+p(D_t-I_t-x_{t,0})^+Ct​=hIt+1​+p(Dt​−It​−xt,0​)+, viewed in the extended nonnegative reals [0,∞][0,\infty][0,∞], and the cost of the policy is J(π)=lim sup⁡T→∞1T+1∑t=0T∫Ct d(P⊗N⊗Leb∣[0,1])J(\pi)=\limsup_{T\to\infty}\frac{1}{T+1}\sum_{t=0}^{T}\int C_t\,\mathrm d(P^{\otimes\mathbb N}\otimes\mathrm{Leb}|_{[0,1]})J(π)=limsupT→∞​T+11​∑t=0T​∫Ct​d(P⊗N⊗Leb∣[0,1]​), with TTT ranging over nonnegative integers and the integrals being nonnegative Lebesgue integrals, possibly infinite. Put O=inf⁡πJ(π)O=\inf_{\pi}J(\pi)O=infπ​J(π) over all such admissible policies. For any finite S,a≥0S,a\ge0S,a≥0, let JS,aJ_{S,a}JS,a​ be the same average cost for the recursion with qt=min⁡{(S−It−∑j=0L−1xt,j)+,a}q_t=\min\{(S-I_t-\sum_{j=0}^{L-1}x_{t,j})^+,a\}qt​=min{(S−It​−∑j=0L−1​xt,j​)+,a}, and put B=inf⁡S≥0inf⁡a≥0JS,aB=\inf_{S\ge0}\inf_{a\ge0}J_{S,a}B=infS≥0​infa≥0​JS,a​; zero values of either parameter are included. For a demand sequence d=(di)i≥0d=(d_i)_{i\ge0}d=(di​)i≥0​, real numbers r,zr,zr,z, and an integer m≥0m\ge0m≥0, set Fm(d,r)=max⁡0≤k≤m∑i=0k−1(r−di)F_m(d,r)=\max_{0\le k\le m}\sum_{i=0}^{k-1}(r-d_i)Fm​(d,r)=max0≤k≤m​∑i=0k−1​(r−di​), including the empty sum 000 at k=0k=0k=0, and Em(d,r,z)=(Fm(d,r)+∑i=0m−1(di−r)−z)+E_m(d,r,z)=(F_m(d,r)+\sum_{i=0}^{m-1}(d_i-r)-z)^+Em​(d,r,z)=(Fm​(d,r)+∑i=0m−1​(di​−r)−z)+, viewed in [0,∞][0,\infty][0,∞]. A pair (r,z)∈R2(r,z)\in\mathbb R^2(r,z)∈R2 is feasible precisely when 0≤r≤μ0\le r\le\mu0≤r≤μ, z≥0z\ge0z≥0, and ∫Em(d,r,z) P⊗N(dd)≤m(μ−r)\int E_m(d,r,z)\,P^{\otimes\mathbb N}(\mathrm dd)\le m(\mu-r)∫Em​(d,r,z)P⊗N(dd)≤m(μ−r) for each of m=Lm=Lm=L and m=L+1m=L+1m=L+1, where each integral is a nonnegative Lebesgue integral and the finite nonnegative right side is embedded into [0,∞][0,\infty][0,∞]. Put A=inf⁡(r,z) feasible[hz+p(μ−r)]A=\inf_{(r,z)\text{ feasible}}[hz+p(\mu-r)]A=inf(r,z) feasible​[hz+p(μ−r)], again as an infimum in [0,∞][0,\infty][0,∞], and K=1+4L2(L+1)(3L−1)K=1+\frac{4L^2}{(L+1)(3L-1)}K=1+(L+1)(3L−1)4L2​. The assertion is B≤KA ∧ B≤KO ∧ KO≤73OB\le KA\ \land\ B\le KO\ \land\ KO\le\frac73OB≤KA ∧ B≤KO ∧ KO≤37​O. All costs, infima, inequalities and products in this assertion use the extended nonnegative reals; the displayed finite nonnegative coefficients are embedded in that space, an empty infimum is +∞+\infty+∞, and 0⋅(+∞)=00\cdot(+\infty)=00⋅(+∞)=0. The infima defining AAA, BBB and OOO assert no attainment by any pair of parameters or policy. The assumption L≥1L\ge1L≥1 includes L=1L=1L=1 and keeps the displayed denominators nonzero; demands may equal zero even though their mean is strictly positive.

Human review
  • Endorsed by Shuze Chen · Sep 8, 2026

  • Endorsed by StellaXin · Sep 8, 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 paper
SKILL.mdTourFAQContactTerms
© 2026 Prove2Me