Theorem 1 — convexity through the Clarke subdifferential
OpenTiltedCvx.Supplement.theorem_1convex-analysisgeneralized-convexitynonsmooth-analysisoptimizationp2o-gran-per-chapterp2o-plan-bookp2o-strengthenedp2o-v1
Let be a proper lower semicontinuous function and let be its Clarke–Rockafellar subdifferential. The following are equivalent:
- for all and , ;
- is –tilted convex;
- is convex;
- the operator is –tilted monotone;
- the operator is monotone: for , .
This restates the classical subgradient and monotonicity characterizations of convexity in terms of the tilt ; it identifies the top classes with the proper lsc convex functions.
Differences from the source:
- generalization: functions on ℝ^d → functions on a finite-dimensional real inner product space E; reason: every such E is isometric to ℝ^d and ℝ^d = EuclideanSpace ℝ (Fin d) is an instance, so the statement specializes to the paper's
- encoding: ℝ̄ = ℝ ∪ {±∞}, dom(f) = {f finite} → EReal and the published TiltedCvx.effDom; reason: Lean convention; same instances
- encoding: f : ℝ^d → ℝ̄ is convex → the published epigraph convexity WassersteinDRO.ConvexLossP1.IsConvexExt f (the epigraph {(x, t) : f x ≤ t} is convex); reason: the standard notion for extended-valued functions; Mathlib's ConvexOn needs an ordered-module codomain, which EReal is not
- encoding: ∂̄f is monotone → the published AccelPPM.FPR.IsMonotoneOp (ClarkeSubdiff f): ⟨x − y, u − v⟩ ≥ 0 for u ∈ ∂̄f(x), v ∈ ∂̄f(y); reason: the same inequality as ⟨y* − x*, y − x⟩ ≥ 0
Preamble
import Mathlib import Definitions.Def_TiltedCvx_effDom import Definitions.Def_TiltedCvx_IsProper import Definitions.Def_TiltedCvx_TiltedConvex import Definitions.Def_TiltedCvx_TiltedMonotone import Definitions.Def_ClarkeStrat_Proj_Setting import Definitions.Def_WassersteinDRO_ConvexLossP1_IsConvexExt import Definitions.Def_AccelPPM_FPR_IsMaximalMonotone open scoped InnerProductSpace open ClarkeStrat.Proj
Formal statement
namespace TiltedCvx.Supplement
/-- **Theorem 1 (convexity through the Clarke subdifferential).** Pedramfar–Khanna–Quinn–Aggarwal,
*Tilted Convexity and an Exact Offline–Online Separation*, arXiv:2611.03472, Theorem 1, p.15. Let
`f : ℝ^d → ℝ̄` be a proper lsc function. Then the following are equivalent:
(i) for all `x, y ∈ dom(f)` and `v ∈ ∂̄f(x)`, `f(y) − f(x) ≥ ⟨y − x, v⟩`;
(ii) `f` is 1–tilted convex;
(iii) `f` is convex;
(iv) the operator `∂̄f` is 1–tilted monotone;
(v) the operator `∂̄f` is monotone.
**Formalization Note** `∂̄f` is the published `ClarkeStrat.Proj.ClarkeSubdiff`. Convexity of
`f : E → ℝ̄` is the published epigraph convexity `WassersteinDRO.ConvexLossP1.IsConvexExt`, and
monotonicity of a set-valued operator is the published `AccelPPM.FPR.IsMonotoneOp`
(`⟨x − y, u − v⟩ ≥ 0` for `u ∈ F x`, `v ∈ F y`, the same as `⟨y* − x*, y − x⟩ ≥ 0`). -/
theorem theorem_1 {E : Type*} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
[FiniteDimensional ℝ E] (f : E → EReal) (hf : TiltedCvx.IsProper f)
(hlsc : LowerSemicontinuous f) :
List.TFAE
[∀ x ∈ TiltedCvx.effDom f, ∀ y ∈ TiltedCvx.effDom f, ∀ v ∈ ClarkeSubdiff f x,
((⟪y - x, v⟫_ℝ : ℝ) : EReal) ≤ f y - f x,
TiltedCvx.TiltedConvex 1 f,
WassersteinDRO.ConvexLossP1.IsConvexExt f,
TiltedCvx.TiltedMonotone 1 (ClarkeSubdiff f),
AccelPPM.FPR.IsMonotoneOp (ClarkeSubdiff f)] := by sorry
end TiltedCvx.Supplement
Source
Pedramfar–Khanna–Quinn–Aggarwal, Tilted Convexity and an Exact Offline–Online Separation, arXiv:2611.03472, Theorem 1, p.15