§2 — consequences of logarithmic homogeneity: , , , ,
ProvedExpConeIPM.Corrector.homogeneity_identitiesLet be a proper cone (pointed, closed, convex, with nonempty interior) and let be a -logarithmically homogeneous self-concordant barrier for , so that in particular for and . Then for every and every ,
Here is the gradient, the Hessian, and is the third derivative contracted once with , a linear map.
These identities are the algebraic engine of the paper's analysis: is what makes the corrector (16) orthogonal to the complementarity gap (Lemma 3), and turns the centering term into (Lemma 4).
Formalization Note is gradient F, is the published hess F (the derivative of the gradient), and is fderiv ℝ (hess F) x x. The barrier is the published IsLogHomBarrier, which contains the paper's two conditions (self-concordance with , logarithmic homogeneity) plus three standard ones the paper uses implicitly (boundary blow-up, the -inequality, positive definite Hessian).
import Mathlib import Definitions.Def_SelfScaledIPM_ShortStep_Setting open scoped InnerProductSpace
namespace ExpConeIPM.Corrector
/-- **§2 homogeneity identities** (Dahl–Andersen, Math. Program. 194 (2022), p. 345). Let `K ⊆ ℝᴺ`
be a proper cone and `F` a `ϑ`-logarithmically homogeneous self-concordant barrier for `K`. Then for
every `x ∈ int K` and `τ > 0`:
`F'(τx) = (1/τ) F'(x)`, `F''(τx) = (1/τ²) F''(x)`, `F''(x)x = −F'(x)`, `F'''(x)[x] = −2F''(x)`,
`⟨F'(x), x⟩ = −ϑ`.
`F' = gradient F`, `F'' = hess F`, and `F'''(x)[x] = fderiv ℝ (hess F) x x`, a linear map. -/
theorem homogeneity_identities {N : ℕ} (K : Set (EuclideanSpace ℝ (Fin N)))
(F : EuclideanSpace ℝ (Fin N) → ℝ) (ϑ : ℝ)
(hK : SelfScaledIPM.ShortStep.IsProperCone K)
(hF : SelfScaledIPM.ShortStep.IsLogHomBarrier K F ϑ)
(x : EuclideanSpace ℝ (Fin N)) (hx : x ∈ interior K) (τ : ℝ) (hτ : 0 < τ) :
gradient F (τ • x) = (1 / τ) • gradient F x ∧
SelfScaledIPM.ShortStep.hess F (τ • x) = (1 / τ ^ 2) • SelfScaledIPM.ShortStep.hess F x ∧
SelfScaledIPM.ShortStep.hess F x x = -gradient F x ∧
fderiv ℝ (SelfScaledIPM.ShortStep.hess F) x x = (-2 : ℝ) • SelfScaledIPM.ShortStep.hess F x ∧
⟪gradient F x, x⟫_ℝ = -ϑ := by sorry
end ExpConeIPM.Corrector
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.