Theorem 5.3 — mean–Gini optimal solutions exist and are SSD-efficient
ProvedDualSSD.MeanRisk.meanGini_optimal_exists_ssdEfficientLet and let be a nonempty, convex, bounded and closed set. Consider the mean–risk problem with the Gini mean difference as risk functional,
Then for every :
- the set of optimal solutions of (5.1) is nonempty, and
- each optimal solution is SSD-efficient in : there is no with .
This is the main result of the mission: the mean–Gini model with trade-off never selects a portfolio that some feasible portfolio dominates in the second-degree sense, so it is consistent with risk-averse preferences.
Formalization Note Nonemptiness of is not written in the paper and is added: without it no optimal solution exists. The paper's is a real number, so is part of its hypothesis. The trade-off coefficient is named lam because λ is a Lean keyword. Optimal solutions are stated as maximizers in , not through a supremum value.
import Mathlib import Definitions.Def_DualSSD_MeanRisk_gini import Definitions.Def_DualSSD_MeanRisk_ssdEfficient
namespace DualSSD.MeanRisk
open MeasureTheory
/-- **Theorem 5.3** (Ogryczak–Ruszczyński 2002, p. 73). Let `Q ⊆ L_q(Ω, P)`, `1 < q < ∞`, be
convex, bounded, closed and nonempty, and let `lam ∈ (0, 1]` (the paper's `λ`; `λ` is a Lean
keyword). Then the mean–Gini problem (5.1) `max_{X ∈ Q} (μ_X − lam Γ_X)` has an optimal solution
in `Q`, and every optimal solution is SSD-efficient in `Q` (no `Y ∈ Q` with `Y ≻_SSD X`).
`Q.Nonempty` is not on the page: the paper assumes it silently, and without it no optimal
solution exists. -/
theorem meanGini_optimal_exists_ssdEfficient {Ω : Type*} [MeasurableSpace Ω] (P : Measure Ω)
[IsProbabilityMeasure P] (q : ENNReal) [Fact (1 ≤ q)] (hq : 1 < q) (hqtop : q ≠ ⊤)
(Q : Set (Lp ℝ q P)) (hQconv : Convex ℝ Q) (hQbdd : Bornology.IsBounded Q)
(hQclosed : IsClosed Q) (hQne : Q.Nonempty) (lam : ℝ) (hlam0 : 0 < lam) (hlam1 : lam ≤ 1) :
(∃ X ∈ Q, ∀ Z ∈ Q, mean P ⇑Z - lam * gini P ⇑Z ≤ mean P ⇑X - lam * gini P ⇑X) ∧
∀ X ∈ Q, (∀ Z ∈ Q, mean P ⇑Z - lam * gini P ⇑Z ≤ mean P ⇑X - lam * gini P ⇑X) →
SSDEfficient P ((fun Z : Lp ℝ q P => (⇑Z : Ω → ℝ)) '' Q) ⇑X := by sorry
end DualSSD.MeanRisk
Read-back
What the Lean code literally says, in plain math · claude-opus-5-5
Fix a set with a -algebra and a probability measure on it, so . Fix an exponent satisfying , and , so in effect . Let be the space of real-valued -integrable functions on modulo -almost-everywhere equality, with its usual norm. Each element of this space is used below through one fixed representative function , written . The statement does not say which representative is chosen.
Hypotheses. Let satisfy all of the following:
- is convex (over ).
- is bounded in the norm.
- is closed in the norm topology.
- is nonempty.
Let be a real number with .
Objective. For , define
- and are functionals defined elsewhere in this bundle. Each takes and a function and returns a real number. Their definitions are not included in the code given here, so their exact formulas cannot be stated from it. In particular, this code does not show whether they depend only on the law of under or on the particular representative chosen.
- is also a bundle definition whose formula is not shown. It is a property of a set of functions , a function , and . Its name suggests efficiency of within with respect to second-order stochastic dominance, but what exactly it requires cannot be read from this code.
The statement asserts two things:
- Existence of a maximiser. There is some such that
- Every maximiser is efficient. Take any with for all . Then
holds. The set here is the collection of chosen representatives of the members of , not the equivalence classes themselves.
The maximisation in both parts is non-strict (). Part 2 is an implication for each maximiser. It does not claim the maximiser is unique, and it does not claim the converse (that efficient elements are maximisers).
Degenerate cases. cannot be empty, because . The hypotheses can be satisfied together, for example by with any admissible and . So part 2 is not vacuous: part 1 guarantees at least one maximiser exists.
- a single point : part 1 holds trivially with that . Part 2 then asserts .
- with one point, or concentrated so that is one-dimensional: every element is effectively a constant . is then a closed bounded interval of constants, and is whatever and return on constants, which the code does not show.
- Division, subtraction in ℕ, suprema: the statement itself uses none of these; its only arithmetic is real subtraction and multiplication.
- Integrals of non-integrable functions: whether or use integrals that could fall back to a default value depends on their unseen definitions. Every here is -integrable for on a probability space, so it is also integrable.
- The value and all are excluded by the hypotheses. The hypothesis is redundant given .
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.