Theorem 19 — basic properties of γ–tilted secant convexity
OpenTiltedCvx.Supplement.theorem_19convex-analysisgeneralized-convexitynonsmooth-analysisoptimizationp2o-gran-per-chapterp2o-plan-bookp2o-strengthenedp2o-v1
Let , convex, and with . Then:
- if is –tilted secant convex, then it is quasiconvex and is convex;
- if is convex, then it is –tilted secant convex;
- the set of –tilted secant convex functions over is closed under arbitrary non-empty suprema;
- if is –tilted secant convex and is convex, then is –tilted secant convex;
- if and is –tilted secant convex, then is –tilted secant convex and on ;
- is –tilted secant convex if and only if, for every line meeting , the restriction is –tilted secant convex.
These are the analogues for of the elementary closure properties of convex functions.
Formalization Note A function is encoded by its proper extension: , never , equal to off . In item 3 the family is a non-empty set of such functions and the supremum is pointwise. A line is with , regarded as through (the published lineRestrict).
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 : X → ℝ ∪ {+∞}, and the functions “over X” of (iii) → f : E → EReal, never ⊥, equal to ⊤ off X; the family is a non-empty set U of such functions and its supremum is sSup U (pointwise); reason: the paper's proper-extension convention (§3); a set-indexed family is an arbitrary family
- encoding: f is quasiconvex (on X) → QuasiconvexOn ℝ X f; reason: convex sublevel sets of f on X
- 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|C : C → ℝ ∪ {+∞} → x ↦ f x on C, ⊤ off C; reason: proper extension of the restriction
- encoding: lsc(f) : ℝ^d → ℝ ∪ {+∞} → lscEnv X f, with lscEnv X f x ≠ ⊥ for every x; reason: the codomain claim is that lsc(f) never takes −∞
- encoding: f|ℓ : ℓ ∩ X → ℝ ∪ {+∞} for a line ℓ meeting X → lineRestrict f p e {t | p + t • e ∈ X} with ‖e‖ = 1; reason: the unit-speed parametrization is an isometry of ℝ onto ℓ, as in the published Theorems 9 and 16
Preamble
import Mathlib import Definitions.Def_TiltedCvx_effDom import Definitions.Def_TiltedCvx_TiltedSecant import Definitions.Def_TiltedCvx_lineRestrict import Definitions.Def_WassersteinDRO_ConvexLossP1_IsConvexExt import Definitions.Def_TiltedCvx_Supplement_lscEnv open Classical
Formal statement
namespace TiltedCvx.Supplement
/-- **Theorem 19 (basic properties of γ–tilted secant convexity).**
Pedramfar–Khanna–Quinn–Aggarwal, arXiv:2611.03472, Theorem 19, pp.25–26. Let `γ ∈ [0, 1]`,
`X ⊆ ℝ^d` a convex set, and `f : X → ℝ ∪ {+∞}` with effective domain `K := dom(f)`. Then:
(i) if `f` is γ–tilted secant convex, then it is quasiconvex, and `K` is convex;
(ii) if `f` is convex, then it is γ–tilted secant convex;
(iii) the set of γ–tilted secant convex functions over `X` is closed under arbitrary non-empty
suprema;
(iv) if `f` is γ–tilted secant convex and `C ⊆ X` is convex, then `f|C : C → ℝ ∪ {+∞}` is
γ–tilted secant convex;
(v) if `γ ∈ (0, 1]` and `f` is γ–tilted secant convex, then `lsc(f) : ℝ^d → ℝ ∪ {+∞}` is γ–tilted
secant convex and `f = lsc(f)` on `ri(K)`;
(vi) `f` is γ–tilted secant convex if and only if, for every line `ℓ` that intersects `X`, the
restriction `f|ℓ : ℓ ∩ X → ℝ ∪ {+∞}` is γ–tilted secant convex.
**Formalization Note** A function `X → ℝ ∪ {+∞}` is encoded by its proper extension: `f : E →
EReal`, never `−∞`, equal to `+∞` off `X`; the functions "over `X`" in (iii) are those of this form,
and their supremum is the pointwise supremum `sSup U`. Quasiconvexity of `f` on `X` is
`QuasiconvexOn ℝ X f`; convexity of `f` is the published epigraph convexity `IsConvexExt`. `f|C` is
`f` on `C` and `+∞` elsewhere. `lsc(f)` is `lscEnv X f`, and "`lsc(f)` maps into `ℝ ∪ {+∞}`" is
stated as `lsc(f) ≠ −∞` everywhere. A line is `{p + t e : t ∈ ℝ}` with `‖e‖ = 1`, identified with `ℝ`
by its unit-speed parametrization; `f|ℓ` on `ℓ ∩ X` is the published `lineRestrict f p e I` with `I`
the parameters of `ℓ ∩ X` (`+∞` off `I`). -/
theorem theorem_19 {E : Type*} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
[FiniteDimensional ℝ E] (γ : ℝ) (hγ0 : 0 ≤ γ) (hγ1 : γ ≤ 1) (X : Set E)
(hX : Convex ℝ X) (f : E → EReal) (hbot : ∀ x, f x ≠ ⊥) (htop : ∀ x ∉ X, f x = ⊤) :
(TiltedCvx.TiltedSecant γ f →
QuasiconvexOn ℝ X f ∧ Convex ℝ (TiltedCvx.effDom f)) ∧
(WassersteinDRO.ConvexLossP1.IsConvexExt f → TiltedCvx.TiltedSecant γ f) ∧
(∀ U : Set (E → EReal), U.Nonempty →
(∀ g ∈ U, (∀ x, g x ≠ ⊥) ∧ (∀ x ∉ X, g x = ⊤) ∧ TiltedCvx.TiltedSecant γ g) →
(∀ x, sSup U x ≠ ⊥) ∧ (∀ x ∉ X, sSup U x = ⊤) ∧ TiltedCvx.TiltedSecant γ (sSup U)) ∧
(∀ C : Set E, C ⊆ X → Convex ℝ C → TiltedCvx.TiltedSecant γ f →
TiltedCvx.TiltedSecant γ (fun x => if x ∈ C then f x else ⊤)) ∧
(0 < γ → TiltedCvx.TiltedSecant γ f →
(∀ x, lscEnv X f x ≠ ⊥) ∧ TiltedCvx.TiltedSecant γ (lscEnv X f) ∧
∀ x ∈ intrinsicInterior ℝ (TiltedCvx.effDom f), f x = lscEnv X f x) ∧
(TiltedCvx.TiltedSecant γ f ↔
∀ p e : E, ‖e‖ = 1 → (∃ t : ℝ, p + t • e ∈ X) →
TiltedCvx.TiltedSecant γ (TiltedCvx.lineRestrict f p e {t | p + t • e ∈ X})) := by sorry
end TiltedCvx.Supplement
Source
Pedramfar–Khanna–Quinn–Aggarwal, Tilted Convexity and an Exact Offline–Online Separation, arXiv:2611.03472, Theorem 19, pp.25–26