Theorem 1: the Lookahead Peak-Shaving policy is robustly optimal (with a no-end-effect condition)
OpenRobustVolumeFlex.lookahead_peak_shaving_optimalTheorem 1 of Gijsbrechts et al. (2023): the Lookahead Peak-Shaving policy is robustly optimal, under one extra no-end-effect hypothesis.
Consider one period of the robust rolling-horizon model with volume flexibility, with data as in the definition RobustVolumeFlex_Model. Let be the lookahead horizon. Assume:
- , and for all ;
- is nonempty and bounded, and every has ;
- (Assumption 1) ;
- (Assumption 2) and ;
- (no end-of-horizon effect) .
Then the linear program (4) has an optimal solution whose current-period order is the Lookahead Peak-Shaving order:
where is the uncapacitated base stock and the shifting need.
The theorem reduces the robust LP to an explicit two-threshold base-stock rule, the same structure as the stochastic optimum. It underlies the paper's closed-form policy parameters and sensitivity analysis.
Formalization Note Hypothesis 5 is not in the paper. Without it the statement fails: , , , , , , , deterministic demand satisfies Assumptions 1–2, but the LP optimum orders 1 unit (cost 59), while LPS orders 10 (best follow-up cost 63.5). Hypothesis 1 and the conditions on are the model's standing conventions. The conclusion is existential: ties may allow other optimal first-period orders.
import Mathlib import Definitions.Def_RobustVolumeFlex_Model
namespace RobustVolumeFlex
theorem lookahead_peak_shaving_optimal {H : ℕ} [NeZero H] (P : Params H)
(h_pos : 0 < P.h) (c_nonneg : 0 ≤ P.c) (k_nonneg : ∀ n, 0 ≤ P.k n)
(Ω_nonempty : P.Ω.Nonempty) (Ω_bounded : Bornology.IsBounded P.Ω)
(demand_nonneg : ∀ d ∈ P.Ω, ∀ i, 0 ≤ d i)
(assumption1 : P.h < P.cp ∧ P.cp < P.b)
(assumption2 : ⌊P.cp / P.h⌋₊ < H ∧ ⌊P.c / P.b⌋₊ < H)
(no_end_effect : P.c + P.cp ≤ P.b * ((H : ℝ) - ⌊P.cp / P.h⌋₊)) :
∃ q y z : Fin H → ℝ, IsRobustlyOptimal P q y z ∧ q 0 = lpsOrder P ⌊P.cp / P.h⌋₊ := by
sorry
end RobustVolumeFlex
Read-back
What the Lean code literally says, in plain math · claude-opus-5-5
Setting. Fix a horizon length with ; this is the typeclass assumption, so the periods are . Fix a parameter bundle made of:
- real numbers (holding), (backorder), (unit cost), (premium) and (starting inventory);
- a vector ;
- a set .
Hypotheses.
- .
- .
- for every .
- .
- is bounded.
- Every has for all .
- .
- Write , the natural-number floor of the real quotient. The hypothesis requires and .
- The "no end effect" condition, with and read as reals:
No hypothesis is placed on , which may be any real number, including a negative one. There is also no hypothesis linking to or to .
From these, , so , and therefore . Also and . The hypotheses can all hold at once; for example , , , , gives .
Definitions, expanded.
- Partial sums. For and ,
where ranges over . For this is the full sum.
- Demand bounds.
Because is nonempty and bounded, these are a genuine real infimum and supremum. They need not be attained, since is not assumed closed. Both are .
- Feasibility. A triple of vectors in is feasible if, for every :
There is no sign constraint on .
- Objective.
- Robust optimality. is robustly optimal if it is feasible and its objective is the objective of every feasible triple. That is, it is a global minimizer of this linear program.
- Base-stock level. For ,
The denominator is positive.
- Uncapacitated order. For :
The comparison is between reals.
- Shifting need. For , is defined as a fold of that starts from :
If , then . Here , so the cap never truncates.
- Lookahead peak-shaving order.
The inequalities are strict, so each boundary point falls into the later case. Note that .
Conclusion. The statement is existential, not universal. There exist vectors such that is robustly optimal in the sense above and
The statement does not claim that every optimal solution has this first-period order, and it says nothing about , or beyond optimality. Proving it does, however, require that an optimal solution exists.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.