Proposition 2 (corrected: Eq. (22) includes the duals of y ≤ Y) — D(x; d) = E u(x, Y; d)
ProvedRLPBidPrice.Directional.directional_derivative_eq_expected_uLet be a 0/1 leg–itinerary incidence matrix with nonzero columns, the fares, and a capacity vector with every component . Let be a measurable random demand vector on a probability space with almost surely, and let , where . For every direction the one-sided directional derivative of at exists and
where
is the least value of over the optimal vectors of dual prices of the capacity constraints.
No condition on the demand distribution is assumed: the result covers the nondifferentiable case, where the dual prices are not unique with positive probability and the averaged dual prices only estimate a supergradient. Averaging over simulated demands gives an unbiased estimator of the directional derivative.
Formalization Note. The limit is one-sided, (nhdsWithin 0 (Set.Ioi 0)); the paper's Eq. (12) writes , whose two-sided version fails in the nondifferentiable case. is implicit in the paper (it differentiates at a capacity vector; at a zero capacity a direction leaving makes the LP infeasible). is a real sInf; for and the optimal dual price set is nonempty and bounded, so is an attained minimum and bounded uniformly in . No integrability of is needed: is bounded by because every itinerary uses a leg. The time index is dropped.
Correction. The paper defines by Eq. (22), subject to , , , which omits the dual variables of ; with that the proposition is false (one leg, one itinerary, , , uniform on , : but the printed (22) gives ). Here is taken over the optimal dual prices of the Lagrangian problem (14).
import Mathlib import Definitions.Def_RLPBidPrice_Directional_Model open MeasureTheory Matrix
namespace RLPBidPrice.Directional
/-- Proposition 2, p. 212, with Eq. (22) corrected: at a capacity vector `x > 0` the one-sided
directional derivative `D(x; d)` of `J^PI` exists for every direction `d` and equals
`E u(x, Y; d)`. -/
theorem directional_derivative_eq_expected_u {m n : ℕ} {Ω : Type*} [MeasurableSpace Ω]
(P : Measure Ω) [IsProbabilityMeasure P]
(A : Matrix (Fin m) (Fin n) ℝ) (r : Fin n → ℝ) (hA : IsIncidence A) (hr : 0 ≤ r)
(x : Fin m → ℝ) (hx : ∀ i, 0 < x i)
(Y : Ω → Fin n → ℝ) (hYm : Measurable Y) (hY : ∀ᵐ ω ∂P, 0 ≤ Y ω)
(d : Fin m → ℝ) :
Filter.Tendsto (fun h : ℝ => h⁻¹ * (JPI P A r Y (x + h • d) - JPI P A r Y x))
(nhdsWithin 0 (Set.Ioi 0)) (nhds (∫ ω, u A r x (Y ω) d ∂P)) := by sorry
end RLPBidPrice.Directional
Read-back
What the Lean code literally says, in plain math · claude-opus-5-5
Fix natural numbers and . Let be a measurable space with a probability measure on it. The statement also takes:
- a real matrix that satisfies the predicate ;
- a vector with for every ;
- a vector with for every (strictly positive in every coordinate);
- a measurable map such that, for -almost every , every coordinate of is ;
- an arbitrary direction , with no sign or size restriction.
and two other objects come from the imported module Def_RLPBidPrice_Directional_Model, and their definitions are not part of the code shown here:
- , a real-valued function of a capacity vector. It depends on , , and .
- , a real number for a vector , a vector and a direction .
This read-back cannot unfold these three definitions. Whatever they contain carries over into the statement's meaning unchanged. The same holds for any conventions they use, such as integrals, suprema, or the meaning of "incidence".
Conclusion. The difference quotient of at in direction converges as tends to from the right ( only):
This claims two things: the one-sided limit exists, and it equals the stated value. The statement makes no claim about the two-sided limit or about . The right-hand side is a Bochner integral. No hypothesis says that is integrable, and no hypothesis says that is integrable. If that function is not integrable (including when it is not measurable), the right-hand side is by convention , and the claim becomes that the quotient tends to .
Degenerate cases.
- . It is excluded by the right-sided limit, so the convention plays no role.
- . The only vector is the empty one, and the positivity hypothesis on holds vacuously. The only direction is = the empty vector, so and the quotient is identically . The statement then asserts exactly that .
- (any ). The quotient is identically , so the statement asserts .
- . and are the empty vector, and their hypotheses hold vacuously. The content then depends entirely on the imported definitions.
- . It is necessarily nonempty, because is a probability measure. It may have a single point, in which case is a single deterministic vector with nonnegative entries.
- Hypotheses on . Whether holds for any particular , including when or is , is decided by the imported definition. This read-back cannot say whether it is ever satisfiable.
- Imported definitions. If is itself defined through integrals or suprema, their default-value conventions apply inside it: for a non-integrable integrand, and junk values for empty or unbounded suprema. None of the hypotheses above rule those situations out.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.