Theorem 1: ALSO-X is at least as good as CVaR
DisprovedJiangALSOX.Dominance.theorem_1Let be an arbitrary probability law on , a positive finite maximum of measurable random losses convex and lower semicontinuous in the decision on a common full-measure set, nonempty closed and convex inside a closed pointed convex cone, , and . Assume A1–A2 only. All random relations below hold almost surely.
Let be the value of (7): the infimum of budgets admitting a hinge-optimal pair whose slack is zero with probability at least . Let be the value of (13): the infimum of budgets admitting a CVaR-optimal triple with . Then
Both values are constructed from the original decision data and the complete lower-level argmin sets. The target concerns the exact bilevel formulations, with existential selection among tied minimizers, rather than every tie-breaking implementation of the bisection algorithm. No certificate-existence assumption, uniqueness, Slater condition, A3, finite-support specialization or tolerance appears. The two documented KKT errors are findings about the source's proof, not assumptions or repairs in this target.
Differences from the source:
- reading: A1–A3 appear in the global preamble, but section 3, Theorem 1 and Appendix A.4 explicitly invoke A1–A2 → A1–A2 only; reason: the three explicit local statements determine the claimed domain
- encoding: a measurable random vector on an arbitrary probability space, with a positive number of constraints → its probability law on a subset of finite-dimensional real space and Fin (I+1) constraint indices; reason: taking the law preserves every expectation, probability and almost-sure assertion for functions of that vector; I+1 represents every positive count
- encoding: random functions and almost-sure constraints → measurable sections and functional decisions, with one full-measure set for A1; reason: measurability makes the displayed expectations and probabilities meaningful and does not impose a moment condition
- encoding: optimal values written as minima → EReal infima of the original feasible budget sets; reason: this agrees with every finite attained value and assigns positive infinity to the empty set and negative infinity to an unbounded-below set, without assuming attainment
- encoding: E[s] and E[s]-(1-epsilon) beta in the lower-level objectives → positive extended expectation and epsilon beta plus the positive expectation of s-beta; reason: the feasibility constraints give s nonnegative or bounded below by beta almost surely, so the formulas agree even at positive infinity
- reading: the pair or triple in (7b) or (13b) is an optimization variable → existential membership in the full lower-level argmin set; reason: this is the literal optimistic bilevel formulation; the source separately discusses unlucky algorithmic choices in Appendix C
import Definitions.Def_JiangALSOX_Dominance_Programs
namespace JiangALSOX.Dominance
theorem theorem_1 {n m I : ℕ} (D : Model n m I) (hD : Assumptions D) :
alsoValue D ≤ cvarValue D := by sorry
end JiangALSOX.Dominance
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.