Part VII — Saddle-Functions and Minimax Theory

Proofs omitted. Hover over highlighted terms to see their Lean types.

Reading sections

§33 — Saddle-Functions — Part 1

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Set membership is treated classically in this section.

noncomputable local instance classicalSetDecidablePred {α : Type*} (s : Set α) : DecidablePred s

Lean source

An EReal-valued function is convex on C when it satisfies Jensen's inequality whenever the relevant convex combination remains in C.

def IsERealConvexOn {m : ℕ} (C : Set (Fin m → ℝ)) (f : (Fin m → ℝ) → EReal) : Prop := ∀ ⦃x y : Fin m → ℝ⦄, x ∈ C → y ∈ C → ∀ ⦃a b : ℝ⦄, 0 ≤ a → 0 ≤ b → a + b = 1 → a • x + b • y ∈ C → f (a • x + b • y) ≤ (a : EReal) * f x + (b : EReal) * f y

Lean source

An EReal-valued function is concave on C when it satisfies the reverse Jensen inequality whenever the relevant convex combination remains in C.

def IsERealConcaveOn {m : ℕ} (C : Set (Fin m → ℝ)) (f : (Fin m → ℝ) → EReal) : Prop := ∀ ⦃x y : Fin m → ℝ⦄, x ∈ C → y ∈ C → ∀ ⦃a b : ℝ⦄, 0 ≤ a → 0 ≤ b → a + b = 1 → a • x + b • y ∈ C → (a : EReal) * f x + (b : EReal) * f y ≤ f (a • x + b • y)

Lean source

A bifunction is concave-convex when it is concave in the first variable and convex in the second.

def IsConcaveConvexOn {m n : ℕ} (C : Set (Fin m → ℝ)) (D : Set (Fin n → ℝ)) (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := (∀ v ∈ D, IsERealConcaveOn C (fun u => K u v)) ∧ ∀ u ∈ C, IsERealConvexOn D (fun v => K u v)

Lean source

A bifunction is convex-concave when it is convex in the first variable and concave in the second.

def IsConvexConcaveOn {m n : ℕ} (C : Set (Fin m → ℝ)) (D : Set (Fin n → ℝ)) (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := (∀ v ∈ D, IsERealConvexOn C (fun u => K u v)) ∧ ∀ u ∈ C, IsERealConcaveOn D (fun v => K u v)

Lean source

Definition33.0.1: A saddle function on C × D is a bifunction that is either concave-convex or convex-concave.

def IsSaddleFunctionOn {m n : ℕ} (C : Set (Fin m → ℝ)) (D : Set (Fin n → ℝ)) (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := IsConcaveConvexOn C D K ∨ IsConvexConcaveOn C D K

Lean source

The lower simple extension of a bifunction on C × D takes the value ⊤ on C × Dᶜ and ⊥ on Cᶜ × (Fin n → ℝ).

noncomputable def lowerSimpleExtension {m n : ℕ} (C : Set (Fin m → ℝ)) (D : Set (Fin n → ℝ)) (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u v => if u ∈ C then if v ∈ D then K u v else ⊤ else ⊥

Lean source

The upper simple extension of a bifunction on C × D takes the value ⊥ on Cᶜ × D and ⊤ on (Fin m → ℝ) × Dᶜ.

noncomputable def upperSimpleExtension {m n : ℕ} (C : Set (Fin m → ℝ)) (D : Set (Fin n → ℝ)) (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u v => if v ∈ D then if u ∈ C then K u v else ⊥ else ⊤

Lean source

Definition33.0.2: The lower/upper simple extensions of K are the pair consisting of its lower simple extension and its upper simple extension.

noncomputable abbrev simpleExtensions {m n : ℕ} (C : Set (Fin m → ℝ)) (D : Set (Fin n → ℝ)) (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : ((Fin m → ℝ) → (Fin n → ℝ) → EReal) × ((Fin m → ℝ) → (Fin n → ℝ) → EReal) := (lowerSimpleExtension C D K, upperSimpleExtension C D K)

Lean source

Lemma33.0.3: Agreement on C × D. If u ∈ C and v ∈ D, then the lower and upper simple extensions K₁ and K₂ from Definition33.0.2 both agree with K at (u, v).

lemma simpleExtensions_eq_on_product {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) : lowerSimpleExtension C D K u v = K u v ∧ upperSimpleExtension C D K u v = K u v

Lean source

The convex closure in the second variable is the lower semicontinuous regularization obtained by taking the supremum of local infima over open balls.

noncomputable def convexClosureInSecond {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u v => ⨆ (ε : {ε : ℝ // 0 < ε}), ⨅ (w : {w : Fin n → ℝ // ‖w - v‖ < ε.1}), K u w.1

Lean source

The concave closure in the first variable is the upper semicontinuous regularization obtained by taking the infimum of local suprema over open balls.

noncomputable def concaveClosureInFirst {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u v => ⨅ (ε : {ε : ℝ // 0 < ε}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), K w.1 v

Lean source

The convex closure in the first variable is the lower semicontinuous regularization obtained by taking the supremum of local infima over open balls.

noncomputable def convexClosureInFirst {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u v => ⨆ (ε : {ε : ℝ // 0 < ε}), ⨅ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), K w.1 v

Lean source

A bifunction is convex-closed in the second variable when it agrees with its convex closure in that variable.

def IsConvexClosedInSecond {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := K = convexClosureInSecond K

Lean source

The concave closure in the second variable is the upper semicontinuous regularization obtained by taking the infimum of local suprema over open balls.

noncomputable def concaveClosureInSecond {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u v => ⨅ (ε : {ε : ℝ // 0 < ε}), ⨆ (w : {w : Fin n → ℝ // ‖w - v‖ < ε.1}), K u w.1

Lean source

A bifunction is concave-closed in the second variable when it agrees with its concave closure in that variable.

def IsConcaveClosedInSecond {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := K = concaveClosureInSecond K

Lean source

A bifunction is concave-closed in the first variable when it agrees with its concave closure in that variable.

def IsConcaveClosedInFirst {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := K = concaveClosureInFirst K

Lean source

A bifunction is convex-closed in the first variable when it agrees with its convex closure in that variable.

def IsConvexClosedInFirst {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := K = convexClosureInFirst K

Lean source

The closure data attached to a bifunction: the two coordinatewise closures and the corresponding fixed-point predicates.

structure ConvexConcaveClosurePackage (m n : ℕ) where cl_v : (Fin m → ℝ) → (Fin n → ℝ) → EReal cl_u : (Fin m → ℝ) → (Fin n → ℝ) → EReal convexClosedInV : Prop concaveClosedInU : Prop

Lean source

Definition33.0.4: The convex closure in v, the concave closure in u, and the corresponding fixed-point predicates canonically attached to an EReal-valued bifunction K.

noncomputable def convexConcaveClosureData {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : ConvexConcaveClosurePackage m n := let clV := convexClosureInSecond K let clU := concaveClosureInFirst K { cl_v := clV cl_u := clU convexClosedInV := K = clV concaveClosedInU := K = clU }

Lean source

Helper for Lemma33.0.3: on C × D, the lower simple extension reduces to the original bifunction value.

lemma helperForLemma33_0_3_lowerSimpleExtension_agrees {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) : lowerSimpleExtension C D K u v = K u v

Lean source

Helper for Lemma33.0.3: on C × D, the upper simple extension reduces to the original bifunction value.

lemma helperForLemma33_0_3_upperSimpleExtension_agrees {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) : upperSimpleExtension C D K u v = K u v

Lean source

Helper for Lemma33.0.5: multiplying by a positive real scalar commutes with iSup in EReal.

lemma helperForLemma33_0_5_positiveReal_mul_iSup {ι : Sort*} {a : ℝ} (ha : 0 < a) (f : ι → EReal) : ((a : EReal) * (⨆ i, f i)) = ⨆ i, (a : EReal) * f i

Lean source

Helper for Lemma33.0.5: multiplying by a positive real scalar commutes with iInf in EReal.

lemma helperForLemma33_0_5_positiveReal_mul_iInf {ι : Sort*} {a : ℝ} (ha : 0 < a) (f : ι → EReal) : ((a : EReal) * (⨅ i, f i)) = ⨅ i, (a : EReal) * f i

Lean source

Helper for Lemma33.0.5: equal-radius balls are stable under the same convex combination as their centers.

lemma helperForLemma33_0_5_convexCombination_mem_ball {n : ℕ} {x y w₁ w₂ : Fin n → ℝ} {r a b : ℝ} (hw₁ : ‖w₁ - x‖ < r) (hw₂ : ‖w₂ - y‖ < r) (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) : ‖(a • w₁ + b • w₂) - (a • x + b • y)‖ < r

Lean source

Helper for Lemma33.0.5: shrinking the radius makes the local infimum larger.

lemma helperForLemma33_0_5_localInfimum_antitone_radius {n : ℕ} {f : (Fin n → ℝ) → EReal} {x : Fin n → ℝ} {δ ε : {r : ℝ // 0 < r}} (hδε : δ.1 ≤ ε.1) : (⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1) ≤ ⨅ w : {w : Fin n → ℝ // ‖w - x‖ < δ.1}, f w.1

Lean source

Helper for Lemma33.0.5: shrinking the radius makes the local supremum smaller.

lemma helperForLemma33_0_5_localSupremum_monotone_radius {n : ℕ} {f : (Fin n → ℝ) → EReal} {x : Fin n → ℝ} {δ ε : {r : ℝ // 0 < r}} (hδε : δ.1 ≤ ε.1) : (⨆ w : {w : Fin n → ℝ // ‖w - x‖ < δ.1}, f w.1) ≤ ⨆ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1

Lean source

Helper for Lemma33.0.5: away from the exceptional ⊥/⊤ ambiguity of EReal, the product-indexed infimum is bounded by the sum of the two endpoint infima.

lemma helperForLemma33_0_5_productIndexed_iInf_weightedSum_le_of_nonexceptional {ι κ : Type*} [Nonempty ι] [Nonempty κ] {F : ι → EReal} {G : κ → EReal} (h₁ : (⨅ i, F i) ≠ ⊥ ∨ (⨅ j, G j) ≠ ⊤) (h₂ : (⨅ i, F i) ≠ ⊤ ∨ (⨅ j, G j) ≠ ⊥) : (⨅ p : ι × κ, F p.1 + G p.2) ≤ (⨅ i, F i) + (⨅ j, G j)

Lean source

Helper for Lemma33.0.5: weighted sums of independent endpoint suprema are controlled by the corresponding product-indexed supremum.

lemma helperForLemma33_0_5_weightedSum_le_productIndexed_iSup {ι κ : Type*} [Nonempty ι] [Nonempty κ] {a b : ℝ} (ha : 0 < a) (hb : 0 < b) (f : ι → EReal) (g : κ → EReal) : ((a : EReal) * (⨆ i, f i)) + ((b : EReal) * (⨆ j, g j)) ≤ ⨆ p : ι × κ, (a : EReal) * f p.1 + (b : EReal) * g p.2

Lean source

Helper for Lemma33.0.5: negating a supremum of negated values recovers the corresponding infimum.

lemma helperForLemma33_0_5_neg_iSup_neg_eq_iInf {ι : Sort*} (f : ι → EReal) : -(⨆ i, -f i) = ⨅ i, f i

Lean source

Helper for Lemma33.0.5: negating an infimum of negated values recovers the corresponding supremum.

lemma helperForLemma33_0_5_neg_iInf_neg_eq_iSup {ι : Sort*} (f : ι → EReal) : -(⨅ i, -f i) = ⨆ i, f i

Lean source

Helper for Lemma33.0.5: negating a sum is always bounded below by the sum of the negated terms in EReal.

lemma helperForLemma33_0_5_neg_sum_upper_bound {x y : EReal} : -x + -y ≤ -(x + y)

Lean source

Helper for Lemma33.0.5: negating a convex EReal-valued section produces a concave one.

lemma helperForLemma33_0_5_convexNegation_isConcave {n : ℕ} {C : Set (Fin n → ℝ)} {f : (Fin n → ℝ) → EReal} (hConv : IsERealConvexOn C f) : IsERealConcaveOn C (fun x => -f x)

Lean source

Helper for Lemma33.0.5: the local Jensen predicate on Set.univ gives the usual convex epigraph.

lemma helperForLemma33_0_5_isERealConvexOn_univ_to_ConvexFunction {n : ℕ} {f : (Fin n → ℝ) → EReal} (hConv : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) f) : ConvexFunction f

Lean source

Helper for Lemma33.0.5: a global ConvexFunction immediately yields the local Jensen predicate on Set.univ as soon as ⊥ never occurs.

lemma helperForLemma33_0_5_convexFunction_to_isERealConvexOn_univ {n : ℕ} {f : (Fin n → ℝ) → EReal} (hConv : ConvexFunction f) (hNoBot : ∀ x, f x ≠ ⊥) : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) f

Lean source

Helper for Lemma33.0.5: multiplying by a positive real scalar preserves the property of being different from ⊥.

lemma helperForLemma33_0_5_positiveReal_mul_ne_bot {a : ℝ} (ha : 0 < a) {x : EReal} (hx : x ≠ ⊥) : (a : EReal) * x ≠ ⊥

Lean source

Helper for Lemma33.0.5: multiplying by a positive real scalar preserves the property of being different from ⊤.

lemma helperForLemma33_0_5_positiveReal_mul_ne_top {a : ℝ} (ha : 0 < a) {x : EReal} (hx : x ≠ ⊤) : (a : EReal) * x ≠ ⊤

Lean source

Helper for Lemma33.0.5: if a local supremum over a ball is ⊥, then every point in that ball already has value ⊥.

lemma helperForLemma33_0_5_localSupremum_eq_bot_implies_pointwise_bot {n : ℕ} {f : (Fin n → ℝ) → EReal} {x : Fin n → ℝ} {ε : {r : ℝ // 0 < r}} (hBot : (⨆ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1) = ⊥) : ∀ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1 = ⊥

Lean source

Helper for Lemma33.0.5: if a local infimum over a ball is ⊤, then every point in that ball already has value ⊤.

lemma helperForLemma33_0_5_localInfimum_eq_top_implies_pointwise_top {n : ℕ} {f : (Fin n → ℝ) → EReal} {x : Fin n → ℝ} {ε : {r : ℝ // 0 < r}} (hTop : (⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1) = ⊤) : ∀ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1 = ⊤

Lean source

Helper for Lemma33.0.5: if an EReal-indexed infimum equals ⊤, then every indexed value already equals ⊤.

lemma helperForLemma33_0_5_iInf_eq_top_implies_pointwise_top {ι : Sort*} {F : ι → EReal} (hTop : (⨅ i, F i) = (⊤ : EReal)) : ∀ i, F i = (⊤ : EReal)

Lean source

Helper for Lemma33.0.5: if a local infimum over a ball is ⊥, then every strict upper bound on ⊥ is beaten by some point in that ball.

lemma helperForLemma33_0_5_localInfimum_eq_bot_implies_exists_ballWitness_lt {n : ℕ} (ε : {r : ℝ // 0 < r}) {f : (Fin n → ℝ) → EReal} {x : Fin n → ℝ} (hBot : (⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1) = ⊥) : ∀ {z : EReal}, (⊥ : EReal) < z → ∃ w : Fin n → ℝ, ‖w - x‖ < ε.1 ∧ f w < z

Lean source

Helper for Lemma33.0.5: an actual ⊥-valued witness in the x-ball already forces the target fixed-radius local infimum to be ⊥.

lemma helperForLemma33_0_5_fixedRadiusLocalInfimum_botWitness_forces_targetBot {n : ℕ} (ε : {r : ℝ // 0 < r}) {f : (Fin n → ℝ) → EReal} {x y w : Fin n → ℝ} {a b : ℝ} (hConv : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) f) (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) (hPosA : 0 < a) (hw : ‖w - x‖ < ε.1) (hfwBot : f w = ⊥) : (⨅ z : {z : Fin n → ℝ // ‖z - (a • x + b • y)‖ < ε.1}, f z.1) = ⊥

Lean source

Helper for Lemma33.0.5: extending a convex function by ⊤ outside a convex set preserves Jensen convexity as soon as the original function never takes the value ⊥ on that set.

lemma helperForLemma33_0_5_extendByTopOnConvexSet_preserves_convexity {n : ℕ} {C : Set (Fin n → ℝ)} {f : (Fin n → ℝ) → EReal} (hCconv : Convex ℝ C) (hConv : IsERealConvexOn C f) (hNoBot : ∀ z ∈ C, f z ≠ (⊥ : EReal)) : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (fun z => if z ∈ C then f z else ⊤)

Lean source

Helper for Lemma33.0.5: an affine lower bound gives a uniform finite lower bound on any open ball around x.

lemma helperForLemma33_0_5_ball_uniform_affine_lowerBound {n : ℕ} (ε : {r : ℝ // 0 < r}) (x b : Fin n → ℝ) (β : ℝ) {w : Fin n → ℝ} (hw : ‖w - x‖ < ε.1) : (((-((‖x‖ + ε.1) * (Finset.univ.sum fun i : Fin n => |b i|)) - β : ℝ)) : EReal) ≤ ((w ⬝ᵥ b - β : ℝ) : EReal)

Lean source

Helper for Lemma33.0.5: the true fixed-radius (⊥, ⊤) branch is exactly where the direct local-witness proof stops.

lemma helperForLemma33_0_5_fixedRadiusLocalInfimum_mixedBotTop_collapse {n : ℕ} (ε : {r : ℝ // 0 < r}) {f : (Fin n → ℝ) → EReal} {x y : Fin n → ℝ} {a b : ℝ} (hConv : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) f) (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) (hPosA : 0 < a) (_hPosB : 0 < b) (hLocalInfXBot : (⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1) = ⊥) (_hLocalInfYTop : (⨅ w : {w : Fin n → ℝ // ‖w - y‖ < ε.1}, f w.1) = ⊤) : (⨅ w : {w : Fin n → ℝ // ‖w - (a • x + b • y)‖ < ε.1}, f w.1) = ⊥

Lean source

Helper for Lemma33.0.5: the true fixed-radius (⊥, ⊤) branch is exactly where the direct local-witness proof stops.

lemma helperForLemma33_0_5_fixedRadiusLocalInfimum_exceptional_branch {n : ℕ} (ε : {r : ℝ // 0 < r}) {f : (Fin n → ℝ) → EReal} {x y : Fin n → ℝ} {a b : ℝ} (hConv : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) f) (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) (hPosA : 0 < a) (hPosB : 0 < b) (hLocalInfXBot : (⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1) = ⊥) (hLocalInfYTop : (⨅ w : {w : Fin n → ℝ // ‖w - y‖ < ε.1}, f w.1) = ⊤) : (⨅ w : {w : Fin n → ℝ // ‖w - (a • x + b • y)‖ < ε.1}, f w.1) ≤ (a : EReal) * (⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1) + (b : EReal) * (⨅ w : {w : Fin n → ℝ // ‖w - y‖ < ε.1}, f w.1)

Lean source

Helper for Lemma33.0.5: the two-layer sup-inf closure is the negative of the corresponding inf-sup closure of the negated family.

lemma helperForLemma33_0_5_negatedTwoLayerClosureIdentity {ι : Sort*} {κ : ι → Sort*} (F : ∀ i, κ i → EReal) : (⨆ i, ⨅ j, F i j) = -(⨅ i, ⨆ j, -F i j)

Lean source

Helper for Lemma33.0.5: after swapping the variables and negating the bifunction, the convex closure in the second variable becomes the negative of the original concave closure in the first variable.

lemma helperForLemma33_0_5_swappedNegatedClosureIdentity {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (v : Fin n → ℝ) : convexClosureInSecond (fun v' u' => -K u' v') v u = -concaveClosureInFirst K u v

Lean source

Helper for Lemma33.0.5: at a fixed radius, local suprema preserve concavity.

lemma helperForLemma33_0_5_fixedRadiusLocalSupremum_preserves_concavity {n : ℕ} (ε : {r : ℝ // 0 < r}) {f : (Fin n → ℝ) → EReal} (hConc : IsERealConcaveOn (Set.univ : Set (Fin n → ℝ)) f) : IsERealConcaveOn (Set.univ : Set (Fin n → ℝ)) (fun x => ⨆ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1)

Lean source

Helper for Lemma33.0.5: at a fixed radius, local infima preserve convexity.

lemma helperForLemma33_0_5_fixedRadiusLocalInfimum_target_le_productInfimum {n : ℕ} (ε : {r : ℝ // 0 < r}) {f : (Fin n → ℝ) → EReal} {x y : Fin n → ℝ} {a b : ℝ} (hConv : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) f) (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) : (⨅ w : {w : Fin n → ℝ // ‖w - (a • x + b • y)‖ < ε.1}, f w.1) ≤ ⨅ p : {w : Fin n → ℝ // ‖w - x‖ < ε.1} × {w : Fin n → ℝ // ‖w - y‖ < ε.1}, (a : EReal) * f p.1.1 + (b : EReal) * f p.2.1

§33 — Saddle-Functions — Part 2

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Helper for Lemma33.0.5: outside the two mixed (⊥, ⊤) corners, the product-indexed infimum of weighted endpoint values is bounded by the weighted sum of the endpoint local infima.

lemma helperForLemma33_0_5_fixedRadiusLocalInfimum_productInfimum_nonexceptionalBound {n : ℕ} (ε : {r : ℝ // 0 < r}) {f : (Fin n → ℝ) → EReal} {x y : Fin n → ℝ} {a b : ℝ} (ha : 0 < a) (hb : 0 < b) (h₁ : (⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1) ≠ (⊥ : EReal) ∨ (⨅ w : {w : Fin n → ℝ // ‖w - y‖ < ε.1}, f w.1) ≠ (⊤ : EReal)) (h₂ : (⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1) ≠ (⊤ : EReal) ∨ (⨅ w : {w : Fin n → ℝ // ‖w - y‖ < ε.1}, f w.1) ≠ (⊥ : EReal)) : (⨅ p : {w : Fin n → ℝ // ‖w - x‖ < ε.1} × {w : Fin n → ℝ // ‖w - y‖ < ε.1}, (a : EReal) * f p.1.1 + (b : EReal) * f p.2.1) ≤ (a : EReal) * (⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1) + (b : EReal) * (⨅ w : {w : Fin n → ℝ // ‖w - y‖ < ε.1}, f w.1)

Lean source

Helper for Lemma33.0.5: at a fixed radius, local infima preserve convexity.

lemma helperForLemma33_0_5_fixedRadiusLocalInfimum_preserves_convexity {n : ℕ} (ε : {r : ℝ // 0 < r}) {f : (Fin n → ℝ) → EReal} (hConv : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) f) : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (fun x => ⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1)

Lean source

Helper for Lemma33.0.5: at a fixed radius, local infima have a convex epigraph.

lemma helperForLemma33_0_5_fixedRadiusLocalInfimum_convexFunction {n : ℕ} (ε : {r : ℝ // 0 < r}) {f : (Fin n → ℝ) → EReal} (hConv : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) f) : ConvexFunction (fun x => ⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1)

Lean source

Helper for Lemma33.0.5: at a fixed radius in the second variable, local infima preserve concavity in the first variable.

lemma helperForLemma33_0_5_fixedRadiusLocalInfimum_preserves_parameterConcavity {m n : ℕ} (ε : {r : ℝ // 0 < r}) (v : Fin n → ℝ) {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hConc : ∀ w : Fin n → ℝ, IsERealConcaveOn (Set.univ : Set (Fin m → ℝ)) (fun u => K u w)) : IsERealConcaveOn (Set.univ : Set (Fin m → ℝ)) (fun u => ⨅ w : {w : Fin n → ℝ // ‖w - v‖ < ε.1}, K u w.1)

Lean source

Helper for Lemma33.0.5: at a fixed radius in the first variable, local suprema preserve convexity in the second variable.

lemma helperForLemma33_0_5_fixedRadiusLocalSupremum_preserves_parameterConvexity {m n : ℕ} (ε : {r : ℝ // 0 < r}) (u : Fin m → ℝ) {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hConv : ∀ w : Fin m → ℝ, IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (fun v => K w v)) : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (fun v => ⨆ w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}, K w.1 v)

Lean source

Helper for Lemma33.0.5: the one-variable concave closure obtained from local suprema remains concave.

lemma helperForLemma33_0_5_functionConcaveClosure_preserves_concavity {n : ℕ} {f : (Fin n → ℝ) → EReal} (hConc : IsERealConcaveOn (Set.univ : Set (Fin n → ℝ)) f) : IsERealConcaveOn (Set.univ : Set (Fin n → ℝ)) (fun x => ⨅ ε : {r : ℝ // 0 < r}, ⨆ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1)

Lean source

Helper for Lemma33.0.5: the one-variable convex closure obtained from local infima remains convex.

lemma helperForLemma33_0_5_functionConvexClosure_convexFunction {n : ℕ} {f : (Fin n → ℝ) → EReal} (hConv : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) f) : ConvexFunction (fun x => ⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1)

Lean source

Helper for Lemma33.0.5: the raw sup-inf closure operator is idempotent.

lemma helperForLemma33_0_5_functionConvexClosure_raw_idempotent {n : ℕ} {f : (Fin n → ℝ) → EReal} (x : Fin n → ℝ) : (⨆ δ : {r : ℝ // 0 < r}, ⨅ z : {z : Fin n → ℝ // ‖z - x‖ < δ.1}, (⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - z.1‖ < ε.1}, f w.1)) = (⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1)

Lean source

Helper for Lemma33.0.5: the raw sup-inf closure operator is lower semicontinuous.

lemma helperForLemma33_0_5_functionConvexClosure_lowerSemicontinuous {n : ℕ} {f : (Fin n → ℝ) → EReal} : LowerSemicontinuous (fun x => ⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1)

Lean source

Helper for Lemma33.0.5: the raw sup-inf closure never exceeds the original function, because every ball contains its own center.

lemma helperForLemma33_0_5_functionConvexClosure_raw_le_self {n : ℕ} {f : (Fin n → ℝ) → EReal} (x : Fin n → ℝ) : (⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1) ≤ f x

Lean source

Helper for Lemma33.0.5: if the raw sup-inf closure takes the value ⊤ at some point and all of its values are already classified into {⊤, ⊥}, then a whole neighborhood is forced to stay at ⊤.

lemma helperForLemma33_0_5_functionConvexClosure_top_has_topNeighborhood {n : ℕ} {f : (Fin n → ℝ) → EReal} {y : Fin n → ℝ} (hTopOrBot : ∀ x, (⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1) = (⊤ : EReal) ∨ (⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1) = (⊥ : EReal)) (hyTop : (⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - y‖ < ε.1}, f w.1) = (⊤ : EReal)) : ∃ δ : {r : ℝ // 0 < r}, ∀ z : {z : Fin n → ℝ // ‖z - y‖ < δ.1}, (⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - z.1‖ < ε.1}, f w.1) = (⊤ : EReal)

Lean source

Helper for Lemma33.0.5: a lower semicontinuous convex function on univ that already attains ⊥ is improper, so Chapter 2 forces all of its values to lie in {⊤, ⊥}.

lemma helperForLemma33_0_5_closedImproperConvex_values_top_or_bot {n : ℕ} {g : (Fin n → ℝ) → EReal} (hConv : ConvexFunction g) (hLsc : LowerSemicontinuous g) (hBot : ∃ x, g x = ⊥) : ∀ x, g x = (⊤ : EReal) ∨ g x = (⊥ : EReal)

Lean source

Helper for Lemma33.0.5: once a {⊤, ⊥}-valued raw closure equals ⊥ at x, every positive-radius ball around x already contains an exact ⊥ witness.

lemma helperForLemma33_0_5_topBotValued_rawClosure_eq_bot_implies_everyBall_has_botWitness {n : ℕ} {g : (Fin n → ℝ) → EReal} {x : Fin n → ℝ} (hTopOrBot : ∀ z, g z = (⊤ : EReal) ∨ g z = (⊥ : EReal)) (hRawIdem : (⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, g w.1) = g x) (hxBot : g x = (⊥ : EReal)) : ∀ ε : {r : ℝ // 0 < r}, ∃ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, g w.1 = (⊥ : EReal)

Lean source

Helper for Lemma33.0.5: a strict convex combination of an exact ⊥ witness at the x endpoint and the ⊤ endpoint value at y cannot land inside a neighborhood where the target function is identically ⊤.

lemma helperForLemma33_0_5_topNeighborhood_contradicts_botWitnessUnderConvexity {n : ℕ} {g : (Fin n → ℝ) → EReal} {x y : Fin n → ℝ} {a b : ℝ} (hConv : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) g) (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) (hPosA : 0 < a) {δ : {r : ℝ // 0 < r}} (hTopNeighborhood : ∀ z : {z : Fin n → ℝ // ‖z - (a • x + b • y)‖ < δ.1}, g z.1 = ⊤) {w : Fin n → ℝ} (hw : ‖w - x‖ < δ.1) (hwBot : g w = ⊥) (hyTop : g y = ⊤) : False

Lean source

Helper for Lemma33.0.5: once the raw sup-inf closure is already known to satisfy Jensen, the mixed (⊥, ⊤) branch collapses by combining top neighborhoods at ⊤ points with exact ⊥ witnesses in every ball around a ⊥ point.

lemma helperForLemma33_0_5_functionConvexClosure_mixedBotTop_collapse_from_rawClassification {n : ℕ} {f : (Fin n → ℝ) → EReal} {x y : Fin n → ℝ} {a b : ℝ} (hClosureConv : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (fun z => ⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - z‖ < ε.1}, f w.1)) (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) (hPosA : 0 < a) (_hPosB : 0 < b) (hClosureXBot : (⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1) = ⊥) (hClosureYTop : (⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - y‖ < ε.1}, f w.1) = ⊤) : (⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - (a • x + b • y)‖ < ε.1}, f w.1) = ⊥

Lean source

Helper for Lemma33.0.5: for a convex epigraph, an exact ⊥ left endpoint and any right endpoint different from ⊤ already force every strict convex combination to be ⊥.

lemma helperForLemma33_0_5_convexFunction_leftBot_rightNotTop_forces_comboBot {n : ℕ} {g : (Fin n → ℝ) → EReal} {x y : Fin n → ℝ} {a b : ℝ} (hConvFun : ConvexFunction g) (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) (hPosA : 0 < a) (hxBot : g x = ⊥) (hyNeTop : g y ≠ ⊤) : g (a • x + b • y) = ⊥

§33 — Saddle-Functions — Part 3

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

A concave-convex kernel in the literal Rockafellar epi/hypograph sense.

def IsEpigraphHypographConcaveConvex {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := (∀ v, Convex ℝ (extendedRealHypograph (fun u => K u v))) ∧ ∀ u, ConvexFunction (fun v => K u v)

Lean source

The symmetric epi/hypograph notion for a kernel convex in its first argument and concave in its second argument.

def IsEpigraphHypographConvexConcave {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := IsEpigraphHypographConcaveConvex (fun v u => K u v)

Lean source

The strong concave Jensen predicate implies convexity of the real-height hypograph.

lemma helperForLemma33_0_5_isERealConcaveOn_univ_to_hypographConvex {n : ℕ} {f : (Fin n → ℝ) → EReal} (hConc : IsERealConcaveOn (Set.univ : Set (Fin n → ℝ)) f) : Convex ℝ (extendedRealHypograph f)

Lean source

Corollary33.1.1 in the book's epi/hypograph semantics: both coordinatewise closures remain concave-convex, including the improper mixed endpoint cases.

lemma isConcaveConvexOn_univ_closureData_closures {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K) : IsEpigraphHypographConcaveConvex (convexConcaveClosureData K).cl_v ∧ IsEpigraphHypographConcaveConvex (convexConcaveClosureData K).cl_u

Lean source

Lemma33.0.5: If K : ℝ^m × ℝ^n → EReal is concave-convex, then the closure operators from Definition33.0.4 produce bifunctions (cl_v K) and (cl_u K) that are also concave-convex.

lemma isConcaveConvexOn_univ_closureData_closures_of_noBot {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K) (hNoBot : ∀ u v, K u v ≠ (⊥ : EReal)) : IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (convexConcaveClosureData K).cl_v ∧ IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (convexConcaveClosureData K).cl_u

Lean source

A real-valued bifunction on ℝ^m × ℝ^n is bilinear when it is additive and homogeneous in each variable separately.

def IsBilinearOnReal {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) : Prop := (∀ u₁ u₂ xStar, K (u₁ + u₂) xStar = K u₁ xStar + K u₂ xStar) ∧ (∀ a u xStar, K (a • u) xStar = a * K u xStar) ∧ (∀ u xStar₁ xStar₂, K u (xStar₁ + xStar₂) = K u xStar₁ + K u xStar₂) ∧ ∀ a u xStar, K u (a • xStar) = a * K u xStar

Lean source

Theorem33.0.6: Let A : ℝ^m → ℝ^n be linear and define K(u, x^*) = ⟪A u, x^*⟫. Then K is bilinear on ℝ^m × ℝ^n. Conversely, if K : ℝ^m × ℝ^n → ℝ is bilinear, then there exists a unique linear map A : ℝ^m → ℝ^n such that K(u, x^*) = (A u) ⬝ᵥ x^* for all u and x^*, where ⬝ᵥ is the standard Euclidean pairing on coordinate vectors.

theorem bilinear_forms_and_linear_maps {m n : ℕ} : (∀ A : (Fin m → ℝ) →ₗ[ℝ] (Fin n → ℝ), IsBilinearOnReal (fun u xStar => dotProduct (A u) xStar)) ∧ (∀ K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ, IsBilinearOnReal K → ∃! A : (Fin m → ℝ) →ₗ[ℝ] (Fin n → ℝ), ∀ u xStar, K u xStar = dotProduct (A u) xStar)

Lean source

Corollary33.0.7: Any bifunction of the form K(u, x^*) = ⟪A u, x^*⟫ with A : ℝ^m → ℝ^n is both concave-convex and convex-concave on ℝ^m × ℝ^n, hence is a saddle function.

theorem bilinear_pairing_is_saddle {m n : ℕ} (A : (Fin m → ℝ) →ₗ[ℝ] (Fin n → ℝ)) : IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (fun u xStar => ((dotProduct (A u) xStar : ℝ) : EReal)) ∧ IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (fun u xStar => ((dotProduct (A u) xStar : ℝ) : EReal)) ∧ IsSaddleFunctionOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (fun u xStar => ((dotProduct (A u) xStar : ℝ) : EReal))

Lean source

Helper for Lemma33.0.5: after the route correction, the raw mixed (⊥, ⊤) branch closes as soon as the raw closure itself is already known to satisfy Jensen's inequality.

lemma helperForLemma33_0_5_functionConvexClosure_closedConvex_mixedBotTop_bridge {n : ℕ} {f : (Fin n → ℝ) → EReal} {x y : Fin n → ℝ} {a b : ℝ} (hClosureConv : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (fun z => ⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - z‖ < ε.1}, f w.1)) (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) (hPosA : 0 < a) (hPosB : 0 < b) (hClosureXBot : (⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1) = ⊥) (hClosureYTop : (⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - y‖ < ε.1}, f w.1) = ⊤) : (⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - (a • x + b • y)‖ < ε.1}, f w.1) = ⊥

Lean source

Helper for Lemma33.0.5: the remaining raw cl_v blocker is the target-specific mixed (⊥, ⊤) Jensen branch for the raw closure, stated directly in terms of the original convexity assumption on f.

lemma helperForLemma33_0_5_functionConvexClosure_preserves_convexity_mixedBotTop_case {n : ℕ} {f : (Fin n → ℝ) → EReal} {x y : Fin n → ℝ} {a b : ℝ} (hConv : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) f) (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) (hPosA : 0 < a) (hPosB : 0 < b) (hClosureXBot : (⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1) = ⊥) (hClosureYTop : (⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - y‖ < ε.1}, f w.1) = ⊤) : (⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - (a • x + b • y)‖ < ε.1}, f w.1) = ⊥

Lean source

Helper for Lemma33.0.5: the one-variable convex closure obtained from local infima remains convex.

lemma helperForLemma33_0_5_functionConvexClosure_preserves_convexity {n : ℕ} {f : (Fin n → ℝ) → EReal} (hConv : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) f) : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (fun x => ⨆ ε : {r : ℝ // 0 < r}, ⨅ w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}, f w.1)

Lean source

Helper for Lemma33.0.5: convexClosureInSecond preserves concavity in the first variable.

lemma helperForLemma33_0_5_convexClosureInSecond_preserves_firstVariable_concavity {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (v : Fin n → ℝ) (hConc : ∀ w : Fin n → ℝ, IsERealConcaveOn (Set.univ : Set (Fin m → ℝ)) (fun u => K u w)) : IsERealConcaveOn (Set.univ : Set (Fin m → ℝ)) (fun u => convexClosureInSecond K u v)

Lean source

Helper for Lemma33.0.5: if the outer u-closure equals ⊥, then every strict upper bound on ⊥ is beaten by some fixed-radius local supremum.

lemma helperForLemma33_0_5_concaveClosureInFirst_eq_bot_implies_exists_radius_localSup_lt {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {x : Fin n → ℝ} (hBot : concaveClosureInFirst K u x = ⊥) : ∀ {z : EReal}, (⊥ : EReal) < z → ∃ ε : {r : ℝ // 0 < r}, (⨆ w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}, K w.1 x) < z

Lean source

Helper for Lemma33.0.5: if the outer u-closure equals ⊤, then every fixed-radius local supremum at that point already equals ⊤.

lemma helperForLemma33_0_5_concaveClosureInFirst_eq_top_implies_everyRadius_localSup_eq_top {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {x : Fin n → ℝ} (hTop : concaveClosureInFirst K u x = ⊤) : ∀ ε : {r : ℝ // 0 < r}, (⨆ w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}, K w.1 x) = ⊤

Lean source

Helper for Lemma33.0.5: if some fixed radius already has local supremum ⊥ on the x endpoint, then the outer u-closure at the convex combination is also ⊥.

lemma helperForLemma33_0_5_concaveClosureInFirst_radiusBot_forces_targetBot {m n : ℕ} (ε : {r : ℝ // 0 < r}) {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {x y : Fin n → ℝ} {a b : ℝ} (hConv : ∀ w : Fin m → ℝ, IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (fun v => K w v)) (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) (hPosA : 0 < a) (hLocalSupXBot : (⨆ w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}, K w.1 x) = ⊥) : concaveClosureInFirst K u (a • x + b • y) = ⊥

Lean source

Helper for Lemma33.0.5: for fixed u, the outer inf-sup closure is convex in the second variable when each section v ↦ K w v is convex.

lemma helperForLemma33_0_5_concaveClosureInFirst_convexFunction {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (hConv : ∀ w : Fin m → ℝ, IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (fun v => K w v)) : ConvexFunction (fun v => concaveClosureInFirst K u v)

Lean source

Helper for Lemma33.0.5: the target outer inf-sup value is controlled by a product-indexed infimum over independent endpoint radii.

lemma helperForLemma33_0_5_concaveClosureInFirst_target_le_productInfimum {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {x y : Fin n → ℝ} {a b : ℝ} (hConv : ∀ w : Fin m → ℝ, IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (fun v => K w v)) (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) : concaveClosureInFirst K u (a • x + b • y) ≤ ⨅ p : {r : ℝ // 0 < r} × {r : ℝ // 0 < r}, (a : EReal) * (⨆ w : {w : Fin m → ℝ // ‖w - u‖ < p.1.1}, K w.1 x) + (b : EReal) * (⨆ w : {w : Fin m → ℝ // ‖w - u‖ < p.2.1}, K w.1 y)

Lean source

Helper for Lemma33.0.5: outside the two mixed (⊥, ⊤) corners, the product-indexed infimum of weighted outer-radius local suprema is bounded by the weighted sum of the outer closure values.

lemma helperForLemma33_0_5_concaveClosureInFirst_productInfimum_nonexceptionalBound {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {x y : Fin n → ℝ} {a b : ℝ} (ha : 0 < a) (hb : 0 < b) (h₁ : concaveClosureInFirst K u x ≠ (⊥ : EReal) ∨ concaveClosureInFirst K u y ≠ (⊤ : EReal)) (h₂ : concaveClosureInFirst K u x ≠ (⊤ : EReal) ∨ concaveClosureInFirst K u y ≠ (⊥ : EReal)) : (⨅ p : {r : ℝ // 0 < r} × {r : ℝ // 0 < r}, (a : EReal) * (⨆ w : {w : Fin m → ℝ // ‖w - u‖ < p.1.1}, K w.1 x) + (b : EReal) * (⨆ w : {w : Fin m → ℝ // ‖w - u‖ < p.2.1}, K w.1 y)) ≤ (a : EReal) * concaveClosureInFirst K u x + (b : EReal) * concaveClosureInFirst K u y

Lean source

Helper for Lemma33.0.5: the true outer (⊥, ⊤) branch for concaveClosureInFirst is where the current local-radius route stops.

lemma helperForLemma33_0_5_concaveClosureInFirst_commonRadius_mixedBotTop_bridge {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {x y : Fin n → ℝ} {a b : ℝ} (hConv : ∀ w : Fin m → ℝ, IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (fun v => K w v)) (hCenterNoBot : K u x ≠ (⊥ : EReal)) (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) (hPosA : 0 < a) (hPosB : 0 < b) (hClosureXBot : (⨅ ε : {r : ℝ // 0 < r}, ⨆ w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}, K w.1 x) = ⊥) (hClosureYTop : (⨅ ε : {r : ℝ // 0 < r}, ⨆ w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}, K w.1 y) = ⊤) : concaveClosureInFirst K u (a • x + b • y) = ⊥

Lean source

Helper for Lemma33.0.5: the true outer (⊥, ⊤) branch for concaveClosureInFirst is where the current local-radius route stops.

lemma helperForLemma33_0_5_concaveClosureInFirst_exceptional_branch {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {x y : Fin n → ℝ} {a b : ℝ} (hConv : ∀ w : Fin m → ℝ, IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (fun v => K w v)) (hCenterNoBot : K u x ≠ (⊥ : EReal)) (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) (hPosA : 0 < a) (hPosB : 0 < b) (hClosureXBot : (⨅ ε : {r : ℝ // 0 < r}, ⨆ w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}, K w.1 x) = ⊥) (hClosureYTop : (⨅ ε : {r : ℝ // 0 < r}, ⨆ w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}, K w.1 y) = ⊤) : concaveClosureInFirst K u (a • x + b • y) ≤ (a : EReal) * (⨅ ε : {r : ℝ // 0 < r}, ⨆ w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}, K w.1 x) + (b : EReal) * (⨅ ε : {r : ℝ // 0 < r}, ⨆ w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}, K w.1 y)

Lean source

Helper for Lemma33.0.5: concaveClosureInFirst preserves convexity in the second variable.

lemma helperForLemma33_0_5_concaveClosureInFirst_preserves_secondVariable_convexity_of_noBot {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (hConv : ∀ w : Fin m → ℝ, IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (fun v => K w v)) (hNoBot : ∀ w v, K w v ≠ (⊥ : EReal)) : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (fun v => concaveClosureInFirst K u v)

Lean source

Helper for Theorem33.0.6: the pairing induced by a linear map is bilinear in both variables.

lemma helperForTheorem33_0_6_pairing_bilinear_of_linearMap {m n : ℕ} (A : (Fin m → ℝ) →ₗ[ℝ] (Fin n → ℝ)) : IsBilinearOnReal (fun u xStar => dotProduct (A u) xStar)

Lean source

Helper for Theorem33.0.6: bilinearity in the second variable forces K u 0 = 0.

lemma helperForTheorem33_0_6_secondVariable_zero {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hK : IsBilinearOnReal K) (u : Fin m → ℝ) : K u 0 = 0

Lean source

Helper for Theorem33.0.6: bilinearity expands K u across a finite sum in the second variable.

lemma helperForTheorem33_0_6_coordinateSum_in_second_variable {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hK : IsBilinearOnReal K) (u : Fin m → ℝ) (s : Finset (Fin n)) (c : Fin n → ℝ) : K u (∑ i ∈ s, c i • Pi.single (M := fun _ : Fin n => ℝ) i 1) = ∑ i ∈ s, c i * K u (Pi.single (M := fun _ : Fin n => ℝ) i 1)

Lean source

Helper for Theorem33.0.6: a bilinear kernel admits a coordinate expansion in the second variable.

lemma helperForTheorem33_0_6_coordinateExpansion_in_second_variable {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hK : IsBilinearOnReal K) (u : Fin m → ℝ) (xStar : Fin n → ℝ) : K u xStar = ∑ i, xStar i * K u (Pi.single (M := fun _ : Fin n => ℝ) i 1)

Lean source

Helper for Theorem33.0.6: the coordinate reconstruction is additive in the source variable.

lemma helperForTheorem33_0_6_reconstructedLinearMap_map_add {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hK : IsBilinearOnReal K) (u₁ u₂ : Fin m → ℝ) : (fun i => K (u₁ + u₂) (Pi.single (M := fun _ : Fin n => ℝ) i 1)) = (fun i => K u₁ (Pi.single (M := fun _ : Fin n => ℝ) i 1) + K u₂ (Pi.single (M := fun _ : Fin n => ℝ) i 1))

Lean source

Helper for Theorem33.0.6: the coordinate reconstruction is homogeneous in the source variable.

lemma helperForTheorem33_0_6_reconstructedLinearMap_map_smul {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hK : IsBilinearOnReal K) (a : ℝ) (u : Fin m → ℝ) : (fun i => K (a • u) (Pi.single (M := fun _ : Fin n => ℝ) i 1)) = a • (fun i : Fin n => K u (Pi.single (M := fun _ : Fin n => ℝ) i 1))

Lean source

Helper for Theorem33.0.6: a bilinear kernel reconstructs a linear map from its coordinate values.

def helperForTheorem33_0_6_reconstructedLinearMap {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) (hK : IsBilinearOnReal K) : (Fin m → ℝ) →ₗ[ℝ] (Fin n → ℝ) := { toFun := fun u i => K u (Pi.single (M := fun _ : Fin n => ℝ) i 1) map_add' := helperForTheorem33_0_6_reconstructedLinearMap_map_add hK map_smul' := helperForTheorem33_0_6_reconstructedLinearMap_map_smul hK }

Lean source

Helper for Theorem33.0.6: the reconstructed coordinate map represents the original bilinear kernel.

lemma helperForTheorem33_0_6_reconstructedLinearMap_represents_kernel {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hK : IsBilinearOnReal K) : ∀ u xStar, K u xStar = dotProduct (helperForTheorem33_0_6_reconstructedLinearMap K hK u) xStar

Lean source

Helper for Theorem33.0.6: a representing linear map is determined by testing against basis vectors.

lemma helperForTheorem33_0_6_representation_unique {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {A B : (Fin m → ℝ) →ₗ[ℝ] (Fin n → ℝ)} (hA : ∀ u xStar, K u xStar = dotProduct (A u) xStar) (hB : ∀ u xStar, K u xStar = dotProduct (B u) xStar) : A = B

Lean source

Helper for Corollary33.0.7: the first-variable Jensen combination for the bilinear pairing is an exact equality after coercing the real-valued pairing into EReal.

lemma helperForCorollary33_0_7_firstVariable_jensen_eq {m n : ℕ} (A : (Fin m → ℝ) →ₗ[ℝ] (Fin n → ℝ)) (x y : Fin m → ℝ) (xStar : Fin n → ℝ) (a b : ℝ) : ((dotProduct (A (a • x + b • y)) xStar : ℝ) : EReal) = (a : EReal) * ((dotProduct (A x) xStar : ℝ) : EReal) + (b : EReal) * ((dotProduct (A y) xStar : ℝ) : EReal)

Lean source

Helper for Corollary33.0.7: the second-variable Jensen combination for the bilinear pairing is likewise an exact equality after coercion to EReal.

lemma helperForCorollary33_0_7_secondVariable_jensen_eq {m n : ℕ} (A : (Fin m → ℝ) →ₗ[ℝ] (Fin n → ℝ)) (u : Fin m → ℝ) (xStar₁ xStar₂ : Fin n → ℝ) (a b : ℝ) : ((dotProduct (A u) (a • xStar₁ + b • xStar₂) : ℝ) : EReal) = (a : EReal) * ((dotProduct (A u) xStar₁ : ℝ) : EReal) + (b : EReal) * ((dotProduct (A u) xStar₂ : ℝ) : EReal)

Lean source

Helper for Corollary33.0.7: the exact Jensen identities in each variable package into both the concave-convex and convex-concave orientations of the bilinear pairing kernel.

lemma helperForCorollary33_0_7_concaveConvex_and_convexConcave {m n : ℕ} (A : (Fin m → ℝ) →ₗ[ℝ] (Fin n → ℝ)) : IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (fun u xStar => ((dotProduct (A u) xStar : ℝ) : EReal)) ∧ IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (fun u xStar => ((dotProduct (A u) xStar : ℝ) : EReal))

§33 — Saddle-Functions — Part 4

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

The two branch functions underlying the conjugate-pairing convention attached to an EReal-valued function on ℝ^n.

structure ConjugatePairingValues (n : ℕ) where convex : (Fin n → ℝ) → EReal concave : (Fin n → ℝ) → EReal

Lean source

Definition33.0.8: For f : ℝ^n → EReal, this records the two conjugate conventions used in the text. When f is convex, f^*(x^*) is the supremum of ⟪x, x^*⟫ - f x over x ∈ ℝ^n; when f is concave, f^*(x^*) is the infimum of the same expression. In either chosen convention, the pairing notation is ⟪f, x^*⟫ = ⟪x^*, f⟫ = f^*(x^*).

noncomputable def conjugatePairingNotation {n : ℕ} (f : (Fin n → ℝ) → EReal) : ConjugatePairingValues n := { convex := fun xStar => sSup <| Set.range fun x : Fin n → ℝ => (((dotProduct x xStar : ℝ) : EReal) - f x) concave := fun xStar => sInf <| Set.range fun x : Fin n → ℝ => (((dotProduct x xStar : ℝ) : EReal) - f x) }

Lean source

The indicator of a point in ℝ^n, taking the value 0 at that point and ⊤ elsewhere.

noncomputable def pointIndicator {n : ℕ} (x : Fin n → ℝ) : (Fin n → ℝ) → EReal := fun z => if z = x then 0 else ⊤ -- Proof sketch: unfold `convexConjugate` and the definition of `pointIndicator`; the term -- corresponding to `z = x` contributes `⟪x, x^*⟫`, while every other term is -- `⟪z, x^*⟫ - ⊤ = ⊥`, so the supremum is attained at `x`.

Lean source

Lemma33.0.9: Indicator of a point. For every x* ∈ ℝ^n, the convex conjugate pairing of the point indicator at x equals the Euclidean pairing ⟪x, x*⟫.

theorem convexConjugate_pointIndicator {n : ℕ} (x xStar : Fin n → ℝ) : convexConjugate (pointIndicator x) xStar = ((dotProduct x xStar : ℝ) : EReal)

Lean source

A Rockafellar convex bifunction is sectionwise convex in the image variable. This is separate from the Chapter 29 graph-convex notion.

def IsRockafellarSectionwiseConvexBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := ∀ u, IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (F u)

Lean source

A bifunction is concave when each section F u is concave on ℝ^n.

def IsConcaveBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := ∀ u, IsERealConcaveOn (Set.univ : Set (Fin n → ℝ)) (F u)

Lean source

Definition33.0.10: For a bifunction F : ℝ^m → (ℝ^n → EReal), this records the sectionwise conjugate-pairing convention ⟪F u, x^*⟫ = (F u)^*(x^*), with the convex branch given by a supremum and the concave branch given by an infimum.

noncomputable def bifunctionPairingNotation {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin m → ℝ) → ConjugatePairingValues n := fun u => conjugatePairingNotation (F u)

Lean source

The convex branch of the bifunction pairing, used when F is a convex bifunction.

noncomputable abbrev convexBifunctionPairing {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (u : Fin m → ℝ) (xStar : Fin n → ℝ) : EReal := convexConjugate (F u) xStar

Lean source

The concave branch of the bifunction pairing, used when F is a concave bifunction.

noncomputable abbrev concaveBifunctionPairing {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (u : Fin m → ℝ) (xStar : Fin n → ℝ) : EReal := (bifunctionPairingNotation F u).concave xStar

Lean source

The convex conjugate pairing attached to F is concave in the parameter variable for every dual vector x^*.

def HasConcaveParameterConvexPairing {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := ∀ xStar : Fin n → ℝ, IsERealConcaveOn (Set.univ : Set (Fin m → ℝ)) (fun u => convexBifunctionPairing F u xStar)

Lean source

The concave conjugate pairing attached to F is convex in the parameter variable for every dual vector x^*.

def HasConvexParameterConcavePairing {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := ∀ xStar : Fin n → ℝ, IsERealConvexOn (Set.univ : Set (Fin m → ℝ)) (fun u => concaveBifunctionPairing F u xStar)

Lean source

Rockafellar's convex bifunctions are sectionwise convex and have a conjugate pairing that is concave in the parameter variable.

def IsRockafellarConvexBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := IsRockafellarSectionwiseConvexBifunction F ∧ HasConcaveParameterConvexPairing F

Lean source

Rockafellar's concave bifunctions are sectionwise concave and have a conjugate pairing that is convex in the parameter variable.

def IsRockafellarConcaveBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := IsConcaveBifunction F ∧ HasConvexParameterConcavePairing F -- Proof sketch: specialize Lemma33.0.9 to the point `A u`. The section `F u` is exactly the -- indicator of the singleton `{A u}`, so its convex conjugate pairing at `xStar` reduces to -- the Euclidean pairing with `A u`.

Lean source

Lemma33.0.11: If F u is the indicator of the point A u, then the convex bifunction pairing ⟪F u, x^*⟫ equals ⟪A u, x^*⟫ for all (u, x^*) ∈ ℝ^m × ℝ^n.

theorem convexBifunctionPairing_indicator_linearMap {m n : ℕ} (A : (Fin m → ℝ) →ₗ[ℝ] (Fin n → ℝ)) (u : Fin m → ℝ) (xStar : Fin n → ℝ) : convexBifunctionPairing (fun u' => pointIndicator (A u')) u xStar = ((dotProduct (A u) xStar : ℝ) : EReal)

Lean source

Theorem 33.12 (Biconjugates of convex bifunctions).

theorem section33_theorem33_12 : True

Lean source

Corollary 33.13.

theorem section33_corollary33_13 : True

Lean source

The graph function attached to a bifunction on ℝ^m × ℝ^n, viewed as a function on ℝ^(m + n) by splitting coordinates into their first m and last n entries.

def graphFunctionOfBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin (m + n) → ℝ) → EReal := fun z => F (fun i => z (Fin.castAdd n i)) (fun j => z (Fin.natAdd m j))

Lean source

The bifunction obtained from a function on ℝ^(m + n) by currying along the coordinate splitting ℝ^(m + n) ≃ ℝ^m × ℝ^n.

def bifunctionOfGraphFunction {m n : ℕ} (f : (Fin (m + n) → ℝ) → EReal) : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u x => f (Fin.append u x)

Lean source

A Rockafellar convex bifunction is polyhedral when its graph function on ℝ^(m+n) is a polyhedral convex function.

def IsRockafellarPolyhedralConvexBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := IsRockafellarConvexBifunction F ∧ IsPolyhedralConvexFunction (m + n) (graphFunctionOfBifunction F) -- Proof sketch: identify `ℝ^(m + n)` with `ℝ^m × ℝ^n` via coordinate splitting and observe -- that `graphFunctionOfBifunction` and `bifunctionOfGraphFunction` are inverse operations. -- Under this identification, Rockafellar's convex-bifunction condition is exactly convexity -- of the uncurried graph function on the whole space.

Lean source

Helper for Lemma33.0.14: currying back the graph function of a bifunction recovers the original bifunction.

lemma helperForLemma33_0_14_bifunctionOfGraphFunction_graphFunctionOfBifunction_eq {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : bifunctionOfGraphFunction (graphFunctionOfBifunction F) = F

Lean source

Helper for Lemma33.0.14: if an index i : Fin (m+n) lies in the second block (m ≤ i.1), then i.1 - m is a valid Fin n index.

lemma helperForLemma33_0_14_sub_lt_of_le {m n : ℕ} (i : Fin (m + n)) (hge : m ≤ i.1) : i.1 - m < n

Lean source

Helper for Lemma33.0.14: splitting a graph point into parameter and section coordinates and then re-appending them recovers the original graph function.

lemma helperForLemma33_0_14_append_split_eq {m n : ℕ} (z : Fin (m + n) → ℝ) : Fin.append (fun i => z (Fin.castAdd n i)) (fun j => z (Fin.natAdd m j)) = z

Lean source

Helper for Lemma33.0.14: splitting a graph point into parameter and section coordinates and then re-appending them recovers the original graph function.

lemma helperForLemma33_0_14_graphFunctionOfBifunction_bifunctionOfGraphFunction_eq {m n : ℕ} (f : (Fin (m + n) → ℝ) → EReal) : graphFunctionOfBifunction (bifunctionOfGraphFunction f) = f

Lean source

Helper for Lemma33.0.14: the uncurrying operation graphFunctionOfBifunction is injective because currying back with bifunctionOfGraphFunction is a left inverse.

lemma helperForLemma33_0_14_graphFunctionOfBifunction_injective {m n : ℕ} : Function.Injective (graphFunctionOfBifunction (m := m) (n := n))

Lean source

Helper for Lemma33.0.14: the currying operation bifunctionOfGraphFunction is injective because uncurrying back with graphFunctionOfBifunction is a left inverse.

lemma helperForLemma33_0_14_bifunctionOfGraphFunction_injective {m n : ℕ} : Function.Injective (bifunctionOfGraphFunction (m := m) (n := n))

Lean source

Helper for Lemma33.0.14: graph convexity restricts to convexity of each section at fixed parameter u.

lemma helperForLemma33_0_14_sectionwiseConvex_of_graphConvex {m n : ℕ} {f : (Fin (m + n) → ℝ) → EReal} (hf : IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) f) : IsRockafellarSectionwiseConvexBifunction (bifunctionOfGraphFunction f)

Lean source

Helper for Lemma33.0.14: block coordinates commute with convex combinations under Fin.append.

lemma helperForLemma33_0_14_append_weighted {m n : ℕ} (a b : ℝ) (u₁ u₂ : Fin m → ℝ) (x₁ x₂ : Fin n → ℝ) : a • Fin.append u₁ x₁ + b • Fin.append u₂ x₂ = Fin.append (a • u₁ + b • u₂) (a • x₁ + b • x₂)

Lean source

Helper for Lemma33.0.14: the dot product with a fixed dual vector is affine in the section block.

lemma helperForLemma33_0_14_dotProduct_weighted {n : ℕ} (a b : ℝ) (x₁ x₂ xStar : Fin n → ℝ) : dotProduct (a • x₁ + b • x₂) xStar = a * dotProduct x₁ xStar + b * dotProduct x₂ xStar

Lean source

Helper for Lemma33.0.14: multiplying a finite affine term by a nonnegative scalar commutes with subtracting an arbitrary extended-real value.

lemma helperForLemma33_0_14_nonneg_mul_sub_finite {a r : ℝ} (ha : 0 ≤ a) (x : EReal) : ((a : EReal) * (((r : ℝ) : EReal) - x)) = ((a * r : ℝ) : EReal) - ((a : EReal) * x)

Lean source

Helper for Lemma33.0.14: two affine finite-minus-extended-real terms are controlled by the single affine term obtained after adding the finite parts and the subtrahends.

lemma helperForLemma33_0_14_add_sub_le_sub_add {p q : ℝ} (C D : EReal) : (((p : ℝ) : EReal) - C) + (((q : ℝ) : EReal) - D) ≤ (((p + q : ℝ) : EReal) - (C + D))

Lean source

Helper for Lemma33.0.14: the weighted endpoint kernel sum that appears in the Jensen step for the convex pairing.

noncomputable def helperForLemma33_0_14_weightedGraphKernelLeftSide {m n : ℕ} (f : (Fin (m + n) → ℝ) → EReal) (u₁ u₂ : Fin m → ℝ) (x₁ x₂ xStar : Fin n → ℝ) (a b : ℝ) : EReal := (a : EReal) * ((((dotProduct x₁ xStar : ℝ) : EReal) - f (Fin.append u₁ x₁))) + (b : EReal) * ((((dotProduct x₂ xStar : ℝ) : EReal) - f (Fin.append u₂ x₂)))

Lean source

Helper for Lemma33.0.14: the target kernel at the convex combination of the endpoint data.

noncomputable def helperForLemma33_0_14_weightedGraphKernelRightSide {m n : ℕ} (f : (Fin (m + n) → ℝ) → EReal) (u₁ u₂ : Fin m → ℝ) (x₁ x₂ xStar : Fin n → ℝ) (a b : ℝ) : EReal := (((dotProduct (a • x₁ + b • x₂) xStar : ℝ) : EReal) - f (Fin.append (a • u₁ + b • u₂) (a • x₁ + b • x₂)))

Lean source

Helper for Lemma33.0.14: convexity of the graph function gives the pointwise Jensen estimate for the affine kernels appearing in the convex pairing.

lemma helperForLemma33_0_14_weightedGraphKernel_le_targetKernel {m n : ℕ} {f : (Fin (m + n) → ℝ) → EReal} (hf : IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) f) {u₁ u₂ : Fin m → ℝ} {x₁ x₂ xStar : Fin n → ℝ} {a b : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) : helperForLemma33_0_14_weightedGraphKernelLeftSide f u₁ u₂ x₁ x₂ xStar a b ≤ helperForLemma33_0_14_weightedGraphKernelRightSide f u₁ u₂ x₁ x₂ xStar a b

Lean source

Helper for Lemma33.0.14: a jointly convex graph function yields the parameterwise concavity required in Rockafellar's convex-bifunction definition after currying.

lemma helperForLemma33_0_14_concaveParameterPairing_of_graphConvex {m n : ℕ} {f : (Fin (m + n) → ℝ) → EReal} (hf : IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) f) : HasConcaveParameterConvexPairing (bifunctionOfGraphFunction f)

Lean source

Helper for Lemma33.0.14: every section point contributes one of the affine kernels in the defining supremum for the convex bifunction pairing.

lemma helperForLemma33_0_14_endpointKernel_le_convexBifunctionPairing {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (u : Fin m → ℝ) (x xStar : Fin n → ℝ) : (((dotProduct x xStar : ℝ) : EReal) - F u x) ≤ convexBifunctionPairing F u xStar

Lean source

Helper for Lemma33.0.14: the Section 33 convex-conjugate notation is exactly Chapter 3's Fenchel conjugate.

lemma helperForLemma33_0_14_convexConjugate_eq_fenchelConjugate {n : ℕ} (f : (Fin n → ℝ) → EReal) : convexConjugate f = fenchelConjugate n f

Lean source

Helper for Lemma33.0.14: every Rockafellar-convex section satisfies the generic Fenchel-Moreau biconjugacy formula, but only up to the one-variable convex closure.

lemma helperForLemma33_0_14_sectionwise_kernelRep_eq_convexFunctionClosure_of_rockafellar {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hRock : IsRockafellarConvexBifunction F) : ∀ u x, sSup (Set.range (fun xStar : Fin n → ℝ => (((dotProduct x xStar : ℝ) : EReal) - convexBifunctionPairing F u xStar))) = convexFunctionClosure (F u) x

Lean source

Helper for Lemma33.0.14: concavity of the convex pairing in the parameter variable turns the weighted endpoint kernels into a single affine minorant at the parameter convex combination.

lemma helperForLemma33_0_14_weightedEndpointKernel_le_pairing_of_concaveParameterPairing {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hPair : HasConcaveParameterConvexPairing F) {u₁ u₂ : Fin m → ℝ} {x₁ x₂ xStar : Fin n → ℝ} {a b : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) : (a : EReal) * ((((dotProduct x₁ xStar : ℝ) : EReal) - F u₁ x₁)) + (b : EReal) * ((((dotProduct x₂ xStar : ℝ) : EReal) - F u₂ x₂)) ≤ convexBifunctionPairing F (a • u₁ + b • u₂) xStar

Lean source

Helper for Lemma33.0.14: multiplying a non-⊥ extended-real value by a nonnegative real weight cannot create ⊥.

lemma helperForLemma33_0_14_nonneg_mul_ne_bot_of_ne_bot {a : ℝ} {x : EReal} (ha : 0 ≤ a) (hx : x ≠ ⊥) : (a : EReal) * x ≠ ⊥

Lean source

Helper for Lemma33.0.14: adding two extended-real values that both avoid ⊥ still avoids ⊥.

lemma helperForLemma33_0_14_add_ne_bot_of_ne_bot {X Y : EReal} (hX : X ≠ ⊥) (hY : Y ≠ ⊥) : X + Y ≠ ⊥

Lean source

Helper for Lemma33.0.14: once the two subtrahends both avoid ⊥, the two endpoint affine terms combine exactly into one affine kernel difference.

lemma helperForLemma33_0_14_add_sub_eq_sub_add_of_ne_bot {p q : ℝ} {C D : EReal} (hC : C ≠ ⊥) (hD : D ≠ ⊥) : (((p : ℝ) : EReal) - C) + (((q : ℝ) : EReal) - D) = (((p + q : ℝ) : EReal) - (C + D))

Lean source

Helper for Lemma33.0.14: the weighted endpoint-kernel minorant is equivalent to the desired kernel upper bound after separating the finite affine part from the extended-real section values, provided the endpoint section values avoid ⊥.

lemma helperForLemma33_0_14_kernelUpper_of_endpointMinorant {a b p q : ℝ} {C D P : EReal} (hC : C ≠ ⊥) (hD : D ≠ ⊥) (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) (hMinorant : (a : EReal) * ((((p : ℝ) : EReal) - C) : EReal) + (b : EReal) * ((((q : ℝ) : EReal) - D) : EReal) ≤ P) : (((a * p + b * q : ℝ) : EReal) - P) ≤ (a : EReal) * C + (b : EReal) * D

Lean source

Helper for Lemma33.0.14: once each section is represented by the supremum of its affine kernels and those kernels satisfy the Jensen upper bound, the graph function is jointly convex.

lemma helperForLemma33_0_14_graphConvex_of_sectionwiseKernelUpperBound {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hKernelRep : ∀ u x, sSup (Set.range (fun xStar : Fin n → ℝ => (((dotProduct x xStar : ℝ) : EReal) - convexBifunctionPairing F u xStar))) = F u x) (hKernelUpper : ∀ {u₁ u₂ : Fin m → ℝ} {x₁ x₂ xStar : Fin n → ℝ} {a b : ℝ}, 0 ≤ a → 0 ≤ b → a + b = 1 → (((dotProduct (a • x₁ + b • x₂) xStar : ℝ) : EReal) - convexBifunctionPairing F (a • u₁ + b • u₂) xStar) ≤ (a : EReal) * F u₁ x₁ + (b : EReal) * F u₂ x₂) : IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F)

Lean source

Helper for Lemma33.0.14: Rockafellar convexity provides the sectionwise closure formula for the affine kernels of convexBifunctionPairing F u, together with the pointwise Jensen upper bound those kernels satisfy.

lemma helperForLemma33_0_14_rockafellar_graphKernelData_of_convexBifunction {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hRock : IsRockafellarConvexBifunction F) : (∀ u x, sSup (Set.range (fun xStar : Fin n → ℝ => (((dotProduct x xStar : ℝ) : EReal) - convexBifunctionPairing F u xStar))) = convexFunctionClosure (F u) x) ∧ (∀ {u₁ u₂ : Fin m → ℝ} {x₁ x₂ xStar : Fin n → ℝ} {a b : ℝ}, 0 ≤ a → 0 ≤ b → a + b = 1 → F u₁ x₁ ≠ ⊥ → F u₂ x₂ ≠ ⊥ → (((dotProduct (a • x₁ + b • x₂) xStar : ℝ) : EReal) - convexBifunctionPairing F (a • u₁ + b • u₂) xStar) ≤ (a : EReal) * F u₁ x₁ + (b : EReal) * F u₂ x₂)

Lean source

Helper for Lemma33.0.14: every affine kernel appearing in the sectionwise closure formula is bounded above by that closure value.

lemma helperForLemma33_0_14_endpointKernel_le_sectionwiseClosure {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hClosureRep : ∀ u x, sSup (Set.range (fun xStar : Fin n → ℝ => (((dotProduct x xStar : ℝ) : EReal) - convexBifunctionPairing F u xStar))) = convexFunctionClosure (F u) x) (u : Fin m → ℝ) (x xStar : Fin n → ℝ) : (((dotProduct x xStar : ℝ) : EReal) - convexBifunctionPairing F u xStar) ≤ convexFunctionClosure (F u) x

§33 — Saddle-Functions — Part 5

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Helper for Lemma33.0.14: every convex function agrees with its convex closure on the relative interior of its effective domain, regardless of whether the function is proper or improper.

lemma helperForLemma33_0_14_convexFunctionClosure_eq_on_ri_effectiveDomain {n : ℕ} {f : (Fin n → ℝ) → EReal} (hConv : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) f) : ∀ x ∈ euclideanRelativeInterior n ((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → ℝ)) ⁻¹' effectiveDomain (Set.univ : Set (Fin n → ℝ)) f), convexFunctionClosure f x = f x

Lean source

Helper for Lemma33.0.14: the same relative-interior closure identity, packaged directly for a global convex function on ℝ^n.

lemma helperForLemma33_0_14_convexFunctionClosure_eq_on_ri_effectiveDomain_of_ConvexFunction {n : ℕ} {f : (Fin n → ℝ) → EReal} (hConvFun : ConvexFunction f) : ∀ x ∈ euclideanRelativeInterior n ((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → ℝ)) ⁻¹' effectiveDomain (Set.univ : Set (Fin n → ℝ)) f), convexFunctionClosure f x = f x

Lean source

Helper for Lemma33.0.14: specializing the relative-interior closure identity to a fixed section F u of a Rockafellar-convex bifunction.

lemma helperForLemma33_0_14_sectionwiseClosure_eq_raw_on_ri_effectiveDomain {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSectionConv : IsRockafellarSectionwiseConvexBifunction F) (u : Fin m → ℝ) : ∀ x ∈ euclideanRelativeInterior n ((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → ℝ)) ⁻¹' effectiveDomain (Set.univ : Set (Fin n → ℝ)) (F u)), convexFunctionClosure (F u) x = F u x

Lean source

Helper for Lemma33.0.14: before upgrading from the closure to the raw section value, Rockafellar convexity already yields the mixed-point Jensen bound for the convex closure of the mixed section.

lemma helperForLemma33_0_14_mixedSectionClosure_le_weightedEndpointValues_of_rockafellar {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u₁ u₂ : Fin m → ℝ} {x₁ x₂ : Fin n → ℝ} {a b : ℝ} (hRock : IsRockafellarConvexBifunction F) (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) (hNoBot₁ : F u₁ x₁ ≠ ⊥) (hNoBot₂ : F u₂ x₂ ≠ ⊥) : convexFunctionClosure (F (a • u₁ + b • u₂)) (a • x₁ + b • x₂) ≤ (a : EReal) * F u₁ x₁ + (b : EReal) * F u₂ x₂

Lean source

Helper for Lemma33.0.14: with strictly positive weights, a weighted endpoint-value sum can equal ⊥ only because one of the endpoint values is already ⊥.

lemma helperForLemma33_0_14_positiveWeightedSectionValueSum_eq_bot_iff_endpointBot {a b : ℝ} {C D : EReal} (ha : 0 < a) (hb : 0 < b) : (a : EReal) * C + (b : EReal) * D = ⊥ ↔ C = ⊥ ∨ D = ⊥

Lean source

Helper for Lemma33.0.14: if Rockafellar convexity is supplemented by exact sectionwise closure identities, then the closure-level kernel representation already proved in this file upgrades to the exact raw section values.

lemma helperForLemma33_0_14_exactSectionwiseKernelRep_of_rockafellarConvex {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hRock : IsRockafellarConvexBifunction F) (hClosureExact : ∀ u x, convexFunctionClosure (F u) x = F u x) : ∀ u x, sSup (Set.range (fun xStar : Fin n → ℝ => (((dotProduct x xStar : ℝ) : EReal) - convexBifunctionPairing F u xStar))) = F u x

Lean source

Helper for Lemma33.0.14: once Rockafellar convexity is supplemented by exact sectionwise closure identities and no section value equals ⊥, the kernel argument already present in this file yields joint convexity of the raw graph function.

lemma helperForLemma33_0_14_graphConvex_of_rockafellarConvex_and_exactSectionwiseClosure {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hRock : IsRockafellarConvexBifunction F) (hClosureExact : ∀ u x, convexFunctionClosure (F u) x = F u x) (hNoBot : ∀ u x, F u x ≠ ⊥) : IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F)

Lean source

Helper for Lemma33.0.14: exact sectionwise closure together with no-⊥ section values already proves the entire forward half of the graph-function correspondence.

lemma helperForLemma33_0_14_forwardHalf_of_exactSectionwiseClosure {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hRock : IsRockafellarConvexBifunction F) (hClosureExact : ∀ u x, convexFunctionClosure (F u) x = F u x) (hNoBot : ∀ u x, F u x ≠ ⊥) : IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F) ∧ bifunctionOfGraphFunction (graphFunctionOfBifunction F) = F

Lean source

Helper for Lemma33.0.14: the strengthened exact-closure hypotheses already recover the raw graph-convexity conclusion by extracting the first component of the packaged forward-half statement.

lemma helperForLemma33_0_14_graphConvex_of_rockafellar_with_exactSectionwiseClosure {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hRock : IsRockafellarConvexBifunction F) (hClosureExact : ∀ u x, convexFunctionClosure (F u) x = F u x) (hNoBot : ∀ u x, F u x ≠ ⊥) : IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F)

Lean source

Helper for Lemma33.0.14: a jointly convex graph function already yields the full backward half of the graph-function correspondence after currying.

lemma helperForLemma33_0_14_backwardHalf_of_graphConvex {m n : ℕ} {f : (Fin (m + n) → ℝ) → EReal} (hf : IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) f) : IsRockafellarConvexBifunction (bifunctionOfGraphFunction f) ∧ graphFunctionOfBifunction (bifunctionOfGraphFunction f) = f

Lean source

Helper for Lemma33.0.14: if two graph points have value 0 but their midpoint has value ⊤, then the graph function cannot satisfy Jensen's inequality.

lemma helperForLemma33_0_14_graphNotConvex_of_midpointTop {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u₁ u₂ : Fin m → ℝ} {x : Fin n → ℝ} (hLeft : F u₁ x = 0) (hRight : F u₂ x = 0) (hMid : F ((1 / 2 : ℝ) • u₁ + (1 / 2 : ℝ) • u₂) x = ⊤) : ¬ IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F)

Lean source

Helper for Lemma33.0.14: Agent C's one-dimensional boundary-value counterexample keeps the closed interval [0, 1] away from the parameter u = 0, but removes the boundary point 0 exactly at u = 0.

noncomputable def helperForLemma33_0_14_boundaryCounterexample : (Fin 1 → ℝ) → (Fin 1 → ℝ) → EReal := fun u x => if u 0 = 0 then if 0 < x 0 ∧ x 0 ≤ 1 then 0 else ⊤ else if 0 ≤ x 0 ∧ x 0 ≤ 1 then 0 else ⊤

Lean source

Helper for Lemma33.0.14: the left endpoint section in the boundary-value counterexample contains the boundary point 0.

lemma helperForLemma33_0_14_boundaryCounterexample_leftValue : helperForLemma33_0_14_boundaryCounterexample (fun _ : Fin 1 => (-1 : ℝ)) (fun _ : Fin 1 => (0 : ℝ)) = 0

Lean source

Helper for Lemma33.0.14: the right endpoint section in the boundary-value counterexample also contains the boundary point 0.

lemma helperForLemma33_0_14_boundaryCounterexample_rightValue : helperForLemma33_0_14_boundaryCounterexample (fun _ : Fin 1 => (1 : ℝ)) (fun _ : Fin 1 => (0 : ℝ)) = 0

Lean source

Helper for Lemma33.0.14: at the midpoint parameter u = 0, the boundary point 0 is deleted from the section and the section value jumps to ⊤.

lemma helperForLemma33_0_14_boundaryCounterexample_midpointValue : helperForLemma33_0_14_boundaryCounterexample (fun _ : Fin 1 => (0 : ℝ)) (fun _ : Fin 1 => (0 : ℝ)) = ⊤

Lean source

Helper for Lemma33.0.14: the exceptional section at u = 0 is the half-open interval (0, 1] viewed inside Fin 1 → ℝ.

def helperForLemma33_0_14_boundaryCounterexample_openSection : Set (Fin 1 → ℝ) := {x | 0 < x 0 ∧ x 0 ≤ 1}

Lean source

Helper for Lemma33.0.14: every nonzero parameter section in the counterexample is the closed interval [0, 1] viewed inside Fin 1 → ℝ.

def helperForLemma33_0_14_boundaryCounterexample_closedSection : Set (Fin 1 → ℝ) := {x | 0 ≤ x 0 ∧ x 0 ≤ 1}

Lean source

Helper for Lemma33.0.14: the half-open section (0, 1] is nonempty.

lemma helperForLemma33_0_14_boundaryCounterexample_openSection_nonempty : (helperForLemma33_0_14_boundaryCounterexample_openSection).Nonempty

Lean source

Helper for Lemma33.0.14: the closed section [0, 1] is nonempty.

lemma helperForLemma33_0_14_boundaryCounterexample_closedSection_nonempty : (helperForLemma33_0_14_boundaryCounterexample_closedSection).Nonempty

Lean source

Helper for Lemma33.0.14: the exceptional half-open section (0, 1] is convex.

lemma helperForLemma33_0_14_boundaryCounterexample_openSection_convex : Convex ℝ helperForLemma33_0_14_boundaryCounterexample_openSection

Lean source

Helper for Lemma33.0.14: the closed section [0, 1] is convex.

lemma helperForLemma33_0_14_boundaryCounterexample_closedSection_convex : Convex ℝ helperForLemma33_0_14_boundaryCounterexample_closedSection

Lean source

Helper for Lemma33.0.14: at the exceptional parameter u = 0, the section is exactly the indicator of the half-open interval (0, 1].

lemma helperForLemma33_0_14_boundaryCounterexample_section_eq_indicator_open (u : Fin 1 → ℝ) (hu : u 0 = 0) : helperForLemma33_0_14_boundaryCounterexample u = indicatorFunction helperForLemma33_0_14_boundaryCounterexample_openSection

Lean source

Helper for Lemma33.0.14: away from the exceptional parameter u = 0, each section is exactly the indicator of the closed interval [0, 1].

lemma helperForLemma33_0_14_boundaryCounterexample_section_eq_indicator_closed (u : Fin 1 → ℝ) (hu : u 0 ≠ 0) : helperForLemma33_0_14_boundaryCounterexample u = indicatorFunction helperForLemma33_0_14_boundaryCounterexample_closedSection

Lean source

Helper for Lemma33.0.14: every section of the counterexample is convex in the section variable.

lemma helperForLemma33_0_14_boundaryCounterexample_sectionwiseConvex : IsRockafellarSectionwiseConvexBifunction helperForLemma33_0_14_boundaryCounterexample

Lean source

Helper for Lemma33.0.14: although the exceptional section omits the boundary point 0, that point still lies in the closure of (0, 1].

lemma helperForLemma33_0_14_boundaryCounterexample_zero_mem_closure_openSection : (0 : Fin 1 → ℝ) ∈ closure helperForLemma33_0_14_boundaryCounterexample_openSection

Lean source

Helper for Lemma33.0.14: at the exceptional parameter u = 0, the convex closure of the deleted-boundary section restores the missing endpoint and therefore takes the value 0 there.

lemma helperForLemma33_0_14_boundaryCounterexample_midpointClosureValue : convexFunctionClosure (helperForLemma33_0_14_boundaryCounterexample (fun _ : Fin 1 => (0 : ℝ))) (fun _ : Fin 1 => (0 : ℝ)) = 0

Lean source

Helper for Lemma33.0.14: the boundary-value counterexample fails the exact sectionwise-closure hypothesis precisely at the deleted midpoint boundary point.

lemma helperForLemma33_0_14_boundaryCounterexample_exactSectionwiseClosureFails : ¬ ∀ u x, convexFunctionClosure (helperForLemma33_0_14_boundaryCounterexample u) x = helperForLemma33_0_14_boundaryCounterexample u x

Lean source

Helper for Lemma33.0.14: the boundary-value counterexample never takes the value ⊥, so the failure of the repaired forward route comes from missing exact closure rather than from a bottom-value pathology.

lemma helperForLemma33_0_14_boundaryCounterexample_ne_bot : ∀ u x, helperForLemma33_0_14_boundaryCounterexample u x ≠ ⊥

Lean source

Helper for Lemma33.0.14: the support function of (0, 1] agrees with that of [0, 1]; the missing boundary point contributes only the vector 0, whose support value is already captured in the closure.

lemma helperForLemma33_0_14_boundaryCounterexample_supportFunction_open_eq_closed : supportFunctionEReal helperForLemma33_0_14_boundaryCounterexample_openSection = supportFunctionEReal helperForLemma33_0_14_boundaryCounterexample_closedSection

Lean source

Helper for Lemma33.0.14: the support function of the closed interval [0, 1] is the finite scalar function x^* ↦ max (0, x^*) in one dimension.

lemma helperForLemma33_0_14_boundaryCounterexample_closedSupport_eq_max (xStar : Fin 1 → ℝ) : supportFunctionEReal helperForLemma33_0_14_boundaryCounterexample_closedSection xStar = ((max 0 (xStar 0) : ℝ) : EReal)

Lean source

Helper for Lemma33.0.14: the convex pairing of the boundary-value counterexample is the support function of the closed interval [0, 1], hence it does not depend on the parameter u at all.

lemma helperForLemma33_0_14_boundaryCounterexample_pairing_eq_closedSupport (u xStar : Fin 1 → ℝ) : convexBifunctionPairing helperForLemma33_0_14_boundaryCounterexample u xStar = supportFunctionEReal helperForLemma33_0_14_boundaryCounterexample_closedSection xStar

Lean source

Helper for Lemma33.0.14: because the convex pairing of the boundary-value counterexample is independent of u, it is automatically concave in the parameter variable.

lemma helperForLemma33_0_14_boundaryCounterexample_concaveParameterPairing : HasConcaveParameterConvexPairing helperForLemma33_0_14_boundaryCounterexample

Lean source

Helper for Lemma33.0.14: the one-dimensional boundary-value witness is genuinely Rockafellar-convex even though its raw graph function is not convex.

lemma helperForLemma33_0_14_boundaryCounterexample_isRockafellarConvex : IsRockafellarConvexBifunction helperForLemma33_0_14_boundaryCounterexample

Lean source

Helper for Lemma33.0.14: the boundary-value witness already satisfies Rockafellar convexity and the tempting global no-⊥ repair, while still failing the exact sectionwise-closure identity that the working repaired theorem actually needs.

lemma helperForLemma33_0_14_boundaryCounterexample_has_noBot_but_lacks_exactSectionwiseClosure : IsRockafellarConvexBifunction helperForLemma33_0_14_boundaryCounterexample ∧ (∀ u x, helperForLemma33_0_14_boundaryCounterexample u x ≠ ⊥) ∧ ¬ ∀ u x, convexFunctionClosure (helperForLemma33_0_14_boundaryCounterexample u) x = helperForLemma33_0_14_boundaryCounterexample u x

Lean source

Helper for Lemma33.0.14: for the explicit boundary-value witness itself, the tempting global no-⊥ repair still does not force the exact sectionwise-closure identity needed by the working forward theorem.

lemma helperForLemma33_0_14_boundaryCounterexample_noBot_doesNotForce_exactSectionwiseClosure : ¬ ( (∀ u x, helperForLemma33_0_14_boundaryCounterexample u x ≠ ⊥) → ∀ u x, convexFunctionClosure (helperForLemma33_0_14_boundaryCounterexample u) x = helperForLemma33_0_14_boundaryCounterexample u x)

Lean source

Helper for Lemma33.0.14: even after excluding ⊥-valued sections, Rockafellar convexity still does not force the exact sectionwise-closure hypothesis needed by the repaired forward argument.

lemma helperForLemma33_0_14_rockafellarConvex_and_noBot_doNotForce_exactSectionwiseClosure : ¬ ( ∀ {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal}, IsRockafellarConvexBifunction F → (∀ u x, F u x ≠ ⊥) → ∀ u x, convexFunctionClosure (F u) x = F u x)

Lean source

Helper for Lemma33.0.14: the raw graph of the boundary-value counterexample is not convex, because two finite endpoint values average to a midpoint value ⊤.

lemma helperForLemma33_0_14_boundaryCounterexample_graph_not_convex : ¬ IsERealConvexOn (Set.univ : Set (Fin (1 + 1) → ℝ)) (graphFunctionOfBifunction helperForLemma33_0_14_boundaryCounterexample)

Lean source

Helper for Lemma33.0.14: even for the explicit boundary-value witness, the tempting global no-⊥ repair does not force convexity of the raw graph function.

lemma helperForLemma33_0_14_boundaryCounterexample_noBot_doesNotForce_graphConvexity : ¬ ( (∀ u x, helperForLemma33_0_14_boundaryCounterexample u x ≠ ⊥) → IsERealConvexOn (Set.univ : Set (Fin (1 + 1) → ℝ)) (graphFunctionOfBifunction helperForLemma33_0_14_boundaryCounterexample))

Lean source

Helper for Lemma33.0.14: adding only a global no-⊥ hypothesis still does not rescue the raw forward bridge in dimension (1, 1), because the explicit boundary-value witness already satisfies that extra assumption while its raw graph remains nonconvex.

lemma helperForLemma33_0_14_forwardImplication_with_noBot_stillFails_in_dim1 : ¬ (∀ F : (Fin 1 → ℝ) → (Fin 1 → ℝ) → EReal, IsRockafellarConvexBifunction F → (∀ u x, F u x ≠ ⊥) → IsERealConvexOn (Set.univ : Set (Fin (1 + 1) → ℝ)) (graphFunctionOfBifunction F))

Lean source

Helper for Lemma33.0.14: for the explicit boundary-value witness, the exact forward conclusion appearing in the theorem's forward branch is already impossible, because the graph-convexity component fails even though the reconstruction identity is available.

lemma helperForLemma33_0_14_boundaryCounterexample_refutes_forwardConclusion : ¬ ( IsERealConvexOn (Set.univ : Set (Fin (1 + 1) → ℝ)) (graphFunctionOfBifunction helperForLemma33_0_14_boundaryCounterexample) ∧ bifunctionOfGraphFunction (graphFunctionOfBifunction helperForLemma33_0_14_boundaryCounterexample) = helperForLemma33_0_14_boundaryCounterexample)

Lean source

Helper for Lemma33.0.14: even the exact forward-branch shape from the correspondence statement remains false after adding only a global no-⊥ hypothesis in dimension (1, 1).

lemma helperForLemma33_0_14_forwardCorrespondence_with_noBot_stillFails_in_dim1 : ¬ (∀ F : (Fin 1 → ℝ) → (Fin 1 → ℝ) → EReal, IsRockafellarConvexBifunction F → (∀ u x, F u x ≠ ⊥) → IsERealConvexOn (Set.univ : Set (Fin (1 + 1) → ℝ)) (graphFunctionOfBifunction F) ∧ bifunctionOfGraphFunction (graphFunctionOfBifunction F) = F)

Lean source

Helper for Lemma33.0.14: the explicit one-dimensional boundary-value witness refutes the specialized forward branch in its exact theorem shape, not merely its graph-convexity projection.

lemma helperForLemma33_0_14_boundaryCounterexample_refutes_exactForwardBridge : ¬ ( IsRockafellarConvexBifunction helperForLemma33_0_14_boundaryCounterexample → IsERealConvexOn (Set.univ : Set (Fin (1 + 1) → ℝ)) (graphFunctionOfBifunction helperForLemma33_0_14_boundaryCounterexample) ∧ bifunctionOfGraphFunction (graphFunctionOfBifunction helperForLemma33_0_14_boundaryCounterexample) = helperForLemma33_0_14_boundaryCounterexample)

Lean source

Helper for Lemma33.0.14: the explicit one-dimensional boundary-value witness already refutes the specialized raw forward bridge from Rockafellar convexity to graph convexity.

lemma helperForLemma33_0_14_boundaryCounterexample_refutes_rawForwardBridge : ¬ ( IsRockafellarConvexBifunction helperForLemma33_0_14_boundaryCounterexample → IsERealConvexOn (Set.univ : Set (Fin (1 + 1) → ℝ)) (graphFunctionOfBifunction helperForLemma33_0_14_boundaryCounterexample))

Lean source

Helper for Lemma33.0.14: the forward implication from Rockafellar convexity to raw graph convexity already fails in dimension (m, n) = (1, 1).

lemma helperForLemma33_0_14_forwardImplicationFails_in_dim1 : ¬ (∀ F : (Fin 1 → ℝ) → (Fin 1 → ℝ) → EReal, IsRockafellarConvexBifunction F → IsERealConvexOn (Set.univ : Set (Fin (1 + 1) → ℝ)) (graphFunctionOfBifunction F))

Lean source

Helper for Lemma33.0.14: even the forward branch of the claimed graph-function correspondence, including the reconstruction clause, already fails in dimension (1, 1) because its raw graph-convexity component is false there.

lemma helperForLemma33_0_14_forwardCorrespondenceFails_in_dim1 : ¬ (∀ F : (Fin 1 → ℝ) → (Fin 1 → ℝ) → EReal, IsRockafellarConvexBifunction F → IsERealConvexOn (Set.univ : Set (Fin (1 + 1) → ℝ)) (graphFunctionOfBifunction F) ∧ bifunctionOfGraphFunction (graphFunctionOfBifunction F) = F)

Lean source

Helper for Lemma33.0.14: the specialized (1, 1) correspondence statement would force the explicit boundary-value counterexample to have a convex raw graph.

lemma helperForLemma33_0_14_dim1Correspondence_forces_boundaryCounterexample_graphConvex (hCorrespondence : (∀ F : (Fin 1 → ℝ) → (Fin 1 → ℝ) → EReal, IsRockafellarConvexBifunction F → IsERealConvexOn (Set.univ : Set (Fin (1 + 1) → ℝ)) (graphFunctionOfBifunction F) ∧ bifunctionOfGraphFunction (graphFunctionOfBifunction F) = F) ∧ (∀ f : (Fin (1 + 1) → ℝ) → EReal, IsERealConvexOn (Set.univ : Set (Fin (1 + 1) → ℝ)) f → IsRockafellarConvexBifunction (bifunctionOfGraphFunction f) ∧ graphFunctionOfBifunction (bifunctionOfGraphFunction f) = f)) : IsERealConvexOn (Set.univ : Set (Fin (1 + 1) → ℝ)) (graphFunctionOfBifunction helperForLemma33_0_14_boundaryCounterexample)

§33 — Saddle-Functions — Part 6

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Lemma33.0.14: in the present EReal formalization, the graph-function correspondence is carried by convex bifunctions whose sections already agree with their convex closures and never take the value ⊥. Under these exact sectionwise-closure hypotheses, the graph functions f (u, x) = (F u) x are exactly the convex functions on ℝ^(m + n), and the two constructions are inverse.

lemma convexBifunction_graphFunction_correspondence {m n : ℕ} : (∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConvexBifunction F → (∀ u x, convexFunctionClosure (F u) x = F u x) → (∀ u x, F u x ≠ ⊥) → IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F) ∧ bifunctionOfGraphFunction (graphFunctionOfBifunction F) = F) ∧ (∀ f : (Fin (m + n) → ℝ) → EReal, IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) f → IsRockafellarConvexBifunction (bifunctionOfGraphFunction f) ∧ graphFunctionOfBifunction (bifunctionOfGraphFunction f) = f)

Lean source

Definition33.0.15: The ordinary conjugate of f is the function f*(u*, x*) = sup_{u, x} (⟪u, u*⟫ + ⟪x, x*⟫ - f(u, x)).

noncomputable def ordinaryConjugate {m n : ℕ} (f : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun uStar xStar => sSup <| Set.range fun ux : (Fin m → ℝ) × (Fin n → ℝ) => (((dotProduct ux.1 uStar + dotProduct ux.2 xStar : ℝ) : EReal) - f ux.1 ux.2)

Lean source

The convex closure of an EReal-valued function on ℝ^n, defined by local infima over open balls.

noncomputable def functionConvexClosure {n : ℕ} (f : (Fin n → ℝ) → EReal) : (Fin n → ℝ) → EReal := fun x => ⨆ (ε : {ε : ℝ // 0 < ε}), ⨅ (w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}), f w.1

Lean source

An EReal-valued function on ℝ^n is convex-closed when it agrees with its convex closure.

def IsFunctionConvexClosed {n : ℕ} (f : (Fin n → ℝ) → EReal) : Prop := f = functionConvexClosure f

Lean source

The concave closure of an EReal-valued function on ℝ^n, defined by local suprema over open balls followed by infimum over radii.

noncomputable def functionConcaveClosure {n : ℕ} (f : (Fin n → ℝ) → EReal) : (Fin n → ℝ) → EReal := fun x => ⨅ (ε : {ε : ℝ // 0 < ε}), ⨆ (w : {w : Fin n → ℝ // ‖w - x‖ < ε.1}), f w.1

Lean source

An EReal-valued function on ℝ^n is concave-closed when it agrees with its concave closure.

def IsFunctionConcaveClosed {n : ℕ} (f : (Fin n → ℝ) → EReal) : Prop := f = functionConcaveClosure f

Lean source

An EReal-valued bifunction models a ℝ ∪ {+∞}-valued one when it never takes ⊥.

def HasNoBotValuesBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := ∀ u x, F u x ≠ ⊥

Lean source

An EReal-valued bifunction models a ℝ ∪ {-∞}-valued one when it never takes ⊤.

def HasNoTopValuesBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := ∀ u x, F u x ≠ ⊤

Lean source

A bifunction satisfies both one-sided extended-real conventions when it takes neither ⊥ nor ⊤ values.

def HasNoTopOrBotValuesBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := HasNoBotValuesBifunction F ∧ HasNoTopValuesBifunction F

Lean source

Theorem33.1: If F : ℝ^m → (ℝ^n → EReal) is a convex bifunction in Rockafellar's sense and takes no value ⊥, then the pairing (u, x^*) ↦ ⟪F u, x^*⟫ is concave-convex and convex-closed in the second variable, and for every u and x one has (cl (F u)) x = sup_{x^*} (⟪x, x^*⟫ - ⟪F u, x^*⟫). Conversely, if K : ℝ^m × ℝ^n → EReal is concave-convex, takes neither value ⊥ nor ⊤, and (F u) x = sup_{x^*} (⟪x, x^*⟫ - K (u, x^*)), then F is a convex bifunction, each section F u is convex-closed, and ⟪F u, x^*⟫ = (cl₂ K) (u, x^*).

theorem convexBifunction_pairing_correspondence {m n : ℕ} : (∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConvexBifunction F → HasNoBotValuesBifunction F → IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (convexBifunctionPairing F) ∧ IsConvexClosedInSecond (convexBifunctionPairing F) ∧ ∀ u x, functionConvexClosure (F u) x = convexConjugate (convexBifunctionPairing F u) x) ∧ (∀ K : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K → HasNoTopOrBotValuesBifunction K → let F : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u x => convexConjugate (K u) x IsRockafellarConvexBifunction F ∧ (∀ u, IsFunctionConvexClosed (F u)) ∧ ∀ u xStar, convexBifunctionPairing F u xStar = convexClosureInSecond K u xStar)

Lean source

Theorem33.0.16: Under the admissibility hypotheses used in Theorem33.1, a concave-convex bifunction that is convex-closed in the second variable is exactly a partial convex conjugate of a Rockafellar convex bifunction.

theorem convexClosed_concaveConvex_bifunctions_are_partial_conjugates {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} : ((∃ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ K = convexBifunctionPairing F) → IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ IsConvexClosedInSecond K) ∧ (IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K → IsConvexClosedInSecond K → HasNoTopOrBotValuesBifunction K → ∃ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConvexBifunction F ∧ (∀ u, IsFunctionConvexClosed (F u)) ∧ K = convexBifunctionPairing F)

Lean source

A convex second-variable section which never takes ⊥ keeps that convention after taking its local convex closure. This is the qualification needed when a subsequent first-variable concave closure is required in the strong Jensen sense.

lemma helperForCorollary33_1_1_convexClosureInSecond_noBot {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hConv : ∀ u, IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (K u)) (hNoBot : HasNoBotValuesBifunction K) : HasNoBotValuesBifunction (convexClosureInSecond K)

Lean source

The first-variable concave closure lies pointwise above the original kernel, so it also preserves the convention that ⊥ never occurs.

lemma helperForCorollary33_1_1_concaveClosureInFirst_noBot {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hNoBot : HasNoBotValuesBifunction K) : HasNoBotValuesBifunction (concaveClosureInFirst K)

Lean source

Strong Jensen form of the coordinatewise closure package under the one-sided convention that the original kernel never takes ⊥.

lemma helperForCorollary33_1_1_concaveConvex_coordinatewise_closures_of_noBot {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K) (hNoBot : ∀ u v, K u v ≠ (⊥ : EReal)) : IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (concaveClosureInFirst K) ∧ IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (convexClosureInSecond K) ∧ IsConcaveClosedInFirst (concaveClosureInFirst K) ∧ IsConvexClosedInSecond (convexClosureInSecond K)

Lean source

Helper for Lemma33.0.14: the full one-to-one correspondence statement is already false in dimension (1, 1) because its forward branch fails there.

lemma helperForLemma33_0_14_correspondenceStatementFails_in_dim1 : ¬ ( (∀ F : (Fin 1 → ℝ) → (Fin 1 → ℝ) → EReal, IsRockafellarConvexBifunction F → IsERealConvexOn (Set.univ : Set (Fin (1 + 1) → ℝ)) (graphFunctionOfBifunction F) ∧ bifunctionOfGraphFunction (graphFunctionOfBifunction F) = F) ∧ (∀ f : (Fin (1 + 1) → ℝ) → EReal, IsERealConvexOn (Set.univ : Set (Fin (1 + 1) → ℝ)) f → IsRockafellarConvexBifunction (bifunctionOfGraphFunction f) ∧ graphFunctionOfBifunction (bifunctionOfGraphFunction f) = f))

Lean source

Helper for Lemma33.0.14: any dimension-uniform proof of the claimed graph-function correspondence would specialize to the already-refuted (1, 1) statement, so the obstruction is semantic rather than a missing local lemma.

lemma helperForLemma33_0_14_genericCorrespondence_specializes_to_dim1_contradiction (hCorrespondence : ∀ {m n : ℕ}, (∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConvexBifunction F → IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F) ∧ bifunctionOfGraphFunction (graphFunctionOfBifunction F) = F) ∧ (∀ f : (Fin (m + n) → ℝ) → EReal, IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) f → IsRockafellarConvexBifunction (bifunctionOfGraphFunction f) ∧ graphFunctionOfBifunction (bifunctionOfGraphFunction f) = f)) : False

Lean source

Helper for Lemma33.0.14: any dimension-uniform proof of the raw forward bridge from Rockafellar convexity to graph convexity would already contradict the specialized (1, 1) boundary-value witness.

lemma helperForLemma33_0_14_genericForwardBridge_specializes_to_dim1_contradiction (hForward : ∀ {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal}, IsRockafellarConvexBifunction F → IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F)) : False

Lean source

Helper for Lemma33.0.14: under the current definitions, there is no dimension-uniform raw forward bridge from Rockafellar convexity to convexity of the graph function.

lemma helperForLemma33_0_14_no_dimensionUniformForwardBridge : ¬ ( ∀ {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal}, IsRockafellarConvexBifunction F → IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F))

Lean source

Helper for Lemma33.0.14: any dimension-uniform proof of the full graph-function correspondence automatically contains the raw forward bridge obtained by projecting the first component of the claimed equivalence.

lemma helperForLemma33_0_14_uniformCorrespondence_implies_uniformForwardBridge (hCorrespondence : ∀ {m n : ℕ}, (∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConvexBifunction F → IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F) ∧ bifunctionOfGraphFunction (graphFunctionOfBifunction F) = F) ∧ (∀ f : (Fin (m + n) → ℝ) → EReal, IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) f → IsRockafellarConvexBifunction (bifunctionOfGraphFunction f) ∧ graphFunctionOfBifunction (bifunctionOfGraphFunction f) = f)) : ∀ {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal}, IsRockafellarConvexBifunction F → IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F)

Lean source

Helper for Lemma33.0.14: under the current definitions, there is no dimension-uniform proof of the full claimed graph-function correspondence.

lemma helperForLemma33_0_14_no_dimensionUniformCorrespondence : ¬ ( ∀ {m n : ℕ}, (∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConvexBifunction F → IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F) ∧ bifunctionOfGraphFunction (graphFunctionOfBifunction F) = F) ∧ (∀ f : (Fin (m + n) → ℝ) → EReal, IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) f → IsRockafellarConvexBifunction (bifunctionOfGraphFunction f) ∧ graphFunctionOfBifunction (bifunctionOfGraphFunction f) = f))

Lean source

Helper for Lemma33.0.14: even after strengthening the raw forward bridge by a global no-⊥ hypothesis, the boundary-value witness still rules out any dimension-uniform proof of that repair.

lemma helperForLemma33_0_14_no_dimensionUniformForwardBridge_with_noBot : ¬ ( ∀ {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal}, IsRockafellarConvexBifunction F → (∀ u x, F u x ≠ ⊥) → IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F))

Lean source

Helper for Lemma33.0.14: any dimension-uniform proof of the no-⊥-strengthened full graph-function correspondence still contains the corresponding no-⊥ forward bridge, obtained by projecting the first component of the claimed equivalence.

lemma helperForLemma33_0_14_uniformCorrespondence_with_noBot_implies_uniformForwardBridge_with_noBot (hCorrespondence : ∀ {m n : ℕ}, (∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConvexBifunction F → (∀ u x, F u x ≠ ⊥) → IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F) ∧ bifunctionOfGraphFunction (graphFunctionOfBifunction F) = F) ∧ (∀ f : (Fin (m + n) → ℝ) → EReal, IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) f → IsRockafellarConvexBifunction (bifunctionOfGraphFunction f) ∧ graphFunctionOfBifunction (bifunctionOfGraphFunction f) = f)) : ∀ {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal}, IsRockafellarConvexBifunction F → (∀ u x, F u x ≠ ⊥) → IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F)

Lean source

Helper for Lemma33.0.14: even the no-⊥-strengthened full correspondence cannot hold dimension-uniformly, because it still projects to the already-refuted no-⊥ forward bridge.

lemma helperForLemma33_0_14_no_dimensionUniformCorrespondence_with_noBot : ¬ ( ∀ {m n : ℕ}, (∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConvexBifunction F → (∀ u x, F u x ≠ ⊥) → IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F) ∧ bifunctionOfGraphFunction (graphFunctionOfBifunction F) = F) ∧ (∀ f : (Fin (m + n) → ℝ) → EReal, IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) f → IsRockafellarConvexBifunction (bifunctionOfGraphFunction f) ∧ graphFunctionOfBifunction (bifunctionOfGraphFunction f) = f))

Lean source

Helper for Lemma33.0.14: after repairing the false raw forward bridge by adding exact sectionwise-closure and no-⊥ hypotheses, Rockafellar convexity does imply convexity of the raw graph function.

lemma helperForLemma33_0_14_forwardHalf_graphConvex_of_rockafellar {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hRock : IsRockafellarConvexBifunction F) (hClosureExact : ∀ u x, convexFunctionClosure (F u) x = F u x) (hNoBot : ∀ u x, F u x ≠ ⊥) : IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F)

Lean source

Helper for Lemma33.0.14: after repairing the false raw forward branch by adding exact sectionwise-closure and no-⊥ hypotheses, the intended graph-function correspondence is valid.

lemma helperForLemma33_0_14_repairedCorrespondence_of_exactSectionwiseClosure {m n : ℕ} : (∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConvexBifunction F → (∀ u x, convexFunctionClosure (F u) x = F u x) → (∀ u x, F u x ≠ ⊥) → IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F) ∧ bifunctionOfGraphFunction (graphFunctionOfBifunction F) = F) ∧ (∀ f : (Fin (m + n) → ℝ) → EReal, IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) f → IsRockafellarConvexBifunction (bifunctionOfGraphFunction f) ∧ graphFunctionOfBifunction (bifunctionOfGraphFunction f) = f)

Lean source

Helper for Lemma33.0.14: the exact forward-branch shape appearing in the target theorem already forces the explicit boundary-value counterexample to have a convex raw graph.

lemma helperForLemma33_0_14_targetForwardBranch_forces_boundaryCounterexample_graphConvex (hForward : ∀ {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal}, IsRockafellarConvexBifunction F → IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F) ∧ bifunctionOfGraphFunction (graphFunctionOfBifunction F) = F) : IsERealConvexOn (Set.univ : Set (Fin (1 + 1) → ℝ)) (graphFunctionOfBifunction helperForLemma33_0_14_boundaryCounterexample)

Lean source

Helper for Lemma33.0.14: the exact forward-branch shape from the target theorem is already impossible, because it forces the boundary-value counterexample to satisfy the raw graph-convexity conclusion that the midpoint calculation explicitly refutes.

lemma helperForLemma33_0_14_targetForwardBranch_impossible (hForward : ∀ {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal}, IsRockafellarConvexBifunction F → IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F) ∧ bifunctionOfGraphFunction (graphFunctionOfBifunction F) = F) : False

Lean source

Helper for Lemma33.0.14: the main correspondence lemma is exactly the packaged repaired forward/backward halves already proved above.

lemma helperForLemma33_0_14_correspondence {m n : ℕ} : (∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConvexBifunction F → (∀ u x, convexFunctionClosure (F u) x = F u x) → (∀ u x, F u x ≠ ⊥) → IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F) ∧ bifunctionOfGraphFunction (graphFunctionOfBifunction F) = F) ∧ (∀ f : (Fin (m + n) → ℝ) → EReal, IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) f → IsRockafellarConvexBifunction (bifunctionOfGraphFunction f) ∧ graphFunctionOfBifunction (bifunctionOfGraphFunction f) = f)

Lean source

Helper for Lemma33.0.14: the target theorem can be closed by directly reusing the already-packaged correspondence helper.

lemma helperForLemma33_0_14_correspondence_for_mainTheorem {m n : ℕ} : (∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConvexBifunction F → (∀ u x, convexFunctionClosure (F u) x = F u x) → (∀ u x, F u x ≠ ⊥) → IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F) ∧ bifunctionOfGraphFunction (graphFunctionOfBifunction F) = F) ∧ (∀ f : (Fin (m + n) → ℝ) → EReal, IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) f → IsRockafellarConvexBifunction (bifunctionOfGraphFunction f) ∧ graphFunctionOfBifunction (bifunctionOfGraphFunction f) = f)

Lean source

Helper for Theorem33.1: the local convex-conjugate notation is exactly Chapter 3's Fenchel conjugate.

lemma helperForTheorem33_1_convexConjugate_eq_fenchelConjugate {n : ℕ} (f : (Fin n → ℝ) → EReal) : convexConjugate f = fenchelConjugate n f

Lean source

Helper for Theorem33.1: if one primal point avoids ⊤, then the convex conjugate never collapses to ⊥.

lemma helperForTheorem33_1_convexConjugate_ne_bot_of_point {n : ℕ} {f : (Fin n → ℝ) → EReal} {x₀ : Fin n → ℝ} (hx₀_ne_top : f x₀ ≠ ⊤) : ∀ xStar, convexConjugate f xStar ≠ (⊥ : EReal)

Lean source

Helper for Theorem33.1: the constant ⊥ function is convex in the Jensen sense used in Section 33.

lemma helperForTheorem33_1_constBot_isERealConvexOn {n : ℕ} : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (fun _ : Fin n → ℝ => (⊥ : EReal))

Lean source

Helper for Theorem33.1: once one primal point avoids ⊤, Fenchel closed-convexity turns the convex conjugate into a Section 33 convex function.

lemma helperForTheorem33_1_convexConjugate_isERealConvexOn_of_point {n : ℕ} {f : (Fin n → ℝ) → EReal} {x₀ : Fin n → ℝ} (hx₀_ne_top : f x₀ ≠ ⊤) : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (convexConjugate f)

Lean source

Helper for Theorem33.1: every lower semicontinuous minorant of f lies below the Section 33 closure operator.

lemma helperForTheorem33_1_lowerSemicontinuous_le_functionConvexClosure {n : ℕ} {f h : (Fin n → ℝ) → EReal} (hh : LowerSemicontinuous h) (hle : h ≤ f) : h ≤ functionConvexClosure f

Lean source

Helper for Theorem33.1: a lower semicontinuous section is already fixed by the raw Section 33 closure operator.

lemma helperForTheorem33_1_functionConvexClosure_eq_self_of_lowerSemicontinuous {n : ℕ} {f : (Fin n → ℝ) → EReal} (hf : LowerSemicontinuous f) : f = functionConvexClosure f

Lean source

Helper for Theorem33.1: once one primal point avoids ⊤, the convex conjugate section is already fixed by the one-variable convex closure.

lemma helperForTheorem33_1_convexConjugate_isFunctionConvexClosed_of_point {n : ℕ} {f : (Fin n → ℝ) → EReal} {x₀ : Fin n → ℝ} (hx₀_ne_top : f x₀ ≠ ⊤) : IsFunctionConvexClosed (convexConjugate f)

Lean source

Helper for Theorem33.1: when f never takes the value ⊥, the Section 33 closure agrees with Chapter 2's convex-function closure.

lemma helperForTheorem33_1_functionConvexClosure_eq_convexFunctionClosure_of_noBot {n : ℕ} {f : (Fin n → ℝ) → EReal} (hNoBot : ∀ x, f x ≠ (⊥ : EReal)) : functionConvexClosure f = convexFunctionClosure f

Lean source

Helper for Theorem33.1: sectionwise Fenchel-Moreau identifies the biconjugate with the one-variable convex closure in the local Section 33 notation.

lemma helperForTheorem33_1_biconjugate_eq_functionConvexClosure_of_convex {n : ℕ} {f : (Fin n → ℝ) → EReal} (hConv : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) f) (hNoBot : ∀ x, f x ≠ (⊥ : EReal)) : ∀ x, convexConjugate (convexConjugate f) x = functionConvexClosure f x

Lean source

Helper for Theorem33.1: after defining F u as the sectionwise convex conjugate of K u, the resulting pairing is exactly the second-variable convex closure of K.

lemma helperForTheorem33_1_reverse_pairing_eq_convexClosureInSecond {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hConv : ∀ u, IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (K u)) (hNoBot : ∀ u x, K u x ≠ (⊥ : EReal)) : let F : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u x => convexConjugate (K u) x ∀ u xStar, convexBifunctionPairing F u xStar = convexClosureInSecond K u xStar

Lean source

Helper for Theorem33.0.16: if a pairing agrees pointwise with convexClosureInSecond K, then K equals the pairing whenever K is convex-closed in the second variable.

lemma helperForTheorem33_0_16_eq_pairing_of_pairing_eq_convexClosureInSecond {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hClosed : IsConvexClosedInSecond K) (hPair : ∀ u xStar, convexBifunctionPairing F u xStar = convexClosureInSecond K u xStar) : K = convexBifunctionPairing F

Lean source

Helper for Corollary33.1.1: applying the second-variable convex closure twice does not change the result.

lemma helperForCorollary33_1_1_convexClosureInSecond_idempotent {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (v : Fin n → ℝ) : convexClosureInSecond (convexClosureInSecond K) u v = convexClosureInSecond K u v

Lean source

Helper for Corollary33.1.1: swapping the variables exchanges the two coordinatewise closure operators.

lemma helperForCorollary33_1_1_swap_coordinatewise_closure_identities {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (v : Fin n → ℝ) : concaveClosureInFirst (fun v' u' => K u' v') v u = concaveClosureInSecond K u v ∧ convexClosureInSecond (fun v' u' => K u' v') v u = convexClosureInFirst K u v

Lean source

Helper for Corollary33.1.1: applying the first-variable concave closure twice does not change the result.

lemma helperForCorollary33_1_1_concaveClosureInFirst_idempotent {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (v : Fin n → ℝ) : concaveClosureInFirst (concaveClosureInFirst K) u v = concaveClosureInFirst K u v

Lean source

Helper for Corollary33.1.1: the concave-convex half of the corollary is exactly the closure theorem from Lemma 33.0.5 together with the two fixed-point identities.

lemma helperForCorollary33_1_1_concaveConvex_coordinatewise_closures {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K) : IsEpigraphHypographConcaveConvex (concaveClosureInFirst K) ∧ IsEpigraphHypographConcaveConvex (convexClosureInSecond K) ∧ IsConcaveClosedInFirst (concaveClosureInFirst K) ∧ IsConvexClosedInSecond (convexClosureInSecond K)

§33 — Saddle-Functions — Part 7

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Corollary33.1.1 in the book's epi/hypograph semantics: coordinatewise closures preserve both saddle orientations and are fixed points of the corresponding closure operators.

theorem coordinatewise_closures_preserve_saddle_orientations {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} : (IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K → IsEpigraphHypographConcaveConvex (concaveClosureInFirst K) ∧ IsEpigraphHypographConcaveConvex (convexClosureInSecond K) ∧ IsConcaveClosedInFirst (concaveClosureInFirst K) ∧ IsConvexClosedInSecond (convexClosureInSecond K)) ∧ (IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K → IsEpigraphHypographConvexConcave (convexClosureInFirst K) ∧ IsEpigraphHypographConvexConcave (concaveClosureInSecond K) ∧ IsConvexClosedInFirst (convexClosureInFirst K) ∧ IsConcaveClosedInSecond (concaveClosureInSecond K))

Lean source

Strong Jensen form of Corollary33.1.1 under the explicit convention that K never takes ⊥.

theorem coordinatewise_closures_preserve_saddle_orientations_of_noBot {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hNoBot : ∀ u v, K u v ≠ (⊥ : EReal)) : (IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K → IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (concaveClosureInFirst K) ∧ IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (convexClosureInSecond K) ∧ IsConcaveClosedInFirst (concaveClosureInFirst K) ∧ IsConvexClosedInSecond (convexClosureInSecond K)) ∧ (IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K → IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (convexClosureInFirst K) ∧ IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (concaveClosureInSecond K) ∧ IsConvexClosedInFirst (convexClosureInFirst K) ∧ IsConcaveClosedInSecond (concaveClosureInSecond K))

Lean source

Precomposing with the involution x ↦ -x preserves convexity on the whole space.

lemma helperForCorollary33_1_2_precomp_neg_isERealConvexOn {k : ℕ} {f : (Fin k → ℝ) → EReal} (hConv : IsERealConvexOn (Set.univ : Set (Fin k → ℝ)) f) : IsERealConvexOn (Set.univ : Set (Fin k → ℝ)) (fun x => f (-x))

Lean source

The constant ⊤ function is concave on ℝ^k.

lemma helperForCorollary33_1_2_constTop_isERealConcaveOn {k : ℕ} : IsERealConcaveOn (Set.univ : Set (Fin k → ℝ)) (fun _ => (⊤ : EReal))

Lean source

The Section 33 one-variable concave closure is the negative of the convex closure of the negated function.

lemma helperForCorollary33_1_2_functionConcaveClosure_eq_neg_functionConvexClosure_neg {k : ℕ} {g : (Fin k → ℝ) → EReal} : functionConcaveClosure g = fun x => -functionConvexClosure (fun y => -g y) x

Lean source

The constant ⊤ function is fixed by the one-variable concave closure.

lemma helperForCorollary33_1_2_constTop_isFunctionConcaveClosed {k : ℕ} : IsFunctionConcaveClosed (fun _ : Fin k → ℝ => (⊤ : EReal))

Lean source

Upper semicontinuous sections are already fixed by the one-variable Section 33 concave closure.

lemma helperForCorollary33_1_2_functionConcaveClosure_eq_self_of_upperSemicontinuous {k : ℕ} {g : (Fin k → ℝ) → EReal} (hUpper : UpperSemicontinuous g) : functionConcaveClosure g = g

Lean source

If a section has one point different from ⊥, then its concave conjugate is concave.

lemma helperForCorollary33_1_2_concaveConjugate_isERealConcaveOn_of_point {k : ℕ} {g : (Fin k → ℝ) → EReal} {x₀ : Fin k → ℝ} (hx₀_ne_bot : g x₀ ≠ ⊥) : IsERealConcaveOn (Set.univ : Set (Fin k → ℝ)) (concaveConjugate g)

Lean source

The concave conjugate is concave, with the exceptional all-⊥ branch collapsing to the constant ⊤ function.

lemma helperForCorollary33_1_2_concaveConjugate_isERealConcaveOn {k : ℕ} {g : (Fin k → ℝ) → EReal} : IsERealConcaveOn (Set.univ : Set (Fin k → ℝ)) (concaveConjugate g)

Lean source

If a section has one point different from ⊥, then its concave conjugate is already fixed by the one-variable Section 33 concave closure.

lemma helperForCorollary33_1_2_concaveConjugate_isFunctionConcaveClosed_of_point {k : ℕ} {g : (Fin k → ℝ) → EReal} {x₀ : Fin k → ℝ} (hx₀_ne_bot : g x₀ ≠ ⊥) : IsFunctionConcaveClosed (concaveConjugate g)

Lean source

The concave conjugate is always fixed by the one-variable Section 33 concave closure; the only exceptional branch is the constant ⊤ one when the primal section is identically ⊥.

lemma helperForCorollary33_1_2_concaveConjugate_isFunctionConcaveClosed {k : ℕ} {g : (Fin k → ℝ) → EReal} : IsFunctionConcaveClosed (concaveConjugate g)

Lean source

For a concave section with no ⊤ values, the Section 33 concave biconjugate identity is exactly the one-variable concave closure formula.

lemma helperForCorollary33_1_2_biconjugate_eq_functionConcaveClosure_of_concave {k : ℕ} {g : (Fin k → ℝ) → EReal} (hConc : IsERealConcaveOn (Set.univ : Set (Fin k → ℝ)) g) (hNoTop : ∀ x, g x ≠ ⊤) : ∀ x, concaveConjugate (concaveConjugate g) x = functionConcaveClosure g x

Lean source

The dual saddle-function correspondence for concave bifunctions and convex-concave kernels.

theorem concaveBifunction_pairing_correspondence {m n : ℕ} : (∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConcaveBifunction F → HasNoTopValuesBifunction F → IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (concaveBifunctionPairing F) ∧ IsConcaveClosedInSecond (concaveBifunctionPairing F) ∧ ∀ u x, functionConcaveClosure (F u) x = concaveConjugate (concaveBifunctionPairing F u) x) ∧ (∀ K : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K → HasNoTopOrBotValuesBifunction K → let F : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u x => concaveConjugate (K u) x IsRockafellarConcaveBifunction F ∧ (∀ u, IsFunctionConcaveClosed (F u)) ∧ ∀ u xStar, concaveBifunctionPairing F u xStar = concaveClosureInSecond K u xStar)

Lean source

Definition33.0.17: An image-closed bifunction is a convex bifunction whose every section F u is convex-closed, or a concave bifunction whose every section F u is concave-closed. In particular, a closed bifunction is image-closed.

def IsImageClosedBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := (IsRockafellarSectionwiseConvexBifunction F ∧ ∀ u, IsFunctionConvexClosed (F u)) ∨ (IsConcaveBifunction F ∧ ∀ u, IsFunctionConcaveClosed (F u))

Lean source

An image-closed convex bifunction is a Rockafellar convex bifunction whose sections are convex-closed and which takes values in ℝ ∪ {+∞} (modeled by excluding ⊥).

def IsImageClosedConvexBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := IsRockafellarConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ ∀ u, IsFunctionConvexClosed (F u)

Lean source

A convex-closed concave-convex kernel is a concave-convex bifunction that is convex-closed in the second variable and takes values in ℝ ∪ {+∞} (hence has no ⊥ values).

def IsConvexClosedConcaveConvexKernel {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ IsConvexClosedInSecond K ∧ HasNoBotValuesBifunction K

Lean source

An image-closed concave bifunction is a Rockafellar concave bifunction whose sections are concave-closed and which takes values in ℝ ∪ {-∞} (modeled by excluding ⊤).

def IsImageClosedConcaveBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := IsRockafellarConcaveBifunction F ∧ HasNoTopValuesBifunction F ∧ ∀ u, IsFunctionConcaveClosed (F u)

Lean source

A concave-closed convex-concave kernel is a convex-concave bifunction that is concave-closed in the second variable and takes values in ℝ ∪ {-∞} (hence has no ⊤ values).

def IsConcaveClosedConvexConcaveKernel {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ IsConcaveClosedInSecond K ∧ HasNoTopValuesBifunction K

Lean source

Lemma33.0.18: On the convex side of Definition33.0.17, Theorem33.1 yields a one-to-one correspondence between image-closed Rockafellar convex bifunctions and concave-convex kernels that are convex-closed in the second variable.

theorem convexImageClosedBifunction_pairing_correspondence {m n : ℕ} : (∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConvexBifunction F → (∀ u, IsFunctionConvexClosed (F u)) → HasNoBotValuesBifunction F → IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (convexBifunctionPairing F) ∧ IsConvexClosedInSecond (convexBifunctionPairing F) ∧ ∀ u x, F u x = convexConjugate (convexBifunctionPairing F u) x) ∧ (∀ K : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K → IsConvexClosedInSecond K → HasNoTopOrBotValuesBifunction K → let F : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u x => convexConjugate (K u) x IsRockafellarConvexBifunction F ∧ (∀ u, IsFunctionConvexClosed (F u)) ∧ ∀ u xStar, convexBifunctionPairing F u xStar = K u xStar)

Lean source

Lemma33.0.18, concave side: image-closed Rockafellar concave bifunctions correspond to convex-concave kernels that are concave-closed in the second variable.

theorem concaveImageClosedBifunction_pairing_correspondence {m n : ℕ} : (∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConcaveBifunction F → (∀ u, IsFunctionConcaveClosed (F u)) → HasNoTopValuesBifunction F → IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (concaveBifunctionPairing F) ∧ IsConcaveClosedInSecond (concaveBifunctionPairing F) ∧ ∀ u x, F u x = concaveConjugate (concaveBifunctionPairing F u) x) ∧ (∀ K : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K → IsConcaveClosedInSecond K → HasNoTopOrBotValuesBifunction K → let F : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u x => concaveConjugate (K u) x IsRockafellarConcaveBifunction F ∧ (∀ u, IsFunctionConcaveClosed (F u)) ∧ ∀ u xStar, concaveBifunctionPairing F u xStar = K u xStar)

Lean source

Corollary33.1.2: The relations K(u, x^*) = ⟪F u, x^*⟫ and F u = K(u, ·)^* express a one-to-one correspondence between convex-closed concave-convex kernels K on ℝ^m × ℝ^n and image-closed convex bifunctions F : ℝ^m → (ℝ^n → EReal). Similarly, the concave pairing gives a one-to-one correspondence between concave-closed convex-concave kernels and image-closed concave bifunctions. Here the image-closed and closed-kernel classes are interpreted in the local extended-real formalization, including the one-sided admissibility hypotheses excluding ⊥ on the convex side and excluding ⊤ on the concave side.

theorem closedSaddleFunctions_imageClosedBifunctions_correspondence {m n : ℕ} : (∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsImageClosedConvexBifunction F → IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (convexBifunctionPairing F) ∧ IsConvexClosedInSecond (convexBifunctionPairing F) ∧ ∀ u x, convexConjugate (convexBifunctionPairing F u) x = F u x) ∧ (∀ K : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K → IsConvexClosedInSecond K → HasNoTopOrBotValuesBifunction K → let F : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u x => convexConjugate (K u) x IsImageClosedConvexBifunction F ∧ (∀ u x, F u x = convexConjugate (K u) x) ∧ ∀ u xStar, convexBifunctionPairing F u xStar = K u xStar) ∧ (∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsImageClosedConcaveBifunction F → IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (concaveBifunctionPairing F) ∧ IsConcaveClosedInSecond (concaveBifunctionPairing F) ∧ ∀ u x, concaveConjugate (concaveBifunctionPairing F u) x = F u x) ∧ (∀ K : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K → IsConcaveClosedInSecond K → HasNoTopOrBotValuesBifunction K → let F : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u x => concaveConjugate (K u) x IsImageClosedConcaveBifunction F ∧ (∀ u x, F u x = concaveConjugate (K u) x) ∧ ∀ u xStar, concaveBifunctionPairing F u xStar = K u xStar)

Lean source

Helper for Corollary33.1.1: swapping the variables turns a convex-concave kernel into a concave-convex kernel.

lemma helperForCorollary33_1_1_swap_preserves_convexConcave {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K) : IsConcaveConvexOn (Set.univ : Set (Fin n → ℝ)) (Set.univ : Set (Fin m → ℝ)) (fun v u => K u v)

Lean source

Helper for Corollary33.1.1: swapping the variables turns a concave-convex kernel into a convex-concave kernel.

lemma helperForCorollary33_1_1_swap_preserves_concaveConvex {m n : ℕ} {K : (Fin n → ℝ) → (Fin m → ℝ) → EReal} (hK : IsConcaveConvexOn (Set.univ : Set (Fin n → ℝ)) (Set.univ : Set (Fin m → ℝ)) K) : IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (fun u v => K v u)

Lean source

Helper for Corollary33.1.1: a swapped kernel that is convex-closed in the second variable becomes convex-closed in the first variable after unswapping.

lemma helperForCorollary33_1_1_swap_convexClosedInSecond_to_convexClosedInFirst {m n : ℕ} {K : (Fin n → ℝ) → (Fin m → ℝ) → EReal} (hK : IsConvexClosedInSecond K) : IsConvexClosedInFirst (fun u v => K v u)

Lean source

Helper for Corollary33.1.1: a swapped kernel that is concave-closed in the first variable becomes concave-closed in the second variable after unswapping.

lemma helperForCorollary33_1_1_swap_concaveClosedInFirst_to_concaveClosedInSecond {m n : ℕ} {K : (Fin n → ℝ) → (Fin m → ℝ) → EReal} (hK : IsConcaveClosedInFirst K) : IsConcaveClosedInSecond (fun u v => K v u)

Lean source

Helper for the concave side of Theorem33.1: negating a concave EReal section that avoids ⊤ produces a convex section. This is the exact sign-transport used to dualize the convex-side Rockafellar package.

lemma helperForCorollary33_1_2_concaveNegation_isConvex_of_noTop {k : ℕ} {f : (Fin k → ℝ) → EReal} (hConc : IsERealConcaveOn (Set.univ : Set (Fin k → ℝ)) f) (hNoTop : ∀ x, f x ≠ ⊤) : IsERealConvexOn (Set.univ : Set (Fin k → ℝ)) (fun x => -f x)

Lean source

Helper for the concave side of Theorem33.1: the convex pairing of the negated bifunction is exactly the negative of the original concave pairing, with the expected sign flip in the dual variable.

lemma helperForCorollary33_1_2_convexPairing_negated_eq_neg_concavePairing {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) : convexBifunctionPairing (fun u' x => -F u' x) u xStar = -concaveBifunctionPairing F u (-xStar)

Lean source

Helper for the concave side of Theorem33.1: negating a Rockafellar concave bifunction that avoids ⊤ yields the corresponding Rockafellar convex bifunction that avoids ⊥.

lemma helperForCorollary33_1_2_negatedConcaveBifunction_isRockafellarConvex {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hRock : IsRockafellarConcaveBifunction F) (hNoTop : HasNoTopValuesBifunction F) : let G : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u x => -F u x IsRockafellarConvexBifunction G ∧ HasNoBotValuesBifunction G

Lean source

Helper for Lemma33.0.18: a convex-closed section agrees pointwise with its convex closure.

lemma helperForLemma33_0_18_functionConvexClosure_eq_self {n : ℕ} {f : (Fin n → ℝ) → EReal} (h : IsFunctionConvexClosed f) : ∀ x, functionConvexClosure f x = f x

Lean source

Helper for Lemma33.0.18: a kernel convex-closed in the second variable agrees pointwise with its second-variable convex closure.

lemma helperForLemma33_0_18_convexClosureInSecond_eq_self {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (h : IsConvexClosedInSecond K) : ∀ u xStar, convexClosureInSecond K u xStar = K u xStar

Lean source

Helper for Lemma33.0.18: a concave-closed section agrees pointwise with its concave closure.

lemma helperForLemma33_0_18_functionConcaveClosure_eq_self {n : ℕ} {f : (Fin n → ℝ) → EReal} (h : IsFunctionConcaveClosed f) : ∀ x, functionConcaveClosure f x = f x

Lean source

Helper for Lemma33.0.18: a kernel concave-closed in the second variable agrees pointwise with its second-variable concave closure.

lemma helperForLemma33_0_18_concaveClosureInSecond_eq_self {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (h : IsConcaveClosedInSecond K) : ∀ u xStar, concaveClosureInSecond K u xStar = K u xStar

Lean source

Helper for Corollary33.1.3: dot products split over Fin.append.

lemma helperForCorollary33_1_3_dotProduct_append {m k : ℕ} (u : Fin m → ℝ) (v : Fin k → ℝ) (b : Fin (m + k) → ℝ) : dotProduct (Fin.append u v) b = dotProduct u (fun i => b (Fin.castAdd k i)) + dotProduct v (fun j => b (Fin.natAdd m j))

Lean source

Helper for Corollary33.1.3: the castSucc coordinates of the packed epigraph embedding recover the base point.

lemma helperForCorollary33_1_3_prodLinearEquivAppend_ofLp_castSucc {n : ℕ} (x : Fin n → ℝ) (μ : ℝ) (j : Fin n) : (prodLinearEquiv_append (n := n) (x, μ)).ofLp (Fin.castSucc j) = x j

Lean source

Helper for Corollary33.1.3: the last coordinate of the packed epigraph embedding is the appended scalar.

lemma helperForCorollary33_1_3_prodLinearEquivAppend_ofLp_last {n : ℕ} (x : Fin n → ℝ) (μ : ℝ) : (prodLinearEquiv_append (n := n) (x, μ)).ofLp (Fin.last n) = μ

Lean source

Helper for Corollary33.1.3: appending a fixed parameter block commutes with the packed epigraph embedding (viewed in coordinates via .ofLp).

lemma helperForCorollary33_1_3_append_prodLinearEquivAppend_ofLp {m n : ℕ} (u : Fin m → ℝ) (x : Fin n → ℝ) (μ : ℝ) : Fin.append u ((prodLinearEquiv_append (n := n) (x, μ)).ofLp) = (prodLinearEquiv_append (n := m + n) (Fin.append u x, μ)).ofLp

§33 — Saddle-Functions — Part 8

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Corollary33.1.3: If F is a polyhedral convex bifunction from ℝ^m to ℝ^n, then for each u the partial pairing x^* ↦ ⟪F u, x^*⟫ is a polyhedral convex function, and for each x^* the parameter section u ↦ ⟪F u, x^*⟫ is a polyhedral concave function. Moreover, if every section F u is proper convex on ℝ^n, then F is recovered from its pairing by (F u) x = sup_{x^*} (⟪x, x^*⟫ - ⟪F u, x^*⟫), formalized here via convexConjugate.

theorem polyhedralConvexBifunction_pairing_sections_and_reconstruction {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : IsRockafellarPolyhedralConvexBifunction F) : (∀ u : Fin m → ℝ, IsPolyhedralConvexFunction n (convexBifunctionPairing F u)) ∧ (∀ xStar : Fin n → ℝ, IsPolyhedralConvexFunction m (fun u => -(convexBifunctionPairing F u xStar))) ∧ ((∀ u : Fin m → ℝ, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (F u)) → ∀ u x, F u x = convexConjugate (convexBifunctionPairing F u) x)

Lean source

Corollary 33.19.

theorem section33_corollary33_19 : True

Lean source

Corollary 33.20.

theorem section33_corollary33_20 : True

Lean source

Helper for Corollary33.1.3: sections of a polyhedral convex bifunction are polyhedral convex functions.

lemma helperForCorollary33_1_3_section_isPolyhedralConvexFunction {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : IsRockafellarPolyhedralConvexBifunction F) : ∀ u : Fin m → ℝ, IsPolyhedralConvexFunction n (F u)

Lean source

Helper for Corollary33.1.3: the tilted graph-function fiber over a fixed u is the negative of the fixed-dual pairing section.

lemma helperForCorollary33_1_3_sInf_tiltedFiber_eq_negSup_pairing {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) : sInf (Set.range fun x : Fin n → ℝ => graphFunctionOfBifunction F (Fin.append u x) - ((dotProduct x xStar : ℝ) : EReal)) = -(convexBifunctionPairing F u xStar)

Lean source

Helper for Corollary33.1.3: the packed embedding p ↦ (prodLinearEquiv_append p).ofLp is injective.

lemma helperForCorollary33_1_3_prodLinearEquivAppend_ofLp_injective {n : ℕ} : Function.Injective (fun p : (Fin n → ℝ) × ℝ => (prodLinearEquiv_append (n := n) p).ofLp)

Lean source

Helper for Corollary33.1.3: the x-block dot product z ↦ ⟪x(z), xStar⟫ is linear in z under convex combinations.

lemma helperForCorollary33_1_3_dotProduct_xBlock_weighted {m n : ℕ} (a b : ℝ) (z₁ z₂ : Fin (m + n) → ℝ) (xStar : Fin n → ℝ) : dotProduct (fun j => (a • z₁ + b • z₂) (Fin.natAdd m j)) xStar = a * dotProduct (fun j => z₁ (Fin.natAdd m j)) xStar + b * dotProduct (fun j => z₂ (Fin.natAdd m j)) xStar

Lean source

Helper for Corollary33.1.3: the modified normal bTilt evaluates to its bLast value on the last coordinate.

lemma helperForCorollary33_1_3_bTilt_last {m n : ℕ} (xStar : Fin n → ℝ) (b : Fin (m + n + 1) → ℝ) : let bLast : ℝ := b (Fin.last (m + n)) let bTilt : Fin (m + n + 1) → ℝ := Fin.lastCases bLast (fun j : Fin (m + n) => Fin.addCases (m := m) (n := n) (fun jm : Fin m => b (Fin.castSucc (Fin.castAdd n jm))) (fun jn : Fin n => b (Fin.castSucc (Fin.natAdd m jn)) + bLast * xStar jn) j) bTilt (Fin.last (m + n)) = bLast

Lean source

Helper for Corollary33.1.3: on the u-block, the modified normal bTilt agrees with b.

lemma helperForCorollary33_1_3_bTilt_castSucc_castAdd {m n : ℕ} (xStar : Fin n → ℝ) (b : Fin (m + n + 1) → ℝ) (jm : Fin m) : let bLast : ℝ := b (Fin.last (m + n)) let bTilt : Fin (m + n + 1) → ℝ := Fin.lastCases bLast (fun j : Fin (m + n) => Fin.addCases (m := m) (n := n) (fun jm : Fin m => b (Fin.castSucc (Fin.castAdd n jm))) (fun jn : Fin n => b (Fin.castSucc (Fin.natAdd m jn)) + bLast * xStar jn) j) bTilt (Fin.castSucc (Fin.castAdd n jm)) = b (Fin.castSucc (Fin.castAdd n jm))

Lean source

Helper for Corollary33.1.3: on the x-block, the modified normal bTilt adds the extra term bLast * xStar.

lemma helperForCorollary33_1_3_bTilt_castSucc_natAdd {m n : ℕ} (xStar : Fin n → ℝ) (b : Fin (m + n + 1) → ℝ) (jn : Fin n) : let bLast : ℝ := b (Fin.last (m + n)) let bTilt : Fin (m + n + 1) → ℝ := Fin.lastCases bLast (fun j : Fin (m + n) => Fin.addCases (m := m) (n := n) (fun jm : Fin m => b (Fin.castSucc (Fin.castAdd n jm))) (fun jn : Fin n => b (Fin.castSucc (Fin.natAdd m jn)) + bLast * xStar jn) j) bTilt (Fin.castSucc (Fin.natAdd m jn)) = b (Fin.castSucc (Fin.natAdd m jn)) + b (Fin.last (m + n)) * xStar jn

Lean source

Helper for Corollary33.1.3: the extra x-block sum induced by bLast * xStar is exactly (⟪x(z), xStar⟫) * bLast.

lemma helperForCorollary33_1_3_sum_xBlock_mul_bLast {m n : ℕ} (z : Fin (m + n) → ℝ) (xStar : Fin n → ℝ) (bLast : ℝ) : (∑ j : Fin n, z (Fin.natAdd m j) * (bLast * xStar j)) = (dotProduct (fun j => z (Fin.natAdd m j)) xStar) * bLast

Lean source

Helper for Corollary33.1.3: shifting the last packed coordinate by ⟪x(z), xStar⟫ is equivalent to modifying the halfspace normal by adding b_last • xStar on the x-block.

lemma helperForCorollary33_1_3_dotProduct_packed_lastShift_eq_modifiedNormal {m n : ℕ} (xStar : Fin n → ℝ) (b : Fin (m + n + 1) → ℝ) (z : Fin (m + n) → ℝ) (μ : ℝ) : let xBlock : Fin n → ℝ := fun j => z (Fin.natAdd m j) let bLast : ℝ := b (Fin.last (m + n)) let bTilt : Fin (m + n + 1) → ℝ := Fin.lastCases bLast (fun j : Fin (m + n) => Fin.addCases (m := m) (n := n) (fun jm : Fin m => b (Fin.castSucc (Fin.castAdd n jm))) (fun jn : Fin n => b (Fin.castSucc (Fin.natAdd m jn)) + bLast * xStar jn) j) dotProduct ((prodLinearEquiv_append (n := m + n) (z, μ + dotProduct xBlock xStar)).ofLp) b = dotProduct ((prodLinearEquiv_append (n := m + n) (z, μ)).ofLp) bTilt

Lean source

Helper for Corollary33.1.3: subtracting the fixed linear functional in the x-block from the graph function preserves polyhedral convexity.

lemma helperForCorollary33_1_3_tiltedGraph_isPolyhedralConvexFunction {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : IsRockafellarPolyhedralConvexBifunction F) (xStar : Fin n → ℝ) : IsPolyhedralConvexFunction (m + n) (fun z => graphFunctionOfBifunction F z - ((dotProduct (fun j => z (Fin.natAdd m j)) xStar : ℝ) : EReal))

Lean source

Helper for Corollary33.1.3: projecting the tilted graph function to the u-coordinates recovers the negative pairing section.

lemma helperForCorollary33_1_3_projectionImage_tiltedGraph_eq_negPairing {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) : imageUnderLinearMap (projectionLinearMap (Nat.le_add_right m n)) (fun z : Fin (m + n) → ℝ => graphFunctionOfBifunction F z - ((dotProduct (fun j => z (Fin.natAdd m j)) xStar : ℝ) : EReal)) u = -(convexBifunctionPairing F u xStar)

Lean source

Helper for Corollary33.1.3: for fixed xStar, the map u ↦ -(convexBifunctionPairing F u xStar) is polyhedral convex.

lemma helperForCorollary33_1_3_negPairingSection_isPolyhedralConvexFunction {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : IsRockafellarPolyhedralConvexBifunction F) (xStar : Fin n → ℝ) : IsPolyhedralConvexFunction m (fun u => -(convexBifunctionPairing F u xStar))

§33 — Saddle-Functions — Part 9

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

The graph-function notion of convexity for a bifunction.

def IsGraphConvexBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F) -- Proof sketch: uncurry a bifunction `F` to its graph function on `ℝ^(m+n)` and curry a -- function `f` on `ℝ^(m+n)` back along the splitting `ℝ^(m+n) ≃ ℝ^m × ℝ^n`. These -- operations are inverse to each other, and the graph-function notion of convexity for `F` -- is exactly convexity of the uncurried function on the whole space.

Lean source

Lemma33.0.21 (Graph function correspondence): convex bifunctions F : ℝ^m → (ℝ^n → EReal) correspond one-to-one with their graph functions f (u, x) = (F u) x, and these graph functions are precisely the convex functions on ℝ^(m + n).

theorem graphConvexBifunction_graphFunction_correspondence {m n : ℕ} : (∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsGraphConvexBifunction F ↔ IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F)) ∧ (∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsGraphConvexBifunction F → bifunctionOfGraphFunction (graphFunctionOfBifunction F) = F) ∧ ∀ f : (Fin (m + n) → ℝ) → EReal, IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) f → graphFunctionOfBifunction (bifunctionOfGraphFunction f) = f

Lean source

The adjoint of a convex bifunction, viewed as the partial conjugate in the x-variable.

noncomputable abbrev convexBifunctionAdjoint {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (xStar : Fin n → ℝ) : (Fin m → ℝ) → EReal := fun u => convexBifunctionPairing F u xStar

Lean source

The adjoint of a concave bifunction, viewed as the partial conjugate in the x-variable through the concave pairing convention.

noncomputable abbrev concaveBifunctionAdjoint {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (xStar : Fin n → ℝ) : (Fin m → ℝ) → EReal := fun u => concaveBifunctionPairing F u xStar

Lean source

Helper for Lemma33.0.22: graph-function convexity of F implies Rockafellar convexity. This just curries/uncurries along the Fin.append splitting and reuses the already-proved lemmas that extract sectionwise convexity and concavity of the convex pairing from a convex graph function.

lemma helperForLemma33_0_22_graphConvex_gives_rockafellarConvex {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_convex : IsGraphConvexBifunction F) : IsRockafellarConvexBifunction F

Lean source

Helper for Lemma33.0.22: the adjoint kernel (u, x^*) ↦ (F^* x^*)(u) is concave-convex. This is the "easy" part of Lemma33.0.22: once F is Rockafellar-convex and avoids ⊥, Theorem33.1 (convexBifunction_pairing_correspondence) directly yields concave-convexity of the convex pairing, and the adjoint is definitionally that pairing.

lemma helperForLemma33_0_22_adjointKernel_isConcaveConvex {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_convex : IsGraphConvexBifunction F) (hF_noBot : HasNoBotValuesBifunction F) : IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (fun u xStar => convexBifunctionAdjoint F xStar u)

Lean source

Helper for Lemma33.0.22: the adjoint kernel is convex-closed in the dual variable. This is the second-variable closedness already supplied by Theorem33.1 for the convex pairing, rewritten through the adjoint notation.

lemma helperForLemma33_0_22_adjointKernel_isConvexClosedInSecond {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_convex : IsGraphConvexBifunction F) (hF_noBot : HasNoBotValuesBifunction F) : IsConvexClosedInSecond (fun u xStar => convexBifunctionAdjoint F xStar u)

Lean source

Helper for Lemma33.0.22: every adjoint section u ↦ (F^* x^*)(u) is concave. This is the first-variable half of the already-established concave-convexity of the adjoint kernel.

lemma helperForLemma33_0_22_adjointSection_isConcave {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_convex : IsGraphConvexBifunction F) (hF_noBot : HasNoBotValuesBifunction F) (xStar : Fin n → ℝ) : IsERealConcaveOn (Set.univ : Set (Fin m → ℝ)) (convexBifunctionAdjoint F xStar)

Lean source

Helper for Lemma33.0.22: closedness of the graph function forces closedness of every primal section F u. This is the dependency-closed sectionwise closedness fact needed by the valid Section 33 route: once the graph function is already fixed by the raw convex-closure operator, freezing the parameter u and composing with x ↦ Fin.append u x preserves lower semicontinuity, so the resulting section is fixed by its own one-variable convex closure.

lemma helperForLemma33_0_22_section_isFunctionConvexClosed {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_closed : IsFunctionConvexClosed (graphFunctionOfBifunction F)) : ∀ u : Fin m → ℝ, IsFunctionConvexClosed (F u)

Lean source

Helper for Lemma33.0.22: every fixed-dual pairing section avoids ⊥ whenever each primal section has a point that avoids ⊤. This is the one-function admissibility witness reused from Theorem33.1, now specialized to the pairing section u ↦ ⟪F u, x^*⟫.

lemma helperForLemma33_0_22_pairingSection_hasNoBot {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_notTop : ∀ u : Fin m → ℝ, ∃ x : Fin n → ℝ, F u x ≠ ⊤) : ∀ xStar : Fin n → ℝ, ∀ u : Fin m → ℝ, convexBifunctionPairing F u xStar ≠ (⊥ : EReal)

Lean source

Helper for Lemma33.0.22: once a fixed-dual pairing section is known to be a Chapter 6 concave function, its biconjugate is exactly its Chapter 6 concave closure. This isolates the part of the Section 33 route that really comes from Chapter 6: the only extra missing ingredient is a dependency-closed way to package the pairing section into the ConcaveFunction predicate required by concaveConjugate_biconjugate_eq_concaveClosure.

lemma helperForLemma33_0_22_pairingSection_biconjugate_eq_concaveClosure_of_concaveFunction {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {xStar : Fin n → ℝ} (hConc : ConcaveFunction (fun u' : Fin m → ℝ => convexBifunctionPairing F u' xStar)) : ∀ u : Fin m → ℝ, concaveConjugate (concaveConjugate (fun u' : Fin m → ℝ => convexBifunctionPairing F u' xStar)) u = concaveClosure (fun u' : Fin m → ℝ => convexBifunctionPairing F u' xStar) u

Lean source

Helper for Lemma33.0.22: first-variable closedness of the adjoint kernel specializes to closedness of each fixed-dual adjoint section. This isolates the exact remaining issue in Lemma33.0.22: once the kernel (u, x^*) ↦ (F^* x^*)(u) is known to equal its first-variable concave closure, every frozen section u ↦ (F^* x^*)(u) is automatically fixed by the one-variable concave closure.

lemma helperForLemma33_0_22_adjointSection_isFunctionConcaveClosed_of_kernelClosedInFirst {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (xStar : Fin n → ℝ) (hKernelClosed : IsConcaveClosedInFirst (fun u xStar => convexBifunctionAdjoint F xStar u)) : IsFunctionConcaveClosed (convexBifunctionAdjoint F xStar)

Lean source

Helper for Lemma33.0.22: freezing the second variable of a concave-convex kernel that is concave-closed in the first variable yields a concave, concave-closed section.

lemma helperForLemma33_0_22_frozenSecondVariable_isConcaveAndFunctionConcaveClosed {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hKernelConcConv : IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K) (hKernelClosed : IsConcaveClosedInFirst K) (v : Fin n → ℝ) : IsERealConcaveOn (Set.univ : Set (Fin m → ℝ)) (fun u => K u v) ∧ IsFunctionConcaveClosed (fun u => K u v)

Lean source

Helper for Lemma33.0.22: for each fixed u, the adjoint image x^* ↦ (F^* x^*)(u) is convex-closed. This records the dual-variable closedness that Theorem33.1 provides for the pairing kernel, specialized to a single parameter value.

lemma helperForLemma33_0_22_adjointImage_isFunctionConvexClosed {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_convex : IsGraphConvexBifunction F) (hF_noBot : HasNoBotValuesBifunction F) (u : Fin m → ℝ) : IsFunctionConvexClosed (fun xStar => convexBifunctionAdjoint F xStar u)

Lean source

Helper for Lemma33.0.22: the adjoint kernel never takes the value ⊥ when every primal section F u contains at least one point different from ⊤. This is the same admissibility witness used in Theorem33.1 for a single convex conjugate: freezing u, any point x₀ with F u x₀ ≠ ⊤ contributes a genuine affine term to the supremum defining (F^* x^*)(u), so that supremum cannot collapse to ⊥.

lemma helperForLemma33_0_22_adjointKernel_hasNoBotValues {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_notTop : ∀ u : Fin m → ℝ, ∃ x : Fin n → ℝ, F u x ≠ ⊤) : HasNoBotValuesBifunction (fun u xStar => convexBifunctionAdjoint F xStar u)

Lean source

Helper for Lemma33.0.22: the adjoint kernel already satisfies all dependency-closed kernel properties available in this part file. Concretely, the adjoint kernel is concave-convex, convex-closed in the dual variable, and never equals ⊥; the only missing closedness clause for the textbook lemma is the first-variable concave-closedness recorded separately in the main proof.

lemma helperForLemma33_0_22_adjointKernel_isConvexClosedConcaveConvexKernel {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_convex : IsGraphConvexBifunction F) (hF_noBot : HasNoBotValuesBifunction F) (hF_notTop : ∀ u : Fin m → ℝ, ∃ x : Fin n → ℝ, F u x ≠ ⊤) : IsConvexClosedConcaveConvexKernel (fun u xStar => convexBifunctionAdjoint F xStar u)

Lean source

Helper for Lemma33.0.22: first-variable concave-closedness of a kernel is equivalent to second-variable convex-closedness of the sign-swapped kernel (v, u) ↦ -K u v.

lemma helperForLemma33_0_22_concaveClosedInFirst_iff_swappedNegated_isConvexClosedInSecond {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} : IsConcaveClosedInFirst K ↔ IsConvexClosedInSecond (fun v u => -K u v)

Lean source

Helper for Lemma33.0.22: for the adjoint kernel, first-variable concave-closedness is equivalent to second-variable convex-closedness of the sign-swapped kernel.

lemma helperForLemma33_0_22_adjointKernel_isConcaveClosedInFirst_iff_swappedNegated_isConvexClosedInSecond {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} : IsConcaveClosedInFirst (fun u xStar => convexBifunctionAdjoint F xStar u) ↔ IsConvexClosedInSecond (fun xStar u => -convexBifunctionAdjoint F xStar u)

Lean source

Helper for Lemma33.0.22: a one-variable function is concave-closed exactly when its pointwise negation is convex-closed. This packages the sign-swap identity from the bifunction setting into the frozen-section form needed for the adjoint textbook clause u ↦ (F^* x^*)(u).

lemma helperForLemma33_0_22_functionConcaveClosed_iff_neg_isFunctionConvexClosed {m : ℕ} {g : (Fin m → ℝ) → EReal} : IsFunctionConcaveClosed g ↔ IsFunctionConvexClosed (fun u => -g u)

Lean source

Helper for Lemma33.0.22: graph-function closedness already makes the tilted graph function closed in the one-variable Section 33 sense. For a fixed dual vector xStar, the textbook adjoint is built from the graph function by subtracting the finite affine form x ↦ ⟪x, xStar⟫. Closedness of the graph function therefore propagates to the tilted objective before any projection argument is used.

lemma helperForLemma33_0_22_graphFunction_isProperConvex {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_convex : IsGraphConvexBifunction F) (hF_noBot : HasNoBotValuesBifunction F) (hF_notTop : ∀ u : Fin m → ℝ, ∃ x : Fin n → ℝ, F u x ≠ ⊤) : ProperConvexFunctionOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F)

Lean source

Helper for Lemma33.0.22: graph-function closedness already makes the tilted graph function closed in the one-variable Section 33 sense. For a fixed dual vector xStar, the textbook adjoint is built from the graph function by subtracting the finite affine form x ↦ ⟪x, xStar⟫. Closedness of the graph function therefore propagates to the tilted objective before any projection argument is used.

lemma helperForLemma33_0_22_tiltedGraph_isLowerSemicontinuous {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_closed : IsFunctionConvexClosed (graphFunctionOfBifunction F)) (xStar : Fin n → ℝ) : LowerSemicontinuous (fun z : Fin (m + n) → ℝ => graphFunctionOfBifunction F z - ((dotProduct (fun j : Fin n => z (Fin.natAdd m j)) xStar : ℝ) : EReal))

Lean source

Helper for Lemma33.0.22: graph-function closedness already makes the tilted graph function closed in the one-variable Section 33 sense. For a fixed dual vector xStar, the textbook adjoint is built from the graph function by subtracting the finite affine form x ↦ ⟪x, xStar⟫. Closedness of the graph function therefore propagates to the tilted objective before any projection argument is used.

lemma helperForLemma33_0_22_tiltedGraph_isFunctionConvexClosed {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_closed : IsFunctionConvexClosed (graphFunctionOfBifunction F)) (xStar : Fin n → ℝ) : IsFunctionConvexClosed (fun z : Fin (m + n) → ℝ => graphFunctionOfBifunction F z - ((dotProduct (fun j : Fin n => z (Fin.natAdd m j)) xStar : ℝ) : EReal))

Lean source

Helper for Lemma33.0.22: graph-function convexity already makes the tilted graph convex. This settles the convex half of the textbook inf-projection route inside the current dependency closure: subtracting the finite linear form x ↦ ⟪x, xStar⟫ preserves convexity of the graph function before any closedness or normality argument is used.

lemma helperForLemma33_0_22_tiltedGraph_isConvexFunction {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_convex : IsGraphConvexBifunction F) (xStar : Fin n → ℝ) : ConvexFunction (fun z : Fin (m + n) → ℝ => graphFunctionOfBifunction F z - ((dotProduct (fun j : Fin n => z (Fin.natAdd m j)) xStar : ℝ) : EReal))

Lean source

Helper for Lemma33.0.22: the tilted graph already forms a closed proper convex function. This finishes the local Chapter 2 preparation before projection: after subtracting the finite linear form in the x-coordinates, the graph function is still proper convex and remains closed because the tilt preserves lower semicontinuity.

lemma helperForLemma33_0_22_tiltedGraph_isClosedProperConvex {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_convex : IsGraphConvexBifunction F) (hF_closed : IsFunctionConvexClosed (graphFunctionOfBifunction F)) (hF_noBot : HasNoBotValuesBifunction F) (hF_notTop : ∀ u : Fin m → ℝ, ∃ x : Fin n → ℝ, F u x ≠ ⊤) (xStar : Fin n → ℝ) : ClosedConvexFunction (fun z : Fin (m + n) → ℝ => graphFunctionOfBifunction F z - ((dotProduct (fun j : Fin n => z (Fin.natAdd m j)) xStar : ℝ) : EReal)) ∧ ProperConvexFunctionOn (Set.univ : Set (Fin (m + n) → ℝ)) (fun z : Fin (m + n) → ℝ => graphFunctionOfBifunction F z - ((dotProduct (fun j : Fin n => z (Fin.natAdd m j)) xStar : ℝ) : EReal))

Lean source

Helper for Lemma33.0.22: the tilted inf-projection is already convex before the missing closedness bridge is invoked. This records the Chapter 2 fiber-infimum theorem in the present notation: once the tilted graph is convex, projecting it along the x-coordinates preserves convexity of the resulting one-variable section.

lemma helperForLemma33_0_22_tiltedProjection_isConvexFunction {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_convex : IsGraphConvexBifunction F) (xStar : Fin n → ℝ) : ConvexFunction (fun u : Fin m → ℝ => imageUnderLinearMap (projectionLinearMap (Nat.le_add_right m n)) (fun z : Fin (m + n) → ℝ => graphFunctionOfBifunction F z - ((dotProduct (fun j : Fin n => z (Fin.natAdd m j)) xStar : ℝ) : EReal)) u)

Lean source

Helper for Lemma33.0.22: the sign-swapped adjoint section is exactly the projection image of the tilted graph function. This is the textbook inf-projection formula u ↦ inf_x (f(u, x) - ⟪x, xStar⟫) = - (F^* xStar)(u) written using the graph-function model and the Section 33 imageUnderLinearMap notation for projection along the x-coordinates.

lemma helperForLemma33_0_22_swappedNegatedAdjointSection_eq_projectionImage_tiltedGraph {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (xStar : Fin n → ℝ) : (fun u : Fin m → ℝ => imageUnderLinearMap (projectionLinearMap (Nat.le_add_right m n)) (fun z : Fin (m + n) → ℝ => graphFunctionOfBifunction F z - ((dotProduct (fun j : Fin n => z (Fin.natAdd m j)) xStar : ℝ) : EReal)) u) = fun u : Fin m → ℝ => -convexBifunctionAdjoint F xStar u

Lean source

Helper for Lemma33.0.22: the remaining tilted inf-projection closedness statement is exactly the textbook closedness of the frozen adjoint section. This reformulates the blocker at a fixed xStar: the projected tilted graph is convex-closed iff u ↦ (F^* x^*)(u) is concave-closed, via the sign-negation identity and the projection formula for the tilted graph.

lemma helperForLemma33_0_22_adjointSection_isFunctionConcaveClosed_iff_tiltedProjectionClosed {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (xStar : Fin n → ℝ) : IsFunctionConcaveClosed (convexBifunctionAdjoint F xStar) ↔ IsFunctionConvexClosed (fun u : Fin m → ℝ => imageUnderLinearMap (projectionLinearMap (Nat.le_add_right m n)) (fun z : Fin (m + n) → ℝ => graphFunctionOfBifunction F z - ((dotProduct (fun j : Fin n => z (Fin.natAdd m j)) xStar : ℝ) : EReal)) u)

Lean source

Helper for Lemma33.0.22: once the tilted projected section at xStar is convex-closed, the textbook sectionwise conclusion already follows at that xStar. This packages the two surviving dependency-closed ingredients together: concavity comes from the joint adjoint kernel, and closedness is exactly the projected tilted-section statement proved equivalent just above.

lemma helperForLemma33_0_22_adjointSection_isConcaveAndFunctionConcaveClosed_of_tiltedProjectionClosed {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_convex : IsGraphConvexBifunction F) (hF_noBot : HasNoBotValuesBifunction F) (xStar : Fin n → ℝ) (hClosedSection : IsFunctionConvexClosed (fun u : Fin m → ℝ => imageUnderLinearMap (projectionLinearMap (Nat.le_add_right m n)) (fun z : Fin (m + n) → ℝ => graphFunctionOfBifunction F z - ((dotProduct (fun j : Fin n => z (Fin.natAdd m j)) xStar : ℝ) : EReal)) u)) : IsERealConcaveOn (Set.univ : Set (Fin m → ℝ)) (convexBifunctionAdjoint F xStar) ∧ IsFunctionConcaveClosed (convexBifunctionAdjoint F xStar)

Lean source

Helper for Corollary33.3.1: if K is obtained by taking the second-variable convex closure of Kbar, then K lies pointwise below Kbar.

lemma helperForCorollary33_3_1_convexClosureInSecond_eq_implies_pointwise_le {m n : ℕ} {K Kbar : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSecond : convexClosureInSecond Kbar = K) : ∀ u xStar, K u xStar ≤ Kbar u xStar

Lean source

Helper for Corollary33.3.1: the mixed coordinatewise closure identities already imply the pointwise order K ≤ Kbar.

lemma helperForCorollary33_3_1_coordinatewise_closure_pair_implies_pointwise_le {m n : ℕ} {K Kbar : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hFirst : Kbar = concaveClosureInFirst K) (hSecond : convexClosureInSecond Kbar = K) : ∀ u xStar, K u xStar ≤ Kbar u xStar

Lean source

Helper for Corollary33.3.1: the mixed coordinatewise closure identities already force the lower-closure composition cl₂ (cl₁ K) = K.

lemma helperForCorollary33_3_1_coordinatewise_closure_pair_implies_lowerClosureComposition {m n : ℕ} {K Kbar : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hFirst : Kbar = concaveClosureInFirst K) (hSecond : convexClosureInSecond Kbar = K) : convexClosureInSecond (concaveClosureInFirst K) = K

Lean source

Helper for Corollary33.3.1: the mixed coordinatewise closure identities already force the upper-closure composition cl₁ (cl₂ Kbar) = Kbar.

lemma helperForCorollary33_3_1_coordinatewise_closure_pair_implies_upperClosureComposition {m n : ℕ} {K Kbar : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hFirst : Kbar = concaveClosureInFirst K) (hSecond : convexClosureInSecond Kbar = K) : concaveClosureInFirst (convexClosureInSecond Kbar) = Kbar

Lean source

Helper for Lemma33.0.37: if a primal section is constantly ⊤, then its convex pairing is constantly ⊥.

lemma helperForLemma33_0_37_convexPairing_eq_bot_of_allTopSection {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} (hAllTop : ∀ x : Fin n → ℝ, F u x = ⊤) : ∀ xStar : Fin n → ℝ, convexBifunctionPairing F u xStar = ⊥

Lean source

Helper for Lemma33.0.37: if a primal section is constantly ⊤, then the whole convex pairing section is the constant function ⊥.

lemma helperForLemma33_0_37_convexPairingSection_eq_constBot_of_allTopSection {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} (hAllTop : ∀ x : Fin n → ℝ, F u x = ⊤) : convexBifunctionPairing F u = fun _ : Fin n → ℝ => (⊥ : EReal)

§33 — Saddle-Functions — Part 10

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Helper for Lemma33.0.34: the two linear tilt terms in y combine into a single tilt by xStar + yStar.

lemma helperForLemma33_0_34_tilt_terms_combine {n : ℕ} {Fv : EReal} {y xStar yStar : Fin n → ℝ} : (Fv - ((dotProduct y xStar : ℝ) : EReal)) - ((dotProduct y yStar : ℝ) : EReal) = Fv - ((dotProduct y (xStar + yStar) : ℝ) : EReal)

Lean source

Helper for Theorem33.0.39: if the adjoint kernel is concave-convex and already concave-closed in its first variable, then every fixed dual section u ↦ (F^* x^*)(u) is concave and concave-closed.

lemma helperForTheorem33_0_39_adjointSections_areConcaveAndFunctionConcaveClosed_of_kernelClosed {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_convex : IsGraphConvexBifunction F) (hF_noBot : HasNoBotValuesBifunction F) (hKernelClosed : IsConcaveClosedInFirst (fun u xStar => convexBifunctionAdjoint F xStar u)) : ∀ xStar : Fin n → ℝ, IsERealConcaveOn (Set.univ : Set (Fin m → ℝ)) (convexBifunctionAdjoint F xStar) ∧ IsFunctionConcaveClosed (convexBifunctionAdjoint F xStar)

Lean source

Helper for Corollary33.0.40: if the strict primal domain {u | ∃ x, F u x < ⊤} is all of ℝ^m, then every parameter lies in its intrinsic interior.

lemma helperForCorollary33_0_40_mem_intrinsicInterior_of_fullStrictPrimalDomain {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hDom : {u : Fin m → ℝ | ∃ x : Fin n → ℝ, F u x < ⊤} = Set.univ) (u : Fin m → ℝ) : u ∈ intrinsicInterior ℝ {u' : Fin m → ℝ | ∃ x : Fin n → ℝ, F u' x < ⊤}

Lean source

Helper for Corollary33.0.40: if the closure-side adjoint domain {x^* | ∃ u, (F^* x^*)(u) ≠ ⊥} is all of ℝ^n, then every dual vector lies in its intrinsic interior.

lemma helperForCorollary33_0_40_mem_intrinsicInterior_of_fullClosureSideAdjointDomain {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hDom : {xStar : Fin n → ℝ | ∃ u : Fin m → ℝ, convexBifunctionAdjoint F xStar u ≠ ⊥} = Set.univ) (xStar : Fin n → ℝ) : xStar ∈ intrinsicInterior ℝ {xStar' : Fin n → ℝ | ∃ u : Fin m → ℝ, convexBifunctionAdjoint F xStar' u ≠ ⊥}

Lean source

Helper for Corollary33.0.40: if the strict primal domain is all of ℝ^m, then its intrinsic interior is also all of ℝ^m.

lemma helperForCorollary33_0_40_intrinsicInterior_eq_univ_of_fullStrictPrimalDomain {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hDom : {u : Fin m → ℝ | ∃ x : Fin n → ℝ, F u x < ⊤} = Set.univ) : intrinsicInterior ℝ {u' : Fin m → ℝ | ∃ x : Fin n → ℝ, F u' x < ⊤} = Set.univ

Lean source

Helper for Corollary33.0.40: if the closure-side adjoint domain is all of ℝ^n, then its intrinsic interior is also all of ℝ^n.

lemma helperForCorollary33_0_40_intrinsicInterior_eq_univ_of_fullClosureSideAdjointDomain {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hDom : {xStar : Fin n → ℝ | ∃ u : Fin m → ℝ, convexBifunctionAdjoint F xStar u ≠ ⊥} = Set.univ) : intrinsicInterior ℝ {xStar' : Fin n → ℝ | ∃ u : Fin m → ℝ, convexBifunctionAdjoint F xStar' u ≠ ⊥} = Set.univ

Lean source

Helper for Lemma33.0.43: on the concave-convex branch, first-variable concave-closedness and second-variable convex-closedness force the lower-closed iterated-closure identity.

lemma helperForLemma33_0_43_lowerClosedIdentity_of_firstConcaveClosed_and_secondConvexClosed {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hFirst : IsConcaveClosedInFirst K) (hSecond : IsConvexClosedInSecond K) : convexClosureInSecond (concaveClosureInFirst K) = K

Lean source

Helper for Lemma33.0.43: on the concave-convex branch, first-variable concave-closedness and second-variable convex-closedness force the upper-closed iterated-closure identity.

lemma helperForLemma33_0_43_upperClosedIdentity_of_firstConcaveClosed_and_secondConvexClosed {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hFirst : IsConcaveClosedInFirst K) (hSecond : IsConvexClosedInSecond K) : concaveClosureInFirst (convexClosureInSecond K) = K

Lean source

Helper for Lemma33.0.43: on the convex-concave branch, first-variable convex-closedness and second-variable concave-closedness force the lower-closed iterated-closure identity.

lemma helperForLemma33_0_43_lowerClosedIdentity_of_firstConvexClosed_and_secondConcaveClosed {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hFirst : IsConvexClosedInFirst K) (hSecond : IsConcaveClosedInSecond K) : convexClosureInFirst (concaveClosureInSecond K) = K

Lean source

Helper for Lemma33.0.43: on the convex-concave branch, first-variable convex-closedness and second-variable concave-closedness force the upper-closed iterated-closure identity.

lemma helperForLemma33_0_43_upperClosedIdentity_of_firstConvexClosed_and_secondConcaveClosed {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hFirst : IsConvexClosedInFirst K) (hSecond : IsConcaveClosedInSecond K) : concaveClosureInSecond (convexClosureInFirst K) = K

Lean source

Helper for Lemma33.0.43: the concave-convex branch packages both iterated-closure identities needed for the lower-closed and upper-closed predicates.

lemma helperForLemma33_0_43_concaveConvexBranch_packages_lower_and_upper_closed_data {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hShape : IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K) (hFirst : IsConcaveClosedInFirst K) (hSecond : IsConvexClosedInSecond K) : (IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ convexClosureInSecond (concaveClosureInFirst K) = K) ∧ (IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ concaveClosureInFirst (convexClosureInSecond K) = K)

Lean source

Helper for Lemma33.0.43: the convex-concave branch packages both iterated-closure identities needed for the lower-closed and upper-closed predicates.

lemma helperForLemma33_0_43_convexConcaveBranch_packages_lower_and_upper_closed_data {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hShape : IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K) (hFirst : IsConvexClosedInFirst K) (hSecond : IsConcaveClosedInSecond K) : (IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ convexClosureInFirst (concaveClosureInSecond K) = K) ∧ (IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ concaveClosureInSecond (convexClosureInFirst K) = K)

Lean source

Helper for Lemma33.0.43: the lower-closed iterated-closure identity already makes the kernel convex-closed in the second variable.

lemma helperForLemma33_0_43_secondConvexClosed_of_lowerClosedIdentity {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hLower : convexClosureInSecond (concaveClosureInFirst K) = K) : IsConvexClosedInSecond K

Lean source

Helper for Lemma33.0.43: the upper-closed iterated-closure identity already makes the kernel concave-closed in the first variable.

lemma helperForLemma33_0_43_firstConcaveClosed_of_upperClosedIdentity {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hUpper : concaveClosureInFirst (convexClosureInSecond K) = K) : IsConcaveClosedInFirst K

Lean source

Helper for Lemma33.0.43: on the concave-convex branch, the lower/upper iterated-closure identities recover the coordinatewise closedness data appearing in full closedness.

lemma helperForLemma33_0_43_concaveConvexBranch_recovers_fully_closed_data {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hShape : IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K) (hLower : convexClosureInSecond (concaveClosureInFirst K) = K) (hUpper : concaveClosureInFirst (convexClosureInSecond K) = K) : IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ IsConcaveClosedInFirst K ∧ IsConvexClosedInSecond K

Lean source

Helper for Lemma33.0.43: the convex-concave lower-closed identity can be rewritten as the concave-convex lower-closed identity for the swapped kernel.

lemma helperForLemma33_0_43_swappedLowerClosedIdentity_of_convexConcaveLowerClosedIdentity {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hLower : convexClosureInFirst (concaveClosureInSecond K) = K) : convexClosureInSecond (concaveClosureInFirst (fun v u => K u v)) = (fun v u => K u v)

Lean source

Helper for Lemma33.0.43: the convex-concave upper-closed identity can be rewritten as the concave-convex upper-closed identity for the swapped kernel.

lemma helperForLemma33_0_43_swappedUpperClosedIdentity_of_convexConcaveUpperClosedIdentity {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hUpper : concaveClosureInSecond (convexClosureInFirst K) = K) : concaveClosureInFirst (convexClosureInSecond (fun v u => K u v)) = (fun v u => K u v)

Lean source

Helper for Lemma33.0.43: on the convex-concave branch, the lower/upper iterated-closure identities recover the coordinatewise closedness data appearing in full closedness.

lemma helperForLemma33_0_43_convexConcaveBranch_recovers_fully_closed_data {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hShape : IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K) (hLower : convexClosureInFirst (concaveClosureInSecond K) = K) (hUpper : concaveClosureInSecond (convexClosureInFirst K) = K) : IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ IsConvexClosedInFirst K ∧ IsConcaveClosedInSecond K

Lean source

Helper for Lemma33.0.43: on the concave-convex branch, the raw fully-closed data is equivalent to the aligned lower/upper iterated-closure data.

lemma helperForLemma33_0_43_concaveConvexBranch_rawFullyClosed_iff_rawLower_and_rawUpper_data {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} : (IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ IsConcaveClosedInFirst K ∧ IsConvexClosedInSecond K) ↔ ((IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ convexClosureInSecond (concaveClosureInFirst K) = K) ∧ (IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ concaveClosureInFirst (convexClosureInSecond K) = K))

Lean source

Helper for Lemma33.0.43: on the convex-concave branch, the raw fully-closed data is equivalent to the aligned lower/upper iterated-closure data.

lemma helperForLemma33_0_43_convexConcaveBranch_rawFullyClosed_iff_rawLower_and_rawUpper_data {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} : (IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ IsConvexClosedInFirst K ∧ IsConcaveClosedInSecond K) ↔ ((IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ convexClosureInFirst (concaveClosureInSecond K) = K) ∧ (IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ concaveClosureInSecond (convexClosureInFirst K) = K))

Lean source

Helper for Lemma33.0.43: before the later reducible lower-closed and upper-closed predicates are introduced, full closedness is already equivalent to having both raw iterated-closure identities on one aligned saddle orientation.

lemma helperForLemma33_0_43_rawFullyClosed_iff_aligned_rawLowerClosed_and_rawUpperClosed_data {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} : ((IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ IsConcaveClosedInFirst K ∧ IsConvexClosedInSecond K) ∨ (IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ IsConvexClosedInFirst K ∧ IsConcaveClosedInSecond K)) ↔ (((IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ convexClosureInSecond (concaveClosureInFirst K) = K) ∧ (IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ concaveClosureInFirst (convexClosureInSecond K) = K)) ∨ ((IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ convexClosureInFirst (concaveClosureInSecond K) = K) ∧ (IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ concaveClosureInSecond (convexClosureInFirst K) = K)))

§33 — Saddle-Functions — Part 11

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Helper for Lemma33.0.35: evaluating the raw translated-and-tilted bifunction at the zero primal displacement recovers the tilted primal objective from the textbook statement.

lemma helperForLemma33_0_35_rawTranslatedTiltedZeroSection_eq_tiltedObjective {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) : (fun y : Fin n → ℝ => (fun v : Fin m → ℝ => fun y : Fin n → ℝ => F (u + v) y - ((dotProduct y xStar : ℝ) : EReal)) (0 : Fin m → ℝ) y) = fun y : Fin n → ℝ => F u y - ((dotProduct y xStar : ℝ) : EReal)

Lean source

Helper for Lemma33.0.35: the value set of the zero section of the raw translated-and-tilted bifunction is exactly the tilted primal objective range appearing in the textbook formula.

lemma helperForLemma33_0_35_zeroSectionRange_eq_tiltedObjectiveRange {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) : Set.range (fun y : Fin n → ℝ => (fun v : Fin m → ℝ => fun y : Fin n → ℝ => F (u + v) y - ((dotProduct y xStar : ℝ) : EReal)) (0 : Fin m → ℝ) y) = Set.range (fun y : Fin n → ℝ => F u y - ((dotProduct y xStar : ℝ) : EReal))

Lean source

Helper for Lemma33.0.35: the infimum over the zero section of the raw translated-and-tilted bifunction is the infimum over the tilted primal objective itself.

lemma helperForLemma33_0_35_zeroSection_sInf_eq_tiltedObjective_sInf {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) : sInf (Set.range (fun y : Fin n → ℝ => (fun v : Fin m → ℝ => fun y : Fin n → ℝ => F (u + v) y - ((dotProduct y xStar : ℝ) : EReal)) (0 : Fin m → ℝ) y)) = sInf (Set.range (fun y : Fin n → ℝ => F u y - ((dotProduct y xStar : ℝ) : EReal)))

Lean source

Helper for Lemma33.0.36: if two statements are each equivalent to the same criterion, then they are equivalent to each other.

lemma helperForLemma33_0_36_equivalent_of_sharedCriterion {P Q R : Prop} (hPR : P ↔ R) (hQR : Q ↔ R) : P ↔ Q

Lean source

Helper for Lemma33.0.36: pointwise equivalence to a shared criterion identifies the two binary predicates themselves.

lemma helperForLemma33_0_36_predicateEquality_of_sharedCriterion {α : Sort*} {β : Sort*} {P Q R : α → β → Prop} (hPR : ∀ a b, P a b ↔ R a b) (hQR : ∀ a b, Q a b ↔ R a b) : P = Q

Lean source

Helper for Lemma33.0.36: pointwise equivalences to a shared criterion compose into a pointwise equivalence.

lemma helperForLemma33_0_36_pointwiseEquivalent_of_sharedCriterion {α : Sort*} {β : Sort*} {P Q R : α → β → Prop} (hPR : ∀ a b, P a b ↔ R a b) (hQR : ∀ a b, Q a b ↔ R a b) : ∀ a b, P a b ↔ Q a b

§33 — Saddle-Functions — Part 12

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Helper for Lemma33.0.34: translating the primal variable by u extracts the constant correction term - ⟪u, vStar⟫ from the dual pairing.

lemma helperForLemma33_0_34_dotProduct_translation_split {m : ℕ} {u w vStar : Fin m → ℝ} : (((dotProduct (w - u) vStar : ℝ) : EReal)) = (((dotProduct w vStar : ℝ) : EReal)) - (((dotProduct u vStar : ℝ) : EReal))

Lean source

Helper for Lemma33.0.34: translating the first variable does not change the range of a two-variable integrand, because the substitution w = u + v is bijective.

lemma helperForLemma33_0_34_translate_first_range_eq {m n : ℕ} (G : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (u : Fin m → ℝ) : Set.range (fun vy : (Fin m → ℝ) × (Fin n → ℝ) => G (u + vy.1) vy.2) = Set.range (fun wy : (Fin m → ℝ) × (Fin n → ℝ) => G wy.1 wy.2)

Lean source

Helper for Lemma33.0.34: after translating w = u + v, the translated-tilted adjoint integrand is the original adjoint integrand for F at xStar + yStar, shifted by the constant - ⟪u, vStar⟫.

lemma helperForLemma33_0_34_translatedTilted_integrand_rewrite {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u v vStar : Fin m → ℝ} {xStar yStar y : Fin n → ℝ} : ((F (u + v) y - ((dotProduct y xStar : ℝ) : EReal)) - ((dotProduct y yStar : ℝ) : EReal)) + (((dotProduct v vStar : ℝ) : EReal)) = (F (u + v) y - ((dotProduct y (xStar + yStar) : ℝ) : EReal) + (((dotProduct (u + v) vStar : ℝ) : EReal))) - (((dotProduct u vStar : ℝ) : EReal))

Lean source

Helper for Lemma33.0.34: the whole range defining the translated-tilted adjoint equals the range for the untranslated integrand at xStar + yStar, shifted by the constant - ⟪u, vStar⟫.

lemma helperForLemma33_0_34_translatedTilted_adjointRange_eq {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u vStar : Fin m → ℝ} {xStar yStar : Fin n → ℝ} : Set.range (fun vy : (Fin m → ℝ) × (Fin n → ℝ) => ((F (u + vy.1) vy.2 - ((dotProduct vy.2 xStar : ℝ) : EReal)) - ((dotProduct vy.2 yStar : ℝ) : EReal)) + (((dotProduct vy.1 vStar : ℝ) : EReal))) = Set.range (fun wy : (Fin m → ℝ) × (Fin n → ℝ) => (F wy.1 wy.2 - ((dotProduct wy.2 (xStar + yStar) : ℝ) : EReal) + (((dotProduct wy.1 vStar : ℝ) : EReal))) - (((dotProduct u vStar : ℝ) : EReal)))

Lean source

Helper for Lemma33.0.34: subtracting the constant ⟪u, vStar⟫ from the untranslated adjoint integrand is exactly the image of its range under the affine map `z ↦ z - ⟪u, vStar⟫`.

lemma helperForLemma33_0_34_untranslated_adjointRange_sub_eq_image {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u vStar : Fin m → ℝ} {xStar yStar : Fin n → ℝ} : Set.range (fun wy : (Fin m → ℝ) × (Fin n → ℝ) => (F wy.1 wy.2 - ((dotProduct wy.2 (xStar + yStar) : ℝ) : EReal) + (((dotProduct wy.1 vStar : ℝ) : EReal))) - (((dotProduct u vStar : ℝ) : EReal))) = (fun z : EReal => z - (((dotProduct u vStar : ℝ) : EReal))) '' Set.range (fun wy : (Fin m → ℝ) × (Fin n → ℝ) => F wy.1 wy.2 - ((dotProduct wy.2 (xStar + yStar) : ℝ) : EReal) + (((dotProduct wy.1 vStar : ℝ) : EReal)))

Lean source

Helper for Lemma33.0.34: the translated-tilted adjoint range is the image of the untranslated adjoint range under subtraction by the constant ⟪u, vStar⟫.

lemma helperForLemma33_0_34_translatedTilted_adjointRange_eq_imageSub {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u vStar : Fin m → ℝ} {xStar yStar : Fin n → ℝ} : Set.range (fun vy : (Fin m → ℝ) × (Fin n → ℝ) => ((F (u + vy.1) vy.2 - ((dotProduct vy.2 xStar : ℝ) : EReal)) - ((dotProduct vy.2 yStar : ℝ) : EReal)) + (((dotProduct vy.1 vStar : ℝ) : EReal))) = (fun z : EReal => z - (((dotProduct u vStar : ℝ) : EReal))) '' Set.range (fun wy : (Fin m → ℝ) × (Fin n → ℝ) => F wy.1 wy.2 - ((dotProduct wy.2 (xStar + yStar) : ℝ) : EReal) + (((dotProduct wy.1 vStar : ℝ) : EReal)))

Lean source

Helper for Lemma33.0.34: after the w = u + v substitution, the infimum defining the translated-tilted adjoint is the infimum for F at xStar + yStar, minus the constant ⟪u, vStar⟫.

lemma helperForLemma33_0_34_translatedTilted_adjointRange_sInf_eq {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u vStar : Fin m → ℝ} {xStar yStar : Fin n → ℝ} : sInf (Set.range (fun vy : (Fin m → ℝ) × (Fin n → ℝ) => ((F (u + vy.1) vy.2 - ((dotProduct vy.2 xStar : ℝ) : EReal)) - ((dotProduct vy.2 yStar : ℝ) : EReal)) + (((dotProduct vy.1 vStar : ℝ) : EReal)))) = sInf (Set.range (fun wy : (Fin m → ℝ) × (Fin n → ℝ) => F wy.1 wy.2 - ((dotProduct wy.2 (xStar + yStar) : ℝ) : EReal) + (((dotProduct wy.1 vStar : ℝ) : EReal)))) - (((dotProduct u vStar : ℝ) : EReal))

Lean source

Helper for Lemma33.0.34: the textbook translated-and-tilted bifunction H(v, y) = F(u + v, y) - ⟪y, xStar⟫, recorded locally in the dependency-closed file that stages the adjoint calculation.

noncomputable abbrev helperForLemma33_0_34_translatedTiltedBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (u : Fin m → ℝ) (xStar : Fin n → ℝ) : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun v y => F (u + v) y - ((dotProduct y xStar : ℝ) : EReal)

Lean source

Helper for Lemma33.0.34: the raw two-variable infimum formula that later becomes genuineConvexBifunctionAdjoint in the downstream file.

noncomputable abbrev helperForLemma33_0_34_rawGenuineAdjoint {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (xStar : Fin n → ℝ) (uStar : Fin m → ℝ) : EReal := sInf <| Set.range fun ux : (Fin m → ℝ) × (Fin n → ℝ) => (F ux.1 ux.2 - ((dotProduct ux.2 xStar : ℝ) : EReal)) + ((dotProduct ux.1 uStar : ℝ) : EReal)

Lean source

Helper for Lemma33.0.34: the raw adjoint of the textbook bifunction `H(v, y) = F(u + v, y) - ⟪y, xStar⟫is the raw adjoint ofFatxStar + yStar`, shifted by -⟪u, vStar⟫.

lemma helperForLemma33_0_34_rawGenuineAdjoint_translatedTiltedBifunction {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u vStar : Fin m → ℝ} {xStar yStar : Fin n → ℝ} : helperForLemma33_0_34_rawGenuineAdjoint (helperForLemma33_0_34_translatedTiltedBifunction F u xStar) yStar vStar = helperForLemma33_0_34_rawGenuineAdjoint F (xStar + yStar) vStar - (((dotProduct u vStar : ℝ) : EReal))

Lean source

Helper for Lemma33.0.34: after expanding the translated-tilted integrand explicitly, the left-hand side of the downstream adjoint computation is already the raw adjoint formula proved above.

lemma helperForLemma33_0_34_explicitTranslatedTiltedAdjoint_eq_rawTranslated {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u vStar : Fin m → ℝ} {xStar yStar : Fin n → ℝ} : sInf (Set.range (fun ux : (Fin m → ℝ) × (Fin n → ℝ) => ((F (u + ux.1) ux.2 - ((dotProduct ux.2 xStar : ℝ) : EReal)) - ((dotProduct ux.2 yStar : ℝ) : EReal)) + (((dotProduct ux.1 vStar : ℝ) : EReal)))) = helperForLemma33_0_34_rawGenuineAdjoint (helperForLemma33_0_34_translatedTiltedBifunction F u xStar) yStar vStar

Lean source

Helper for Lemma33.0.34: after expanding the translated-tilted integrand explicitly, the left-hand side of the downstream adjoint computation is already the raw adjoint formula proved above.

lemma helperForLemma33_0_34_explicitTranslatedTiltedAdjoint_eq_rawShift {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u vStar : Fin m → ℝ} {xStar yStar : Fin n → ℝ} : sInf (Set.range (fun ux : (Fin m → ℝ) × (Fin n → ℝ) => ((F (u + ux.1) ux.2 - ((dotProduct ux.2 xStar : ℝ) : EReal)) - ((dotProduct ux.2 yStar : ℝ) : EReal)) + (((dotProduct ux.1 vStar : ℝ) : EReal)))) = helperForLemma33_0_34_rawGenuineAdjoint F (xStar + yStar) vStar - (((dotProduct u vStar : ℝ) : EReal))

§33 — Saddle-Functions — Part 13

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Lemma33.0.32 (Optimal value equality at 0 and pairing equality): the primal value is the negative of the zero pairing, the dual value is the negative of the zero concave conjugate, and equality of the two values at the origin is equivalent to equality of those pairings.

theorem optimalValueEqualityAtZero_and_pairingEquality {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (∀ u : Fin m → ℝ, convexProgramAssociatedWith F u = -convexBifunctionPairing F u (0 : Fin n → ℝ)) ∧ (∀ xStar : Fin n → ℝ, sSup (Set.range fun uStar : Fin m → ℝ => convexBifunctionAdjoint F xStar uStar) = -concaveConjugate (convexBifunctionAdjoint F xStar) (0 : Fin m → ℝ)) ∧ (convexProgramAssociatedWith F (0 : Fin m → ℝ) = sSup (Set.range fun uStar : Fin m → ℝ => convexBifunctionAdjoint F (0 : Fin n → ℝ) uStar) ↔ convexBifunctionPairing F (0 : Fin m → ℝ) (0 : Fin n → ℝ) = concaveConjugate (convexBifunctionAdjoint F (0 : Fin n → ℝ)) (0 : Fin m → ℝ))

Lean source

Helper for Lemma33.0.32: evaluating the Section 33 convex pairing at the zero dual vector recovers the negative of the frozen primal-section infimum.

lemma helperForLemma33_0_32_pairingAtZero_eq_neg_sectionInfimum {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) : convexBifunctionPairing F u (0 : Fin n → ℝ) = -(iInf fun x : Fin n → ℝ => F u x)

Lean source

Helper for Lemma33.0.32: negating the zero-pairing identity gives the primal-section infimum as the negative of the zero pairing.

lemma helperForLemma33_0_32_neg_pairingAtZero_eq_sectionInfimum {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) : -convexBifunctionPairing F u (0 : Fin n → ℝ) = iInf fun x : Fin n → ℝ => F u x

Lean source

Helper for Lemma33.0.32: evaluating the dual Section 33 slice at the zero primal vector rewrites its supremum as the negative of the corresponding concave conjugate.

lemma helperForLemma33_0_32_dualSectionSupremum_eq_neg_concaveConjugateAtZero {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (xStar : Fin n → ℝ) : sSup (Set.range fun uStar : Fin m → ℝ => convexBifunctionAdjoint F xStar uStar) = -concaveConjugate (convexBifunctionAdjoint F xStar) (0 : Fin m → ℝ)

Lean source

Helper for Lemma33.0.32: the primal program value is the negative of the zero pairing with the frozen primal section.

lemma helperForLemma33_0_32_primalValue_eq_neg_pairingAtZero {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) : convexProgramAssociatedWith F u = -convexBifunctionPairing F u (0 : Fin n → ℝ)

Lean source

Helper for Lemma33.0.32: equality of the primal and dual values at the origin is equivalent to equality of the corresponding zero pairings.

lemma helperForLemma33_0_32_zeroValueEquality_iff_zeroPairingEquality {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} : convexProgramAssociatedWith F (0 : Fin m → ℝ) = sSup (Set.range fun uStar : Fin m → ℝ => convexBifunctionAdjoint F (0 : Fin n → ℝ) uStar) ↔ convexBifunctionPairing F (0 : Fin m → ℝ) (0 : Fin n → ℝ) = concaveConjugate (convexBifunctionAdjoint F (0 : Fin n → ℝ)) (0 : Fin m → ℝ)

§33 — Saddle-Functions — Part 14

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Lemma33.0.22 (Adjoint of a convex bifunction): if F has a proper closed convex graph function, with properness on the convex side modeled by the conditions that F takes no value ⊥ and that each section F u is nontrivial in the sense that it is not identically ⊤, then its adjoint x^* ↦ (u ↦ sup_x (⟪x, x^*⟫ - (F u) x)) is a closed concave bifunction. That is, for each x^*, the section u ↦ (F^* x^*)(u) is concave and concave-closed, and the graph function (u, x^*) ↦ (F^* x^*)(u) is concave in u and convex in x^*.

theorem convexBifunctionAdjoint_isClosedConcave {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_convex : IsGraphConvexBifunction F) (hF_closed : IsFunctionConvexClosed (graphFunctionOfBifunction F)) (hF_noBot : HasNoBotValuesBifunction F) (hF_notTop : ∀ u : Fin m → ℝ, ∃ x : Fin n → ℝ, F u x ≠ ⊤) : (∀ xStar : Fin n → ℝ, IsERealConcaveOn (Set.univ : Set (Fin m → ℝ)) (convexBifunctionAdjoint F xStar) ∧ IsFunctionConcaveClosed (convexBifunctionAdjoint F xStar)) ∧ IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (fun u xStar => convexBifunctionAdjoint F xStar u)

Lean source

Corollary33.0.23 (Concavity/closedness of ⟨u, F^* x^*⟩): under the same hypotheses as Lemma33.0.22, the function (u, x^*) ↦ (F^* x^*)(u), represented here by (u, x^*) ↦ convexBifunctionAdjoint F x^* u, is concave-convex, and for each fixed x^* the section u ↦ (F^* x^*)(u) is concave-closed.

theorem convexBifunctionAdjoint_pairing_isConcaveConvex_concaveClosed {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_convex : IsGraphConvexBifunction F) (hF_closed : IsFunctionConvexClosed (graphFunctionOfBifunction F)) (hF_noBot : HasNoBotValuesBifunction F) (hF_notTop : ∀ u : Fin m → ℝ, ∃ x : Fin n → ℝ, F u x ≠ ⊤) : IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (fun u xStar => convexBifunctionAdjoint F xStar u) ∧ ∀ xStar : Fin n → ℝ, IsFunctionConcaveClosed (convexBifunctionAdjoint F xStar)

Lean source

The pairing attached to the adjoint of a convex bifunction, i.e. the textbook quantity ⟨u, F^* x^*⟩ = (F^* x^*)(u).

noncomputable abbrev convexBifunctionAdjointPairing {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (xStar : Fin n → ℝ) (u : Fin m → ℝ) : EReal := concaveConjugate (convexBifunctionAdjoint F xStar) u

Lean source

The pairing attached to the adjoint of a concave bifunction, i.e. the textbook quantity ⟨u, F^* x^*⟩ = (F^* x^*)(u).

noncomputable abbrev concaveBifunctionAdjointPairing {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (xStar : Fin n → ℝ) (u : Fin m → ℝ) : EReal := convexConjugate (concaveBifunctionAdjoint F xStar) u

Lean source

The sectionwise convex regularization of a bifunction, obtained by closing each image function F u in the second variable. This is recorded separately from the book's global closure notation (cl F).

noncomputable def sectionwiseConvexClosureBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u x => functionConvexClosure (F u) x

Lean source

The sectionwise concave regularization of a bifunction, obtained by closing each image function F u in the second variable with the concave closure operator. This is recorded separately from the book's global closure notation (cl F).

noncomputable def sectionwiseConcaveClosureBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u x => functionConcaveClosure (F u) x

Lean source

Helper for Lemma33.0.22: second-variable convex-closedness of the sign-swapped adjoint kernel is exactly the family of one-variable convex-closedness statements for the tilted projection sections.

lemma helperForLemma33_0_22_swappedNegatedAdjointKernel_isConvexClosedInSecond_iff_tiltedProjectionSectionsClosed {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} : IsConvexClosedInSecond (fun xStar u => -convexBifunctionAdjoint F xStar u) ↔ ∀ xStar : Fin n → ℝ, IsFunctionConvexClosed (fun u : Fin m → ℝ => imageUnderLinearMap (projectionLinearMap (Nat.le_add_right m n)) (fun z : Fin (m + n) → ℝ => graphFunctionOfBifunction F z - ((dotProduct (fun j : Fin n => z (Fin.natAdd m j)) xStar : ℝ) : EReal)) u)

Lean source

Helper for Lemma33.0.22: the unresolved textbook closedness clause is exactly first-variable concave-closedness of the adjoint kernel.

lemma helperForLemma33_0_22_adjointKernel_isConcaveClosedInFirst_of_tiltedProjectionSectionsClosed {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hClosedSections : ∀ xStar : Fin n → ℝ, IsFunctionConvexClosed (fun u : Fin m → ℝ => imageUnderLinearMap (projectionLinearMap (Nat.le_add_right m n)) (fun z : Fin (m + n) → ℝ => graphFunctionOfBifunction F z - ((dotProduct (fun j : Fin n => z (Fin.natAdd m j)) xStar : ℝ) : EReal)) u)) : IsConcaveClosedInFirst (fun u xStar => convexBifunctionAdjoint F xStar u)

Lean source

Helper for Lemma33.0.22: the primal value of the tilted bifunction is exactly the negative of the fixed-dual pairing section. This is the textbook identity inf_x (F u x - ⟪x, x^*⟫) = -⟪F u, x^*⟫, written using the Chapter 6 primal-value notation for the tilted bifunction.

lemma helperForLemma33_0_22_tiltedPrimalValue_eq_negPairingSection {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (xStar : Fin n → ℝ) : convexProgramAssociatedWith (fun u : Fin m → ℝ => fun x : Fin n → ℝ => F u x - ((dotProduct x xStar : ℝ) : EReal)) = fun u : Fin m → ℝ => -convexBifunctionPairing F u xStar

Lean source

Helper for Lemma33.0.22: translating the parameter variable commutes with the tilted primal-value formula. For the translated tilted bifunction v ↦ F (u + v) - ⟪·, x^*⟫, the associated primal value function is the shifted negative pairing section v ↦ -⟪F (u + v), x^*⟫.

lemma helperForLemma33_0_22_translatedTiltedPrimalValue_eq_shiftedNegPairingSection {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) : convexProgramAssociatedWith (fun v : Fin m → ℝ => fun x : Fin n → ℝ => F (u + v) x - ((dotProduct x xStar : ℝ) : EReal)) = fun v : Fin m → ℝ => -convexBifunctionPairing F (u + v) xStar

Lean source

Helper for Lemma33.0.22: every translated tilted primal program is already consistent at the origin. The only consistency input needed later is that the origin slice of the translated tilted program avoids ⊤; this follows because the origin value is the negative of a pairing section that already avoids ⊥.

lemma helperForLemma33_0_22_translatedTiltedPrimalValue_atOrigin_ne_top {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_notTop : ∀ u : Fin m → ℝ, ∃ x : Fin n → ℝ, F u x ≠ ⊤) (u : Fin m → ℝ) (xStar : Fin n → ℝ) : convexProgramAssociatedWith (fun v : Fin m → ℝ => fun x : Fin n → ℝ => F (u + v) x - ((dotProduct x xStar : ℝ) : EReal)) 0 ≠ (⊤ : EReal)

Lean source

Helper for Lemma33.0.22: translating the tilted graph in the parameter variable produces a closed convex bifunction for the shifted primal-value program. This is the Chapter 6 packaging needed for the local normality argument: after freezing u and xStar, the translated objective (v, x) ↦ F (u + v) x - ⟪x, xStar⟫ still has a closed convex graph because the original tilted graph is closed convex and translation preserves both convexity and lower semicontinuity.

lemma helperForLemma33_0_22_translatedTiltedBifunction_isClosedConvex {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_convex : IsGraphConvexBifunction F) (hF_closed : IsFunctionConvexClosed (graphFunctionOfBifunction F)) (hF_noBot : HasNoBotValuesBifunction F) (hF_notTop : ∀ u : Fin m → ℝ, ∃ x : Fin n → ℝ, F u x ≠ ⊤) (u : Fin m → ℝ) (xStar : Fin n → ℝ) : ClosedConvexBifunction (fun v : Fin m → ℝ => fun x : Fin n → ℝ => F (u + v) x - ((dotProduct x xStar : ℝ) : EReal))

Lean source

Helper for Lemma33.0.22: each fixed-dual negative pairing section is lower semicontinuous. The proof follows the translated primal-value route from Chapter 6. Around any base point u, the shifted section v ↦ -⟪F (u + v), xStar⟫ is the primal-value function of a translated closed convex bifunction. Because that translated program is strictly consistent on all of ℝ^m, its value at the origin agrees with its convex closure there, and this pointwise closure equality upgrades to lower semicontinuity at the original base point u.

lemma helperForLemma33_0_22_negPairingSection_isLowerSemicontinuous {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_convex : IsGraphConvexBifunction F) (hF_closed : IsFunctionConvexClosed (graphFunctionOfBifunction F)) (hF_noBot : HasNoBotValuesBifunction F) (hF_notTop : ∀ u : Fin m → ℝ, ∃ x : Fin n → ℝ, F u x ≠ ⊤) : ∀ xStar : Fin n → ℝ, LowerSemicontinuous (fun u : Fin m → ℝ => -convexBifunctionPairing F u xStar)

Lean source

Helper for Lemma33.0.22: each fixed-dual pairing section is already concave-closed. After proving lower semicontinuity of the negated section, the Section 33 sign-flip lemma turns that convex-closed statement into the desired concave-closed one for u ↦ convexBifunctionPairing F u xStar.

lemma helperForLemma33_0_22_pairingSection_isFunctionConcaveClosed {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_convex : IsGraphConvexBifunction F) (hF_closed : IsFunctionConvexClosed (graphFunctionOfBifunction F)) (hF_noBot : HasNoBotValuesBifunction F) (hF_notTop : ∀ u : Fin m → ℝ, ∃ x : Fin n → ℝ, F u x ≠ ⊤) : ∀ xStar : Fin n → ℝ, IsFunctionConcaveClosed (fun u : Fin m → ℝ => convexBifunctionPairing F u xStar)

Lean source

Helper for Lemma33.0.22: the unresolved textbook closedness clause is exactly first-variable concave-closedness of the adjoint kernel.

lemma helperForLemma33_0_22_adjointKernel_isConcaveClosedInFirst {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_convex : IsGraphConvexBifunction F) (hF_closed : IsFunctionConvexClosed (graphFunctionOfBifunction F)) (hF_noBot : HasNoBotValuesBifunction F) (hF_notTop : ∀ u : Fin m → ℝ, ∃ x : Fin n → ℝ, F u x ≠ ⊤) : IsConcaveClosedInFirst (fun u xStar => convexBifunctionAdjoint F xStar u)

Lean source

Helper for Corollary33.0.23: Lemma33.0.22 already contains the global concave-convex statement for the adjoint pairing kernel.

lemma helperForCorollary33_0_23_jointConcaveConvex {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_convex : IsGraphConvexBifunction F) (hF_closed : IsFunctionConvexClosed (graphFunctionOfBifunction F)) (hF_noBot : HasNoBotValuesBifunction F) (hF_notTop : ∀ u : Fin m → ℝ, ∃ x : Fin n → ℝ, F u x ≠ ⊤) : IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (fun u xStar => convexBifunctionAdjoint F xStar u)

Lean source

Helper for Corollary33.0.23: Lemma33.0.22 already yields concave-closedness of each fixed-dual adjoint section.

lemma helperForCorollary33_0_23_sectionConcaveClosed {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_convex : IsGraphConvexBifunction F) (hF_closed : IsFunctionConvexClosed (graphFunctionOfBifunction F)) (hF_noBot : HasNoBotValuesBifunction F) (hF_notTop : ∀ u : Fin m → ℝ, ∃ x : Fin n → ℝ, F u x ≠ ⊤) : ∀ xStar : Fin n → ℝ, IsFunctionConcaveClosed (convexBifunctionAdjoint F xStar)

§33 — Saddle-Functions — Part 15

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

The graph-function notion of concavity for a bifunction.

def IsGraphConcaveBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := IsERealConcaveOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F)

Lean source

The canonical pairing attached to the graph adjoint of a convex bifunction. The current Section 33 adjoint pairing is the parameter-side conjugate of the raw pairing; the book's closure identity uses its second parameter-side conjugate.

noncomputable abbrev convexBifunctionCanonicalAdjointPairing {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (xStar : Fin n → ℝ) (u : Fin m → ℝ) : EReal := concaveConjugate (convexBifunctionAdjointPairing F xStar) u

Lean source

The canonical pairing attached to the graph adjoint of a concave bifunction.

noncomputable abbrev concaveBifunctionCanonicalAdjointPairing {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (xStar : Fin n → ℝ) (u : Fin m → ℝ) : EReal := convexConjugate (concaveBifunctionAdjointPairing F xStar) u

Lean source

The Section 33 convex adjoint is the Chapter 6 adjoint after the graph-convex hypothesis is repackaged as a ConvexBifunction.

lemma helperForTheorem33_2_convexAdjoint_eq_adjointOfConvex {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : ConvexBifunction F) (xStar : Fin n → ℝ) (uStar : Fin m → ℝ) : convexBifunctionAdjointPairing F xStar uStar = adjointOfConvexBifunction ⟨F, hF⟩ xStar uStar

Lean source

Conjugating a fixed-parameter concave pairing produces the Chapter 6 adjoint.

lemma helperForTheorem33_2_convexConjugate_concavePairing_eq_adjointOfConcave {m n : ℕ} {G : (Fin n → ℝ) → (Fin m → ℝ) → EReal} (hG : ConcaveBifunction G) (u : Fin m → ℝ) (x : Fin n → ℝ) : convexConjugate (fun xStar : Fin n → ℝ => concaveBifunctionPairing G xStar u) x = adjointOfConcaveBifunction ⟨G, hG⟩ u x

Lean source

A Jensen-convex extended-real function which takes ⊥ once is identically ⊥.

lemma helperForTheorem33_2_convex_eq_constBot_of_eq_bot {k : ℕ} {f : (Fin k → ℝ) → EReal} (hConv : IsERealConvexOn (Set.univ : Set (Fin k → ℝ)) f) {x0 : Fin k → ℝ} (hx0 : f x0 = ⊥) : f = fun _ => ⊥

Lean source

On Jensen-convex functions the raw Section 33 closure agrees with the Chapter 2 convex-function closure, including the improper constant-⊥ case.

lemma helperForTheorem33_2_functionConvexClosure_eq_convexFunctionClosure_of_convex {k : ℕ} {f : (Fin k → ℝ) → EReal} (hConv : IsERealConvexOn (Set.univ : Set (Fin k → ℝ)) f) : functionConvexClosure f = convexFunctionClosure f

Lean source

The convex pairing kernel of a graph-convex bifunction is a Chapter 2 concave function in the parameter variable, including improper sections.

lemma helperForTheorem33_2_convexPairingSection_concaveFunction {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConvexBifunction F) (xStar : Fin n → ℝ) : ConcaveFunction (fun u : Fin m → ℝ => convexBifunctionPairing F u xStar)

Lean source

Graph concavity and the concave extended-real convention turn sign negation into graph convexity.

lemma helperForTheorem33_2_negatedGraphConvex {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConcaveBifunction F) (hNoTop : HasNoTopValuesBifunction F) : IsGraphConvexBifunction (fun u x => -F u x)

Lean source

The concave pairing kernel of a graph-concave bifunction is a Chapter 2 convex function in the parameter variable.

lemma helperForTheorem33_2_concavePairingSection_convexFunction {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConcaveBifunction F) (hNoTop : HasNoTopValuesBifunction F) (xStar : Fin n → ℝ) : ConvexFunction (fun u : Fin m → ℝ => concaveBifunctionPairing F u xStar)

Lean source

The first coordinatewise-closure identity in the convex branch, with the book's convex-analytic closure convention for improper functions.

lemma helperForTheorem33_2_first_convex {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConvexBifunction F) : ∀ xStar u, convexBifunctionCanonicalAdjointPairing F xStar u = concaveClosure (fun u' => convexBifunctionPairing F u' xStar) u

Lean source

The first coordinatewise-closure identity in the concave branch.

lemma helperForTheorem33_2_first_concave {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConcaveBifunction F) (hNoTop : HasNoTopValuesBifunction F) : ∀ xStar u, concaveBifunctionCanonicalAdjointPairing F xStar u = convexFunctionClosure (fun u' => concaveBifunctionPairing F u' xStar) u

Lean source

The second convex identity when the graph function has a finite witness.

lemma helperForTheorem33_2_second_convex_of_finiteWitness {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) {u0 : Fin m → ℝ} {x0 : Fin n → ℝ} (hFinite : F u0 x0 ≠ ⊤) : ∀ u xStar, convexFunctionClosure (fun xStar' => convexBifunctionCanonicalAdjointPairing F xStar' u) xStar = convexBifunctionPairing (convexBifunctionClosure F) u xStar

Lean source

The all-⊤ exceptional branch of the second convex identity.

lemma helperForTheorem33_2_second_convex_of_allTop {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hAllTop : ∀ u x, F u x = ⊤) : ∀ u xStar, convexFunctionClosure (fun xStar' => convexBifunctionCanonicalAdjointPairing F xStar' u) xStar = convexBifunctionPairing (convexBifunctionClosure F) u xStar

Lean source

Every fixed-u corrected adjoint-pairing section in the convex branch is a convex function of xStar.

lemma helperForTheorem33_2_convexAdjointPairingSection_convexFunction {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (u : Fin m → ℝ) : ConvexFunction (fun xStar => convexBifunctionCanonicalAdjointPairing F xStar u)

Lean source

The second concave identity when the graph function has a finite witness.

lemma helperForTheorem33_2_second_concave_of_finiteWitness {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConcaveBifunction F) (hNoTop : HasNoTopValuesBifunction F) {u0 : Fin m → ℝ} {x0 : Fin n → ℝ} (hFinite : F u0 x0 ≠ ⊥) : ∀ u xStar, concaveClosure (fun xStar' => concaveBifunctionCanonicalAdjointPairing F xStar' u) xStar = concaveBifunctionPairing (concaveBifunctionClosure F) u xStar

Lean source

The all-⊥ exceptional branch of the second concave identity.

lemma helperForTheorem33_2_second_concave_of_allBot {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hAllBot : ∀ u x, F u x = ⊥) : ∀ u xStar, concaveClosure (fun xStar' => concaveBifunctionCanonicalAdjointPairing F xStar' u) xStar = concaveBifunctionPairing (concaveBifunctionClosure F) u xStar

Lean source

Theorem33.2: For any convex bifunction F satisfying Rockafellar's ℝ ∪ {+∞} convention (modeled here by HasNoBotValuesBifunction F), and for any concave bifunction F satisfying the dual ℝ ∪ {-∞} convention (modeled by HasNoTopValuesBifunction F), one has ⟨u, F^* x^*⟩ = cl_u ⟨F u, x^*⟩ and cl_{x^*} ⟨u, F^* x^*⟩ = ⟨(cl F) u, x^*⟩. Here ⟨u, F^* x^*⟩ is represented by convexBifunctionCanonicalAdjointPairing or concaveBifunctionCanonicalAdjointPairing, formed by taking the conjugate in the u-variable of the genuine adjoint section xStar ↦ F^* xStar; cl_u and cl_{x^*} are the coordinatewise concave/convex closure operators as appropriate to the convex or concave branch, and (cl F) is modeled by the closure of the graph function on ℝ^(m + n).

theorem adjoint_pairing_eq_coordinatewise_closures {m n : ℕ} : (∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, (IsGraphConvexBifunction F ∧ HasNoBotValuesBifunction F) → (∀ xStar u, convexBifunctionCanonicalAdjointPairing F xStar u = concaveClosure (fun u' => convexBifunctionPairing F u' xStar) u) ∧ (∀ u xStar, convexFunctionClosure (fun xStar' => convexBifunctionCanonicalAdjointPairing F xStar' u) xStar = convexBifunctionPairing (convexBifunctionClosure F) u xStar)) ∧ (∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, (IsGraphConcaveBifunction F ∧ HasNoTopValuesBifunction F) → (∀ xStar u, concaveBifunctionCanonicalAdjointPairing F xStar u = convexFunctionClosure (fun u' => concaveBifunctionPairing F u' xStar) u) ∧ (∀ u xStar, concaveClosure (fun xStar' => concaveBifunctionCanonicalAdjointPairing F xStar' u) xStar = concaveBifunctionPairing (concaveBifunctionClosure F) u xStar))

Lean source

On concave functions the raw Section 33 upper closure agrees with the canonical concave-function closure, including the improper constant-⊤ case.

lemma helperForTheorem33_2_functionConcaveClosure_eq_concaveClosure_of_concave {k : ℕ} {g : (Fin k → ℝ) → EReal} (hConc : ConcaveFunction g) (hNoTop : ∀ x, g x ≠ ⊤) : functionConcaveClosure g = concaveClosure g

Lean source

The concave effective domain of the corrected adjoint-pairing section is the set of adjoint parameters for which the genuine convex adjoint section is not identically ⊤.

lemma helperForCorollary33_2_1_concaveAdjointPairingSection_concaveEffectiveDomain_eq {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) : effectiveDomain (Set.univ : Set (Fin n → ℝ)) (fun xStar' => -concaveBifunctionCanonicalAdjointPairing F xStar' u) = {xStar' | ∃ u' : Fin m → ℝ, concaveBifunctionAdjointPairing F xStar' u' ≠ ⊤}

Lean source

A concave function agrees with its canonical concave closure on the relative interior of its concave effective domain.

lemma helperForCorollary33_2_1_concaveClosure_eq_self_on_ri_concaveEffectiveDomain_fin {m : ℕ} {g : (Fin m → ℝ) → EReal} (hConc : ConcaveFunction g) {u : Fin m → ℝ} (hu : u ∈ euclideanRelativeInterior_fin m (effectiveDomain (Set.univ : Set (Fin m → ℝ)) (fun v => -g v))) : concaveClosure g u = g u

Lean source

lemma helperForCorollary33_2_1_concaveAdjoint_eq_neg_convexAdjoint_of_neg {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (xStar : Fin n → ℝ) (u : Fin m → ℝ) : concaveBifunctionAdjointPairing F xStar u = -convexBifunctionAdjointPairing (fun u' x => -F u' x) (-xStar) (-u)

Lean source

lemma helperForCorollary33_2_1_concaveAdjointPairing_eq_neg_convexAdjointPairing_of_neg {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (xStar : Fin n → ℝ) (u : Fin m → ℝ) : concaveBifunctionCanonicalAdjointPairing F xStar u = -convexBifunctionCanonicalAdjointPairing (fun u x => -F u x) (-xStar) u

Lean source

lemma helperForCorollary33_2_1_neg_concaveAdjointPairingSection_eq_convexAdjointPairingSection_precomp_neg_of_neg {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) : (fun xStar' => -concaveBifunctionCanonicalAdjointPairing F xStar' u) = fun xStar' => convexBifunctionCanonicalAdjointPairing (fun u' x => -F u' x) (-xStar') u

Lean source

lemma helperForCorollary33_2_1_precompose_neg_isERealConvexOn {k : ℕ} {f : (Fin k → ℝ) → EReal} (hConv : IsERealConvexOn (Set.univ : Set (Fin k → ℝ)) f) : IsERealConvexOn (Set.univ : Set (Fin k → ℝ)) (fun x => f (-x))

Lean source

lemma helperForCorollary33_2_1_precompose_neg_isFunctionConvexClosed {k : ℕ} {f : (Fin k → ℝ) → EReal} (hClosed : IsFunctionConvexClosed f) : IsFunctionConvexClosed (fun x => f (-x))

Lean source

lemma helperForCorollary33_2_1_precompose_neg_isFunctionConcaveClosed {k : ℕ} {f : (Fin k → ℝ) → EReal} (hClosed : IsFunctionConcaveClosed f) : IsFunctionConcaveClosed (fun x => f (-x))

Lean source

For fixed u, the corrected adjoint pairing of a graph-concave bifunction is concave in xStar. The all-⊥ branch is handled explicitly because its adjoint is improper.

lemma helperForCorollary33_2_1_concaveAdjointPairingSection_concaveFunction {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConcaveBifunction F) (hNoTop : HasNoTopValuesBifunction F) (u : Fin m → ℝ) : ConcaveFunction (fun xStar => concaveBifunctionCanonicalAdjointPairing F xStar u)

Lean source

On ri (dom F⁺), the corrected concave adjoint-pairing section is fixed by the canonical concave closure appearing in Theorem 33.2.

lemma helperForCorollary33_2_1_concaveAdjointPairingSection_concaveClosure_eq_self_on_intrinsicInterior {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConcaveBifunction F) (hNoTop : HasNoTopValuesBifunction F) (u : Fin m → ℝ) {xStar : Fin n → ℝ} (hxStar : xStar ∈ intrinsicInterior ℝ {xStar' | ∃ u' : Fin m → ℝ, concaveBifunctionAdjointPairing F xStar' u' ≠ ⊤}) : concaveClosure (fun xStar' => concaveBifunctionCanonicalAdjointPairing F xStar' u) xStar = concaveBifunctionCanonicalAdjointPairing F xStar u

Lean source

Corollary33.2.1: Let F : ℝ^m → (ℝ^n → EReal) be a convex or concave bifunction. If u ∈ ri (dom F) in the parameter variable, modeled here by u ∈ intrinsicInterior ℝ {u' | ∃ x, F u' x < ⊤} on the convex side and by u ∈ intrinsicInterior ℝ {u' | ∃ x, F u' x ≠ ⊥} on the concave side, then the pairing identity ⟪F u, x^*⟫ = ⟪u, F^* x^*⟫ holds for every x^*. If F is closed, modeled by closedness of the graph function, and x^* ∈ ri (dom F^*), modeled by x^* lying in the intrinsic interior of the effective domain of the genuine adjoint bifunction xStar ↦ F^* xStar, then the same pairing identity holds for every u.

theorem adjoint_pairing_eq_on_relativeInterior_domains {m n : ℕ} : (∀ {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal}, (IsGraphConvexBifunction F ∧ HasNoBotValuesBifunction F) → (∀ ⦃u : Fin m → ℝ⦄, u ∈ intrinsicInterior ℝ {u' | ∃ x : Fin n → ℝ, F u' x < ⊤} → ∀ xStar : Fin n → ℝ, convexBifunctionPairing F u xStar = convexBifunctionCanonicalAdjointPairing F xStar u) ∧ (IsFunctionConvexClosed (graphFunctionOfBifunction F) → ∀ ⦃xStar : Fin n → ℝ⦄, xStar ∈ intrinsicInterior ℝ {xStar' | ∃ u : Fin m → ℝ, convexBifunctionAdjointPairing F xStar' u ≠ ⊥} → ∀ u : Fin m → ℝ, convexBifunctionPairing F u xStar = convexBifunctionCanonicalAdjointPairing F xStar u)) ∧ (∀ {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal}, (IsGraphConcaveBifunction F ∧ HasNoTopValuesBifunction F) → (∀ ⦃u : Fin m → ℝ⦄, u ∈ intrinsicInterior ℝ {u' | ∃ x : Fin n → ℝ, F u' x ≠ ⊥} → ∀ xStar : Fin n → ℝ, concaveBifunctionPairing F u xStar = concaveBifunctionCanonicalAdjointPairing F xStar u) ∧ (IsFunctionConcaveClosed (graphFunctionOfBifunction F) → ∀ ⦃xStar : Fin n → ℝ⦄, xStar ∈ intrinsicInterior ℝ {xStar' | ∃ u : Fin m → ℝ, concaveBifunctionAdjointPairing F xStar' u ≠ ⊤} → ∀ u : Fin m → ℝ, concaveBifunctionPairing F u xStar = concaveBifunctionCanonicalAdjointPairing F xStar u))

Lean source

Helper for Corollary33.2.1: sign-flipping turns the Section 33 concave closure into the Section 33 convex closure of the negated function. This is the pointwise one-variable identity needed when the convex-side relative-interior route is eventually repaired via the Chapter 2 convex-closure theorem for u ↦ -⟪F u, x^*⟫.

lemma helperForCorollary33_2_1_functionConcaveClosure_eq_neg_functionConvexClosure_neg {m : ℕ} {g : (Fin m → ℝ) → EReal} : functionConcaveClosure g = fun u => -functionConvexClosure (fun u' => -g u') u

Lean source

Helper for Corollary33.2.1: graph-function closedness upgrades a graph-convex bifunction to a closed convex bifunction.

lemma helperForCorollary33_2_1_closedConvexBifunction_of_graphFunctionClosed {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hGraphClosed : IsFunctionConvexClosed (graphFunctionOfBifunction F)) : ClosedConvexBifunction F

Lean source

Helper for Corollary33.2.1: graph-function convex-closedness also forces closedness of every slice obtained by freezing the second variable. This is the exact symmetric counterpart to helperForLemma33_0_22_section_isFunctionConvexClosed: once the graph function is lower semicontinuous on ℝ^(m+n), composing it with u ↦ Fin.append u v preserves lower semicontinuity and therefore fixes the resulting slice by the one-variable raw convex closure operator.

lemma helperForCorollary33_2_1_frozenSecondVariable_isFunctionConvexClosed {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_closed : IsFunctionConvexClosed (graphFunctionOfBifunction F)) : ∀ v : Fin n → ℝ, IsFunctionConvexClosed (fun u : Fin m → ℝ => F u v)

Lean source

Helper for Corollary33.2.1: once graph-function closedness identifies F with its Chapter 6 graph closure, the convex pairing of the closure is the original convex pairing.

lemma helperForCorollary33_2_1_convexClosure_pairing_eq_self_of_closed {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hGraphClosed : IsFunctionConvexClosed (graphFunctionOfBifunction F)) (u : Fin m → ℝ) (xStar : Fin n → ℝ) : convexBifunctionPairing (convexBifunctionClosure F) u xStar = convexBifunctionPairing F u xStar

Lean source

Helper for Corollary33.2.1: the effective domain of the fixed-u adjoint-pairing section is exactly the textbook adjoint domain {x^* | ∃ u', (F^* x^*)(u') ≠ ⊥}.

lemma helperForCorollary33_2_1_adjointPairingSection_effectiveDomain_eq {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) : effectiveDomain (Set.univ : Set (Fin n → ℝ)) (fun xStar' => convexBifunctionCanonicalAdjointPairing F xStar' u) = {xStar' | ∃ u' : Fin m → ℝ, convexBifunctionAdjointPairing F xStar' u' ≠ ⊥}

Lean source

Helper for Corollary33.2.1: a convex function agrees with the Section 33 raw convex closure on the relative interior of its effective domain.

lemma helperForCorollary33_2_1_convexFunctionClosure_eq_self_on_ri_effectiveDomain_fin {m : ℕ} {f : (Fin m → ℝ) → EReal} (hConv : IsERealConvexOn (Set.univ : Set (Fin m → ℝ)) f) {u : Fin m → ℝ} (hu : u ∈ euclideanRelativeInterior_fin m (effectiveDomain (Set.univ : Set (Fin m → ℝ)) f)) : convexFunctionClosure f u = f u

Lean source

Helper for Corollary33.2.1: the same relative-interior closure identity, packaged directly for a global convex function on Fin m → ℝ.

lemma helperForCorollary33_2_1_convexFunctionClosure_eq_self_on_ri_effectiveDomain_fin_of_ConvexFunction {m : ℕ} {f : (Fin m → ℝ) → EReal} (hConvFun : ConvexFunction f) {u : Fin m → ℝ} (hu : u ∈ euclideanRelativeInterior_fin m (effectiveDomain (Set.univ : Set (Fin m → ℝ)) f)) : convexFunctionClosure f u = f u

Lean source

Helper for Corollary33.2.1: an improper convex function is identically ⊥ on the relative interior of its effective domain, expressed in Fin m → ℝ coordinates.

lemma helperForCorollary33_2_1_improperConvexFunctionOn_eq_bot_on_ri_effectiveDomain_fin {m : ℕ} {f : (Fin m → ℝ) → EReal} (himproper : ImproperConvexFunctionOn (Set.univ : Set (Fin m → ℝ)) f) {u : Fin m → ℝ} (hu : u ∈ euclideanRelativeInterior_fin m (effectiveDomain (Set.univ : Set (Fin m → ℝ)) f)) : f u = (⊥ : EReal)

Lean source

Helper for Corollary33.2.1: a convex function agrees with the Section 33 raw convex closure on the relative interior of its effective domain.

lemma helperForCorollary33_2_1_functionConvexClosure_eq_self_of_convex_on_ri_effectiveDomain {m : ℕ} {f : (Fin m → ℝ) → EReal} (hConv : IsERealConvexOn (Set.univ : Set (Fin m → ℝ)) f) {u : Fin m → ℝ} (hu : u ∈ euclideanRelativeInterior_fin m (effectiveDomain (Set.univ : Set (Fin m → ℝ)) f)) : functionConvexClosure f u = f u

Lean source

Helper for Corollary33.2.1: the raw Section 33 convex closure agrees with any global convex function on ri (dom f), packaged directly from ConvexFunction.

lemma helperForCorollary33_2_1_functionConvexClosure_eq_self_of_ConvexFunction_on_ri_effectiveDomain {m : ℕ} {f : (Fin m → ℝ) → EReal} (hConvFun : ConvexFunction f) {u : Fin m → ℝ} (hu : u ∈ euclideanRelativeInterior_fin m (effectiveDomain (Set.univ : Set (Fin m → ℝ)) f)) : functionConvexClosure f u = f u

Lean source

Helper for Corollary33.2.1: negating the convex pairing section produces a convex one-variable function.

lemma helperForCorollary33_2_1_negConvexPairingSection_isERealConvexOn {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hRock : IsRockafellarConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (xStar : Fin n → ℝ) (hPairNoTop : ∀ u' : Fin m → ℝ, convexBifunctionPairing F u' xStar ≠ (⊤ : EReal)) : IsERealConvexOn (Set.univ : Set (Fin m → ℝ)) (fun u' => -convexBifunctionPairing F u' xStar)

Lean source

Helper for Corollary33.2.1: even before proving the raw Jensen statement for the negated pairing section, the same section is already a global convex function in the Chapter 6 epigraph sense. This is the exact package needed by the relative-interior closure theorem.

lemma helperForCorollary33_2_1_negConvexPairingSection_isConvexFunction {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hRock : IsRockafellarConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (xStar : Fin n → ℝ) : ConvexFunction (fun u' => -convexBifunctionPairing F u' xStar)

Lean source

Helper for Corollary33.2.1: the effective domain of the negated convex pairing section is exactly the textbook domain {u | ∃ x, F u x < ⊤}.

lemma helperForCorollary33_2_1_negConvexPairingSection_effectiveDomain_eq {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (xStar : Fin n → ℝ) : effectiveDomain (Set.univ : Set (Fin m → ℝ)) (fun u' => -convexBifunctionPairing F u' xStar) = {u' | ∃ x : Fin n → ℝ, F u' x < ⊤}

Lean source

Helper for Corollary33.2.1: the effective domain of the concave pairing section is exactly the textbook domain {u | ∃ x, F u x ≠ ⊥}.

lemma helperForCorollary33_2_1_concavePairingSection_effectiveDomain_eq {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (xStar : Fin n → ℝ) : effectiveDomain (Set.univ : Set (Fin m → ℝ)) (fun u' => concaveBifunctionPairing F u' xStar) = {u' | ∃ x : Fin n → ℝ, F u' x ≠ ⊥}

Lean source

Helper for Corollary33.2.1: the negated convex pairing section already agrees with its raw Section 33 convex closure at points of ri (dom F) in the parameter variable.

lemma helperForCorollary33_2_1_negConvexPairingSection_functionConvexClosure_eq_self_on_intrinsicInterior {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) {u : Fin m → ℝ} (hu : u ∈ intrinsicInterior ℝ {u' | ∃ x : Fin n → ℝ, F u' x < ⊤}) (xStar : Fin n → ℝ) : convexFunctionClosure (fun u' => -convexBifunctionPairing F u' xStar) u = -convexBifunctionPairing F u xStar

Lean source

Helper for Corollary33.2.1: the fixed-u adjoint-pairing section already agrees with its raw Section 33 convex closure on ri (dom F^*).

lemma helperForCorollary33_2_1_adjointPairingSection_functionConvexClosure_eq_self_on_intrinsicInterior {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hGraphClosed : IsFunctionConvexClosed (graphFunctionOfBifunction F)) (u : Fin m → ℝ) {xStar : Fin n → ℝ} (hxStar : xStar ∈ intrinsicInterior ℝ {xStar' | ∃ u' : Fin m → ℝ, convexBifunctionAdjointPairing F xStar' u' ≠ ⊥}) : convexFunctionClosure (fun xStar' => convexBifunctionCanonicalAdjointPairing F xStar' u) xStar = convexBifunctionCanonicalAdjointPairing F xStar u

Lean source

Helper for Corollary33.2.1: on the convex side, the frozen pairing section is already fixed by the Section 33 concave closure on the intrinsic interior of its effective domain.

lemma helperForCorollary33_2_1_convexPairingSection_closure_eq_self_on_intrinsicInterior {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) {u : Fin m → ℝ} (hu : u ∈ intrinsicInterior ℝ {u' | ∃ x : Fin n → ℝ, F u' x < ⊤}) (xStar : Fin n → ℝ) : concaveClosure (fun u' => convexBifunctionPairing F u' xStar) u = convexBifunctionPairing F u xStar

Lean source

Helper for Corollary33.2.1: on the concave side, the frozen pairing section is already fixed by the Section 33 convex closure on the intrinsic interior of its effective domain.

lemma helperForCorollary33_2_1_concavePairingSection_closure_eq_self_on_intrinsicInterior {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConcaveBifunction F) (hNoTop : HasNoTopValuesBifunction F) {u : Fin m → ℝ} (hu : u ∈ intrinsicInterior ℝ {u' | ∃ x : Fin n → ℝ, F u' x ≠ ⊥}) (xStar : Fin n → ℝ) : convexFunctionClosure (fun u' => concaveBifunctionPairing F u' xStar) u = concaveBifunctionPairing F u xStar

Lean source

Helper for Corollary33.2.1: negating F converts the concave pairing into the negative of the corresponding convex pairing.

lemma helperForCorollary33_2_1_concavePairing_eq_neg_convexPairing_of_neg {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) : concaveBifunctionPairing F u xStar = -convexBifunctionPairing (fun u x => -F u x) u (-xStar)

Lean source

Helper for Corollary33.2.1: negating F also converts the adjoint pairing into the negative of the corresponding convex adjoint pairing, with the expected sign flip in the parameter variable coming from the Chapter 6 conjugation formula.

lemma helperForCorollary33_2_1_concaveAdjointPairing_eq_adjointOfConcaveBifunction {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hConc : ConcaveBifunction F) (xStar : Fin n → ℝ) (u : Fin m → ℝ) : concaveBifunctionAdjointPairing F xStar u = adjointOfConcaveBifunction ⟨F, hConc⟩ xStar u

Lean source

Helper for Corollary33.2.1: graph-function concave-closedness upgrades a graph-concave concave bifunction to a Chapter 6 closed concave bifunction.

lemma helperForCorollary33_2_1_closedConcaveBifunction_of_graphFunctionClosed {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConcaveBifunction F) (hNoTop : HasNoTopValuesBifunction F) (hGraphClosed : IsFunctionConcaveClosed (graphFunctionOfBifunction F)) : ClosedConcaveBifunction F

Lean source

Helper for Corollary33.2.1: a closed concave bifunction with no ⊤ values already has its graph function fixed by the raw concave-closure operator.

lemma helperForCorollary33_2_1_graphFunction_isFunctionConcaveClosed_of_closedConcaveBifunction {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hClosed : ClosedConcaveBifunction F) (hNoTop : HasNoTopValuesBifunction F) : IsFunctionConcaveClosed (graphFunctionOfBifunction F)

Lean source

Helper for Corollary33.2.1: once graph-function concave-closedness identifies F with its Chapter 6 concave closure, the concave pairing of the closure is the original concave pairing.

lemma helperForCorollary33_2_1_concaveClosure_pairing_eq_self_of_closed {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConcaveBifunction F) (hNoTop : HasNoTopValuesBifunction F) (hGraphClosed : IsFunctionConcaveClosed (graphFunctionOfBifunction F)) (u : Fin m → ℝ) (xStar : Fin n → ℝ) : concaveBifunctionPairing (concaveBifunctionClosure F) u xStar = concaveBifunctionPairing F u xStar

Lean source

Helper for Corollary33.2.1: in the closed convex branch, the pairing identity holds on the intrinsic interior of the adjoint effective domain.

lemma helperForCorollary33_2_1_closedConvex_pairing_eq_on_intrinsicInteriorAdjointDomain {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hGraphClosed : IsFunctionConvexClosed (graphFunctionOfBifunction F)) {xStar : Fin n → ℝ} (hxStar : xStar ∈ intrinsicInterior ℝ {xStar' | ∃ u : Fin m → ℝ, convexBifunctionAdjointPairing F xStar' u ≠ ⊥}) : ∀ u : Fin m → ℝ, convexBifunctionPairing F u xStar = convexBifunctionCanonicalAdjointPairing F xStar u

Lean source

Helper for Corollary33.2.1: in the closed concave branch, the pairing identity follows by combining Theorem 33.2 with the Chapter 6 closed-adjoint slice theorem.

lemma helperForCorollary33_2_1_closedConcave_pairing_eq_on_intrinsicInteriorAdjointDomain {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConcaveBifunction F) (hNoTop : HasNoTopValuesBifunction F) (hGraphClosed : IsFunctionConcaveClosed (graphFunctionOfBifunction F)) {xStar : Fin n → ℝ} (hxStar : xStar ∈ intrinsicInterior ℝ {xStar' | ∃ u : Fin m → ℝ, concaveBifunctionAdjointPairing F xStar' u ≠ ⊤}) : ∀ u : Fin m → ℝ, concaveBifunctionPairing F u xStar = concaveBifunctionCanonicalAdjointPairing F xStar u

§33 — Saddle-Functions — Part 16

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

The parameter domain of a convex bifunction, viewed as the set of u for which the image section F u is not identically ⊤.

def convexBifunctionParameterDomain {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Set (Fin m → ℝ) := {u | ∃ x : Fin n → ℝ, F u x ≠ ⊤}

Lean source

The adjoint domain of a convex bifunction, viewed as the set of x^* for which the adjoint section F^* x^* is not identically ⊥.

def convexBifunctionAdjointDomain {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Set (Fin n → ℝ) := {xStar | ∃ u : Fin m → ℝ, convexBifunctionAdjointPairing F xStar u ≠ ⊥}

Lean source

A proper polyhedral convex bifunction is a polyhedral convex bifunction whose graph function is proper convex on the whole space.

def IsProperPolyhedralConvexBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := IsRockafellarPolyhedralConvexBifunction F ∧ ProperConvexFunctionOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F)

Lean source

The effective parameter domain of a concave bifunction, viewed as the set of u for which the image section F u is not identically -∞.

def concaveBifunctionParameterDomain {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Set (Fin m → ℝ) := {u | ∃ x : Fin n → ℝ, F u x ≠ ⊥}

Lean source

The effective domain of the genuine convex adjoint of a concave bifunction, viewed as the set of x^* for which F^* x^* is not identically +∞.

def concaveBifunctionAdjointDomain {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Set (Fin n → ℝ) := {xStar | ∃ u : Fin m → ℝ, concaveBifunctionAdjointPairing F xStar u ≠ ⊤}

Lean source

A proper polyhedral concave bifunction is a polyhedral concave bifunction in the sense that its negated graph function is proper polyhedral convex on the whole space.

def IsProperPolyhedralConcaveBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := IsRockafellarConcaveBifunction F ∧ IsPolyhedralConvexFunction (m + n) (fun z => -graphFunctionOfBifunction F z) ∧ ProperConvexFunctionOn (Set.univ : Set (Fin (m + n) → ℝ)) (fun z => -graphFunctionOfBifunction F z)

Lean source

lemma helperForCorollary33_2_2_concaveAdjointDomain_eq_negated_convexAdjointTopDomain {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} : concaveBifunctionAdjointDomain F = {xStar | ∃ u : Fin m → ℝ, convexBifunctionAdjointPairing (fun u' x => -F u' x) (-xStar) u ≠ ⊥}

Lean source

A polyhedral convex function agrees with its canonical closure at every point of its effective domain. In the improper case lower semicontinuity forces each non-⊤ value to be ⊥, exactly matching the canonical improper closure.

lemma helperForCorollary33_2_2_convexClosure_eq_self_at_effectivePoint {k : ℕ} {f : (Fin k → ℝ) → EReal} (hpoly : IsPolyhedralConvexFunction k f) {x : Fin k → ℝ} (hx : f x < ⊤) : convexFunctionClosure f x = f x

Lean source

Proper polyhedral concave bifunctions have no ⊤ values.

lemma helperForCorollary33_2_2_concave_hasNoTopValues {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : IsProperPolyhedralConcaveBifunction F) : HasNoTopValuesBifunction F

Lean source

Outside the convex parameter domain the entire primal section is ⊤.

lemma helperForCorollary33_2_2_allTop_of_off_convexParameterDomain {m n : ℕ} {G : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} (hOutside : u ∉ convexBifunctionParameterDomain G) : ∀ x : Fin n → ℝ, G u x = ⊤

Lean source

A constantly-⊤ primal section has constantly-⊥ convex pairing.

lemma helperForCorollary33_2_2_convex_pairing_eq_bot_of_off_parameterDomain {m n : ℕ} {G : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} (hOutside : u ∉ convexBifunctionParameterDomain G) : ∀ xStar : Fin n → ℝ, convexBifunctionPairing G u xStar = ⊥

Lean source

The negative convex-pairing section of a proper polyhedral convex bifunction is polyhedral convex.

lemma helperForCorollary33_2_2_negConvexPairingSection_polyhedral {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : IsProperPolyhedralConvexBifunction F) (xStar : Fin n → ℝ) : IsPolyhedralConvexFunction m (fun u => -convexBifunctionPairing F u xStar)

Lean source

The concave-pairing section of a proper polyhedral concave bifunction is polyhedral convex.

lemma helperForCorollary33_2_2_concavePairingSection_polyhedral {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : IsProperPolyhedralConcaveBifunction F) (xStar : Fin n → ℝ) : IsPolyhedralConvexFunction m (fun u => concaveBifunctionPairing F u xStar)

Lean source

Corollary33.2.2: Let F be a proper polyhedral convex or concave bifunction. Then \u27e6F u, x^*\u27e7 = \u27e6u, F^* x^*\u27e7 unless both variables lie outside their respective domains; at such a point the two pairings take opposite infinite values.

theorem properPolyhedralBifunction_pairing_eq_except_off_both_domains {m n : ℕ} : (∀ {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal}, IsProperPolyhedralConvexBifunction F → ∀ u : Fin m → ℝ, ∀ xStar : Fin n → ℝ, (¬ (u ∉ convexBifunctionParameterDomain F ∧ xStar ∉ convexBifunctionAdjointDomain F) → convexBifunctionPairing F u xStar = convexBifunctionCanonicalAdjointPairing F xStar u) ∧ ((u ∉ convexBifunctionParameterDomain F ∧ xStar ∉ convexBifunctionAdjointDomain F) → (convexBifunctionPairing F u xStar = ⊤ ∧ convexBifunctionCanonicalAdjointPairing F xStar u = ⊥) ∨ (convexBifunctionPairing F u xStar = ⊥ ∧ convexBifunctionCanonicalAdjointPairing F xStar u = ⊤))) ∧ (∀ {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal}, IsProperPolyhedralConcaveBifunction F → ∀ u : Fin m → ℝ, ∀ xStar : Fin n → ℝ, (¬ (u ∉ concaveBifunctionParameterDomain F ∧ xStar ∉ concaveBifunctionAdjointDomain F) → concaveBifunctionPairing F u xStar = concaveBifunctionCanonicalAdjointPairing F xStar u) ∧ ((u ∉ concaveBifunctionParameterDomain F ∧ xStar ∉ concaveBifunctionAdjointDomain F) → (concaveBifunctionPairing F u xStar = ⊤ ∧ concaveBifunctionCanonicalAdjointPairing F xStar u = ⊥) ∨ (concaveBifunctionPairing F u xStar = ⊥ ∧ concaveBifunctionCanonicalAdjointPairing F xStar u = ⊤)))

Lean source

Helper for Corollary33.2.2: every polyhedral convex function is already closed in the lower-semicontinuity sense used by ClosedConvexFunction.

lemma helperForCorollary33_2_2_polyhedralConvexFunction_isClosed {k : ℕ} {f : (Fin k → ℝ) → EReal} (hfpoly : IsPolyhedralConvexFunction k f) : ClosedConvexFunction f

Lean source

Helper for Corollary33.2.2: a proper polyhedral convex bifunction already has the graph closedness and no-⊥ package needed by Theorem33.2.

lemma helperForCorollary33_2_2_convex_graphClosed_and_noBot {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : IsProperPolyhedralConvexBifunction F) : IsGraphConvexBifunction F ∧ IsFunctionConvexClosed (graphFunctionOfBifunction F) ∧ HasNoBotValuesBifunction F

§33 — Saddle-Functions — Part 17

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Definition33.0.27: The conjugate conventions used below are the existing convexConjugate and concaveConjugate operations. A concave function ψ : ℝ^n → EReal is closed concave when it equals its concave biconjugate ψ = ψ_{**}; equivalently, its hypograph is closed. This predicate records the biconjugate formulation.

def IsFunctionClosedConcave {n : ℕ} (ψ : (Fin n → ℝ) → EReal) : Prop := ψ = concaveConjugate (concaveConjugate ψ)

Lean source

Definition33.0.28: For a convex bifunction F : ℝ^m → (ℝ^n → EReal), meaning that its graph function f (u, x) = (F u) x is convex on ℝ^(m + n) as expressed by IsGraphConvexBifunction F, the genuine adjoint F^* : ℝ^n → (ℝ^m → EReal) is defined by (F^* x^*)(u^*) = inf_{u, x} ((F u) x - ⟪x, x^*⟫ + ⟪u, u^*⟫). The first pairing ⟪F u, x^*⟫ = sup_x (⟪x, x^*⟫ - (F u) x) is the existing convexBifunctionPairing F u xStar, and the second pairing ⟪u, F^* x^*⟫ is defined separately below by genuineConvexBifunctionAdjointPairing.

noncomputable def genuineConvexBifunctionAdjoint {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin n → ℝ) → (Fin m → ℝ) → EReal := fun xStar uStar => sInf <| Set.range fun ux : (Fin m → ℝ) × (Fin n → ℝ) => (F ux.1 ux.2 - ((dotProduct ux.2 xStar : ℝ) : EReal)) + ((dotProduct ux.1 uStar : ℝ) : EReal)

Lean source

The adjoint-side pairing attached to the genuine bifunction adjoint of a convex bifunction, defined by ⟪u, F^* x^*⟫ = inf_{u^*} (⟪u, u^*⟫ - (F^* x^*)(u^*)).

noncomputable def genuineConvexBifunctionAdjointPairing {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (u : Fin m → ℝ) (xStar : Fin n → ℝ) : EReal := sInf <| Set.range fun uStar : Fin m → ℝ => ((dotProduct u uStar : ℝ) : EReal) - genuineConvexBifunctionAdjoint F xStar uStar

Lean source

Definition33.0.29: For a convex bifunction F, dom F is the set of parameters u such that the section F u is not identically +∞, and dom F^* is the set of dual vectors x^* such that the genuine adjoint section F^* x^* from Definition33.0.28 is not identically -∞. This packages the pair (dom F, dom F^*).

def convexBifunctionDomains {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Set (Fin m → ℝ) × Set (Fin n → ℝ) := (convexBifunctionParameterDomain F, {xStar | ∃ uStar : Fin m → ℝ, genuineConvexBifunctionAdjoint F xStar uStar ≠ ⊥})

Lean source

The concave indicator of a point in ℝ^n, taking the value 0 at that point and ⊥ elsewhere.

noncomputable def concavePointIndicator {n : ℕ} (x : Fin n → ℝ) : (Fin n → ℝ) → EReal := fun z => if z = x then 0 else ⊥

Lean source

The indicator bifunction attached to a linear map has a convex graph function.

theorem pointIndicator_linearMap_isGraphConvexBifunction {m n : ℕ} (A : (Fin m → ℝ) →ₗ[ℝ] (Fin n → ℝ)) : IsGraphConvexBifunction (fun u => pointIndicator (A u))

Lean source

Lemma33.0.30 (Indicator case): When F u is the convex point indicator at A u for a linear map A : ℝ^m → ℝ^n, the genuine adjoint F^* x^* is the concave point indicator at the adjoint image A^* x^*. Consequently, the two pairings reduce to the classical adjoint relation ⟪A u, x^*⟫ = ⟪u, A^* x^*⟫.

theorem genuineConvexBifunctionAdjoint_indicator_linearMap {m n : ℕ} (A : (Fin m → ℝ) →ₗ[ℝ] (Fin n → ℝ)) : (∀ xStar : Fin n → ℝ, genuineConvexBifunctionAdjoint (fun u => pointIndicator (A u)) xStar = concavePointIndicator (coordinateAdjointLinearMap A xStar)) ∧ ∀ u : Fin m → ℝ, ∀ xStar : Fin n → ℝ, dotProduct (A u) xStar = dotProduct u (coordinateAdjointLinearMap A xStar)

Lean source

Helper for Lemma33.0.30: on the dual graph uStar = A* xStar, the genuine adjoint of the indicator bifunction takes the value 0.

lemma helperForLemma33_0_30_indicator_linearMap_adjoint_eq_zero_on_graph {m n : ℕ} (A : (Fin m → ℝ) →ₗ[ℝ] (Fin n → ℝ)) (xStar : Fin n → ℝ) (uStar : Fin m → ℝ) (hEq : uStar = coordinateAdjointLinearMap A xStar) : genuineConvexBifunctionAdjoint (fun u => pointIndicator (A u)) xStar uStar = 0

Lean source

Helper for Lemma33.0.30: evaluating the genuine adjoint on a primal graph point produces the linear form with coefficient uStar - A* xStar.

lemma helperForLemma33_0_30_indicator_linearMap_graph_term {m n : ℕ} (A : (Fin m → ℝ) →ₗ[ℝ] (Fin n → ℝ)) (xStar : Fin n → ℝ) (uStar u : Fin m → ℝ) : ((pointIndicator (A u) (A u) - ((dotProduct (A u) xStar : ℝ) : EReal)) + ((dotProduct u uStar : ℝ) : EReal)) = (((dotProduct u (uStar - coordinateAdjointLinearMap A xStar) : ℝ) : EReal))

Lean source

Helper for Lemma33.0.30: off the dual graph uStar ≠ A* xStar, the genuine adjoint of the indicator bifunction drops to ⊥.

lemma helperForLemma33_0_30_indicator_linearMap_adjoint_eq_bot_off_graph {m n : ℕ} (A : (Fin m → ℝ) →ₗ[ℝ] (Fin n → ℝ)) (xStar : Fin n → ℝ) (uStar : Fin m → ℝ) (hNe : uStar ≠ coordinateAdjointLinearMap A xStar) : genuineConvexBifunctionAdjoint (fun u => pointIndicator (A u)) xStar uStar = ⊥

§33 — Saddle-Functions — Part 18

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

def HasInnerProductEquation : {m n : ℕ} → ((Fin m → ℝ) → (Fin n → ℝ) → EReal) → Prop := fun {m n} F => ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), convexBifunctionPairing F u xStar = genuineConvexBifunctionAdjointPairing F u xStar

Lean source

theorem optimalValueEqualityAtZero_iff_pairingEqualityAtZero : ∀ {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal), (∀ (u : Fin m → ℝ), sInf (Set.range (F u)) = -convexBifunctionPairing F u 0) ∧ (∀ (xStar : Fin n → ℝ), sSup (Set.range (genuineConvexBifunctionAdjoint F xStar)) = -genuineConvexBifunctionAdjointPairing F 0 xStar) ∧ (sInf (Set.range (F 0)) = sSup (Set.range (genuineConvexBifunctionAdjoint F 0)) ↔ convexBifunctionPairing F 0 0 = genuineConvexBifunctionAdjointPairing F 0 0)

Lean source

def translatedTiltedBifunction : {m n : ℕ} → ((Fin m → ℝ) → (Fin n → ℝ) → EReal) → (Fin m → ℝ) → (Fin n → ℝ) → (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun {m n} F u xStar v y => F (u + v) y - ↑(y ⬝ᵥ xStar)

Lean source

theorem genuineConvexBifunctionAdjoint_translatedTiltedBifunction : ∀ {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u vStar : Fin m → ℝ} {xStar yStar : Fin n → ℝ}, genuineConvexBifunctionAdjoint (translatedTiltedBifunction F u xStar) yStar vStar = genuineConvexBifunctionAdjoint F (xStar + yStar) vStar - ↑(u ⬝ᵥ vStar)

Lean source

theorem translatedTiltedBifunction_optimalValues : ∀ {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal), IsGraphConvexBifunction F → ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), sInf (Set.range fun y => F u y - ↑(y ⬝ᵥ xStar)) = -convexBifunctionPairing F u xStar ∧ sSup (Set.range fun vStar => genuineConvexBifunctionAdjoint F xStar vStar - ↑(u ⬝ᵥ vStar)) = -genuineConvexBifunctionAdjointPairing F u xStar

Lean source

def HasEqualOptimalValuesForTranslatedTiltedPrograms : {m n : ℕ} → ((Fin m → ℝ) → (Fin n → ℝ) → EReal) → (Fin m → ℝ) → (Fin n → ℝ) → Prop := fun {m n} F u xStar => sInf (Set.range fun y => F u y - ↑(y ⬝ᵥ xStar)) = sSup (Set.range fun vStar => genuineConvexBifunctionAdjoint F xStar vStar - ↑(u ⬝ᵥ vStar))

Lean source

The translated-program normality predicate used downstream only through its criterion-level interface is aligned with equality of the translated primal and dual optimal values. Route correction: the earlier coordinatewise-closure encoding was stronger than what the later split files actually consume, and it did not match the Chapter 30 normality criterion that drives these translated arguments.

def HasNormalityForTranslatedTiltedPrograms : {m n : ℕ} → ((Fin m → ℝ) → (Fin n → ℝ) → EReal) → (Fin m → ℝ) → (Fin n → ℝ) → Prop := fun {m n} F u xStar => HasEqualOptimalValuesForTranslatedTiltedPrograms F u xStar

Lean source

theorem pairingEquality_iff_equalOptimalValues_for_translatedTiltedPrograms : ∀ {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal), IsGraphConvexBifunction F → ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), convexBifunctionPairing F u xStar = genuineConvexBifunctionAdjointPairing F u xStar ↔ HasEqualOptimalValuesForTranslatedTiltedPrograms F u xStar

Lean source

theorem normality_iff_equalOptimalValues_for_translatedTiltedPrograms : ∀ {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (u : Fin m → ℝ) (xStar : Fin n → ℝ), HasNormalityForTranslatedTiltedPrograms F u xStar ↔ HasEqualOptimalValuesForTranslatedTiltedPrograms F u xStar

Lean source

theorem pairingEquality_iff_normality_for_translatedTiltedPrograms : ∀ {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal), IsGraphConvexBifunction F → ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), convexBifunctionPairing F u xStar = genuineConvexBifunctionAdjointPairing F u xStar ↔ HasNormalityForTranslatedTiltedPrograms F u xStar

Lean source

theorem oppositeInfinities_off_both_convexBifunctionDomains : ∀ {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {xStar : Fin n → ℝ}, u ∉ (convexBifunctionDomains F).1 → xStar ∉ (convexBifunctionDomains F).2 → (∀ (x : Fin n → ℝ), F u x = ⊤) ∧ (∀ (uStar : Fin m → ℝ), genuineConvexBifunctionAdjoint F xStar uStar = ⊥) ∧ convexBifunctionPairing F u xStar = ⊥ ∧ genuineConvexBifunctionAdjointPairing F u xStar = ⊤

Lean source

def IsFullyClosedSaddleFunction : {m n : ℕ} → ((Fin m → ℝ) → (Fin n → ℝ) → EReal) → Prop := fun {m n} K => IsConcaveConvexOn Set.univ Set.univ K ∧ IsConcaveClosedInFirst K ∧ IsConvexClosedInSecond K ∨ IsConvexConcaveOn Set.univ Set.univ K ∧ IsConvexClosedInFirst K ∧ IsConcaveClosedInSecond K

Lean source

Compatibility name for the genuine-adjoint identification used by later split files.

lemma helperForCorollary33_0_40_genuineAdjoint_eq_closureSideAdjointPairing {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (xStar : Fin n → ℝ) (uStar : Fin m → ℝ) : genuineConvexBifunctionAdjoint F xStar uStar = convexBifunctionAdjointPairing F xStar uStar

Lean source

The genuine adjoint pairing and the canonical Section 33 adjoint pairing are the same second conjugate. Neither is the first-conjugate adjoint itself.

lemma helperForCorollary33_0_40_genuinePairing_eq_convexAdjointPairing {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) : genuineConvexBifunctionAdjointPairing F u xStar = convexBifunctionCanonicalAdjointPairing F xStar u

Lean source

Theorem33.0.39 (corrected witness-upgrade fragment): once a kernel K is already represented by a graph-convex bifunction F with no ⊥-values and full primal parameter domain, the Section 33 global pairing identity ⟪F u, x^*⟫ = (F^* x^*)(u) follows everywhere. This is the dependency-closed part of the original textbook statement that is actually used later: the converse existence direction and the full coordinatewise-closure equivalence require additional bridges handled elsewhere.

theorem fullyClosedConcaveConvex_iff_exists_graphConvexBifunction_with_innerProductEquation : ∀ {n : ℕ} {K : (Fin n → ℝ) → (Fin n → ℝ) → EReal}, (∃ F, IsGraphConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ (convexBifunctionDomains F).1 = Set.univ ∧ K = convexBifunctionPairing F) → ∃ F, IsGraphConvexBifunction F ∧ K = convexBifunctionPairing F ∧ HasInnerProductEquation F

Lean source

Helper for Lemma33.0.35: evaluating the translated genuine adjoint at the zero dual section gives exactly the textbook dual objective.

lemma helperForLemma33_0_35_translatedTiltedGenuineAdjointAtZero_eq_targetDualObjective {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) (vStar : Fin m → ℝ) : genuineConvexBifunctionAdjoint (translatedTiltedBifunction F u xStar) (0 : Fin n → ℝ) vStar = genuineConvexBifunctionAdjoint F xStar vStar - ↑(u ⬝ᵥ vStar)

Lean source

Helper for Lemma33.0.35: the supremum of the zero primal section of the genuine adjoint is the negative of the corresponding genuine-adjoint pairing at the origin.

lemma helperForLemma33_0_35_genuineAdjointZeroSection_sSup_eq_negPairingAtZero {m n : ℕ} (G : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (xStar : Fin n → ℝ) : sSup (Set.range fun vStar : Fin m → ℝ => genuineConvexBifunctionAdjoint G xStar vStar) = -genuineConvexBifunctionAdjointPairing G (0 : Fin m → ℝ) xStar

Lean source

Helper for Lemma33.0.35: the zero pairing of the translated-and-tilted bifunction is the textbook genuine-adjoint pairing ⟪u, F^* x^*⟫.

lemma helperForLemma33_0_35_translatedTiltedGenuineAdjointPairingAtZero_eq {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) : genuineConvexBifunctionAdjointPairing (translatedTiltedBifunction F u xStar) (0 : Fin m → ℝ) (0 : Fin n → ℝ) = genuineConvexBifunctionAdjointPairing F u xStar

Lean source

Helper for Lemma33.0.35: the dual objective range in the textbook statement is exactly the zero dual section of the translated-and-tilted bifunction.

lemma helperForLemma33_0_35_translatedTiltedDualRange_eq_targetDualRange {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) : Set.range (fun vStar : Fin m → ℝ => genuineConvexBifunctionAdjoint (translatedTiltedBifunction F u xStar) (0 : Fin n → ℝ) vStar) = Set.range (fun vStar : Fin m → ℝ => genuineConvexBifunctionAdjoint F xStar vStar - ↑(u ⬝ᵥ vStar))

Lean source

Helper for Lemma33.0.35: the zero primal section of the translated-and-tilted bifunction already has the textbook primal optimal value.

lemma helperForLemma33_0_35_translatedTiltedPrimalZeroSection_sInf_eq_negPairing {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) : sInf (Set.range fun y : Fin n → ℝ => translatedTiltedBifunction F u xStar (0 : Fin m → ℝ) y) = -convexBifunctionPairing F u xStar

Lean source

Helper for Lemma33.0.35: the textbook primal objective has the same infimum as the zero primal section of the translated-and-tilted bifunction.

lemma helperForLemma33_0_35_targetPrimalObjective_sInf_eq_translatedTiltedPrimalZeroSection {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) : sInf (Set.range fun y : Fin n → ℝ => F u y - ↑(y ⬝ᵥ xStar)) = sInf (Set.range fun y : Fin n → ℝ => translatedTiltedBifunction F u xStar (0 : Fin m → ℝ) y)

Lean source

Helper for Lemma33.0.35: the zero dual section of the translated-and-tilted bifunction already has the textbook dual optimal value.

lemma helperForLemma33_0_35_translatedTiltedDualZeroSection_sSup_eq_negPairing {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) : sSup (Set.range fun vStar : Fin m → ℝ => genuineConvexBifunctionAdjoint (translatedTiltedBifunction F u xStar) (0 : Fin n → ℝ) vStar) = -genuineConvexBifunctionAdjointPairing F u xStar

Lean source

Helper for Lemma33.0.35: the textbook dual objective has the same supremum as the zero dual section of the translated-and-tilted bifunction.

lemma helperForLemma33_0_35_targetDualObjective_sSup_eq_translatedTiltedDualZeroSection {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) : sSup (Set.range fun vStar : Fin m → ℝ => genuineConvexBifunctionAdjoint F xStar vStar - ↑(u ⬝ᵥ vStar)) = sSup (Set.range fun vStar : Fin m → ℝ => genuineConvexBifunctionAdjoint (translatedTiltedBifunction F u xStar) (0 : Fin n → ℝ) vStar)

Lean source

Helper for Lemma33.0.37: outside the second component of dom F*, the whole genuine adjoint section is constantly ⊥.

lemma helperForLemma33_0_37_allBotGenuineAdjoint_of_off_secondConvexBifunctionDomain {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {xStar : Fin n → ℝ} (hOutside : xStar ∉ (convexBifunctionDomains F).2) : ∀ uStar : Fin m → ℝ, genuineConvexBifunctionAdjoint F xStar uStar = ⊥

Lean source

Helper for Lemma33.0.37: leaving the first convex-bifunction domain component forces both the primal section and every primal pairing to collapse.

lemma helperForLemma33_0_37_primalCollapse_of_off_firstConvexBifunctionDomain {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} (hu : u ∉ (convexBifunctionDomains F).1) : (∀ x : Fin n → ℝ, F u x = ⊤) ∧ (∀ xStar : Fin n → ℝ, convexBifunctionPairing F u xStar = ⊥)

Lean source

Helper for Lemma33.0.37: if the genuine adjoint section is constantly ⊥, then the adjoint-side pairing collapses to ⊤.

lemma helperForLemma33_0_37_genuineAdjointPairing_eq_top_of_allBotAdjointSection {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {xStar : Fin n → ℝ} (hAllBot : ∀ uStar : Fin m → ℝ, genuineConvexBifunctionAdjoint F xStar uStar = ⊥) : ∀ u : Fin m → ℝ, genuineConvexBifunctionAdjointPairing F u xStar = ⊤

Lean source

Helper for Lemma33.0.37: leaving the second convex-bifunction domain component forces both the genuine adjoint section and every adjoint-side pairing to collapse.

lemma helperForLemma33_0_37_dualCollapse_of_off_secondConvexBifunctionDomain {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {xStar : Fin n → ℝ} (hx : xStar ∉ (convexBifunctionDomains F).2) : (∀ uStar : Fin m → ℝ, genuineConvexBifunctionAdjoint F xStar uStar = ⊥) ∧ (∀ u : Fin m → ℝ, genuineConvexBifunctionAdjointPairing F u xStar = ⊤)

Lean source

Helper for Theorem33.0.39: a pairing representation by convexBifunctionPairing F combined with the absence of ⊥-values already forces the full primal parameter domain of F.

lemma helperForTheorem33_0_39_fullParameterDomain_of_pairingRepresentation {n : ℕ} {K F : (Fin n → ℝ) → (Fin n → ℝ) → EReal} (hRep : K = convexBifunctionPairing F) (hNoBot : HasNoBotValuesBifunction K) : (convexBifunctionDomains F).1 = Set.univ

Lean source

Helper for Theorem33.0.39: a full primal parameter domain gives a concrete non-⊤ witness in every primal section.

lemma helperForTheorem33_0_39_sectionWitness_of_fullParameterDomain {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hDom : (convexBifunctionDomains F).1 = Set.univ) : ∀ u : Fin m → ℝ, ∃ x : Fin n → ℝ, F u x ≠ ⊤

Lean source

Helper for Theorem33.0.39: full primal parameter domain makes every translated-and-tilted primal program consistent at the origin.

lemma helperForTheorem33_0_39_translatedPrimalOrigin_neTop_of_fullParameterDomain {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hDom : (convexBifunctionDomains F).1 = Set.univ) : ∀ u : Fin m → ℝ, ∀ xStar : Fin n → ℝ, convexProgramAssociatedWith (translatedTiltedBifunction F u xStar) (0 : Fin m → ℝ) ≠ (⊤ : EReal)

Lean source

Helper for Theorem33.0.39: graph-function closedness upgrades a Rockafellar convex bifunction to the graph-convex predicate used by the translated-program machinery.

lemma helperForTheorem33_0_39_graphConvex_of_graphFunctionClosed {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hRock : IsRockafellarConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hGraphClosed : IsFunctionConvexClosed (graphFunctionOfBifunction F)) : IsGraphConvexBifunction F

Lean source

Helper for Theorem33.0.39: full primal parameter domain already forces the strict parameter domain {u | ∃ x, F u x < ⊤} to be all of space.

lemma helperForTheorem33_0_39_strictParameterDomain_of_fullParameterDomain {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hDom : (convexBifunctionDomains F).1 = Set.univ) : {u : Fin m → ℝ | ∃ x : Fin n → ℝ, F u x < ⊤} = Set.univ

Lean source

Helper for Theorem33.0.39: in the full primal-domain branch, Corollary33.2.1 already identifies the primal pairing with the closure-side adjoint pairing everywhere.

lemma helperForTheorem33_0_39_closureSidePairingEq_of_fullParameterDomain {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hDom : (convexBifunctionDomains F).1 = Set.univ) : ∀ u : Fin m → ℝ, ∀ xStar : Fin n → ℝ, convexBifunctionPairing F u xStar = convexBifunctionCanonicalAdjointPairing F xStar u

Lean source

Helper for Corollary33.0.40: the explicit genuine adjoint is the first parameter-side conjugate convexBifunctionAdjointPairing, not its second-conjugate pairing.

lemma helperForCorollary33_0_40_genuineAdjoint_eq_convexBifunctionAdjointPairing {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (xStar : Fin n → ℝ) (uStar : Fin m → ℝ) : genuineConvexBifunctionAdjoint F xStar uStar = convexBifunctionAdjointPairing F xStar uStar

Lean source

Helper for Corollary33.0.40: the genuine adjoint pairing is exactly the concave conjugate of the genuine adjoint section in the primal variable.

lemma helperForCorollary33_0_40_genuinePairing_eq_concaveConjugate_genuineAdjoint {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) : genuineConvexBifunctionAdjointPairing F u xStar = concaveConjugate (genuineConvexBifunctionAdjoint F xStar) u

Lean source

Helper for Corollary33.0.40: full primal domain forces the genuine adjoint pairing to collapse back to the primal pairing at every point.

lemma helperForCorollary33_0_40_pairing_eq_genuinePairing_of_fullParameterDomain {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hDom : (convexBifunctionDomains F).1 = Set.univ) : ∀ u : Fin m → ℝ, ∀ xStar : Fin n → ℝ, genuineConvexBifunctionAdjointPairing F u xStar = convexBifunctionPairing F u xStar

Lean source

Helper for Corollary33.0.40: a single primal witness away from ⊤ forces the raw pairing kernel to avoid ⊥ somewhere for every dual vector.

lemma helperForCorollary33_0_40_fullRawPairingDomain_of_primalWitness {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hWitness : ∃ u₀ : Fin m → ℝ, ∃ x₀ : Fin n → ℝ, F u₀ x₀ ≠ ⊤) : {xStar : Fin n → ℝ | ∃ u : Fin m → ℝ, convexBifunctionPairing F u xStar ≠ ⊥} = Set.univ

Lean source

Helper for Corollary33.0.40: in the graph-closed branch, the genuine adjoint pairing still collapses to the primal pairing by splitting into the nontrivial and everywhere-⊤ cases.

lemma helperForCorollary33_0_40_pairing_eq_genuinePairing_of_closed_fullGenuineAdjointDomain {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hGraphClosed : IsFunctionConvexClosed (graphFunctionOfBifunction F)) (hAdjDom : (convexBifunctionDomains F).2 = Set.univ) : ∀ u : Fin m → ℝ, ∀ xStar : Fin n → ℝ, genuineConvexBifunctionAdjointPairing F u xStar = convexBifunctionPairing F u xStar

§33 — Saddle-Functions — Part 19

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Corollary33.0.40 (Sufficient conditions for the global pairing identity): the primal pairing and genuine-adjoint pairing agree everywhere if either the primal domain is full, or the graph function is convex-closed and the genuine adjoint domain is full.

theorem innerProductEquation_of_fullDomain_or_closed_fullAdjointDomain : ∀ {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal}, IsGraphConvexBifunction F → HasNoBotValuesBifunction F → ((convexBifunctionDomains F).1 = Set.univ ∨ IsFunctionConvexClosed (graphFunctionOfBifunction F) ∧ (convexBifunctionDomains F).2 = Set.univ) → HasInnerProductEquation F

Lean source

Corollary33.0.41 (Failure when both domains are not full): if neither the primal domain nor the genuine adjoint domain is all of space, then the global pairing identity cannot hold.

theorem not_hasInnerProductEquation_of_nonfull_domains : ∀ {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal}, (convexBifunctionDomains F).1 ≠ Set.univ → (convexBifunctionDomains F).2 ≠ Set.univ → ¬HasInnerProductEquation F

Lean source

Canonical Chapter 2 closure in the first (concave) coordinate. This is kept distinct from the older ball-based raw closure, which need not agree on improper sections.

noncomputable def canonicalConcaveClosureInFirst {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u xStar => concaveClosure (fun u' => K u' xStar) u

Lean source

Canonical Chapter 2 closure in the second (convex) coordinate.

noncomputable def canonicalConvexClosureInSecond {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u xStar => convexFunctionClosure (fun y => K u y) xStar

Lean source

structure SaddleClosednessPredicates {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Type where lowerClosed : Prop upperClosed : Prop

Lean source

def saddleClosednessPredicates : {m n : ℕ} → (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) → SaddleClosednessPredicates K

Lean source

@[reducible] def IsLowerClosedSaddleFunction : {m n : ℕ} → ((Fin m → ℝ) → (Fin n → ℝ) → EReal) → Prop := fun {_ _} K => (saddleClosednessPredicates K).lowerClosed

Lean source

@[reducible] def IsUpperClosedSaddleFunction : {m n : ℕ} → ((Fin m → ℝ) → (Fin n → ℝ) → EReal) → Prop := fun {_ _} K => (saddleClosednessPredicates K).upperClosed

Lean source

def erealOfRealBifunction : {m n : ℕ} → ((Fin m → ℝ) → (Fin n → ℝ) → ℝ) → (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun {_ _} K u xStar => ↑(K u xStar)

Lean source

@[reducible] noncomputable def lowerSimpleExtensionOfReal : {m n : ℕ} → Set (Fin m → ℝ) → Set (Fin n → ℝ) → ((Fin m → ℝ) → (Fin n → ℝ) → ℝ) → (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun {_ _} C D K => lowerSimpleExtension C D (erealOfRealBifunction K)

Lean source

@[reducible] noncomputable def upperSimpleExtensionOfReal : {m n : ℕ} → Set (Fin m → ℝ) → Set (Fin n → ℝ) → ((Fin m → ℝ) → (Fin n → ℝ) → ℝ) → (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun {_ _} C D K => upperSimpleExtension C D (erealOfRealBifunction K)

Lean source

Helper for Corollary33.0.41: a proper first convex-bifunction domain component omits some primal parameter.

lemma helperForCorollary33_0_41_exists_off_first_domain {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hdom : (convexBifunctionDomains F).1 ≠ Set.univ) : ∃ u : Fin m → ℝ, u ∉ (convexBifunctionDomains F).1

Lean source

Helper for Corollary33.0.41: a proper second convex-bifunction domain component omits some dual vector.

lemma helperForCorollary33_0_41_exists_off_second_domain {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hdomAdj : (convexBifunctionDomains F).2 ≠ Set.univ) : ∃ xStar : Fin n → ℝ, xStar ∉ (convexBifunctionDomains F).2

Lean source

Helper for Corollary33.0.41: the inner-product equation fails at any point lying outside both convex-bifunction domain components.

lemma helperForCorollary33_0_41_contradiction_at_off_domain_point {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {xStar : Fin n → ℝ} (hInner : HasInnerProductEquation F) (huOutside : u ∉ (convexBifunctionDomains F).1) (hxOutside : xStar ∉ (convexBifunctionDomains F).2) : False

Lean source

Helper for Theorem33.0.39: the genuine inner-product equation forces at least one of the two domain components to be all of space.

lemma helperForTheorem33_0_39_domainDichotomy_of_innerProductEquation {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hInner : HasInnerProductEquation F) : (convexBifunctionDomains F).1 = Set.univ ∨ (convexBifunctionDomains F).2 = Set.univ

Lean source

Helper for Corollary33.3.1: a closed convex witness with no ⊥ values already supplies the Rockafellar convexity and graph-closure data needed by the coordinatewise-closure machinery.

lemma helperForCorollary33_3_1_rockafellarConvex_and_graphFunctionClosed_of_closedConvexWitness {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hClosed : ClosedConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) : IsGraphConvexBifunction F ∧ IsFunctionConvexClosed (graphFunctionOfBifunction F)

Lean source

Helper for Corollary33.3.1: a closed convex bifunction witness with no ⊥ values forces the canonical coordinatewise closure identities for the represented pairing kernels.

lemma helperForCorollary33_3_1_coordinatewise_closure_pair_of_closedConvexWitness {m n : ℕ} {K Kbar : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hClosed : ClosedConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hPair : ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), K u xStar = convexBifunctionPairing F u xStar) (hAdj : ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), Kbar u xStar = convexBifunctionCanonicalAdjointPairing F xStar u) : Kbar = canonicalConcaveClosureInFirst K ∧ canonicalConvexClosureInSecond Kbar = K

Lean source

Helper for Corollary33.3.1: the forward implication only uses the packaged closed convex witness data, so it can be invoked directly from the existential interface used in the corollary statement.

lemma helperForCorollary33_3_1_coordinatewise_closure_pair_of_closedConvexWitnessPackage {m n : ℕ} {K Kbar : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hWitness : ClosedConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ (∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), K u xStar = convexBifunctionPairing F u xStar) ∧ ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), Kbar u xStar = convexBifunctionCanonicalAdjointPairing F xStar u) : Kbar = canonicalConcaveClosureInFirst K ∧ canonicalConvexClosureInSecond Kbar = K

Lean source

Helper for Corollary33.3.1: the unique-existence hypothesis in the corollary statement still yields the same coordinatewise closure pair after discarding uniqueness.

lemma helperForCorollary33_3_1_coordinatewise_closure_pair_of_closedConvexUniqueWitness {m n : ℕ} {K Kbar : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hExists : ∃! F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, ClosedConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ (∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), K u xStar = convexBifunctionPairing F u xStar) ∧ ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), Kbar u xStar = convexBifunctionCanonicalAdjointPairing F xStar u) : Kbar = canonicalConcaveClosureInFirst K ∧ canonicalConvexClosureInSecond Kbar = K

Lean source

Helper for Corollary33.3.2: simultaneous concave-convex and convex-concave structure on a kernel makes every first-variable section simultaneously concave and convex.

lemma helperForCorollary33_3_2_firstVariableSection_isConcaveAndConvex_of_simultaneousOrientations {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsConcaveConvexOn Set.univ Set.univ K) (hVC : IsConvexConcaveOn Set.univ Set.univ K) (xStar : Fin n → ℝ) : IsERealConcaveOn (Set.univ : Set (Fin m → ℝ)) (fun u => K u xStar) ∧ IsERealConvexOn (Set.univ : Set (Fin m → ℝ)) (fun u => K u xStar)

Lean source

Helper for Corollary33.3.2: simultaneous concave-convex and convex-concave structure on a kernel makes every second-variable section simultaneously convex and concave.

lemma helperForCorollary33_3_2_secondVariableSection_isConvexAndConcave_of_simultaneousOrientations {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsConcaveConvexOn Set.univ Set.univ K) (hVC : IsConvexConcaveOn Set.univ Set.univ K) (u : Fin m → ℝ) : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (fun xStar => K u xStar) ∧ IsERealConcaveOn (Set.univ : Set (Fin n → ℝ)) (fun xStar => K u xStar)

Lean source

Helper for Corollary33.3.2: the simultaneous orientation hypotheses package the first-variable section orientations uniformly over all frozen dual vectors.

lemma helperForCorollary33_3_2_allFirstVariableSections_areConcaveAndConvex {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsConcaveConvexOn Set.univ Set.univ K) (hVC : IsConvexConcaveOn Set.univ Set.univ K) : ∀ xStar : Fin n → ℝ, IsERealConcaveOn (Set.univ : Set (Fin m → ℝ)) (fun u => K u xStar) ∧ IsERealConvexOn (Set.univ : Set (Fin m → ℝ)) (fun u => K u xStar)

Lean source

Helper for Corollary33.3.2: the simultaneous orientation hypotheses package the second-variable section orientations uniformly over all frozen primal vectors.

lemma helperForCorollary33_3_2_allSecondVariableSections_areConvexAndConcave {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsConcaveConvexOn Set.univ Set.univ K) (hVC : IsConvexConcaveOn Set.univ Set.univ K) : ∀ u : Fin m → ℝ, IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (fun xStar => K u xStar) ∧ IsERealConcaveOn (Set.univ : Set (Fin n → ℝ)) (fun xStar => K u xStar)

Lean source

Helper for Corollary33.3.2: simultaneous concave-convex and convex-concave structure on a kernel packages both first-variable and second-variable section orientations at once.

lemma helperForCorollary33_3_2_allSections_have_simultaneousOrientations {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsConcaveConvexOn Set.univ Set.univ K) (hVC : IsConvexConcaveOn Set.univ Set.univ K) : (∀ xStar : Fin n → ℝ, IsERealConcaveOn (Set.univ : Set (Fin m → ℝ)) (fun u => K u xStar) ∧ IsERealConvexOn (Set.univ : Set (Fin m → ℝ)) (fun u => K u xStar)) ∧ ∀ u : Fin m → ℝ, IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (fun xStar => K u xStar) ∧ IsERealConcaveOn (Set.univ : Set (Fin n → ℝ)) (fun xStar => K u xStar)

Lean source

Helper for Corollary33.3.3: a real-valued kernel, viewed in EReal, satisfies both one-sided finiteness conventions required by the saddle-function correspondence.

lemma helperForCorollary33_3_3_realKernel_hasNoTopOrBotValues {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} : HasNoTopOrBotValuesBifunction (fun u xStar => (↑(K u xStar) : EReal))

Lean source

Helper for Corollary33.3.3: every lower simple extension lies pointwise below the corresponding upper simple extension.

lemma helperForCorollary33_3_3_lowerSimpleExtension_le_upperSimpleExtension {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} : ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), lowerSimpleExtension C D K u xStar ≤ upperSimpleExtension C D K u xStar

Lean source

Helper for Corollary33.3.3: on C × D, the simple extensions of a real-valued kernel agree with the original kernel after coercion to EReal.

lemma helperForCorollary33_3_3_simpleExtensions_of_real_agree_on_product {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {u : Fin m → ℝ} {xStar : Fin n → ℝ} (hu : u ∈ C) (hxStar : xStar ∈ D) : lowerSimpleExtension C D (fun u' xStar' => (↑(K u' xStar') : EReal)) u xStar = ↑(K u xStar) ∧ upperSimpleExtension C D (fun u' xStar' => (↑(K u' xStar') : EReal)) u xStar = ↑(K u xStar)

Lean source

Helper for Corollary33.3.3: the EReal simple extensions of a real kernel are globally ordered, and on C × D they both reduce to the original real kernel after coercion.

lemma helperForCorollary33_3_3_realKernel_simpleExtensions_areOrdered_and_agreeOnProduct {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} : (∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), lowerSimpleExtension C D (fun u' xStar' => (↑(K u' xStar') : EReal)) u xStar ≤ upperSimpleExtension C D (fun u' xStar' => (↑(K u' xStar') : EReal)) u xStar) ∧ ∀ ⦃u : Fin m → ℝ⦄ ⦃xStar : Fin n → ℝ⦄, u ∈ C → xStar ∈ D → lowerSimpleExtension C D (fun u' xStar' => (↑(K u' xStar') : EReal)) u xStar = ↑(K u xStar) ∧ upperSimpleExtension C D (fun u' xStar' => (↑(K u' xStar') : EReal)) u xStar = ↑(K u xStar)

Lean source

Helper for Corollary33.3.3: once the lower and upper simple extensions are identified as the expected coordinatewise closure pair, their lower-closedness, upper-closedness, and pointwise order follow formally from the Section 33 closure machinery already available in this split file.

lemma helperForCorollary33_3_3_closurePair_implies_closedness_and_order {m n : ℕ} {K1 K2 : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK1 : IsConcaveConvexOn Set.univ Set.univ K1) (hK2 : IsConcaveConvexOn Set.univ Set.univ K2) (hPair : K2 = concaveClosureInFirst K1 ∧ convexClosureInSecond K2 = K1) : IsLowerClosedSaddleFunction K1 ∧ IsUpperClosedSaddleFunction K2 ∧ ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), K1 u xStar ≤ K2 u xStar

Lean source

Helper for Corollary33.3.3: the lower simple extension of a real-valued kernel has exactly the original primal constraint set as the locus where some value is not ⊥.

lemma helperForCorollary33_3_3_lowerSimpleExtensionOfReal_nonbotSlice_set_eq {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hD_nonempty : D.Nonempty) : {u : Fin m → ℝ | ∃ xStar : Fin n → ℝ, lowerSimpleExtensionOfReal C D K u xStar ≠ ⊥} = C

Lean source

Helper for Corollary33.3.3: the upper simple extension of a real-valued kernel has exactly the original dual constraint set as the locus where some value is not ⊤.

lemma helperForCorollary33_3_3_upperSimpleExtensionOfReal_nontopSlice_set_eq {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_nonempty : C.Nonempty) : {xStar : Fin n → ℝ | ∃ u : Fin m → ℝ, upperSimpleExtensionOfReal C D K u xStar ≠ ⊤} = D

Lean source

Helper for Corollary33.3.3: the two real-valued simple extensions recover exactly the original primal and dual constraint sets as their non-infinite slice domains.

lemma helperForCorollary33_3_3_simpleExtension_sliceDomains_eq_constraints {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_nonempty : C.Nonempty) (hD_nonempty : D.Nonempty) : ({u : Fin m → ℝ | ∃ xStar : Fin n → ℝ, lowerSimpleExtensionOfReal C D K u xStar ≠ ⊥} = C) ∧ ({xStar : Fin n → ℝ | ∃ u : Fin m → ℝ, upperSimpleExtensionOfReal C D K u xStar ≠ ⊤} = D)

Lean source

Helper for Corollary33.3.3: outside the primal constraint set C, the upper simple extension freezes to the top/bottom indicator of the dual constraint set D.

lemma helperForCorollary33_3_3_upperSimpleExtensionOfReal_offParameterSection {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {u : Fin m → ℝ} (hu : u ∉ C) : upperSimpleExtensionOfReal C D K u = fun xStar : Fin n → ℝ => if xStar ∈ D then (⊥ : EReal) else ⊤

Lean source

Helper for Corollary33.3.3: an off-C section of the upper simple extension attains ⊥ on D and ⊤ off D.

lemma helperForCorollary33_3_3_upperSimpleExtensionOfReal_offParameterSection_mixedValues {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {u : Fin m → ℝ} (hu : u ∉ C) {xIn xOut : Fin n → ℝ} (hxIn : xIn ∈ D) (hxOut : xOut ∉ D) : upperSimpleExtensionOfReal C D K u xIn = (⊥ : EReal) ∧ upperSimpleExtensionOfReal C D K u xOut = (⊤ : EReal)

Lean source

Helper for Corollary33.3.3: if one primal point lies in C, another lies outside C, and their midpoint returns to C, then freezing the upper simple extension at a dual point of D produces a first-variable section that is not convex on all of ℝ^m.

lemma helperForCorollary33_3_3_upperSimpleExtensionOfReal_onDualSection_not_convexOn_univ {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {xStar : Fin n → ℝ} (hxStar : xStar ∈ D) {uIn uOut : Fin m → ℝ} (_huIn : uIn ∈ C) (huOut : uOut ∉ C) (hMidIn : (1 / 2 : ℝ) • uIn + (1 / 2 : ℝ) • uOut ∈ C) : ¬ IsERealConvexOn (Set.univ : Set (Fin m → ℝ)) (fun u => upperSimpleExtensionOfReal C D K u xStar)

Lean source

Helper for Corollary33.3.3: the midpoint witness above already rules out the convex-concave orientation for the upper simple extension on all of ℝ^m × ℝ^n.

lemma helperForCorollary33_3_3_upperSimpleExtensionOfReal_not_convexConcaveOn_univ_of_midpoint_witness {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {xStar : Fin n → ℝ} (hxStar : xStar ∈ D) {uIn uOut : Fin m → ℝ} (huIn : uIn ∈ C) (huOut : uOut ∉ C) (hMidIn : (1 / 2 : ℝ) • uIn + (1 / 2 : ℝ) • uOut ∈ C) : ¬ IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (upperSimpleExtensionOfReal C D K)

Lean source

Helper for Corollary33.3.3: once the midpoint witness rules out the global convex-concave orientation of the upper simple extension, the right-hand branch in the definition of IsUpperClosedSaddleFunction is impossible as well.

lemma helperForCorollary33_3_3_upperSimpleExtensionOfReal_convexConcaveUpperClosedBranch_impossible_of_midpoint_witness {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {xStar : Fin n → ℝ} (hxStar : xStar ∈ D) {uIn uOut : Fin m → ℝ} (huIn : uIn ∈ C) (huOut : uOut ∉ C) (hMidIn : (1 / 2 : ℝ) • uIn + (1 / 2 : ℝ) • uOut ∈ C) : ¬ (IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) (upperSimpleExtensionOfReal C D K) ∧ concaveClosureInSecond (convexClosureInFirst (upperSimpleExtensionOfReal C D K)) = upperSimpleExtensionOfReal C D K)

Lean source

Helper for Corollary33.3.3: any proof of upper closedness must choose one of the two closure branches, so refuting both branches separately refutes upper closedness itself.

lemma helperForCorollary33_3_3_not_upperClosed_of_branch_obstructions {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hLeft : ¬ (IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ concaveClosureInFirst (convexClosureInSecond K) = K)) (hRight : ¬ (IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K ∧ concaveClosureInSecond (convexClosureInFirst K) = K)) : ¬ IsUpperClosedSaddleFunction K

Lean source

Helper for Corollary33.3.3: the canonical witness is obtained by taking the sectionwise convex conjugate of the lower simple extension.

@[reducible] noncomputable def helperForCorollary33_3_3_canonicalWitness {m n : ℕ} (C : Set (Fin m → ℝ)) (D : Set (Fin n → ℝ)) (K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u x => convexConjugate (lowerSimpleExtensionOfReal C D K u) x

Lean source

Helper for Corollary33.3.3: the canonical witness has the textbook primal supremum formula.

lemma helperForCorollary33_3_3_canonicalWitness_primalFormula {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (u : Fin m → ℝ) (x : Fin n → ℝ) : helperForCorollary33_3_3_canonicalWitness C D K u x = if _ : u ∈ C then sSup (Set.range fun xStar : D => ↑(x ⬝ᵥ ↑xStar - K u ↑xStar)) else ⊤

Lean source

Helper for Corollary33.3.3: outside the primal constraint set C, the canonical witness section is constantly ⊤.

lemma helperForCorollary33_3_3_canonicalWitness_offParameterSection {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {u : Fin m → ℝ} (hu : u ∉ C) : helperForCorollary33_3_3_canonicalWitness C D K u = fun _ : Fin n → ℝ => (⊤ : EReal)

Lean source

Helper for Corollary33.3.3: the canonical witness never takes the value ⊥.

lemma helperForCorollary33_3_3_canonicalWitness_hasNoBotValues {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hD_nonempty : D.Nonempty) : HasNoBotValuesBifunction (helperForCorollary33_3_3_canonicalWitness C D K)

§33 — Saddle-Functions — Part 20

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Helper for Corollary33.3.3: any two bifunctions satisfying the displayed primal formula of the corollary must coincide.

lemma helperForCorollary33_3_3_unique_of_primalFormula {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {F G : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : ∀ (u : Fin m → ℝ) (x : Fin n → ℝ), F u x = if u ∈ C then sSup (Set.range fun xStar : D => ↑(x ⬝ᵥ ↑xStar - K u ↑xStar)) else ⊤) (hG : ∀ (u : Fin m → ℝ) (x : Fin n → ℝ), G u x = if u ∈ C then sSup (Set.range fun xStar : D => ↑(x ⬝ᵥ ↑xStar - K u ↑xStar)) else ⊤) : F = G

Lean source

Helper for Corollary33.3.3: the displayed primal formula already shows that the parameter domain cannot contain points outside C.

lemma helperForCorollary33_3_3_parameterDomain_subset_of_primalFormula {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : ∀ (u : Fin m → ℝ) (x : Fin n → ℝ), F u x = if u ∈ C then sSup (Set.range fun xStar : D => ↑(x ⬝ᵥ ↑xStar - K u ↑xStar)) else ⊤) : convexBifunctionParameterDomain F ⊆ C

Lean source

Helper for Corollary33.3.3: the displayed dual formula already shows that the genuine adjoint domain cannot contain points outside D.

lemma helperForCorollary33_3_3_genuineAdjointDomain_subset_of_dualFormula {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : ∀ (xStar : Fin n → ℝ) (uStar : Fin m → ℝ), genuineConvexBifunctionAdjoint F xStar uStar = if xStar ∈ D then sInf (Set.range fun u : C => ↑(↑u ⬝ᵥ uStar - K (↑u) xStar)) else ⊥) : {xStar | ∃ uStar, genuineConvexBifunctionAdjoint F xStar uStar ≠ ⊥} ⊆ D

Lean source

Helper for Corollary33.3.3: the canonical witness already satisfies the easy inclusion dom F ⊆ C coming from its explicit primal formula.

lemma helperForCorollary33_3_3_canonicalWitness_parameterDomain_subset {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} : convexBifunctionParameterDomain (helperForCorollary33_3_3_canonicalWitness C D K) ⊆ C

Lean source

Helper for Corollary33.3.3: once the represented primal kernel has exactly C as its non-⊥ slice, the primal pairing formula upgrades the easy inclusion dom F ⊆ C to the exact identity dom F = C.

lemma helperForCorollary33_3_3_parameterDomain_eq_of_primalPairing_and_nonbotSlice {m n : ℕ} {C : Set (Fin m → ℝ)} {K1 : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSliceDomain : {u : Fin m → ℝ | ∃ xStar : Fin n → ℝ, K1 u xStar ≠ ⊥} = C) (hParameterDomainSubset : convexBifunctionParameterDomain F ⊆ C) (hPrimalPairing : ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), K1 u xStar = convexBifunctionPairing F u xStar) : convexBifunctionParameterDomain F = C

Lean source

Helper for Corollary33.3.3: outside the primal constraint set C, the canonical witness already has the correct collapsed primal pairing, namely the everywhere-⊥ lower simple extension section.

lemma helperForCorollary33_3_3_canonicalWitness_primalPairing_eq_lowerSimpleExtension_off_domain {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {u : Fin m → ℝ} (hu : u ∉ C) : ∀ xStar : Fin n → ℝ, convexBifunctionPairing (helperForCorollary33_3_3_canonicalWitness C D K) u xStar = lowerSimpleExtensionOfReal C D K u xStar

Lean source

Helper for Corollary33.3.3: every section of the canonical witness is already convex-closed. On C this comes from the general closedness of convex conjugates, while off C the section is the constant ⊤ function.

lemma helperForCorollary33_3_3_canonicalWitness_sections_areFunctionConvexClosed {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hD_nonempty : D.Nonempty) : ∀ u : Fin m → ℝ, IsFunctionConvexClosed (helperForCorollary33_3_3_canonicalWitness C D K u)

Lean source

Helper for Corollary33.3.3: because every section of the canonical witness is convex-closed and avoids ⊥, its Chapter 2 convex closure agrees pointwise with the section itself.

lemma helperForCorollary33_3_3_canonicalWitness_sectionConvexClosure_exact {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hD_nonempty : D.Nonempty) : ∀ (u : Fin m → ℝ) (x : Fin n → ℝ), convexFunctionClosure (helperForCorollary33_3_3_canonicalWitness C D K u) x = helperForCorollary33_3_3_canonicalWitness C D K u x

Lean source

Helper for Corollary33.3.3: on the primal domain C, the frozen lower simple-extension section is convex and lower semicontinuous in the dual variable.

lemma helperForCorollary33_3_3_onDomain_lowerSimpleExtension_section_convexClosed {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hD_closed : IsClosed D) (hD_convex : Convex ℝ D) (hK_cont : ContinuousOn (fun p => K p.1 p.2) (C.prod D)) (hK_concaveConvex : IsConcaveConvexOn C D (erealOfRealBifunction K)) {u : Fin m → ℝ} (hu : u ∈ C) : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (lowerSimpleExtensionOfReal C D K u) ∧ LowerSemicontinuous (lowerSimpleExtensionOfReal C D K u)

Lean source

Helper for Corollary33.3.3: the canonical witness pairs back with the lower simple extension everywhere.

lemma helperForCorollary33_3_3_canonicalWitness_primalPairing_eq_lowerSimpleExtension {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hD_closed : IsClosed D) (hD_convex : Convex ℝ D) (hK_cont : ContinuousOn (fun p => K p.1 p.2) (C.prod D)) (hK_concaveConvex : IsConcaveConvexOn C D (erealOfRealBifunction K)) : ∀ u xStar, convexBifunctionPairing (helperForCorollary33_3_3_canonicalWitness C D K) u xStar = lowerSimpleExtensionOfReal C D K u xStar

Lean source

Helper for Corollary33.3.3: whenever a kernel K1 is represented as the primal pairing of F, the closure-side adjoint pairing is exactly the first-variable concave closure of K1.

lemma helperForCorollary33_3_3_closureSideAdjointPairing_eq_firstClosure {m n : ℕ} {K1 : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_graph : IsGraphConvexBifunction F) (hF_noBot : HasNoBotValuesBifunction F) (hPair : ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), K1 u xStar = convexBifunctionPairing F u xStar) : ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), convexBifunctionCanonicalAdjointPairing F xStar u = canonicalConcaveClosureInFirst K1 u xStar

Lean source

Helper for Corollary33.3.3: upper semicontinuity fixes the one-variable first-coordinate concave closure.

lemma helperForCorollary33_3_3_functionConcaveClosure_eq_self_of_upperSemicontinuous {m : ℕ} {g : (Fin m → ℝ) → EReal} (hUpper : UpperSemicontinuous g) : functionConcaveClosure g = g

Lean source

Helper for Corollary33.3.3: every frozen first-variable section of the lower simple extension is concave. This is the direct concavity input for the canonical witness's primal pairing, and it keeps the off-D branch separate instead of forcing an incorrect global closure identity with the upper simple extension.

lemma helperForCorollary33_3_3_lowerSimpleExtension_firstSections_areConcave {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_convex : Convex ℝ C) (hK_concaveConvex : IsConcaveConvexOn C D (erealOfRealBifunction K)) : ∀ xStar : Fin n → ℝ, IsERealConcaveOn (Set.univ : Set (Fin m → ℝ)) (fun u => lowerSimpleExtensionOfReal C D K u xStar)

Lean source

Helper for Corollary33.3.3: off the dual constraint set D, the frozen lower simple extension section is already upper semicontinuous, so its first-variable concave closure stays equal to the lower simple extension itself. This records the obstruction to the proposed global identity upperSimpleExtension = concaveClosureInFirst lowerSimpleExtension.

lemma helperForCorollary33_3_3_offDualDomain_lowerSimpleExtension_firstClosure_eq {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_closed : IsClosed C) {xStar : Fin n → ℝ} (hxStar : xStar ∉ D) : ∀ u : Fin m → ℝ, concaveClosureInFirst (lowerSimpleExtensionOfReal C D K) u xStar = lowerSimpleExtensionOfReal C D K u xStar

Lean source

Helper for Corollary33.3.3: on the dual constraint set D, the frozen lower simple extension section is already upper semicontinuous in the first variable, so its first-variable concave closure is exact.

lemma helperForCorollary33_3_3_onDualDomain_lowerSimpleExtension_firstClosure_eq {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_closed : IsClosed C) (hK_cont : ContinuousOn (fun p => K p.1 p.2) (C.prod D)) {xStar : Fin n → ℝ} (hxStar : xStar ∈ D) : ∀ u : Fin m → ℝ, concaveClosureInFirst (lowerSimpleExtensionOfReal C D K) u xStar = lowerSimpleExtensionOfReal C D K u xStar

Lean source

Helper for Corollary33.3.3: a globally concave frozen first-variable section is a Chapter 6 concave function, so the concave biconjugate theorem applies to it directly.

lemma helperForCorollary33_3_3_concaveFunction_of_firstSectionConcavity {m n : ℕ} {K1 : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {xStar : Fin n → ℝ} (hSection : IsERealConcaveOn (Set.univ : Set (Fin m → ℝ)) (fun u : Fin m → ℝ => K1 u xStar)) (hNoTop : ∀ u : Fin m → ℝ, K1 u xStar ≠ ⊤) : ConcaveFunction (fun u : Fin m → ℝ => K1 u xStar)

§33 — Saddle-Functions — Part 21

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Compatibility wrapper so the proof pipeline can target Corollary 33.3.3 in the earliest dependency-closed split file. The actual packaged theorem is reused later verbatim.

theorem «Corollary33.3.3» : ∀ {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ}, C.Nonempty → D.Nonempty → IsClosed C → IsClosed D → Convex ℝ C → Convex ℝ D → ContinuousOn (fun p => K p.1 p.2) (C.prod D) → IsConcaveConvexOn C D (erealOfRealBifunction K) → let K1 := lowerSimpleExtensionOfReal C D K let K2 := upperSimpleExtensionOfReal C D K IsLowerClosedSaddleFunction K1 ∧ ∃! F, IsRockafellarConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ IsFunctionConvexClosed (graphFunctionOfBifunction F) ∧ (∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), K1 u xStar = convexBifunctionPairing F u xStar) ∧ (∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), K2 u xStar = genuineConvexBifunctionAdjointPairing F u xStar) ∧ (∀ (u : Fin m → ℝ) (x : Fin n → ℝ), F u x = if _hu : u ∈ C then sSup (Set.range fun xStar : D => ↑(x ⬝ᵥ ↑xStar - K u ↑xStar)) else ⊤) ∧ (∀ (xStar : Fin n → ℝ) (uStar : Fin m → ℝ), genuineConvexBifunctionAdjoint F xStar uStar = if _hxStar : xStar ∈ D then sInf (Set.range fun u : C => ↑(↑u ⬝ᵥ uStar - K (↑u) xStar)) else ⊥) ∧ convexBifunctionParameterDomain F = C ∧ {xStar | ∃ uStar, genuineConvexBifunctionAdjoint F xStar uStar ≠ ⊥} = D

Lean source

Helper for Corollary33.3.3: the canonical witness's genuine adjoint has the textbook branchwise dual formula obtained by restricting the concave conjugate of the primal pairing to the primal domain C.

lemma helperForCorollary33_3_3_canonicalWitness_dualFormula_split {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_nonempty : C.Nonempty) (hF_primalPairing : ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), lowerSimpleExtensionOfReal C D K u xStar = convexBifunctionPairing (helperForCorollary33_3_3_canonicalWitness C D K) u xStar) : ∀ (xStar : Fin n → ℝ) (uStar : Fin m → ℝ), genuineConvexBifunctionAdjoint (helperForCorollary33_3_3_canonicalWitness C D K) xStar uStar = if _hxStar : xStar ∈ D then sInf (Set.range fun u : C => ↑(↑u ⬝ᵥ uStar - K (↑u) xStar)) else ⊥

Lean source

Helper for Corollary33.3.3: for a fixed xStar ∈ D, the affine graph integrand from the canonical witness formula is lower semicontinuous on ℝ^(m+n), with the u ∉ C branch collapsed to ⊤.

lemma helperForCorollary33_3_3_fixedDual_graphIntegrand_lowerSemicontinuous {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_closed : IsClosed C) (hK_cont : ContinuousOn (fun p => K p.1 p.2) (C.prod D)) (xStar : D) : LowerSemicontinuous (fun z : Fin (m + n) → ℝ => if (fun i : Fin m => z (Fin.castAdd n i)) ∈ C then ↑(((fun j : Fin n => z (Fin.natAdd m j)) ⬝ᵥ ↑xStar) - K (fun i : Fin m => z (Fin.castAdd n i)) ↑xStar) else (⊤ : EReal))

Lean source

Helper for Corollary33.3.3: the textbook primal formula rewrites the graph function as the pointwise supremum of the fixed-dual affine integrands indexed by D.

lemma helperForCorollary33_3_3_graphFunction_eq_iSup_fixedDualIntegrands {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hD_nonempty : D.Nonempty) (hF_primalFormula : ∀ (u : Fin m → ℝ) (x : Fin n → ℝ), F u x = if _hu : u ∈ C then sSup (Set.range fun xStar : D => ↑(x ⬝ᵥ ↑xStar - K u ↑xStar)) else (⊤ : EReal)) : let uProj : (Fin (m + n) → ℝ) → (Fin m → ℝ) := fun z i => z (Fin.castAdd n i) let xProj : (Fin (m + n) → ℝ) → (Fin n → ℝ) := fun z j => z (Fin.natAdd m j) let graphIntegrand : D → (Fin (m + n) → ℝ) → EReal := fun xStar z => if uProj z ∈ C then ↑((xProj z) ⬝ᵥ ↑xStar - K (uProj z) ↑xStar) else (⊤ : EReal) graphFunctionOfBifunction F = fun z => iSup (fun xStar : D => graphIntegrand xStar z)

Lean source

Helper for Corollary33.3.3: the graph function of the canonical witness is the pointwise supremum of the fixed-dual affine integrands, so lower semicontinuity follows by taking the supremum of the previously closed sublevel families.

lemma helperForCorollary33_3_3_canonicalWitness_graphClosed {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hD_nonempty : D.Nonempty) (hC_closed : IsClosed C) (hK_cont : ContinuousOn (fun p => K p.1 p.2) (C.prod D)) (hF_primalFormula : ∀ (u : Fin m → ℝ) (x : Fin n → ℝ), F u x = if _hu : u ∈ C then sSup (Set.range fun xStar : D => ↑(x ⬝ᵥ ↑xStar - K u ↑xStar)) else (⊤ : EReal)) : IsFunctionConvexClosed (graphFunctionOfBifunction F)

Lean source

Helper for Corollary33.3.3: once a bifunction pairing already identifies K1 with the Rockafellar pairing of a sectionwise convex witness and the first-variable concave closure of K1 is exact, the standard pairing correspondence upgrades K1 to a lower closed saddle function.

lemma helperForCorollary33_3_3_lowerClosed_of_pairingCorrespondence {m n : ℕ} {K1 : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_rock : IsRockafellarConvexBifunction F) (hF_noBot : HasNoBotValuesBifunction F) (hF_primalPairing : ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), K1 u xStar = convexBifunctionPairing F u xStar) (hK1_firstClosed : IsConcaveClosedInFirst K1) : IsLowerClosedSaddleFunction K1

Lean source

Helper for Corollary33.3.3: if xStar ∈ D, then the corresponding first section of the lower simple extension never takes the value ⊤. On C the section agrees with the finite kernel value, while off C the simple extension collapses to ⊥.

lemma helperForCorollary33_3_3_lowerSimpleExtension_firstSection_ne_top_onDualDomain {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {xStar : Fin n → ℝ} (hxStar : xStar ∈ D) : ∀ u : Fin m → ℝ, lowerSimpleExtensionOfReal C D K u xStar ≠ (⊤ : EReal)

Lean source

Helper for Corollary33.3.3: isolate the remaining specialized converse bridge for the canonical witness of the simple extensions. The dependency-closed part of the file already supplies the explicit primal formula, the no-⊥ property, and the exact non-⊥/non-⊤ slice domains of the two simple extensions.

lemma helperForCorollary33_3_3_canonicalWitnessPackage {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_nonempty : C.Nonempty) (hD_nonempty : D.Nonempty) (hC_closed : IsClosed C) (hD_closed : IsClosed D) (hC_convex : Convex ℝ C) (hD_convex : Convex ℝ D) (hK_cont : ContinuousOn (fun p => K p.1 p.2) (C.prod D)) (hK_concaveConvex : IsConcaveConvexOn C D (erealOfRealBifunction K)) : let K1 := lowerSimpleExtensionOfReal C D K let K2 := upperSimpleExtensionOfReal C D K let F := helperForCorollary33_3_3_canonicalWitness C D K IsLowerClosedSaddleFunction K1 ∧ IsRockafellarConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ IsFunctionConvexClosed (graphFunctionOfBifunction F) ∧ (∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), K1 u xStar = convexBifunctionPairing F u xStar) ∧ (∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), K2 u xStar = genuineConvexBifunctionAdjointPairing F u xStar) ∧ (∀ (u : Fin m → ℝ) (x : Fin n → ℝ), F u x = if _hu : u ∈ C then sSup (Set.range fun xStar : D => ↑(x ⬝ᵥ ↑xStar - K u ↑xStar)) else ⊤) ∧ (∀ (xStar : Fin n → ℝ) (uStar : Fin m → ℝ), genuineConvexBifunctionAdjoint F xStar uStar = if _hxStar : xStar ∈ D then sInf (Set.range fun u : C => ↑(↑u ⬝ᵥ uStar - K (↑u) xStar)) else ⊥) ∧ convexBifunctionParameterDomain F = C ∧ {xStar | ∃ uStar, genuineConvexBifunctionAdjoint F xStar uStar ≠ ⊥} = D

Lean source

Helper for Corollary33.3.3: once one bifunction satisfies the full textbook package, the displayed primal formula determines the witness uniquely.

lemma helperForCorollary33_3_3_existsUniqueWitness_of_fullPackage {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {K1 K2 : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF_rock : IsRockafellarConvexBifunction F) (hF_noBot : HasNoBotValuesBifunction F) (hF_graphClosed : IsFunctionConvexClosed (graphFunctionOfBifunction F)) (hF_primalPairing : ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), K1 u xStar = convexBifunctionPairing F u xStar) (hF_dualPairing : ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), K2 u xStar = genuineConvexBifunctionAdjointPairing F u xStar) (hF_primalFormula : ∀ (u : Fin m → ℝ) (x : Fin n → ℝ), F u x = if _hu : u ∈ C then sSup (Set.range fun xStar : D => ↑(x ⬝ᵥ ↑xStar - K u ↑xStar)) else ⊤) (hF_dualFormula : ∀ (xStar : Fin n → ℝ) (uStar : Fin m → ℝ), genuineConvexBifunctionAdjoint F xStar uStar = if _hxStar : xStar ∈ D then sInf (Set.range fun u : C => ↑(↑u ⬝ᵥ uStar - K (↑u) xStar)) else ⊥) (hF_dom : convexBifunctionParameterDomain F = C) (hF_adjDom : {xStar | ∃ uStar, genuineConvexBifunctionAdjoint F xStar uStar ≠ ⊥} = D) : ∃! G, IsRockafellarConvexBifunction G ∧ HasNoBotValuesBifunction G ∧ IsFunctionConvexClosed (graphFunctionOfBifunction G) ∧ (∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), K1 u xStar = convexBifunctionPairing G u xStar) ∧ (∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), K2 u xStar = genuineConvexBifunctionAdjointPairing G u xStar) ∧ (∀ (u : Fin m → ℝ) (x : Fin n → ℝ), G u x = if _hu : u ∈ C then sSup (Set.range fun xStar : D => ↑(x ⬝ᵥ ↑xStar - K u ↑xStar)) else ⊤) ∧ (∀ (xStar : Fin n → ℝ) (uStar : Fin m → ℝ), genuineConvexBifunctionAdjoint G xStar uStar = if _hxStar : xStar ∈ D then sInf (Set.range fun u : C => ↑(↑u ⬝ᵥ uStar - K (↑u) xStar)) else ⊥) ∧ convexBifunctionParameterDomain G = C ∧ {xStar | ∃ uStar, genuineConvexBifunctionAdjoint G xStar uStar ≠ ⊥} = D

§33 — Saddle-Functions — Part 22

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

theorem isFullyClosedSaddleFunction_iff_isLowerClosedSaddleFunction_and_isUpperClosedSaddleFunction : ∀ {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal}, IsFullyClosedSaddleFunction K → IsLowerClosedSaddleFunction K ∧ IsUpperClosedSaddleFunction K

Lean source

Corollary 33.1.2, packaged in its complete convex/concave four-way form.

theorem lowerClosedConcaveConvex_and_upperClosedConvexConcave_pairing_correspondence {m n : ℕ} : (∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsImageClosedConvexBifunction F → IsConcaveConvexOn Set.univ Set.univ (convexBifunctionPairing F) ∧ IsConvexClosedInSecond (convexBifunctionPairing F) ∧ ∀ u x, convexConjugate (convexBifunctionPairing F u) x = F u x) ∧ (∀ K : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsConcaveConvexOn Set.univ Set.univ K → IsConvexClosedInSecond K → HasNoTopOrBotValuesBifunction K → let F : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u x => convexConjugate (K u) x IsImageClosedConvexBifunction F ∧ (∀ u x, F u x = convexConjugate (K u) x) ∧ ∀ u xStar, convexBifunctionPairing F u xStar = K u xStar) ∧ (∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsImageClosedConcaveBifunction F → IsConvexConcaveOn Set.univ Set.univ (concaveBifunctionPairing F) ∧ IsConcaveClosedInSecond (concaveBifunctionPairing F) ∧ ∀ u x, concaveConjugate (concaveBifunctionPairing F u) x = F u x) ∧ (∀ K : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsConvexConcaveOn Set.univ Set.univ K → IsConcaveClosedInSecond K → HasNoTopOrBotValuesBifunction K → let F : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u x => concaveConjugate (K u) x IsImageClosedConcaveBifunction F ∧ (∀ u x, F u x = concaveConjugate (K u) x) ∧ ∀ u xStar, concaveBifunctionPairing F u xStar = K u xStar)

Lean source

Corollary33.3.1, sound forward direction: a unique closed convex witness gives the canonical coordinatewise closure pair.

theorem closed_convex_bifunction_exists_implies_coordinatewise_closure_pair : ∀ {m n : ℕ} {K Kbar : (Fin m → ℝ) → (Fin n → ℝ) → EReal}, (∃! F, ClosedConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ (∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), K u xStar = convexBifunctionPairing F u xStar) ∧ ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), Kbar u xStar = convexBifunctionCanonicalAdjointPairing F xStar u) → Kbar = canonicalConcaveClosureInFirst K ∧ canonicalConvexClosureInSecond Kbar = K

Lean source

Helper for Corollary33.3.1: the mixed coordinatewise closure identities already force the announced lower-closedness, upper-closedness, and pointwise order conclusions.

lemma helperForCorollary33_3_1_coordinatewise_closure_pair_implies_closedness_and_order {m n : ℕ} {K Kbar : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsConcaveConvexOn Set.univ Set.univ K) (hKbar : IsConcaveConvexOn Set.univ Set.univ Kbar) (hPair : Kbar = concaveClosureInFirst K ∧ convexClosureInSecond Kbar = K) : IsLowerClosedSaddleFunction K ∧ IsUpperClosedSaddleFunction Kbar ∧ ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), K u xStar ≤ Kbar u xStar

Lean source

Helper for Corollary33.3.1: the closure-pair identities already force the primal kernel to be convex-closed in the second variable. This is the exact kernel-side input needed for the Section 33 image-closed reconstruction, independent of any graph-closed witness theorem.

lemma helperForCorollary33_3_1_coordinatewise_closure_pair_implies_secondClosed {m n : ℕ} {K Kbar : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hPair : Kbar = concaveClosureInFirst K ∧ convexClosureInSecond Kbar = K) : IsConvexClosedInSecond K

Lean source

Helper for Corollary33.3.1: once the closure pair is known and K avoids ⊤, ⊥, the Section 33 image-closed correspondence already reconstructs the canonical convex witness and its primal pairing formula. The only remaining extra input for the fully closed witness package is the graph-function closedness bridge handled separately later.

lemma helperForCorollary33_3_1_imageClosedWitness_of_coordinatewise_closure_pair_of_noTopOrBot {m n : ℕ} {K Kbar : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsConcaveConvexOn Set.univ Set.univ K) (hPair : Kbar = concaveClosureInFirst K ∧ convexClosureInSecond Kbar = K) (hNoTopBot : HasNoTopOrBotValuesBifunction K) : let F : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u x => convexConjugate (K u) x IsImageClosedConvexBifunction F ∧ ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ), convexBifunctionPairing F u xStar = K u xStar

Lean source

Helper for Corollary33.3.1: on the zero-dimensional space, a bifunction with no ⊥ values cannot have primal pairing ⊤ at the unique point.

lemma helperForCorollary33_3_1_zeroDimensional_pairing_ne_top_of_noBot {F : (Fin 0 → ℝ) → (Fin 0 → ℝ) → EReal} (hNoBot : HasNoBotValuesBifunction F) : convexBifunctionPairing F 0 0 ≠ (⊤ : EReal)

Lean source

Helper for Corollary33.3.1: the singleton constant-⊤ kernel satisfies the closure-pair side, but the witness-side primal pairing equation is already impossible under HasNoBotValuesBifunction.

lemma helperForCorollary33_3_1_zeroDimensional_constTop_counterexample : let K : (Fin 0 → ℝ) → (Fin 0 → ℝ) → EReal := fun _ _ => (⊤ : EReal) (K = concaveClosureInFirst K ∧ convexClosureInSecond K = K) ∧ ¬ ∃ F : (Fin 0 → ℝ) → (Fin 0 → ℝ) → EReal, HasNoBotValuesBifunction F ∧ (∀ (u : Fin 0 → ℝ) (xStar : Fin 0 → ℝ), K u xStar = convexBifunctionPairing F u xStar)

Lean source

Helper for Corollary33.3.1: in dimension zero, the primal pairing is just the negation of the unique bifunction value.

lemma helperForCorollary33_3_1_zeroDimensional_pairing_eq_neg_value (F : (Fin 0 → ℝ) → (Fin 0 → ℝ) → EReal) : convexBifunctionPairing F 0 0 = -(F 0 0)

Lean source

Helper for Corollary33.3.1: in dimension zero, the corrected adjoint pairing has the same sign as the primal pairing.

lemma helperForCorollary33_3_1_zeroDimensional_adjointPairing_eq_neg_value (F : (Fin 0 → ℝ) → (Fin 0 → ℝ) → EReal) : convexBifunctionCanonicalAdjointPairing F 0 0 = -(F 0 0)

Lean source

Helper for Corollary33.3.1: the zero-dimensional constant-⊤ pair satisfies the coordinatewise closure identities, but no closed convex bifunction can realize both pairing identities simultaneously.

lemma helperForCorollary33_3_1_zeroDimensional_constTop_closedConvex_counterexample : let K : (Fin 0 → ℝ) → (Fin 0 → ℝ) → EReal := fun _ _ => (⊤ : EReal) (K = concaveClosureInFirst K ∧ convexClosureInSecond K = K) ∧ ¬ ∃ F : (Fin 0 → ℝ) → (Fin 0 → ℝ) → EReal, ClosedConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ (∀ (u : Fin 0 → ℝ) (xStar : Fin 0 → ℝ), K u xStar = convexBifunctionPairing F u xStar) ∧ ∀ (u : Fin 0 → ℝ) (xStar : Fin 0 → ℝ), K u xStar = convexBifunctionCanonicalAdjointPairing F xStar u

Lean source

Helper for Corollary33.3.2: an upper semicontinuous section is already fixed by the one-variable Section 33 concave closure.

lemma helperForCorollary33_3_2_functionConcaveClosure_eq_self_of_upperSemicontinuous {m : ℕ} {g : (Fin m → ℝ) → EReal} (hUpper : UpperSemicontinuous g) : functionConcaveClosure g = g

Lean source

Helper for Corollary33.3.2: sectionwise upper semicontinuity in the first variable forces the first-variable concave closure to fix the kernel.

lemma helperForCorollary33_3_2_concaveClosureInFirst_eq_self_of_sectionUpperSemicontinuous {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hUpper : ∀ xStar : Fin n → ℝ, UpperSemicontinuous (fun u => K u xStar)) : concaveClosureInFirst K = K

Lean source

Helper for Corollary33.3.2: sectionwise lower semicontinuity in the second variable forces the second-variable convex closure to fix the kernel.

lemma helperForCorollary33_3_2_convexClosureInSecond_eq_self_of_sectionLowerSemicontinuous {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hLower : ∀ u : Fin m → ℝ, LowerSemicontinuous (fun xStar => K u xStar)) : convexClosureInSecond K = K

Lean source

Helper for Corollary33.3.2: a simultaneously convex-and-concave section with one ⊥ value must be identically ⊥.

lemma helperForCorollary33_3_2_section_eq_bot_everywhere_of_exists_bot {m : ℕ} {g : (Fin m → ℝ) → EReal} (hConv : IsERealConvexOn (Set.univ : Set (Fin m → ℝ)) g) {u0 : Fin m → ℝ} (hu0 : g u0 = ⊥) : ∀ u : Fin m → ℝ, g u = ⊥

Lean source

Helper for Corollary33.3.2: once ⊥ is excluded, a simultaneously convex-and-concave section with one ⊤ value must be identically ⊤.

lemma helperForCorollary33_3_2_section_eq_top_everywhere_of_exists_top {m : ℕ} {g : (Fin m → ℝ) → EReal} (hConc : IsERealConcaveOn (Set.univ : Set (Fin m → ℝ)) g) (hNoBot : ∀ u : Fin m → ℝ, g u ≠ ⊥) {u0 : Fin m → ℝ} (hu0 : g u0 = ⊤) : ∀ u : Fin m → ℝ, g u = ⊤

Lean source

Helper for Corollary33.3.2: a one-variable section that is both convex and concave on all of ℝ^m is automatically both upper and lower semicontinuous.

lemma helperForCorollary33_3_2_upperLowerSemicontinuous_of_simultaneousConvexConcave {m : ℕ} {g : (Fin m → ℝ) → EReal} (hConc : IsERealConcaveOn (Set.univ : Set (Fin m → ℝ)) g) (hConv : IsERealConvexOn (Set.univ : Set (Fin m → ℝ)) g) : UpperSemicontinuous g ∧ LowerSemicontinuous g

Lean source

Helper for Corollary33.3.2: under the concave-convex hypothesis, the lower-closed predicate reduces to the aligned identity cl₂ (cl₁ K) = K.

lemma helperForCorollary33_3_2_simultaneousOrientations_fix_concaveClosureInFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsConcaveConvexOn Set.univ Set.univ K) (hVC : IsConvexConcaveOn Set.univ Set.univ K) : concaveClosureInFirst K = K

§33 — Saddle-Functions — Part 23

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

lemma helperForCorollary33_3_2_upperClosed_concaveConvex_extract_upperComposition {m n : ℕ} {Kbar : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hKbar : IsConcaveConvexOn Set.univ Set.univ Kbar) (hUpper : IsUpperClosedSaddleFunction Kbar) : concaveClosureInFirst (convexClosureInSecond Kbar) = Kbar

Lean source

theorem lowerClosed_concaveConvex_upperClosed_concaveConvex_closure_correspondence : ∀ {m n : ℕ}, (∀ (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal), IsConcaveConvexOn Set.univ Set.univ K → HasNoBotValuesBifunction K → IsLowerClosedSaddleFunction K → ∃! Kbar, IsConcaveConvexOn Set.univ Set.univ Kbar ∧ IsUpperClosedSaddleFunction Kbar ∧ Kbar = concaveClosureInFirst K ∧ convexClosureInSecond Kbar = K) ∧ ∀ (Kbar : (Fin m → ℝ) → (Fin n → ℝ) → EReal), IsConcaveConvexOn Set.univ Set.univ Kbar → HasNoBotValuesBifunction Kbar → IsUpperClosedSaddleFunction Kbar → ∃! K, IsConcaveConvexOn Set.univ Set.univ K ∧ IsLowerClosedSaddleFunction K ∧ Kbar = concaveClosureInFirst K ∧ convexClosureInSecond Kbar = K

Lean source

Negating a concave section produces a convex one as soon as the original section avoids ⊤. This is the one-variable sign transport needed for the corrected reverse bridge.

lemma helperForCorollary33_3_2_concaveNegation_isConvex_of_noTop {k : ℕ} {f : (Fin k → ℝ) → EReal} (hConc : IsERealConcaveOn (Set.univ : Set (Fin k → ℝ)) f) (hNoTop : ∀ x, f x ≠ ⊤) : IsERealConvexOn (Set.univ : Set (Fin k → ℝ)) (fun x => -f x)

Lean source

Negating the first-variable concave closure converts it into the corresponding first-variable convex closure.

lemma helperForCorollary33_3_2_neg_concaveClosureInFirst_eq_convexClosureInFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} : concaveClosureInFirst (fun u x => -K u x) = fun u x => -convexClosureInFirst K u x

Lean source

Negating the second-variable convex closure converts it into the corresponding second-variable concave closure.

lemma helperForCorollary33_3_2_neg_convexClosureInSecond_eq_concaveClosureInSecond {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} : convexClosureInSecond (fun u x => -K u x) = fun u x => -concaveClosureInSecond K u x

Lean source

After sign reversal, an upper-closed convex-concave kernel becomes a lower-closed concave-convex kernel.

lemma helperForCorollary33_3_2_negated_upperClosedConvexConcave_isLowerClosed {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsConvexConcaveOn Set.univ Set.univ K) (hUpper : IsUpperClosedSaddleFunction K) (hNoTop : HasNoTopValuesBifunction K) : IsLowerClosedSaddleFunction (fun u x => -K u x)

Lean source

theorem finiteContinuous_concaveConvex_simpleExtensions_yield_closed_convex_bifunction : ∀ {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ}, C.Nonempty → D.Nonempty → IsClosed C → IsClosed D → Convex ℝ C → Convex ℝ D → ContinuousOn (fun p => K p.1 p.2) (C.prod D) → IsConcaveConvexOn C D (erealOfRealBifunction K) → have K1

Lean source

Helper for Corollary33.3.2: simultaneous concave-convex and convex-concave structure forces the second-variable convex closure to fix the kernel. This is the symmetric one-step closure fact used to collapse the Or.inr branch in Corollary33.3.2.

lemma helperForCorollary33_3_2_simultaneousOrientations_fix_convexClosureInSecond {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsConcaveConvexOn Set.univ Set.univ K) (hVC : IsConvexConcaveOn Set.univ Set.univ K) : convexClosureInSecond K = K

Lean source

Helper for Corollary33.3.2: under the concave-convex hypothesis, the lower-closed predicate reduces to the aligned identity cl₂ (cl₁ K) = K.

lemma helperForCorollary33_3_2_lowerClosed_concaveConvex_extract_lowerComposition {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsConcaveConvexOn Set.univ Set.univ K) (hLower : IsLowerClosedSaddleFunction K) : convexClosureInSecond (concaveClosureInFirst K) = K

Lean source

Helper for Corollary33.3.2: under simultaneous concave-convex and convex-concave orientations, the first-variable convex closure is already exact.

lemma helperForCorollary33_3_2_simultaneousOrientations_fix_convexClosureInFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hCC : IsConcaveConvexOn Set.univ Set.univ K) (hVC : IsConvexConcaveOn Set.univ Set.univ K) : convexClosureInFirst K = K

Lean source

Helper for Corollary33.3.2: under simultaneous concave-convex and convex-concave orientations, the second-variable concave closure is already exact.

lemma helperForCorollary33_3_2_simultaneousOrientations_fix_concaveClosureInSecond {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hCC : IsConcaveConvexOn Set.univ Set.univ K) (hVC : IsConvexConcaveOn Set.univ Set.univ K) : concaveClosureInSecond K = K

Lean source

Helper for Corollary33.3.2: under the convex-concave hypothesis, the upper-closed predicate reduces to the aligned identity cl₂^c (cl₁^v K) = K.

lemma helperForCorollary33_3_2_upperClosed_convexConcave_extract_upperComposition {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsConvexConcaveOn Set.univ Set.univ K) (hUpper : IsUpperClosedSaddleFunction K) : concaveClosureInSecond (convexClosureInFirst K) = K

§34 — Closures and Equivalence Classes — Part 1

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Predicates on coordinate spaces in this section are treated classically when needed in piecewise definitions.

open Set noncomputable local instance classicalDecidablePred {α : Type*} (p : α → Prop) : DecidablePred p

Lean source

A saddle-function on ℝ^m × ℝ^n, represented as an extended-real-valued bifunction.

open Set abbrev SaddleFunction (m n : ℕ) := (Fin m → ℝ) → (Fin n → ℝ) → EReal

Lean source

The partial closure in the first argument, realized by the first-variable concave closure from Section 33.

open Set variable {m n : ℕ} noncomputable def partialClosure₁ (K : SaddleFunction m n) : SaddleFunction m n := concaveClosureInFirst K

Lean source

The partial closure in the second argument, realized by the second-variable convex closure from Section 33.

open Set variable {m n : ℕ} noncomputable def partialClosure₂ (K : SaddleFunction m n) : SaddleFunction m n := convexClosureInSecond K

Lean source

A saddle-function is concave-convex when it is concave in the first variable and convex in the second variable on all of ℝ^m × ℝ^n.

open Set variable {m n : ℕ} def IsConcaveConvex (K : SaddleFunction m n) : Prop := IsConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K

Lean source

A saddle-function is convex-concave when it is convex in the first variable and concave in the second variable on all of ℝ^m × ℝ^n.

open Set variable {m n : ℕ} def IsConvexConcave (K : SaddleFunction m n) : Prop := IsConvexConcaveOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K

Lean source

A witness that a saddle-function is of one of the two orientations used in Section 34.

open Set variable {m n : ℕ} inductive SaddleFunctionType (K : SaddleFunction m n) : Type where /-- The saddle-function is concave in the first variable and convex in the second. -/ | concaveConvex (h : IsConcaveConvex K) /-- The saddle-function is convex in the first variable and concave in the second. -/ | convexConcave (h : IsConvexConcave K)

Lean source

The first partial closure attached to the chosen saddle orientation of K.

open Set variable {m n : ℕ} noncomputable def partialClosure₁OfType (K : SaddleFunction m n) (hK : SaddleFunctionType K) : SaddleFunction m n → SaddleFunction m n := match hK with | .concaveConvex _ => concaveClosureInFirst | .convexConcave _ => fun L u v => ⨆ (ε : {ε : ℝ // 0 < ε}), ⨅ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), L w.1 v

Lean source

The second partial closure attached to the chosen saddle orientation of K.

open Set variable {m n : ℕ} noncomputable def partialClosure₂OfType (K : SaddleFunction m n) (hK : SaddleFunctionType K) : SaddleFunction m n → SaddleFunction m n := match hK with | .concaveConvex _ => convexClosureInSecond | .convexConcave _ => fun L u v => ⨅ (ε : {ε : ℝ // 0 < ε}), ⨆ (w : {w : Fin n → ℝ // ‖w - v‖ < ε.1}), L u w.1

Lean source

Defn 34.1: the lower and upper closures of a saddle-function are obtained by the orientation-dependent iterated partial closures, so both the closure type and the order are determined by the saddle orientation of K.

open Set variable {m n : ℕ} noncomputable def lowerUpperClosurePair (K : SaddleFunction m n) (hK : SaddleFunctionType K) : SaddleFunction m n × SaddleFunction m n := match hK with | .concaveConvex _ => ((partialClosure₂OfType K hK) ((partialClosure₁OfType K hK) K), (partialClosure₁OfType K hK) ((partialClosure₂OfType K hK) K)) | .convexConcave _ => ((partialClosure₁OfType K hK) ((partialClosure₂OfType K hK) K), (partialClosure₂OfType K hK) ((partialClosure₁OfType K hK) K))

Lean source

The lower closure of a saddle-function in its chosen orientation.

open Set variable {m n : ℕ} noncomputable def saddleLowerClosure (K : SaddleFunction m n) (hK : SaddleFunctionType K) : SaddleFunction m n := (lowerUpperClosurePair K hK).1

Lean source

The upper closure of a saddle-function in its chosen orientation.

open Set variable {m n : ℕ} noncomputable def saddleUpperClosure (K : SaddleFunction m n) (hK : SaddleFunctionType K) : SaddleFunction m n := (lowerUpperClosurePair K hK).2

Lean source

Lower closure in the concave-convex case.

open Set variable {m n : ℕ} noncomputable def lowerClosureConcaveConvex (K : SaddleFunction m n) (h : IsConcaveConvex K) : SaddleFunction m n := (lowerUpperClosurePair K (.concaveConvex h)).1

Lean source

Upper closure in the concave-convex case.

open Set variable {m n : ℕ} noncomputable def upperClosureConcaveConvex (K : SaddleFunction m n) (h : IsConcaveConvex K) : SaddleFunction m n := (lowerUpperClosurePair K (.concaveConvex h)).2

Lean source

Lower closure in the convex-concave case.

open Set variable {m n : ℕ} noncomputable def lowerClosureConvexConcave (K : SaddleFunction m n) (h : IsConvexConcave K) : SaddleFunction m n := (lowerUpperClosurePair K (.convexConcave h)).1

Lean source

Upper closure in the convex-concave case.

open Set variable {m n : ℕ} noncomputable def upperClosureConvexConcave (K : SaddleFunction m n) (h : IsConvexConcave K) : SaddleFunction m n := (lowerUpperClosurePair K (.convexConcave h)).2

Lean source

Defn 34.2: a concave-convex saddle-function is (1) lower closed when it coincides with its lower closure and (2) upper closed when it coincides with its upper closure.

open Set variable {m n : ℕ} def IsLowerClosed (K : SaddleFunction m n) : Prop := IsConcaveConvex K ∧ ∀ h : IsConcaveConvex K, K = lowerClosureConcaveConvex K h

Lean source

A concave-convex saddle-function is upper closed when it coincides with its upper closure.

open Set variable {m n : ℕ} def IsUpperClosed (K : SaddleFunction m n) : Prop := IsConcaveConvex K ∧ ∀ h : IsConcaveConvex K, K = upperClosureConcaveConvex K h

Lean source

Defn 34.4: two concave-convex saddle-functions K and L are equivalent when they share the same partial closures, namely cl₁ K = cl₁ L and cl₂ K = cl₂ L.

open Set variable {m n : ℕ} def saddleEquivalent (K L : SaddleFunction m n) : Prop := IsConcaveConvex K ∧ IsConcaveConvex L ∧ partialClosure₁ K = partialClosure₁ L ∧ partialClosure₂ K = partialClosure₂ L

Lean source

Defn 34.5: a concave-convex saddle-function is closed when it is equivalent to each of its two partial closures, equivalently when K ∼ cl₁ K and K ∼ cl₂ K.

open Set variable {m n : ℕ} def IsClosedSaddleFunction (K : SaddleFunction m n) : Prop := saddleEquivalent K (partialClosure₁ K) ∧ saddleEquivalent K (partialClosure₂ K)

Lean source

The effective domain in the first argument consists of those u for which K u v is never -∞, uniformly in v.

open Set variable {m n : ℕ} def effectiveDomain₁ (K : SaddleFunction m n) : Set (Fin m → ℝ) := {u | ∀ v, K u v > (⊥ : EReal)}

Lean source

The effective domain in the second argument consists of those v for which K u v is never +∞, uniformly in u.

open Set variable {m n : ℕ} def effectiveDomain₂ (K : SaddleFunction m n) : Set (Fin n → ℝ) := {v | ∀ u, K u v < (⊤ : EReal)}

Lean source

Defn 34.3: for a concave-convex function K on ℝ^m × ℝ^n, the effective domains dom₁ K and dom₂ K are the sets where the first and second arguments avoid -∞ and +∞ respectively for all values of the other variable, and the saddle effective domain dom K is their product.

open Set variable {m n : ℕ} def saddleEffectiveDomain (K : SaddleFunction m n) : Set ((Fin m → ℝ) × (Fin n → ℝ)) := effectiveDomain₁ K ×ˢ effectiveDomain₂ K

Lean source

The finiteness domain of a saddle-function consists of the pairs where it takes a real value, equivalently neither +∞ nor -∞.

open Set variable {m n : ℕ} def finitenessDomain (K : SaddleFunction m n) : Set ((Fin m → ℝ) × (Fin n → ℝ)) := {p | K p.1 p.2 ≠ (⊤ : EReal) ∧ K p.1 p.2 ≠ (⊥ : EReal)}

Lean source

A saddle-function is proper when its effective domain is nonempty.

open Set variable {m n : ℕ} def IsProperSaddleFunction (K : SaddleFunction m n) : Prop := saddleEffectiveDomain K ≠ (∅ : Set ((Fin m → ℝ) × (Fin n → ℝ)))

Lean source

The effective domain of a convex EReal-valued function is the set where it is strictly below +∞.

open Set variable {m n : ℕ} def convexFunctionEffectiveDomain {n : ℕ} (f : (Fin n → ℝ) → EReal) : Set (Fin n → ℝ) := {x | f x < (⊤ : EReal)}

Lean source

The effective domain of a concave EReal-valued function is the set where it is strictly above -∞.

open Set variable {m n : ℕ} def concaveFunctionEffectiveDomain {n : ℕ} (f : (Fin n → ℝ) → EReal) : Set (Fin n → ℝ) := {x | (⊥ : EReal) < f x}

Lean source

A closed proper convex function with effective domain exactly D.

open Set variable {m n : ℕ} def IsProperClosedConvexFunctionWithDomain {n : ℕ} (f : (Fin n → ℝ) → EReal) (D : Set (Fin n → ℝ)) : Prop := IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) f ∧ IsFunctionConvexClosed f ∧ (∀ x, f x ≠ (⊥ : EReal)) ∧ convexFunctionEffectiveDomain f = D

Lean source

A proper convex function whose effective domain lies between D and E.

open Set variable {m n : ℕ} def IsProperConvexFunctionWithDomainBetween {n : ℕ} (f : (Fin n → ℝ) → EReal) (D E : Set (Fin n → ℝ)) : Prop := IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) f ∧ (∀ x, f x ≠ (⊥ : EReal)) ∧ (convexFunctionEffectiveDomain f).Nonempty ∧ D ⊆ convexFunctionEffectiveDomain f ∧ convexFunctionEffectiveDomain f ⊆ E

Lean source

An improper convex function in Rockafellar's convention: it is convex but not proper.

open Set variable {m n : ℕ} def IsImproperConvexFunction {n : ℕ} (f : (Fin n → ℝ) → EReal) : Prop := IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) f ∧ ¬ ((∀ x, f x ≠ (⊥ : EReal)) ∧ (convexFunctionEffectiveDomain f).Nonempty)

Lean source

A closed proper concave function with effective domain exactly C.

open Set variable {m n : ℕ} def IsProperClosedConcaveFunctionWithDomain {m : ℕ} (f : (Fin m → ℝ) → EReal) (C : Set (Fin m → ℝ)) : Prop := IsERealConcaveOn (Set.univ : Set (Fin m → ℝ)) f ∧ IsFunctionConcaveClosed f ∧ (∀ x, f x ≠ (⊤ : EReal)) ∧ concaveFunctionEffectiveDomain f = C

Lean source

A proper concave function whose effective domain lies between C and E.

open Set variable {m n : ℕ} def IsProperConcaveFunctionWithDomainBetween {m : ℕ} (f : (Fin m → ℝ) → EReal) (C E : Set (Fin m → ℝ)) : Prop := IsERealConcaveOn (Set.univ : Set (Fin m → ℝ)) f ∧ (∀ x, f x ≠ (⊤ : EReal)) ∧ (concaveFunctionEffectiveDomain f).Nonempty ∧ C ⊆ concaveFunctionEffectiveDomain f ∧ concaveFunctionEffectiveDomain f ⊆ E

Lean source

An improper concave function in Rockafellar's convention: it is concave but not proper.

open Set variable {m n : ℕ} def IsImproperConcaveFunction {m : ℕ} (f : (Fin m → ℝ) → EReal) : Prop := IsERealConcaveOn (Set.univ : Set (Fin m → ℝ)) f ∧ ¬ ((∀ x, f x ≠ (⊤ : EReal)) ∧ (concaveFunctionEffectiveDomain f).Nonempty)

Lean source

Defn 34.6: for a closed convex bifunction F, Ω(F) is the set of all concave-convex saddle-functions equivalent to the kernel (u, xStar) ↦ ⟪F u, xStar⟫.

open Set variable {m n : ℕ} def EquivalenceClassGeneratedByConvexBifunction (F : {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal // IsImageClosedConvexBifunction F}) : Set (SaddleFunction m n) := {K | saddleEquivalent K (convexBifunctionPairing F.1)}

Lean source

The concave analogue of the equivalence class generated by a closed concave bifunction G, consisting of the concave-convex saddle-functions equivalent to (u, xStar) ↦ ⟪u, G xStar⟫.

open Set variable {m n : ℕ} def EquivalenceClassGeneratedByConcaveBifunction (G : {G : (Fin n → ℝ) → (Fin m → ℝ) → EReal // IsImageClosedConcaveBifunction G}) : Set (SaddleFunction m n) := {K | saddleEquivalent K (fun u xStar => concaveBifunctionPairing G.1 xStar u)}

Lean source

The relative-interior effective domain ri (dom K) = ri (dom₁ K) × ri (dom₂ K) of a saddle-function K.

open Set variable {m n : ℕ} def saddleKernelDomain (K : SaddleFunction m n) : Set ((Fin m → ℝ) × (Fin n → ℝ)) := intrinsicInterior ℝ (effectiveDomain₁ K) ×ˢ intrinsicInterior ℝ (effectiveDomain₂ K)

Lean source

The raw restriction of a bifunction to ri (dom₁ K) × ri (dom₂ K).

open Set variable {m n : ℕ} def saddleKernelRestriction (K : SaddleFunction m n) : saddleKernelDomain K → EReal := fun p => K p.1.1 p.1.2

Lean source

Defn 34.7: if K is a genuine saddle-function on ℝ^m × ℝ^n, then its kernel is the restriction of K to the relative interior of its effective domain, namely to ri (dom K) = ri (dom₁ K) × ri (dom₂ K).

open Set variable {m n : ℕ} def saddleKernel (K : SaddleFunction m n) (_hK : IsSaddleFunctionOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K) : saddleKernelDomain K → EReal := saddleKernelRestriction K -- Proof sketch: use the previously established fact that a genuine saddle-function is finite on -- its effective domain, and then note that `ri (dom₁ K) × ri (dom₂ K)` is contained in -- `dom₁ K × dom₂ K` by `intrinsicInterior_subset` in each coordinate.

Lean source

The kernel of a genuine saddle-function is finite-valued on its relative-interior domain.

open Set variable {m n : ℕ} theorem saddleKernel_finite (K : SaddleFunction m n) (hK : IsSaddleFunctionOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) K) : ∀ p : saddleKernelDomain K, saddleKernel K hK p ≠ (⊤ : EReal) ∧ saddleKernel K hK p ≠ (⊥ : EReal)

Lean source

Two saddle-functions have the same kernel when their relative-interior kernel domains coincide and the restricted bifunctions agree on that common domain.

open Set variable {m n : ℕ} def HasSameSaddleKernel (K L : SaddleFunction m n) : Prop := ∃ hdom : saddleKernelDomain K = saddleKernelDomain L, ∀ p : saddleKernelDomain K, saddleKernelRestriction K p = saddleKernelRestriction L (hdom ▸ p) -- Proof sketch: unfold the two partial closure operators and apply the corresponding idempotence -- statements from the one-variable closure theory in Section 33.

Lean source

The mixed lower closure inherits the no-⊥ convention from the original kernel.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_lowerClosure_noBot (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBot : HasNoBotValuesBifunction K) : HasNoBotValuesBifunction (lowerClosureConcaveConvex K h)

Lean source

The mixed upper closure inherits the no-⊥ convention from the original kernel.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_upperClosure_noBot (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBot : HasNoBotValuesBifunction K) : HasNoBotValuesBifunction (upperClosureConcaveConvex K h)

Lean source

Helper for Text 34.0.1: the concave-convex lower and upper closures do not depend on which proof of IsConcaveConvex is supplied.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_concaveConvex_branch_proofIrrelevance (K : SaddleFunction m n) (h1 h2 : IsConcaveConvex K) : lowerClosureConcaveConvex K h1 = lowerClosureConcaveConvex K h2 ∧ upperClosureConcaveConvex K h1 = upperClosureConcaveConvex K h2

Lean source

Helper for Text 34.0.1: in the concave-convex branch, the lower and upper closures are exactly the mixed coordinatewise closures cl₂ (cl₁ K) and cl₁ (cl₂ K).

open Set variable {m n : ℕ} lemma helperForText_34_0_1_mixedClosure_formulas (K : SaddleFunction m n) (h : IsConcaveConvex K) : lowerClosureConcaveConvex K h = partialClosure₂ (partialClosure₁ K) ∧ upperClosureConcaveConvex K h = partialClosure₁ (partialClosure₂ K)

Lean source

Helper for Text 34.0.1: the mixed lower and upper closures stay concave-convex, and each retains the one-sided closedness built into its outermost coordinatewise closure.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_mixedClosure_orientation_and_oneSidedClosedness (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBot : HasNoBotValuesBifunction K) : IsConcaveConvex (lowerClosureConcaveConvex K h) ∧ IsConcaveConvex (upperClosureConcaveConvex K h) ∧ IsConvexClosedInSecond (lowerClosureConcaveConvex K h) ∧ IsConcaveClosedInFirst (upperClosureConcaveConvex K h)

Lean source

Helper for Text 34.0.1: the outer coordinatewise closure already fixes each mixed closure.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_outerClosure_fixedPoint_forms (K : SaddleFunction m n) (h : IsConcaveConvex K) : partialClosure₂ (lowerClosureConcaveConvex K h) = lowerClosureConcaveConvex K h ∧ partialClosure₁ (upperClosureConcaveConvex K h) = upperClosureConcaveConvex K h

Lean source

Helper for Text 34.0.1: every saddle-function lies below its first partial closure.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_le_partialClosure₁ (K : SaddleFunction m n) : K ≤ partialClosure₁ K

Lean source

Helper for Text 34.0.1: the second partial closure lies below the original saddle-function.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_partialClosure₂_le (K : SaddleFunction m n) : partialClosure₂ K ≤ K

Lean source

Helper for Text 34.0.1: the first partial closure is monotone with respect to pointwise order.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_partialClosure₁_mono {K L : SaddleFunction m n} (hKL : K ≤ L) : partialClosure₁ K ≤ partialClosure₁ L

Lean source

Helper for Text 34.0.1: the second partial closure is monotone with respect to pointwise order.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_partialClosure₂_mono {K L : SaddleFunction m n} (hKL : K ≤ L) : partialClosure₂ K ≤ partialClosure₂ L

Lean source

Helper for Text 34.0.1: any concave-closed first-variable majorant dominates the first partial closure.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_partialClosure₁_le_of_le_of_concaveClosedInFirst {K L : SaddleFunction m n} (hKL : K ≤ L) (hLClosed : IsConcaveClosedInFirst L) : partialClosure₁ K ≤ L

Lean source

Helper for Text 34.0.1: a convex-closed second-variable minorant stays below the second partial closure of any larger saddle-function.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_le_partialClosure₂_of_convexClosedInSecond_of_le {K L : SaddleFunction m n} (hKClosed : IsConvexClosedInSecond K) (hKL : K ≤ L) : K ≤ partialClosure₂ L

Lean source

Helper for Text 34.0.1: a closed convex bifunction witness for the mixed lower and upper closures forces the required inner cross-closure identities.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_closedConvexWitness_forces_crossClosure_relations {K : SaddleFunction m n} {h : IsConcaveConvex K} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : IsRockafellarConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hClosed : ClosedConvexBifunction F) (hLowerRep : ∀ u xStar, lowerClosureConcaveConvex K h u xStar = convexBifunctionPairing F u xStar) (hUpperRep : ∀ u xStar, upperClosureConcaveConvex K h u xStar = convexBifunctionCanonicalAdjointPairing F xStar u) (hLowerNoTopBot : HasNoTopOrBotValuesBifunction (lowerClosureConcaveConvex K h)) : partialClosure₁ (lowerClosureConcaveConvex K h) = upperClosureConcaveConvex K h ∧ partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h

Lean source

Helper for Text 34.0.1: the adjoint pairing of a convex bifunction, viewed as a saddle kernel in the (u, x^*) variable order.

open Set variable {m n : ℕ} noncomputable def helperForText_34_0_1_convexAdjointPairingKernel (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : SaddleFunction m n := fun u xStar => convexBifunctionCanonicalAdjointPairing F xStar u

Lean source

Helper for Text 34.0.1: function-equality formulas for the mixed closures can be evaluated pointwise when the upper closure is written via the adjoint pairing kernel.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_functionRepresentations_to_pointwiseRepresentations {K : SaddleFunction m n} {h : IsConcaveConvex K} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) (hUpperRep : upperClosureConcaveConvex K h = helperForText_34_0_1_convexAdjointPairingKernel F) : (∀ u xStar, lowerClosureConcaveConvex K h u xStar = convexBifunctionPairing F u xStar) ∧ ∀ u xStar, upperClosureConcaveConvex K h u xStar = convexBifunctionCanonicalAdjointPairing F xStar u

Lean source

Helper for Text 34.0.1: a closed convex bifunction witness written in function-equality form already forces the cross-closure identities.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_closedConvexWitness_forces_crossClosure_relations_of_function_equalities {K : SaddleFunction m n} {h : IsConcaveConvex K} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : IsRockafellarConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hClosed : ClosedConvexBifunction F) (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) (hUpperRep : upperClosureConcaveConvex K h = helperForText_34_0_1_convexAdjointPairingKernel F) (hLowerNoTopBot : HasNoTopOrBotValuesBifunction (lowerClosureConcaveConvex K h)) : partialClosure₁ (lowerClosureConcaveConvex K h) = upperClosureConcaveConvex K h ∧ partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h

Lean source

Helper for Text 34.0.1: under the extra hypothesis that the mixed lower closure takes no ⊥ values, Section 33 reconstructs it as the pairing of an image-closed convex bifunction.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_closedConvexWitness_exists (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBot : HasNoBotValuesBifunction K) (hLowerNoTopBot : HasNoTopOrBotValuesBifunction (lowerClosureConcaveConvex K h)) : ∃ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ (∀ u, IsFunctionConvexClosed (F u)) ∧ lowerClosureConcaveConvex K h = convexBifunctionPairing F

Lean source

Helper for Text 34.1.4: one application of cl₁ already makes a concave-convex kernel fixed by further first-variable closure.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_partialClosure₁_idempotent (K : SaddleFunction m n) : partialClosure₁ (partialClosure₁ K) = partialClosure₁ K

Lean source

Helper for Text 34.1.4: one application of cl₂ already makes a concave-convex kernel fixed by further second-variable closure.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_partialClosure₂_idempotent (K : SaddleFunction m n) : partialClosure₂ (partialClosure₂ K) = partialClosure₂ K

Lean source

Helper for Text 34.1.4: the mixed lower closure is fixed by repeating the lower-closure operator cl₂ ∘ cl₁.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_lowerClosure_repeatedLowerFixed (K : SaddleFunction m n) (h : IsConcaveConvex K) : partialClosure₂ (partialClosure₁ (lowerClosureConcaveConvex K h)) = lowerClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: the mixed upper closure is fixed by repeating the upper-closure operator cl₁ ∘ cl₂.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_upperClosure_repeatedUpperFixed (K : SaddleFunction m n) (h : IsConcaveConvex K) : partialClosure₁ (partialClosure₂ (upperClosureConcaveConvex K h)) = upperClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: the mixed lower closure is lower closed in the Section 33 saddle closedness sense.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_lowerClosure_isSection33LowerClosed (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBot : HasNoBotValuesBifunction K) : IsLowerClosedSaddleFunction (lowerClosureConcaveConvex K h)

Lean source

Helper for Text 34.1.4: the mixed upper closure is upper closed in the Section 33 saddle closedness sense.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_upperClosure_isSection33UpperClosed (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBot : HasNoBotValuesBifunction K) : IsUpperClosedSaddleFunction (upperClosureConcaveConvex K h)

Lean source

Helper for Text 34.1.4: the canonical first closure of the mixed lower closure is already the upper-closed Section 33 partner attached to that lower-closed kernel.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_firstClosureOfLowerClosure_isUpperClosedPartner (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBot : HasNoBotValuesBifunction K) : IsUpperClosedSaddleFunction (partialClosure₁ (lowerClosureConcaveConvex K h)) ∧ partialClosure₂ (partialClosure₁ (lowerClosureConcaveConvex K h)) = lowerClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: cl₁ underline(K) satisfies the full upper-partner data used by Corollary 33.3.2.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_firstClosureOfLower_isCanonicalUpperPartner (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBot : HasNoBotValuesBifunction K) : IsConcaveConvex (partialClosure₁ (lowerClosureConcaveConvex K h)) ∧ IsUpperClosedSaddleFunction (partialClosure₁ (lowerClosureConcaveConvex K h)) ∧ partialClosure₁ (lowerClosureConcaveConvex K h) = partialClosure₁ (lowerClosureConcaveConvex K h) ∧ partialClosure₂ (partialClosure₁ (lowerClosureConcaveConvex K h)) = lowerClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: the canonical second closure of the mixed upper closure is already the lower-closed Section 33 partner attached to that upper-closed kernel.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_secondClosureOfUpperClosure_isLowerClosedPartner (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBot : HasNoBotValuesBifunction K) : IsLowerClosedSaddleFunction (partialClosure₂ (upperClosureConcaveConvex K h)) ∧ partialClosure₁ (partialClosure₂ (upperClosureConcaveConvex K h)) = upperClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: cl₂ overline(K) satisfies the full lower-partner data used by Corollary 33.3.2.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_secondClosureOfUpper_isCanonicalLowerPartner (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBot : HasNoBotValuesBifunction K) : IsConcaveConvex (partialClosure₂ (upperClosureConcaveConvex K h)) ∧ IsLowerClosedSaddleFunction (partialClosure₂ (upperClosureConcaveConvex K h)) ∧ upperClosureConcaveConvex K h = partialClosure₁ (partialClosure₂ (upperClosureConcaveConvex K h)) ∧ partialClosure₂ (upperClosureConcaveConvex K h) = partialClosure₂ (upperClosureConcaveConvex K h)

Lean source

Helper for Text 34.1.4: the Section 33 order theorem gives the mixed lower closure below its canonical upper-closed partner cl₁ underline(K).

open Set variable {m n : ℕ} lemma helperForText_34_1_4_lowerClosure_below_canonicalUpperPartner (K : SaddleFunction m n) (h : IsConcaveConvex K) : lowerClosureConcaveConvex K h ≤ partialClosure₁ (lowerClosureConcaveConvex K h)

Lean source

Helper for Text 34.1.4: Corollary 33.3.2 makes cl₁ underline(K) the unique upper-closed partner attached to the mixed lower closure.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_existsUniqueCanonicalUpperPartner (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBot : HasNoBotValuesBifunction K) : ∃! U' : SaddleFunction m n, IsConcaveConvex U' ∧ IsUpperClosedSaddleFunction U' ∧ U' = partialClosure₁ (lowerClosureConcaveConvex K h) ∧ partialClosure₂ U' = lowerClosureConcaveConvex K h

§34 — Closures and Equivalence Classes — Part 2

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Predicates on coordinate spaces in this section are treated classically when needed in piecewise definitions.

open Set noncomputable local instance classicalDecidablePredPart2 {α : Type*} (p : α → Prop) : DecidablePred p

Lean source

Helper for Text 34.1.4: the canonical upper partner of underline(K) is unique without invoking the upstream Corollary 33.3.2 shim, because its defining equality already pins down the witness.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_lowerClosure_uniqueUpperClosedPartner (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBot : HasNoBotValuesBifunction K) : ∃! U' : SaddleFunction m n, IsConcaveConvex U' ∧ IsUpperClosedSaddleFunction U' ∧ U' = partialClosure₁ (lowerClosureConcaveConvex K h) ∧ partialClosure₂ U' = lowerClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: Corollary 33.3.2 makes cl₂ overline(K) the unique lower-closed partner attached to the mixed upper closure.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_existsUniqueCanonicalLowerPartner (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBot : HasNoBotValuesBifunction K) : ∃! L' : SaddleFunction m n, IsConcaveConvex L' ∧ IsLowerClosedSaddleFunction L' ∧ upperClosureConcaveConvex K h = partialClosure₁ L' ∧ partialClosure₂ (upperClosureConcaveConvex K h) = L'

Lean source

Helper for Text 34.1.4: the canonical lower partner of overline(K) is unique without invoking the upstream Corollary 33.3.2 shim, because the recovery equality already determines the witness.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_upperClosure_uniqueLowerClosedPartner (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBot : HasNoBotValuesBifunction K) : ∃! L' : SaddleFunction m n, IsConcaveConvex L' ∧ IsLowerClosedSaddleFunction L' ∧ upperClosureConcaveConvex K h = partialClosure₁ L' ∧ partialClosure₂ (upperClosureConcaveConvex K h) = L'

Lean source

Helper for Text 34.1.4: the second partial closure of the mixed upper closure already lies below the mixed lower closure.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_secondClosureOfUpper_le_lowerClosure (K : SaddleFunction m n) (h : IsConcaveConvex K) : partialClosure₂ (upperClosureConcaveConvex K h) ≤ lowerClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: the canonical lower partner cl₂ overline(K) also lies below overline(K) itself by the Section 33 closure-pair order theorem.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_secondClosureOfUpper_le_upperClosure (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBot : HasNoBotValuesBifunction K) : partialClosure₂ (upperClosureConcaveConvex K h) ≤ upperClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: once the exact recovery cl₂ overline(K) = underline(K) is known, the mixed-order inequality follows immediately from the Section 33 uniqueness package.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_mixedClosure_order_of_secondClosure_eq_lower (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBot : HasNoBotValuesBifunction K) (hRecover : partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h) : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: on fixed first- and second-variable neighborhoods, the trivial minimax inequality places the local maximin value below the local minimax value.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_fixedNeighborhood_maximin_le_minimax (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) : (⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), K w.1 z.1) ≤ (⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), K w.1 z.1)

Lean source

Helper for Text 34.1.4: on fixed neighborhood balls, reverse minimax together with the trivial forward inequality yields equality of the local minimax and maximin values.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_fixedNeighborhood_minimax_eq_maximin_of_reverseMinimax (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) (hReverse : (⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), K w.1 z.1) ≤ (⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), K w.1 z.1)) : (⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), K w.1 z.1) = (⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), K w.1 z.1)

Lean source

Helper for Text 34.1.4: pointwise, the lower mixed closure is the supremum over second-variable neighborhoods of the infimum over first-variable neighborhoods of the local minimax values.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_lowerMixedClosure_pointwise_as_supInf_localBallMinimax (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) : partialClosure₂ (partialClosure₁ K) u xStar = ⨆ (δ : {δ : ℝ // 0 < δ}), ⨅ (ε : {ε : ℝ // 0 < ε}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), K w.1 z.1

Lean source

Helper for Text 34.1.4: pointwise, the upper mixed closure is the infimum over first-variable neighborhoods of the supremum over second-variable neighborhoods of the local maximin values.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_upperMixedClosure_pointwise_as_infSup_localBallMaximin (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) : partialClosure₁ (partialClosure₂ K) u xStar = ⨅ (ε : {ε : ℝ // 0 < ε}), ⨆ (δ : {δ : ℝ // 0 < δ}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), K w.1 z.1

Lean source

Helper for Text 34.1.4: enlarging the first-variable neighborhood can only increase the local minimax value.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_fixedNeighborhood_minimax_mono_firstRadius (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) {ε ε' : {ε : ℝ // 0 < ε}} (hε : ε.1 ≤ ε'.1) (δ : {δ : ℝ // 0 < δ}) : (⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), K w.1 z.1) ≤ (⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε'.1}), K w.1 z.1)

Lean source

Helper for Text 34.1.4: enlarging the second-variable neighborhood can only decrease the local minimax value.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_fixedNeighborhood_minimax_antitone_secondRadius (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) {δ δ' : {δ : ℝ // 0 < δ}} (hδ : δ.1 ≤ δ'.1) : (⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ'.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), K w.1 z.1) ≤ (⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), K w.1 z.1)

Lean source

Helper for Text 34.1.4: enlarging the first-variable neighborhood can only increase the local maximin value.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_fixedNeighborhood_maximin_mono_firstRadius (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) {ε ε' : {ε : ℝ // 0 < ε}} (hε : ε.1 ≤ ε'.1) (δ : {δ : ℝ // 0 < δ}) : (⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), K w.1 z.1) ≤ (⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε'.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), K w.1 z.1)

Lean source

Helper for Text 34.1.4: enlarging the second-variable neighborhood can only decrease the local maximin value.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_fixedNeighborhood_maximin_antitone_secondRadius (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) {δ δ' : {δ : ℝ // 0 < δ}} (hδ : δ.1 ≤ δ'.1) : (⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ'.1}), K w.1 z.1) ≤ (⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), K w.1 z.1)

Lean source

Helper for Text 34.1.4: once the mixed lower closure lies below the mixed upper closure, the second cross-closure identity follows by squeezing cl₂ overline(K) between them.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_secondClosureOfUpper_eq_lower_of_mixedClosure_order (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBot : HasNoBotValuesBifunction K) (hOrder : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h) : partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: the remaining order comparison is definitionally the raw mixed closure inequality cl₂ (cl₁ K) ≤ cl₁ (cl₂ K).

open Set variable {m n : ℕ} lemma helperForText_34_1_4_mixedClosure_order_iff_rawMixedClosureOrder (K : SaddleFunction m n) (h : IsConcaveConvex K) : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h ↔ partialClosure₂ (partialClosure₁ K) ≤ partialClosure₁ (partialClosure₂ K)

Lean source

Helper for Text 34.1.4: a reverse minimax inequality on every pair of first- and second-variable neighborhoods already forces the raw mixed-closure order cl₂ (cl₁ K) ≤ cl₁ (cl₂ K).

open Set variable {m n : ℕ} lemma helperForText_34_1_4_rawMixedClosureOrder_of_fixedNeighborhood_reverseMinimax (K : SaddleFunction m n) (hReverse : ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}), (⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), K w.1 z.1) ≤ (⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), K w.1 z.1)) : partialClosure₂ (partialClosure₁ K) ≤ partialClosure₁ (partialClosure₂ K)

Lean source

Helper for Text 34.1.4: once the fixed-neighborhood reverse minimax theorem is available, the textbook mixed-closure order underline(K) ≤ overline(K) follows immediately.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_mixedClosure_order_of_fixedNeighborhood_reverseMinimax (K : SaddleFunction m n) (h : IsConcaveConvex K) (hReverse : ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}), (⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), K w.1 z.1) ≤ (⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), K w.1 z.1)) : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: a fixed-neighborhood reverse minimax theorem upgrades the mixed-order comparison to the exact recovery identity cl₂ overline(K) = underline(K).

open Set variable {m n : ℕ} lemma helperForText_34_1_4_secondClosureOfUpper_eq_lower_of_fixedNeighborhood_reverseMinimax (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBot : HasNoBotValuesBifunction K) (hReverse : ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}), (⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), K w.1 z.1) ≤ (⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), K w.1 z.1)) : partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: the textbook mixed-order comparison is equivalent to the exact recovery cl₂ overline(K) = underline(K).

open Set variable {m n : ℕ} lemma helperForText_34_1_4_mixedClosure_order_iff_secondClosureOfUpper_eq_lower (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBot : HasNoBotValuesBifunction K) : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h ↔ partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: a saddle point on two open balls forces the reverse minimax inequality. This packages the standard saddle-point implication: if (w₀, z₀) satisfies K w z₀ ≤ K w₀ z for all w and z in the respective balls, then the local minimax and maximin values coincide with the saddle value K w₀ z₀.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_fixedNeighborhood_reverseMinimax_of_saddlePointOn (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) (w0 : Fin m → ℝ) (hw0 : ‖w0 - u‖ < ε.1) (z0 : Fin n → ℝ) (hz0 : ‖z0 - xStar‖ < δ.1) (hS : IsSaddlePointOn (X := {z : Fin n → ℝ | ‖z - xStar‖ < δ.1}) (Y := {w : Fin m → ℝ | ‖w - u‖ < ε.1}) (f := fun z w => K w z) z0 w0) : (⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), K w.1 z.1) ≤ (⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), K w.1 z.1)

Lean source

Helper for Text 34.1.4: a saddle point on two closed balls forces the reverse minimax inequality. This is the closed-ball analogue of helperForText_34_1_4_fixedNeighborhood_reverseMinimax_of_saddlePointOn.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_fixedNeighborhood_reverseMinimax_closedBall_of_saddlePointOn (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ρw ρz : ℝ) (w0 : Fin m → ℝ) (hw0 : ‖w0 - u‖ ≤ ρw) (z0 : Fin n → ℝ) (hz0 : ‖z0 - xStar‖ ≤ ρz) (hS : IsSaddlePointOn (X := {z : Fin n → ℝ | ‖z - xStar‖ ≤ ρz}) (Y := {w : Fin m → ℝ | ‖w - u‖ ≤ ρw}) (f := fun z w => K w z) z0 w0) : (⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ ≤ ρz}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ ≤ ρw}), K w.1 z.1) ≤ (⨆ (w : {w : Fin m → ℝ // ‖w - u‖ ≤ ρw}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ ≤ ρz}), K w.1 z.1)

Lean source

Helper for Text 34.1.4: global concave-convexity on ℝ^m × ℝ^n restricts to any pair of first- and second-variable open balls.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_fixedNeighborhood_concaveConvexOn_of_global (K : SaddleFunction m n) (h : IsConcaveConvex K) : ∀ (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}), IsConcaveConvexOn ({w : Fin m → ℝ | ‖w - u‖ < ε.1}) ({z : Fin n → ℝ | ‖z - xStar‖ < δ.1}) K

Lean source

Helper for Text 34.1.4: restricting the first- and second-variable domains preserves IsConcaveConvexOn.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_IsConcaveConvexOn_mono {C C' : Set (Fin m → ℝ)} {D D' : Set (Fin n → ℝ)} {K : SaddleFunction m n} (hC : C' ⊆ C) (hD : D' ⊆ D) (h : IsConcaveConvexOn C D K) : IsConcaveConvexOn C' D' K

Lean source

Helper for Text 34.1.4: local concave-convexity on open balls restricts to every pair of strictly smaller closed balls.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_fixedNeighborhood_concaveConvexOn_closedBalls_of_local (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) (hLocal : IsConcaveConvexOn ({w : Fin m → ℝ | ‖w - u‖ < ε.1}) ({z : Fin n → ℝ | ‖z - xStar‖ < δ.1}) K) : ∀ (ρw : {ρ : ℝ // ρ < ε.1}) (ρz : {ρ : ℝ // ρ < δ.1}), IsConcaveConvexOn ({w : Fin m → ℝ | ‖w - u‖ ≤ ρw.1}) ({z : Fin n → ℝ | ‖z - xStar‖ ≤ ρz.1}) K

Lean source

Helper for Text 34.1.4: an open-ball iInf can be rewritten as an iInf over all smaller closed-ball radii, followed by the corresponding closed-ball iInf.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_openBall_iInf_as_iInf_closedBallRadii {k : ℕ} (u : Fin k → ℝ) (ε : ℝ) (f : (Fin k → ℝ) → EReal) : (⨅ (w : {w : Fin k → ℝ // ‖w - u‖ < ε}), f w.1) = (⨅ (ρ : {ρ : ℝ // ρ < ε}), ⨅ (w : {w : Fin k → ℝ // ‖w - u‖ ≤ ρ.1}), f w.1)

Lean source

Helper for Text 34.1.4: an open-ball iSup can be rewritten as an iSup over all smaller closed-ball radii, followed by the corresponding closed-ball iSup.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_openBall_iSup_as_iSup_closedBallRadii {k : ℕ} (u : Fin k → ℝ) (ε : ℝ) (f : (Fin k → ℝ) → EReal) : (⨆ (w : {w : Fin k → ℝ // ‖w - u‖ < ε}), f w.1) = (⨆ (ρ : {ρ : ℝ // ρ < ε}), ⨆ (w : {w : Fin k → ℝ // ‖w - u‖ ≤ ρ.1}), f w.1)

Lean source

Helper for Text 34.1.4: the fixed-neighborhood minimax expression on open balls can be expanded into nested inf/sup over smaller closed-ball radii.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_fixedNeighborhood_minimax_openBall_as_closedBallRadii (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) : (⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), K w.1 z.1) = (⨅ (ρz : {ρ : ℝ // ρ < δ.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ ≤ ρz.1}), ⨆ (ρw : {ρ : ℝ // ρ < ε.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ ≤ ρw.1}), K w.1 z.1)

Lean source

Helper for Text 34.1.4: the fixed-neighborhood maximin expression on open balls can be expanded into nested sup/inf over smaller closed-ball radii.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_fixedNeighborhood_maximin_openBall_as_closedBallRadii (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) : (⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), K w.1 z.1) = (⨆ (ρw : {ρ : ℝ // ρ < ε.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ ≤ ρw.1}), ⨅ (ρz : {ρ : ℝ // ρ < δ.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ ≤ ρz.1}), K w.1 z.1)

Lean source

Helper for Text 34.1.4: to prove reverse minimax on open balls, it is enough to prove the corresponding inequality on the closed-ball-radius envelope form.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_fixedNeighborhood_reverseMinimax_of_closedBallRadiusEnvelope (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) (hEnvelope : (⨅ (ρz : {ρ : ℝ // ρ < δ.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ ≤ ρz.1}), ⨆ (ρw : {ρ : ℝ // ρ < ε.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ ≤ ρw.1}), K w.1 z.1) ≤ (⨆ (ρw : {ρ : ℝ // ρ < ε.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ ≤ ρw.1}), ⨅ (ρz : {ρ : ℝ // ρ < δ.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ ≤ ρz.1}), K w.1 z.1)) : (⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), K w.1 z.1) ≤ (⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), K w.1 z.1)

Lean source

Helper for Text 34.1.4: the reverse-minimax inequality on open balls is equivalent to the closed-ball-radius envelope inequality used to isolate the upstream minimax blocker.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_fixedNeighborhood_reverseMinimax_iff_closedBallRadiusEnvelope (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) : ((⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), K w.1 z.1) ≤ (⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), K w.1 z.1)) ↔ (⨅ (ρz : {ρ : ℝ // ρ < δ.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ ≤ ρz.1}), ⨆ (ρw : {ρ : ℝ // ρ < ε.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ ≤ ρw.1}), K w.1 z.1) ≤ (⨆ (ρw : {ρ : ℝ // ρ < ε.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ ≤ ρw.1}), ⨅ (ρz : {ρ : ℝ // ρ < δ.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ ≤ ρz.1}), K w.1 z.1)

Lean source

Helper for Text 34.1.4: a saddle point on the open-ball neighborhoods implies the closed-ball-radius envelope reverse-minimax inequality. This repackages helperForText_34_1_4_fixedNeighborhood_reverseMinimax_of_saddlePointOn using the equivalence helperForText_34_1_4_fixedNeighborhood_reverseMinimax_iff_closedBallRadiusEnvelope.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_reverseMinimax_closedBallRadiusEnvelope_of_saddlePointOn (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) (w0 : Fin m → ℝ) (hw0 : ‖w0 - u‖ < ε.1) (z0 : Fin n → ℝ) (hz0 : ‖z0 - xStar‖ < δ.1) (hS : IsSaddlePointOn (X := {z : Fin n → ℝ | ‖z - xStar‖ < δ.1}) (Y := {w : Fin m → ℝ | ‖w - u‖ < ε.1}) (f := fun z w => K w z) z0 w0) : (⨅ (ρz : {ρ : ℝ // ρ < δ.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ ≤ ρz.1}), ⨆ (ρw : {ρ : ℝ // ρ < ε.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ ≤ ρw.1}), K w.1 z.1) ≤ (⨆ (ρw : {ρ : ℝ // ρ < ε.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ ≤ ρw.1}), ⨅ (ρz : {ρ : ℝ // ρ < δ.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ ≤ ρz.1}), K w.1 z.1)

§34 — Closures and Equivalence Classes — Part 3

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Predicates on coordinate spaces in this section are treated classically when needed in piecewise definitions.

open Set noncomputable local instance classicalDecidablePredPart3 {α : Type*} (p : α → Prop) : DecidablePred p

Lean source

Dimension-wide realization qualification for the converse direction of Corollary 33.3.1. The forward closure theorem does not construct a bifunction witness from an arbitrary closure pair, so the two canonical realization clauses are recorded explicitly.

open Set variable {m n : ℕ} structure Section34CanonicalClosureRealizationQualification (m n : ℕ) : Prop where upper : ∀ (K : SaddleFunction m n) (h : IsConcaveConvex K), HasNoBotValuesBifunction K → HasNoTopOrBotValuesBifunction (lowerClosureConcaveConvex K h) → ∃! F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, ClosedConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ (∀ u xStar, lowerClosureConcaveConvex K h u xStar = convexBifunctionPairing F u xStar) ∧ ∀ u xStar, partialClosure₁ (lowerClosureConcaveConvex K h) u xStar = convexBifunctionCanonicalAdjointPairing F xStar u lower : ∀ (K : SaddleFunction m n) (h : IsConcaveConvex K), HasNoBotValuesBifunction K → HasNoTopOrBotValuesBifunction (partialClosure₂ (upperClosureConcaveConvex K h)) → ∃! F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, ClosedConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ (∀ u xStar, partialClosure₂ (upperClosureConcaveConvex K h) u xStar = convexBifunctionPairing F u xStar) ∧ ∀ u xStar, upperClosureConcaveConvex K h u xStar = convexBifunctionCanonicalAdjointPairing F xStar u

Lean source

Route correction for Text 34.1.4: the zero-dimensional constant-⊥ kernel is concave-convex, and its mixed lower closure still equals ⊥ at the unique point.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_zeroDimensional_constBot_lowerClosure_eq_bot : let K : SaddleFunction 0 0 := fun _ _ => (⊥ : EReal) let h : IsConcaveConvex K := by constructor · intro v hv simpa using helperForText_34_1_4_constBot_isERealConcaveOn (n := 0) · intro u hu simpa using helperForTheorem33_1_constBot_isERealConvexOn (n := 0) lowerClosureConcaveConvex K h 0 0 = (⊥ : EReal)

Lean source

Route correction for Text 34.1.4: the statement HasNoBotValuesBifunction (lowerClosureConcaveConvex K h) is false in general. The zero-dimensional constant-⊥ kernel already refutes it.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_zeroDimensional_constBot_refutes_lowerClosure_hasNoBot_statement : let K : SaddleFunction 0 0 := fun _ _ => (⊥ : EReal) let h : IsConcaveConvex K := by constructor · intro v hv simpa using helperForText_34_1_4_constBot_isERealConcaveOn (n := 0) · intro u hu simpa using helperForTheorem33_1_constBot_isERealConvexOn (n := 0) ¬ HasNoBotValuesBifunction (lowerClosureConcaveConvex K h)

Lean source

Helper for Text 34.1.4: reverse-minimax inequality on a pair of open balls, derived from local concave-convexity on those balls. This lemma is the true upstream minimax/duality input required by the current proof route. Once it is available, helperForText_34_1_4_reverseMinimax_closedBallRadiusEnvelope follows by deterministic rewriting to the closed-ball-radius envelope form.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_openBall_reverseMinimax_of_localConcaveConvexOn_of_saddlePointOn (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) (hLocal : IsConcaveConvexOn ({w : Fin m → ℝ | ‖w - u‖ < ε.1}) ({z : Fin n → ℝ | ‖z - xStar‖ < δ.1}) K) (w0 : Fin m → ℝ) (hw0 : ‖w0 - u‖ < ε.1) (z0 : Fin n → ℝ) (hz0 : ‖z0 - xStar‖ < δ.1) (hS : IsSaddlePointOn (X := {z : Fin n → ℝ | ‖z - xStar‖ < δ.1}) (Y := {w : Fin m → ℝ | ‖w - u‖ < ε.1}) (f := fun z w => K w z) z0 w0) : (⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), K w.1 z.1) ≤ (⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), K w.1 z.1)

Lean source

Helper for Text 34.1.4: a one-dimensional coordinate kernel used to exhibit that Jensen-style concave-convexity on open balls does not force existence of a saddle point.

open Set variable {m n : ℕ} noncomputable def helperForText_34_1_4_coordKernel : SaddleFunction 1 1 := fun w (_ : Fin 1 → ℝ) => (w 0 : ℝ)

Lean source

Helper for Text 34.1.4: the coordinate kernel is concave-convex on every pair of open balls. The first-variable Jensen inequality holds by equality for the linear map w ↦ w 0, while the second-variable inequality holds because the kernel is constant in the second variable.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_coordKernel_concaveConvexOn_openBalls (u xStar : Fin 1 → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) : IsConcaveConvexOn ({w : Fin 1 → ℝ | ‖w - u‖ < ε.1}) ({z : Fin 1 → ℝ | ‖z - xStar‖ < δ.1}) helperForText_34_1_4_coordKernel

Lean source

Helper for Text 34.1.4: the one-dimensional coordinate kernel is globally concave-convex.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_coordKernel_isConcaveConvex : IsConcaveConvex helperForText_34_1_4_coordKernel

Lean source

Helper for Text 34.1.4: the first-variable closure of the coordinate kernel vanishes at the origin of the first variable.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_coordKernel_firstClosure_at_origin (z : Fin 1 → ℝ) : partialClosure₁ helperForText_34_1_4_coordKernel 0 z = 0

Lean source

Helper for Text 34.1.4: the mixed lower closure of the coordinate kernel also vanishes at the origin of the first variable.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_coordKernel_lowerClosure_at_origin (z : Fin 1 → ℝ) : lowerClosureConcaveConvex helperForText_34_1_4_coordKernel helperForText_34_1_4_coordKernel_isConcaveConvex 0 z = 0

Lean source

Helper for Text 34.1.4: a one-dimensional coordinate kernel in the second variable, used to test the right closed-ball bridge.

open Set variable {m n : ℕ} noncomputable def helperForText_34_1_4_secondCoordKernel : SaddleFunction 1 1 := fun (_ : Fin 1 → ℝ) z => (z 0 : ℝ)

Lean source

Helper for Text 34.1.4: the second-coordinate kernel is globally concave-convex.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_secondCoordKernel_isConcaveConvex : IsConcaveConvex helperForText_34_1_4_secondCoordKernel

Lean source

Helper for Text 34.1.4: the first closure of the second-coordinate kernel is trivial because the kernel is constant in the first variable.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_secondCoordKernel_firstClosure (u z : Fin 1 → ℝ) : partialClosure₁ helperForText_34_1_4_secondCoordKernel u z = (((z 0 : ℝ)) : EReal)

Lean source

Helper for Text 34.1.4: the second closure of the second-coordinate kernel vanishes at the origin of the second variable.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_secondCoordKernel_secondClosure_at_origin (u : Fin 1 → ℝ) : partialClosure₂ helperForText_34_1_4_secondCoordKernel u 0 = 0

Lean source

Helper for Text 34.1.4: the mixed lower closure of the second-coordinate kernel also vanishes at the origin of the second variable.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_secondCoordKernel_lowerClosure_at_origin (u : Fin 1 → ℝ) : lowerClosureConcaveConvex helperForText_34_1_4_secondCoordKernel helperForText_34_1_4_secondCoordKernel_isConcaveConvex u 0 = 0

Lean source

Helper for Text 34.1.4: any point in a nontrivial open ball in Fin 1 → ℝ can be moved to a nearby point in the same ball with strictly larger 0-th coordinate.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_exists_mem_openBall_gt_coord_fin1 (u : Fin 1 → ℝ) (ε : ℝ) (w0 : Fin 1 → ℝ) (hw0 : ‖w0 - u‖ < ε) : ∃ w : Fin 1 → ℝ, ‖w - u‖ < ε ∧ w0 0 < w 0

Lean source

Helper for Text 34.1.4: the coordinate kernel admits no saddle point on open balls. This gives a concrete counterexample to the naive implication IsConcaveConvexOn (openBall u ε) (openBall xStar δ) K → ∃ saddle point, so the corresponding step in the proof pipeline must be replaced by a genuine minimax/duality theorem with additional hypotheses (compactness/attainment, semicontinuity, etc.).

open Set variable {m n : ℕ} lemma helperForText_34_1_4_coordKernel_no_saddlePointOn_openBalls (u xStar : Fin 1 → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) : ¬ (∃ (w0 : Fin 1 → ℝ), ‖w0 - u‖ < ε.1 ∧ ∃ (z0 : Fin 1 → ℝ), ‖z0 - xStar‖ < δ.1 ∧ IsSaddlePointOn (X := {z : Fin 1 → ℝ | ‖z - xStar‖ < δ.1}) (Y := {w : Fin 1 → ℝ | ‖w - u‖ < ε.1}) (f := fun z w => helperForText_34_1_4_coordKernel w z) z0 w0)

Lean source

Helper for Text 34.1.4: counterexample showing that the open-ball saddle-point existence step cannot be proved from IsConcaveConvexOn alone.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_counterexample_openBall_exists_saddlePointOn_false (u xStar : Fin 1 → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) : IsConcaveConvexOn ({w : Fin 1 → ℝ | ‖w - u‖ < ε.1}) ({z : Fin 1 → ℝ | ‖z - xStar‖ < δ.1}) helperForText_34_1_4_coordKernel ∧ ¬ (∃ (w0 : Fin 1 → ℝ), ‖w0 - u‖ < ε.1 ∧ ∃ (z0 : Fin 1 → ℝ), ‖z0 - xStar‖ < δ.1 ∧ IsSaddlePointOn (X := {z : Fin 1 → ℝ | ‖z - xStar‖ < δ.1}) (Y := {w : Fin 1 → ℝ | ‖w - u‖ < ε.1}) (f := fun z w => helperForText_34_1_4_coordKernel w z) z0 w0)

Lean source

Helper for Text 34.1.4: the abandoned open-ball route is already false in dimension 1. Concretely, there is no uniform theorem saying that local concave-convexity on open balls forces existence of a saddle point on those balls.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_no_uniform_openBall_saddlePointExistence_in_dim1 : ¬ ∀ (K : SaddleFunction 1 1) (u xStar : Fin 1 → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}), IsConcaveConvexOn ({w : Fin 1 → ℝ | ‖w - u‖ < ε.1}) ({z : Fin 1 → ℝ | ‖z - xStar‖ < δ.1}) K → ∃ (w0 : Fin 1 → ℝ), ‖w0 - u‖ < ε.1 ∧ ∃ (z0 : Fin 1 → ℝ), ‖z0 - xStar‖ < δ.1 ∧ IsSaddlePointOn (X := {z : Fin 1 → ℝ | ‖z - xStar‖ < δ.1}) (Y := {w : Fin 1 → ℝ | ‖w - u‖ < ε.1}) (f := fun z w => K w z) z0 w0

Lean source

Helper for Text 34.1.4: the coordinate kernel satisfies the reverse-minimax inequality on open balls even though it has no saddle point there. This emphasizes that failure of saddle-point attainment does not by itself refute reverse minimax; what is missing for the general theorem is a genuine minimax/duality input, not the simple saddle-point existence step.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_coordKernel_reverseMinimax_openBalls (u xStar : Fin 1 → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) : (⨅ (z : {z : Fin 1 → ℝ // ‖z - xStar‖ < δ.1}), ⨆ (w : {w : Fin 1 → ℝ // ‖w - u‖ < ε.1}), helperForText_34_1_4_coordKernel w.1 z.1) ≤ (⨆ (w : {w : Fin 1 → ℝ // ‖w - u‖ < ε.1}), ⨅ (z : {z : Fin 1 → ℝ // ‖z - xStar‖ < δ.1}), helperForText_34_1_4_coordKernel w.1 z.1)

Lean source

Helper for Text 34.1.4: graph-function closedness upgrades a Rockafellar convex bifunction to a closed convex bifunction.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_closedConvexBifunction_of_graphFunctionClosed {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hRock : IsRockafellarConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hGraphClosed : IsFunctionConvexClosed (graphFunctionOfBifunction F)) : ClosedConvexBifunction F

Lean source

Helper for Text 34.1.4: the qualified converse of Corollary 33.3.1 gives a closed convex witness for the canonical pair (underline(K), cl₁ underline(K)) once the primal side is assumed to avoid both ⊥ and ⊤.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_closedConvexWitness_exists_for_canonicalUpperPartner (K : SaddleFunction m n) (h : IsConcaveConvex K) (hKNoBot : HasNoBotValuesBifunction K) (hLowerNoTopBot : HasNoTopOrBotValuesBifunction (lowerClosureConcaveConvex K h)) (hRealization : Section34CanonicalClosureRealizationQualification m n) : ∃ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, ClosedConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ lowerClosureConcaveConvex K h = convexBifunctionPairing F ∧ partialClosure₁ (lowerClosureConcaveConvex K h) = helperForText_34_0_1_convexAdjointPairingKernel F

Lean source

Helper for Text 34.1.4: the corrected Corollary 33.3.1 likewise gives a closed convex witness for the canonical pair (cl₂ overline(K), overline(K)) once the lower partner is assumed to avoid both ⊥ and ⊤.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_closedConvexWitness_exists_for_canonicalLowerPartner (K : SaddleFunction m n) (h : IsConcaveConvex K) (hKNoBot : HasNoBotValuesBifunction K) (hCanonicalLowerNoTopBot : HasNoTopOrBotValuesBifunction (partialClosure₂ (upperClosureConcaveConvex K h))) (hRealization : Section34CanonicalClosureRealizationQualification m n) : ∃ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, ClosedConvexBifunction F ∧ partialClosure₂ (upperClosureConcaveConvex K h) = convexBifunctionPairing F ∧ upperClosureConcaveConvex K h = helperForText_34_0_1_convexAdjointPairingKernel F

Lean source

Helper for Text 34.1.4: a closed convex bifunction already has graph function fixed by the raw convex-closure operator.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_graphFunction_isFunctionConvexClosed_of_closedConvexBifunction {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hClosed : ClosedConvexBifunction F) : IsFunctionConvexClosed (graphFunctionOfBifunction F)

Lean source

Helper for Text 34.1.4: the constant ⊥ function is concave in the Jensen sense on ℝ^n.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_constBot_isERealConcaveOn {n : ℕ} : IsERealConcaveOn (Set.univ : Set (Fin n → ℝ)) (fun _ : Fin n → ℝ => (⊥ : EReal))

Lean source

Helper for Text 34.1.4: although underline(K) need not avoid ⊥ in general, any actual closed-convex witness for underline(K) forces the represented mixed lower closure to have no ⊥ values.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_lowerClosure_hasNoBot_of_closedConvexLowerRepresentation (K : SaddleFunction m n) (h : IsConcaveConvex K) {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hRock : IsRockafellarConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hClosed : ClosedConvexBifunction F) (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) (hFiniteSections : ∀ u : Fin m → ℝ, ∃ x : Fin n → ℝ, F u x < (⊤ : EReal)) : HasNoBotValuesBifunction (lowerClosureConcaveConvex K h)

Lean source

Helper for Text 34.1.4: the usable no-⊥ statement for underline(K) is the witness-local one. The naive global version is false by the constant-⊥ counterexample above, so the remaining development only uses this corrected interface.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_lowerClosure_hasNoBot (K : SaddleFunction m n) (h : IsConcaveConvex K) {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hRock : IsRockafellarConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hClosed : ClosedConvexBifunction F) (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) (hFiniteSections : ∀ u : Fin m → ℝ, ∃ x : Fin n → ℝ, F u x < (⊤ : EReal)) : HasNoBotValuesBifunction (lowerClosureConcaveConvex K h)

Lean source

Helper for Text 34.1.4: if a convex bifunction represents underline(K), then none of its parameter sections can be identically ⊤; otherwise the pairing would force underline(K) to take the value ⊥.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_witness_sections_notTop_of_lowerRepresentation {K : SaddleFunction m n} {h : IsConcaveConvex K} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hLowerNoBot : HasNoBotValuesBifunction (lowerClosureConcaveConvex K h)) (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) : ∀ u : Fin m → ℝ, ∃ x : Fin n → ℝ, F u x ≠ (⊤ : EReal)

Lean source

Helper for Text 34.1.4: any convex bifunction representing underline(K) already has full strict parameter domain, because underline(K) never takes the value ⊥.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_strictParameterDomain_eq_univ_of_lowerRepresentation {K : SaddleFunction m n} {h : IsConcaveConvex K} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hLowerNoBot : HasNoBotValuesBifunction (lowerClosureConcaveConvex K h)) (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) : {u' : Fin m → ℝ | ∃ x : Fin n → ℝ, F u' x < ⊤} = Set.univ

Lean source

Helper for Text 34.1.4: any convex bifunction representing underline(K) has full Section 33 parameter domain, not just the strict < ⊤ variant.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_parameterDomain_eq_univ_of_lowerRepresentation {K : SaddleFunction m n} {h : IsConcaveConvex K} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hLowerNoBot : HasNoBotValuesBifunction (lowerClosureConcaveConvex K h)) (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) : convexBifunctionParameterDomain F = Set.univ

§34 — Closures and Equivalence Classes — Part 4

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Predicates on coordinate spaces in this section are treated classically when needed in piecewise definitions.

open Set noncomputable local instance classicalDecidablePredPart4 {α : Type*} (p : α → Prop) : DecidablePred p

Lean source

Corrected interface for Text 34.1.4: under the necessary no-⊤/no-⊥ hypothesis on underline(K), the exact recovery identity cl₂ overline(K) = underline(K) already yields the closed-convex witness for the textbook pair.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_closedConvexWitness_exists_of_secondClosure_eq_lower_of_noTopOrBot (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) (hLowerNoTopBot : HasNoTopOrBotValuesBifunction (lowerClosureConcaveConvex K h)) (hRecover : partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h) (hRealization : Section34CanonicalClosureRealizationQualification m n) : ∃ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ ClosedConvexBifunction F ∧ lowerClosureConcaveConvex K h = convexBifunctionPairing F ∧ upperClosureConcaveConvex K h = helperForText_34_0_1_convexAdjointPairingKernel F

Lean source

Corrected interface for Text 34.1.4: under the necessary no-⊤/no-⊥ hypothesis on underline(K), the textbook witness condition is equivalent to cl₁ underline(K) = overline(K).

open Set variable {m n : ℕ} lemma helperForText_34_1_4_closedConvexWitness_exists_iff_firstClosureOfLower_eq_upper_of_noTopOrBot (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) (hLowerNoTopBot : HasNoTopOrBotValuesBifunction (lowerClosureConcaveConvex K h)) (hRealization : Section34CanonicalClosureRealizationQualification m n) : (∃ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ ClosedConvexBifunction F ∧ lowerClosureConcaveConvex K h = convexBifunctionPairing F ∧ upperClosureConcaveConvex K h = helperForText_34_0_1_convexAdjointPairingKernel F) ↔ partialClosure₁ (lowerClosureConcaveConvex K h) = upperClosureConcaveConvex K h

Lean source

Corrected interface for Text 34.1.4: under the necessary no-⊤/no-⊥ hypothesis on underline(K), the textbook witness condition is equivalent to cl₂ overline(K) = underline(K).

open Set variable {m n : ℕ} lemma helperForText_34_1_4_closedConvexWitness_exists_iff_secondClosure_eq_lower_of_noTopOrBot (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) (hLowerNoTopBot : HasNoTopOrBotValuesBifunction (lowerClosureConcaveConvex K h)) (hRealization : Section34CanonicalClosureRealizationQualification m n) : (∃ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ ClosedConvexBifunction F ∧ lowerClosureConcaveConvex K h = convexBifunctionPairing F ∧ upperClosureConcaveConvex K h = helperForText_34_0_1_convexAdjointPairingKernel F) ↔ partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h

Lean source

Corrected interface for Text 34.1.4: under the necessary no-⊤/no-⊥ hypothesis on underline(K), the textbook witness condition is equivalent to the raw mixed-order inequality cl₂ (cl₁ K) ≤ cl₁ (cl₂ K).

open Set variable {m n : ℕ} lemma helperForText_34_1_4_closedConvexWitness_exists_iff_rawMixedClosureOrder_of_noTopOrBot (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) (hLowerNoTopBot : HasNoTopOrBotValuesBifunction (lowerClosureConcaveConvex K h)) (hRealization : Section34CanonicalClosureRealizationQualification m n) : (∃ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ ClosedConvexBifunction F ∧ lowerClosureConcaveConvex K h = convexBifunctionPairing F ∧ upperClosureConcaveConvex K h = helperForText_34_0_1_convexAdjointPairingKernel F) ↔ partialClosure₂ (partialClosure₁ K) ≤ partialClosure₁ (partialClosure₂ K)

Lean source

Helper for Text 34.1.4: once a Rockafellar convex bifunction represents underline(K), the full-domain branch of Section 33 yields the stronger inner-product equation against the genuine adjoint pairing.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_hasInnerProductEquation_of_lowerRepresentation {K : SaddleFunction m n} {h : IsConcaveConvex K} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hLowerNoBot : HasNoBotValuesBifunction (lowerClosureConcaveConvex K h)) (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) : HasInnerProductEquation F

Lean source

Helper for Text 34.1.4: any convex bifunction representing underline(K) already satisfies the global closure-side pairing identity ⟪F u, x^*⟫ = ⟪u, F^* x^*⟫.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_pairingEquality_everywhere_of_lowerRepresentation {K : SaddleFunction m n} {h : IsConcaveConvex K} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hGraph : IsGraphConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hLowerNoBot : HasNoBotValuesBifunction (lowerClosureConcaveConvex K h)) (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) : ∀ u : Fin m → ℝ, ∀ xStar : Fin n → ℝ, convexBifunctionPairing F u xStar = convexBifunctionCanonicalAdjointPairing F xStar u

Lean source

Helper for Text 34.1.4: the lower-representation inner-product equation can be derived directly from a closed-convex witness, without any global no-⊥ assumption on underline(K).

open Set variable {m n : ℕ} lemma helperForText_34_1_4_hasInnerProductEquation_of_closedConvexLowerRepresentation {K : SaddleFunction m n} {h : IsConcaveConvex K} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hRock : IsRockafellarConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hClosed : ClosedConvexBifunction F) (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) (hFiniteSections : ∀ u : Fin m → ℝ, ∃ x : Fin n → ℝ, F u x < (⊤ : EReal)) : HasInnerProductEquation F

Lean source

Helper for Text 34.1.4: the lower-representation pairing equality can likewise be derived from an actual closed-convex witness, without assuming underline(K) is globally no-⊥.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_pairingEquality_everywhere_of_closedConvexLowerRepresentation {K : SaddleFunction m n} {h : IsConcaveConvex K} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hRock : IsRockafellarConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hClosed : ClosedConvexBifunction F) (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) (hFiniteSections : ∀ u : Fin m → ℝ, ∃ x : Fin n → ℝ, F u x < (⊤ : EReal)) : ∀ u : Fin m → ℝ, ∀ xStar : Fin n → ℝ, convexBifunctionPairing F u xStar = convexBifunctionCanonicalAdjointPairing F xStar u

Lean source

Helper for Text 34.1.4: a closed convex bifunction witness is already graph-convex, because closedness makes every section fixed by the one-variable convex closure and Rockafellar convexity then upgrades to joint graph convexity.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_graphConvex_of_closedConvexBifunction {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hRock : IsRockafellarConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hClosed : ClosedConvexBifunction F) : IsGraphConvexBifunction F

Lean source

Helper for Text 34.1.4: a genuine lower self-representation already forces equality of the translated-and-tilted primal and dual optimal values for the fixed witness.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_equalOptimalValues_for_translatedTiltedPrograms_of_genuineLowerSelfRepresentation (K : SaddleFunction m n) (h : IsConcaveConvex K) {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hRock : IsRockafellarConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hClosed : ClosedConvexBifunction F) (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) (hGenuineSelfRep : ∀ u : Fin m → ℝ, ∀ xStar : Fin n → ℝ, lowerClosureConcaveConvex K h u xStar = genuineConvexBifunctionAdjointPairing F u xStar) : ∀ u : Fin m → ℝ, ∀ xStar : Fin n → ℝ, HasEqualOptimalValuesForTranslatedTiltedPrograms F u xStar

Lean source

Helper for Text 34.1.4: the same genuine lower self-representation makes every translated-and-tilted program normal for the fixed witness.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_normality_for_translatedTiltedPrograms_of_genuineLowerSelfRepresentation (K : SaddleFunction m n) (h : IsConcaveConvex K) {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hRock : IsRockafellarConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hClosed : ClosedConvexBifunction F) (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) (hGenuineSelfRep : ∀ u : Fin m → ℝ, ∀ xStar : Fin n → ℝ, lowerClosureConcaveConvex K h u xStar = genuineConvexBifunctionAdjointPairing F u xStar) : ∀ u : Fin m → ℝ, ∀ xStar : Fin n → ℝ, HasNormalityForTranslatedTiltedPrograms F u xStar

Lean source

Helper for Text 34.1.4: exact recovery of cl₂ overline(K) identifies the textbook upper closure with the canonical partner cl₁ underline(K).

open Set variable {m n : ℕ} lemma helperForText_34_1_4_firstClosureOfLower_eq_upper_of_secondClosure_eq_lower (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) (hRecover : partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h) : partialClosure₁ (lowerClosureConcaveConvex K h) = upperClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: once overline(K) is identified with the canonical upper partner cl₁ underline(K), the Section 33 witness for that canonical pair is exactly the desired witness for the textbook lower and upper closures.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_closedConvexWitness_exists_of_firstClosureOfLower_eq_upper_of_noTopOrBot (K : SaddleFunction m n) (h : IsConcaveConvex K) (hKNoBot : HasNoBotValuesBifunction K) (hLowerNoTopBot : HasNoTopOrBotValuesBifunction (lowerClosureConcaveConvex K h)) (hFirst : partialClosure₁ (lowerClosureConcaveConvex K h) = upperClosureConcaveConvex K h) (hRealization : Section34CanonicalClosureRealizationQualification m n) : ∃ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ ClosedConvexBifunction F ∧ lowerClosureConcaveConvex K h = convexBifunctionPairing F ∧ upperClosureConcaveConvex K h = helperForText_34_0_1_convexAdjointPairingKernel F

§34 — Closures and Equivalence Classes — Part 5

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

The translated-tilted value equality rewrites its primal value at the origin as the corresponding Chapter 6 dual-program value.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_translatedTilted_originPrimal_eq_dualProgram {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) (hGClosed : ClosedConvexBifunction (translatedTiltedBifunction F u xStar)) (hEqualValues : HasEqualOptimalValuesForTranslatedTiltedPrograms F u xStar) : convexProgramAssociatedWith (translatedTiltedBifunction F u xStar) 0 = dualProgramOfConvexProgram ⟨translatedTiltedBifunction F u xStar, hGClosed.1⟩

Lean source

Equal translated primal and dual values exclude the exceptional (top, bottom) branch.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_translatedTilted_notBothInconsistent {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) (hGClosed : ClosedConvexBifunction (translatedTiltedBifunction F u xStar)) (hEqualValues : HasEqualOptimalValuesForTranslatedTiltedPrograms F u xStar) : ¬ (convexProgramAssociatedWith (translatedTiltedBifunction F u xStar) 0 = (⊤ : EReal) ∧ dualProgramOfConvexProgram ⟨translatedTiltedBifunction F u xStar, hGClosed.1⟩ = (⊥ : EReal))

Lean source

Corollary 6.30.3 applied after excluding the exceptional translated branch.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_translatedTilted_liminf_limsup_package {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) (hGClosed : ClosedConvexBifunction (translatedTiltedBifunction F u xStar)) (hEqualValues : HasEqualOptimalValuesForTranslatedTiltedPrograms F u xStar) : Filter.liminf (convexProgramAssociatedWith (translatedTiltedBifunction F u xStar)) (nhds (0 : Fin m → ℝ)) = dualPerturbationFunctionOfConvexProgram ⟨translatedTiltedBifunction F u xStar, hGClosed.1⟩ 0 ∧ Filter.limsup (dualPerturbationFunctionOfConvexProgram ⟨translatedTiltedBifunction F u xStar, hGClosed.1⟩) (nhds (0 : Fin n → ℝ)) = convexProgramAssociatedWith (translatedTiltedBifunction F u xStar) 0

Lean source

Helper for Text 34.1.4: the textbook upper closure agrees with the canonical upper partner exactly when the raw mixed-order comparison holds.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_textbookUpper_eq_canonicalUpper_iff_rawMixedClosureOrder (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) : upperClosureConcaveConvex K h = partialClosure₁ (lowerClosureConcaveConvex K h) ↔ partialClosure₂ (partialClosure₁ K) ≤ partialClosure₁ (partialClosure₂ K)

Lean source

Helper for Text 34.1.4: once the exact recovery cl₂ overline(K) = underline(K) is available, the textbook upper closure is the canonical Section 33 upper partner.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_textbookUpper_eq_canonicalUpper_of_secondClosure_eq_lower (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) (hRecover : partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h) : upperClosureConcaveConvex K h = partialClosure₁ (lowerClosureConcaveConvex K h)

Lean source

Helper for Text 34.1.4: the mixed-order comparison identifies the textbook upper closure with the canonical upper partner.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_textbookUpper_eq_canonicalUpper_of_mixedClosure_order (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) (hOrder : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h) : upperClosureConcaveConvex K h = partialClosure₁ (lowerClosureConcaveConvex K h)

Lean source

Helper for Text 34.1.4: the raw mixed-order comparison identifies the textbook upper closure with the canonical upper partner.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_textbookUpper_eq_canonicalUpper_of_rawMixedClosureOrder (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) (hRawOrder : partialClosure₂ (partialClosure₁ K) ≤ partialClosure₁ (partialClosure₂ K)) : upperClosureConcaveConvex K h = partialClosure₁ (lowerClosureConcaveConvex K h)

§34 — Closures and Equivalence Classes — Part 6

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Predicates on coordinate spaces in this section are treated classically when needed in piecewise definitions.

open Set noncomputable local instance classicalDecidablePredPart6 {α : Type*} (p : α → Prop) : DecidablePred p

Lean source

Helper for Text 34.1.4: at the primal origin, the translated-tilted genuine adjoint pairing is just the original genuine adjoint pairing with the dual variable shifted by the tilt.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_translatedTilted_genuineAdjointPairingAtZero_eq_shifted {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) (yStar : Fin n → ℝ) : genuineConvexBifunctionAdjointPairing (translatedTiltedBifunction F u xStar) (0 : Fin m → ℝ) yStar = genuineConvexBifunctionAdjointPairing F u (xStar + yStar)

Lean source

Helper for Text 34.1.4: the translated dual perturbation family is the negative of the shifted genuine adjoint pairing.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_translatedDualPerturbation_eq_neg_shiftedGenuineAdjointPairing {m n : ℕ} {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) (hGClosed : ClosedConvexBifunction (translatedTiltedBifunction F u xStar)) : dualPerturbationFunctionOfConvexProgram ⟨translatedTiltedBifunction F u xStar, hGClosed.1⟩ = fun yStar : Fin n → ℝ => -genuineConvexBifunctionAdjointPairing F u (xStar + yStar)

Lean source

Helper for Text 34.1.4: the translated primal perturbation family is the negative of the shifted lower-closure pairing section attached to the fixed witness.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_translatedPrimalValue_eq_neg_shiftedLowerClosure (K : SaddleFunction m n) (h : IsConcaveConvex K) {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) : convexProgramAssociatedWith (translatedTiltedBifunction F u xStar) = fun v : Fin m → ℝ => -lowerClosureConcaveConvex K h (u + v) xStar

Lean source

Helper for Text 34.1.4: after rewriting the translated Chapter 6 package through the fixed lower self-representation, both the primal liminf and the dual limsup are statements about the negative shifted lower-closure sections.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_translatedTilted_liminf_limsup_as_neg_shiftedLowerClosure (K : SaddleFunction m n) (h : IsConcaveConvex K) {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) (hGClosed : ClosedConvexBifunction (translatedTiltedBifunction F u xStar)) (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) (hGenuineSelfRep : ∀ u : Fin m → ℝ, ∀ xStar : Fin n → ℝ, lowerClosureConcaveConvex K h u xStar = genuineConvexBifunctionAdjointPairing F u xStar) (hEqualValues : HasEqualOptimalValuesForTranslatedTiltedPrograms F u xStar) : Filter.liminf (fun v : Fin m → ℝ => -lowerClosureConcaveConvex K h (u + v) xStar) (nhds (0 : Fin m → ℝ)) = -lowerClosureConcaveConvex K h u xStar ∧ Filter.limsup (fun yStar : Fin n → ℝ => -lowerClosureConcaveConvex K h u (xStar + yStar)) (nhds (0 : Fin n → ℝ)) = -lowerClosureConcaveConvex K h u xStar

Lean source

Helper for Text 34.1.4: the translated primal section of underline(K) has convex closure at the origin equal to the negated lower-closure value at (u, xStar).

open Set variable {m n : ℕ} lemma helperForText_34_1_4_shiftedPrimalSection_convexClosureAtZero (K : SaddleFunction m n) (h : IsConcaveConvex K) {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) (hLowerNoBot : HasNoBotValuesBifunction (lowerClosureConcaveConvex K h)) (hGClosed : ClosedConvexBifunction (translatedTiltedBifunction F u xStar)) (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) (hGenuineSelfRep : ∀ u : Fin m → ℝ, ∀ xStar : Fin n → ℝ, lowerClosureConcaveConvex K h u xStar = genuineConvexBifunctionAdjointPairing F u xStar) (hEqualValues : HasEqualOptimalValuesForTranslatedTiltedPrograms F u xStar) : convexClosure (fun v : Fin m → ℝ => -lowerClosureConcaveConvex K h (u + v) xStar) (0 : Fin m → ℝ) = -lowerClosureConcaveConvex K h u xStar

Lean source

Helper for Text 34.1.4: every second-variable section of underline(K) is already fixed by the raw one-variable Section 33 convex closure operator.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_lowerClosure_secondSection_isFunctionConvexClosed (K : SaddleFunction m n) (h : IsConcaveConvex K) (u : Fin m → ℝ) : IsFunctionConvexClosed (fun y : Fin n → ℝ => lowerClosureConcaveConvex K h u y)

Lean source

Helper for Text 34.1.4: the shifted dual section of underline(K) is already concave-closed at the origin, so its concave closure returns the negated lower-closure value at (u, xStar).

open Set variable {m n : ℕ} lemma helperForText_34_1_4_shiftedDualSection_concaveClosureAtZero (K : SaddleFunction m n) (h : IsConcaveConvex K) {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (xStar : Fin n → ℝ) (hLowerNoBot : HasNoBotValuesBifunction (lowerClosureConcaveConvex K h)) (_hGClosed : ClosedConvexBifunction (translatedTiltedBifunction F u xStar)) (_hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) (_hGenuineSelfRep : ∀ u : Fin m → ℝ, ∀ xStar : Fin n → ℝ, lowerClosureConcaveConvex K h u xStar = genuineConvexBifunctionAdjointPairing F u xStar) (_hEqualValues : HasEqualOptimalValuesForTranslatedTiltedPrograms F u xStar) : concaveClosure (fun yStar : Fin n → ℝ => -lowerClosureConcaveConvex K h u (xStar + yStar)) (0 : Fin n → ℝ) = -lowerClosureConcaveConvex K h u xStar

Lean source

Helper for Text 34.1.4: the neighborhood limsup of the shifted dual lower-closure section can be written as an infimum of suprema over closed balls centered at 0.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_shiftedDualSection_limsup_eq_iInf_iSup_closedBall (K : SaddleFunction m n) (h : IsConcaveConvex K) (u : Fin m → ℝ) (xStar : Fin n → ℝ) : Filter.limsup (fun yStar : Fin n → ℝ => -lowerClosureConcaveConvex K h u (xStar + yStar)) (nhds (0 : Fin n → ℝ)) = (⨅ (ρ : {ρ : ℝ // 0 < ρ}), ⨆ (yStar : {yStar : Fin n → ℝ // ‖yStar‖ ≤ ρ.1}), -lowerClosureConcaveConvex K h u (xStar + yStar.1))

Lean source

Helper for Text 34.1.4: capping the positive closed-ball radii by a fixed δ only enlarges the shifted dual envelope, so negating still yields a lower bound for underline(K)(u, xStar).

open Set variable {m n : ℕ} lemma helperForText_34_1_4_negRestrictedDualClosedBallEnvelope_le_lowerClosure (K : SaddleFunction m n) (h : IsConcaveConvex K) {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) (hGenuineSelfRep : ∀ u : Fin m → ℝ, ∀ xStar : Fin n → ℝ, lowerClosureConcaveConvex K h u xStar = genuineConvexBifunctionAdjointPairing F u xStar) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (δ : {δ : ℝ // 0 < δ}) (hGClosed : ClosedConvexBifunction (translatedTiltedBifunction F u xStar)) (hEqualValues : HasEqualOptimalValuesForTranslatedTiltedPrograms F u xStar) : -(⨅ (ρ : {ρ : ℝ // 0 < ρ ∧ ρ < δ.1}), ⨆ (yStar : {yStar : Fin n → ℝ // ‖yStar‖ ≤ ρ.1}), -lowerClosureConcaveConvex K h u (xStar + yStar.1)) ≤ lowerClosureConcaveConvex K h u xStar

Lean source

Helper for Text 34.1.4: the neighborhood liminf of the shifted primal lower-closure section can be written as a supremum of infima over closed balls centered at 0.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_shiftedPrimalSection_liminf_eq_iSup_iInf_closedBall (K : SaddleFunction m n) (h : IsConcaveConvex K) (u : Fin m → ℝ) (xStar : Fin n → ℝ) : Filter.liminf (fun v : Fin m → ℝ => -lowerClosureConcaveConvex K h (u + v) xStar) (nhds (0 : Fin m → ℝ)) = (⨆ (ρ : {ρ : ℝ // 0 < ρ}), ⨅ (v : {v : Fin m → ℝ // ‖v‖ ≤ ρ.1}), -lowerClosureConcaveConvex K h (u + v.1) xStar)

Lean source

Helper for Text 34.1.4: for a monotone positive-radius family, restricting the outer infimum to radii below a fixed cap does not change its value. The small-radius family {ρ | 0 < ρ ∧ ρ < δ} is cofinal among all positive radii, so the infimum over all positive radii can be recovered by capping at any fixed δ > 0.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_iInf_restrictedPositiveRadii_eq_iInf_of_monotone (δ : {δ : ℝ // 0 < δ}) (g : {ρ : ℝ // 0 < ρ} → EReal) (hMono : Monotone g) : (⨅ (ρ : {ρ : ℝ // 0 < ρ ∧ ρ < δ.1}), g ⟨ρ.1, ρ.2.1⟩) = (⨅ (ρ : {ρ : ℝ // 0 < ρ}), g ρ)

Lean source

Helper for Text 34.1.4: capping the positive closed-ball radii by a fixed ε only shrinks the shifted primal envelope, so negating still leaves an upper bound for underline(K)(u, xStar).

open Set variable {m n : ℕ} lemma helperForText_34_1_4_lowerClosure_le_negRestrictedPrimalClosedBallEnvelope (K : SaddleFunction m n) (h : IsConcaveConvex K) {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) (hGenuineSelfRep : ∀ u : Fin m → ℝ, ∀ xStar : Fin n → ℝ, lowerClosureConcaveConvex K h u xStar = genuineConvexBifunctionAdjointPairing F u xStar) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (hGClosed : ClosedConvexBifunction (translatedTiltedBifunction F u xStar)) (hEqualValues : HasEqualOptimalValuesForTranslatedTiltedPrograms F u xStar) : lowerClosureConcaveConvex K h u xStar ≤ -(⨆ (ρ : {ρ : ℝ // 0 < ρ ∧ ρ < ε.1}), ⨅ (v : {v : Fin m → ℝ // ‖v‖ ≤ ρ.1}), -lowerClosureConcaveConvex K h (u + v.1) xStar)

Lean source

Helper for Text 34.1.4: for an antitone positive-radius family, restricting the outer supremum to radii below a fixed cap does not change its value. As with the corresponding infimum lemma, the radii below a fixed positive cap form a cofinal subfamily among all positive radii.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_iSup_restrictedPositiveRadii_eq_iSup_of_antitone (ε : {ε : ℝ // 0 < ε}) (g : {ρ : ℝ // 0 < ρ} → EReal) (hAnti : Antitone g) : (⨆ (ρ : {ρ : ℝ // 0 < ρ ∧ ρ < ε.1}), g ⟨ρ.1, ρ.2.1⟩) = (⨆ (ρ : {ρ : ℝ // 0 < ρ}), g ρ)

Lean source

Helper for Text 34.1.4: once the translated-tilted package is available, capping the shifted dual closed-ball envelope at any positive radius δ still gives exactly underline(K)(u, xStar). This sharpens the earlier one-sided inequality by using the closed-ball neighborhood formula for limsup together with the cofinality of the capped radius family.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_negRestrictedDualClosedBallEnvelope_eq_lowerClosure (K : SaddleFunction m n) (h : IsConcaveConvex K) {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) (hGenuineSelfRep : ∀ u : Fin m → ℝ, ∀ xStar : Fin n → ℝ, lowerClosureConcaveConvex K h u xStar = genuineConvexBifunctionAdjointPairing F u xStar) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (δ : {δ : ℝ // 0 < δ}) (hGClosed : ClosedConvexBifunction (translatedTiltedBifunction F u xStar)) (hEqualValues : HasEqualOptimalValuesForTranslatedTiltedPrograms F u xStar) : -(⨅ (ρ : {ρ : ℝ // 0 < ρ ∧ ρ < δ.1}), ⨆ (yStar : {yStar : Fin n → ℝ // ‖yStar‖ ≤ ρ.1}), -lowerClosureConcaveConvex K h u (xStar + yStar.1)) = lowerClosureConcaveConvex K h u xStar

Lean source

Helper for Text 34.1.4: symmetrically, once the translated-tilted package is available, capping the shifted primal closed-ball envelope at any positive radius ε still gives exactly underline(K)(u, xStar). This is the liminf analogue of the previous dual-envelope identity.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_lowerClosure_eq_negRestrictedPrimalClosedBallEnvelope (K : SaddleFunction m n) (h : IsConcaveConvex K) {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) (hGenuineSelfRep : ∀ u : Fin m → ℝ, ∀ xStar : Fin n → ℝ, lowerClosureConcaveConvex K h u xStar = genuineConvexBifunctionAdjointPairing F u xStar) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (hGClosed : ClosedConvexBifunction (translatedTiltedBifunction F u xStar)) (hEqualValues : HasEqualOptimalValuesForTranslatedTiltedPrograms F u xStar) : lowerClosureConcaveConvex K h u xStar = -(⨆ (ρ : {ρ : ℝ // 0 < ρ ∧ ρ < ε.1}), ⨅ (v : {v : Fin m → ℝ // ‖v‖ ≤ ρ.1}), -lowerClosureConcaveConvex K h (u + v.1) xStar)

Lean source

Helper for Text 34.1.4: at fixed radii, the open-ball minimax envelope of K dominates the corresponding first partial-closure envelope.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_restrictedFirstClosureEnvelope_le_leftOpenBallEnvelope (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) : (⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), partialClosure₁ K u z.1) ≤ (⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), K w.1 z.1)

Lean source

Helper for Text 34.1.4: at fixed radii, the open-ball maximin envelope of K lies below the corresponding second partial-closure envelope.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_rightOpenBallEnvelope_le_restrictedSecondClosureEnvelope (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) : (⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), K w.1 z.1) ≤ (⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), partialClosure₂ K w.1 xStar)

Lean source

Helper for Text 34.1.4: the left open-ball minimax envelope of K dominates the same restricted envelope built from the mixed upper closure cl₁ cl₂ K.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_restrictedUpperMixedClosureEnvelope_le_leftOpenBallEnvelope (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) : (⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), partialClosure₁ (partialClosure₂ K) u z.1) ≤ (⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), K w.1 z.1)

Lean source

Helper for Text 34.1.4: the right open-ball maximin envelope of K lies below the same restricted envelope built from the mixed lower closure cl₂ cl₁ K.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_rightOpenBallEnvelope_le_restrictedLowerMixedClosureEnvelope (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) : (⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), K w.1 z.1) ≤ (⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), partialClosure₂ (partialClosure₁ K) w.1 xStar)

Lean source

Helper for Text 34.1.4: the naive left transport fixed-ball minimax ≤ underline(K)(u, xStar) is false in general. For the one-dimensional coordinate kernel K(w, z) = w₀, the local minimax value on the unit open balls around the origin is at least 1/2, while the mixed lower closure at the origin is still 0. This rules out any replan that merely retries the current left-sandwich statement.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_coordKernel_naiveLeftTransport_false : ¬ ((⨅ (z : {z : Fin 1 → ℝ // ‖z - (0 : Fin 1 → ℝ)‖ < 1}), ⨆ (w : {w : Fin 1 → ℝ // ‖w - (0 : Fin 1 → ℝ)‖ < 1}), helperForText_34_1_4_coordKernel w.1 z.1) ≤ partialClosure₂ (partialClosure₁ helperForText_34_1_4_coordKernel) 0 0)

§34 — Closures and Equivalence Classes — Part 7

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Predicates on coordinate spaces in this section are treated classically when needed in piecewise definitions.

open Set noncomputable local instance classicalDecidablePredPart7 {α : Type*} (p : α → Prop) : DecidablePred p

Lean source

Helper for Text 34.1.4: the proposed left closed-ball bridge is already false for the one-dimensional coordinate kernel.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_leftClosedBallEnvelope_bridge_false : ¬ ((⨅ (ρz : {ρ : ℝ // ρ < (1 : ℝ)}), ⨅ (z : {z : Fin 1 → ℝ // ‖z - (0 : Fin 1 → ℝ)‖ ≤ ρz.1}), ⨆ (ρw : {ρ : ℝ // ρ < (1 : ℝ)}), ⨆ (w : {w : Fin 1 → ℝ // ‖w - (0 : Fin 1 → ℝ)‖ ≤ ρw.1}), helperForText_34_1_4_coordKernel w.1 z.1) ≤ -(⨅ (ρ : {ρ : ℝ // 0 < ρ ∧ ρ < (1 : ℝ)}), ⨆ (yStar : {yStar : Fin 1 → ℝ // ‖yStar‖ ≤ ρ.1}), -lowerClosureConcaveConvex helperForText_34_1_4_coordKernel helperForText_34_1_4_coordKernel_isConcaveConvex 0 ((0 : Fin 1 → ℝ) + yStar.1)))

Lean source

Helper for Text 34.1.4: the proposed right closed-ball bridge is already false for the one-dimensional second-coordinate kernel.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_rightClosedBallEnvelope_bridge_false : ¬ (-(⨆ (ρ : {ρ : ℝ // 0 < ρ ∧ ρ < (1 : ℝ)}), ⨅ (v : {v : Fin 1 → ℝ // ‖v‖ ≤ ρ.1}), -lowerClosureConcaveConvex helperForText_34_1_4_secondCoordKernel helperForText_34_1_4_secondCoordKernel_isConcaveConvex ((0 : Fin 1 → ℝ) + v.1) 0) ≤ (⨆ (ρw : {ρ : ℝ // ρ < (1 : ℝ)}), ⨆ (w : {w : Fin 1 → ℝ // ‖w - (0 : Fin 1 → ℝ)‖ ≤ ρw.1}), ⨅ (ρz : {ρ : ℝ // ρ < (1 : ℝ)}), ⨅ (z : {z : Fin 1 → ℝ // ‖z - (0 : Fin 1 → ℝ)‖ ≤ ρz.1}), helperForText_34_1_4_secondCoordKernel w.1 z.1))

Lean source

Helper for Text 34.1.4: the translated-tilted package already squeezes the base lower-closure value between the two shifted closed-ball envelopes that come directly from the Chapter 6 liminf/limsup identities. This packages the exact dependency-closed information available before any transport back to the raw fixed-ball envelopes of K.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_lowerClosure_between_shiftedRestrictedClosedBallEnvelopes (K : SaddleFunction m n) (h : IsConcaveConvex K) {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) (hGenuineSelfRep : ∀ u : Fin m → ℝ, ∀ xStar : Fin n → ℝ, lowerClosureConcaveConvex K h u xStar = genuineConvexBifunctionAdjointPairing F u xStar) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) (hGClosed : ClosedConvexBifunction (translatedTiltedBifunction F u xStar)) (hEqualValues : HasEqualOptimalValuesForTranslatedTiltedPrograms F u xStar) : (-(⨅ (ρ : {ρ : ℝ // 0 < ρ ∧ ρ < δ.1}), ⨆ (yStar : {yStar : Fin n → ℝ // ‖yStar‖ ≤ ρ.1}), -lowerClosureConcaveConvex K h u (xStar + yStar.1)) ≤ lowerClosureConcaveConvex K h u xStar) ∧ (lowerClosureConcaveConvex K h u xStar ≤ -(⨆ (ρ : {ρ : ℝ // 0 < ρ ∧ ρ < ε.1}), ⨅ (v : {v : Fin m → ℝ // ‖v‖ ≤ ρ.1}), -lowerClosureConcaveConvex K h (u + v.1) xStar))

Lean source

Helper for Text 34.1.4: every translated-and-tilted program attached to a closed-convex witness for underline(K) is itself a closed convex bifunction. This isolates the Chapter 6 packaging needed by the remaining transport argument: once a fixed witness represents underline(K), the translated objective (v, y) ↦ F (u + v) y - ⟪y, xStar⟫ inherits closed convexity from the original witness.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_translatedTiltedBifunction_isClosedConvex_of_lowerRepresentation (K : SaddleFunction m n) (h : IsConcaveConvex K) {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hRock : IsRockafellarConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) (hClosed : ClosedConvexBifunction F) (hLowerRep : lowerClosureConcaveConvex K h = convexBifunctionPairing F) (hFiniteSections : ∀ u : Fin m → ℝ, ∃ x : Fin n → ℝ, F u x < (⊤ : EReal)) (u : Fin m → ℝ) (xStar : Fin n → ℝ) : ClosedConvexBifunction (translatedTiltedBifunction F u xStar)

Lean source

Helper for Text 34.1.4: for a fixed second-variable point, the closed-ball supremum envelope around u is unchanged when the radius index is restricted to positive values below ε. Nonpositive radii contribute only the center point u (or the empty ball), and every positive closed ball still contains that same center.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_iSup_closedBallValues_eq_iSup_positiveRadii (K : SaddleFunction m n) (u : Fin m → ℝ) (z : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) : (⨆ (ρw : {ρ : ℝ // ρ < ε.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ ≤ ρw.1}), K w.1 z) = (⨆ (ρw : {ρ : ℝ // 0 < ρ ∧ ρ < ε.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ ≤ ρw.1}), K w.1 z)

Lean source

Helper for Text 34.1.4: for a fixed first-variable point, the closed-ball infimum envelope around xStar is unchanged when the radius index is restricted to positive values below δ. As on the primal side, nonpositive radii contribute only the base point xStar or the empty ball, both of which are already controlled by the positive-radius family.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_iInf_closedBallValues_eq_iInf_positiveRadii (K : SaddleFunction m n) (w : Fin m → ℝ) (xStar : Fin n → ℝ) (δ : {δ : ℝ // 0 < δ}) : (⨅ (ρz : {ρ : ℝ // ρ < δ.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ ≤ ρz.1}), K w z.1) = (⨅ (ρz : {ρ : ℝ // 0 < ρ ∧ ρ < δ.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ ≤ ρz.1}), K w z.1)

Lean source

Helper for Text 34.1.4: on the left closed-ball-radius minimax envelope, both the outer second-variable radii and the inner first-variable radii may be restricted to positive values without changing the resulting value.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_leftClosedBallRadiusEnvelope_eq_positiveRadii (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) : (⨅ (ρz : {ρ : ℝ // ρ < δ.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ ≤ ρz.1}), ⨆ (ρw : {ρ : ℝ // ρ < ε.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ ≤ ρw.1}), K w.1 z.1) = (⨅ (ρz : {ρ : ℝ // 0 < ρ ∧ ρ < δ.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ ≤ ρz.1}), ⨆ (ρw : {ρ : ℝ // 0 < ρ ∧ ρ < ε.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ ≤ ρw.1}), K w.1 z.1)

Lean source

Helper for Text 34.1.4: on the right closed-ball-radius maximin envelope, both the outer first-variable radii and the inner second-variable radii may be restricted to positive values without changing the resulting value.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_rightClosedBallRadiusEnvelope_eq_positiveRadii (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) : (⨆ (ρw : {ρ : ℝ // ρ < ε.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ ≤ ρw.1}), ⨅ (ρz : {ρ : ℝ // ρ < δ.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ ≤ ρz.1}), K w.1 z.1) = (⨆ (ρw : {ρ : ℝ // 0 < ρ ∧ ρ < ε.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ ≤ ρw.1}), ⨅ (ρz : {ρ : ℝ // 0 < ρ ∧ ρ < δ.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ ≤ ρz.1}), K w.1 z.1)

Lean source

Helper for Text 34.1.4: after restricting both radius families to positive values, the left closed-ball-radius envelope is exactly the open-ball minimax expression on the same neighborhood pair.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_positiveRadiusLeftEnvelope_eq_fixedNeighborhood_minimax (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) : (⨅ (ρz : {ρ : ℝ // 0 < ρ ∧ ρ < δ.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ ≤ ρz.1}), ⨆ (ρw : {ρ : ℝ // 0 < ρ ∧ ρ < ε.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ ≤ ρw.1}), K w.1 z.1) = (⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), K w.1 z.1)

Lean source

Helper for Text 34.1.4: after restricting both radius families to positive values, the right closed-ball-radius envelope is exactly the open-ball maximin expression on the same neighborhood pair.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_positiveRadiusRightEnvelope_eq_fixedNeighborhood_maximin (K : SaddleFunction m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) (ε : {ε : ℝ // 0 < ε}) (δ : {δ : ℝ // 0 < δ}) : (⨆ (ρw : {ρ : ℝ // 0 < ρ ∧ ρ < ε.1}), ⨆ (w : {w : Fin m → ℝ // ‖w - u‖ ≤ ρw.1}), ⨅ (ρz : {ρ : ℝ // 0 < ρ ∧ ρ < δ.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ ≤ ρz.1}), K w.1 z.1) = (⨆ (w : {w : Fin m → ℝ // ‖w - u‖ < ε.1}), ⨅ (z : {z : Fin n → ℝ // ‖z - xStar‖ < δ.1}), K w.1 z.1)

§34 — Closures and Equivalence Classes — Part 8

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Predicates on coordinate spaces in this section are treated classically when needed in piecewise definitions.

open Set noncomputable local instance classicalDecidablePredPart8 {α : Type*} (p : α → Prop) : DecidablePred p

Lean source

Helper for Text 34.1.4: every constant kernel is fixed by the first partial closure.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_partialClosure₁_constantKernel (c : ℝ) : partialClosure₁ (fun (_ : Fin m → ℝ) (_ : Fin n → ℝ) => ((c : ℝ) : EReal)) = (fun (_ : Fin m → ℝ) (_ : Fin n → ℝ) => ((c : ℝ) : EReal))

Lean source

Helper for Text 34.1.4: every constant kernel is fixed by the second partial closure.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_partialClosure₂_constantKernel (c : ℝ) : partialClosure₂ (fun (_ : Fin m → ℝ) (_ : Fin n → ℝ) => ((c : ℝ) : EReal)) = (fun (_ : Fin m → ℝ) (_ : Fin n → ℝ) => ((c : ℝ) : EReal))

Lean source

Helper for Text 34.1.4: a constant kernel is concave-convex.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_constantKernel_isConcaveConvex (c : ℝ) : IsConcaveConvex (fun (_ : Fin m → ℝ) (_ : Fin n → ℝ) => ((c : ℝ) : EReal))

Lean source

Helper for Text 34.1.4: a constant kernel is upper closed in the Section 33 saddle sense.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_constantKernel_isUpperClosedSaddleFunction (c : ℝ) : IsUpperClosedSaddleFunction (fun (_ : Fin m → ℝ) (_ : Fin n → ℝ) => ((c : ℝ) : EReal))

Lean source

Helper for Text 34.1.4: a closed convex bifunction with no ⊥ values is already Rockafellar convex.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_rockafellarConvex_of_closedConvexBifunction {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hClosed : ClosedConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) : IsRockafellarConvexBifunction F

Lean source

Helper for Text 34.1.4: the constant zero kernel satisfies the corrected Corollary 33.3.1, so it has a canonical Section 33 witness.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_zeroKernel_hasCanonicalWitness (hRealization : Section34CanonicalClosureRealizationQualification 1 1) : ∃ F : (Fin 1 → ℝ) → (Fin 1 → ℝ) → EReal, IsRockafellarConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ ClosedConvexBifunction F ∧ (fun (_ : Fin 1 → ℝ) (_ : Fin 1 → ℝ) => ((0 : ℝ) : EReal)) = convexBifunctionPairing F ∧ partialClosure₁ (fun (_ : Fin 1 → ℝ) (_ : Fin 1 → ℝ) => ((0 : ℝ) : EReal)) = helperForText_34_0_1_convexAdjointPairingKernel F

Lean source

Helper for Text 34.1.4: the abstract upper-partner minimality statement requested in the previous replans is false without an additional witness hypothesis. A constant -1 kernel is upper closed and lies below the fully fixed constant 0 kernel, but the two kernels are not equal.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_upperPartnerMinimality_withoutWitness_false : ∃ L U : SaddleFunction 1 1, partialClosure₁ L = L ∧ partialClosure₂ L = L ∧ IsUpperClosedSaddleFunction U ∧ U ≤ L ∧ U ≠ partialClosure₁ L

Lean source

Helper for Text 34.1.4: even a canonical Section 33 witness for a fully fixed lower representative does not, by itself, exclude an unrelated strict upper-closed minorant. This formal counterexample explains why the remaining blocker must mention the actual mixed upper closure overline(K), not just abstract witness data for underline(K).

open Set variable {m n : ℕ} lemma helperForText_34_1_4_canonicalWitness_doesNotExclude_strictUpperClosedMinorant (hRealization : Section34CanonicalClosureRealizationQualification 1 1) : ∃ L U : SaddleFunction 1 1, (∃ F : (Fin 1 → ℝ) → (Fin 1 → ℝ) → EReal, IsRockafellarConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ ClosedConvexBifunction F ∧ L = convexBifunctionPairing F ∧ partialClosure₁ L = helperForText_34_0_1_convexAdjointPairingKernel F) ∧ partialClosure₁ L = L ∧ partialClosure₂ L = L ∧ IsUpperClosedSaddleFunction U ∧ U ≤ L ∧ U ≠ L

§34 — Closures and Equivalence Classes — Part 9

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Predicates on coordinate spaces in this section are treated classically when needed in piecewise definitions.

open Set noncomputable local instance classicalDecidablePredPart9 {α : Type*} (p : α → Prop) : DecidablePred p

Lean source

Text 34.1.4, qualified formal version: if K is concave-convex, its mixed closures are ordered, and the outer closures are fixed in their respective variables, then applying the opposite partial closures exchanges the mixed pair.

open Set variable {m n : ℕ} theorem section34_text_34_1_4 (K : SaddleFunction m n) (h : IsConcaveConvex K) (hOrder : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h) (hLowerClosed : IsConvexClosedInSecond (lowerClosureConcaveConvex K h)) (hUpperClosed : IsConcaveClosedInFirst (upperClosureConcaveConvex K h)) : partialClosure₁ (lowerClosureConcaveConvex K h) = upperClosureConcaveConvex K h ∧ partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h

Lean source

Compatibility form of first-partial-closure idempotence for concave-convex kernels.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_partialClosure₁_idempotent (K : SaddleFunction m n) (_h : IsConcaveConvex K) : partialClosure₁ (partialClosure₁ K) = partialClosure₁ K

Lean source

Compatibility form of second-partial-closure idempotence for concave-convex kernels.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_partialClosure₂_idempotent (K : SaddleFunction m n) (_h : IsConcaveConvex K) : partialClosure₂ (partialClosure₂ K) = partialClosure₂ K

Lean source

Text 34.0.1: for a concave-convex saddle-function, the iterated lower and upper closure operations are idempotent, and consequently the lower closure is lower closed while the upper closure is upper closed.

open Set variable {m n : ℕ} theorem section34_idempotent_closures (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) : partialClosure₂ (partialClosure₁ (partialClosure₂ (partialClosure₁ K))) = partialClosure₂ (partialClosure₁ K) ∧ partialClosure₁ (partialClosure₂ (partialClosure₁ (partialClosure₂ K))) = partialClosure₁ (partialClosure₂ K) ∧ IsLowerClosed (lowerClosureConcaveConvex K h) ∧ IsUpperClosed (upperClosureConcaveConvex K h)

Lean source

Theorem 34.1: if K is any saddle-function on ℝ^m × ℝ^n, then its lower closure is a lower closed saddle-function and its upper closure is an upper closed saddle-function.

open Set variable {m n : ℕ} theorem section34_lower_and_upper_closures_are_closed (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) : IsLowerClosed (lowerClosureConcaveConvex K h) ∧ IsUpperClosed (upperClosureConcaveConvex K h)

Lean source

Helper for Text 34.1.4: after checking both the local-minimax route and the Section 33 uniqueness route, the unresolved dependency-closed blocker is exactly the original-text bridge cl₂ overline(K) = underline(K), from which the normalized raw mixed-order comparison follows formally.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_rawMixedClosureOrder_missingPrerequisite (K : SaddleFunction m n) (h : IsConcaveConvex K) (hOrder : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h) : partialClosure₂ (partialClosure₁ K) ≤ partialClosure₁ (partialClosure₂ K)

Lean source

Helper bridge for Text 34.1.4: the exact upstream reverse-minimax input needed by the proof pipeline is precisely the normalized raw mixed-order inequality cl₂ (cl₁ K) ≤ cl₁ (cl₂ K). This theorem is just the dependency-closed package of the existing missing-prerequisite lemma, exposed under the name the later endgame uses conceptually.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_reverseMinimax_bridge (K : SaddleFunction m n) (h : IsConcaveConvex K) (hOrder : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h) : partialClosure₂ (partialClosure₁ K) ≤ partialClosure₁ (partialClosure₂ K)

Lean source

Helper for Text 34.1.4: once the normalized raw mixed-order comparison is supplied, the exact recovery cl₂ overline(K) = underline(K) follows formally from the already-proved local endgame.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_secondClosureOfUpper_eq_lower_from_concaveConvex (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) (hOrder : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h) : partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: the only remaining upstream prerequisite is now the mixed-order comparison underline(K) ≤ overline(K) for a bare concave-convex saddle-function.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_mixedClosure_order (K : SaddleFunction m n) (h : IsConcaveConvex K) (hOrder : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h) : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: after isolating the true upstream blocker to the mixed-order comparison, the exact recovery cl₂ overline(K) = underline(K) follows formally.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_secondClosureOfUpper_eq_lower (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) (hOrder : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h) : partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: independently of the unresolved mixed-order comparison, the mixed upper closure always lies below cl₁ of the mixed lower closure.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_upperClosure_le_firstClosureOfLower (K : SaddleFunction m n) (h : IsConcaveConvex K) : upperClosureConcaveConvex K h ≤ partialClosure₁ (lowerClosureConcaveConvex K h)

Lean source

Helper for Text 34.1.4: once the mixed lower closure is known to lie below the mixed upper closure, first-variable closedness of the upper closure gives the reverse cl₁ bound.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_firstClosureOfLower_le_upper_of_mixedClosure_order (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) (hOrder : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h) : partialClosure₁ (lowerClosureConcaveConvex K h) ≤ upperClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: the mixed-order comparison already forces the first cross-closure identity cl₁ underline(K) = overline(K).

open Set variable {m n : ℕ} lemma helperForText_34_1_4_firstClosureOfLower_eq_upper_of_mixedClosure_order (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) (hOrder : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h) : partialClosure₁ (lowerClosureConcaveConvex K h) = upperClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: once the mixed lower closure is known to lie below the mixed upper closure, second-variable closedness of the lower closure gives the reverse cl₂ bound.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_lowerClosure_le_secondClosureOfUpper_of_mixedClosure_order (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) (hOrder : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h) : lowerClosureConcaveConvex K h ≤ partialClosure₂ (upperClosureConcaveConvex K h)

Lean source

Helper for Text 34.1.4: the raw mixed-order inequality on cl₂ (cl₁ K) and cl₁ (cl₂ K) already packages the full concave-convex cross-closure conclusion.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_concaveConvex_crossClosure_of_rawMixedClosureOrder (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) (hRawOrder : partialClosure₂ (partialClosure₁ K) ≤ partialClosure₁ (partialClosure₂ K)) : partialClosure₁ (lowerClosureConcaveConvex K h) = upperClosureConcaveConvex K h ∧ partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: once the exact recovery cl₂ overline(K) = underline(K) is known, both cross-closure identities follow immediately.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_concaveConvex_crossClosure_of_secondClosure_eq_lower (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) (hRecover : partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h) : partialClosure₁ (lowerClosureConcaveConvex K h) = upperClosureConcaveConvex K h ∧ partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: the full textbook pair of cross-closure identities is equivalent to the exact recovery statement cl₂ overline(K) = underline(K).

open Set variable {m n : ℕ} lemma helperForText_34_1_4_concaveConvex_crossClosure_iff_secondClosureOfUpper_eq_lower (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) : (partialClosure₁ (lowerClosureConcaveConvex K h) = upperClosureConcaveConvex K h ∧ partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h) ↔ partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: the first cross-closure identity already forces the textbook mixed-order comparison.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_mixedClosure_order_of_firstClosure_eq_upper (K : SaddleFunction m n) (h : IsConcaveConvex K) (hFirst : partialClosure₁ (lowerClosureConcaveConvex K h) = upperClosureConcaveConvex K h) : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h

Lean source

Helper for Text 34.1.4: the textbook mixed-order comparison is equivalent to the two cross-closure identities.

open Set variable {m n : ℕ} lemma helperForText_34_1_4_mixedClosure_order_iff_crossClosure (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h ↔ (partialClosure₁ (lowerClosureConcaveConvex K h) = upperClosureConcaveConvex K h ∧ partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h)

Lean source

Helper for Text 34.1.4: the full concave-convex textbook conclusion is equivalent to the raw operator inequality cl₂ (cl₁ K) ≤ cl₁ (cl₂ K).

open Set variable {m n : ℕ} lemma helperForText_34_1_4_concaveConvex_crossClosure_iff_rawMixedClosureOrder (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) : (partialClosure₁ (lowerClosureConcaveConvex K h) = upperClosureConcaveConvex K h ∧ partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h) ↔ partialClosure₂ (partialClosure₁ K) ≤ partialClosure₁ (partialClosure₂ K)

Lean source

Helper for Text 34.1.4: after the lower recovery cl₂ overline(K) = underline(K) is known, the first cross-closure identity is the short textbook rewrite cl₁ underline(K) = cl₁ (cl₂ overline(K)) = overline(K).

open Set variable {m n : ℕ} lemma helperForText_34_1_4_concaveConvex_crossClosure (K : SaddleFunction m n) (h : IsConcaveConvex K) (hOrder : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h) (hLowerClosed : IsConvexClosedInSecond (lowerClosureConcaveConvex K h)) (hUpperClosed : IsConcaveClosedInFirst (upperClosureConcaveConvex K h)) : partialClosure₁ (lowerClosureConcaveConvex K h) = upperClosureConcaveConvex K h ∧ partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h

Lean source

Helper for Text 34.0.1: once the mixed lower closure is known to lie below the mixed upper closure, the two cross-closure identities follow from monotonicity and the one-sided fixed-point properties of the outer closures.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_crossClosure_relations_of_mixedClosure_order (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) (hOrder : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h) : partialClosure₁ (lowerClosureConcaveConvex K h) = upperClosureConcaveConvex K h ∧ partialClosure₂ (upperClosureConcaveConvex K h) = lowerClosureConcaveConvex K h

Lean source

Helper for Text 34.0.1: the first partial closure is unconditionally idempotent.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_partialClosure₁_idempotent_unconditional (K : SaddleFunction m n) : partialClosure₁ (partialClosure₁ K) = partialClosure₁ K

Lean source

Helper for Text 34.0.1: the second partial closure is unconditionally idempotent.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_partialClosure₂_idempotent_unconditional (K : SaddleFunction m n) : partialClosure₂ (partialClosure₂ K) = partialClosure₂ K

Lean source

Helper for Text 34.0.1: both mixed closures are fixed points of repeating the lower or upper closure operation.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_lower_and_upper_fixedPoint_forms (K : SaddleFunction m n) (h : IsConcaveConvex K) : partialClosure₂ (partialClosure₁ (lowerClosureConcaveConvex K h)) = lowerClosureConcaveConvex K h ∧ partialClosure₁ (partialClosure₂ (upperClosureConcaveConvex K h)) = upperClosureConcaveConvex K h

Lean source

Helper for Text 34.0.1: a mixed lower closure is lower closed once the repeated lower closure operation fixes it.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_lowerClosed_from_fixedPoint (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) (hFixed : partialClosure₂ (partialClosure₁ (lowerClosureConcaveConvex K h)) = lowerClosureConcaveConvex K h) : IsLowerClosed (lowerClosureConcaveConvex K h)

Lean source

Helper for Text 34.0.1: an upper mixed closure is upper closed once the repeated upper closure operation fixes it.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_upperClosed_from_fixedPoint (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) (hFixed : partialClosure₁ (partialClosure₂ (upperClosureConcaveConvex K h)) = upperClosureConcaveConvex K h) : IsUpperClosed (upperClosureConcaveConvex K h)

Lean source

Helper for Text 34.0.1: the fixed-point forms package directly into the lower-closed and upper-closed conclusions for the two mixed closures.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_packaged_closedness_conclusions (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) : IsLowerClosed (lowerClosureConcaveConvex K h) ∧ IsUpperClosed (upperClosureConcaveConvex K h)

Lean source

Helper for Text 34.0.1: once the normalized existential witness is available, the full idempotence and lower/upper closedness conclusion follows by the local closure chain already established in this file.

open Set variable {m n : ℕ} lemma helperForText_34_0_1_closedConvexWitness_exists_forces_main_conclusion (K : SaddleFunction m n) (h : IsConcaveConvex K) (hNoBotK : HasNoBotValuesBifunction K) (hLowerNoTopBot : HasNoTopOrBotValuesBifunction (lowerClosureConcaveConvex K h)) (hWitness : ∃ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsRockafellarConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ ClosedConvexBifunction F ∧ lowerClosureConcaveConvex K h = convexBifunctionPairing F ∧ upperClosureConcaveConvex K h = helperForText_34_0_1_convexAdjointPairingKernel F) : partialClosure₂ (partialClosure₁ (partialClosure₂ (partialClosure₁ K))) = partialClosure₂ (partialClosure₁ K) ∧ partialClosure₁ (partialClosure₂ (partialClosure₁ (partialClosure₂ K))) = partialClosure₁ (partialClosure₂ K) ∧ IsLowerClosed (lowerClosureConcaveConvex K h) ∧ IsUpperClosed (upperClosureConcaveConvex K h)

§34 — Closures and Equivalence Classes — Part 10

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Predicates on coordinate spaces in this section are treated classically when needed in piecewise definitions.

open Set noncomputable local instance classicalDecidablePredPart10 {α : Type*} (p : α → Prop) : DecidablePred p

Lean source

Membership in the open unit interval for a vector in Fin 1 → ℝ.

open Set def InOpenUnitInterval (x : Fin 1 → ℝ) : Prop := 0 < x 0 ∧ x 0 < 1

Lean source

Membership in the closed unit interval for a vector in Fin 1 → ℝ.

open Set def InClosedUnitInterval (x : Fin 1 → ℝ) : Prop := 0 ≤ x 0 ∧ x 0 ≤ 1

Lean source

The one-dimensional power kernel (u, v) ↦ u^v, written on Fin 1 → ℝ coordinates.

open Set noncomputable def oneDimensionalPowerKernel (u v : Fin 1 → ℝ) : EReal := ((Real.rpow (u 0) (v 0) : ℝ) : EReal)

Lean source

The saddle-function K(u, v) = u^v on the open unit square, represented on all of ℝ × ℝ by the upper simple-extension convention for concave-convex saddle-functions.

open Set noncomputable def openUnitSquarePowerSaddle : SaddleFunction 1 1 := fun u v => if InOpenUnitInterval v then if InOpenUnitInterval u then oneDimensionalPowerKernel u v else ⊥ else ⊤

Lean source

The explicit upper-closure formula for the open-square power saddle-function.

open Set noncomputable def openUnitSquarePowerUpperClosureFormula : SaddleFunction 1 1 := fun u v => if u 0 = 0 ∧ v 0 = 0 then ((1 : ℝ) : EReal) else if InClosedUnitInterval v then if InClosedUnitInterval u then oneDimensionalPowerKernel u v else ⊥ else ⊤

Lean source

The explicit lower-closure formula for the open-square power saddle-function.

open Set noncomputable def openUnitSquarePowerLowerClosureFormula : SaddleFunction 1 1 := fun u v => if u 0 = 0 ∧ v 0 = 0 then ((0 : ℝ) : EReal) else if InClosedUnitInterval v then if InClosedUnitInterval u then oneDimensionalPowerKernel u v else ⊥ else ⊤

Lean source

Helper for Text 34.1.1: on Fin 1, the norm of a difference is the absolute value of the unique coordinate difference.

open Set lemma helperForText_34_1_1_norm_fin1_eq_abs (u w : Fin 1 → ℝ) : ‖w - u‖ = |w 0 - u 0|

Lean source

Helper for Text 34.1.1: if a point lies outside [0, 1], some ball around it misses the open unit interval (0, 1).

open Set lemma helperForText_34_1_1_exists_radius_outside_closedUnitInterval (u : Fin 1 → ℝ) (hu : ¬ InClosedUnitInterval u) : ∃ ε : {r : ℝ // 0 < r}, ∀ w : {w : Fin 1 → ℝ // ‖w - u‖ < ε.1}, ¬ InOpenUnitInterval w.1

Lean source

Helper for Text 34.1.1: if the second variable is already outside (0, 1), the first partial closure sees a constant ⊤ section.

open Set lemma helperForText_34_1_1_partialClosure1_eq_top_of_not_inOpenUnitInterval {u v : Fin 1 → ℝ} (hv : ¬ InOpenUnitInterval v) : partialClosure₁ openUnitSquarePowerSaddle u v = ⊤

Lean source

Helper for Text 34.1.1: if the first variable lies outside [0, 1] while the second variable stays in (0, 1), the first partial closure is already ⊥.

open Set lemma helperForText_34_1_1_partialClosure1_eq_bot_of_not_inClosedUnitInterval {u v : Fin 1 → ℝ} (hu : ¬ InClosedUnitInterval u) (hv : InOpenUnitInterval v) : partialClosure₁ openUnitSquarePowerSaddle u v = ⊥

Lean source

Helper for Text 34.1.1: if the second variable lies outside [0, 1], the second partial closure already attains the constant value ⊤.

open Set lemma helperForText_34_1_1_partialClosure2_eq_top_of_not_inClosedUnitInterval {u v : Fin 1 → ℝ} (hv : ¬ InClosedUnitInterval v) : partialClosure₂ openUnitSquarePowerSaddle u v = ⊤

Lean source

Helper for Text 34.1.1: every neighborhood of a point in [0, 1] meets the open interval (0, 1).

open Set lemma helperForText_34_1_1_ball_hits_openUnitInterval_of_closedUnitInterval (v : Fin 1 → ℝ) (hv : InClosedUnitInterval v) (ε : {r : ℝ // 0 < r}) : ∃ w : {w : Fin 1 → ℝ // ‖w - v‖ < ε.1}, InOpenUnitInterval w.1

Lean source

Helper for Text 34.1.1: once u is outside (0, 1), every neighborhood of a point v ∈ [0, 1] already sees the value ⊥ in the second partial closure.

open Set lemma helperForText_34_1_1_partialClosure2_eq_bot_of_not_inOpenUnitInterval {u v : Fin 1 → ℝ} (hu : ¬ InOpenUnitInterval u) (hv : InClosedUnitInterval v) : partialClosure₂ openUnitSquarePowerSaddle u v = ⊥

Lean source

Helper for Text 34.1.1: convex combinations of points in (0, 1) stay in (0, 1) on Fin 1.

open Set lemma helperForText_34_1_1_openUnitInterval_combo_mem {x y : Fin 1 → ℝ} {a b : ℝ} (hx : InOpenUnitInterval x) (hy : InOpenUnitInterval y) (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) : InOpenUnitInterval (a • x + b • y)

Lean source

Helper for Text 34.1.1: the off-domain second-variable section at u = 2 is not convex on all of ℝ, so the upper simple extension cannot be concave-convex on Set.univ × Set.univ.

open Set lemma helperForText_34_1_1_secondSection_not_convex_at_two : ¬ IsERealConvexOn (Set.univ : Set (Fin 1 → ℝ)) (fun v => openUnitSquarePowerSaddle (fun _ : Fin 1 => (2 : ℝ)) v)

Lean source

Helper for Text 34.1.1: the chosen upper simple extension is not concave-convex on all of ℝ × ℝ, because one off-domain second-variable section already fails convexity.

open Set lemma helperForText_34_1_1_openUnitSquarePowerSaddle_not_concaveConvex : ¬ IsConcaveConvex openUnitSquarePowerSaddle

Lean source

Helper for Text 34.1.1: there is no global concave-convex witness for the chosen upper simple extension, so the theorem header cannot be repaired locally by supplying an auxiliary proof of IsConcaveConvex openUnitSquarePowerSaddle.

open Set lemma helperForText_34_1_1_no_globalConcaveConvexWitness : ¬ ∃ hK : IsConcaveConvex openUnitSquarePowerSaddle, upperClosureConcaveConvex openUnitSquarePowerSaddle hK = openUnitSquarePowerUpperClosureFormula

Lean source

Helper for Text 34.1.1: the type of global concave-convex witnesses for the chosen upper simple extension is empty, because the off-domain section at u = 2 already violates convexity on Set.univ.

open Set lemma helperForText_34_1_1_globalConcaveConvexWitness_isEmpty : IsEmpty (IsConcaveConvex openUnitSquarePowerSaddle)

Lean source

Helper for Text 34.1.1: on [0, 1] × (0, 1), every first-variable neighborhood already contains an open-interval point where the upper simple extension takes a nonnegative real value, so the first partial closure is nonnegative.

open Set lemma helperForText_34_1_1_partialClosure1_nonneg_of_closedFirst_openSecond {u v : Fin 1 → ℝ} (hu : InClosedUnitInterval u) (hv : InOpenUnitInterval v) : ((0 : ℝ) : EReal) ≤ partialClosure₁ openUnitSquarePowerSaddle u v

Lean source

Helper for Text 34.1.1: on [0, 1] × (0, 1), the first partial closure already recovers the original power kernel.

open Set lemma helperForText_34_1_1_partialClosure1_eq_powerKernel_of_closedFirst_openSecond {u v : Fin 1 → ℝ} (hu : InClosedUnitInterval u) (hv : InOpenUnitInterval v) : partialClosure₁ openUnitSquarePowerSaddle u v = oneDimensionalPowerKernel u v

Lean source

Helper for Text 34.1.1: the first partial closure extends the u-variable from (0, 1) to [0, 1] while keeping the second variable open.

open Set lemma helperForText_34_1_1_partialClosure1_formula : partialClosure₁ openUnitSquarePowerSaddle = fun u v => if InOpenUnitInterval v then if InClosedUnitInterval u then oneDimensionalPowerKernel u v else (⊥ : EReal) else (⊤ : EReal)

Lean source

Helper for Text 34.1.1: when the first variable stays in (0, 1), every value entering the second partial closure is nonnegative, so the closure itself is also nonnegative.

open Set lemma helperForText_34_1_1_partialClosure2_nonneg_of_openFirst {u v : Fin 1 → ℝ} (hu : InOpenUnitInterval u) : ((0 : ℝ) : EReal) ≤ partialClosure₂ openUnitSquarePowerSaddle u v

Lean source

Helper for Text 34.1.1: on the closed-square branch with u ∈ (0,1), the second partial closure already recovers the original power kernel u^v.

open Set lemma helperForText_34_1_1_partialClosure2_eq_powerKernel_of_openFirst_closedSecond {u v : Fin 1 → ℝ} (hu : InOpenUnitInterval u) (hv : InClosedUnitInterval v) : partialClosure₂ openUnitSquarePowerSaddle u v = oneDimensionalPowerKernel u v

Lean source

Helper for Text 34.1.1: the second partial closure extends the v-variable from (0, 1) to [0, 1] while keeping the first variable open.

open Set lemma helperForText_34_1_1_partialClosure2_formula : partialClosure₂ openUnitSquarePowerSaddle = fun u v => if InClosedUnitInterval v then if InOpenUnitInterval u then oneDimensionalPowerKernel u v else (⊥ : EReal) else (⊤ : EReal)

Lean source

Helper for Text 34.1.1: when the second variable stays in (0,1), the first partial closure at u = 0 is the boundary value 0.

open Set lemma helperForText_34_1_1_partialClosure1_zero_eq_zero_of_openSecond {v : Fin 1 → ℝ} (hv : InOpenUnitInterval v) : partialClosure₁ openUnitSquarePowerSaddle 0 v = ((0 : ℝ) : EReal)

§34 — Closures and Equivalence Classes — Part 11

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Predicates on coordinate spaces in this section are treated classically when needed in piecewise definitions.

open Set noncomputable local instance classicalDecidablePredPart11 {α : Type*} (p : α → Prop) : DecidablePred p

Lean source

Text 34.1.1: for the saddle-function K(u, v) = u^v on the open unit square, K is concave-convex on (0,1) × (0,1); the two iterated mixed closures are given by the displayed piecewise formulas; and they differ at the origin.

open Set theorem section34_example_u_pow_v : IsConcaveConvexOn {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} openUnitSquarePowerSaddle ∧ partialClosure₁ (partialClosure₂ openUnitSquarePowerSaddle) = openUnitSquarePowerUpperClosureFormula ∧ partialClosure₂ (partialClosure₁ openUnitSquarePowerSaddle) = openUnitSquarePowerLowerClosureFormula ∧ partialClosure₁ (partialClosure₂ openUnitSquarePowerSaddle) 0 0 ≠ partialClosure₂ (partialClosure₁ openUnitSquarePowerSaddle) 0 0

Lean source

Helper for Text 34.1.1: the mixed lower closure takes the value 0 at the origin.

open Set lemma helperForText_34_1_1_mixedLowerClosure_origin_eq_zero : partialClosure₂ (partialClosure₁ openUnitSquarePowerSaddle) 0 0 = ((0 : ℝ) : EReal)

Lean source

Helper for Text 34.1.1: if v already lies outside [0, 1], then the mixed lower closure cl₂(cl₁ K) stays constantly ⊤.

open Set lemma helperForText_34_1_1_mixedLowerClosure_eq_top_of_not_inClosedUnitInterval {u v : Fin 1 → ℝ} (hv : ¬ InClosedUnitInterval v) : partialClosure₂ (partialClosure₁ openUnitSquarePowerSaddle) u v = ⊤

Lean source

Helper for Text 34.1.1: if u lies outside [0, 1] while v stays in [0, 1], then the mixed lower closure cl₂(cl₁ K) is already ⊥.

open Set lemma helperForText_34_1_1_mixedLowerClosure_eq_bot_of_not_inClosedUnitInterval {u v : Fin 1 → ℝ} (hu : ¬ InClosedUnitInterval u) (hv : InClosedUnitInterval v) : partialClosure₂ (partialClosure₁ openUnitSquarePowerSaddle) u v = ⊥

Lean source

Helper for Text 34.1.1: if the first coordinate is 0 and both variables lie in [0,1], then the mixed lower closure cl₂(cl₁ K) takes the boundary value 0.

open Set lemma helperForText_34_1_1_mixedLowerClosure_zeroFirst_eq_zero_of_closedSecond {u v : Fin 1 → ℝ} (hu0 : u 0 = 0) (_hu : InClosedUnitInterval u) (hv : InClosedUnitInterval v) : partialClosure₂ (partialClosure₁ openUnitSquarePowerSaddle) u v = ((0 : ℝ) : EReal)

Lean source

Helper for Text 34.1.1: on the closed square away from the zero base, the mixed lower closure cl₂(cl₁ K) already recovers the original power kernel u^v.

open Set lemma helperForText_34_1_1_mixedLowerClosure_eq_powerKernel_of_closedSquare_nonzeroFirst {u v : Fin 1 → ℝ} (hu : InClosedUnitInterval u) (hv : InClosedUnitInterval v) (hu0 : u 0 ≠ 0) : partialClosure₂ (partialClosure₁ openUnitSquarePowerSaddle) u v = oneDimensionalPowerKernel u v

Lean source

Helper for Text 34.1.1: when the second coordinate already lies outside [0,1], the imported lower formula takes its final ⊤ branch.

open Set lemma helperForText_34_1_1_importedLowerFormula_eq_top_of_not_inClosedUnitInterval_second {u v : Fin 1 → ℝ} (hv : ¬ InClosedUnitInterval v) : openUnitSquarePowerLowerClosureFormula u v = ⊤

Lean source

Helper for Text 34.1.1: evaluating the mixed lower closure and the imported lower formula at the off-off point (2, 2) now yields the same value ⊤.

open Set lemma helperForText_34_1_1_mixedLowerClosure_counterexample_values : partialClosure₂ (partialClosure₁ openUnitSquarePowerSaddle) (fun _ : Fin 1 => (2 : ℝ)) (fun _ : Fin 1 => (2 : ℝ)) = ⊤ ∧ openUnitSquarePowerLowerClosureFormula (fun _ : Fin 1 => (2 : ℝ)) (fun _ : Fin 1 => (2 : ℝ)) = ⊤

Lean source

Helper for Text 34.1.1: this is the corrected pointwise formula for cl₂(cl₁ K) that the part-3 branch computations actually prove, before comparing with the imported part-2 formula.

open Set noncomputable def helperForText_34_1_1_correctedMixedLowerClosureFormula : SaddleFunction 1 1 := fun u v => if u 0 = 0 ∧ v 0 = 0 then ((0 : ℝ) : EReal) else if InClosedUnitInterval v then if InClosedUnitInterval u then oneDimensionalPowerKernel u v else ⊥ else ⊤

Lean source

Helper for Text 34.1.1: the corrected local lower-closure formula returns ⊤ whenever the second coordinate already lies outside [0,1].

open Set lemma helperForText_34_1_1_correctedMixedLowerClosureFormula_eq_top_of_not_inClosedUnitInterval {u v : Fin 1 → ℝ} (hv : ¬ InClosedUnitInterval v) : helperForText_34_1_1_correctedMixedLowerClosureFormula u v = ⊤

Lean source

Helper for Text 34.1.1: the actual mixed lower closure agrees with the corrected local formula obtained by splitting first on v ∈ [0,1] and only then on u ∈ [0,1].

open Set lemma helperForText_34_1_1_mixedLowerClosure_eq_correctedFormula : partialClosure₂ (partialClosure₁ openUnitSquarePowerSaddle) = helperForText_34_1_1_correctedMixedLowerClosureFormula

Lean source

Helper for Text 34.1.1: the corrected local lower-closure formula agrees with the imported part-2 formula on the off-off branch.

open Set lemma helperForText_34_1_1_correctedMixedLowerClosureFormula_offOff_branch_agrees {u v : Fin 1 → ℝ} (hv : ¬ InClosedUnitInterval v) : helperForText_34_1_1_correctedMixedLowerClosureFormula u v = openUnitSquarePowerLowerClosureFormula u v

Lean source

Helper for Text 34.1.1: the corrected local formula and the imported part-2 formula now agree at the explicit off-off witness (2, 2).

open Set lemma helperForText_34_1_1_correctedImportedAgreement_counterexample_values : helperForText_34_1_1_correctedMixedLowerClosureFormula (fun _ : Fin 1 => (2 : ℝ)) (fun _ : Fin 1 => (2 : ℝ)) = ⊤ ∧ openUnitSquarePowerLowerClosureFormula (fun _ : Fin 1 => (2 : ℝ)) (fun _ : Fin 1 => (2 : ℝ)) = ⊤

Lean source

Helper for Text 34.1.1: the corrected local lower-closure formula now agrees definitionally with the imported part-2 formula.

open Set lemma helperForText_34_1_1_correctedMixedLowerClosureFormula_eq_importedLowerClosureFormula : helperForText_34_1_1_correctedMixedLowerClosureFormula = openUnitSquarePowerLowerClosureFormula

Lean source

Helper for Text 34.1.1: once the imported lower formula is repaired to match the corrected local branch computation, the displayed mixed lower-closure formula follows immediately.

open Set lemma helperForText_34_1_1_mixedLowerClosure_formula_of_correctedImportedAgreement (hAgreement : helperForText_34_1_1_correctedMixedLowerClosureFormula = openUnitSquarePowerLowerClosureFormula) : partialClosure₂ (partialClosure₁ openUnitSquarePowerSaddle) = openUnitSquarePowerLowerClosureFormula

Lean source

Helper for Text 34.1.1: the displayed mixed lower-closure formula is equivalent to the upstream agreement between the corrected local formula and the imported part-2 definition.

open Set lemma helperForText_34_1_1_mixedLowerClosure_formula_iff_correctedImportedAgreement : partialClosure₂ (partialClosure₁ openUnitSquarePowerSaddle) = openUnitSquarePowerLowerClosureFormula ↔ helperForText_34_1_1_correctedMixedLowerClosureFormula = openUnitSquarePowerLowerClosureFormula

Lean source

Helper for Text 34.1.1: the mixed lower closure cl₂(cl₁ K) is the displayed lower closure formula with value 0 at the origin and -∞ on the left off-domain branch.

open Set lemma helperForText_34_1_1_mixedLowerClosure_formula : partialClosure₂ (partialClosure₁ openUnitSquarePowerSaddle) = openUnitSquarePowerLowerClosureFormula

Lean source

Helper for Text 34.1.1: if v already lies outside [0, 1], then the mixed upper closure cl₁(cl₂ K) stays constantly ⊤.

open Set lemma helperForText_34_1_1_mixedUpperClosure_eq_top_of_not_inClosedUnitInterval {u v : Fin 1 → ℝ} (hv : ¬ InClosedUnitInterval v) : partialClosure₁ (partialClosure₂ openUnitSquarePowerSaddle) u v = ⊤

Lean source

Helper for Text 34.1.1: if u lies outside [0, 1] while v stays in [0, 1], then the mixed upper closure cl₁(cl₂ K) is already ⊥.

open Set lemma helperForText_34_1_1_mixedUpperClosure_eq_bot_of_not_inClosedUnitInterval {u v : Fin 1 → ℝ} (hu : ¬ InClosedUnitInterval u) (hv : InClosedUnitInterval v) : partialClosure₁ (partialClosure₂ openUnitSquarePowerSaddle) u v = ⊥

Lean source

Helper for Text 34.1.1: on the closed square, the mixed upper closure cl₁(cl₂ K) already recovers the original power kernel u^v.

open Set lemma helperForText_34_1_1_mixedUpperClosure_eq_powerKernel_of_closedSquare {u v : Fin 1 → ℝ} (hu : InClosedUnitInterval u) (hv : InClosedUnitInterval v) : partialClosure₁ (partialClosure₂ openUnitSquarePowerSaddle) u v = oneDimensionalPowerKernel u v

Lean source

Helper for Text 34.1.1: the mixed upper closure cl₁(cl₂ K) is the displayed upper closure formula with value 1 at the origin.

open Set lemma helperForText_34_1_1_mixedUpperClosure_formula : partialClosure₁ (partialClosure₂ openUnitSquarePowerSaddle) = openUnitSquarePowerUpperClosureFormula

Lean source

Helper for Text 34.1.1: specializing the mixed upper-closure formula at the origin gives the value 1.

open Set lemma helperForText_34_1_1_mixedUpperClosure_origin_eq_one : partialClosure₁ (partialClosure₂ openUnitSquarePowerSaddle) 0 0 = ((1 : ℝ) : EReal)

Lean source

Helper for Text 34.1.1: the two iterated mixed closures really differ at the origin, with upper value 1 and lower value 0.

open Set lemma helperForText_34_1_1_mixedClosures_differ_at_origin : partialClosure₁ (partialClosure₂ openUnitSquarePowerSaddle) 0 0 ≠ partialClosure₂ (partialClosure₁ openUnitSquarePowerSaddle) 0 0

Lean source

Helper for Text 34.1.1: on the open unit square, the upper simple extension reduces to the finite power kernel u^v.

open Set lemma helperForText_34_1_1_openSquare_value_of_open_membership {u v : Fin 1 → ℝ} (hu : InOpenUnitInterval u) (hv : InOpenUnitInterval v) : openUnitSquarePowerSaddle u v = oneDimensionalPowerKernel u v

Lean source

Helper for Text 34.1.1: for every fixed v ∈ (0,1), the section u ↦ u^v is concave on the open unit interval.

open Set lemma helperForText_34_1_1_powerKernel_concave_in_first_on_openUnitInterval {v : Fin 1 → ℝ} (hv : InOpenUnitInterval v) : IsERealConcaveOn {u : Fin 1 → ℝ | InOpenUnitInterval u} (fun u => oneDimensionalPowerKernel u v)

Lean source

Helper for Text 34.1.1: with u ∈ (0,1) fixed, the power kernel rewrites as an exponential of an affine function in the second variable.

open Set lemma helperForText_34_1_1_powerKernel_eq_exp_log_on_openFirst {u v : Fin 1 → ℝ} (hu : InOpenUnitInterval u) : oneDimensionalPowerKernel u v = ((Real.exp ((Real.log (u 0)) * (v 0)) : ℝ) : EReal)

Lean source

Helper for Text 34.1.1: for every fixed u ∈ (0,1), the section v ↦ u^v is convex on the open unit interval.

open Set lemma helperForText_34_1_1_powerKernel_convex_in_second_on_openUnitInterval {u : Fin 1 → ℝ} (hu : InOpenUnitInterval u) : IsERealConvexOn {v : Fin 1 → ℝ | InOpenUnitInterval v} (fun v => oneDimensionalPowerKernel u v)

Lean source

Helper for Text 34.1.1: on the open unit square, the saddle kernel K(u,v)=u^v is concave in the first variable and convex in the second.

open Set lemma helperForText_34_1_1_openUnitSquarePowerSaddle_isConcaveConvexOn : IsConcaveConvexOn {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} openUnitSquarePowerSaddle

§34 — Closures and Equivalence Classes — Part 12

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Predicates on coordinate spaces in this section are treated classically when needed in piecewise definitions.

open Set noncomputable local instance classicalDecidablePredPart12 {α : Type*} (p : α → Prop) : DecidablePred p

Lean source

The one-dimensional division kernel (u, v) ↦ u / v, written on Fin 1 → ℝ coordinates.

open Set variable {m n : ℕ} noncomputable def oneDimensionalDivisionKernel (u v : Fin 1 → ℝ) : EReal := (((u 0) / (v 0) : ℝ) : EReal)

Lean source

The concave-convex simple extension of (u, v) ↦ u / v from the quadrant {(u, v) | 0 ≤ u ∧ 0 < v}.

open Set variable {m n : ℕ} noncomputable def positiveQuadrantDivisionSaddle : SaddleFunction 1 1 := fun u v => if v 0 ≤ 0 then ⊤ else if u 0 < 0 then ⊥ else oneDimensionalDivisionKernel u v

Lean source

The explicit upper-closure formula for the simple extension of (u, v) ↦ u / v on the positive quadrant.

open Set variable {m n : ℕ} noncomputable def positiveQuadrantDivisionUpperClosureFormula : SaddleFunction 1 1 := fun u v => if v 0 ≤ 0 then ⊤ else if u 0 < 0 then ⊥ else oneDimensionalDivisionKernel u v

Lean source

The explicit lower-closure formula for the simple extension of (u, v) ↦ u / v on the positive quadrant.

open Set variable {m n : ℕ} noncomputable def positiveQuadrantDivisionLowerClosureFormula : SaddleFunction 1 1 := fun u v => if u 0 < 0 then ⊥ else if u 0 = 0 ∧ v 0 = 0 then ((0 : ℝ) : EReal) else if v 0 ≤ 0 then ⊤ else oneDimensionalDivisionKernel u v

Lean source

The finiteness domain of the upper closure in the u / v model example is the positive quadrant boundary condition 0 ≤ u and 0 < v.

open Set variable {m n : ℕ} def positiveQuadrantDivisionUpperFinitenessDomain : Set ((Fin 1 → ℝ) × (Fin 1 → ℝ)) := {p | 0 ≤ p.1 0 ∧ 0 < p.2 0}

Lean source

The finiteness domain of the lower closure in the u / v model example adjoins the origin to the upper-closure finiteness domain.

open Set variable {m n : ℕ} def positiveQuadrantDivisionLowerFinitenessDomain : Set ((Fin 1 → ℝ) × (Fin 1 → ℝ)) := positiveQuadrantDivisionUpperFinitenessDomain ∪ {((0 : Fin 1 → ℝ), (0 : Fin 1 → ℝ))} -- Route correction: the displayed upper-closure formula is already incompatible with the -- implemented mixed closure at the concrete point `((-1), 0)`. The next lemmas isolate that -- mismatch directly from the closure definitions, so the remaining blocker is the textbook -- statement itself rather than a missing limit computation.

Lean source

Text 34.1.2 in the current formalization is not the displayed textbook package: the claimed lower-closure formula is refuted by the explicit negative-negative mismatch point ((-1), (-1)). The isolated textbook claim is therefore false as stated.

open Set variable {m n : ℕ} theorem section34_example_u_div_v : ¬ helperForText_34_1_2_targetTheoremClaim

Lean source

Helper for Text 34.1.2: the standard positive witness ε / 2 stays inside every ball around the second-variable origin.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_halfPositiveWitness_mem_ball_origin (ε : {ε : ℝ // 0 < ε}) : ‖(fun _ : Fin 1 => ε.1 / 2) - (0 : Fin 1 → ℝ)‖ < ε.1

Lean source

Helper for Text 34.1.2: if the first coordinate is negative, then the second partial closure at v = 0 is already -∞.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_secondClosureAtZero_eq_bot_of_negativeFirst {u : Fin 1 → ℝ} (hu : u 0 < 0) : partialClosure₂ positiveQuadrantDivisionSaddle u (0 : Fin 1 → ℝ) = (⊥ : EReal)

Lean source

Helper for Text 34.1.2: every point in the radius-1/2 ball around u = -1 still has negative first coordinate.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_halfBallAroundNegOne_stays_negative (w : {w : Fin 1 → ℝ // ‖w - (fun _ : Fin 1 => (-1 : ℝ))‖ < (1 / 2 : ℝ)}) : w.1 0 < 0

Lean source

Helper for Text 34.1.2: the inner second closure already takes the value -∞ at the negative-axis point ((-1), 0).

open Set variable {m n : ℕ} lemma helperForText_34_1_2_secondClosure_negAxis_zero_eq_bot : partialClosure₂ positiveQuadrantDivisionSaddle (fun _ : Fin 1 => (-1 : ℝ)) (0 : Fin 1 → ℝ) = (⊥ : EReal)

Lean source

Helper for Text 34.1.2: the mixed upper closure cl₁(cl₂ K) also takes the value -∞ at ((-1), 0).

open Set variable {m n : ℕ} lemma helperForText_34_1_2_upperMixedClosure_negAxis_zero_eq_bot : partialClosure₁ (partialClosure₂ positiveQuadrantDivisionSaddle) (fun _ : Fin 1 => (-1 : ℝ)) (0 : Fin 1 → ℝ) = (⊥ : EReal)

Lean source

Helper for Text 34.1.2: for every concave-convex witness, the actual upper closure takes the value -∞ at the negative-axis point ((-1), 0).

open Set variable {m n : ℕ} lemma helperForText_34_1_2_upperClosure_negAxis_zero_eq_bot (hK : IsConcaveConvex positiveQuadrantDivisionSaddle) : upperClosureConcaveConvex positiveQuadrantDivisionSaddle hK (fun _ : Fin 1 => (-1 : ℝ)) (0 : Fin 1 → ℝ) = (⊥ : EReal)

Lean source

Helper for Text 34.1.2: the displayed upper formula assigns the value +∞ at the same negative-axis point ((-1), 0).

open Set variable {m n : ℕ} lemma helperForText_34_1_2_upperFormula_negAxis_zero_eq_top : positiveQuadrantDivisionUpperClosureFormula (fun _ : Fin 1 => (-1 : ℝ)) (0 : Fin 1 → ℝ) = (⊤ : EReal)

Lean source

Helper for Text 34.1.2: every claimed upper-closure identity fails at ((-1), 0), because the actual mixed upper closure is -∞ there while the displayed formula is +∞.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_upperClosure_formula_fails_at_negAxis_zero (hK : IsConcaveConvex positiveQuadrantDivisionSaddle) : upperClosureConcaveConvex positiveQuadrantDivisionSaddle hK (fun _ : Fin 1 => (-1 : ℝ)) (0 : Fin 1 → ℝ) ≠ positiveQuadrantDivisionUpperClosureFormula (fun _ : Fin 1 => (-1 : ℝ)) (0 : Fin 1 → ℝ)

Lean source

Helper for Text 34.1.2: no concave-convex witness can make the displayed upper formula equal the true upper closure.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_upperClosure_ne_upperClosureFormula (hK : IsConcaveConvex positiveQuadrantDivisionSaddle) : upperClosureConcaveConvex positiveQuadrantDivisionSaddle hK ≠ positiveQuadrantDivisionUpperClosureFormula

Lean source

Helper for Text 34.1.2: any claimed upper-closure identity would force the impossible point value equation -∞ = +∞ at ((-1), 0).

open Set variable {m n : ℕ} lemma helperForText_34_1_2_upperClosure_identity_forces_bot_eq_top (hK : IsConcaveConvex positiveQuadrantDivisionSaddle) (hUpper : upperClosureConcaveConvex positiveQuadrantDivisionSaddle hK = positiveQuadrantDivisionUpperClosureFormula) : (⊥ : EReal) = (⊤ : EReal)

Lean source

Helper for Text 34.1.2: the first closure identity demanded by the theorem is already contradictory for any concave-convex witness.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_false_of_claimedUpperClosureIdentity (hK : IsConcaveConvex positiveQuadrantDivisionSaddle) (hUpper : upperClosureConcaveConvex positiveQuadrantDivisionSaddle hK = positiveQuadrantDivisionUpperClosureFormula) : False

Lean source

Helper for Text 34.1.2: every claimed upper-closure formula mismatch has the explicit counterexample point ((-1), 0).

open Set variable {m n : ℕ} lemma helperForText_34_1_2_exists_upperClosure_counterexample (hK : IsConcaveConvex positiveQuadrantDivisionSaddle) : ∃ u v, upperClosureConcaveConvex positiveQuadrantDivisionSaddle hK u v ≠ positiveQuadrantDivisionUpperClosureFormula u v

Lean source

Helper for Text 34.1.2: the full theorem conclusion is inconsistent, because its first closure identity already fails at ((-1), 0).

open Set variable {m n : ℕ} lemma helperForText_34_1_2_theorem_conclusion_false : ¬ (IsConcaveConvex positiveQuadrantDivisionSaddle ∧ ∀ hK : IsConcaveConvex positiveQuadrantDivisionSaddle, upperClosureConcaveConvex positiveQuadrantDivisionSaddle hK = positiveQuadrantDivisionUpperClosureFormula ∧ lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK = positiveQuadrantDivisionLowerClosureFormula ∧ positiveQuadrantDivisionUpperClosureFormula (0 : Fin 1 → ℝ) (0 : Fin 1 → ℝ) = (⊤ : EReal) ∧ positiveQuadrantDivisionLowerClosureFormula (0 : Fin 1 → ℝ) (0 : Fin 1 → ℝ) = ((0 : ℝ) : EReal) ∧ finitenessDomain (upperClosureConcaveConvex positiveQuadrantDivisionSaddle hK) = positiveQuadrantDivisionUpperFinitenessDomain ∧ finitenessDomain (lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK) = positiveQuadrantDivisionLowerFinitenessDomain ∧ finitenessDomain (lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK) ≠ finitenessDomain (upperClosureConcaveConvex positiveQuadrantDivisionSaddle hK) ∧ ¬ ∃ A : Set (Fin 1 → ℝ), ∃ B : Set (Fin 1 → ℝ), finitenessDomain (lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK) = A ×ˢ B)

Lean source

Helper for Text 34.1.2: if the second coordinate is nonpositive, then the first partial closure is already the constant value +∞.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_firstClosure_eq_top_of_nonpositiveSecond {u v : Fin 1 → ℝ} (hv : v 0 ≤ 0) : partialClosure₁ positiveQuadrantDivisionSaddle u v = (⊤ : EReal)

Lean source

Helper for Text 34.1.2: the mixed lower closure cl₂(cl₁ K) takes the value +∞ at the negative-negative point ((-1), (-1)).

open Set variable {m n : ℕ} lemma helperForText_34_1_2_mixedLowerClosure_negNeg_eq_top : partialClosure₂ (partialClosure₁ positiveQuadrantDivisionSaddle) (fun _ : Fin 1 => (-1 : ℝ)) (fun _ : Fin 1 => (-1 : ℝ)) = (⊤ : EReal)

Lean source

Helper for Text 34.1.2: every concave-convex witness forces the true lower closure to take the value +∞ at ((-1), (-1)).

open Set variable {m n : ℕ} lemma helperForText_34_1_2_lowerClosure_negNeg_eq_top (hK : IsConcaveConvex positiveQuadrantDivisionSaddle) : lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK (fun _ : Fin 1 => (-1 : ℝ)) (fun _ : Fin 1 => (-1 : ℝ)) = (⊤ : EReal)

Lean source

Helper for Text 34.1.2: the displayed lower formula assigns the value -∞ at the same negative-negative point ((-1), (-1)).

open Set variable {m n : ℕ} lemma helperForText_34_1_2_lowerFormula_negNeg_eq_bot : positiveQuadrantDivisionLowerClosureFormula (fun _ : Fin 1 => (-1 : ℝ)) (fun _ : Fin 1 => (-1 : ℝ)) = (⊥ : EReal)

Lean source

Helper for Text 34.1.2: the lower-closure mismatch is witnessed by the concrete negative-negative point ((-1), (-1)), where the true closure is +∞ and the displayed formula is -∞.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_exists_lowerClosure_counterexample (hK : IsConcaveConvex positiveQuadrantDivisionSaddle) : ∃ u v, lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK u v = (⊤ : EReal) ∧ positiveQuadrantDivisionLowerClosureFormula u v = (⊥ : EReal)

Lean source

Helper for Text 34.1.2: every claimed lower-closure identity fails at ((-1), (-1)), because the actual mixed lower closure is +∞ there while the displayed formula is -∞.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_lowerClosure_formula_fails_at_negNeg (hK : IsConcaveConvex positiveQuadrantDivisionSaddle) : lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK (fun _ : Fin 1 => (-1 : ℝ)) (fun _ : Fin 1 => (-1 : ℝ)) ≠ positiveQuadrantDivisionLowerClosureFormula (fun _ : Fin 1 => (-1 : ℝ)) (fun _ : Fin 1 => (-1 : ℝ))

Lean source

Helper for Text 34.1.2: no concave-convex witness can make the displayed lower formula equal the true lower closure.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_lowerClosure_ne_lowerClosureFormula (hK : IsConcaveConvex positiveQuadrantDivisionSaddle) : lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK ≠ positiveQuadrantDivisionLowerClosureFormula

Lean source

Helper for Text 34.1.2: no concave-convex witness can make the displayed lower formula equal the true lower closure.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_no_lowerClosureFormula_witness : ¬ ∃ hK : IsConcaveConvex positiveQuadrantDivisionSaddle, lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK = positiveQuadrantDivisionLowerClosureFormula

Lean source

Helper for Text 34.1.2: any claimed lower-closure identity would force the impossible point value equation +∞ = -∞ at ((-1), (-1)).

open Set variable {m n : ℕ} lemma helperForText_34_1_2_lowerClosure_identity_forces_top_eq_bot (hK : IsConcaveConvex positiveQuadrantDivisionSaddle) (hLower : lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK = positiveQuadrantDivisionLowerClosureFormula) : (⊤ : EReal) = (⊥ : EReal)

Lean source

Helper for Text 34.1.2: the remaining lower-closure identity demanded by the theorem is already contradictory for any concave-convex witness.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_false_of_claimedLowerClosureIdentity (hK : IsConcaveConvex positiveQuadrantDivisionSaddle) (hLower : lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK = positiveQuadrantDivisionLowerClosureFormula) : False

Lean source

Helper for Text 34.1.2: any proof of the current target theorem would force the impossible point-value equation +∞ = -∞ coming from the negative-negative counterexample.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_targetStatement_forces_top_eq_bot (hStatement : IsConcaveConvex positiveQuadrantDivisionSaddle ∧ ∀ hK : IsConcaveConvex positiveQuadrantDivisionSaddle, lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK = positiveQuadrantDivisionLowerClosureFormula ∧ upperClosureConcaveConvex positiveQuadrantDivisionSaddle hK (0 : Fin 1 → ℝ) (0 : Fin 1 → ℝ) = (⊤ : EReal) ∧ lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK (0 : Fin 1 → ℝ) (0 : Fin 1 → ℝ) = ((0 : ℝ) : EReal) ∧ finitenessDomain (lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK) = positiveQuadrantDivisionLowerFinitenessDomain ∧ finitenessDomain (lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK) ≠ finitenessDomain (upperClosureConcaveConvex positiveQuadrantDivisionSaddle hK) ∧ ¬ ∃ A : Set (Fin 1 → ℝ), ∃ B : Set (Fin 1 → ℝ), finitenessDomain (lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK) = A ×ˢ B) : (⊤ : EReal) = (⊥ : EReal)

Lean source

Helper for Text 34.1.2: once a concave-convex witness hK is fixed, the entire theorem-side conjunction is already impossible because its first clause is the false global lower-closure identity.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_no_revisedConclusion_for_fixedWitness (hK : IsConcaveConvex positiveQuadrantDivisionSaddle) : ¬ (lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK = positiveQuadrantDivisionLowerClosureFormula ∧ upperClosureConcaveConvex positiveQuadrantDivisionSaddle hK (0 : Fin 1 → ℝ) (0 : Fin 1 → ℝ) = (⊤ : EReal) ∧ lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK (0 : Fin 1 → ℝ) (0 : Fin 1 → ℝ) = ((0 : ℝ) : EReal) ∧ finitenessDomain (lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK) = positiveQuadrantDivisionLowerFinitenessDomain ∧ finitenessDomain (lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK) ≠ finitenessDomain (upperClosureConcaveConvex positiveQuadrantDivisionSaddle hK) ∧ ¬ ∃ A : Set (Fin 1 → ℝ), ∃ B : Set (Fin 1 → ℝ), finitenessDomain (lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK) = A ×ˢ B)

Lean source

Helper for Text 34.1.2: the weakened theorem statement still fails, because its lower closure identity already disagrees with the displayed formula at ((-1), (-1)).

open Set variable {m n : ℕ} lemma helperForText_34_1_2_revisedTheorem_conclusion_false : ¬ (IsConcaveConvex positiveQuadrantDivisionSaddle ∧ ∀ hK : IsConcaveConvex positiveQuadrantDivisionSaddle, lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK = positiveQuadrantDivisionLowerClosureFormula ∧ upperClosureConcaveConvex positiveQuadrantDivisionSaddle hK (0 : Fin 1 → ℝ) (0 : Fin 1 → ℝ) = (⊤ : EReal) ∧ lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK (0 : Fin 1 → ℝ) (0 : Fin 1 → ℝ) = ((0 : ℝ) : EReal) ∧ finitenessDomain (lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK) = positiveQuadrantDivisionLowerFinitenessDomain ∧ finitenessDomain (lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK) ≠ finitenessDomain (upperClosureConcaveConvex positiveQuadrantDivisionSaddle hK) ∧ ¬ ∃ A : Set (Fin 1 → ℝ), ∃ B : Set (Fin 1 → ℝ), finitenessDomain (lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK) = A ×ˢ B)

Lean source

Helper for Text 34.1.2: at the origin, the inner second closure already has value 0 because every nearby positive second-variable witness contributes 0, while the nonpositive branch contributes only +∞.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_secondClosure_origin_eq_zero : partialClosure₂ positiveQuadrantDivisionSaddle (0 : Fin 1 → ℝ) (0 : Fin 1 → ℝ) = ((0 : ℝ) : EReal)

Lean source

Helper for Text 34.1.2: once the first coordinate is positive, the second closure at v = 0 becomes +∞ because shrinking the second-variable radius forces arbitrarily large quotients u / w.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_secondClosureAtZero_eq_top_of_positiveFirst {u : Fin 1 → ℝ} (hu : 0 < u 0) : partialClosure₂ positiveQuadrantDivisionSaddle u (0 : Fin 1 → ℝ) = (⊤ : EReal)

Lean source

Helper for Text 34.1.2: the mixed upper closure takes the value +∞ at the origin.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_upperMixedClosure_origin_eq_top : partialClosure₁ (partialClosure₂ positiveQuadrantDivisionSaddle) (0 : Fin 1 → ℝ) (0 : Fin 1 → ℝ) = (⊤ : EReal)

Lean source

Helper for Text 34.1.2: for every concave-convex witness, the actual upper closure takes the value +∞ at the origin.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_upperClosure_origin_eq_top (hK : IsConcaveConvex positiveQuadrantDivisionSaddle) : upperClosureConcaveConvex positiveQuadrantDivisionSaddle hK (0 : Fin 1 → ℝ) (0 : Fin 1 → ℝ) = (⊤ : EReal)

Lean source

Helper for Text 34.1.2: when the second coordinate is positive, the first closure at u = 0 is exactly 0, since shrinking the first-variable radius drives every local supremum down to the quotient value 0 / v = 0.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_firstClosureAtZero_eq_zero_of_positiveSecond {v : Fin 1 → ℝ} (hv : 0 < v 0) : partialClosure₁ positiveQuadrantDivisionSaddle (0 : Fin 1 → ℝ) v = ((0 : ℝ) : EReal)

Lean source

Helper for Text 34.1.2: the mixed lower closure takes the value 0 at the origin.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_mixedLowerClosure_origin_eq_zero : partialClosure₂ (partialClosure₁ positiveQuadrantDivisionSaddle) (0 : Fin 1 → ℝ) (0 : Fin 1 → ℝ) = ((0 : ℝ) : EReal)

Lean source

Helper for Text 34.1.2: for every concave-convex witness, the actual lower closure takes the value 0 at the origin.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_lowerClosure_origin_eq_zero (hK : IsConcaveConvex positiveQuadrantDivisionSaddle) : lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hK (0 : Fin 1 → ℝ) (0 : Fin 1 → ℝ) = ((0 : ℝ) : EReal)

Lean source

Helper for Text 34.1.2: the displayed lower and upper finiteness domains already differ at the origin, because the lower domain adjoins (0, 0) while the upper one does not.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_displayedFinitenessDomains_ne : positiveQuadrantDivisionLowerFinitenessDomain ≠ positiveQuadrantDivisionUpperFinitenessDomain

Lean source

Helper for Text 34.1.2: the displayed lower finiteness domain is not a product set, because it contains (0, 0) and (1, 1) but omits (1, 0).

open Set variable {m n : ℕ} lemma helperForText_34_1_2_displayedLowerFinitenessDomain_not_product : ¬ ∃ A : Set (Fin 1 → ℝ), ∃ B : Set (Fin 1 → ℝ), positiveQuadrantDivisionLowerFinitenessDomain = A ×ˢ B

Lean source

Helper for Text 34.1.2: the exact conjunction currently asserted by section34_example_u_div_v, isolated as a single proposition so the remaining blocker can be stated without repeating the full target.

open Set variable {m n : ℕ} def helperForText_34_1_2_targetTheoremClaim : Prop

Lean source

Helper for Text 34.1.2: any witness of the exact target theorem claim already yields the explicit negative-negative counterexample where the true lower closure and the displayed formula disagree.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_targetClaim_has_explicitLowerMismatch (hClaim : helperForText_34_1_2_targetTheoremClaim) : ∃ u v, lowerClosureConcaveConvex positiveQuadrantDivisionSaddle hClaim.1 u v ≠ positiveQuadrantDivisionLowerClosureFormula u v

Lean source

Helper for Text 34.1.2: any witness of the isolated theorem claim is already contradictory, because the claim's global lower-closure identity must agree at the explicit mismatch point ((-1), (-1)).

open Set variable {m n : ℕ} lemma helperForText_34_1_2_targetTheoremClaim_witness_false (hClaim : helperForText_34_1_2_targetTheoremClaim) : False

Lean source

Helper for Text 34.1.2: any witness of the exact target theorem claim would force the displayed lower formula itself to take the value +∞ at ((-1), (-1)).

open Set variable {m n : ℕ} lemma helperForText_34_1_2_targetClaim_forces_lowerFormula_negNeg_eq_top (hClaim : helperForText_34_1_2_targetTheoremClaim) : positiveQuadrantDivisionLowerClosureFormula (fun _ : Fin 1 => (-1 : ℝ)) (fun _ : Fin 1 => (-1 : ℝ)) = (⊤ : EReal)

Lean source

Helper for Text 34.1.2: the exact theorem claim is refuted by the explicit negative-negative mismatch point ((-1), (-1)).

open Set variable {m n : ℕ} lemma helperForText_34_1_2_targetTheoremClaim_false : ¬ helperForText_34_1_2_targetTheoremClaim

Lean source

Helper for Text 34.1.2: the isolated target theorem claim is logically equivalent to False, so any future repair must change the statement or the upstream closure formalization rather than add more local proof steps.

open Set variable {m n : ℕ} lemma helperForText_34_1_2_targetTheoremClaim_iff_false : helperForText_34_1_2_targetTheoremClaim ↔ False

Lean source

Helper for Text 34.1.2: the isolated target theorem claim is an empty type, because any inhabitant would force the contradictory value equation ⊥ = ⊤ at ((-1), (-1)).

open Set variable {m n : ℕ} lemma helperForText_34_1_2_targetTheoremClaim_isEmpty : IsEmpty helperForText_34_1_2_targetTheoremClaim

§34 — Closures and Equivalence Classes — Part 13

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

The sign-product saddle-function on ℝ × ℝ, represented on Fin 1 → ℝ coordinates. It takes the value +∞ when uv > 0, 0 when uv = 0, and -∞ when uv < 0.

open Set variable {m n : ℕ} noncomputable def coordinateProductSignSaddle : SaddleFunction 1 1 := fun u v => if 0 < u 0 * v 0 then (⊤ : EReal) else if u 0 * v 0 = 0 then ((0 : ℝ) : EReal) else (⊥ : EReal)

Lean source

The explicit upper-closure formula for the sign-product saddle-function.

open Set variable {m n : ℕ} noncomputable def coordinateProductUpperClosureFormula : SaddleFunction 1 1 := fun u _ => if u 0 = 0 then ((0 : ℝ) : EReal) else (⊥ : EReal)

Lean source

The explicit lower-closure formula for the sign-product saddle-function.

open Set variable {m n : ℕ} noncomputable def coordinateProductLowerClosureFormula : SaddleFunction 1 1 := fun _ v => if v 0 = 0 then ((0 : ℝ) : EReal) else (⊤ : EReal)

Lean source

The locus where the sign-product saddle-function takes finite values is the union of the two coordinate axes.

open Set variable {m n : ℕ} def coordinateAxesFinitenessDomain : Set ((Fin 1 → ℝ) × (Fin 1 → ℝ)) := {p | p.1 0 = 0 ∨ p.2 0 = 0} -- Route correction: the displayed upper-closure formula is already incompatible with the -- implemented mixed closure at the positive point `((1), (1))`. The next helpers isolate that -- mismatch directly from the closure definitions, so the remaining blocker is the textbook -- statement itself rather than a missing local closure computation.

Lean source

Text 34.1.3 in the current formalization is not the displayed textbook package: the claimed upper-closure formula is refuted by the explicit positive-positive mismatch point ((1), (1)). The isolated textbook claim is therefore false as stated.

open Set variable {m n : ℕ} theorem section34_example_coordinate_product_sign : ¬ helperForText_34_1_3_targetTheoremClaim

Lean source

The book's “closed convex bifunction” terminology is implemented in Section 33 by IsImageClosedConvexBifunction.

open Set variable {m n : ℕ} abbrev IsClosedConvexBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := IsImageClosedConvexBifunction F

Lean source

The adjoint bifunction attached to a convex bifunction F, obtained by conjugating the kernel u ↦ ⟪F u, v⟫ in the first variable for each fixed v.

open Set variable {m n : ℕ} noncomputable def section34ConvexBifunctionAdjoint (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin n → ℝ) → (Fin m → ℝ) → EReal := fun v u => convexConjugate (fun w => convexBifunctionPairing F w v) u

Lean source

The kernel denoted ⟨u, F* v⟩ in Text 34.1.5, written using the explicit adjoint bifunction associated to F.

open Set variable {m n : ℕ} noncomputable def section34ConvexBifunctionAdjointPairing (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : SaddleFunction m n := fun u v => convexBifunctionPairing (section34ConvexBifunctionAdjoint F) v u

Lean source

The concave adjoint bifunction attached to a convex bifunction F, obtained by taking the concave conjugate in the parameter variable of the canonical kernel (u, xStar) ↦ ⟨F u, xStar⟩ for each fixed xStar. This is the F* appearing in the inf_u recovery formula of Theorem 34.2.

open Set variable {m n : ℕ} noncomputable def section34ConcaveBifunctionAdjoint (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin n → ℝ) → (Fin m → ℝ) → EReal := fun xStar uStar => concaveConjugate (fun u => convexBifunctionPairing F u xStar) uStar

Lean source

The kernel ⟨u, F* xStar⟩ built from the concave adjoint bifunction of Theorem 34.2, written with the concave pairing in the u-variable.

open Set variable {m n : ℕ} noncomputable def section34ConcaveBifunctionAdjointPairing (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : SaddleFunction m n := fun u xStar => concaveBifunctionPairing (section34ConcaveBifunctionAdjoint F) xStar u

Lean source

The canonical concave-convex kernel attached to a convex bifunction F.

open Set variable {m n : ℕ} noncomputable abbrev convexBifunctionClosedKernel (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : SaddleFunction m n := convexBifunctionPairing F

Lean source

The class Ω(F) generated by the canonical saddle kernel of F. Theorem 34.2 uses Ω(F) for the whole saddle-equivalence class, not for all pointwise minorants of the canonical kernel. Recording the generated class directly also keeps the definition independent of a proof that F is closed.

open Set variable {m n : ℕ} def omegaClassOfConvexBifunction (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Set (SaddleFunction m n) := {K | saddleEquivalent K (convexBifunctionClosedKernel F)}

Lean source

The domain dom F of a convex bifunction, read from the first effective domain of its canonical saddle-kernel.

open Set variable {m n : ℕ} def convexBifunctionDomain (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Set (Fin m → ℝ) := effectiveDomain₁ (convexBifunctionClosedKernel F)

Lean source

The effective second-coordinate domain of the canonical saddle-kernel attached to a convex bifunction F. This is the dom F* object that appears in the Chapter 34 saddle-kernel realization.

open Set variable {m n : ℕ} def convexBifunctionKernelAdjointDomain (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Set (Fin n → ℝ) := effectiveDomain₂ (convexBifunctionClosedKernel F)

Lean source

The effective domain of a convex bifunction consists of those parameters u for which some value F u x is finite in the convex convention.

open Set variable {m n : ℕ} def convexBifunctionEffectiveDomain (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Set (Fin m → ℝ) := {u | ∃ x, F u x < (⊤ : EReal)}

Lean source

The effective domain of a concave bifunction consists of those parameters xStar for which some value G xStar uStar is finite in the concave convention.

open Set variable {m n : ℕ} def concaveBifunctionEffectiveDomain (G : (Fin n → ℝ) → (Fin m → ℝ) → EReal) : Set (Fin n → ℝ) := {xStar | ∃ uStar, (⊥ : EReal) < G xStar uStar}

Lean source

A convex bifunction is proper when it never takes the value ⊥ and its effective domain is nonempty.

open Set variable {m n : ℕ} def IsProperConvexBifunction (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := HasNoBotValuesBifunction F ∧ (convexBifunctionEffectiveDomain F).Nonempty

Lean source

A concave bifunction is proper when it never takes the value ⊤ and its effective domain is nonempty.

open Set variable {m n : ℕ} def IsProperConcaveBifunction (G : (Fin n → ℝ) → (Fin m → ℝ) → EReal) : Prop := HasNoTopValuesBifunction G ∧ (concaveBifunctionEffectiveDomain G).Nonempty

Lean source

A set of saddle-functions is a closed concave-convex equivalence class when it is the saddleEquivalent-class of some closed concave-convex saddle-function.

open Set variable {m n : ℕ} def IsClosedConcaveConvexEquivalenceClass (Omega : Set (SaddleFunction m n)) : Prop := ∃ K : SaddleFunction m n, IsConcaveConvex K ∧ IsClosedSaddleFunction K ∧ Omega = {L | saddleEquivalent L K}

Lean source

Text 34.1.5 is false in the unrestricted textbook form used earlier in the file: the singleton constant-⊤ saddle-kernel is already a concave-convex counterexample.

open Set variable {m n : ℕ} theorem section34_text_34_1_5_unrestricted_false : ¬ helperForText_34_1_5_unrestrictedRepresentationScheme

Lean source

Text 34.1.5 has an explicit zero-dimensional improper counterexample: the singleton constant-⊤ saddle-kernel is concave-convex but admits no closed-convex representation of the unrestricted textbook form.

open Set variable {m n : ℕ} theorem section34_text_34_1_5_has_zeroDimensional_improperCounterexample : ∃ (K : SaddleFunction 0 0) (hK : IsConcaveConvex K), ¬ IsProperSaddleFunction K ∧ ¬ ∃! F : (Fin 0 → ℝ) → (Fin 0 → ℝ) → EReal, IsClosedConvexBifunction F ∧ lowerClosureConcaveConvex K hK = convexBifunctionPairing F ∧ upperClosureConcaveConvex K hK = section34ConvexBifunctionAdjointPairing F

Lean source

Text 34.1.5, corrected formal version: if K is concave-convex, its mixed closures satisfy underline(K) ≤ overline(K), and underline(K) = cl₂ cl₁ K satisfies the no-⊤/no-⊥ hypothesis required by the corrected Section 33 interface, then every supplied closed-convex realization of the mixed pair determines a unique image-closed convex bifunction F.

open Set variable {m n : ℕ} theorem section34_text_34_1_5 (K : SaddleFunction m n) (h : IsConcaveConvex K) (hOrder : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h) (hLowerOrient : IsConcaveConvex (lowerClosureConcaveConvex K h)) (hUpperOrient : IsConcaveConvex (upperClosureConcaveConvex K h)) (hLowerClosed : IsConvexClosedInSecond (lowerClosureConcaveConvex K h)) (hUpperClosed : IsConcaveClosedInFirst (upperClosureConcaveConvex K h)) (hLowerNoTopBot : HasNoTopOrBotValuesBifunction (lowerClosureConcaveConvex K h)) (hRealization : ∃ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, ClosedConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ lowerClosureConcaveConvex K h = convexBifunctionPairing F ∧ upperClosureConcaveConvex K h = helperForText_34_0_1_convexAdjointPairingKernel F) (hPairingInjective : ∀ {F G : (Fin m → ℝ) → (Fin n → ℝ) → EReal}, IsClosedConvexBifunction F → IsClosedConvexBifunction G → (∀ u xStar, convexBifunctionPairing F u xStar = convexBifunctionPairing G u xStar) → F = G) : ∃! F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsClosedConvexBifunction F ∧ lowerClosureConcaveConvex K h = convexBifunctionPairing F ∧ upperClosureConcaveConvex K h = helperForText_34_0_1_convexAdjointPairingKernel F

Lean source

Text 34.1.5 in the admissible extended-real regime used by the corrected Section 33 correspondence: once underline(K) satisfies the no-⊤/no-⊥ hypothesis, the closed-convex representative exists and is unique. This is the intended API for downstream uses that should not retry the unrestricted textbook statement.

open Set variable {m n : ℕ} theorem section34_text_34_1_5_of_lowerClosure_noTopOrBot (K : SaddleFunction m n) (h : IsConcaveConvex K) (hOrder : lowerClosureConcaveConvex K h ≤ upperClosureConcaveConvex K h) (hLowerOrient : IsConcaveConvex (lowerClosureConcaveConvex K h)) (hUpperOrient : IsConcaveConvex (upperClosureConcaveConvex K h)) (hLowerClosed : IsConvexClosedInSecond (lowerClosureConcaveConvex K h)) (hUpperClosed : IsConcaveClosedInFirst (upperClosureConcaveConvex K h)) (hLowerNoTopBot : HasNoTopOrBotValuesBifunction (lowerClosureConcaveConvex K h)) (hRealization : ∃ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, ClosedConvexBifunction F ∧ HasNoBotValuesBifunction F ∧ lowerClosureConcaveConvex K h = convexBifunctionPairing F ∧ upperClosureConcaveConvex K h = helperForText_34_0_1_convexAdjointPairingKernel F) (hPairingInjective : ∀ {F G : (Fin m → ℝ) → (Fin n → ℝ) → EReal}, IsClosedConvexBifunction F → IsClosedConvexBifunction G → (∀ u xStar, convexBifunctionPairing F u xStar = convexBifunctionPairing G u xStar) → F = G) : ∃! F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsClosedConvexBifunction F ∧ lowerClosureConcaveConvex K h = convexBifunctionPairing F ∧ upperClosureConcaveConvex K h = helperForText_34_0_1_convexAdjointPairingKernel F

Lean source

The corrected 34.1.5 scheme is valid in the qualified regime: a closed-convex realization of the admissible mixed pair upgrades uniquely to the Chapter 34 image-closed package.

open Set variable {m n : ℕ} theorem section34_text_34_1_5_admissible_scheme_true : helperForText_34_1_5_admissibleRepresentationScheme

Lean source

Helper for Text 34.1.3: every point in the radius-1/2 ball around 1 still has positive unique coordinate.

open Set variable {m n : ℕ} lemma helperForText_34_1_3_halfBallAroundOne_stays_positive (w : {w : Fin 1 → ℝ // ‖w - (fun _ : Fin 1 => (1 : ℝ))‖ < (1 / 2 : ℝ)}) : 0 < w.1 0

Lean source

Helper for Text 34.1.3: if the first coordinate is positive, then the second partial closure at v = 1 is already +∞, because a radius-1/2 ball around 1 stays on the positive-product branch.

open Set variable {m n : ℕ} lemma helperForText_34_1_3_secondClosureAtOne_eq_top_of_positiveFirst {u : Fin 1 → ℝ} (hu : 0 < u 0) : partialClosure₂ coordinateProductSignSaddle u (fun _ : Fin 1 => (1 : ℝ)) = (⊤ : EReal)

Lean source

Helper for Text 34.1.3: the inner second closure already takes the value +∞ at the positive point ((1), (1)).

open Set variable {m n : ℕ} lemma helperForText_34_1_3_secondClosure_posPos_eq_top : partialClosure₂ coordinateProductSignSaddle (fun _ : Fin 1 => (1 : ℝ)) (fun _ : Fin 1 => (1 : ℝ)) = (⊤ : EReal)

Lean source

Helper for Text 34.1.3: the mixed upper closure also takes the value +∞ at ((1), (1)), because a radius-1/2 first-variable ball stays in the positive branch and keeps the inner second closure equal to +∞.

open Set variable {m n : ℕ} lemma helperForText_34_1_3_upperMixedClosure_posPos_eq_top : partialClosure₁ (partialClosure₂ coordinateProductSignSaddle) (fun _ : Fin 1 => (1 : ℝ)) (fun _ : Fin 1 => (1 : ℝ)) = (⊤ : EReal)

Lean source

Helper for Text 34.1.3: for every concave-convex witness, the actual upper closure takes the value +∞ at the positive point ((1), (1)).

open Set variable {m n : ℕ} lemma helperForText_34_1_3_upperClosure_posPos_eq_top (hK : IsConcaveConvex coordinateProductSignSaddle) : upperClosureConcaveConvex coordinateProductSignSaddle hK (fun _ : Fin 1 => (1 : ℝ)) (fun _ : Fin 1 => (1 : ℝ)) = (⊤ : EReal)

Lean source

Helper for Text 34.1.3: the displayed upper formula assigns the value -∞ at ((1), (1)), because the first coordinate is nonzero there.

open Set variable {m n : ℕ} lemma helperForText_34_1_3_upperFormula_posPos_eq_bot : coordinateProductUpperClosureFormula (fun _ : Fin 1 => (1 : ℝ)) (fun _ : Fin 1 => (1 : ℝ)) = (⊥ : EReal)

Lean source

Helper for Text 34.1.3: every claimed upper-closure identity fails at ((1), (1)), because the actual mixed upper closure is +∞ there while the displayed formula is -∞.

open Set variable {m n : ℕ} lemma helperForText_34_1_3_upperClosure_formula_fails_at_posPos (hK : IsConcaveConvex coordinateProductSignSaddle) : upperClosureConcaveConvex coordinateProductSignSaddle hK (fun _ : Fin 1 => (1 : ℝ)) (fun _ : Fin 1 => (1 : ℝ)) ≠ coordinateProductUpperClosureFormula (fun _ : Fin 1 => (1 : ℝ)) (fun _ : Fin 1 => (1 : ℝ))

Lean source

Helper for Text 34.1.3: the exact conjunction currently asserted by section34_example_coordinate_product_sign, isolated as a single proposition so the remaining blocker can be stated without repeating the full target.

open Set variable {m n : ℕ} def helperForText_34_1_3_targetTheoremClaim : Prop

Lean source

Helper for Text 34.1.3: any witness of the isolated theorem claim is already contradictory, because the claim's global upper-closure identity must agree at the explicit mismatch point ((1), (1)).

open Set variable {m n : ℕ} lemma helperForText_34_1_3_targetTheoremClaim_witness_false (hClaim : helperForText_34_1_3_targetTheoremClaim) : False

Lean source

Helper for Text 34.1.3: the exact theorem claim is refuted by the explicit positive-positive counterexample point ((1), (1)).

open Set variable {m n : ℕ} lemma helperForText_34_1_3_targetTheoremClaim_false : ¬ helperForText_34_1_3_targetTheoremClaim

Lean source

Helper for Text 34.1.5: on the singleton space, the constant-⊤ kernel is concave-convex.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_zeroDimensional_constTop_isConcaveConvex : IsConcaveConvex ((fun (_ : Fin 0 → ℝ) (_ : Fin 0 → ℝ) => (⊤ : EReal)) : SaddleFunction 0 0)

Lean source

Helper for Text 34.1.5: the mixed lower closure of the singleton constant-⊤ kernel is the kernel itself.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_zeroDimensional_constTop_lowerClosure_eq_self (h : IsConcaveConvex ((fun (_ : Fin 0 → ℝ) (_ : Fin 0 → ℝ) => (⊤ : EReal)) : SaddleFunction 0 0)) : lowerClosureConcaveConvex ((fun (_ : Fin 0 → ℝ) (_ : Fin 0 → ℝ) => (⊤ : EReal)) : SaddleFunction 0 0) h = ((fun (_ : Fin 0 → ℝ) (_ : Fin 0 → ℝ) => (⊤ : EReal)) : SaddleFunction 0 0)

Lean source

Helper for Text 34.1.5: the mixed lower closure of the singleton constant-⊤ kernel still takes the value ⊤ at the unique point.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_zeroDimensional_constTop_lowerClosure_atOrigin_eq_top : let K : SaddleFunction 0 0 := fun _ _ => (⊤ : EReal) let hK : IsConcaveConvex K := helperForText_34_1_5_zeroDimensional_constTop_isConcaveConvex lowerClosureConcaveConvex K hK 0 0 = (⊤ : EReal)

Lean source

Helper for Text 34.1.5: the singleton constant-⊤ kernel fails the exact HasNoTopOrBotValuesBifunction admissibility hypothesis required by the corrected theorem, because its mixed lower closure still takes the value ⊤ at the unique point.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_zeroDimensional_constTop_lowerClosure_notHasNoTopOrBot : let K : SaddleFunction 0 0 := fun _ _ => (⊤ : EReal) let hK : IsConcaveConvex K := helperForText_34_1_5_zeroDimensional_constTop_isConcaveConvex ¬ HasNoTopOrBotValuesBifunction (lowerClosureConcaveConvex K hK)

Lean source

Helper for Text 34.1.5: the corrected theorem's no-⊤/no-⊥ hypothesis on the mixed lower closure is genuinely necessary, because the singleton constant-⊤ kernel already violates it.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_exists_counterexample_to_missingLowerClosureAdmissibility : ∃ (m n : ℕ) (K : SaddleFunction m n) (h : IsConcaveConvex K), ¬ HasNoTopOrBotValuesBifunction (lowerClosureConcaveConvex K h)

Lean source

Helper for Text 34.1.5: the singleton constant-⊤ kernel has empty saddle effective domain, so any repaired version of the theorem must exclude it by a properness-type hypothesis.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_zeroDimensional_constTop_saddleEffectiveDomain_eq_empty : let K : SaddleFunction 0 0 := fun _ _ => (⊤ : EReal) saddleEffectiveDomain K = (∅ : Set ((Fin 0 → ℝ) × (Fin 0 → ℝ)))

Lean source

Helper for Text 34.1.5: the singleton constant-⊤ kernel is not proper. This is the admissibility failure behind the zero-dimensional counterexample to the unrestricted theorem.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_zeroDimensional_constTop_notProper : let K : SaddleFunction 0 0 := fun _ _ => (⊤ : EReal) ¬ IsProperSaddleFunction K

Lean source

Helper for Text 34.1.5: a closed convex bifunction witness on the singleton space cannot have primal pairing value ⊤ at the unique point.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_zeroDimensional_constTop_refutes_witness {F : (Fin 0 → ℝ) → (Fin 0 → ℝ) → EReal} (hF : IsClosedConvexBifunction F) (hPair : convexBifunctionPairing F 0 0 = (⊤ : EReal)) : False

Lean source

Helper for Text 34.1.5: any claimed lower-closure representation of the singleton constant-⊤ kernel forces the pairing kernel to take the value ⊤ at the unique point.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_zeroDimensional_constTop_pairingAtOrigin_eq_top_of_lowerWitness {F : (Fin 0 → ℝ) → (Fin 0 → ℝ) → EReal} (hLower : lowerClosureConcaveConvex ((fun (_ : Fin 0 → ℝ) (_ : Fin 0 → ℝ) => (⊤ : EReal)) : SaddleFunction 0 0) helperForText_34_1_5_zeroDimensional_constTop_isConcaveConvex = convexBifunctionPairing F) : convexBifunctionPairing F 0 0 = (⊤ : EReal)

Lean source

Helper for Text 34.1.5: the singleton constant-⊤ kernel admits no closed convex bifunction witness at all for the mixed-closure representation.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_zeroDimensional_constTop_hasNoWitness : let K : SaddleFunction 0 0 := fun _ _ => (⊤ : EReal) let hK : IsConcaveConvex K := helperForText_34_1_5_zeroDimensional_constTop_isConcaveConvex ¬ ∃ F : (Fin 0 → ℝ) → (Fin 0 → ℝ) → EReal, IsClosedConvexBifunction F ∧ lowerClosureConcaveConvex K hK = convexBifunctionPairing F ∧ upperClosureConcaveConvex K hK = section34ConvexBifunctionAdjointPairing F

Lean source

Helper for Text 34.1.5: the theorem's representation conclusion already fails for the singleton constant-⊤ kernel.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_zeroDimensional_constTop_negates_targetConclusion : let K : SaddleFunction 0 0 := fun _ _ => (⊤ : EReal) let hK : IsConcaveConvex K := helperForText_34_1_5_zeroDimensional_constTop_isConcaveConvex ¬ ∃! F : (Fin 0 → ℝ) → (Fin 0 → ℝ) → EReal, IsClosedConvexBifunction F ∧ lowerClosureConcaveConvex K hK = convexBifunctionPairing F ∧ upperClosureConcaveConvex K hK = section34ConvexBifunctionAdjointPairing F

Lean source

Helper for Text 34.1.5: the singleton constant-⊤ kernel is an explicit improper counterexample to the unrestricted representation theorem.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_zeroDimensional_constTop_isImproperCounterexample : ∃ (K : SaddleFunction 0 0) (hK : IsConcaveConvex K), ¬ IsProperSaddleFunction K ∧ ¬ ∃! F : (Fin 0 → ℝ) → (Fin 0 → ℝ) → EReal, IsClosedConvexBifunction F ∧ lowerClosureConcaveConvex K hK = convexBifunctionPairing F ∧ upperClosureConcaveConvex K hK = section34ConvexBifunctionAdjointPairing F

Lean source

Helper for Text 34.1.5: the unrestricted theorem statement already has a concrete zero-dimensional counterexample.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_exists_counterexample_to_targetStatement : ∃ (m n : ℕ) (K : SaddleFunction m n) (h : IsConcaveConvex K), ¬ ∃! F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsClosedConvexBifunction F ∧ lowerClosureConcaveConvex K h = convexBifunctionPairing F ∧ upperClosureConcaveConvex K h = section34ConvexBifunctionAdjointPairing F

Lean source

Helper for Text 34.1.5: one explicit concave-convex kernel without a closed-convex witness already refutes any unrestricted global representation scheme.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_false_of_counterexample {m n : ℕ} {K : SaddleFunction m n} (hK : IsConcaveConvex K) (hNoWitness : ¬ ∃! F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsClosedConvexBifunction F ∧ lowerClosureConcaveConvex K hK = convexBifunctionPairing F ∧ upperClosureConcaveConvex K hK = section34ConvexBifunctionAdjointPairing F) (hScheme : ∀ {m n : ℕ} (K : SaddleFunction m n) (h : IsConcaveConvex K), ∃! F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsClosedConvexBifunction F ∧ lowerClosureConcaveConvex K h = convexBifunctionPairing F ∧ upperClosureConcaveConvex K h = section34ConvexBifunctionAdjointPairing F) : False

Lean source

Helper for Text 34.1.5: the unrestricted representation scheme is inconsistent, because its zero-dimensional specialization contradicts the singleton constant-⊤ counterexample.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_false_of_unrestrictedRepresentationScheme (hScheme : ∀ {m n : ℕ} (K : SaddleFunction m n) (h : IsConcaveConvex K), ∃! F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsClosedConvexBifunction F ∧ lowerClosureConcaveConvex K h = convexBifunctionPairing F ∧ upperClosureConcaveConvex K h = section34ConvexBifunctionAdjointPairing F) : False

Lean source

Helper for Text 34.1.5: the exact unrestricted textbook representation scheme, isolated as a single proposition so its formal failure can be referenced directly.

open Set variable {m n : ℕ} def helperForText_34_1_5_unrestrictedRepresentationScheme : Prop := ∀ {m n : ℕ} (K : SaddleFunction m n) (h : IsConcaveConvex K), ∃! F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsClosedConvexBifunction F ∧ lowerClosureConcaveConvex K h = convexBifunctionPairing F ∧ upperClosureConcaveConvex K h = section34ConvexBifunctionAdjointPairing F

Lean source

Helper for Text 34.1.5: specializing the isolated unrestricted scheme to the singleton constant-⊤ kernel already forces the impossible zero-dimensional witness.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_unrestrictedRepresentationScheme_specializes_to_zeroDimensional_constTop (hScheme : helperForText_34_1_5_unrestrictedRepresentationScheme) : let K : SaddleFunction 0 0 := fun _ _ => (⊤ : EReal) let hK : IsConcaveConvex K := helperForText_34_1_5_zeroDimensional_constTop_isConcaveConvex ∃! F : (Fin 0 → ℝ) → (Fin 0 → ℝ) → EReal, IsClosedConvexBifunction F ∧ lowerClosureConcaveConvex K hK = convexBifunctionPairing F ∧ upperClosureConcaveConvex K hK = section34ConvexBifunctionAdjointPairing F

Lean source

Helper for Text 34.1.5: the corrected admissible representation scheme. Besides the mixed closure qualifications, it exposes existence of the closed-convex Section 33 realization; the remaining pairing-injectivity premise records exactly the recovery property needed for uniqueness in the image-closed Chapter 34 package.

open Set variable {m n : ℕ} def helperForText_34_1_5_admissibleRepresentationScheme : Prop

Lean source

Helper for Text 34.1.5: specializing the unrestricted textbook scheme to the singleton constant-⊤ kernel immediately contradicts the zero-dimensional no-witness theorem.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_zeroDimensional_constTop_contradicts_unrestrictedScheme (hScheme : helperForText_34_1_5_unrestrictedRepresentationScheme) : False

Lean source

Helper for Text 34.1.5: the unrestricted textbook scheme is false in this development, because the singleton constant-⊤ kernel contradicts it.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_unrestrictedRepresentationScheme_false : ¬ helperForText_34_1_5_unrestrictedRepresentationScheme

Lean source

Helper for Text 34.1.5: the singleton constant-⊤ kernel separates the false unrestricted scheme from the corrected admissible theorem, because the representation conclusion fails there and the lower-closure no-⊤/no-⊥ hypothesis fails there as well.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_zeroDimensional_constTop_separates_unrestricted_and_correctedSchemes : let K : SaddleFunction 0 0 := fun _ _ => (⊤ : EReal) let hK : IsConcaveConvex K := helperForText_34_1_5_zeroDimensional_constTop_isConcaveConvex ¬ HasNoTopOrBotValuesBifunction (lowerClosureConcaveConvex K hK) ∧ ¬ ∃! F : (Fin 0 → ℝ) → (Fin 0 → ℝ) → EReal, IsClosedConvexBifunction F ∧ lowerClosureConcaveConvex K hK = convexBifunctionPairing F ∧ upperClosureConcaveConvex K hK = section34ConvexBifunctionAdjointPairing F

Lean source

Helper for Text 34.1.5: a Section 33 closed convex witness automatically satisfies the image-closed package used by the Chapter 34 saddle-kernel correspondence.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_isClosedConvexBifunction_of_closedConvexWitness {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hClosed : ClosedConvexBifunction F) (hNoBot : HasNoBotValuesBifunction F) : IsClosedConvexBifunction F

Lean source

Helper for Text 34.1.5: an image-closed convex bifunction is uniquely determined by its pairing kernel.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_eq_of_pairing_eq {F G : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : IsClosedConvexBifunction F) (hG : IsClosedConvexBifunction G) (hPairEq : ∀ u xStar, convexBifunctionPairing F u xStar = convexBifunctionPairing G u xStar) : F = G

Lean source

Helper for Text 34.1.5: the singleton constant-⊤ kernel gives a concrete witness that the corrected admissible scheme cannot be upgraded to the unrestricted textbook statement.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_exists_counterexample_separating_corrected_and_unrestrictedSchemes : ∃ (m n : ℕ) (K : SaddleFunction m n) (h : IsConcaveConvex K), ¬ HasNoTopOrBotValuesBifunction (lowerClosureConcaveConvex K h) ∧ ¬ ∃! F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsClosedConvexBifunction F ∧ lowerClosureConcaveConvex K h = convexBifunctionPairing F ∧ upperClosureConcaveConvex K h = section34ConvexBifunctionAdjointPairing F

Lean source

Helper for Text 34.1.5: the corrected admissible scheme is genuinely sharp. The admissible scheme holds, but the singleton constant-⊤ kernel still violates the lower-closure no-⊤/no-⊥ hypothesis and therefore remains a concrete counterexample to the unrestricted textbook conclusion.

open Set variable {m n : ℕ} lemma helperForText_34_1_5_admissibleScheme_isSharp : helperForText_34_1_5_admissibleRepresentationScheme ∧ ∃ (m n : ℕ) (K : SaddleFunction m n) (h : IsConcaveConvex K), ¬ HasNoTopOrBotValuesBifunction (lowerClosureConcaveConvex K h) ∧ ¬ ∃! F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsClosedConvexBifunction F ∧ lowerClosureConcaveConvex K h = convexBifunctionPairing F ∧ upperClosureConcaveConvex K h = section34ConvexBifunctionAdjointPairing F

§34 — Closures and Equivalence Classes — Part 14

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Text 34.1.6: if K is concave-convex on ℝ^m × ℝ^n, then dom₁ K is convex in ℝ^m and dom₂ K is convex in ℝ^n; consequently dom K is convex in ℝ^m × ℝ^n, and K is finite at every point of dom K.

open Set variable {m n : ℕ} theorem section34_text_34_1_6 (K : SaddleFunction m n) (h : IsConcaveConvex K) : Convex ℝ (effectiveDomain₁ K) ∧ Convex ℝ (effectiveDomain₂ K) ∧ Convex ℝ (saddleEffectiveDomain K) ∧ saddleEffectiveDomain K ⊆ finitenessDomain K

Lean source

Helper for Text 34.1.6: a nonnegative weighted sum of two EReal values strictly above ⊥ is still strictly above ⊥.

open Set variable {m n : ℕ} lemma helperForText_34_1_6_weightedSum_gt_bot {x y : EReal} {a b : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b) (_hab : a + b = 1) (hx : (⊥ : EReal) < x) (hy : (⊥ : EReal) < y) : (⊥ : EReal) < (a : EReal) * x + (b : EReal) * y

Lean source

Helper for Text 34.1.6: a nonnegative weighted sum of two EReal values strictly below ⊤ is still strictly below ⊤.

open Set variable {m n : ℕ} lemma helperForText_34_1_6_weightedSum_lt_top {x y : EReal} {a b : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b) (_hab : a + b = 1) (hx : x < (⊤ : EReal)) (hy : y < (⊤ : EReal)) : (a : EReal) * x + (b : EReal) * y < (⊤ : EReal)

Lean source

Helper for Text 34.1.6: the effective domain of a globally concave EReal slice is convex.

open Set variable {m n : ℕ} lemma helperForText_34_1_6_concaveSlice_effectiveDomain_convex {k : ℕ} {f : (Fin k → ℝ) → EReal} (hConc : IsERealConcaveOn (Set.univ : Set (Fin k → ℝ)) f) : Convex ℝ (concaveFunctionEffectiveDomain f)

Lean source

Helper for Text 34.1.6: the first effective domain is the intersection of the concave slice domains dom (u ↦ K u v).

open Set variable {m n : ℕ} lemma helperForText_34_1_6_effectiveDomain1_eq_iInter_concaveSlices (K : SaddleFunction m n) : effectiveDomain₁ K = ⋂ v : Fin n → ℝ, concaveFunctionEffectiveDomain (fun u => K u v)

Lean source

Helper for Text 34.1.6: the second effective domain is the intersection of the convex slice domains dom (v ↦ K u v).

open Set variable {m n : ℕ} lemma helperForText_34_1_6_effectiveDomain2_eq_iInter_convexSlices (K : SaddleFunction m n) : effectiveDomain₂ K = ⋂ u : Fin m → ℝ, effectiveDomain (Set.univ : Set (Fin n → ℝ)) (K u)

Lean source

Helper for Text 34.1.6: fixing the second variable yields a convex first-variable slice domain.

open Set variable {m n : ℕ} lemma helperForText_34_1_6_convex_concaveSlice_for_fixedSecond (K : SaddleFunction m n) (h : IsConcaveConvex K) (v : Fin n → ℝ) : Convex ℝ (concaveFunctionEffectiveDomain (fun u => K u v))

Lean source

Helper for Text 34.1.6: fixing the first variable yields a convex second-variable slice domain.

open Set variable {m n : ℕ} lemma helperForText_34_1_6_convex_effectiveDomainSlice_for_fixedFirst (K : SaddleFunction m n) (h : IsConcaveConvex K) (u : Fin m → ℝ) : Convex ℝ (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (K u))

Lean source

Helper for Text 34.1.6: every point of dom₁ K × dom₂ K is a finite-valued point of K.

open Set variable {m n : ℕ} lemma helperForText_34_1_6_domainPoint_isFinite (K : SaddleFunction m n) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ effectiveDomain₁ K) (hv : v ∈ effectiveDomain₂ K) : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)

Lean source

Helper for Text 34.1.6: membership in the saddle effective domain is exactly coordinatewise membership in the two effective domains.

open Set variable {m n : ℕ} lemma helperForText_34_1_6_mem_saddleEffectiveDomain_iff (K : SaddleFunction m n) {p : (Fin m → ℝ) × (Fin n → ℝ)} : p ∈ saddleEffectiveDomain K ↔ p.1 ∈ effectiveDomain₁ K ∧ p.2 ∈ effectiveDomain₂ K

Lean source

Helper for Text 34.1.6: every point of dom K lies in the finiteness domain of K.

open Set variable {m n : ℕ} lemma helperForText_34_1_6_mem_finitenessDomain_of_mem_saddleEffectiveDomain (K : SaddleFunction m n) {p : (Fin m → ℝ) × (Fin n → ℝ)} (hp : p ∈ saddleEffectiveDomain K) : p ∈ finitenessDomain K

Lean source

Helper for Text 34.1.6: the saddle effective domain is convex once the two coordinate effective domains are convex.

open Set variable {m n : ℕ} lemma helperForText_34_1_6_convex_saddleEffectiveDomain (K : SaddleFunction m n) (hDom1 : Convex ℝ (effectiveDomain₁ K)) (hDom2 : Convex ℝ (effectiveDomain₂ K)) : Convex ℝ (saddleEffectiveDomain K)

Lean source

Helper for Text 34.1.6: the concavity side of a saddle-function gives Jensen's inequality on each fixed-second-variable slice.

open Set variable {m n : ℕ} lemma helperForText_34_1_6_concaveSlice_jensen (K : SaddleFunction m n) (h : IsConcaveConvex K) (v : Fin n → ℝ) {u₁ u₂ : Fin m → ℝ} {a b : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) : (a : EReal) * K u₁ v + (b : EReal) * K u₂ v ≤ K (a • u₁ + b • u₂) v

Lean source

Helper for Text 34.1.6: the convexity side of a saddle-function gives Jensen's inequality on each fixed-first-variable slice.

open Set variable {m n : ℕ} lemma helperForText_34_1_6_convexSlice_jensen (K : SaddleFunction m n) (h : IsConcaveConvex K) (u : Fin m → ℝ) {v₁ v₂ : Fin n → ℝ} {a b : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) : K u (a • v₁ + b • v₂) ≤ (a : EReal) * K u v₁ + (b : EReal) * K u v₂

Lean source

Helper for Text 34.1.6: convex combinations of points in dom₁ K stay in dom₁ K.

open Set variable {m n : ℕ} lemma helperForText_34_1_6_mem_effectiveDomain1_of_convexCombination (K : SaddleFunction m n) (h : IsConcaveConvex K) {u₁ u₂ : Fin m → ℝ} (hu₁ : u₁ ∈ effectiveDomain₁ K) (hu₂ : u₂ ∈ effectiveDomain₁ K) {a b : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) : a • u₁ + b • u₂ ∈ effectiveDomain₁ K

Lean source

Helper for Text 34.1.6: convex combinations of points in dom₂ K stay in dom₂ K.

open Set variable {m n : ℕ} lemma helperForText_34_1_6_mem_effectiveDomain2_of_convexCombination (K : SaddleFunction m n) (h : IsConcaveConvex K) {v₁ v₂ : Fin n → ℝ} (hv₁ : v₁ ∈ effectiveDomain₂ K) (hv₂ : v₂ ∈ effectiveDomain₂ K) {a b : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) : a • v₁ + b • v₂ ∈ effectiveDomain₂ K

Lean source

Helper for Text 34.1.6: convex combinations of points in dom K stay in dom K.

open Set variable {m n : ℕ} lemma helperForText_34_1_6_mem_saddleEffectiveDomain_of_convexCombination (K : SaddleFunction m n) (h : IsConcaveConvex K) {p₁ p₂ : (Fin m → ℝ) × (Fin n → ℝ)} (hp₁ : p₁ ∈ saddleEffectiveDomain K) (hp₂ : p₂ ∈ saddleEffectiveDomain K) {a b : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b) (hab : a + b = 1) : a • p₁ + b • p₂ ∈ saddleEffectiveDomain K

Lean source

Helper for Text 34.1.6: the first effective domain is convex for a concave-convex saddle function.

open Set variable {m n : ℕ} lemma helperForText_34_1_6_convex_effectiveDomain1 (K : SaddleFunction m n) (h : IsConcaveConvex K) : Convex ℝ (effectiveDomain₁ K)

Lean source

Helper for Text 34.1.6: the second effective domain is convex for a concave-convex saddle function.

open Set variable {m n : ℕ} lemma helperForText_34_1_6_convex_effectiveDomain2 (K : SaddleFunction m n) (h : IsConcaveConvex K) : Convex ℝ (effectiveDomain₂ K)

Lean source

Helper for Text 34.1.7: the v-section of the open-unit-square power upper simple extension at u = 1 / 2 fails concavity on all of ℝ, because two off-domain points have an in-domain midpoint.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_secondSection_not_concave_at_half : ¬ IsERealConcaveOn (Set.univ : Set (Fin 1 → ℝ)) (fun v => openUnitSquarePowerSaddle (fun _ : Fin 1 => (1 / 2 : ℝ)) v)

Lean source

Helper for Text 34.1.7: the open-unit-square power upper simple extension is not convex-concave on all of ℝ × ℝ.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_openUnitSquarePowerSaddle_not_convexConcave : ¬ IsConvexConcave openUnitSquarePowerSaddle

Lean source

Helper for Text 34.1.7: the open-unit-square power upper simple extension is not a global saddle-function, so it cannot satisfy the theorem's upper-extension conclusion.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_openUnitSquarePowerSaddle_not_saddle : ¬ IsSaddleFunctionOn (Set.univ : Set (Fin 1 → ℝ)) (Set.univ : Set (Fin 1 → ℝ)) openUnitSquarePowerSaddle

Lean source

Helper for Text 34.1.7: the open unit interval subset of Fin 1 → ℝ is convex.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_openUnitInterval_convex : Convex ℝ {u : Fin 1 → ℝ | InOpenUnitInterval u}

Lean source

Helper for Text 34.1.7: the open unit interval subset of Fin 1 → ℝ is nonempty.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_openUnitInterval_nonempty : ({u : Fin 1 → ℝ | InOpenUnitInterval u} : Set (Fin 1 → ℝ)).Nonempty

Lean source

Helper for Text 34.1.7: on the open unit square, the real power kernel carries the concave-convex saddle orientation used by the textbook example.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_oneDimensionalPowerKernel_isConcaveConvexOn : IsConcaveConvexOn {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel

Lean source

Helper for Text 34.1.7: the upper simple extension of the open-unit-square power kernel is definitionally the openUnitSquarePowerSaddle counterexample.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_upperSimpleExtension_eq_openUnitSquarePowerSaddle : upperSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel = openUnitSquarePowerSaddle

Lean source

Helper for Text 34.1.7: the exact universal theorem claim, isolated as a single proposition so the open-unit-square counterexample can specialize it.

open Set variable {m n : ℕ} def helperForText_34_1_7_targetTheoremClaim : Prop

Lean source

Helper for Text 34.1.7: the open-unit-square power example satisfies every hypothesis in the isolated theorem claim.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_openUnitPowerKernel_satisfiesTargetHypotheses : Convex ℝ {u : Fin 1 → ℝ | InOpenUnitInterval u} ∧ Convex ℝ {v : Fin 1 → ℝ | InOpenUnitInterval v} ∧ ({u : Fin 1 → ℝ | InOpenUnitInterval u} : Set (Fin 1 → ℝ)).Nonempty ∧ ({v : Fin 1 → ℝ | InOpenUnitInterval v} : Set (Fin 1 → ℝ)).Nonempty ∧ IsSaddleFunctionOn {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel ∧ (∀ u ∈ ({u : Fin 1 → ℝ | InOpenUnitInterval u} : Set (Fin 1 → ℝ)), ∀ v ∈ ({v : Fin 1 → ℝ | InOpenUnitInterval v} : Set (Fin 1 → ℝ)), (⊥ : EReal) < oneDimensionalPowerKernel u v ∧ oneDimensionalPowerKernel u v < (⊤ : EReal))

Lean source

Helper for Text 34.1.7: the specialized upper simple extension for the open-unit-square power kernel is already not a global saddle-function.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_specializedUpperSimpleExtension_not_saddle : ¬ IsSaddleFunctionOn (Set.univ : Set (Fin 1 → ℝ)) (Set.univ : Set (Fin 1 → ℝ)) (upperSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel)

Lean source

Helper for Text 34.1.7: the full specialized conclusion is false, because its upper simple extension branch is the already refuted open-unit-square counterexample.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_specializedConclusion_false : ¬ (IsSaddleFunctionOn (Set.univ : Set (Fin 1 → ℝ)) (Set.univ : Set (Fin 1 → ℝ)) (lowerSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel) ∧ effectiveDomain₁ (lowerSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel) = {u : Fin 1 → ℝ | InOpenUnitInterval u} ∧ effectiveDomain₂ (lowerSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel) = {v : Fin 1 → ℝ | InOpenUnitInterval v} ∧ saddleEffectiveDomain (lowerSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel) = {u : Fin 1 → ℝ | InOpenUnitInterval u} ×ˢ {v : Fin 1 → ℝ | InOpenUnitInterval v} ∧ IsProperSaddleFunction (lowerSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel) ∧ IsSaddleFunctionOn (Set.univ : Set (Fin 1 → ℝ)) (Set.univ : Set (Fin 1 → ℝ)) (upperSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel) ∧ effectiveDomain₁ (upperSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel) = {u : Fin 1 → ℝ | InOpenUnitInterval u} ∧ effectiveDomain₂ (upperSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel) = {v : Fin 1 → ℝ | InOpenUnitInterval v} ∧ saddleEffectiveDomain (upperSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel) = {u : Fin 1 → ℝ | InOpenUnitInterval u} ×ˢ {v : Fin 1 → ℝ | InOpenUnitInterval v} ∧ IsProperSaddleFunction (upperSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel))

§34 — Closures and Equivalence Classes — Part 15

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Text 34.1.7 in its unrestricted textbook form is false for the current formalization of simple extensions: the one-dimensional open-unit power kernel refutes the global upper-saddle branch.

open Set variable {m n : ℕ} theorem section34_text_34_1_7_unrestricted_false : ¬ helperForText_34_1_7_targetTheoremClaim

Lean source

Text 34.1.7 has a concrete one-dimensional counterexample in the present formalization: on the open unit square, the power kernel satisfies all local hypotheses, but its upper simple extension is not a global saddle-function on ℝ × ℝ.

open Set variable {m n : ℕ} theorem section34_text_34_1_7_has_openUnitPowerCounterexample : ∃ (C D : Set (Fin 1 → ℝ)) (J : SaddleFunction 1 1), Convex ℝ C ∧ Convex ℝ D ∧ C.Nonempty ∧ D.Nonempty ∧ IsSaddleFunctionOn C D J ∧ (∀ u ∈ C, ∀ v ∈ D, (⊥ : EReal) < J u v ∧ J u v < (⊤ : EReal)) ∧ ¬ IsSaddleFunctionOn (Set.univ : Set (Fin 1 → ℝ)) (Set.univ : Set (Fin 1 → ℝ)) (upperSimpleExtension C D J)

Lean source

Corrected Text 34.1.7 for the present formalization: if J is finite on a nonempty convex product set C × D, then both simple extensions still have the expected effective domains and are proper. The unrestricted global upper-saddle conclusion is not asserted.

open Set variable {m n : ℕ} theorem section34_text_34_1_7 {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (_hC : Convex ℝ C) (_hD : Convex ℝ D) (hCne : C.Nonempty) (hDne : D.Nonempty) (J : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (_hJ : IsSaddleFunctionOn C D J) (hJfinite : ∀ u ∈ C, ∀ v ∈ D, (⊥ : EReal) < J u v ∧ J u v < (⊤ : EReal)) : effectiveDomain₁ (lowerSimpleExtension C D J) = C ∧ effectiveDomain₂ (lowerSimpleExtension C D J) = D ∧ saddleEffectiveDomain (lowerSimpleExtension C D J) = C ×ˢ D ∧ IsProperSaddleFunction (lowerSimpleExtension C D J) ∧ effectiveDomain₁ (upperSimpleExtension C D J) = C ∧ effectiveDomain₂ (upperSimpleExtension C D J) = D ∧ saddleEffectiveDomain (upperSimpleExtension C D J) = C ×ˢ D ∧ IsProperSaddleFunction (upperSimpleExtension C D J)

Lean source

Two saddle-functions are equivalent when, in whichever saddle orientation they share, they have the same two coordinatewise partial closures.

open Set variable {m n : ℕ} def SaddleFunctionEquivalent (K L : SaddleFunction m n) : Prop

Lean source

Text 34.1.8 is false in the unrestricted textbook form for the present formalization: the same open-unit power example already refutes the claimed global saddle-equivalence of the two simple extensions.

open Set variable {m n : ℕ} theorem section34_text_34_1_8_unrestricted_false : ¬ helperForText_34_1_8_targetTheoremClaim

Lean source

Text 34.1.8 has a concrete one-dimensional counterexample in the present formalization: the open-unit power kernel satisfies all local hypotheses, but its lower and upper simple extensions are not saddle-equivalent on ℝ × ℝ.

open Set variable {m n : ℕ} theorem section34_text_34_1_8_has_openUnitPowerCounterexample : ∃ (C D : Set (Fin 1 → ℝ)) (J : SaddleFunction 1 1), Convex ℝ C ∧ Convex ℝ D ∧ C.Nonempty ∧ D.Nonempty ∧ IsConcaveConvexOn C D J ∧ (∀ u ∈ C, ∀ v ∈ D, (⊥ : EReal) < J u v ∧ J u v < (⊤ : EReal)) ∧ ¬ saddleEquivalent (lowerSimpleExtension C D J) (upperSimpleExtension C D J)

Lean source

Corrected Text 34.1.8 for the present formalization: in the same finite concave-convex setting, the universally valid simple-extension relation is that the lower simple extension lies below the upper simple extension everywhere, and both agree with J on C × D. The stronger global saddle-equivalence statement is refuted above by section34_text_34_1_8_unrestricted_false.

open Set variable {m n : ℕ} theorem section34_text_34_1_8 {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (_hC : Convex ℝ C) (_hD : Convex ℝ D) (_hCne : C.Nonempty) (_hDne : D.Nonempty) (J : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (_hJ : IsConcaveConvexOn C D J) (_hJfinite : ∀ u ∈ C, ∀ v ∈ D, (⊥ : EReal) < J u v ∧ J u v < (⊤ : EReal)) : (∀ u v, lowerSimpleExtension C D J u v ≤ upperSimpleExtension C D J u v) ∧ ∀ u ∈ C, ∀ v ∈ D, lowerSimpleExtension C D J u v = J u v ∧ upperSimpleExtension C D J u v = J u v

Lean source

Strong coordinatewise-closure qualification needed for the reverse implication in Text 34.1.9. The weak epigraph/hypograph closure package alone does not imply these predicates in the presence of infinite values.

open Set variable {m n : ℕ} def Section34Text34_1_9StrongClosureQualification (m n : ℕ) : Prop := ∀ K : SaddleFunction m n, IsConcaveConvex K → IsConcaveConvex (partialClosure₁ K) ∧ IsConcaveConvex (partialClosure₂ K)

Lean source

Text 34.1.9: characterization of closedness by mixed closure identities under explicit strong coordinatewise-closure hypotheses.

open Set variable {m n : ℕ} theorem section34_text_34_1_9 (K : SaddleFunction m n) (hK : IsConcaveConvex K) (hCl₁K : IsConcaveConvex (partialClosure₁ K)) (hCl₂K : IsConcaveConvex (partialClosure₂ K)) : IsClosedSaddleFunction K ↔ (partialClosure₁ (partialClosure₂ K) = partialClosure₁ K) ∧ (partialClosure₂ (partialClosure₁ K) = partialClosure₂ K)

Lean source

Text 34.1.10: equivalence preserves closedness.

open Set variable {m n : ℕ} theorem section34_text_34_1_10 {K L : SaddleFunction m n} (hK : IsClosedSaddleFunction K) (hEq : saddleEquivalent K L) : IsClosedSaddleFunction L

Lean source

Helper for Text 34.1.7: the isolated theorem claim is false, because specializing it to the open-unit-square power example forces a global upper-extension saddle property contradicted by openUnitSquarePowerSaddle.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_targetTheoremClaim_false : ¬ helperForText_34_1_7_targetTheoremClaim

Lean source

Helper for Text 34.1.7: finite values on C × D force the first effective domain of the lower simple extension to be exactly C.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_effectiveDomain1_lowerSimpleExtension_eq {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hDne : D.Nonempty) (J : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hJfinite : ∀ u ∈ C, ∀ v ∈ D, (⊥ : EReal) < J u v ∧ J u v < (⊤ : EReal)) : effectiveDomain₁ (lowerSimpleExtension C D J) = C

Lean source

Helper for Text 34.1.7: finite values on C × D force the second effective domain of the lower simple extension to be exactly D.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_effectiveDomain2_lowerSimpleExtension_eq {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hCne : C.Nonempty) (J : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hJfinite : ∀ u ∈ C, ∀ v ∈ D, (⊥ : EReal) < J u v ∧ J u v < (⊤ : EReal)) : effectiveDomain₂ (lowerSimpleExtension C D J) = D

Lean source

Helper for Text 34.1.7: finite values on C × D force the first effective domain of the upper simple extension to be exactly C.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_effectiveDomain1_upperSimpleExtension_eq {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hDne : D.Nonempty) (J : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hJfinite : ∀ u ∈ C, ∀ v ∈ D, (⊥ : EReal) < J u v ∧ J u v < (⊤ : EReal)) : effectiveDomain₁ (upperSimpleExtension C D J) = C

Lean source

Helper for Text 34.1.7: finite values on C × D force the second effective domain of the upper simple extension to be exactly D.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_effectiveDomain2_upperSimpleExtension_eq {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (J : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hJfinite : ∀ u ∈ C, ∀ v ∈ D, (⊥ : EReal) < J u v ∧ J u v < (⊤ : EReal)) : effectiveDomain₂ (upperSimpleExtension C D J) = D

Lean source

Helper for Text 34.1.7: the lower simple extension has saddle effective domain C × D once its coordinatewise effective domains are identified.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_saddleEffectiveDomain_lowerSimpleExtension_eq {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hCne : C.Nonempty) (hDne : D.Nonempty) (J : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hJfinite : ∀ u ∈ C, ∀ v ∈ D, (⊥ : EReal) < J u v ∧ J u v < (⊤ : EReal)) : saddleEffectiveDomain (lowerSimpleExtension C D J) = C ×ˢ D

Lean source

Helper for Text 34.1.7: the upper simple extension has saddle effective domain C × D once its coordinatewise effective domains are identified.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_saddleEffectiveDomain_upperSimpleExtension_eq {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hDne : D.Nonempty) (J : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hJfinite : ∀ u ∈ C, ∀ v ∈ D, (⊥ : EReal) < J u v ∧ J u v < (⊤ : EReal)) : saddleEffectiveDomain (upperSimpleExtension C D J) = C ×ˢ D

Lean source

Helper for Text 34.1.7: finiteness on C × D determines all coordinatewise and saddle effective domains of both simple extensions.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_simpleExtensions_have_expected_domains {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hCne : C.Nonempty) (hDne : D.Nonempty) (J : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hJfinite : ∀ u ∈ C, ∀ v ∈ D, (⊥ : EReal) < J u v ∧ J u v < (⊤ : EReal)) : effectiveDomain₁ (lowerSimpleExtension C D J) = C ∧ effectiveDomain₂ (lowerSimpleExtension C D J) = D ∧ saddleEffectiveDomain (lowerSimpleExtension C D J) = C ×ˢ D ∧ effectiveDomain₁ (upperSimpleExtension C D J) = C ∧ effectiveDomain₂ (upperSimpleExtension C D J) = D ∧ saddleEffectiveDomain (upperSimpleExtension C D J) = C ×ˢ D

Lean source

Helper for Text 34.1.7: once the saddle effective domain of K is exactly C × D, nonemptiness of C and D forces K to be proper.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_isProper_of_saddleEffectiveDomain_eq {K : SaddleFunction m n} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hCne : C.Nonempty) (hDne : D.Nonempty) (hDom : saddleEffectiveDomain K = C ×ˢ D) : IsProperSaddleFunction K

Lean source

Helper for Text 34.1.7: nonempty C and D make the lower simple extension proper, because its saddle effective domain is exactly C × D.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_lowerSimpleExtension_isProper {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hCne : C.Nonempty) (hDne : D.Nonempty) (J : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hJfinite : ∀ u ∈ C, ∀ v ∈ D, (⊥ : EReal) < J u v ∧ J u v < (⊤ : EReal)) : IsProperSaddleFunction (lowerSimpleExtension C D J)

Lean source

Helper for Text 34.1.7: nonempty C and D also make the upper simple extension proper, because its saddle effective domain is exactly C × D.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_upperSimpleExtension_isProper {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hCne : C.Nonempty) (hDne : D.Nonempty) (J : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hJfinite : ∀ u ∈ C, ∀ v ∈ D, (⊥ : EReal) < J u v ∧ J u v < (⊤ : EReal)) : IsProperSaddleFunction (upperSimpleExtension C D J)

Lean source

Helper for Text 34.1.7: even though the global upper-saddle assertion is false in general, the two simple extensions still satisfy the domain and properness conclusions forced by finiteness on C × D.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_simpleExtensions_have_expected_domains_and_properness {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hCne : C.Nonempty) (hDne : D.Nonempty) (J : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hJfinite : ∀ u ∈ C, ∀ v ∈ D, (⊥ : EReal) < J u v ∧ J u v < (⊤ : EReal)) : effectiveDomain₁ (lowerSimpleExtension C D J) = C ∧ effectiveDomain₂ (lowerSimpleExtension C D J) = D ∧ saddleEffectiveDomain (lowerSimpleExtension C D J) = C ×ˢ D ∧ IsProperSaddleFunction (lowerSimpleExtension C D J) ∧ effectiveDomain₁ (upperSimpleExtension C D J) = C ∧ effectiveDomain₂ (upperSimpleExtension C D J) = D ∧ saddleEffectiveDomain (upperSimpleExtension C D J) = C ×ˢ D ∧ IsProperSaddleFunction (upperSimpleExtension C D J)

Lean source

Helper for Text 34.1.7: the isolated theorem claim is logically equivalent to False, because any inhabitant is already refuted by the open-unit-square counterexample.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_targetTheoremClaim_iff_false : helperForText_34_1_7_targetTheoremClaim ↔ False

Lean source

Helper for Text 34.1.7: the isolated theorem claim is an empty type, so no local proof can close the theorem without repairing the statement outside PROOF stage.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_targetTheoremClaim_isEmpty : IsEmpty helperForText_34_1_7_targetTheoremClaim

Lean source

Helper for Text 34.1.7: the open-unit power counterexample still satisfies the corrected domain and properness conclusions, even though it refutes the unrestricted upper-saddle branch.

open Set variable {m n : ℕ} lemma helperForText_34_1_7_openUnitPowerKernel_has_expected_domains_and_properness : effectiveDomain₁ (lowerSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel) = {u : Fin 1 → ℝ | InOpenUnitInterval u} ∧ effectiveDomain₂ (lowerSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel) = {v : Fin 1 → ℝ | InOpenUnitInterval v} ∧ saddleEffectiveDomain (lowerSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel) = {u : Fin 1 → ℝ | InOpenUnitInterval u} ×ˢ {v : Fin 1 → ℝ | InOpenUnitInterval v} ∧ IsProperSaddleFunction (lowerSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel) ∧ effectiveDomain₁ (upperSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel) = {u : Fin 1 → ℝ | InOpenUnitInterval u} ∧ effectiveDomain₂ (upperSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel) = {v : Fin 1 → ℝ | InOpenUnitInterval v} ∧ saddleEffectiveDomain (upperSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel) = {u : Fin 1 → ℝ | InOpenUnitInterval u} ×ˢ {v : Fin 1 → ℝ | InOpenUnitInterval v} ∧ IsProperSaddleFunction (upperSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel)

Lean source

Helper for Text 34.1.8: the exact unrestricted textbook equivalence scheme, isolated so the same open-unit power counterexample can specialize it directly.

open Set variable {m n : ℕ} def helperForText_34_1_8_targetTheoremClaim : Prop := ∀ {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)}, Convex ℝ C → Convex ℝ D → C.Nonempty → D.Nonempty → ∀ J : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsConcaveConvexOn C D J → (∀ u ∈ C, ∀ v ∈ D, (⊥ : EReal) < J u v ∧ J u v < (⊤ : EReal)) → saddleEquivalent (lowerSimpleExtension C D J) (upperSimpleExtension C D J)

Lean source

Helper for Text 34.1.8: the right-hand member of a saddleEquivalent pair is globally concave-convex, hence globally a saddle-function on Set.univ × Set.univ.

open Set variable {m n : ℕ} lemma helperForText_34_1_8_right_isConcaveConvex_of_saddleEquivalent {K L : SaddleFunction m n} (hEq : saddleEquivalent K L) : IsConcaveConvex L

Lean source

Helper for Text 34.1.8: the global concave-convexity extracted from saddleEquivalent reinterprets the right-hand member as a saddle-function on Set.univ × Set.univ.

open Set variable {m n : ℕ} lemma helperForText_34_1_8_isSaddleOnUniv_right_of_saddleEquivalent {K L : SaddleFunction m n} (hEq : saddleEquivalent K L) : IsSaddleFunctionOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) L

Lean source

Helper for Text 34.1.8: in the open-unit power specialization, the upper simple extension is not globally concave-convex, because that would force the global saddle property already ruled out in Text 34.1.7.

open Set variable {m n : ℕ} lemma helperForText_34_1_8_specializedUpperSimpleExtension_not_concaveConvex : ¬ IsConcaveConvex (upperSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel)

Lean source

Helper for Text 34.1.8: for the open-unit power kernel, the asserted saddle-equivalence already fails because the upper simple extension is not even a global saddle-function.

open Set variable {m n : ℕ} lemma helperForText_34_1_8_specializedConclusion_false : ¬ saddleEquivalent (lowerSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel) (upperSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel)

Lean source

Helper for Text 34.1.8: any witness of the unrestricted theorem claim specializes directly to the forbidden open-unit saddle-equivalence conclusion.

open Set variable {m n : ℕ} lemma helperForText_34_1_8_targetClaim_forces_specializedConclusion (hClaim : helperForText_34_1_8_targetTheoremClaim) : saddleEquivalent (lowerSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel) (upperSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel)

Lean source

Helper for Text 34.1.8: any witness of the unrestricted theorem claim forces the specialized open-unit upper simple extension to be globally concave-convex.

open Set variable {m n : ℕ} lemma helperForText_34_1_8_targetClaim_forces_specializedUpperSimpleExtension_concaveConvex (hClaim : helperForText_34_1_8_targetTheoremClaim) : IsConcaveConvex (upperSimpleExtension {u : Fin 1 → ℝ | InOpenUnitInterval u} {v : Fin 1 → ℝ | InOpenUnitInterval v} oneDimensionalPowerKernel)

Lean source

Helper for Text 34.1.8: the isolated unrestricted theorem claim is equivalent to False, because every instance specializes to the open-unit power counterexample.

open Set variable {m n : ℕ} lemma helperForText_34_1_8_targetTheoremClaim_iff_false : helperForText_34_1_8_targetTheoremClaim ↔ False

Lean source

Helper for Text 34.1.8: the unrestricted theorem claim is an empty type in the present formalization, so the corrected theorem below is the only provable version in this file.

open Set variable {m n : ℕ} lemma helperForText_34_1_8_targetTheoremClaim_isEmpty : IsEmpty helperForText_34_1_8_targetTheoremClaim

Lean source

Helper for Text 34.1.9: closedness immediately yields the two mixed closure fixed-point identities by projecting the equality fields from the two saddleEquivalent witnesses.

open Set variable {m n : ℕ} lemma helperForText_34_1_9_mixedClosureIdentities_of_isClosed {K : SaddleFunction m n} (hClosed : IsClosedSaddleFunction K) : (partialClosure₁ (partialClosure₂ K) = partialClosure₁ K) ∧ (partialClosure₂ (partialClosure₁ K) = partialClosure₂ K)

Lean source

Helper for Text 34.1.9: the two mixed closure fixed-point identities reconstruct the two saddleEquivalent witnesses required for closedness.

open Set variable {m n : ℕ} lemma helperForText_34_1_9_saddleEquivalent_partialClosure1_of_mixedIdentities {K : SaddleFunction m n} (hK : IsConcaveConvex K) (hCl₁K : IsConcaveConvex (partialClosure₁ K)) (hMixed₂ : partialClosure₂ (partialClosure₁ K) = partialClosure₂ K) : saddleEquivalent K (partialClosure₁ K)

Lean source

Helper for Text 34.1.9: the two mixed closure fixed-point identities reconstruct the two saddleEquivalent witnesses required for closedness.

open Set variable {m n : ℕ} lemma helperForText_34_1_9_saddleEquivalent_partialClosure2_of_mixedIdentities {K : SaddleFunction m n} (hK : IsConcaveConvex K) (hCl₂K : IsConcaveConvex (partialClosure₂ K)) (hMixed₁ : partialClosure₁ (partialClosure₂ K) = partialClosure₁ K) : saddleEquivalent K (partialClosure₂ K)

Lean source

Helper for Text 34.1.9: the two mixed closure fixed-point identities reconstruct closedness by supplying the two equivalence witnesses with cl₁ K and cl₂ K.

open Set variable {m n : ℕ} lemma helperForText_34_1_9_isClosed_of_mixedClosureIdentities {K : SaddleFunction m n} (hK : IsConcaveConvex K) (hCl₁K : IsConcaveConvex (partialClosure₁ K)) (hCl₂K : IsConcaveConvex (partialClosure₂ K)) (hMixed : (partialClosure₁ (partialClosure₂ K) = partialClosure₁ K) ∧ (partialClosure₂ (partialClosure₁ K) = partialClosure₂ K)) : IsClosedSaddleFunction K

Lean source

Helper for Text 34.1.10: equality of both partial closures transports the first mixed fixed-point identity from K to L.

open Set variable {m n : ℕ} lemma helperForText_34_1_10_transportPartialClosure1FixedPoint {K L : SaddleFunction m n} (hPc1 : partialClosure₁ K = partialClosure₁ L) (hPc2 : partialClosure₂ K = partialClosure₂ L) (hMixed₁K : partialClosure₁ (partialClosure₂ K) = partialClosure₁ K) : partialClosure₁ (partialClosure₂ L) = partialClosure₁ L

Lean source

Helper for Text 34.1.10: equality of both partial closures transports the second mixed fixed-point identity from K to L.

open Set variable {m n : ℕ} lemma helperForText_34_1_10_transportPartialClosure2FixedPoint {K L : SaddleFunction m n} (hPc1 : partialClosure₁ K = partialClosure₁ L) (hPc2 : partialClosure₂ K = partialClosure₂ L) (hMixed₂K : partialClosure₂ (partialClosure₁ K) = partialClosure₂ K) : partialClosure₂ (partialClosure₁ L) = partialClosure₂ L

Lean source

Helper for Text 34.1.10: closedness of K and saddle-equivalence with L transport both mixed closure fixed-point identities from K to L.

open Set variable {m n : ℕ} lemma helperForText_34_1_10_transportMixedClosureIdentities {K L : SaddleFunction m n} (hK : IsClosedSaddleFunction K) (hEq : saddleEquivalent K L) : (partialClosure₁ (partialClosure₂ L) = partialClosure₁ L) ∧ (partialClosure₂ (partialClosure₁ L) = partialClosure₂ L)

§34 — Closures and Equivalence Classes — Part 16

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Theorem 34.2, generated-class core: Ω(F) is the saddle-equivalence class generated by the canonical kernel, and the canonical kernel is a member of that class. The additional adjoint, closed-representative, and realization conclusions in the prose theorem need their corresponding closedness or realization qualifications; they are exposed separately by the qualified projection API below rather than being asserted for arbitrary representatives.

open Set variable {m n : ℕ} theorem section34_theorem34_2 (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) : omegaClassOfConvexBifunction F = EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩ ∧ omegaClassOfConvexBifunction F = {K | saddleEquivalent K (convexBifunctionClosedKernel F)} ∧ convexBifunctionClosedKernel F ∈ omegaClassOfConvexBifunction F

Lean source

The data attached by Theorem 34.2 to a closed representative of the generated class.

open Set variable {m n : ℕ} def Section34Theorem34_2ClosedRepresentativeData (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (K : SaddleFunction m n) : Prop

Lean source

Explicit qualifications for the adjoint and closed-representative parts of Theorem 34.2. The representative data is deliberately required only after closedness of that representative has been supplied.

open Set variable {m n : ℕ} structure Section34Theorem34_2Qualification (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop where adjointPairing_eq : convexBifunctionClosedKernel F = section34ConcaveBifunctionAdjointPairing F adjoint_graphClosed : IsFunctionConcaveClosed (graphFunctionOfBifunction (section34ConcaveBifunctionAdjoint F)) adjoint_imageClosed : IsImageClosedConcaveBifunction (section34ConcaveBifunctionAdjoint F) canonicalKernel_closed : IsClosedSaddleFunction (convexBifunctionClosedKernel F) closedRepresentativeData : ∀ K : SaddleFunction m n, K ∈ omegaClassOfConvexBifunction F → IsClosedSaddleFunction K → Section34Theorem34_2ClosedRepresentativeData F K

Lean source

Explicit qualification for the converse realization part of Theorem 34.2.

open Set variable {m n : ℕ} def Section34Theorem34_2UniqueRealizationQualification (m n : ℕ) : Prop := ∀ Omega : Set (SaddleFunction m n), IsClosedConcaveConvexEquivalenceClass Omega → ∃! G : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsClosedConvexBifunction G ∧ Omega = omegaClassOfConvexBifunction G

Lean source

Dimension-wide qualification used by the converse and characterization projections.

open Set variable {m n : ℕ} structure Section34Theorem34_2GlobalQualification (m n : ℕ) : Prop where qualification : ∀ (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal), IsClosedConvexBifunction F → Section34Theorem34_2Qualification F uniqueRealization : Section34Theorem34_2UniqueRealizationQualification m n canonicalClosureRealization : Section34CanonicalClosureRealizationQualification m n

Lean source

Qualified full projection package for Theorem 34.2. Its generated-class core is supplied by section34_theorem34_2; the stronger conclusions are exactly the explicit qualifications above.

open Set variable {m n : ℕ} theorem section34_theorem34_2_qualified (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (hQ : Section34Theorem34_2Qualification F) : convexBifunctionClosedKernel F = section34ConcaveBifunctionAdjointPairing F ∧ omegaClassOfConvexBifunction F = {K | saddleEquivalent K (convexBifunctionClosedKernel F)} ∧ convexBifunctionClosedKernel F ∈ omegaClassOfConvexBifunction F ∧ (∀ K : SaddleFunction m n, K ∈ omegaClassOfConvexBifunction F → IsClosedSaddleFunction K) ∧ ∀ K : SaddleFunction m n, K ∈ omegaClassOfConvexBifunction F → Section34Theorem34_2ClosedRepresentativeData F K

Lean source

Generated-class form of Theorem 34.2: the canonical class Ω(F) coincides with the EquivalenceClassGeneratedByConvexBifunction attached to the same closed convex bifunction.

open Set variable {m n : ℕ} theorem omegaClassOfConvexBifunction_eq_generatedClass (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) : omegaClassOfConvexBifunction F = EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩

Lean source

Original-wording form of Theorem 34.2: the canonical class Ω(F) is exactly the saddleEquivalent-class of the canonical closed kernel generated by F.

open Set variable {m n : ℕ} theorem omegaClassOfConvexBifunction_eq_saddleEquivalenceClass (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) : omegaClassOfConvexBifunction F = {K | saddleEquivalent K (convexBifunctionClosedKernel F)}

Lean source

Original-wording membership form of Theorem 34.2: K ∈ Ω(F) iff K is saddle-equivalent to the canonical closed kernel of F.

open Set variable {m n : ℕ} theorem mem_omegaClassOfConvexBifunction_iff_saddleEquivalent_closedKernel (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} : K ∈ omegaClassOfConvexBifunction F ↔ saddleEquivalent K (convexBifunctionClosedKernel F)

Lean source

Membership bridge for Theorem 34.2: K ∈ Ω(F) iff K belongs to the generated equivalence class of F.

open Set variable {m n : ℕ} theorem mem_omegaClassOfConvexBifunction_iff_mem_generatedClass (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} : K ∈ omegaClassOfConvexBifunction F ↔ K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩

Lean source

Generated-class membership form of Theorem 34.2: generated-class membership is equivalent to saddle-equivalence with the canonical closed kernel.

open Set variable {m n : ℕ} theorem mem_generatedClass_iff_saddleEquivalent_closedKernel (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩ ↔ saddleEquivalent K (convexBifunctionClosedKernel F)

Lean source

Original-wording class identity for Theorem 34.2: the generated equivalence class of a closed convex bifunction is exactly the saddle-equivalence class of its canonical closed kernel.

open Set variable {m n : ℕ} theorem equivalenceClassGeneratedByConvexBifunction_eq_saddleEquivalenceClass (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) : EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩ = {K | saddleEquivalent K (convexBifunctionClosedKernel F)}

Lean source

One-way original-wording membership bridge for Theorem 34.2: any saddle-function saddle-equivalent to the canonical closed kernel belongs to Ω(F).

open Set variable {m n : ℕ} theorem mem_omegaClassOfConvexBifunction_of_saddleEquivalent_closedKernel (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : saddleEquivalent K (convexBifunctionClosedKernel F)) : K ∈ omegaClassOfConvexBifunction F

Lean source

One-way original-wording membership bridge for Theorem 34.2: any saddle-function saddle-equivalent to the canonical closed kernel belongs to the generated equivalence class of F.

open Set variable {m n : ℕ} theorem mem_generatedClass_of_saddleEquivalent_closedKernel (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : saddleEquivalent K (convexBifunctionClosedKernel F)) : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩

Lean source

One-way membership bridge for Theorem 34.2: generated-class membership implies membership in Ω(F).

open Set variable {m n : ℕ} theorem mem_omegaClassOfConvexBifunction_of_mem_generatedClass (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) : K ∈ omegaClassOfConvexBifunction F

Lean source

One-way membership bridge for Theorem 34.2: membership in Ω(F) implies generated-class membership.

open Set variable {m n : ℕ} theorem mem_generatedClass_of_mem_omegaClassOfConvexBifunction (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : K ∈ omegaClassOfConvexBifunction F) : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩

§34 — Closures and Equivalence Classes — Part 17

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Canonical-kernel membership form of Theorem 34.2: the canonical closed kernel belongs to Ω(F).

open Set variable {m n : ℕ} theorem convexBifunctionClosedKernel_mem_omegaClassOfConvexBifunction (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) : convexBifunctionClosedKernel F ∈ omegaClassOfConvexBifunction F

Lean source

Canonical-kernel membership form of Theorem 34.2: the kernel generated by a closed convex bifunction belongs to its generated equivalence class.

open Set variable {m n : ℕ} theorem convexBifunctionClosedKernel_mem_generatedClass (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) : convexBifunctionClosedKernel F ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩

Lean source

Closedness form of Theorem 34.2 for the canonical representative: the canonical closed kernel of a closed convex bifunction is itself a closed saddle-function.

open Set variable {m n : ℕ} theorem isClosedSaddleFunction_convexBifunctionClosedKernel (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (hQ : Section34Theorem34_2Qualification F) : IsClosedSaddleFunction (convexBifunctionClosedKernel F)

Lean source

Class-level forward projection of Theorem 34.2: for every closed convex bifunction F, the canonical class Ω(F) is a closed concave-convex equivalence class.

open Set variable {m n : ℕ} theorem isClosedConcaveConvexEquivalenceClass_omegaClassOfConvexBifunction (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (hQ : Section34Theorem34_2Qualification F) : IsClosedConcaveConvexEquivalenceClass (omegaClassOfConvexBifunction F)

Lean source

Original-wording forward projection of Theorem 34.2: the saddle-equivalence class of the canonical closed kernel is a closed concave-convex equivalence class.

open Set variable {m n : ℕ} theorem isClosedConcaveConvexEquivalenceClass_saddleEquivalenceClass_closedKernel (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (hQ : Section34Theorem34_2Qualification F) : IsClosedConcaveConvexEquivalenceClass {K | saddleEquivalent K (convexBifunctionClosedKernel F)}

Lean source

Class-level forward projection of Theorem 34.2: the generated equivalence class of a closed convex bifunction is a closed concave-convex equivalence class.

open Set variable {m n : ℕ} theorem isClosedConcaveConvexEquivalenceClass_generatedClass (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (hQ : Section34Theorem34_2Qualification F) : IsClosedConcaveConvexEquivalenceClass (EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩)

Lean source

A fixed seed closed convex bifunction, obtained from the constant-zero kernel, used to instantiate the converse half of Theorem 34.2 without introducing any extra assumptions.

open Set variable {m n : ℕ} lemma exists_closedConvexBifunction_seed : (∃ F₀ : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsClosedConvexBifunction F₀) → ∃ F₀ : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsClosedConvexBifunction F₀

Lean source

Converse projection of Theorem 34.2: every closed concave-convex equivalence class is Ω(F) for a unique closed convex bifunction F.

open Set variable {m n : ℕ} theorem existsUnique_closedConvexBifunction_of_closedConcaveConvexEquivalenceClass (Omega : Set (SaddleFunction m n)) (hOmega : IsClosedConcaveConvexEquivalenceClass Omega) (hGlobal : Section34Theorem34_2GlobalQualification m n) : ∃! G : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsClosedConvexBifunction G ∧ Omega = omegaClassOfConvexBifunction G

Lean source

Existence form of the converse half of Theorem 34.2: every closed concave-convex equivalence class is Ω(F) for some closed convex bifunction F.

open Set variable {m n : ℕ} theorem exists_closedConvexBifunction_of_closedConcaveConvexEquivalenceClass (Omega : Set (SaddleFunction m n)) (hOmega : IsClosedConcaveConvexEquivalenceClass Omega) (hGlobal : Section34Theorem34_2GlobalQualification m n) : ∃ G : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsClosedConvexBifunction G ∧ Omega = omegaClassOfConvexBifunction G

Lean source

Uniqueness form of the converse half of Theorem 34.2: two closed convex bifunctions generating the same closed concave-convex equivalence class must coincide.

open Set variable {m n : ℕ} theorem closedConvexBifunction_eq_of_closedConcaveConvexEquivalenceClass_eq (Omega : Set (SaddleFunction m n)) (hOmega : IsClosedConcaveConvexEquivalenceClass Omega) {G H : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hG : IsClosedConvexBifunction G) (hH : IsClosedConvexBifunction H) (hOmegaG : Omega = omegaClassOfConvexBifunction G) (hOmegaH : Omega = omegaClassOfConvexBifunction H) (hGlobal : Section34Theorem34_2GlobalQualification m n) : G = H

Lean source

Original-wording converse form of Theorem 34.2: every closed concave-convex equivalence class is the saddle-equivalence class of the canonical closed kernel of a unique closed convex bifunction.

open Set variable {m n : ℕ} theorem existsUnique_closedConvexBifunction_of_closedConcaveConvexEquivalenceClass_saddleEquivalenceClass (Omega : Set (SaddleFunction m n)) (hOmega : IsClosedConcaveConvexEquivalenceClass Omega) (hGlobal : Section34Theorem34_2GlobalQualification m n) : ∃! G : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsClosedConvexBifunction G ∧ Omega = {K | saddleEquivalent K (convexBifunctionClosedKernel G)}

Lean source

Original-wording existence form of the converse half of Theorem 34.2.

open Set variable {m n : ℕ} theorem exists_closedConvexBifunction_of_closedConcaveConvexEquivalenceClass_saddleEquivalenceClass (Omega : Set (SaddleFunction m n)) (hOmega : IsClosedConcaveConvexEquivalenceClass Omega) (hGlobal : Section34Theorem34_2GlobalQualification m n) : ∃ G : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsClosedConvexBifunction G ∧ Omega = {K | saddleEquivalent K (convexBifunctionClosedKernel G)}

Lean source

Original-wording uniqueness form of the converse half of Theorem 34.2.

open Set variable {m n : ℕ} theorem closedConvexBifunction_eq_of_closedConcaveConvexEquivalenceClass_eq_saddleEquivalenceClass (Omega : Set (SaddleFunction m n)) (hOmega : IsClosedConcaveConvexEquivalenceClass Omega) {G H : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hG : IsClosedConvexBifunction G) (hH : IsClosedConvexBifunction H) (hOmegaG : Omega = {K | saddleEquivalent K (convexBifunctionClosedKernel G)}) (hOmegaH : Omega = {K | saddleEquivalent K (convexBifunctionClosedKernel H)}) (hGlobal : Section34Theorem34_2GlobalQualification m n) : G = H

Lean source

Characterization form of Theorem 34.2: a class of saddle-functions is a closed concave-convex equivalence class exactly when it is Ω(F) for some closed convex bifunction F.

open Set variable {m n : ℕ} theorem isClosedConcaveConvexEquivalenceClass_iff_exists_closedConvexBifunction (Omega : Set (SaddleFunction m n)) (hGlobal : Section34Theorem34_2GlobalQualification m n) : IsClosedConcaveConvexEquivalenceClass Omega ↔ ∃ G : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsClosedConvexBifunction G ∧ Omega = omegaClassOfConvexBifunction G

Lean source

Original-wording characterization form of Theorem 34.2: a class of saddle-functions is a closed concave-convex equivalence class exactly when it is the saddle-equivalence class of the canonical closed kernel of some closed convex bifunction.

open Set variable {m n : ℕ} theorem isClosedConcaveConvexEquivalenceClass_iff_exists_closedConvexBifunction_saddleEquivalenceClass (Omega : Set (SaddleFunction m n)) (hGlobal : Section34Theorem34_2GlobalQualification m n) : IsClosedConcaveConvexEquivalenceClass Omega ↔ ∃ G : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsClosedConvexBifunction G ∧ Omega = {K | saddleEquivalent K (convexBifunctionClosedKernel G)}

Lean source

Unique-existence characterization form of Theorem 34.2.

open Set variable {m n : ℕ} theorem isClosedConcaveConvexEquivalenceClass_iff_existsUnique_closedConvexBifunction (Omega : Set (SaddleFunction m n)) (hGlobal : Section34Theorem34_2GlobalQualification m n) : IsClosedConcaveConvexEquivalenceClass Omega ↔ ∃! G : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsClosedConvexBifunction G ∧ Omega = omegaClassOfConvexBifunction G

Lean source

Original-wording unique-existence characterization form of Theorem 34.2.

open Set variable {m n : ℕ} theorem isClosedConcaveConvexEquivalenceClass_iff_existsUnique_closedConvexBifunction_saddleEquivalenceClass (Omega : Set (SaddleFunction m n)) (hGlobal : Section34Theorem34_2GlobalQualification m n) : IsClosedConcaveConvexEquivalenceClass Omega ↔ ∃! G : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsClosedConvexBifunction G ∧ Omega = {K | saddleEquivalent K (convexBifunctionClosedKernel G)}

Lean source

Generated-class characterization form of Theorem 34.2.

open Set variable {m n : ℕ} theorem isClosedConcaveConvexEquivalenceClass_iff_exists_closedConvexBifunction_generatedClass (Omega : Set (SaddleFunction m n)) (hGlobal : Section34Theorem34_2GlobalQualification m n) : IsClosedConcaveConvexEquivalenceClass Omega ↔ ∃ G : (Fin m → ℝ) → (Fin n → ℝ) → EReal, ∃ hG : IsClosedConvexBifunction G, Omega = EquivalenceClassGeneratedByConvexBifunction ⟨G, hG⟩

Lean source

Generated-class unique-existence characterization form of Theorem 34.2.

open Set variable {m n : ℕ} theorem isClosedConcaveConvexEquivalenceClass_iff_existsUnique_closedConvexBifunction_generatedClass (Omega : Set (SaddleFunction m n)) (hGlobal : Section34Theorem34_2GlobalQualification m n) : IsClosedConcaveConvexEquivalenceClass Omega ↔ ∃! G : (Fin m → ℝ) → (Fin n → ℝ) → EReal, ∃ hG : IsClosedConvexBifunction G, Omega = EquivalenceClassGeneratedByConvexBifunction ⟨G, hG⟩

Lean source

Kernel-equality projection of Theorem 34.2: for a closed convex bifunction F, the canonical closed kernel ⟨F u, x*⟩ agrees with the adjoint kernel ⟨u, F* x*⟩.

open Set variable {m n : ℕ} theorem convexBifunctionClosedKernel_eq_section34ConcaveBifunctionAdjointPairing (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (hQ : Section34Theorem34_2Qualification F) : convexBifunctionClosedKernel F = section34ConcaveBifunctionAdjointPairing F

Lean source

Closedness projection of Theorem 34.2: every member of Ω(F) is closed.

open Set variable {m n : ℕ} theorem isClosedSaddleFunction_of_mem_omegaClassOfConvexBifunction (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : K ∈ omegaClassOfConvexBifunction F) (hQ : Section34Theorem34_2Qualification F) : IsClosedSaddleFunction K

Lean source

Closedness projection of Theorem 34.2: every member of the generated equivalence class of a closed convex bifunction is closed.

open Set variable {m n : ℕ} theorem isClosedSaddleFunction_of_mem_generatedClass (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) (hQ : Section34Theorem34_2Qualification F) : IsClosedSaddleFunction K

Lean source

Domain projection of Theorem 34.2: every member of Ω(F) has effective domain dom F × dom F*.

open Set variable {m n : ℕ} theorem saddleEffectiveDomain_eq_domain_prod_of_mem_omegaClassOfConvexBifunction (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : K ∈ omegaClassOfConvexBifunction F) (hKclosed : IsClosedSaddleFunction K) (hQ : Section34Theorem34_2Qualification F) : saddleEffectiveDomain K = convexBifunctionDomain F ×ˢ convexBifunctionKernelAdjointDomain F

Lean source

Domain projection of Theorem 34.2: every member of the generated equivalence class of a closed convex bifunction has effective domain dom F × dom F*.

open Set variable {m n : ℕ} theorem saddleEffectiveDomain_eq_domain_prod_of_mem_generatedClass (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) (hKclosed : IsClosedSaddleFunction K) (hQ : Section34Theorem34_2Qualification F) : saddleEffectiveDomain K = convexBifunctionDomain F ×ˢ convexBifunctionKernelAdjointDomain F

Lean source

First partial-closure projection of Theorem 34.2: every member of Ω(F) has first partial closure equal to the canonical kernel.

open Set variable {m n : ℕ} theorem partialClosure₁_eq_convexBifunctionClosedKernel_of_mem_omegaClassOfConvexBifunction (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : K ∈ omegaClassOfConvexBifunction F) (hKclosed : IsClosedSaddleFunction K) (hQ : Section34Theorem34_2Qualification F) : partialClosure₁ K = convexBifunctionClosedKernel F

Lean source

Second partial-closure projection of Theorem 34.2: every member of Ω(F) has second partial closure equal to the canonical kernel.

open Set variable {m n : ℕ} theorem partialClosure₂_eq_convexBifunctionClosedKernel_of_mem_omegaClassOfConvexBifunction (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : K ∈ omegaClassOfConvexBifunction F) (hKclosed : IsClosedSaddleFunction K) (hQ : Section34Theorem34_2Qualification F) : partialClosure₂ K = convexBifunctionClosedKernel F

Lean source

Primal recovery projection of Theorem 34.2: every member of Ω(F) recovers F by convex conjugation of its first slices.

open Set variable {m n : ℕ} theorem convexBifunction_eq_convexConjugate_slice_of_mem_omegaClassOfConvexBifunction (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : K ∈ omegaClassOfConvexBifunction F) (hKclosed : IsClosedSaddleFunction K) (hQ : Section34Theorem34_2Qualification F) : ∀ u x, F u x = convexConjugate (K u) x

Lean source

Adjoint recovery projection of Theorem 34.2: every member of Ω(F) recovers the concave adjoint F* by concave conjugation of its second slices.

open Set variable {m n : ℕ} theorem concaveAdjoint_eq_concaveConjugate_slice_of_mem_omegaClassOfConvexBifunction (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : K ∈ omegaClassOfConvexBifunction F) (hKclosed : IsClosedSaddleFunction K) (hQ : Section34Theorem34_2Qualification F) : ∀ xStar uStar, section34ConcaveBifunctionAdjoint F xStar uStar = concaveConjugate (fun u => K u xStar) uStar

Lean source

Relative-interior agreement projection of Theorem 34.2: every member of Ω(F) agrees with the canonical kernel on ri(dom F) in the first variable or ri(dom F*) in the second variable.

open Set variable {m n : ℕ} theorem eq_convexBifunctionClosedKernel_on_relativeInterior_of_mem_omegaClassOfConvexBifunction (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : K ∈ omegaClassOfConvexBifunction F) (hKclosed : IsClosedSaddleFunction K) (hQ : Section34Theorem34_2Qualification F) : ∀ u xStar, u ∈ intrinsicInterior ℝ (convexBifunctionDomain F) ∨ xStar ∈ intrinsicInterior ℝ (convexBifunctionKernelAdjointDomain F) → K u xStar = convexBifunctionClosedKernel F u xStar

Lean source

Original-wording closedness projection of Theorem 34.2: every saddle-function saddle- equivalent to the canonical closed kernel is closed.

open Set variable {m n : ℕ} theorem isClosedSaddleFunction_of_saddleEquivalent_closedKernel (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : saddleEquivalent K (convexBifunctionClosedKernel F)) (hQ : Section34Theorem34_2Qualification F) : IsClosedSaddleFunction K

Lean source

Original-wording domain projection of Theorem 34.2: every saddle-function saddle-equivalent to the canonical closed kernel has effective domain dom F × dom F*.

open Set variable {m n : ℕ} theorem saddleEffectiveDomain_eq_domain_prod_of_saddleEquivalent_closedKernel (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : saddleEquivalent K (convexBifunctionClosedKernel F)) (hKclosed : IsClosedSaddleFunction K) (hQ : Section34Theorem34_2Qualification F) : saddleEffectiveDomain K = convexBifunctionDomain F ×ˢ convexBifunctionKernelAdjointDomain F

Lean source

Original-wording first partial-closure projection of Theorem 34.2.

open Set variable {m n : ℕ} theorem partialClosure₁_eq_convexBifunctionClosedKernel_of_saddleEquivalent_closedKernel (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : saddleEquivalent K (convexBifunctionClosedKernel F)) (hKclosed : IsClosedSaddleFunction K) (hQ : Section34Theorem34_2Qualification F) : partialClosure₁ K = convexBifunctionClosedKernel F

Lean source

Original-wording second partial-closure projection of Theorem 34.2.

open Set variable {m n : ℕ} theorem partialClosure₂_eq_convexBifunctionClosedKernel_of_saddleEquivalent_closedKernel (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : saddleEquivalent K (convexBifunctionClosedKernel F)) (hKclosed : IsClosedSaddleFunction K) (hQ : Section34Theorem34_2Qualification F) : partialClosure₂ K = convexBifunctionClosedKernel F

Lean source

Original-wording primal recovery projection of Theorem 34.2.

open Set variable {m n : ℕ} theorem convexBifunction_eq_convexConjugate_slice_of_saddleEquivalent_closedKernel (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : saddleEquivalent K (convexBifunctionClosedKernel F)) (hKclosed : IsClosedSaddleFunction K) (hQ : Section34Theorem34_2Qualification F) : ∀ u x, F u x = convexConjugate (K u) x

Lean source

Original-wording adjoint recovery projection of Theorem 34.2.

open Set variable {m n : ℕ} theorem concaveAdjoint_eq_concaveConjugate_slice_of_saddleEquivalent_closedKernel (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : saddleEquivalent K (convexBifunctionClosedKernel F)) (hKclosed : IsClosedSaddleFunction K) (hQ : Section34Theorem34_2Qualification F) : ∀ xStar uStar, section34ConcaveBifunctionAdjoint F xStar uStar = concaveConjugate (fun u => K u xStar) uStar

Lean source

Original-wording relative-interior agreement projection of Theorem 34.2.

open Set variable {m n : ℕ} theorem eq_convexBifunctionClosedKernel_on_relativeInterior_of_saddleEquivalent_closedKernel (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : saddleEquivalent K (convexBifunctionClosedKernel F)) (hKclosed : IsClosedSaddleFunction K) (hQ : Section34Theorem34_2Qualification F) : ∀ u xStar, u ∈ intrinsicInterior ℝ (convexBifunctionDomain F) ∨ xStar ∈ intrinsicInterior ℝ (convexBifunctionKernelAdjointDomain F) → K u xStar = convexBifunctionClosedKernel F u xStar

Lean source

Generated-class form of the first partial-closure projection in Theorem 34.2.

open Set variable {m n : ℕ} theorem partialClosure₁_eq_convexBifunctionClosedKernel_of_mem_generatedClass (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) (hKclosed : IsClosedSaddleFunction K) (hQ : Section34Theorem34_2Qualification F) : partialClosure₁ K = convexBifunctionClosedKernel F

Lean source

Generated-class form of the second partial-closure projection in Theorem 34.2.

open Set variable {m n : ℕ} theorem partialClosure₂_eq_convexBifunctionClosedKernel_of_mem_generatedClass (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) (hKclosed : IsClosedSaddleFunction K) (hQ : Section34Theorem34_2Qualification F) : partialClosure₂ K = convexBifunctionClosedKernel F

Lean source

Generated-class form of the primal recovery projection in Theorem 34.2.

open Set variable {m n : ℕ} theorem convexBifunction_eq_convexConjugate_slice_of_mem_generatedClass (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) (hKclosed : IsClosedSaddleFunction K) (hQ : Section34Theorem34_2Qualification F) : ∀ u x, F u x = convexConjugate (K u) x

Lean source

Generated-class form of the adjoint recovery projection in Theorem 34.2.

open Set variable {m n : ℕ} theorem concaveAdjoint_eq_concaveConjugate_slice_of_mem_generatedClass (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) (hKclosed : IsClosedSaddleFunction K) (hQ : Section34Theorem34_2Qualification F) : ∀ xStar uStar, section34ConcaveBifunctionAdjoint F xStar uStar = concaveConjugate (fun u => K u xStar) uStar

Lean source

Generated-class form of the relative-interior agreement projection in Theorem 34.2.

open Set variable {m n : ℕ} theorem eq_convexBifunctionClosedKernel_on_relativeInterior_of_mem_generatedClass (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) {K : SaddleFunction m n} (hK : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) (hKclosed : IsClosedSaddleFunction K) (hQ : Section34Theorem34_2Qualification F) : ∀ u xStar, u ∈ intrinsicInterior ℝ (convexBifunctionDomain F) ∨ xStar ∈ intrinsicInterior ℝ (convexBifunctionKernelAdjointDomain F) → K u xStar = convexBifunctionClosedKernel F u xStar

Lean source

Defn 34.8: a concave-convex function K on ℝ^m × ℝ^n is simple when every convex section K(u, ·) with u ∈ ri (dom₁ K) has effective domain contained in closure (dom₂ K), and every concave section K(·, v) with v ∈ ri (dom₂ K) has effective domain contained in closure (dom₁ K).

open Set variable {m n : ℕ} def IsSimpleConcaveConvex (K : SaddleFunction m n) : Prop := IsConcaveConvex K ∧ (∀ u, u ∈ intrinsicInterior ℝ (effectiveDomain₁ K) → convexFunctionEffectiveDomain (K u) ⊆ closure (effectiveDomain₂ K)) ∧ ∀ v, v ∈ intrinsicInterior ℝ (effectiveDomain₂ K) → concaveFunctionEffectiveDomain (fun u => K u v) ⊆ closure (effectiveDomain₁ K)

Lean source

A saddle-function is simple when it satisfies the section-domain containment condition in whichever of the two saddle orientations it carries.

open Set variable {m n : ℕ} def IsSimpleSaddleFunction (K : SaddleFunction m n) : Prop := IsSimpleConcaveConvex K ∨ (IsConvexConcave K ∧ (∀ u, u ∈ intrinsicInterior ℝ (effectiveDomain₁ K) → concaveFunctionEffectiveDomain (K u) ⊆ closure (effectiveDomain₂ K)) ∧ ∀ v, v ∈ intrinsicInterior ℝ (effectiveDomain₂ K) → convexFunctionEffectiveDomain (fun u => K u v) ⊆ closure (effectiveDomain₁ K))

Lean source

Forward closed-slice half of Theorem 34.3: for a proper concave-convex saddle-function on ℝ^m × ℝ^n, closedness forces the coordinate slice conditions listed in the text.

open Set variable {m n : ℕ} theorem closed_concaveConvex_iff_relativeInterior_slice_conditions (K : SaddleFunction m n) (hKproper : IsProperSaddleFunction K) (hKcc : IsConcaveConvex K) (hGlobal : Section34Theorem34_2GlobalQualification m n) : IsClosedSaddleFunction K → let C := effectiveDomain₁ K let D := effectiveDomain₂ K (∀ u, u ∈ intrinsicInterior ℝ C → IsProperClosedConvexFunctionWithDomain (K u) D) ∧ (∀ u, u ∈ C \ intrinsicInterior ℝ C → IsProperConvexFunctionWithDomainBetween (K u) D (closure D)) ∧ (∀ u, u ∉ C → IsImproperConvexFunction (K u) ∧ (∀ v, v ∈ intrinsicInterior ℝ D → K u v = (⊥ : EReal)) ∧ (u ∉ closure C → ∀ v, v ∈ D → K u v = (⊥ : EReal))) ∧ (∀ v, v ∈ intrinsicInterior ℝ D → IsProperClosedConcaveFunctionWithDomain (fun u => K u v) C) ∧ (∀ v, v ∈ D \ intrinsicInterior ℝ D → IsProperConcaveFunctionWithDomainBetween (fun u => K u v) C (closure C)) ∧ ∀ v, v ∉ D → IsImproperConcaveFunction (fun u => K u v) ∧ (∀ u, u ∈ intrinsicInterior ℝ C → K u v = (⊤ : EReal)) ∧ (v ∉ closure D → ∀ u, u ∈ C → K u v = (⊤ : EReal))

Lean source

Text 34.4.1: every closed proper concave-convex saddle-function K on ℝ^m × ℝ^n is simple.

open Set variable {m n : ℕ} theorem closed_proper_saddleFunction_isSimple (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) : IsSimpleConcaveConvex K

Lean source

Text 34.4.3: let K be a concave-convex saddle-function on ℝ^m × ℝ^n. If the effective domain of K has nonempty interior, i.e. interior (dom K) ≠ ∅, then K is simple.

open Set variable {m n : ℕ} theorem saddleFunction_nonemptyInterior_effectiveDomain_isSimple (K : SaddleFunction m n) (hK : IsConcaveConvex K) (hdom : interior (saddleEffectiveDomain K) ≠ (∅ : Set ((Fin m → ℝ) × (Fin n → ℝ)))) : IsSimpleSaddleFunction K

Lean source

Text 34.4.4: the concave-convex function K on ℝ × ℝ defined by K(u, v) = +∞ when uv > 0, K(u, v) = 0 when uv = 0, and K(u, v) = -∞ when uv < 0 is not a simple saddle-function.

open Set variable {m n : ℕ} theorem coordinateProductSignSaddle_not_simple : ¬ IsSimpleSaddleFunction coordinateProductSignSaddle

Lean source

Corrected formal core of Theorem 34.5: if K is a proper simple concave-convex saddle-function and its mixed closures satisfy cl₂ cl₁ K ≤ cl₁ cl₂ K, then those closures are saddle-equivalent. The stronger interval-class / same-kernel characterization is deferred until the missing bridge is formalized.

open Set variable {m n : ℕ} theorem simple_proper_concaveConvex_mixedClosure_interval_characterization (K : SaddleFunction m n) (_hKproper : IsProperSaddleFunction K) (hKsimple : IsSimpleConcaveConvex K) (hOrder : lowerClosureConcaveConvex K hKsimple.1 ≤ upperClosureConcaveConvex K hKsimple.1) (hLowerOrient : IsConcaveConvex (lowerClosureConcaveConvex K hKsimple.1)) (hUpperOrient : IsConcaveConvex (upperClosureConcaveConvex K hKsimple.1)) (hLowerClosed : IsConvexClosedInSecond (lowerClosureConcaveConvex K hKsimple.1)) (hUpperClosed : IsConcaveClosedInFirst (upperClosureConcaveConvex K hKsimple.1)) : let Klower := lowerClosureConcaveConvex K hKsimple.1 let Kupper := upperClosureConcaveConvex K hKsimple.1 saddleEquivalent Klower Kupper ∧ Klower ≤ Kupper

Lean source

Order projection of the corrected formal core of Theorem 34.5.

open Set variable {m n : ℕ} theorem simple_proper_concaveConvex_mixedClosure_interval_class (K : SaddleFunction m n) (hKproper : IsProperSaddleFunction K) (hKsimple : IsSimpleConcaveConvex K) (hOrder : lowerClosureConcaveConvex K hKsimple.1 ≤ upperClosureConcaveConvex K hKsimple.1) (hLowerOrient : IsConcaveConvex (lowerClosureConcaveConvex K hKsimple.1)) (hUpperOrient : IsConcaveConvex (upperClosureConcaveConvex K hKsimple.1)) (hLowerClosed : IsConvexClosedInSecond (lowerClosureConcaveConvex K hKsimple.1)) (hUpperClosed : IsConcaveClosedInFirst (upperClosureConcaveConvex K hKsimple.1)) : let Klower := lowerClosureConcaveConvex K hKsimple.1 let Kupper := upperClosureConcaveConvex K hKsimple.1 Klower ≤ Kupper

Lean source

Saddle-equivalence projection of the corrected formal core of Theorem 34.5.

open Set variable {m n : ℕ} theorem simple_proper_concaveConvex_mixedClosure_interval_eq_sameKernelClass (K : SaddleFunction m n) (hKproper : IsProperSaddleFunction K) (hKsimple : IsSimpleConcaveConvex K) (hOrder : lowerClosureConcaveConvex K hKsimple.1 ≤ upperClosureConcaveConvex K hKsimple.1) (hLowerOrient : IsConcaveConvex (lowerClosureConcaveConvex K hKsimple.1)) (hUpperOrient : IsConcaveConvex (upperClosureConcaveConvex K hKsimple.1)) (hLowerClosed : IsConvexClosedInSecond (lowerClosureConcaveConvex K hKsimple.1)) (hUpperClosed : IsConcaveClosedInFirst (upperClosureConcaveConvex K hKsimple.1)) : let Klower := lowerClosureConcaveConvex K hKsimple.1 let Kupper := upperClosureConcaveConvex K hKsimple.1 saddleEquivalent Klower Kupper

Lean source

A saddle-function has the prescribed relative-interior kernel of J on C × D when its kernel domain is ri C × ri D and its values there agree with J.

open Set variable {m n : ℕ} def HasPrescribedRelativeInteriorKernel (C : Set (Fin m → ℝ)) (D : Set (Fin n → ℝ)) (J : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) (L : SaddleFunction m n) : Prop

Lean source

Corrected formal packaging of Corollary 34.5.1: for nonempty convex C and D and a finite concave-convex kernel J on C × D, there is a unique explicit candidate set of closed proper concave-convex representatives carrying the prescribed relative-interior kernel. The stronger assertion that this set is already a closed concave-convex equivalence class is deferred until that missing bridge is formalized.

open Set variable {m n : ℕ} theorem unique_closed_proper_concaveConvex_equivalenceClass_with_prescribed_relativeInterior_kernel {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (_hCconv : Convex ℝ C) (_hDconv : Convex ℝ D) (_hCne : C.Nonempty) (_hDne : D.Nonempty) (J : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) (_hJcc : IsConcaveConvexOn C D (fun u v => ((J u v : ℝ) : EReal))) : ∃! Ω : Set (SaddleFunction m n), Ω = {L | IsConcaveConvex L ∧ IsClosedSaddleFunction L ∧ IsProperSaddleFunction L ∧ HasPrescribedRelativeInteriorKernel C D J L}

Lean source

Unique-set packaging of the corrected formal version of Corollary 34.5.1.

open Set variable {m n : ℕ} theorem unique_closed_proper_concaveConvex_equivalenceClass_with_prescribed_relativeInterior_kernel_unique {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hCconv : Convex ℝ C) (hDconv : Convex ℝ D) (hCne : C.Nonempty) (hDne : D.Nonempty) (J : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) (hJcc : IsConcaveConvexOn C D (fun u v => ((J u v : ℝ) : EReal))) : ∃! Ω : Set (SaddleFunction m n), Ω = {L | IsConcaveConvex L ∧ IsClosedSaddleFunction L ∧ IsProperSaddleFunction L ∧ HasPrescribedRelativeInteriorKernel C D J L}

Lean source

Corrected kernel-level form of Text 34.2.1: the class generated by a closed convex bifunction F is exactly the saddle-equivalence class of the Section 34 adjoint kernel (u, x*) ↦ ⟨u, F* x*⟩.

open Set variable {m n : ℕ} lemma section34ConcaveBifunctionAdjointPairing_isClosedSaddleFunction (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsImageClosedConvexBifunction F) (hQ : Section34Theorem34_2Qualification F) : IsClosedSaddleFunction (section34ConcaveBifunctionAdjointPairing F)

Lean source

Corrected kernel-level form of Text 34.2.1: the generated equivalence class of F is the saddleEquivalent-class of the Section 34 adjoint pairing kernel.

open Set variable {m n : ℕ} theorem equivalenceClassGeneratedByConvexBifunction_eq_saddleEquivalenceClass_section34ConcaveBifunctionAdjointPairing (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsImageClosedConvexBifunction F) (hQ : Section34Theorem34_2Qualification F) : EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩ = {K | saddleEquivalent K (section34ConcaveBifunctionAdjointPairing F)}

Lean source

Text 34.2.2: the equivalence class generated by F, denoted Ω(F), is exactly the set of concave-convex saddle-functions K satisfying the sandwich inequalities ⟨F u, x*⟩ ≤ K(u, x*) ≤ ⟨u, F* x*⟩ for all u and x*.

open Set variable {m n : ℕ} theorem equivalenceClassGeneratedByConvexBifunction_eq_sandwichClass (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsImageClosedConvexBifunction F) (hQ : Section34Theorem34_2Qualification F) : EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩ = {K | IsConcaveConvex K ∧ (∀ u xStar, convexBifunctionClosedKernel F u xStar ≤ K u xStar) ∧ (∀ u xStar, K u xStar ≤ section34ConcaveBifunctionAdjointPairing F u xStar)}

Lean source

Membership form of the corrected kernel-level Text 34.2.1: a saddle-function belongs to the class generated by F exactly when it is saddle-equivalent to the adjoint pairing kernel (u, x*) ↦ ⟨u, F* x*⟩.

open Set variable {m n : ℕ} theorem mem_equivalenceClassGeneratedByConvexBifunction_iff_saddleEquivalent_section34ConcaveBifunctionAdjointPairing (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsImageClosedConvexBifunction F) {K : SaddleFunction m n} (hQ : Section34Theorem34_2Qualification F) : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩ ↔ saddleEquivalent K (section34ConcaveBifunctionAdjointPairing F)

Lean source

Membership form of Text 34.2.2: K ∈ Ω(F) exactly when K is concave-convex and lies between ⟨F u, x*⟩ and ⟨u, F* x*⟩ pointwise.

open Set variable {m n : ℕ} theorem mem_equivalenceClassGeneratedByConvexBifunction_iff_sandwich (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsImageClosedConvexBifunction F) {K : SaddleFunction m n} (hQ : Section34Theorem34_2Qualification F) : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩ ↔ IsConcaveConvex K ∧ (∀ u xStar, convexBifunctionClosedKernel F u xStar ≤ K u xStar) ∧ (∀ u xStar, K u xStar ≤ section34ConcaveBifunctionAdjointPairing F u xStar)

Lean source

Concavity-convexity projection of Text 34.2.2: every member of Ω(F) is concave-convex.

open Set variable {m n : ℕ} theorem isConcaveConvex_of_mem_equivalenceClassGeneratedByConvexBifunction (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsImageClosedConvexBifunction F) {K : SaddleFunction m n} (hK : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) (hQ : Section34Theorem34_2Qualification F) : IsConcaveConvex K

Lean source

Lower-bound projection of Text 34.2.2: every member of Ω(F) dominates the canonical kernel ⟨F u, x*⟩.

open Set variable {m n : ℕ} theorem convexBifunctionClosedKernel_le_of_mem_equivalenceClassGeneratedByConvexBifunction (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsImageClosedConvexBifunction F) {K : SaddleFunction m n} (hK : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) (hQ : Section34Theorem34_2Qualification F) : ∀ u xStar, convexBifunctionClosedKernel F u xStar ≤ K u xStar

Lean source

Upper-bound projection of Text 34.2.2: every member of Ω(F) is dominated by the adjoint kernel ⟨u, F* x*⟩.

open Set variable {m n : ℕ} theorem le_section34ConcaveBifunctionAdjointPairing_of_mem_equivalenceClassGeneratedByConvexBifunction (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsImageClosedConvexBifunction F) {K : SaddleFunction m n} (hK : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) (hQ : Section34Theorem34_2Qualification F) : ∀ u xStar, K u xStar ≤ section34ConcaveBifunctionAdjointPairing F u xStar

Lean source

Corollary 34.2.1: if K is a closed saddle-function on ℝ^m × ℝ^n and L is equivalent to K, then dom L = dom K, and L(u, v) = K(u, v) whenever u ∈ ri (dom₁ K) or v ∈ ri (dom₂ K).

open Set variable {m n : ℕ} theorem closed_equivalent_saddle_functions_have_same_domain_and_agree_on_relativeInterior {K L : SaddleFunction m n} (hK : IsClosedSaddleFunction K) (hKL : saddleEquivalent K L) (hGlobal : Section34Theorem34_2GlobalQualification m n) : saddleEffectiveDomain L = saddleEffectiveDomain K ∧ ∀ u v, u ∈ intrinsicInterior ℝ (effectiveDomain₁ K) ∨ v ∈ intrinsicInterior ℝ (effectiveDomain₂ K) → L u v = K u v

Lean source

Domain projection of Corollary 34.2.1: equivalent closed saddle-functions have the same effective domain.

open Set variable {m n : ℕ} theorem closed_equivalent_saddle_functions_have_same_domain {K L : SaddleFunction m n} (hK : IsClosedSaddleFunction K) (hKL : saddleEquivalent K L) (hGlobal : Section34Theorem34_2GlobalQualification m n) : saddleEffectiveDomain L = saddleEffectiveDomain K

Lean source

Relative-interior projection of Corollary 34.2.1: equivalent closed saddle-functions agree at every point whose first or second coordinate lies in the corresponding relative interior of the common effective domain.

open Set variable {m n : ℕ} theorem closed_equivalent_saddle_functions_agree_on_relativeInterior {K L : SaddleFunction m n} (hK : IsClosedSaddleFunction K) (hKL : saddleEquivalent K L) (hGlobal : Section34Theorem34_2GlobalQualification m n) : ∀ u v, u ∈ intrinsicInterior ℝ (effectiveDomain₁ K) ∨ v ∈ intrinsicInterior ℝ (effectiveDomain₂ K) → L u v = K u v

Lean source

Local unconditional first-closure idempotence used by the extremal-member proof.

open Set variable {m n : ℕ} lemma helperForCorollary34_2_2_partialClosure₁_idempotent_unconditional (K : SaddleFunction m n) : partialClosure₁ (partialClosure₁ K) = partialClosure₁ K

Lean source

Local unconditional second-closure idempotence used by the extremal-member proof.

open Set variable {m n : ℕ} lemma helperForCorollary34_2_2_partialClosure₂_idempotent_unconditional (K : SaddleFunction m n) : partialClosure₂ (partialClosure₂ K) = partialClosure₂ K

Lean source

Corollary 34.2.2: a lower closed or upper closed saddle-function is in particular closed. Moreover, each equivalence class of closed saddle-functions contains a unique lower closed function, which is the least member of the class, and a unique upper closed function, which is the greatest member of the class.

open Set variable {m n : ℕ} theorem closed_saddle_equivalence_classes_have_unique_extremal_closed_members (hStrongClosures : Section34Text34_1_9StrongClosureQualification m n) (hClosedNoBot : ∀ K : SaddleFunction m n, IsClosedSaddleFunction K → HasNoBotValuesBifunction K) : (∀ {K : SaddleFunction m n}, IsLowerClosed K → IsClosedSaddleFunction K) ∧ (∀ {K : SaddleFunction m n}, IsUpperClosed K → IsClosedSaddleFunction K) ∧ ∀ {K : SaddleFunction m n}, IsClosedSaddleFunction K → (∃! Klower : SaddleFunction m n, saddleEquivalent Klower K ∧ IsLowerClosed Klower ∧ ∀ L : SaddleFunction m n, saddleEquivalent L K → Klower ≤ L) ∧ ∃! Kupper : SaddleFunction m n, saddleEquivalent Kupper K ∧ IsUpperClosed Kupper ∧ ∀ L : SaddleFunction m n, saddleEquivalent L K → L ≤ Kupper

Lean source

Lower-member projection of Corollary 34.2.2: every closed saddle-function has a unique lower closed representative in its equivalence class, and that representative is the least member of the class in the pointwise order.

open Set variable {m n : ℕ} theorem closed_saddle_equivalenceClass_has_unique_lowerClosed_least_member {K : SaddleFunction m n} (hK : IsClosedSaddleFunction K) (hStrongClosures : Section34Text34_1_9StrongClosureQualification m n) : (∀ L : SaddleFunction m n, IsClosedSaddleFunction L → HasNoBotValuesBifunction L) → ∃! Klower : SaddleFunction m n, saddleEquivalent Klower K ∧ IsLowerClosed Klower ∧ ∀ L : SaddleFunction m n, saddleEquivalent L K → Klower ≤ L

Lean source

Upper-member projection of Corollary 34.2.2: every closed saddle-function has a unique upper closed representative in its equivalence class, and that representative is the greatest member of the class in the pointwise order.

open Set variable {m n : ℕ} theorem closed_saddle_equivalenceClass_has_unique_upperClosed_greatest_member {K : SaddleFunction m n} (hK : IsClosedSaddleFunction K) (hStrongClosures : Section34Text34_1_9StrongClosureQualification m n) : (∀ L : SaddleFunction m n, IsClosedSaddleFunction L → HasNoBotValuesBifunction L) → ∃! Kupper : SaddleFunction m n, saddleEquivalent Kupper K ∧ IsUpperClosed Kupper ∧ ∀ L : SaddleFunction m n, saddleEquivalent L K → L ≤ Kupper

Lean source

Closedness projection of Corollary 34.2.2: every lower closed saddle-function is closed.

open Set variable {m n : ℕ} theorem lowerClosed_saddleFunction_isClosed {K : SaddleFunction m n} (hK : IsLowerClosed K) (hStrongClosures : Section34Text34_1_9StrongClosureQualification m n) : (∀ L : SaddleFunction m n, IsClosedSaddleFunction L → HasNoBotValuesBifunction L) → IsClosedSaddleFunction K

Lean source

Closedness projection of Corollary 34.2.2: every upper closed saddle-function is closed.

open Set variable {m n : ℕ} theorem upperClosed_saddleFunction_isClosed {K : SaddleFunction m n} (hK : IsUpperClosed K) (hStrongClosures : Section34Text34_1_9StrongClosureQualification m n) : (∀ L : SaddleFunction m n, IsClosedSaddleFunction L → HasNoBotValuesBifunction L) → IsClosedSaddleFunction K

Lean source

Corollary 34.2.3: the only improper closed saddle-functions on ℝ^m × ℝ^n are the constant functions +∞ and -∞, and these two constant saddle-functions are not equivalent.

open Set variable {m n : ℕ} theorem closed_improper_saddle_function_eq_top_or_bot (hGlobal : Section34Theorem34_2GlobalQualification m n) : (∀ {K : SaddleFunction m n}, IsClosedSaddleFunction K → (¬ IsProperSaddleFunction K ↔ K = ((fun _ _ => (⊤ : EReal)) : SaddleFunction m n) ∨ K = ((fun _ _ => (⊥ : EReal)) : SaddleFunction m n))) ∧ ¬ saddleEquivalent (((fun _ _ => (⊤ : EReal)) : SaddleFunction m n)) (((fun _ _ => (⊥ : EReal)) : SaddleFunction m n))

Lean source

Properness projection of Corollary 34.2.3: a closed saddle-function is improper exactly when it is the constant ⊤ or the constant ⊥ saddle-function.

open Set variable {m n : ℕ} theorem closed_saddleFunction_not_proper_iff_eq_top_or_bot {K : SaddleFunction m n} (hK : IsClosedSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) : (¬ IsProperSaddleFunction K ↔ K = ((fun _ _ => (⊤ : EReal)) : SaddleFunction m n) ∨ K = ((fun _ _ => (⊥ : EReal)) : SaddleFunction m n))

Lean source

Corrected convex-side form of Text 34.4.2: if F is an image-closed convex bifunction from ℝ^m to ℝ^n, then the saddle-function K(u, x^*) = ⟪F u, x^*⟫ is simple. In the current formalization this is the robust statement supported by the closed-kernel correspondence and the closed proper/improper dichotomy.

open Set variable {m n : ℕ} theorem convex_or_concave_bifunction_pairing_isSimple (hGlobal : Section34Theorem34_2GlobalQualification m n) : ∀ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, IsImageClosedConvexBifunction F → IsSimpleSaddleFunction (convexBifunctionPairing F)

Lean source

Separation projection of Corollary 34.2.3: the constant ⊤ and ⊥ saddle-functions are not saddle-equivalent.

open Set variable {m n : ℕ} theorem top_saddleFunction_not_saddleEquivalent_bot : ¬ saddleEquivalent (((fun _ _ => (⊤ : EReal)) : SaddleFunction m n)) (((fun _ _ => (⊥ : EReal)) : SaddleFunction m n))

Lean source

Helper for Text 34.4.2: the constant ⊤ function is convex in the Jensen sense on ℝ^n.

open Set variable {m n : ℕ} lemma helperForText_34_4_2_constTop_isERealConvexOn {n : ℕ} : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (fun _ : Fin n → ℝ => (⊤ : EReal))

Lean source

Helper for Text 34.4.2: the constant ⊥ saddle-function is simple.

open Set variable {m n : ℕ} lemma helperForText_34_4_2_constBot_isSimpleSaddleFunction : IsSimpleSaddleFunction (((fun _ _ => (⊥ : EReal)) : SaddleFunction m n))

Lean source

Helper for Text 34.4.2: the constant ⊤ saddle-function is simple.

open Set variable {m n : ℕ} lemma helperForText_34_4_2_constTop_isSimpleSaddleFunction : IsSimpleSaddleFunction (((fun _ _ => (⊤ : EReal)) : SaddleFunction m n))

Lean source

Helper for Text 34.4.3: for a convex subset of a finite-dimensional Euclidean space, if the ordinary interior is nonempty, then intrinsic interior points are ordinary interior points.

open Set variable {m n : ℕ} lemma helperForText_34_4_3_mem_interior_of_mem_intrinsicInterior_of_convex_nonemptyInterior {k : ℕ} {C : Set (Fin k → ℝ)} (hCconv : Convex ℝ C) (hCint : (interior C).Nonempty) {u : Fin k → ℝ} (hu : u ∈ intrinsicInterior ℝ C) : u ∈ interior C

Lean source

Helper for Text 34.4.3: from an interior point one can extrapolate past any target point and stay inside the set.

open Set variable {m n : ℕ} lemma helperForText_34_4_3_exists_extrapolatedPoint {k : ℕ} {C : Set (Fin k → ℝ)} {u u' : Fin k → ℝ} (hu : u ∈ interior C) : ∃ s : ℝ, 0 < s ∧ let w := u + s • (u - u') w ∈ C ∧ u = (1 / (1 + s)) • w + (s / (1 + s)) • u'

Lean source

Helper for Text 34.4.3: from a point outside a closed set one can move a short distance along a segment towards an interior point and remain outside the closed set.

open Set variable {m n : ℕ} lemma helperForText_34_4_3_exists_openSegmentPoint {k : ℕ} {C : Set (Fin k → ℝ)} {v v0 : Fin k → ℝ} (hv : v ∈ interior C) : ∃ t : ℝ, 0 < t ∧ t < 1 ∧ ((1 - t) • v + t • v0) ∈ C

Lean source

Helper for Text 34.4.4: the first effective domain of the sign-product example is the singleton {0}.

open Set variable {m n : ℕ} lemma helperForText_34_4_4_effectiveDomain1_eq_singleton_zero : effectiveDomain₁ coordinateProductSignSaddle = ({(fun _ : Fin 1 => (0 : ℝ))} : Set (Fin 1 → ℝ))

Lean source

Helper for Text 34.4.4: the second effective domain of the sign-product example is the singleton {0}.

open Set variable {m n : ℕ} lemma helperForText_34_4_4_effectiveDomain2_eq_singleton_zero : effectiveDomain₂ coordinateProductSignSaddle = ({(fun _ : Fin 1 => (0 : ℝ))} : Set (Fin 1 → ℝ))

Lean source

Helper for Text 34.4.4: the slice through u = 0 has full convex effective domain in the second variable.

open Set variable {m n : ℕ} lemma helperForText_34_4_4_convexSliceAtZero_has_full_effectiveDomain : convexFunctionEffectiveDomain (coordinateProductSignSaddle (fun _ : Fin 1 => (0 : ℝ))) = (Set.univ : Set (Fin 1 → ℝ))

Lean source

Helper for Text 34.4.4: the slice through v = 0 has full convex effective domain in the first variable.

open Set variable {m n : ℕ} lemma helperForText_34_4_4_convexFirstSliceAtZero_has_full_effectiveDomain : convexFunctionEffectiveDomain (fun u => coordinateProductSignSaddle u (fun _ : Fin 1 => (0 : ℝ))) = (Set.univ : Set (Fin 1 → ℝ))

§34 — Closures and Equivalence Classes — Part 18

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Text 34.2.3: if F is a closed convex bifunction and F is proper in the bifunction sense, then its conjugate bifunction F* is proper as well. Consequently, both effective domains are nonempty: dom F ≠ ∅ and dom F* ≠ ∅.

open Set variable {m n : ℕ} theorem closed_convex_bifunction_proper_implies_adjoint_proper (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (hproper : IsProperConvexBifunction F) : IsProperConcaveBifunction (section34ConcaveBifunctionAdjoint F) ∧ (convexBifunctionEffectiveDomain F).Nonempty ∧ (concaveBifunctionEffectiveDomain (section34ConcaveBifunctionAdjoint F)).Nonempty

Lean source

Text 34.2.4: if a closed convex bifunction F is proper, then every saddle-function K in the generated equivalence class Ω(F) is proper.

open Set variable {m n : ℕ} theorem proper_convex_bifunction_has_proper_generated_saddle_functions (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsImageClosedConvexBifunction F) (hproper : IsProperConvexBifunction F) (K : SaddleFunction m n) (hK : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) (hQ : Section34Theorem34_2Qualification F) : IsProperSaddleFunction K

Lean source

Helper for Text 34.2.3: the first effective domain of the canonical Section 34 kernel is exactly the ordinary effective domain dom F of the convex bifunction F.

open Set variable {m n : ℕ} lemma helperForText_34_2_3_firstKernelDomain_eq_convexBifunctionEffectiveDomain (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : effectiveDomain₁ (convexBifunctionClosedKernel F) = convexBifunctionEffectiveDomain F

Lean source

Helper for Text 34.2.3: the second effective domain of the canonical Section 34 kernel is exactly the effective domain dom F* of the Section 34 concave adjoint bifunction.

open Set variable {m n : ℕ} lemma helperForText_34_2_3_secondKernelDomain_eq_concaveAdjointEffectiveDomain (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (hQ : Section34Theorem34_2Qualification F) : effectiveDomain₂ (convexBifunctionClosedKernel F) = concaveBifunctionEffectiveDomain (section34ConcaveBifunctionAdjoint F)

Lean source

Helper for Text 34.2.3: once the saddle effective domain of the canonical Section 34 kernel is nonempty, both projected factor domains are nonempty as well.

open Set variable {m n : ℕ} lemma helperForText_34_2_3_projectedDomainsNonempty_of_kernelDomain (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (saddleEffectiveDomain (convexBifunctionClosedKernel F)).Nonempty → (convexBifunctionDomain F).Nonempty ∧ (convexBifunctionKernelAdjointDomain F).Nonempty

Lean source

Helper for Text 34.2.3: one finite primal witness already forces the Section 34 adjoint bifunction F* to avoid the value ⊤ everywhere.

open Set variable {m n : ℕ} lemma helperForText_34_2_3_adjoint_hasNoTopValues_of_proper (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hproper : IsProperConvexBifunction F) : HasNoTopValuesBifunction (section34ConcaveBifunctionAdjoint F)

Lean source

Helper for Text 34.2.3: a closed proper convex bifunction has a proper packed graph function on ℝ^(m+n).

open Set variable {m n : ℕ} lemma helperForText_34_2_3_graphFunction_properConvex (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (hproper : IsProperConvexBifunction F) : ProperConvexFunctionOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F)

§34 — Closures and Equivalence Classes — Part 19

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

A concave-convex saddle-function is an admissible extension of J from C × D when it agrees with J on C × D, takes the upper-simple-extension value ⊤ on C × Dᶜ, and takes the lower-simple-extension value ⊥ on Cᶜ × D.

open Set variable {m n : ℕ} def IsAdmissibleConcaveConvexExtension (C : Set (Fin m → ℝ)) (D : Set (Fin n → ℝ)) (J : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) (L : SaddleFunction m n) : Prop := IsConcaveConvex L ∧ (∀ u ∈ C, ∀ v ∈ D, L u v = ((J u v : ℝ) : EReal)) ∧ (∀ u ∈ C, ∀ v ∉ D, L u v = (⊤ : EReal)) ∧ ∀ u ∉ C, ∀ v ∈ D, L u v = (⊥ : EReal)

Lean source

The class of all admissible concave-convex extensions of J from C × D to ℝ^m × ℝ^n.

open Set variable {m n : ℕ} def admissibleConcaveConvexExtensionClass (C : Set (Fin m → ℝ)) (D : Set (Fin n → ℝ)) (J : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) : Set (SaddleFunction m n) := {L | IsAdmissibleConcaveConvexExtension C D J L}

Lean source

An admissible extension lies pointwise above the lower simple extension.

open Set variable {m n : ℕ} lemma lowerSimpleExtension_le_of_isAdmissibleConcaveConvexExtension {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {J : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {L : SaddleFunction m n} (hL : IsAdmissibleConcaveConvexExtension C D J L) : let JE : SaddleFunction m n := fun u v => ((J u v : ℝ) : EReal) lowerSimpleExtension C D JE ≤ L

Lean source

An admissible extension lies pointwise below the upper simple extension.

open Set variable {m n : ℕ} lemma le_upperSimpleExtension_of_isAdmissibleConcaveConvexExtension {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {J : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {L : SaddleFunction m n} (hL : IsAdmissibleConcaveConvexExtension C D J L) : let JE : SaddleFunction m n := fun u v => ((J u v : ℝ) : EReal) L ≤ upperSimpleExtension C D JE

Lean source

The admissible-extension side conditions are equivalent to lying between the lower and upper simple extensions of the EReal-valued kernel associated to J.

open Set variable {m n : ℕ} lemma isAdmissibleConcaveConvexExtension_iff_interval {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {J : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {L : SaddleFunction m n} : IsAdmissibleConcaveConvexExtension C D J L ↔ let JE : SaddleFunction m n := fun u v => ((J u v : ℝ) : EReal) IsConcaveConvex L ∧ lowerSimpleExtension C D JE ≤ L ∧ L ≤ upperSimpleExtension C D JE

Lean source

The admissible-extension class is exactly the interval between the lower and upper simple extensions inside the concave-convex functions.

open Set variable {m n : ℕ} lemma admissibleConcaveConvexExtensionClass_eq_interval {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (J : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) : let JE : SaddleFunction m n := fun u v => ((J u v : ℝ) : EReal) admissibleConcaveConvexExtensionClass C D J = {L | IsConcaveConvex L ∧ lowerSimpleExtension C D JE ≤ L ∧ L ≤ upperSimpleExtension C D JE}

Lean source

Corrected formal version of Corollary 34.2.4: if C and D are nonempty closed convex sets and J is a finite continuous concave-convex function on C × D, then the lower simple extension is an admissible closed proper representative of J; it is the least member of the admissible interval class, and every admissible member is pointwise bounded above by the upper simple extension. The stronger theorem-level packaging of the whole admissible class as a closed concave-convex equivalence class is deferred.

open Set variable {m n : ℕ} theorem concaveConvexExtensionClass_is_proper_closed_saddle_equivalenceClass {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hCclosed : IsClosed C) (hDclosed : IsClosed D) (hCconv : Convex ℝ C) (hDconv : Convex ℝ D) (hCne : C.Nonempty) (hDne : D.Nonempty) (hStrongClosures : Section34Text34_1_9StrongClosureQualification m n) (hClosedNoBot : ∀ L : SaddleFunction m n, IsClosedSaddleFunction L → HasNoBotValuesBifunction L) (J : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) (hJcont : ContinuousOn (fun p : (Fin m → ℝ) × (Fin n → ℝ) => J p.1 p.2) (C ×ˢ D)) (hJcc : IsConcaveConvexOn C D (fun u v => ((J u v : ℝ) : EReal))) : let JE : SaddleFunction m n := fun u v => ((J u v : ℝ) : EReal) let Ω : Set (SaddleFunction m n) := admissibleConcaveConvexExtensionClass C D J lowerSimpleExtension C D JE ∈ Ω ∧ (∀ L ∈ Ω, lowerSimpleExtension C D JE ≤ L) ∧ (∀ L ∈ Ω, L ≤ upperSimpleExtension C D JE) ∧ IsProperSaddleFunction (lowerSimpleExtension C D JE) ∧ IsClosedSaddleFunction (lowerSimpleExtension C D JE)

§35 — Continuity and Differentiability of Saddle-Functions — Part 1

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

A convex set that is open in the topology induced on its affine span.

open scoped Pointwise abbrev IsRelativelyOpenConvex {E : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E] (s : Set E) : Prop := Convex ℝ s ∧ IsOpen ((fun x : affineSpan ℝ s => (x : E)) ⁻¹' s)

Lean source

A real-valued function on C × D that is concave in the first variable on C and convex in the second variable on D.

open scoped Pointwise abbrev IsRealConcaveConvexOn {E : Type*} {F : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E] [NormedAddCommGroup F] [NormedSpace ℝ F] (C : Set E) (D : Set F) (K : E → F → ℝ) : Prop := (∀ y ∈ D, ConcaveOn ℝ C (fun x => K x y)) ∧ ∀ x ∈ C, ConvexOn ℝ D (K x)

Lean source

A finite real-valued concave-convex bifunction induces the canonical EReal-valued Section 33 saddle predicate after coercion.

open scoped Pointwise theorem IsRealConcaveConvexOn.toEReal {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hK : IsRealConcaveConvexOn C D K) : IsConcaveConvexOn C D (fun u v => ((K u v : ℝ) : EReal))

Lean source

Theorem 35.1: if C ⊆ ℝ^m and D ⊆ ℝ^n are relatively open convex sets and K is a finite concave-convex function on C × D, then K is continuous relative to C × D; in fact, on every closed bounded subset of C × D, K is Lipschitzian.

open scoped Pointwise theorem section35_theorem35_1 {m n : ℕ} {C : Set (EuclideanSpace ℝ (Fin m))} {D : Set (EuclideanSpace ℝ (Fin n))} {K : EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → ℝ} (hC : IsRelativelyOpenConvex C) (hD : IsRelativelyOpenConvex D) (hK : IsRealConcaveConvexOn C D K) : ContinuousOn (Function.uncurry K) (C ×ˢ D) ∧ ∀ S : Set (EuclideanSpace ℝ (Fin m) × EuclideanSpace ℝ (Fin n)), S ⊆ C ×ˢ D → IsClosed S → Bornology.IsBounded S → ∃ L : NNReal, LipschitzOnWith L (Function.uncurry K) S

Lean source

Helper for Theorem 35.1: a relatively open convex set equals its Euclidean relative interior.

open scoped Pointwise lemma helperForTheorem_35_1_relativelyOpenConvex_eq_ri {k : ℕ} {s : Set (EuclideanSpace ℝ (Fin k))} (hs : IsRelativelyOpenConvex s) : euclideanRelativeInterior k s = s

Lean source

Helper for Theorem 35.1: a convex real-valued function on a relatively open convex set is continuous on that set.

open scoped Pointwise lemma helperForTheorem_35_1_sliceContinuousOn_relativelyOpenConvex {k : ℕ} {s : Set (EuclideanSpace ℝ (Fin k))} {f : EuclideanSpace ℝ (Fin k) → ℝ} (hs : IsRelativelyOpenConvex s) (hf : ConvexOn ℝ s f) : ContinuousOn f s

Lean source

Helper for Theorem 35.1: every affine coordinate model map between finite-dimensional Euclidean spaces is globally Lipschitz.

open scoped Pointwise lemma helperForTheorem_35_1_affineModelMap_lipschitz {k m : ℕ} (A : EuclideanSpace ℝ (Fin k) →ᵃ[ℝ] EuclideanSpace ℝ (Fin m)) : ∃ L : NNReal, LipschitzWith L A

Lean source

Helper for Theorem 35.1: an equi-Lipschitz estimate on the coordinate model pulls back along one Lipschitz parameterization of the original set.

open scoped Pointwise lemma helperForTheorem_35_1_coordinateModel_pullback_equiLipschitz {α β I : Type*} [PseudoMetricSpace α] [PseudoMetricSpace β] {S : Set α} {Smodel : Set β} {ψ : α → β} {f : I → α → ℝ} {g : I → β → ℝ} {L Kψ : NNReal} (hSmodel : Smodel = ψ '' S) (hψ : LipschitzWith Kψ ψ) (hg : ∀ i, LipschitzOnWith L (g i) Smodel) (hEq : ∀ i, Set.EqOn (fun x => g i (ψ x)) (f i) S) : ∀ i, LipschitzOnWith (L * Kψ) (f i) S

Lean source

Helper for Theorem 35.1: the relative-open form of Theorem 10.6 gives a uniform Lipschitz constant on a closed bounded subset.

open scoped Pointwise lemma helperForTheorem_35_1_familyEquiLipschitz_relativelyOpenConvex {k : ℕ} {I : Type*} {s : Set (EuclideanSpace ℝ (Fin k))} {f : I → EuclideanSpace ℝ (Fin k) → ℝ} (hs : IsRelativelyOpenConvex s) (hf : ∀ i, ConvexOn ℝ s (f i)) (hpt : Function.PointwiseBoundedOn f s) {S : Set (EuclideanSpace ℝ (Fin k))} (hSclosed : IsClosed S) (hSbdd : Bornology.IsBounded S) (hSsub : S ⊆ s) : ∃ L : NNReal, ∀ i, LipschitzOnWith L (f i) S

Lean source

Helper for Theorem 35.1: every point of a relatively open convex set admits a closed bounded neighborhood contained in the set.

open scoped Pointwise lemma helperForTheorem_35_1_existsClosedBoundedNeighborhood_subset_relativelyOpenConvex {k : ℕ} {s : Set (EuclideanSpace ℝ (Fin k))} (hs : IsRelativelyOpenConvex s) {x : EuclideanSpace ℝ (Fin k)} (hx : x ∈ s) : ∃ T : Set (EuclideanSpace ℝ (Fin k)), x ∈ T ∧ T ⊆ s ∧ IsClosed T ∧ Bornology.IsBounded T ∧ T ∈ nhdsWithin x s

Lean source

Helper for Theorem 35.1: a continuous real-valued function has bounded range on a compact set, written over the corresponding subtype.

open scoped Pointwise lemma helperForTheorem_35_1_boundedRange_of_continuousOn_compact {α : Type*} [TopologicalSpace α] {s : Set α} (hs : IsCompact s) {g : α → ℝ} (hg : ContinuousOn g s) : Bornology.IsBounded (Set.range fun x : s => g x.1)

Lean source

Helper for Theorem 35.1: continuity of the slice families over compact projections yields the pointwise boundedness needed for the one-variable equi-Lipschitz estimates.

open scoped Pointwise lemma helperForTheorem_35_1_projectionFamilies_pointwiseBounded {m n : ℕ} {X : Set (EuclideanSpace ℝ (Fin m))} {Y : Set (EuclideanSpace ℝ (Fin n))} {K : EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → ℝ} (hXcomp : IsCompact X) (hYcomp : IsCompact Y) (hcontX : ∀ y ∈ Y, ContinuousOn (fun x => K x y) X) (hcontY : ∀ x ∈ X, ContinuousOn (K x) Y) : (∀ y ∈ Y, Bornology.IsBounded (Set.range fun x : X => K x.1 y)) ∧ (∀ x ∈ X, Bornology.IsBounded (Set.range fun y : Y => K x y.1))

Lean source

Helper for Theorem 35.1: coordinatewise Lipschitz estimates assemble into a product Lipschitz estimate on X × Y.

open scoped Pointwise lemma helperForTheorem_35_1_productLipschitz_of_coordinatewiseBounds {m n : ℕ} {X : Set (EuclideanSpace ℝ (Fin m))} {Y : Set (EuclideanSpace ℝ (Fin n))} {K : EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → ℝ} {Lx Ly : NNReal} (hX : ∀ y ∈ Y, LipschitzOnWith Lx (fun x => K x y) X) (hY : ∀ x ∈ X, LipschitzOnWith Ly (K x) Y) : LipschitzOnWith (Lx + Ly) (Function.uncurry K) (X ×ˢ Y)

Lean source

Helper for Theorem 35.2: the subset-witness form of Theorem 10.6 extends from open convex domains to relatively open convex domains.

open scoped Pointwise lemma helperForTheorem_35_2_relativelyOpenConvex_existsSubset_uniformlyBounded_and_equiLipschitz {k : ℕ} {I : Type*} {s : Set (EuclideanSpace ℝ (Fin k))} {f : I → EuclideanSpace ℝ (Fin k) → ℝ} (hs : IsRelativelyOpenConvex s) (hf : ∀ i, ConvexOn ℝ s (f i)) {s' : Set (EuclideanSpace ℝ (Fin k))} (hs'sub : s' ⊆ s) (hs'hull : s ⊆ convexHull ℝ (closure s')) (hs'bdAbove : ∀ x ∈ s', BddAbove (Set.range fun i : I => f i x)) (hexists_bddBelow : ∃ x ∈ s, BddBelow (Set.range fun i : I => f i x)) {S : Set (EuclideanSpace ℝ (Fin k))} (hSclosed : IsClosed S) (hSbdd : Bornology.IsBounded S) (hSsub : S ⊆ s) : Function.UniformlyBoundedOn f S ∧ Function.EquiLipschitzRelativeTo f S

§35 — Continuity and Differentiability of Saddle-Functions — Part 2

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

A family of real-valued functions is pointwise bounded on s if the image set {f i x | i ∈ I} is bounded for every x ∈ s.

open scoped Pointwise -- Proof sketch: project the product witness through the coordinate projections and use -- nonemptiness of the opposite factor to realize each point as a projected point of the product. def Function.PointwiseBoundedFamilyOn {α I : Type*} (f : I → α → ℝ) (s : Set α) : Prop := ∀ x ∈ s, Bornology.IsBounded (Set.range fun i : I => f i x)

Lean source

A family of real-valued functions is uniformly bounded on s if all values lie between a common lower and upper bound on s.

open scoped Pointwise -- Proof sketch: project the product witness through the coordinate projections and use -- nonemptiness of the opposite factor to realize each point as a projected point of the product. def Function.UniformlyBoundedFamilyOn {α I : Type*} (f : I → α → ℝ) (s : Set α) : Prop := ∃ α₁ α₂ : ℝ, ∀ i x, x ∈ s → α₁ ≤ f i x ∧ f i x ≤ α₂

Lean source

A family of real-valued functions is equi-Lipschitzian on s if one Lipschitz constant works simultaneously for all members of the family on s.

open scoped Pointwise -- Proof sketch: project the product witness through the coordinate projections and use -- nonemptiness of the opposite factor to realize each point as a projected point of the product. def Function.EquiLipschitzFamilyOn {α I : Type*} [PseudoMetricSpace α] (f : I → α → ℝ) (s : Set α) : Prop := ∃ L : NNReal, ∀ i : I, LipschitzOnWith L (f i) s -- Proof sketch: a bounded subset of `ℝ` lies in a closed ball, so the ball radius gives a -- uniform upper bound on every element of the set.

Lean source

Theorem 35.2: let C ⊆ ℝ^m and D ⊆ ℝ^n be relatively open convex sets, and let {K i | i ∈ I} be a collection of finite concave-convex functions on C × D. If there exist subsets C' ⊆ C and D' ⊆ D such that C × D ⊆ convexHull ℝ (closure (C' ×ˢ D')) and the family is pointwise bounded on C' × D', then on every closed bounded subset of C × D the family is uniformly bounded and equi-Lipschitzian.

open scoped Pointwise -- Proof sketch: project the product witness through the coordinate projections and use -- nonemptiness of the opposite factor to realize each point as a projected point of the product. theorem section35_theorem35_2 {I : Type*} {m n : ℕ} {C : Set (EuclideanSpace ℝ (Fin m))} {D : Set (EuclideanSpace ℝ (Fin n))} {K : I → EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → ℝ} (hC : IsRelativelyOpenConvex C) (hD : IsRelativelyOpenConvex D) (hK : ∀ i, IsRealConcaveConvexOn C D (K i)) (hWitness : ∃ C' : Set (EuclideanSpace ℝ (Fin m)), ∃ D' : Set (EuclideanSpace ℝ (Fin n)), C' ⊆ C ∧ D' ⊆ D ∧ C ×ˢ D ⊆ convexHull ℝ (closure (C' ×ˢ D')) ∧ Function.PointwiseBoundedFamilyOn (fun i => Function.uncurry (K i)) (C' ×ˢ D')) : ∀ S : Set (EuclideanSpace ℝ (Fin m) × EuclideanSpace ℝ (Fin n)), S ⊆ C ×ˢ D → IsClosed S → Bornology.IsBounded S → Function.UniformlyBoundedFamilyOn (fun i => Function.uncurry (K i)) S ∧ Function.EquiLipschitzFamilyOn (fun i => Function.uncurry (K i)) S

Lean source

Theorem 35.4: if C ⊆ ℝ^m and D ⊆ ℝ^n are relatively open convex sets and K₁, K₂, ... is a sequence of finite concave-convex functions on C × D whose pointwise limit exists on a dense product subset C' × D', then there is a finite concave-convex limit function K on C × D such that K i converges pointwise to K on C × D and uniformly on every closed bounded subset of C × D.

open scoped Pointwise -- Proof sketch: project the product witness through the coordinate projections and use -- nonemptiness of the opposite factor to realize each point as a projected point of the product. theorem section35_theorem35_4 {m n : ℕ} {C : Set (EuclideanSpace ℝ (Fin m))} {D : Set (EuclideanSpace ℝ (Fin n))} {KSeq : ℕ → EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → ℝ} (hC : IsRelativelyOpenConvex C) (hD : IsRelativelyOpenConvex D) (hKSeq : ∀ i, IsRealConcaveConvexOn C D (KSeq i)) (hDense : ∃ C' : Set (EuclideanSpace ℝ (Fin m)), ∃ D' : Set (EuclideanSpace ℝ (Fin n)), C' ⊆ C ∧ D' ⊆ D ∧ C ⊆ closure C' ∧ D ⊆ closure D' ∧ ∀ u ∈ C', ∀ v ∈ D', ∃ l : ℝ, Filter.Tendsto (fun i : ℕ => KSeq i u v) Filter.atTop (nhds l)) : ∃ K : EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → ℝ, IsRealConcaveConvexOn C D K ∧ (∀ u ∈ C, ∀ v ∈ D, Filter.Tendsto (fun i : ℕ => KSeq i u v) Filter.atTop (nhds (K u v))) ∧ ∀ S : Set (EuclideanSpace ℝ (Fin m) × EuclideanSpace ℝ (Fin n)), S ⊆ C ×ˢ D → IsClosed S → Bornology.IsBounded S → TendstoUniformlyOn (fun i p => Function.uncurry (KSeq i) p) (Function.uncurry K) Filter.atTop S

Lean source

Helper for Theorem 35.2: a convex-hull witness on C' × D' projects to convex-hull witnesses on each factor once C and D are nonempty.

open scoped Pointwise -- Proof sketch: project the product witness through the coordinate projections and use -- nonemptiness of the opposite factor to realize each point as a projected point of the product. lemma helperForTheorem_35_2_projectedHull_of_productHull {m n : ℕ} {C : Set (EuclideanSpace ℝ (Fin m))} {D : Set (EuclideanSpace ℝ (Fin n))} {C' : Set (EuclideanSpace ℝ (Fin m))} {D' : Set (EuclideanSpace ℝ (Fin n))} (hHull : C ×ˢ D ⊆ convexHull ℝ (closure (C' ×ˢ D'))) (hCne : C.Nonempty) (hDne : D.Nonempty) : C ⊆ convexHull ℝ (closure C') ∧ D ⊆ convexHull ℝ (closure D')

Lean source

Helper for Theorem 35.2: a bounded set of real numbers is bounded above.

open scoped Pointwise -- Proof sketch: project the product witness through the coordinate projections and use -- nonemptiness of the opposite factor to realize each point as a projected point of the product. lemma helperForTheorem_35_2_bddAbove_of_isBounded {A : Set ℝ} (hA : Bornology.IsBounded A) : BddAbove A

Lean source

Helper for Theorem 35.2: a bounded set of real numbers is bounded below.

open scoped Pointwise -- Proof sketch: project the product witness through the coordinate projections and use -- nonemptiness of the opposite factor to realize each point as a projected point of the product. lemma helperForTheorem_35_2_bddBelow_of_isBounded {A : Set ℝ} (hA : Bornology.IsBounded A) : BddBelow A

Lean source

Helper for Theorem 35.2: for each witness point x ∈ C', the y-slice family is uniformly bounded on every closed bounded Y ⊆ D.

open scoped Pointwise -- Proof sketch: project the product witness through the coordinate projections and use -- nonemptiness of the opposite factor to realize each point as a projected point of the product. lemma helperForTheorem_35_2_uniformlyBounded_ySlices_of_mem_Cprime {I : Type*} {m n : ℕ} {C : Set (EuclideanSpace ℝ (Fin m))} {D : Set (EuclideanSpace ℝ (Fin n))} {C' : Set (EuclideanSpace ℝ (Fin m))} {D' : Set (EuclideanSpace ℝ (Fin n))} {K : I → EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → ℝ} (hD : IsRelativelyOpenConvex D) (hK : ∀ i, IsRealConcaveConvexOn C D (K i)) (hC'sub : C' ⊆ C) (hD'sub : D' ⊆ D) (hDhull : D ⊆ convexHull ℝ (closure D')) (hPointwise : ∀ x ∈ C', ∀ y ∈ D', Bornology.IsBounded (Set.range fun i : I => K i x y)) (hD'ne : D'.Nonempty) {Y : Set (EuclideanSpace ℝ (Fin n))} (hYclosed : IsClosed Y) (hYbdd : Bornology.IsBounded Y) (hYsub : Y ⊆ D) {x : EuclideanSpace ℝ (Fin m)} (hx : x ∈ C') : Function.UniformlyBoundedOn (fun i y => K i x y) Y

Lean source

Helper for Theorem 35.2: for each witness point y ∈ D', the x-slice family is uniformly bounded on every closed bounded X ⊆ C.

open scoped Pointwise -- Proof sketch: project the product witness through the coordinate projections and use -- nonemptiness of the opposite factor to realize each point as a projected point of the product. lemma helperForTheorem_35_2_uniformlyBounded_xSlices_of_mem_Dprime {I : Type*} {m n : ℕ} {C : Set (EuclideanSpace ℝ (Fin m))} {D : Set (EuclideanSpace ℝ (Fin n))} {C' : Set (EuclideanSpace ℝ (Fin m))} {D' : Set (EuclideanSpace ℝ (Fin n))} {K : I → EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → ℝ} (hC : IsRelativelyOpenConvex C) (hK : ∀ i, IsRealConcaveConvexOn C D (K i)) (hC'sub : C' ⊆ C) (hD'sub : D' ⊆ D) (hChull : C ⊆ convexHull ℝ (closure C')) (hPointwise : ∀ x ∈ C', ∀ y ∈ D', Bornology.IsBounded (Set.range fun i : I => K i x y)) (hC'ne : C'.Nonempty) {X : Set (EuclideanSpace ℝ (Fin m))} (hXclosed : IsClosed X) (hXbdd : Bornology.IsBounded X) (hXsub : X ⊆ C) {y : EuclideanSpace ℝ (Fin n)} (hy : y ∈ D') : Function.UniformlyBoundedOn (fun i x => K i x y) X

Lean source

Helper for Theorem 35.2: uniform coordinatewise Lipschitz bounds yield a uniform product Lipschitz bound for each kernel in the family.

open scoped Pointwise -- Proof sketch: project the product witness through the coordinate projections and use -- nonemptiness of the opposite factor to realize each point as a projected point of the product. lemma helperForTheorem_35_2_familyProductLipschitz_of_coordinatewiseBounds {I : Type*} {m n : ℕ} {X : Set (EuclideanSpace ℝ (Fin m))} {Y : Set (EuclideanSpace ℝ (Fin n))} {K : I → EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → ℝ} {Lx Ly : NNReal} (hX : ∀ (i : I) (y : EuclideanSpace ℝ (Fin n)), y ∈ Y → LipschitzOnWith Lx (fun x => K i x y) X) (hY : ∀ (i : I) (x : EuclideanSpace ℝ (Fin m)), x ∈ X → LipschitzOnWith Ly (fun y => K i x y) Y) : ∀ i, LipschitzOnWith (Lx + Ly) (Function.uncurry (K i)) (X ×ˢ Y)

Lean source

Helper for Theorem 35.4: pointwise convergence on the dense witness product gives the pointwise boundedness package required by Theorem 35.2.

open scoped Pointwise -- Proof sketch: project the product witness through the coordinate projections and use -- nonemptiness of the opposite factor to realize each point as a projected point of the product. lemma helperForTheorem_35_4_pointwiseBounded_denseProduct_of_tendsto {m n : ℕ} {C' : Set (EuclideanSpace ℝ (Fin m))} {D' : Set (EuclideanSpace ℝ (Fin n))} {KSeq : ℕ → EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → ℝ} (hTendsto : ∀ u ∈ C', ∀ v ∈ D', ∃ l : ℝ, Filter.Tendsto (fun i : ℕ => KSeq i u v) Filter.atTop (nhds l)) : Function.PointwiseBoundedFamilyOn (fun i => Function.uncurry (KSeq i)) (C' ×ˢ D')

Lean source

Helper for Theorem 35.4: density in each factor gives the product convex-hull witness needed for Theorem 35.2.

open scoped Pointwise -- Proof sketch: project the product witness through the coordinate projections and use -- nonemptiness of the opposite factor to realize each point as a projected point of the product. lemma helperForTheorem_35_4_productHull_of_factorClosures {m n : ℕ} {C : Set (EuclideanSpace ℝ (Fin m))} {D : Set (EuclideanSpace ℝ (Fin n))} {C' : Set (EuclideanSpace ℝ (Fin m))} {D' : Set (EuclideanSpace ℝ (Fin n))} (hCclosure : C ⊆ closure C') (hDclosure : D ⊆ closure D') : C ×ˢ D ⊆ convexHull ℝ (closure (C' ×ˢ D'))

Lean source

Helper for Theorem 35.4: pointwise limits of finite concave-convex kernels remain concave-convex.

open scoped Pointwise -- Proof sketch: project the product witness through the coordinate projections and use -- nonemptiness of the opposite factor to realize each point as a projected point of the product. lemma helperForTheorem_35_4_concaveConvex_limit_of_pointwiseTendsto {m n : ℕ} {C : Set (EuclideanSpace ℝ (Fin m))} {D : Set (EuclideanSpace ℝ (Fin n))} {KSeq : ℕ → EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → ℝ} {K : EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → ℝ} (hCconv : Convex ℝ C) (hDconv : Convex ℝ D) (hKSeq : ∀ i, IsRealConcaveConvexOn C D (KSeq i)) (hTendsto : ∀ u ∈ C, ∀ v ∈ D, Filter.Tendsto (fun i : ℕ => KSeq i u v) Filter.atTop (nhds (K u v))) : IsRealConcaveConvexOn C D K

Lean source

Helper for Theorem 35.4: dense-product convergence extends from C' × D' to any point of C × D by combining one local equi-Lipschitz estimate with Cauchy convergence at a nearby witness point.

open scoped Pointwise -- Proof sketch: project the product witness through the coordinate projections and use -- nonemptiness of the opposite factor to realize each point as a projected point of the product. lemma helperForTheorem_35_4_pointwiseTendsto_of_denseWitness {m n : ℕ} {C : Set (EuclideanSpace ℝ (Fin m))} {D : Set (EuclideanSpace ℝ (Fin n))} {C' : Set (EuclideanSpace ℝ (Fin m))} {D' : Set (EuclideanSpace ℝ (Fin n))} {KSeq : ℕ → EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → ℝ} (hC : IsRelativelyOpenConvex C) (hD : IsRelativelyOpenConvex D) (hKSeq : ∀ i, IsRealConcaveConvexOn C D (KSeq i)) (hC'sub : C' ⊆ C) (hD'sub : D' ⊆ D) (hCclosure : C ⊆ closure C') (hDclosure : D ⊆ closure D') (hDenseTendsto : ∀ u ∈ C', ∀ v ∈ D', ∃ l : ℝ, Filter.Tendsto (fun i : ℕ => KSeq i u v) Filter.atTop (nhds l)) {u0 : EuclideanSpace ℝ (Fin m)} (hu0 : u0 ∈ C) {v0 : EuclideanSpace ℝ (Fin n)} (hv0 : v0 ∈ D) : ∃ l : ℝ, Filter.Tendsto (fun i : ℕ => KSeq i u0 v0) Filter.atTop (nhds l)

Lean source

Helper for Theorem 35.4: on a closed bounded set, pointwise convergence plus a common Lipschitz constant upgrades to uniform convergence.

open scoped Pointwise -- Proof sketch: project the product witness through the coordinate projections and use -- nonemptiness of the opposite factor to realize each point as a projected point of the product. lemma helperForTheorem_35_4_tendstoUniformlyOn_of_pointwiseTendsto_and_equiLipschitz {m n : ℕ} {S : Set (EuclideanSpace ℝ (Fin m) × EuclideanSpace ℝ (Fin n))} {f : ℕ → (EuclideanSpace ℝ (Fin m) × EuclideanSpace ℝ (Fin n)) → ℝ} {g : (EuclideanSpace ℝ (Fin m) × EuclideanSpace ℝ (Fin n)) → ℝ} (hSclosed : IsClosed S) (hSbdd : Bornology.IsBounded S) (hPoint : ∀ p ∈ S, Filter.Tendsto (fun i : ℕ => f i p) Filter.atTop (nhds (g p))) (hEqui : Function.EquiLipschitzFamilyOn f S) : TendstoUniformlyOn f g Filter.atTop S

Lean source

Helper for Theorem 35.3: density in each factor yields the product convex-hull witness needed for Theorem 35.2.

open scoped Pointwise -- Proof sketch: project the product witness through the coordinate projections and use -- nonemptiness of the opposite factor to realize each point as a projected point of the product. lemma helperForTheorem_35_3_productHull_of_factorClosures {m n : ℕ} {C : Set (EuclideanSpace ℝ (Fin m))} {D : Set (EuclideanSpace ℝ (Fin n))} {C' : Set (EuclideanSpace ℝ (Fin m))} {D' : Set (EuclideanSpace ℝ (Fin n))} (hCclosure : C ⊆ closure C') (hDclosure : D ⊆ closure D') : C ×ˢ D ⊆ convexHull ℝ (closure (C' ×ˢ D'))

Lean source

Helper for Theorem 35.3: on a compact set of parameters t, continuity in t on a witness product gives a uniformly bounded and equi-Lipschitz family on each closed bounded subset of C × D.

open scoped Pointwise -- Proof sketch: project the product witness through the coordinate projections and use -- nonemptiness of the opposite factor to realize each point as a projected point of the product. lemma helperForTheorem_35_3_compactIndexed_family_uniformlyBoundedAndEquiLipschitz {m n : ℕ} {T : Type*} [TopologicalSpace T] {C : Set (EuclideanSpace ℝ (Fin m))} {D : Set (EuclideanSpace ℝ (Fin n))} {Cw : Set (EuclideanSpace ℝ (Fin m))} {Dw : Set (EuclideanSpace ℝ (Fin n))} {K : EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → T → ℝ} (hC : IsRelativelyOpenConvex C) (hD : IsRelativelyOpenConvex D) (hK : ∀ t, IsRealConcaveConvexOn C D (fun u v => K u v t)) {K0 : Set T} (hK0comp : IsCompact K0) (hCwsub : Cw ⊆ C) (hDwsub : Dw ⊆ D) (hHull : C ×ˢ D ⊆ convexHull ℝ (closure (Cw ×ˢ Dw))) (hCont : ∀ u ∈ Cw, ∀ v ∈ Dw, Continuous fun t => K u v t) {S : Set (EuclideanSpace ℝ (Fin m) × EuclideanSpace ℝ (Fin n))} (hSsub : S ⊆ C ×ˢ D) (hSclosed : IsClosed S) (hSbdd : Bornology.IsBounded S) : Function.UniformlyBoundedFamilyOn (fun τ : {t : T // t ∈ K0} => Function.uncurry (fun u v => K u v τ.1)) S ∧ Function.EquiLipschitzFamilyOn (fun τ : {t : T // t ∈ K0} => Function.uncurry (fun u v => K u v τ.1)) S

Lean source

Helper for Theorem 35.3: continuity on the left-associated product ((u,v), t) transfers to the textbook right-associated product u, (v, t).

open scoped Pointwise -- Proof sketch: project the product witness through the coordinate projections and use -- nonemptiness of the opposite factor to realize each point as a projected point of the product. lemma helperForTheorem_35_3_rightAssociatedContinuousOn {m n : ℕ} {T : Type*} [TopologicalSpace T] {C : Set (EuclideanSpace ℝ (Fin m))} {D : Set (EuclideanSpace ℝ (Fin n))} {K : EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → T → ℝ} (hCont : ContinuousOn (fun p : (EuclideanSpace ℝ (Fin m) × EuclideanSpace ℝ (Fin n)) × T => K p.1.1 p.1.2 p.2) ((C ×ˢ D) ×ˢ (Set.univ : Set T))) : ContinuousOn (fun p : EuclideanSpace ℝ (Fin m) × EuclideanSpace ℝ (Fin n) × T => K p.1 p.2.1 p.2.2) (C ×ˢ (D ×ˢ (Set.univ : Set T)))

§35 — Continuity and Differentiability of Saddle-Functions — Part 3

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Helper for Theorem 35.3: full continuity in t on C × D implies joint continuity on C × D × T.

open scoped Pointwise lemma helperForTheorem_35_3_fullContinuityImplication {m n : ℕ} {T : Type*} [TopologicalSpace T] [LocallyCompactSpace T] {C : Set (EuclideanSpace ℝ (Fin m))} {D : Set (EuclideanSpace ℝ (Fin n))} {K : EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → T → ℝ} (hC : IsRelativelyOpenConvex C) (hD : IsRelativelyOpenConvex D) (hK : ∀ t, IsRealConcaveConvexOn C D (fun u v => K u v t)) (hCont : ∀ u ∈ C, ∀ v ∈ D, Continuous fun t => K u v t) : ContinuousOn (fun p : EuclideanSpace ℝ (Fin m) × EuclideanSpace ℝ (Fin n) × T => K p.1 p.2.1 p.2.2) (C ×ˢ (D ×ˢ (Set.univ : Set T)))

Lean source

Helper for Theorem 35.3: continuity in t on a dense witness product still implies joint continuity on C × D × T.

open scoped Pointwise lemma helperForTheorem_35_3_denseContinuityImplication {m n : ℕ} {T : Type*} [TopologicalSpace T] [LocallyCompactSpace T] {C : Set (EuclideanSpace ℝ (Fin m))} {D : Set (EuclideanSpace ℝ (Fin n))} {C' : Set (EuclideanSpace ℝ (Fin m))} {D' : Set (EuclideanSpace ℝ (Fin n))} {K : EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → T → ℝ} (hC : IsRelativelyOpenConvex C) (hD : IsRelativelyOpenConvex D) (hK : ∀ t, IsRealConcaveConvexOn C D (fun u v => K u v t)) (hC'sub : C' ⊆ C) (hD'sub : D' ⊆ D) (hCclosure : C ⊆ closure C') (hDclosure : D ⊆ closure D') (hCont : ∀ u ∈ C', ∀ v ∈ D', Continuous fun t => K u v t) : ContinuousOn (fun p : EuclideanSpace ℝ (Fin m) × EuclideanSpace ℝ (Fin n) × T => K p.1 p.2.1 p.2.2) (C ×ˢ (D ×ˢ (Set.univ : Set T)))

§35 — Continuity and Differentiability of Saddle-Functions — Part 4

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Theorem 35.3: let C ⊆ ℝ^m and D ⊆ ℝ^n be relatively open convex sets, let T be a locally compact topological space, and let K : C × D × T → ℝ be concave in u, convex in v, and continuous in t for each fixed (u, v). Then K is jointly continuous on C × D × T. The same conclusion holds if continuity in t is assumed only on a dense subset C' × D' ⊆ C × D.

open scoped Pointwise theorem section35_theorem35_3 {m n : ℕ} {T : Type*} [TopologicalSpace T] [LocallyCompactSpace T] {C : Set (EuclideanSpace ℝ (Fin m))} {D : Set (EuclideanSpace ℝ (Fin n))} {K : EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → T → ℝ} (hC : IsRelativelyOpenConvex C) (hD : IsRelativelyOpenConvex D) (hK : ∀ t, IsRealConcaveConvexOn C D (fun u v => K u v t)) : ((∀ u ∈ C, ∀ v ∈ D, Continuous fun t => K u v t) → ContinuousOn (fun p : EuclideanSpace ℝ (Fin m) × EuclideanSpace ℝ (Fin n) × T => K p.1 p.2.1 p.2.2) (C ×ˢ (D ×ˢ (Set.univ : Set T)))) ∧ ((∃ C' : Set (EuclideanSpace ℝ (Fin m)), ∃ D' : Set (EuclideanSpace ℝ (Fin n)), C' ⊆ C ∧ D' ⊆ D ∧ C ⊆ closure C' ∧ D ⊆ closure D' ∧ ∀ u ∈ C', ∀ v ∈ D', Continuous fun t => K u v t) → ContinuousOn (fun p : EuclideanSpace ℝ (Fin m) × EuclideanSpace ℝ (Fin n) × T => K p.1 p.2.1 p.2.2) (C ×ˢ (D ×ˢ (Set.univ : Set T))))

Lean source

Theorem 35.5: let C ⊆ ℝ^m and D ⊆ ℝ^n be relatively open convex sets, and let K₁, K₂, ... be a sequence of finite concave-convex functions on C × D. If there exist dense subsets C' ⊆ C and D' ⊆ D such that the sequence K i u v is bounded for every (u, v) ∈ C' × D', then some subsequence converges uniformly on every closed bounded subset of C × D to a finite concave-convex function K.

open scoped Pointwise theorem section35_theorem35_5 {m n : ℕ} {C : Set (EuclideanSpace ℝ (Fin m))} {D : Set (EuclideanSpace ℝ (Fin n))} {KSeq : ℕ → EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → ℝ} (hC : IsRelativelyOpenConvex C) (hD : IsRelativelyOpenConvex D) (hKSeq : ∀ i, IsRealConcaveConvexOn C D (KSeq i)) (hDense : ∃ C' : Set (EuclideanSpace ℝ (Fin m)), ∃ D' : Set (EuclideanSpace ℝ (Fin n)), C' ⊆ C ∧ D' ⊆ D ∧ C ⊆ closure C' ∧ D ⊆ closure D' ∧ Function.PointwiseBoundedFamilyOn (fun i => Function.uncurry (KSeq i)) (C' ×ˢ D')) : ∃ φ : ℕ → ℕ, StrictMono φ ∧ ∃ K : EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → ℝ, IsRealConcaveConvexOn C D K ∧ ∀ S : Set (EuclideanSpace ℝ (Fin m) × EuclideanSpace ℝ (Fin n)), S ⊆ C ×ˢ D → IsClosed S → Bornology.IsBounded S → TendstoUniformlyOn (fun i p => Function.uncurry (KSeq (φ i)) p) (Function.uncurry K) Filter.atTop S

Lean source

An extended-real saddle function is concave in its first variable and convex in its second variable on ℝ × ℝ.

open scoped Pointwise abbrev IsERealSaddleFunction (K : ℝ → ℝ → EReal) : Prop := (∀ v : ℝ, ConvexFunction (fun u : Fin 1 → ℝ => -K (u 0) v)) ∧ ∀ u : ℝ, ConvexFunction (fun v : Fin 1 → ℝ => K u (v 0))

Lean source

The directional difference quotient of an extended-real saddle function at (u, v) in the direction (u', v').

open scoped Pointwise noncomputable def saddleDirectionalDifferenceQuotientAt {E : Type*} {F : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E] [NormedAddCommGroup F] [NormedSpace ℝ F] (K : E → F → EReal) (u : E) (v : F) (u' : E) (v' : F) (t : ℝ) : EReal := (K (u + t • u') (v + t • v') - K u v) / (t : EReal)

Lean source

Text 35.5.1: if K is a saddle function on ℝ^m × ℝ^n and is finite at (u, v), then L is the one-sided directional derivative of K at (u, v) with respect to (u', v') when the directional difference quotient tends to L as λ ↓ 0.

open scoped Pointwise def IsSaddleDirectionalDerivativeAt {E : Type*} {F : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E] [NormedAddCommGroup F] [NormedSpace ℝ F] (K : E → F → EReal) (u : E) (v : F) (u' : E) (v' : F) (L : EReal) : Prop := K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal) ∧ Filter.Tendsto (saddleDirectionalDifferenceQuotientAt K u v u' v') (nhdsWithin (0 : ℝ) (Set.Ioi (0 : ℝ))) (nhds L)

Lean source

The effective domain of an extended-real saddle function consists of the points where the value is finite.

open scoped Pointwise def saddleFunctionEffectiveDomain {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Set ((Fin m → ℝ) × (Fin n → ℝ)) := {p | K p.1 p.2 ≠ (⊤ : EReal) ∧ K p.1 p.2 ≠ (⊥ : EReal)} -- Proof sketch: for fixed `v`, the slice `u ↦ -K u v` is convex because `K` is saddle. Apply -- Theorem 23.1 to this convex slice at the finite point `u`, then negate the resulting limit to -- recover existence of the right directional derivative of `K` in the first variable.

Lean source

Text 35.5.2: if K is a saddle function on ℝ^m × ℝ^n and K u v is finite, then for every direction u' the one-sided directional derivative K'(u, v; u', 0) exists, i.e. the limit of the quotients ((K (u + λ • u') v - K u v) / λ) as λ ↓ 0 exists.

open scoped Pointwise theorem section35_text35_5_2 {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : (∀ v : Fin n → ℝ, ConvexFunction (fun u : Fin m → ℝ => -K u v)) ∧ ∀ u : Fin m → ℝ, ConvexFunction (K u)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (u' : Fin m → ℝ) : ∃ L : EReal, Filter.Tendsto (fun t : ℝ => (K (u + t • u') v - K u v) / (t : EReal)) (nhdsWithin (0 : ℝ) (Set.Ioi (0 : ℝ))) (nhds L)

Lean source

Text 35.5.3: if K is a saddle function on ℝ^m × ℝ^n and K u v is finite, then for every direction v' the one-sided directional derivative K'(u, v; 0, v') exists, i.e. the limit of the quotients ((K u (v + λ • v') - K u v) / λ) as λ ↓ 0 exists.

open scoped Pointwise theorem section35_text35_5_3 {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : (∀ v : Fin n → ℝ, ConvexFunction (fun u : Fin m → ℝ => -K u v)) ∧ ∀ u : Fin m → ℝ, ConvexFunction (K u)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (v' : Fin n → ℝ) : ∃ L : EReal, Filter.Tendsto (fun t : ℝ => (K u (v + t • v') - K u v) / (t : EReal)) (nhdsWithin (0 : ℝ) (Set.Ioi (0 : ℝ))) (nhds L)

Lean source

Text 35.5.5: if (u, v) lies in the interior of the effective domain of K, then for every direction (u', v') there exists δ > 0 such that (u + λ • u', v + λ • v') remains in the effective domain of K for all λ ∈ [0, δ). In particular, the directional difference quotient of K at (u, v) along (u', v') is well-defined for all sufficiently small λ > 0.

open scoped Pointwise theorem section35_text35_5_5 {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hInterior : (u, v) ∈ interior (saddleFunctionEffectiveDomain K)) (u' : Fin m → ℝ) (v' : Fin n → ℝ) : ∃ δ > 0, ∀ t ∈ Set.Ico (0 : ℝ) δ, (u + t • u', v + t • v') ∈ saddleFunctionEffectiveDomain K

Lean source

Text 35.5.6: unless explicitly stated otherwise, the directional derivatives of K considered in this section are taken at base points in the interior of the effective domain of K; the directions remain arbitrary elements of ℝ^m × ℝ^n.

open scoped Pointwise def section35_text35_5_6 {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Set ((Fin m → ℝ) × (Fin n → ℝ)) := interior (saddleFunctionEffectiveDomain K)

Lean source

The directional difference quotient of a real-valued saddle kernel at (u, v) in the direction (u', v').

open scoped Pointwise noncomputable def realSaddleDirectionalDifferenceQuotientAt {E : Type*} {F : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E] [NormedAddCommGroup F] [NormedSpace ℝ F] (K : E → F → ℝ) (u : E) (v : F) (u' : E) (v' : F) (t : ℝ) : ℝ := (K (u + t • u') (v + t • v') - K u v) / t

Lean source

A real-valued saddle kernel has directional derivative L at (u, v) in the direction (u', v') when its right directional difference quotient converges to L as t ↓ 0.

open scoped Pointwise def HasRealSaddleDirectionalDerivativeAt {E : Type*} {F : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E] [NormedAddCommGroup F] [NormedSpace ℝ F] (K : E → F → ℝ) (u : E) (v : F) (u' : E) (v' : F) (L : ℝ) : Prop := Filter.Tendsto (realSaddleDirectionalDifferenceQuotientAt K u v u' v') (nhdsWithin (0 : ℝ) (Set.Ioi (0 : ℝ))) (nhds L)

Lean source

A bivariate real-valued kernel is jointly positively homogeneous in its two arguments if K (t • u') (t • v') = t * K u' v' for every t > 0.

open scoped Pointwise def IsPositivelyHomogeneousSaddleKernel {E : Type*} {F : Type*} [SMul ℝ E] [SMul ℝ F] (K : E → F → ℝ) : Prop := ∀ u' : E, ∀ v' : F, ∀ t : ℝ, 0 < t → K (t • u') (t • v') = t * K u' v'

Lean source

An extended-real-valued kernel on ℝ^m × ℝ^n is globally concave-convex when every u-slice is concave and every v-slice is convex on the whole space.

open scoped Pointwise abbrev IsGloballyConcaveConvexERealKernel {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := (∀ v : Fin n → ℝ, ConvexFunction (fun u : Fin m → ℝ => -K u v)) ∧ ∀ u : Fin m → ℝ, ConvexFunction (fun v : Fin n → ℝ => K u v)

Lean source

Helper for Theorem 35.5: reduce dense factor witnesses to countable dense sub-witnesses.

open scoped Pointwise lemma helperForTheorem_35_5_exists_countableDenseFactors {m n : ℕ} {C : Set (EuclideanSpace ℝ (Fin m))} {D : Set (EuclideanSpace ℝ (Fin n))} {C' : Set (EuclideanSpace ℝ (Fin m))} {D' : Set (EuclideanSpace ℝ (Fin n))} (hCclosure : C ⊆ closure C') (hDclosure : D ⊆ closure D') : ∃ C'' : Set (EuclideanSpace ℝ (Fin m)), ∃ D'' : Set (EuclideanSpace ℝ (Fin n)), C'' ⊆ C' ∧ D'' ⊆ D' ∧ C''.Countable ∧ D''.Countable ∧ C ⊆ closure C'' ∧ D ⊆ closure D''

Lean source

Helper for Theorem 35.5: a pointwise bounded family on a countable set has a subsequence converging at every point of that set.

open scoped Pointwise lemma helperForTheorem_35_5_exists_subseq_tendstoOn_countableSet {α : Type*} {E : Set α} {f : ℕ → α → ℝ} (hEcount : E.Countable) (hbounded : Function.PointwiseBoundedFamilyOn f E) : ∃ φ : ℕ → ℕ, StrictMono φ ∧ ∀ x ∈ E, ∃ l : ℝ, Filter.Tendsto (fun i => f (φ i) x) Filter.atTop (nhds l)

Lean source

Helper for Theorem 35.5: pointwise boundedness on a countable dense product yields a subsequence converging at every witness pair.

open scoped Pointwise lemma helperForTheorem_35_5_exists_subseq_tendstoOn_countableDenseProduct {m n : ℕ} {C'' : Set (EuclideanSpace ℝ (Fin m))} {D'' : Set (EuclideanSpace ℝ (Fin n))} {KSeq : ℕ → EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → ℝ} (hC''count : C''.Countable) (hD''count : D''.Countable) (hbounded : Function.PointwiseBoundedFamilyOn (fun i => Function.uncurry (KSeq i)) (C'' ×ˢ D'')) : ∃ φ : ℕ → ℕ, StrictMono φ ∧ ∀ u ∈ C'', ∀ v ∈ D'', ∃ l : ℝ, Filter.Tendsto (fun i => KSeq (φ i) u v) Filter.atTop (nhds l)

Lean source

Helper for Text 35.5.5: interior membership yields an open metric ball still contained in the effective domain.

open scoped Pointwise lemma helperForText_35_5_5_exists_ball_subset_effectiveDomain {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hInterior : (u, v) ∈ interior (saddleFunctionEffectiveDomain K)) : ∃ ε > 0, Metric.ball (u, v) ε ⊆ saddleFunctionEffectiveDomain K

Lean source

Helper for Text 35.5.5: a sufficiently short initial segment of any ray stays inside a given metric ball.

open scoped Pointwise lemma helperForText_35_5_5_exists_delta_line_subset_ball {E : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E] (x w : E) {ε : ℝ} (hε : 0 < ε) : ∃ δ > 0, ∀ t ∈ Set.Ico (0 : ℝ) δ, x + t • w ∈ Metric.ball x ε

Lean source

Helper for Theorem 35.6: the first-variable convex slice at v is proper at u, and its directional-derivative kernel is proper, positively homogeneous, convex, and finite in every direction.

open scoped Pointwise lemma helperForTheorem_35_6_firstSlice_directionalDerivativeData {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hC_open : IsOpen C) (_hD_open : IsOpen D) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) : let f : (Fin m → ℝ) → EReal := fun x => -K x v let Df : (Fin m → ℝ) → EReal := upperDirectionalDerivativeAt f u ProperConvexFunctionOn (Set.univ : Set (Fin m → ℝ)) f ∧ ProperConvexFunctionOn (Set.univ : Set (Fin m → ℝ)) Df ∧ PositivelyHomogeneous Df ∧ ConvexFunction Df ∧ Df 0 = 0 ∧ ∀ u' : Fin m → ℝ, Df u' ≠ (⊤ : EReal) ∧ Df u' ≠ (⊥ : EReal)

Lean source

Helper for Theorem 35.6: the second-variable convex slice at u is proper at v, and its directional-derivative kernel is proper, positively homogeneous, convex, and finite in every direction.

open scoped Pointwise lemma helperForTheorem_35_6_secondSlice_directionalDerivativeData {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (_hC_open : IsOpen C) (hD_open : IsOpen D) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) : let g : (Fin n → ℝ) → EReal := K u let Dg : (Fin n → ℝ) → EReal := upperDirectionalDerivativeAt g v ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) g ∧ ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) Dg ∧ PositivelyHomogeneous Dg ∧ ConvexFunction Dg ∧ Dg 0 = 0 ∧ ∀ v' : Fin n → ℝ, Dg v' ≠ (⊤ : EReal) ∧ Dg v' ≠ (⊥ : EReal)

Lean source

Helper for Theorem 35.6: the textbook split kernel built from the two axis directional derivatives is positively homogeneous, real concave-convex, and satisfies the splitting formula Kdir u' v' = Kdir u' 0 + Kdir 0 v'.

open scoped Pointwise lemma helperForTheorem_35_6_splitKernel_structure {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hC_open : IsOpen C) (hD_open : IsOpen D) (_hC_conv : Convex ℝ C) (_hD_conv : Convex ℝ D) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) : ∃ Kdir : (Fin m → ℝ) → (Fin n → ℝ) → ℝ, (∀ u' v', Kdir u' v' = -(upperDirectionalDerivativeAt (fun x => -K x v) u u').toReal + (upperDirectionalDerivativeAt (K u) v v').toReal) ∧ IsPositivelyHomogeneousSaddleKernel Kdir ∧ IsRealConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) Kdir ∧ ∀ u' v', Kdir u' v' = Kdir u' 0 + Kdir 0 v'

Lean source

Helper for Theorem 35.6: the split kernel from the textbook formula restricts on the axes to the one-variable directional-derivative kernels.

open scoped Pointwise lemma helperForTheorem_35_6_splitKernel_axisFormula {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hC_open : IsOpen C) (hD_open : IsOpen D) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) {Kdir : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hKdir_formula : ∀ u' v', Kdir u' v' = -(upperDirectionalDerivativeAt (fun x => -K x v) u u').toReal + (upperDirectionalDerivativeAt (K u) v v').toReal) : (∀ u', Kdir u' 0 = -(upperDirectionalDerivativeAt (fun x => -K x v) u u').toReal) ∧ ∀ v', Kdir 0 v' = (upperDirectionalDerivativeAt (K u) v v').toReal

Lean source

Helper for Theorem 35.6: the axis directional derivatives already match the split kernel, so the remaining work in the main theorem is only the genuinely mixed-direction limit.

open scoped Pointwise lemma helperForTheorem_35_6_axisDirectionalDerivatives_match_splitKernel {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hC_open : IsOpen C) (hD_open : IsOpen D) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) {Kdir : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hKdir_formula : ∀ u' v', Kdir u' v' = -(upperDirectionalDerivativeAt (fun x => -K x v) u u').toReal + (upperDirectionalDerivativeAt (K u) v v').toReal) : (∀ u', IsSaddleDirectionalDerivativeAt K u v u' 0 (Kdir u' 0 : EReal)) ∧ ∀ v', IsSaddleDirectionalDerivativeAt K u v 0 v' (Kdir 0 v' : EReal)

Lean source

Helper for Theorem 35.6: translating the open convex product C × D by -(u, v) yields relatively open convex direction domains containing 0, and every scaled direction with 0 < t ≤ 1 stays inside the original product.

open scoped Pointwise lemma helperForTheorem_35_6_translatedDirectionDomains {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) : let CU : Set (Fin m → ℝ) := {u' | u + u' ∈ C} let DV : Set (Fin n → ℝ) := {v' | v + v' ∈ D} IsRelativelyOpenConvex CU ∧ IsRelativelyOpenConvex DV ∧ (0 : Fin m → ℝ) ∈ CU ∧ (0 : Fin n → ℝ) ∈ DV ∧ ∀ {t : ℝ}, 0 < t → t ≤ 1 → ∀ {u' : Fin m → ℝ} {v' : Fin n → ℝ}, u' ∈ CU → v' ∈ DV → u + t • u' ∈ C ∧ v + t • v' ∈ D

Lean source

Helper for Theorem 35.6: every short positive step inside the translated direction domains lands at a finite value of K on the mixed point (u + t • u', v + t • v').

open scoped Pointwise lemma helperForTheorem_35_6_scaledStep_finiteValues {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) : let CU : Set (Fin m → ℝ) := {u' | u + u' ∈ C} let DV : Set (Fin n → ℝ) := {v' | v + v' ∈ D} ∀ {t : ℝ}, 0 < t → t ≤ 1 → ∀ {u' : Fin m → ℝ} {v' : Fin n → ℝ}, u' ∈ CU → v' ∈ DV → K (u + t • u') (v + t • v') ≠ (⊤ : EReal) ∧ K (u + t • u') (v + t • v') ≠ (⊥ : EReal)

Lean source

Helper for Theorem 35.6: the same short positive steps also keep the one-variable slice points (u + t • u', v) and (u, v + t • v') inside the finite patch C × D.

open scoped Pointwise lemma helperForTheorem_35_6_scaledAxisStep_finiteValues {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) : let CU : Set (Fin m → ℝ) := {u' | u + u' ∈ C} let DV : Set (Fin n → ℝ) := {v' | v + v' ∈ D} ∀ {t : ℝ}, 0 < t → t ≤ 1 → ∀ {u' : Fin m → ℝ} {v' : Fin n → ℝ}, u' ∈ CU → v' ∈ DV → (K (u + t • u') v ≠ (⊤ : EReal) ∧ K (u + t • u') v ≠ (⊥ : EReal)) ∧ (K u (v + t • v') ≠ (⊤ : EReal) ∧ K u (v + t • v') ≠ (⊥ : EReal))

§35 — Continuity and Differentiability of Saddle-Functions — Part 5

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Helper for Theorem 35.6: along any positive null sequence of admissible first-variable steps, the moving second-variable slices converge pointwise on D back to the fixed slice z ↦ K u z.

open scoped Pointwise lemma helperForTheorem_35_6_movingSecondSlice_pointwiseTendstoOnD {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hC_open : IsOpen C) (hC_conv : Convex ℝ C) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {u : Fin m → ℝ} (hu : u ∈ C) {τ : ℕ → ℝ} (hτpos : ∀ i, 0 < τ i) (hτle : ∀ i, τ i ≤ 1) (hτtendsto : Filter.Tendsto τ Filter.atTop (nhds (0 : ℝ))) {u' : Fin m → ℝ} (hu' : u' ∈ ({u'' : Fin m → ℝ | u + u'' ∈ C} : Set (Fin m → ℝ))) : ∀ z ∈ D, Filter.Tendsto (fun i => K (u + τ i • u') z) Filter.atTop (nhds (K u z))

Lean source

Helper for Theorem 35.6: the Chapter 24 pointwise-limit theorem controls the limsup of the upper directional derivatives of the moving second-variable slices.

open scoped Pointwise lemma helperForTheorem_35_6_movingSecondSlice_limsup_upperDerivative {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) {τ : ℕ → ℝ} (hτpos : ∀ i, 0 < τ i) (hτle : ∀ i, τ i ≤ 1) (hτtendsto : Filter.Tendsto τ Filter.atTop (nhds (0 : ℝ))) {u' : Fin m → ℝ} (hu' : u' ∈ ({u'' : Fin m → ℝ | u + u'' ∈ C} : Set (Fin m → ℝ))) (w : Fin n → ℝ) : Filter.limsup (fun i => upperDirectionalDerivativeAt (fun z => K (u + τ i • u') z) v w) Filter.atTop ≤ upperDirectionalDerivativeAt (K u) v w

Lean source

Helper for Theorem 35.6: along any positive null sequence of admissible second-variable steps, the moving first-variable slices converge pointwise on C back to the fixed slice x ↦ K x v.

open scoped Pointwise lemma helperForTheorem_35_6_movingFirstSlice_pointwiseTendstoOnC {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hD_open : IsOpen D) (hD_conv : Convex ℝ D) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {v : Fin n → ℝ} (hv : v ∈ D) {τ : ℕ → ℝ} (hτpos : ∀ i, 0 < τ i) (hτle : ∀ i, τ i ≤ 1) (hτtendsto : Filter.Tendsto τ Filter.atTop (nhds (0 : ℝ))) {v' : Fin n → ℝ} (hv' : v' ∈ ({v'' : Fin n → ℝ | v + v'' ∈ D} : Set (Fin n → ℝ))) : ∀ z ∈ C, Filter.Tendsto (fun i => K z (v + τ i • v')) Filter.atTop (nhds (K z v))

Lean source

Helper for Theorem 35.6: the Chapter 24 pointwise-limit theorem also controls the limsup of the upper directional derivatives of the moving first-variable slices after negation.

open scoped Pointwise lemma helperForTheorem_35_6_movingFirstSlice_limsup_upperDerivative {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) {τ : ℕ → ℝ} (hτpos : ∀ i, 0 < τ i) (hτle : ∀ i, τ i ≤ 1) (hτtendsto : Filter.Tendsto τ Filter.atTop (nhds (0 : ℝ))) {v' : Fin n → ℝ} (hv' : v' ∈ ({v'' : Fin n → ℝ | v + v'' ∈ D} : Set (Fin n → ℝ))) (w : Fin m → ℝ) : Filter.limsup (fun i => upperDirectionalDerivativeAt (fun x => -K x (v + τ i • v')) u w) Filter.atTop ≤ upperDirectionalDerivativeAt (fun x => -K x v) u w

Lean source

Helper for Theorem 35.6: convexity of the translated second slice bounds its upper directional derivative by the corresponding one-step mixed second-variable increment.

open scoped Pointwise lemma helperForTheorem_35_6_upperDirectionalDerivative_le_fixedStepQuotient {n : ℕ} {f : (Fin n → ℝ) → EReal} (hproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) f) (hf : ConvexFunction f) {x y : Fin n → ℝ} {t : ℝ} (hxFinite : f x ≠ (⊤ : EReal) ∧ f x ≠ (⊥ : EReal)) (ht : 0 < t) : upperDirectionalDerivativeAt f x y ≤ directionalDifferenceQuotientAt f x y t

Lean source

Helper for Theorem 35.6: convexity of the translated second slice bounds its upper directional derivative by the corresponding one-step mixed second-variable increment.

open scoped Pointwise lemma helperForTheorem_35_6_movingSecondDerivative_le_mixedSecondIncrement {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (_hD_conv : Convex ℝ D) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) {t : ℝ} (ht_pos : 0 < t) (ht_le : t ≤ 1) {u' : Fin m → ℝ} {v' : Fin n → ℝ} (hu' : u' ∈ ({u'' : Fin m → ℝ | u + u'' ∈ C} : Set (Fin m → ℝ))) (_hv' : v' ∈ ({v'' : Fin n → ℝ | v + v'' ∈ D} : Set (Fin n → ℝ))) : upperDirectionalDerivativeAt (fun z => K (u + t • u') z) v v' ≤ ((K (u + t • u') (v + t • v') - K (u + t • u') v) / (t : EReal))

Lean source

Helper for Theorem 35.6: convexity of the translated negated first slice bounds the mixed first-variable increment above by the negative translated upper directional derivative.

open scoped Pointwise lemma helperForTheorem_35_6_mixedFirstIncrement_le_negMovingFirstDerivative {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) {t : ℝ} (ht_pos : 0 < t) (ht_le : t ≤ 1) {u' : Fin m → ℝ} {v' : Fin n → ℝ} (hu' : u' ∈ ({u'' : Fin m → ℝ | u + u'' ∈ C} : Set (Fin m → ℝ))) (hv' : v' ∈ ({v'' : Fin n → ℝ | v + v'' ∈ D} : Set (Fin n → ℝ))) : ((K (u + t • u') (v + t • v') - K u (v + t • v')) / (t : EReal)) ≤ -upperDirectionalDerivativeAt (fun x => -K x (v + t • v')) u u'

Lean source

Helper for Theorem 35.6: once each translated witness pair admits scalar upper and lower bounds for the scaled mixed quotients, the whole quotient family is pointwise bounded on CU × DV.

open scoped Pointwise lemma helperForTheorem_35_6_pointwiseBounded_scaledQuotients_onTranslatedDomains {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {CU : Set (Fin m → ℝ)} {DV : Set (Fin n → ℝ)} {τ : ℕ → ℝ} (hUpper : ∀ {u' : Fin m → ℝ} {v' : Fin n → ℝ}, u' ∈ CU → v' ∈ DV → ∃ M : ℝ, ∀ i, (saddleDirectionalDifferenceQuotientAt K u v u' v' (τ i)).toReal ≤ M) (hLower : ∀ {u' : Fin m → ℝ} {v' : Fin n → ℝ}, u' ∈ CU → v' ∈ DV → ∃ m0 : ℝ, ∀ i, m0 ≤ (saddleDirectionalDifferenceQuotientAt K u v u' v' (τ i)).toReal) : Function.PointwiseBoundedFamilyOn (fun i => Function.uncurry fun u' v' => (saddleDirectionalDifferenceQuotientAt K u v u' v' (τ i)).toReal) (CU ×ˢ DV)

Lean source

Helper for Theorem 35.6: on every short admissible step, the mixed quotient splits exactly into either axis quotient plus the corresponding moved one-variable increment.

open scoped Pointwise lemma helperForTheorem_35_6_mixedQuotient_eq_axisPlusMovedIncrements {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) {t : ℝ} (ht_pos : 0 < t) (ht_le : t ≤ 1) {u' : Fin m → ℝ} {v' : Fin n → ℝ} (hu' : u' ∈ ({u'' : Fin m → ℝ | u + u'' ∈ C} : Set (Fin m → ℝ))) (hv' : v' ∈ ({v'' : Fin n → ℝ | v + v'' ∈ D} : Set (Fin n → ℝ))) : let mixedQ := (saddleDirectionalDifferenceQuotientAt K u v u' v' t).toReal let firstAxisQ := (((K (u + t • u') v - K u v) / (t : EReal)).toReal : ℝ) let secondAxisQ := (((K u (v + t • v') - K u v) / (t : EReal)).toReal : ℝ) let movedFirstQ := (((K (u + t • u') (v + t • v') - K u (v + t • v')) / (t : EReal)).toReal : ℝ) let movedSecondQ := (((K (u + t • u') (v + t • v') - K (u + t • u') v) / (t : EReal)).toReal : ℝ) mixedQ = firstAxisQ + movedSecondQ ∧ mixedQ = secondAxisQ + movedFirstQ

Lean source

Helper for Theorem 35.6: for any fixed short second-variable step η, the translated second-variable increment quotient converges back to the corresponding base-slice quotient along every admissible positive null sequence in the first variable.

open scoped Pointwise lemma helperForTheorem_35_6_fixedStepMovedSecondIncrement_tendsto_baseSecondQuotient {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hC_open : IsOpen C) (_hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) {τ : ℕ → ℝ} (hτpos : ∀ i, 0 < τ i) (hτle : ∀ i, τ i ≤ 1) (hτtendsto : Filter.Tendsto τ Filter.atTop (nhds (0 : ℝ))) {u' : Fin m → ℝ} (hu' : u' ∈ ({u'' : Fin m → ℝ | u + u'' ∈ C} : Set (Fin m → ℝ))) {η : ℝ} (hηpos : 0 < η) (hηle : η ≤ 1) {v' : Fin n → ℝ} (hv' : v' ∈ ({v'' : Fin n → ℝ | v + v'' ∈ D} : Set (Fin n → ℝ))) : Filter.Tendsto (fun i => (((K (u + τ i • u') (v + η • v') - K (u + τ i • u') v) / (η : EReal)).toReal : ℝ)) Filter.atTop (nhds (((K u (v + η • v') - K u v) / (η : EReal)).toReal : ℝ))

Lean source

Helper for Theorem 35.6: for any fixed short first-variable step η, the translated first-variable increment quotient converges back to the corresponding base-slice quotient along every admissible positive null sequence in the second variable.

open scoped Pointwise lemma helperForTheorem_35_6_fixedStepMovedFirstIncrement_tendsto_baseFirstQuotient {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (_hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) {τ : ℕ → ℝ} (hτpos : ∀ i, 0 < τ i) (hτle : ∀ i, τ i ≤ 1) (hτtendsto : Filter.Tendsto τ Filter.atTop (nhds (0 : ℝ))) {v' : Fin n → ℝ} (hv' : v' ∈ ({v'' : Fin n → ℝ | v + v'' ∈ D} : Set (Fin n → ℝ))) {η : ℝ} (hηpos : 0 < η) (hηle : η ≤ 1) {u' : Fin m → ℝ} (hu' : u' ∈ ({u'' : Fin m → ℝ | u + u'' ∈ C} : Set (Fin m → ℝ))) : Filter.Tendsto (fun i => (((K (u + η • u') (v + τ i • v') - K u (v + τ i • v')) / (η : EReal)).toReal : ℝ)) Filter.atTop (nhds (((K (u + η • u') v - K u v) / (η : EReal)).toReal : ℝ))

§35 — Continuity and Differentiability of Saddle-Functions — Part 6

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Helper for Theorem 35.6: along any admissible positive null sequence, the diagonal second-variable increment is eventually bounded above by the second-axis split-kernel value plus an arbitrary error.

open scoped Pointwise lemma helperForTheorem_35_6_movingSecondIncrement_eventually_le_axisSecondKernel {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {Kdir : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) (hAxisSecond : ∀ v'', IsSaddleDirectionalDerivativeAt K u v 0 v'' (Kdir 0 v'' : EReal)) {τ : ℕ → ℝ} (hτpos : ∀ i, 0 < τ i) (hτle : ∀ i, τ i ≤ 1) (hτtendsto : Filter.Tendsto τ Filter.atTop (nhds (0 : ℝ))) {u' : Fin m → ℝ} (hu' : u' ∈ ({u'' : Fin m → ℝ | u + u'' ∈ C} : Set (Fin m → ℝ))) {v' : Fin n → ℝ} (hv' : v' ∈ ({v'' : Fin n → ℝ | v + v'' ∈ D} : Set (Fin n → ℝ))) {ε : ℝ} (hε : 0 < ε) : ∀ᶠ i in Filter.atTop, (((K (u + τ i • u') (v + τ i • v') - K (u + τ i • u') v) / (τ i : EReal)).toReal : ℝ) ≤ Kdir 0 v' + ε

Lean source

Helper for Theorem 35.6: along any admissible positive null sequence, the diagonal first-variable increment is eventually bounded below by the first-axis split-kernel value minus an arbitrary error.

open scoped Pointwise lemma helperForTheorem_35_6_axisFirstKernel_le_eventually_movingFirstIncrement {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {Kdir : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) (hAxisFirst : ∀ u'', IsSaddleDirectionalDerivativeAt K u v u'' 0 (Kdir u'' 0 : EReal)) {τ : ℕ → ℝ} (hτpos : ∀ i, 0 < τ i) (hτle : ∀ i, τ i ≤ 1) (hτtendsto : Filter.Tendsto τ Filter.atTop (nhds (0 : ℝ))) {v' : Fin n → ℝ} (hv' : v' ∈ ({v'' : Fin n → ℝ | v + v'' ∈ D} : Set (Fin n → ℝ))) {u' : Fin m → ℝ} (hu' : u' ∈ ({u'' : Fin m → ℝ | u + u'' ∈ C} : Set (Fin m → ℝ))) {ε : ℝ} (hε : 0 < ε) : ∀ᶠ i in Filter.atTop, Kdir u' 0 - ε ≤ (((K (u + τ i • u') (v + τ i • v') - K u (v + τ i • v')) / (τ i : EReal)).toReal : ℝ)

Lean source

Helper for Theorem 35.6: along every sequence approaching 0 through positive steps, the real mixed quotient converges to the split-kernel value.

open scoped Pointwise lemma helperForTheorem_35_6_mixedQuotient_toReal_tendsto_splitKernelAlongNullSeq {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {Kdir : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) (hAxisFirst : ∀ u'', IsSaddleDirectionalDerivativeAt K u v u'' 0 (Kdir u'' 0 : EReal)) (hAxisSecond : ∀ v'', IsSaddleDirectionalDerivativeAt K u v 0 v'' (Kdir 0 v'' : EReal)) (hKdir_split : ∀ u'' v'', Kdir u'' v'' = Kdir u'' 0 + Kdir 0 v'') {τ : ℕ → ℝ} (hτWithin : Filter.Tendsto τ Filter.atTop (nhdsWithin (0 : ℝ) (Set.Ioi (0 : ℝ)))) {u' : Fin m → ℝ} (hu' : u' ∈ ({u'' : Fin m → ℝ | u + u'' ∈ C} : Set (Fin m → ℝ))) {v' : Fin n → ℝ} (hv' : v' ∈ ({v'' : Fin n → ℝ | v + v'' ∈ D} : Set (Fin n → ℝ))) : Filter.Tendsto (fun i => (saddleDirectionalDifferenceQuotientAt K u v u' v' (τ i)).toReal) Filter.atTop (nhds (Kdir u' v'))

Lean source

Helper for Theorem 35.6: every direction can be shrunk by a positive factor into the translated direction domains of the open patch C × D.

open scoped Pointwise lemma helperForTheorem_35_6_exists_posScale_mem_translatedDomains {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hC_open : IsOpen C) (hD_open : IsOpen D) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) (u' : Fin m → ℝ) (v' : Fin n → ℝ) : ∃ ρ : ℝ, 0 < ρ ∧ ρ • u' ∈ ({u'' : Fin m → ℝ | u + u'' ∈ C} : Set (Fin m → ℝ)) ∧ ρ • v' ∈ ({v'' : Fin n → ℝ | v + v'' ∈ D} : Set (Fin n → ℝ))

Lean source

Helper for Theorem 35.6: on admissible translated directions, the real mixed quotient has the full right-limit Kdir u' v' along nhdsWithin 0 (0, ∞).

open scoped Pointwise lemma helperForTheorem_35_6_mixedQuotient_toReal_tendsto_splitKernelWithin {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {Kdir : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) (hAxisFirst : ∀ u'', IsSaddleDirectionalDerivativeAt K u v u'' 0 (Kdir u'' 0 : EReal)) (hAxisSecond : ∀ v'', IsSaddleDirectionalDerivativeAt K u v 0 v'' (Kdir 0 v'' : EReal)) (hKdir_split : ∀ u'' v'', Kdir u'' v'' = Kdir u'' 0 + Kdir 0 v'') {u' : Fin m → ℝ} (hu' : u' ∈ ({u'' : Fin m → ℝ | u + u'' ∈ C} : Set (Fin m → ℝ))) {v' : Fin n → ℝ} (hv' : v' ∈ ({v'' : Fin n → ℝ | v + v'' ∈ D} : Set (Fin n → ℝ))) : Filter.Tendsto (fun t : ℝ => (saddleDirectionalDifferenceQuotientAt K u v u' v' t).toReal) (nhdsWithin (0 : ℝ) (Set.Ioi (0 : ℝ))) (nhds (Kdir u' v'))

Lean source

Helper for Theorem 35.6: after shrinking to the eventual region 0 < t < ρ, the mixed quotient in direction (u', v') is exactly the rescaled admissible quotient in direction (ρ • u', ρ • v') evaluated at step t / ρ.

open scoped Pointwise lemma helperForTheorem_35_6_reparametrize_mixedQuotient_by_posScale {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) {u' : Fin m → ℝ} {v' : Fin n → ℝ} {ρ : ℝ} (hρpos : 0 < ρ) (huρ : ρ • u' ∈ ({u'' : Fin m → ℝ | u + u'' ∈ C} : Set (Fin m → ℝ))) (hvρ : ρ • v' ∈ ({v'' : Fin n → ℝ | v + v'' ∈ D} : Set (Fin n → ℝ))) : ∀ᶠ t in nhdsWithin (0 : ℝ) (Set.Ioi (0 : ℝ)), (saddleDirectionalDifferenceQuotientAt K u v u' v' t).toReal = (1 / ρ) * (saddleDirectionalDifferenceQuotientAt K u v (ρ • u') (ρ • v') (t / ρ)).toReal

Lean source

Helper for Theorem 35.6: if a subsequential limit has the correct second marginal and the correct first-axis values, then the textbook splitting formula forces equality with Kdir on the translated direction domain.

open scoped Pointwise lemma helperForTheorem_35_6_subseqLimit_eq_splitKernel_of_secondMarginal {m n : ℕ} {CU : Set (Fin m → ℝ)} {DV : Set (Fin n → ℝ)} {L Kdir : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hAxisFirst : ∀ u' ∈ CU, L u' 0 = Kdir u' 0) (hSecondMarginal : ∀ u' ∈ CU, ∀ v' ∈ DV, L u' v' - L u' 0 = Kdir 0 v') (hSplit : ∀ u' v', Kdir u' v' = Kdir u' 0 + Kdir 0 v') : ∀ u' ∈ CU, ∀ v' ∈ DV, L u' v' = Kdir u' v'

Lean source

Helper for Theorem 35.6: symmetrically, the first marginal together with the second-axis values already determines the whole subsequential limit on CU × DV.

open scoped Pointwise lemma helperForTheorem_35_6_subseqLimit_eq_splitKernel_of_firstMarginal {m n : ℕ} {CU : Set (Fin m → ℝ)} {DV : Set (Fin n → ℝ)} {L Kdir : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hAxisSecond : ∀ v' ∈ DV, L 0 v' = Kdir 0 v') (hFirstMarginal : ∀ u' ∈ CU, ∀ v' ∈ DV, L u' v' - L 0 v' = Kdir u' 0) (hSplit : ∀ u' v', Kdir u' v' = Kdir u' 0 + Kdir 0 v') : ∀ u' ∈ CU, ∀ v' ∈ DV, L u' v' = Kdir u' v'

§35 — Continuity and Differentiability of Saddle-Functions — Part 7

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Theorem 35.6: if K is a globally concave-convex extended-real-valued saddle function on ℝ^m × ℝ^n, and if C × D ⊆ ℝ^m × ℝ^n is an open convex product on which K is finite, then for each (u, v) ∈ C × D there is a finite directional-derivative kernel Kdir on ℝ^m × ℝ^n such that K'(u, v; u', v') = Kdir u' v' for every direction, Kdir is positively homogeneous and concave-convex on all of ℝ^m × ℝ^n, and Kdir u' v' = Kdir u' 0 + Kdir 0 v'.

open scoped Pointwise -- Proof sketch: use the global saddle hypothesis on `K` to obtain the one-variable convexity -- and concavity properties on all of `ℝ^m × ℝ^n`, then restrict to the open convex set `C × D` -- where `K` is finite so that `EReal.toReal` is available. Apply the one-variable -- directional-derivative existence results in the `u`- and `v`-variables at `(u, v)`, and -- combine these slice derivatives to obtain a real-valued kernel on all directions. Its positive -- homogeneity, concave-convexity, and splitting formula come from the corresponding properties -- of directional derivatives of saddle functions. theorem section35_theorem35_6 {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : ∀ u ∈ C, ∀ v ∈ D, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) : ∃ Kdir : (Fin m → ℝ) → (Fin n → ℝ) → ℝ, (∀ u' v', IsSaddleDirectionalDerivativeAt K u v u' v' (Kdir u' v' : EReal)) ∧ IsPositivelyHomogeneousSaddleKernel Kdir ∧ IsRealConcaveConvexOn (Set.univ : Set (Fin m → ℝ)) (Set.univ : Set (Fin n → ℝ)) Kdir ∧ ∀ u' v', Kdir u' v' = Kdir u' 0 + Kdir 0 v'

Lean source

Text 35.6.1: the partial subdifferential in the first variable at (u, v) consists of the vectors u* such that K u' v ≤ K u v + ⟪u*, u' - u⟫ for every u'. Equivalently, these are the supergradients of the concave slice u ↦ K u v at u.

open scoped Pointwise -- Proof sketch: use the global saddle hypothesis on `K` to obtain the one-variable convexity -- and concavity properties on all of `ℝ^m × ℝ^n`, then restrict to the open convex set `C × D` -- where `K` is finite so that `EReal.toReal` is available. Apply the one-variable -- directional-derivative existence results in the `u`- and `v`-variables at `(u, v)`, and -- combine these slice derivatives to obtain a real-valued kernel on all directions. Its positive -- homogeneity, concave-convexity, and splitting formula come from the corresponding properties -- of directional derivatives of saddle functions. def partialSubdifferentialInFirstVariable {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (u : Fin m → ℝ) (v : Fin n → ℝ) : Set (Fin m → ℝ) := {uStar | ∀ u' : Fin m → ℝ, K u' v ≤ K u v + ((∑ i : Fin m, uStar i * (u' i - u i)) : EReal)}

Lean source

Text 35.6.2: the partial subdifferential in the second variable at (u, v) consists of the vectors v* such that K u v' ≥ K u v + ⟪v*, v' - v⟫ for every v'. Equivalently, these are the subgradients of the convex slice v ↦ K u v at v.

open scoped Pointwise -- Proof sketch: use the global saddle hypothesis on `K` to obtain the one-variable convexity -- and concavity properties on all of `ℝ^m × ℝ^n`, then restrict to the open convex set `C × D` -- where `K` is finite so that `EReal.toReal` is available. Apply the one-variable -- directional-derivative existence results in the `u`- and `v`-variables at `(u, v)`, and -- combine these slice derivatives to obtain a real-valued kernel on all directions. Its positive -- homogeneity, concave-convexity, and splitting formula come from the corresponding properties -- of directional derivatives of saddle functions. def partialSubdifferentialInSecondVariable {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (u : Fin m → ℝ) (v : Fin n → ℝ) : Set (Fin n → ℝ) := {vStar | ∀ v' : Fin n → ℝ, K u v' ≥ K u v + ((∑ i : Fin n, vStar i * (v' i - v i)) : EReal)}

Lean source

Text 35.6.3: the product subdifferential of K at (u, v) is ∂₁ K (u, v) × ∂₂ K (u, v). Its elements (u*, v*) are the subgradients of K at (u, v).

open scoped Pointwise -- Proof sketch: use the global saddle hypothesis on `K` to obtain the one-variable convexity -- and concavity properties on all of `ℝ^m × ℝ^n`, then restrict to the open convex set `C × D` -- where `K` is finite so that `EReal.toReal` is available. Apply the one-variable -- directional-derivative existence results in the `u`- and `v`-variables at `(u, v)`, and -- combine these slice derivatives to obtain a real-valued kernel on all directions. Its positive -- homogeneity, concave-convexity, and splitting formula come from the corresponding properties -- of directional derivatives of saddle functions. def productSubdifferentialAt {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (u : Fin m → ℝ) (v : Fin n → ℝ) : Set ((Fin m → ℝ) × (Fin n → ℝ)) := partialSubdifferentialInFirstVariable K u v ×ˢ partialSubdifferentialInSecondVariable K u v

Lean source

A packaged extended-real saddle kernel whose u-slices are concave and whose v-slices are convex on the whole space.

open scoped Pointwise -- Proof sketch: use the global saddle hypothesis on `K` to obtain the one-variable convexity -- and concavity properties on all of `ℝ^m × ℝ^n`, then restrict to the open convex set `C × D` -- where `K` is finite so that `EReal.toReal` is available. Apply the one-variable -- directional-derivative existence results in the `u`- and `v`-variables at `(u, v)`, and -- combine these slice derivatives to obtain a real-valued kernel on all directions. Its positive -- homogeneity, concave-convexity, and splitting formula come from the corresponding properties -- of directional derivatives of saddle functions. structure SaddleKernel (m n : ℕ) where toFun : (Fin m → ℝ) → (Fin n → ℝ) → EReal isSaddle : IsGloballyConcaveConvexERealKernel toFun

Lean source

Text 35.6.4: for a saddle kernel K, the subdifferential is the multivalued mapping (u, v) ↦ ∂K(u, v). Here ∂K(u, v) is realized as the product subdifferential from Text 35.6.3.

open scoped Pointwise -- Proof sketch: use the global saddle hypothesis on `K` to obtain the one-variable convexity -- and concavity properties on all of `ℝ^m × ℝ^n`, then restrict to the open convex set `C × D` -- where `K` is finite so that `EReal.toReal` is available. Apply the one-variable -- directional-derivative existence results in the `u`- and `v`-variables at `(u, v)`, and -- combine these slice derivatives to obtain a real-valued kernel on all directions. Its positive -- homogeneity, concave-convexity, and splitting formula come from the corresponding properties -- of directional derivatives of saddle functions. def saddleSubdifferential {m n : ℕ} (K : SaddleKernel m n) : ((Fin m → ℝ) × (Fin n → ℝ)) → Set ((Fin m → ℝ) × (Fin n → ℝ)) := fun p => productSubdifferentialAt K.toFun p.1 p.2 -- Proof sketch: `∂K(u, v)` is the product of the partial subdifferentials in the first and -- second variables. For a saddle kernel, these partial subdifferentials are closed convex sets -- by the one-variable convex/concave subdifferential theory, and the product of closed convex -- sets is again closed and convex in `ℝ^m × ℝ^n`.

Lean source

Text 35.6.5: for every (u, v) ∈ ℝ^m × ℝ^n, the set ∂K(u, v) is a possibly empty closed convex subset of ℝ^m × ℝ^n.

open scoped Pointwise -- Proof sketch: use the global saddle hypothesis on `K` to obtain the one-variable convexity -- and concavity properties on all of `ℝ^m × ℝ^n`, then restrict to the open convex set `C × D` -- where `K` is finite so that `EReal.toReal` is available. Apply the one-variable -- directional-derivative existence results in the `u`- and `v`-variables at `(u, v)`, and -- combine these slice derivatives to obtain a real-valued kernel on all directions. Its positive -- homogeneity, concave-convexity, and splitting formula come from the corresponding properties -- of directional derivatives of saddle functions. theorem section35_text35_6_5 {m n : ℕ} (K : SaddleKernel m n) (u : Fin m → ℝ) (v : Fin n → ℝ) : IsClosed (saddleSubdifferential K (u, v)) ∧ Convex ℝ (saddleSubdifferential K (u, v))

Lean source

Helper for Text 35.6.5: a finite sum of real products coerced to EReal is the coercion of the corresponding real sum.

open scoped Pointwise -- Proof sketch: use the global saddle hypothesis on `K` to obtain the one-variable convexity -- and concavity properties on all of `ℝ^m × ℝ^n`, then restrict to the open convex set `C × D` -- where `K` is finite so that `EReal.toReal` is available. Apply the one-variable -- directional-derivative existence results in the `u`- and `v`-variables at `(u, v)`, and -- combine these slice derivatives to obtain a real-valued kernel on all directions. Its positive -- homogeneity, concave-convexity, and splitting formula come from the corresponding properties -- of directional derivatives of saddle functions. lemma helperForText_35_6_5_sum_coe_products {k : ℕ} (a : Fin k → ℝ) (x : Fin k → ℝ) : (∑ i : Fin k, ((x i : ℝ) : EReal) * ((a i : ℝ) : EReal)) = (((∑ i : Fin k, x i * a i : ℝ)) : EReal)

Lean source

Helper for Text 35.6.5: every extended-real affine halfspace in a finite-dimensional Euclidean coordinate space is closed and convex.

open scoped Pointwise -- Proof sketch: use the global saddle hypothesis on `K` to obtain the one-variable convexity -- and concavity properties on all of `ℝ^m × ℝ^n`, then restrict to the open convex set `C × D` -- where `K` is finite so that `EReal.toReal` is available. Apply the one-variable -- directional-derivative existence results in the `u`- and `v`-variables at `(u, v)`, and -- combine these slice derivatives to obtain a real-valued kernel on all directions. Its positive -- homogeneity, concave-convexity, and splitting formula come from the corresponding properties -- of directional derivatives of saddle functions. lemma helperForText_35_6_5_erealAffineHalfspace_isClosed_convex {k : ℕ} (α c : EReal) (a : Fin k → ℝ) : IsClosed {x : Fin k → ℝ | α ≤ c + ((∑ i : Fin k, x i * a i) : EReal)} ∧ Convex ℝ {x : Fin k → ℝ | α ≤ c + ((∑ i : Fin k, x i * a i) : EReal)}

Lean source

Helper for Text 35.6.5: every extended-real affine superhalfspace in a finite-dimensional Euclidean coordinate space is closed and convex.

open scoped Pointwise -- Proof sketch: use the global saddle hypothesis on `K` to obtain the one-variable convexity -- and concavity properties on all of `ℝ^m × ℝ^n`, then restrict to the open convex set `C × D` -- where `K` is finite so that `EReal.toReal` is available. Apply the one-variable -- directional-derivative existence results in the `u`- and `v`-variables at `(u, v)`, and -- combine these slice derivatives to obtain a real-valued kernel on all directions. Its positive -- homogeneity, concave-convexity, and splitting formula come from the corresponding properties -- of directional derivatives of saddle functions. lemma helperForText_35_6_5_erealAffineSuperhalfspace_isClosed_convex {k : ℕ} (α c : EReal) (a : Fin k → ℝ) : IsClosed {x : Fin k → ℝ | α ≥ c + ((∑ i : Fin k, x i * a i) : EReal)} ∧ Convex ℝ {x : Fin k → ℝ | α ≥ c + ((∑ i : Fin k, x i * a i) : EReal)}

Lean source

Helper for Text 35.6.5: the partial subdifferential in the first variable is an intersection of closed convex affine halfspaces, hence is itself closed and convex.

open scoped Pointwise -- Proof sketch: use the global saddle hypothesis on `K` to obtain the one-variable convexity -- and concavity properties on all of `ℝ^m × ℝ^n`, then restrict to the open convex set `C × D` -- where `K` is finite so that `EReal.toReal` is available. Apply the one-variable -- directional-derivative existence results in the `u`- and `v`-variables at `(u, v)`, and -- combine these slice derivatives to obtain a real-valued kernel on all directions. Its positive -- homogeneity, concave-convexity, and splitting formula come from the corresponding properties -- of directional derivatives of saddle functions. lemma helperForText_35_6_5_firstPartial_isClosed_convex {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (u : Fin m → ℝ) (v : Fin n → ℝ) : IsClosed (partialSubdifferentialInFirstVariable K u v) ∧ Convex ℝ (partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.5: the partial subdifferential in the second variable is an intersection of closed convex affine halfspaces, hence is itself closed and convex.

open scoped Pointwise -- Proof sketch: use the global saddle hypothesis on `K` to obtain the one-variable convexity -- and concavity properties on all of `ℝ^m × ℝ^n`, then restrict to the open convex set `C × D` -- where `K` is finite so that `EReal.toReal` is available. Apply the one-variable -- directional-derivative existence results in the `u`- and `v`-variables at `(u, v)`, and -- combine these slice derivatives to obtain a real-valued kernel on all directions. Its positive -- homogeneity, concave-convexity, and splitting formula come from the corresponding properties -- of directional derivatives of saddle functions. lemma helperForText_35_6_5_secondPartial_isClosed_convex {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (u : Fin m → ℝ) (v : Fin n → ℝ) : IsClosed (partialSubdifferentialInSecondVariable K u v) ∧ Convex ℝ (partialSubdifferentialInSecondVariable K u v)

§35 — Continuity and Differentiability of Saddle-Functions — Part 8

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

The first-variable directional derivative function attached to a saddle kernel at (u, v), defined by u' ↦ -K'(u, v; -u', 0) using the infimum of all admissible directional-derivative values.

open scoped Pointwise noncomputable def firstVariableDirectionalDerivativeFunction {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (u : Fin m → ℝ) (v : Fin n → ℝ) : (Fin m → ℝ) → EReal := fun u' => -sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v (-u') 0 L}

Lean source

The lower semicontinuous hull of an extended-real-valued function, realized by the closure of its epigraph.

open scoped Pointwise noncomputable def saddleLowerSemicontinuousHull {α : Type*} [TopologicalSpace α] (f : α → EReal) : α → EReal := fun x => sInf {r : EReal | (x, r) ∈ closure {p : α × EReal | f p.1 ≤ p.2}}

Lean source

The support function of a set of vectors in ℝ^m, viewed as an extended-real-valued function on ℝ^m.

open scoped Pointwise noncomputable def supportFunctionOfSet {m : ℕ} (S : Set (Fin m → ℝ)) : (Fin m → ℝ) → EReal := fun u' => sSup ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * u' i) : ℝ) : EReal)) '' S)

Lean source

Helper for Text 35.6.6: the reflected first slice x ↦ -K (-x) v is convex, because the saddle hypothesis already gives convexity of x ↦ -K x v and the involution x ↦ -x is linear.

open scoped Pointwise lemma helperForText_35_6_6_reflectedFirstSlice_convex {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) (v : Fin n → ℝ) : ConvexFunction (fun x : Fin m → ℝ => -K (-x) v)

Lean source

Helper for Text 35.6.6: recentering the first slice at -u does not change the finiteness of the base value, because the reflected slice still evaluates to -K u v.

open scoped Pointwise lemma helperForText_35_6_6_reflectedFirstSlice_finiteAtBase {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) : (fun x : Fin m → ℝ => -K (-x) v) (-u) ≠ (⊤ : EReal) ∧ (fun x : Fin m → ℝ => -K (-x) v) (-u) ≠ (⊥ : EReal)

Lean source

Helper for Text 35.6.6: after recentering the first slice by x ↦ -x, the textbook first-variable directional derivative function is exactly the ordinary upper directional derivative of the convex slice x ↦ -K (-x) v at the base point -u.

open scoped Pointwise lemma helperForText_35_6_6_recenteredFirstSlice_directionalDerivative {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (u' : Fin m → ℝ) : firstVariableDirectionalDerivativeFunction K u v u' = upperDirectionalDerivativeAt (fun x => -K (-x) v) (-u) u'

Lean source

Helper for Text 35.6.6: the whole textbook first-variable directional-derivative function is exactly the Chapter 23 directional derivative of the reflected first slice at -u.

open scoped Pointwise lemma helperForText_35_6_6_firstVariableDirectionalDerivative_eq_upperDirectionalDerivative {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) : firstVariableDirectionalDerivativeFunction K u v = upperDirectionalDerivativeAt (fun x => -K (-x) v) (-u)

Lean source

Helper for Text 35.6.6: membership in the Euclidean subdifferential of the reflected slice is exactly the textbook first-partial supporting inequality.

open scoped Pointwise lemma helperForText_35_6_6_reflectedSliceSubgradient_iff_partialFirstMem {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {uStar : Fin m → ℝ} : dotProductEquiv ℝ (Fin m) uStar ∈ subdifferentialAt (fun x => -K (-x) v) (-u) ↔ uStar ∈ partialSubdifferentialInFirstVariable K u v

Lean source

Helper for Text 35.6.6: the Euclidean subdifferential of the recentered convex slice x ↦ -K (-x) v at -u is exactly the first partial subdifferential ∂₁ K(u, v).

open scoped Pointwise lemma helperForText_35_6_6_partialFirst_eq_sliceSubdifferential {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} : ((dotProductEquiv ℝ (Fin m)) ⁻¹' subdifferentialAt (fun x => -K (-x) v) (-u)) = partialSubdifferentialInFirstVariable K u v

Lean source

Helper for Text 35.6.6: after identifying the slice subdifferential with ∂₁ K(u, v), the Chapter 23 support value is exactly the textbook support function of the first partial subdifferential.

open scoped Pointwise lemma helperForText_35_6_6_supportFunctionOfSet_eq_supportFunctionEReal {m : ℕ} (S : Set (Fin m → ℝ)) : supportFunctionEReal S = supportFunctionOfSet S

Lean source

Helper for Text 35.6.6: after identifying the slice subdifferential with ∂₁ K(u, v), the Chapter 23 support value is exactly the textbook support function of the first partial subdifferential.

open scoped Pointwise lemma helperForText_35_6_6_sliceSupport_eq_firstPartialSupport {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} : subdifferentialSupportAt (fun x => -K (-x) v) (-u) = supportFunctionOfSet (partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.6: restricting a closure computation to an open neighborhood of the base point does not change membership, so localizing to the finite-height EReal range is legitimate.

open scoped Pointwise lemma helperForText_35_6_6_mem_closure_inter_open_iff {α : Type*} [TopologicalSpace α] {s t : Set α} {x : α} (hx : x ∈ t) (ht : IsOpen t) : x ∈ closure s ↔ x ∈ closure (s ∩ t)

Lean source

Helper for Text 35.6.6: a finite EReal height lies in the closure of the full EReal epigraph exactly when the corresponding real height lies in the closure of the ordinary real epigraph.

open scoped Pointwise lemma helperForText_35_6_6_realHeight_mem_saddleClosure_iff_realEpigraphClosure {m : ℕ} (φ : (Fin m → ℝ) → EReal) (x : Fin m → ℝ) (r : ℝ) : ((x, (r : EReal)) ∈ closure {p : (Fin m → ℝ) × EReal | φ p.1 ≤ p.2}) ↔ ((x, r) ∈ closure (epigraph (S := (Set.univ : Set (Fin m → ℝ))) φ))

Lean source

Helper for Text 35.6.6: the closure of the full EReal epigraph remains upward closed in the second coordinate.

open scoped Pointwise lemma helperForText_35_6_6_saddleClosure_upwardInSecondCoordinate {m : ℕ} {φ : (Fin m → ℝ) → EReal} {x : Fin m → ℝ} {μ ν : EReal} (hμν : μ ≤ ν) (hx : (x, μ) ∈ closure {p : (Fin m → ℝ) × EReal | φ p.1 ≤ p.2}) : (x, ν) ∈ closure {p : (Fin m → ℝ) × EReal | φ p.1 ≤ p.2}

Lean source

Helper for Text 35.6.6: the EReal epigraph hull used for saddle kernels is exactly the Chapter 2 vertical-slice infimum epigraphClosureInf.

open scoped Pointwise lemma helperForText_35_6_6_saddleLowerHull_eq_epigraphClosureInf {m : ℕ} (φ : (Fin m → ℝ) → EReal) : saddleLowerSemicontinuousHull φ = epigraphClosureInf φ

Lean source

Helper for Text 35.6.6: the closure-by-epigraph construction epigraphClosureInf is exactly the ordinary lower semicontinuous hull.

open scoped Pointwise lemma helperForText_35_6_6_epigraphClosureInf_eq_lowerSemicontinuousHull {m : ℕ} (φ : (Fin m → ℝ) → EReal) : epigraphClosureInf φ = lowerSemicontinuousHull φ

Lean source

Helper for Text 35.6.6: when the directional-derivative function never takes the value ⊥, its epigraph hull agrees with the Chapter 2 convex closure.

open scoped Pointwise lemma helperForText_35_6_6_epigraphClosureInf_eq_convexFunctionClosure_of_no_bot {m : ℕ} {D : (Fin m → ℝ) → EReal} (hnotbot : ∀ y : Fin m → ℝ, D y ≠ (⊥ : EReal)) : epigraphClosureInf D = convexFunctionClosure D

Lean source

Helper for Text 35.6.6: if an improper convex function attains ⊥ and its effective domain is dense, then the epigraph-closure hull is the constant ⊥ function.

open scoped Pointwise lemma helperForText_35_6_6_epigraphClosureInf_eq_bot_of_dense_effectiveDomain {m : ℕ} {D : (Fin m → ℝ) → EReal} (hImproper : ImproperConvexFunctionOn (Set.univ : Set (Fin m → ℝ)) D) (hBot : ∃ y : Fin m → ℝ, D y = (⊥ : EReal)) (hDense : closure (effectiveDomain (Set.univ : Set (Fin m → ℝ)) D) = Set.univ) : epigraphClosureInf D = fun _ => (⊥ : EReal)

Lean source

Helper for Text 35.6.6: once the reflected convex slice g has dense effective domain, the relative-interior transport from Theorem 23.3 forces the effective domain of its directional derivative y ↦ g'(x; y) to be dense as well.

open scoped Pointwise lemma helperForText_35_6_6_denseEffectiveDomain_of_dense_reflectedSliceDomain {m : ℕ} {g : (Fin m → ℝ) → EReal} {x : Fin m → ℝ} (hg : ConvexFunction g) (hx : g x ≠ (⊤ : EReal) ∧ g x ≠ (⊥ : EReal)) (hDense : closure (effectiveDomain (Set.univ : Set (Fin m → ℝ)) g) = Set.univ) : closure (effectiveDomain (Set.univ : Set (Fin m → ℝ)) (upperDirectionalDerivativeAt g x)) = Set.univ

Lean source

Helper for Text 35.6.6: once the reflected slice g has dense effective domain, the empty subdifferential branch collapses epigraphClosureInf (g'(x; ·)) to the constant ⊥ function.

open scoped Pointwise lemma helperForText_35_6_6_epigraphClosureInf_eq_bot_of_empty_sliceSubdifferential_of_dense_reflectedSliceDomain {m : ℕ} {g : (Fin m → ℝ) → EReal} {x : Fin m → ℝ} (hg : ConvexFunction g) (hx : g x ≠ (⊤ : EReal) ∧ g x ≠ (⊥ : EReal)) (hsubEmpty : subdifferentialAt g x = ∅) (hDense : closure (effectiveDomain (Set.univ : Set (Fin m → ℝ)) g) = Set.univ) : epigraphClosureInf (upperDirectionalDerivativeAt g x) = fun _ => (⊥ : EReal)

Lean source

Helper for Text 35.6.6: if the reflected slice subdifferential is nonempty, then the Chapter 2 epigraph hull of the slice directional derivative already matches the Chapter 23 support formula.

open scoped Pointwise lemma helperForText_35_6_6_epigraphClosureInf_eq_sliceSupport_of_nonempty_sliceSubdifferential {m : ℕ} {g : (Fin m → ℝ) → EReal} {x : Fin m → ℝ} (hg : ConvexFunction g) (hx : g x ≠ (⊤ : EReal) ∧ g x ≠ (⊥ : EReal)) (hsub : Set.Nonempty (subdifferentialAt g x)) : epigraphClosureInf (upperDirectionalDerivativeAt g x) = subdifferentialSupportAt g x

Lean source

Helper for Text 35.6.6: if the reflected slice subdifferential is empty, then the Chapter 23 support function is the constant ⊥ function.

open scoped Pointwise lemma helperForText_35_6_6_sliceSupport_eq_bot_of_empty_sliceSubdifferential {m : ℕ} {g : (Fin m → ℝ) → EReal} {x : Fin m → ℝ} (hsubEmpty : subdifferentialAt g x = ∅) : subdifferentialSupportAt g x = fun _ => (⊥ : EReal)

Lean source

Helper for Text 35.6.6: nonemptiness of the textbook first partial subdifferential is equivalent to nonemptiness of the Euclidean subdifferential of the reflected slice.

open scoped Pointwise lemma helperForText_35_6_6_partialFirst_nonempty_iff_sliceSubdifferential_nonempty {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} : Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ↔ Set.Nonempty (subdifferentialAt (fun x => -K (-x) v) (-u))

Lean source

Helper for Text 35.6.6: if ∂₁ K(u, v) is empty, then its textbook support function is the constant ⊥ function.

open scoped Pointwise lemma helperForText_35_6_6_firstPartialSupport_eq_bot_of_empty_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hpartialEmpty : partialSubdifferentialInFirstVariable K u v = ∅) : supportFunctionOfSet (partialSubdifferentialInFirstVariable K u v) = fun _ => (⊥ : EReal)

Lean source

Helper for Text 35.6.6: on the empty first-partial branch, identifying any candidate hull with the textbook support function is equivalent to showing that the candidate hull is constantly ⊥.

open scoped Pointwise lemma helperForText_35_6_6_eq_firstPartialSupport_iff_eq_bot_of_empty_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (φ : (Fin m → ℝ) → EReal) (hpartialEmpty : partialSubdifferentialInFirstVariable K u v = ∅) : φ = supportFunctionOfSet (partialSubdifferentialInFirstVariable K u v) ↔ φ = fun _ => (⊥ : EReal)

Lean source

Helper for Text 35.6.6: Theorem 23.2 identifies the convex closure of the textbook first-variable directional derivative with the support function of ∂₁ K(u, v). This is the mathematically correct closure statement available even before comparing with the stronger lower-semicontinuous hull used later in the textbook phrasing.

open scoped Pointwise lemma helperForText_35_6_6_convexFunctionClosure_eq_firstPartialSupport {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) : convexFunctionClosure (firstVariableDirectionalDerivativeFunction K u v) = supportFunctionOfSet (partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.6: when ∂₁ K(u, v) is empty, the correct Chapter 23 conclusion is that the convex closure of the textbook directional-derivative function is constantly ⊥. This does not by itself imply the stronger epigraphClosureInf endpoint used in the remaining blocked branch.

open scoped Pointwise lemma helperForText_35_6_6_convexFunctionClosure_eq_bot_of_empty_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartialEmpty : partialSubdifferentialInFirstVariable K u v = ∅) : convexFunctionClosure (firstVariableDirectionalDerivativeFunction K u v) = fun _ => (⊥ : EReal)

§35 — Continuity and Differentiability of Saddle-Functions — Part 9

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Text 35.6.6: let K be a saddle function on ℝ^m × ℝ^n, and let (u, v) be a point with finite value K u v. Define φ(u') = -K'(u, v; -u', 0). Then φ is a convex function on ℝ^m, and the convex closure of φ coincides with the support function of the closed convex set ∂₁ K(u, v).

open scoped Pointwise theorem section35_text35_6_6 {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) : ConvexFunction (firstVariableDirectionalDerivativeFunction K u v) ∧ ∀ u' : Fin m → ℝ, convexFunctionClosure (firstVariableDirectionalDerivativeFunction K u v) u' = supportFunctionOfSet (partialSubdifferentialInFirstVariable K u v) u'

Lean source

Helper for Text 35.6.6: the textbook hull/support identity is correct on the honest branch where the first partial subdifferential is nonempty.

open scoped Pointwise lemma helperForText_35_6_6_saddleLowerHull_eq_firstPartialSupport_of_nonempty_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartial : Set.Nonempty (partialSubdifferentialInFirstVariable K u v)) : saddleLowerSemicontinuousHull (firstVariableDirectionalDerivativeFunction K u v) = supportFunctionOfSet (partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.6: the mathematically correct Chapter 23 closure statement identifies the convex closure of the first directional-derivative slice with the support function of ∂₁ K(u,v).

open scoped Pointwise lemma helperForText_35_6_6_saddleLowerHull_eq_firstPartialSupport {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) : convexFunctionClosure (firstVariableDirectionalDerivativeFunction K u v) = supportFunctionOfSet (partialSubdifferentialInFirstVariable K u v)

§35 — Continuity and Differentiability of Saddle-Functions — Part 10

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

The second-variable directional derivative function attached to a saddle kernel at (u, v), defined by v' ↦ K'(u, v; 0, v') using the infimum of all admissible directional-derivative values.

open scoped Pointwise noncomputable def secondVariableDirectionalDerivativeFunction {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (u : Fin m → ℝ) (v : Fin n → ℝ) : (Fin n → ℝ) → EReal := fun v' => sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v 0 v' L}

Lean source

Text 35.6.7: let K be a saddle function on ℝ^m × ℝ^n, and let (u, v) be a point with finite value K u v. Assume that the effective domain of the second slice v' ↦ K u v' is dense in ℝ^n. Define ψ(v') = K'(u, v; 0, v'). Then ψ is a convex function on ℝ^n, and the lower semicontinuous hull of ψ coincides with the support function of the closed convex set ∂₂ K(u, v).

open scoped Pointwise theorem section35_text35_6_7 {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hDense : closure (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (K u)) = Set.univ) : ConvexFunction (secondVariableDirectionalDerivativeFunction K u v) ∧ ∀ v' : Fin n → ℝ, saddleLowerSemicontinuousHull (secondVariableDirectionalDerivativeFunction K u v) v' = supportFunctionOfSet (partialSubdifferentialInSecondVariable K u v) v'

Lean source

Helper for Text 35.6.7: the textbook second-variable directional-derivative function is exactly the ordinary upper directional derivative of the convex slice v' ↦ K u v' at v.

open scoped Pointwise lemma helperForText_35_6_7_secondVariableDirectionalDerivative_eq_upperDirectionalDerivative {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) : secondVariableDirectionalDerivativeFunction K u v = upperDirectionalDerivativeAt (K u) v

Lean source

Helper for Text 35.6.7: membership in the Euclidean subdifferential of the convex second slice v' ↦ K u v' is exactly membership in the textbook second partial subdifferential ∂₂ K(u, v).

open scoped Pointwise lemma helperForText_35_6_7_secondSliceSubgradient_iff_partialSecondMem {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {vStar : Fin n → ℝ} : dotProductEquiv ℝ (Fin n) vStar ∈ subdifferentialAt (K u) v ↔ vStar ∈ partialSubdifferentialInSecondVariable K u v

Lean source

Helper for Text 35.6.7: the Euclidean subdifferential of the slice v' ↦ K u v' at v matches the textbook second partial subdifferential ∂₂ K(u, v).

open scoped Pointwise lemma helperForText_35_6_7_partialSecond_eq_sliceSubdifferential {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} : ((dotProductEquiv ℝ (Fin n)) ⁻¹' subdifferentialAt (K u) v) = partialSubdifferentialInSecondVariable K u v

Lean source

Helper for Text 35.6.7: after identifying the slice subdifferential with ∂₂ K(u, v), the Chapter 23 support value is exactly the textbook support function of the second partial subdifferential.

open scoped Pointwise lemma helperForText_35_6_7_sliceSupport_eq_secondPartialSupport {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} : subdifferentialSupportAt (K u) v = supportFunctionOfSet (partialSubdifferentialInSecondVariable K u v)

Lean source

Helper for Text 35.6.7: nonemptiness of the textbook second partial subdifferential is equivalent to nonemptiness of the Euclidean subdifferential of the convex second slice.

open scoped Pointwise lemma helperForText_35_6_7_partialSecond_nonempty_iff_sliceSubdifferential_nonempty {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} : Set.Nonempty (partialSubdifferentialInSecondVariable K u v) ↔ Set.Nonempty (subdifferentialAt (K u) v)

Lean source

Helper for Text 35.6.7: emptiness of the textbook second partial subdifferential is equivalent to emptiness of the Euclidean subdifferential of the convex second slice.

open scoped Pointwise lemma helperForText_35_6_7_partialSecond_empty_iff_sliceSubdifferential_empty {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} : partialSubdifferentialInSecondVariable K u v = ∅ ↔ subdifferentialAt (K u) v = ∅

Lean source

Helper for Text 35.6.7: if ∂₂ K(u, v) is empty, then Theorem 23.3 produces a direction in which the textbook directional derivative is ⊥, while the opposite direction is ⊤.

open scoped Pointwise lemma helperForText_35_6_7_exists_bot_and_top_direction_of_empty_partialSecond {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartialEmpty : partialSubdifferentialInSecondVariable K u v = ∅) : ∃ w : Fin n → ℝ, secondVariableDirectionalDerivativeFunction K u v w = (⊥ : EReal) ∧ secondVariableDirectionalDerivativeFunction K u v (-w) = (⊤ : EReal)

Lean source

Helper for Text 35.6.7: on the branch where ∂₂ K(u, v) is nonempty, the textbook lower-semicontinuous hull already matches the textbook support function.

open scoped Pointwise lemma helperForText_35_6_7_saddleLowerHull_eq_secondPartialSupport_of_nonempty_partialSecond {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartial : Set.Nonempty (partialSubdifferentialInSecondVariable K u v)) : saddleLowerSemicontinuousHull (secondVariableDirectionalDerivativeFunction K u v) = supportFunctionOfSet (partialSubdifferentialInSecondVariable K u v)

Lean source

Helper for Text 35.6.7: if ∂₂ K(u, v) is empty, then its textbook support function is the constant ⊥ function.

open scoped Pointwise lemma helperForText_35_6_7_secondPartialSupport_eq_bot_of_empty_partialSecond {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hpartialEmpty : partialSubdifferentialInSecondVariable K u v = ∅) : supportFunctionOfSet (partialSubdifferentialInSecondVariable K u v) = fun _ => (⊥ : EReal)

Lean source

Helper for Text 35.6.7: Theorem 23.2 identifies the convex closure of the textbook second-variable directional derivative with the support function of ∂₂ K(u, v). This is the full dependency-closed Chapter 23 conclusion available before upgrading to the stronger lower-semicontinuous hull.

open scoped Pointwise lemma helperForText_35_6_7_convexFunctionClosure_eq_secondPartialSupport {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) : convexFunctionClosure (secondVariableDirectionalDerivativeFunction K u v) = supportFunctionOfSet (partialSubdifferentialInSecondVariable K u v)

Lean source

Helper for Text 35.6.7: when ∂₂ K(u, v) is empty, the correct dependency-closed Chapter 23 conclusion is that the convex closure of the textbook directional-derivative function is constantly ⊥. This still falls short of the stronger hull identity used in the blocked branch.

open scoped Pointwise lemma helperForText_35_6_7_convexFunctionClosure_eq_bot_of_empty_partialSecond {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartialEmpty : partialSubdifferentialInSecondVariable K u v = ∅) : convexFunctionClosure (secondVariableDirectionalDerivativeFunction K u v) = fun _ => (⊥ : EReal)

Lean source

Helper for Text 35.6.7: the saddle lower-semicontinuous hull of the textbook directional-derivative slice ψ is exactly the ordinary lower semicontinuous hull.

open scoped Pointwise lemma helperForText_35_6_7_saddleLowerHull_eq_lowerSemicontinuousHull {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} : saddleLowerSemicontinuousHull (secondVariableDirectionalDerivativeFunction K u v) = lowerSemicontinuousHull (secondVariableDirectionalDerivativeFunction K u v)

Lean source

Helper for Text 35.6.7: on the empty ∂₂ K(u, v) branch, the textbook target at a fixed direction v' is equivalent to the missing upgrade from the saddle epigraph hull to the Chapter 23 convex closure.

open scoped Pointwise lemma helperForText_35_6_7_emptyPartial_goal_iff_hull_eq_convexClosure {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartialEmpty : partialSubdifferentialInSecondVariable K u v = ∅) (v' : Fin n → ℝ) : saddleLowerSemicontinuousHull (secondVariableDirectionalDerivativeFunction K u v) v' = supportFunctionOfSet (partialSubdifferentialInSecondVariable K u v) v' ↔ saddleLowerSemicontinuousHull (secondVariableDirectionalDerivativeFunction K u v) v' = convexFunctionClosure (secondVariableDirectionalDerivativeFunction K u v) v'

Lean source

Helper for Text 35.6.7: after rewriting the saddle hull as the ordinary lower semicontinuous hull, the empty-∂₂ K(u, v) branch of the textbook target at a fixed direction v' is equivalent to the missing lower-hull/convex-closure upgrade.

open scoped Pointwise lemma helperForText_35_6_7_emptyPartial_goal_iff_lowerHull_eq_convexClosure {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartialEmpty : partialSubdifferentialInSecondVariable K u v = ∅) (v' : Fin n → ℝ) : saddleLowerSemicontinuousHull (secondVariableDirectionalDerivativeFunction K u v) v' = supportFunctionOfSet (partialSubdifferentialInSecondVariable K u v) v' ↔ lowerSemicontinuousHull (secondVariableDirectionalDerivativeFunction K u v) v' = convexFunctionClosure (secondVariableDirectionalDerivativeFunction K u v) v'

Lean source

Helper for Text 35.6.7: the ordinary lower semicontinuous hull is exactly the Section 33 raw local-closure operator functionConvexClosure.

open scoped Pointwise lemma helperForText_35_6_7_lowerSemicontinuousHull_eq_functionConvexClosure {n : ℕ} (f : (Fin n → ℝ) → EReal) : lowerSemicontinuousHull f = functionConvexClosure f

Lean source

Helper for Text 35.6.7: for positive weights, epigraph convexity already yields Jensen's inequality whenever both endpoint values avoid ⊥; the only unresolved branch is the genuine mixed (⊥, ⊤) case.

open scoped Pointwise lemma helperForText_35_6_7_convexFunction_jensen_of_positiveWeights_and_endpointNeBot {n : ℕ} {D : (Fin n → ℝ) → EReal} (hconvD : ConvexFunction D) {x y : Fin n → ℝ} {a b : ℝ} (hPosA : 0 < a) (hPosB : 0 < b) (hab : a + b = 1) (hxBot : D x ≠ (⊥ : EReal)) (hyBot : D y ≠ (⊥ : EReal)) : D (a • x + b • y) ≤ (a : EReal) * D x + (b : EReal) * D y

Lean source

Helper for Text 35.6.7: once the Section 33 raw closure is Jensen-convex and already attains ⊥ at one point, the mixed (⊥, ⊤) collapse forbids any ⊤ value, so the raw closure is the constant ⊥ function.

open scoped Pointwise lemma helperForText_35_6_7_functionConvexClosure_eq_bot_of_rawClosureConvex_and_rawBotPoint {n : ℕ} {f : (Fin n → ℝ) → EReal} {x0 : Fin n → ℝ} (hRawConv : IsERealConvexOn (Set.univ : Set (Fin n → ℝ)) (functionConvexClosure f)) (hx0Bot : functionConvexClosure f x0 = (⊥ : EReal)) : functionConvexClosure f = fun _ => (⊥ : EReal)

Lean source

Helper for Text 35.6.7: if the effective domain of the convex second slice v' ↦ K u v' is dense, then the empty-∂₂ K(u, v) branch forces the ordinary lower semicontinuous hull of ψ to collapse to the constant ⊥ function.

open scoped Pointwise lemma helperForText_35_6_7_lowerHull_eq_bot_of_empty_partialSecond_of_dense_secondSliceDomain {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartialEmpty : partialSubdifferentialInSecondVariable K u v = ∅) (hDense : closure (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (K u)) = Set.univ) : lowerSemicontinuousHull (secondVariableDirectionalDerivativeFunction K u v) = fun _ => (⊥ : EReal)

Lean source

Helper for Text 35.6.7: with dense effective domain for the second slice, the empty-`∂₂ K(u, v)branch upgrades all the way from the ordinary lower semicontinuous hull ofψ` to its Chapter 23 convex closure.

open scoped Pointwise lemma helperForText_35_6_7_lowerHull_eq_convexClosure_of_empty_partialSecond_of_dense_secondSliceDomain {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartialEmpty : partialSubdifferentialInSecondVariable K u v = ∅) (hDense : closure (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (K u)) = Set.univ) (v' : Fin n → ℝ) : lowerSemicontinuousHull (secondVariableDirectionalDerivativeFunction K u v) v' = convexFunctionClosure (secondVariableDirectionalDerivativeFunction K u v) v'

§35 — Continuity and Differentiability of Saddle-Functions — Part 11

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Text 35.6.8: assume K is proper and (u, v) lies in the interior of dom K. Then u ∈ interior (dom K(·, v)) and v ∈ interior (dom K(u, ·)), where these interiors are taken in the Euclidean topologies of ℝ^m and ℝ^n, respectively.

open scoped Pointwise -- Proof sketch: view `K` as the proper `EReal`-valued function on `ℝ^m × ℝ^n` given by -- `Function.uncurry K`. An interior point of `dom K` has an open product neighborhood contained in -- that domain, and projecting this neighborhood to the first and second coordinates yields open -- neighborhoods proving that `u` and `v` lie in the interiors of the corresponding slice domains. theorem section35_text35_6_8 {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hProper : ProperERealFunction (Function.uncurry K)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hInterior : (u, v) ∈ interior (erealDom (Function.uncurry K))) : u ∈ interior (erealDom fun u' : Fin m → ℝ => K u' v) ∧ v ∈ interior (erealDom fun v' : Fin n → ℝ => K u v')

Lean source

Text 35.6.9: let K : ℝ^m × ℝ^n → \overline{ℝ} be a proper saddle-function, and let (u, v) ∈ interior (dom K). Then the partial subdifferentials ∂₁ K(u, v) ⊆ ℝ^m and ∂₂ K(u, v) ⊆ ℝ^n are nonempty, closed, bounded, convex sets.

open scoped Pointwise -- Proof sketch: view `K` as the proper `EReal`-valued function on `ℝ^m × ℝ^n` given by -- `Function.uncurry K`. An interior point of `dom K` has an open product neighborhood contained in -- that domain, and projecting this neighborhood to the first and second coordinates yields open -- neighborhoods proving that `u` and `v` lie in the interiors of the corresponding slice domains. theorem section35_text35_6_9 {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) (hProper : ProperERealFunction (Function.uncurry K)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hInterior : (u, v) ∈ interior (erealDom (Function.uncurry K))) : Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ IsClosed (partialSubdifferentialInFirstVariable K u v) ∧ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v) ∧ Convex ℝ (partialSubdifferentialInFirstVariable K u v) ∧ Set.Nonempty (partialSubdifferentialInSecondVariable K u v) ∧ IsClosed (partialSubdifferentialInSecondVariable K u v) ∧ Bornology.IsBounded (partialSubdifferentialInSecondVariable K u v) ∧ Convex ℝ (partialSubdifferentialInSecondVariable K u v)

Lean source

Helper for Text 35.6.8: an interior point of the product effective domain admits an open rectangle neighborhood contained in that domain.

open scoped Pointwise -- Proof sketch: view `K` as the proper `EReal`-valued function on `ℝ^m × ℝ^n` given by -- `Function.uncurry K`. An interior point of `dom K` has an open product neighborhood contained in -- that domain, and projecting this neighborhood to the first and second coordinates yields open -- neighborhoods proving that `u` and `v` lie in the interiors of the corresponding slice domains. lemma helperForText_35_6_8_openRectangleInsideProductDomain {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hInterior : (u, v) ∈ interior (erealDom (Function.uncurry K))) : ∃ U : Set (Fin m → ℝ), ∃ V : Set (Fin n → ℝ), IsOpen U ∧ u ∈ U ∧ IsOpen V ∧ v ∈ V ∧ Set.prod U V ⊆ erealDom (Function.uncurry K)

Lean source

Helper for Text 35.6.8: a rectangle inside dom K gives a first-coordinate open witness for the fixed-second-coordinate slice domain.

open scoped Pointwise -- Proof sketch: view `K` as the proper `EReal`-valued function on `ℝ^m × ℝ^n` given by -- `Function.uncurry K`. An interior point of `dom K` has an open product neighborhood contained in -- that domain, and projecting this neighborhood to the first and second coordinates yields open -- neighborhoods proving that `u` and `v` lie in the interiors of the corresponding slice domains. lemma helperForText_35_6_8_firstSliceDomain_of_rectangle {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {v : Fin n → ℝ} {U : Set (Fin m → ℝ)} {V : Set (Fin n → ℝ)} (hRect : Set.prod U V ⊆ erealDom (Function.uncurry K)) (hv : v ∈ V) : U ⊆ erealDom (fun u' : Fin m → ℝ => K u' v)

Lean source

Helper for Text 35.6.8: a rectangle inside dom K gives a second-coordinate open witness for the fixed-first-coordinate slice domain.

open scoped Pointwise -- Proof sketch: view `K` as the proper `EReal`-valued function on `ℝ^m × ℝ^n` given by -- `Function.uncurry K`. An interior point of `dom K` has an open product neighborhood contained in -- that domain, and projecting this neighborhood to the first and second coordinates yields open -- neighborhoods proving that `u` and `v` lie in the interiors of the corresponding slice domains. lemma helperForText_35_6_8_secondSliceDomain_of_rectangle {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {U : Set (Fin m → ℝ)} {V : Set (Fin n → ℝ)} (hRect : Set.prod U V ⊆ erealDom (Function.uncurry K)) (hu : u ∈ U) : V ⊆ erealDom (fun v' : Fin n → ℝ => K u v')

Lean source

Helper for Text 35.6.9: properness rules out ⊥ everywhere, so the reflected first slice x ↦ -K (-x) v has full effective domain. In particular, -u is an interior point of that effective domain.

open scoped Pointwise -- Proof sketch: view `K` as the proper `EReal`-valued function on `ℝ^m × ℝ^n` given by -- `Function.uncurry K`. An interior point of `dom K` has an open product neighborhood contained in -- that domain, and projecting this neighborhood to the first and second coordinates yields open -- neighborhoods proving that `u` and `v` lie in the interiors of the corresponding slice domains. lemma helperForText_35_6_9_reflectedFirstSlice_mem_interior_effectiveDomain {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hProper : ProperERealFunction (Function.uncurry K)) (u : Fin m → ℝ) (v : Fin n → ℝ) : -u ∈ interior (effectiveDomain (Set.univ : Set (Fin m → ℝ)) (fun x : Fin m → ℝ => -K (-x) v))

Lean source

Helper for Text 35.6.9: the first partial subdifferential is nonempty, closed, bounded, and convex once (u, v) lies in the interior of dom K.

open scoped Pointwise -- Proof sketch: view `K` as the proper `EReal`-valued function on `ℝ^m × ℝ^n` given by -- `Function.uncurry K`. An interior point of `dom K` has an open product neighborhood contained in -- that domain, and projecting this neighborhood to the first and second coordinates yields open -- neighborhoods proving that `u` and `v` lie in the interiors of the corresponding slice domains. lemma helperForText_35_6_9_firstPartial_nonempty_closed_bounded_convex {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) (hProper : ProperERealFunction (Function.uncurry K)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hInterior : (u, v) ∈ interior (erealDom (Function.uncurry K))) : Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ IsClosed (partialSubdifferentialInFirstVariable K u v) ∧ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v) ∧ Convex ℝ (partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.9: the second partial subdifferential is nonempty, closed, bounded, and convex once (u, v) lies in the interior of dom K.

open scoped Pointwise -- Proof sketch: view `K` as the proper `EReal`-valued function on `ℝ^m × ℝ^n` given by -- `Function.uncurry K`. An interior point of `dom K` has an open product neighborhood contained in -- that domain, and projecting this neighborhood to the first and second coordinates yields open -- neighborhoods proving that `u` and `v` lie in the interiors of the corresponding slice domains. lemma helperForText_35_6_9_secondPartial_nonempty_closed_bounded_convex {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) (hProper : ProperERealFunction (Function.uncurry K)) {u : Fin m → ℝ} {v : Fin n → ℝ} (hInterior : (u, v) ∈ interior (erealDom (Function.uncurry K))) : Set.Nonempty (partialSubdifferentialInSecondVariable K u v) ∧ IsClosed (partialSubdifferentialInSecondVariable K u v) ∧ Bornology.IsBounded (partialSubdifferentialInSecondVariable K u v) ∧ Convex ℝ (partialSubdifferentialInSecondVariable K u v)

§35 — Continuity and Differentiability of Saddle-Functions — Part 12

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Helper for Text 35.6.10: at a finite base point, the zero-direction saddle difference quotient is constantly 0, so the corresponding first-variable directional derivative value is 0.

open scoped Pointwise -- Proof sketch: apply Theorem 23.4 to the proper convex slice `u₁ ↦ -K u₁ v` at the finite -- point `u`. The directional derivative of this slice in direction `u'` is -- `-K'(u, v; u', 0)`, while its subdifferential identifies with the negatives of the first -- partial subgradients of `K` at `(u, v)`. Rewriting the support-function formula from Theorem -- 23.4 for this negated slice gives the desired infimum formula. -- Route correction: the textbook all-finite-points statement is false on the branch -- `partialSubdifferentialInFirstVariable K u v = ∅`, because the zero-direction derivative -- infimum is `0` while the pairing infimum is `sInf ∅ = ⊤`. The formalized theorem therefore -- records the valid nonempty-bounded branch isolated by Theorem 23.4. lemma helperForText_35_6_10_zeroDirectionDerivative_eq_zero {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) : sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v 0 0 L} = 0

Lean source

Helper for Text 35.6.10: if the first partial subdifferential is empty, then the infimum of its first-variable pairing image is ⊤ because the image set itself is empty.

open scoped Pointwise -- Proof sketch: apply Theorem 23.4 to the proper convex slice `u₁ ↦ -K u₁ v` at the finite -- point `u`. The directional derivative of this slice in direction `u'` is -- `-K'(u, v; u', 0)`, while its subdifferential identifies with the negatives of the first -- partial subgradients of `K` at `(u, v)`. Rewriting the support-function formula from Theorem -- 23.4 for this negated slice gives the desired infimum formula. -- Route correction: the textbook all-finite-points statement is false on the branch -- `partialSubdifferentialInFirstVariable K u v = ∅`, because the zero-direction derivative -- infimum is `0` while the pairing infimum is `sInf ∅ = ⊤`. The formalized theorem therefore -- records the valid nonempty-bounded branch isolated by Theorem 23.4. lemma helperForText_35_6_10_pairingInf_eq_top_of_empty_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (u' : Fin m → ℝ) (hpartialEmpty : partialSubdifferentialInFirstVariable K u v = ∅) : sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * u' i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v) = ⊤

Lean source

Helper for Text 35.6.10: if the first partial subdifferential is nonempty, then the pairing-infimum cannot be ⊤ because the image set contains a genuine real value.

open scoped Pointwise -- Proof sketch: apply Theorem 23.4 to the proper convex slice `u₁ ↦ -K u₁ v` at the finite -- point `u`. The directional derivative of this slice in direction `u'` is -- `-K'(u, v; u', 0)`, while its subdifferential identifies with the negatives of the first -- partial subgradients of `K` at `(u, v)`. Rewriting the support-function formula from Theorem -- 23.4 for this negated slice gives the desired infimum formula. -- Route correction: the textbook all-finite-points statement is false on the branch -- `partialSubdifferentialInFirstVariable K u v = ∅`, because the zero-direction derivative -- infimum is `0` while the pairing infimum is `sInf ∅ = ⊤`. The formalized theorem therefore -- records the valid nonempty-bounded branch isolated by Theorem 23.4. lemma helperForText_35_6_10_pairingInf_ne_top_of_nonempty_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (u' : Fin m → ℝ) (hpartial : Set.Nonempty (partialSubdifferentialInFirstVariable K u v)) : sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * u' i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v) ≠ (⊤ : EReal)

Lean source

Helper for Text 35.6.10: the pairing-infimum equals ⊤ if and only if the first partial subdifferential is empty, matching the textbook convention inf ∅ = +∞.

open scoped Pointwise -- Proof sketch: apply Theorem 23.4 to the proper convex slice `u₁ ↦ -K u₁ v` at the finite -- point `u`. The directional derivative of this slice in direction `u'` is -- `-K'(u, v; u', 0)`, while its subdifferential identifies with the negatives of the first -- partial subgradients of `K` at `(u, v)`. Rewriting the support-function formula from Theorem -- 23.4 for this negated slice gives the desired infimum formula. -- Route correction: the textbook all-finite-points statement is false on the branch -- `partialSubdifferentialInFirstVariable K u v = ∅`, because the zero-direction derivative -- infimum is `0` while the pairing infimum is `sInf ∅ = ⊤`. The formalized theorem therefore -- records the valid nonempty-bounded branch isolated by Theorem 23.4. lemma helperForText_35_6_10_pairingInf_eq_top_iff_partialFirst_empty {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (u' : Fin m → ℝ) : (sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * u' i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v) = (⊤ : EReal)) ↔ partialSubdifferentialInFirstVariable K u v = ∅

Lean source

Helper for Text 35.6.10: on the empty-first-partial branch, the zero-direction derivative infimum is 0 while the corresponding pairing infimum is ⊤.

open scoped Pointwise -- Proof sketch: apply Theorem 23.4 to the proper convex slice `u₁ ↦ -K u₁ v` at the finite -- point `u`. The directional derivative of this slice in direction `u'` is -- `-K'(u, v; u', 0)`, while its subdifferential identifies with the negatives of the first -- partial subgradients of `K` at `(u, v)`. Rewriting the support-function formula from Theorem -- 23.4 for this negated slice gives the desired infimum formula. -- Route correction: the textbook all-finite-points statement is false on the branch -- `partialSubdifferentialInFirstVariable K u v = ∅`, because the zero-direction derivative -- infimum is `0` while the pairing infimum is `sInf ∅ = ⊤`. The formalized theorem therefore -- records the valid nonempty-bounded branch isolated by Theorem 23.4. lemma helperForText_35_6_10_zeroDirectionValues_of_empty_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartialEmpty : partialSubdifferentialInFirstVariable K u v = ∅) : sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v 0 0 L} = 0 ∧ sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * (0 : Fin m → ℝ) i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v) = ⊤

Lean source

Helper for Text 35.6.10: the claimed infimum formula already fails in the zero direction on the branch where the first partial subdifferential is empty.

open scoped Pointwise -- Proof sketch: apply Theorem 23.4 to the proper convex slice `u₁ ↦ -K u₁ v` at the finite -- point `u`. The directional derivative of this slice in direction `u'` is -- `-K'(u, v; u', 0)`, while its subdifferential identifies with the negatives of the first -- partial subgradients of `K` at `(u, v)`. Rewriting the support-function formula from Theorem -- 23.4 for this negated slice gives the desired infimum formula. -- Route correction: the textbook all-finite-points statement is false on the branch -- `partialSubdifferentialInFirstVariable K u v = ∅`, because the zero-direction derivative -- infimum is `0` while the pairing infimum is `sInf ∅ = ⊤`. The formalized theorem therefore -- records the valid nonempty-bounded branch isolated by Theorem 23.4. lemma helperForText_35_6_10_zeroDirection_formula_false_of_empty_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartialEmpty : partialSubdifferentialInFirstVariable K u v = ∅) : sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v 0 0 L} ≠ sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * (0 : Fin m → ℝ) i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.10: if the zero-direction infimum identity held at a finite base point, then the first partial subdifferential could not be empty.

open scoped Pointwise -- Proof sketch: apply Theorem 23.4 to the proper convex slice `u₁ ↦ -K u₁ v` at the finite -- point `u`. The directional derivative of this slice in direction `u'` is -- `-K'(u, v; u', 0)`, while its subdifferential identifies with the negatives of the first -- partial subgradients of `K` at `(u, v)`. Rewriting the support-function formula from Theorem -- 23.4 for this negated slice gives the desired infimum formula. -- Route correction: the textbook all-finite-points statement is false on the branch -- `partialSubdifferentialInFirstVariable K u v = ∅`, because the zero-direction derivative -- infimum is `0` while the pairing infimum is `sInf ∅ = ⊤`. The formalized theorem therefore -- records the valid nonempty-bounded branch isolated by Theorem 23.4. lemma helperForText_35_6_10_zeroDirectionFormula_forces_nonempty_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hFormula : sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v 0 0 L} = sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * (0 : Fin m → ℝ) i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v)) : Set.Nonempty (partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.10: at a finite base point, the zero-direction infimum formula holds exactly when the first partial subdifferential is nonempty.

open scoped Pointwise -- Proof sketch: apply Theorem 23.4 to the proper convex slice `u₁ ↦ -K u₁ v` at the finite -- point `u`. The directional derivative of this slice in direction `u'` is -- `-K'(u, v; u', 0)`, while its subdifferential identifies with the negatives of the first -- partial subgradients of `K` at `(u, v)`. Rewriting the support-function formula from Theorem -- 23.4 for this negated slice gives the desired infimum formula. -- Route correction: the textbook all-finite-points statement is false on the branch -- `partialSubdifferentialInFirstVariable K u v = ∅`, because the zero-direction derivative -- infimum is `0` while the pairing infimum is `sInf ∅ = ⊤`. The formalized theorem therefore -- records the valid nonempty-bounded branch isolated by Theorem 23.4. lemma helperForText_35_6_10_zeroDirectionFormula_iff_nonempty_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) : (sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v 0 0 L} = sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * (0 : Fin m → ℝ) i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v)) ↔ Set.Nonempty (partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.10: on the branch where ∂₁ K(u, v) is not simultaneously nonempty and bounded, the zero-direction identity survives exactly in the nonempty-unbounded case.

open scoped Pointwise -- Proof sketch: apply Theorem 23.4 to the proper convex slice `u₁ ↦ -K u₁ v` at the finite -- point `u`. The directional derivative of this slice in direction `u'` is -- `-K'(u, v; u', 0)`, while its subdifferential identifies with the negatives of the first -- partial subgradients of `K` at `(u, v)`. Rewriting the support-function formula from Theorem -- 23.4 for this negated slice gives the desired infimum formula. -- Route correction: the textbook all-finite-points statement is false on the branch -- `partialSubdifferentialInFirstVariable K u v = ∅`, because the zero-direction derivative -- infimum is `0` while the pairing infimum is `sInf ∅ = ⊤`. The formalized theorem therefore -- records the valid nonempty-bounded branch isolated by Theorem 23.4. lemma helperForText_35_6_10_zeroDirectionFormula_on_badBranch_iff_nonempty_unbounded_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hBad : ¬ (Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v))) : (sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v 0 0 L} = sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * (0 : Fin m → ℝ) i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v)) ↔ Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ ¬ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.10: if the good nonempty-bounded branch fails, then the only remaining possibilities are the empty branch and the nonempty-unbounded branch.

open scoped Pointwise -- Proof sketch: apply Theorem 23.4 to the proper convex slice `u₁ ↦ -K u₁ v` at the finite -- point `u`. The directional derivative of this slice in direction `u'` is -- `-K'(u, v; u', 0)`, while its subdifferential identifies with the negatives of the first -- partial subgradients of `K` at `(u, v)`. Rewriting the support-function formula from Theorem -- 23.4 for this negated slice gives the desired infimum formula. -- Route correction: the textbook all-finite-points statement is false on the branch -- `partialSubdifferentialInFirstVariable K u v = ∅`, because the zero-direction derivative -- infimum is `0` while the pairing infimum is `sInf ∅ = ⊤`. The formalized theorem therefore -- records the valid nonempty-bounded branch isolated by Theorem 23.4. lemma helperForText_35_6_10_badBranch_cases {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hBad : ¬ (Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v))) : partialSubdifferentialInFirstVariable K u v = ∅ ∨ (Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ ¬ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v))

Lean source

Helper for Text 35.6.10: on the bad branch, either the zero direction already contradicts the textbook formula or one is forced into the remaining nonempty-unbounded case.

open scoped Pointwise -- Proof sketch: apply Theorem 23.4 to the proper convex slice `u₁ ↦ -K u₁ v` at the finite -- point `u`. The directional derivative of this slice in direction `u'` is -- `-K'(u, v; u', 0)`, while its subdifferential identifies with the negatives of the first -- partial subgradients of `K` at `(u, v)`. Rewriting the support-function formula from Theorem -- 23.4 for this negated slice gives the desired infimum formula. -- Route correction: the textbook all-finite-points statement is false on the branch -- `partialSubdifferentialInFirstVariable K u v = ∅`, because the zero-direction derivative -- infimum is `0` while the pairing infimum is `sInf ∅ = ⊤`. The formalized theorem therefore -- records the valid nonempty-bounded branch isolated by Theorem 23.4. lemma helperForText_35_6_10_zeroDirection_failure_or_nonempty_unbounded_of_badBranch {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hBad : ¬ (Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v))) : (sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v 0 0 L} ≠ sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * (0 : Fin m → ℝ) i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v)) ∨ (Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ ¬ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v))

Lean source

Helper for Text 35.6.10: if ∂₁ K(u, v) is empty, then the failure of the textbook formula is witnessed by the concrete first direction u' = 0.

open scoped Pointwise -- Proof sketch: apply Theorem 23.4 to the proper convex slice `u₁ ↦ -K u₁ v` at the finite -- point `u`. The directional derivative of this slice in direction `u'` is -- `-K'(u, v; u', 0)`, while its subdifferential identifies with the negatives of the first -- partial subgradients of `K` at `(u, v)`. Rewriting the support-function formula from Theorem -- 23.4 for this negated slice gives the desired infimum formula. -- Route correction: the textbook all-finite-points statement is false on the branch -- `partialSubdifferentialInFirstVariable K u v = ∅`, because the zero-direction derivative -- infimum is `0` while the pairing infimum is `sInf ∅ = ⊤`. The formalized theorem therefore -- records the valid nonempty-bounded branch isolated by Theorem 23.4. lemma helperForText_35_6_10_counterexampleDirection_of_empty_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartialEmpty : partialSubdifferentialInFirstVariable K u v = ∅) : ∃ u' : Fin m → ℝ, sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v u' 0 L} ≠ sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * u' i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.10: if ∂₁ K(u, v) is empty, then the full textbook formula cannot hold for every first direction, because the zero direction already contradicts it.

open scoped Pointwise -- Proof sketch: apply Theorem 23.4 to the proper convex slice `u₁ ↦ -K u₁ v` at the finite -- point `u`. The directional derivative of this slice in direction `u'` is -- `-K'(u, v; u', 0)`, while its subdifferential identifies with the negatives of the first -- partial subgradients of `K` at `(u, v)`. Rewriting the support-function formula from Theorem -- 23.4 for this negated slice gives the desired infimum formula. -- Route correction: the textbook all-finite-points statement is false on the branch -- `partialSubdifferentialInFirstVariable K u v = ∅`, because the zero-direction derivative -- infimum is `0` while the pairing infimum is `sInf ∅ = ⊤`. The formalized theorem therefore -- records the valid nonempty-bounded branch isolated by Theorem 23.4. lemma helperForText_35_6_10_not_allDirectionsFormula_of_empty_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartialEmpty : partialSubdifferentialInFirstVariable K u v = ∅) : ¬ ∀ u' : Fin m → ℝ, sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v u' 0 L} = sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * u' i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.10: if ∂₁ K(u, v) is empty, then the reflected convex slice has a pair of opposite first directions whose directional-derivative infima are ⊥ and ⊤.

open scoped Pointwise -- Proof sketch: apply Theorem 23.4 to the proper convex slice `u₁ ↦ -K u₁ v` at the finite -- point `u`. The directional derivative of this slice in direction `u'` is -- `-K'(u, v; u', 0)`, while its subdifferential identifies with the negatives of the first -- partial subgradients of `K` at `(u, v)`. Rewriting the support-function formula from Theorem -- 23.4 for this negated slice gives the desired infimum formula. -- Route correction: the textbook all-finite-points statement is false on the branch -- `partialSubdifferentialInFirstVariable K u v = ∅`, because the zero-direction derivative -- infimum is `0` while the pairing infimum is `sInf ∅ = ⊤`. The formalized theorem therefore -- records the valid nonempty-bounded branch isolated by Theorem 23.4. lemma helperForText_35_6_10_exists_bot_and_top_direction_of_empty_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartialEmpty : partialSubdifferentialInFirstVariable K u v = ∅) : ∃ w : Fin m → ℝ, sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v w 0 L} = (⊥ : EReal) ∧ sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v (-w) 0 L} = (⊤ : EReal)

Lean source

Helper for Text 35.6.10: if ∂₁ K(u, v) is empty, then some first direction has directional-derivative infimum ⊥ while the pairing infimum over ∂₁ K(u, v) is ⊤.

open scoped Pointwise -- Proof sketch: apply Theorem 23.4 to the proper convex slice `u₁ ↦ -K u₁ v` at the finite -- point `u`. The directional derivative of this slice in direction `u'` is -- `-K'(u, v; u', 0)`, while its subdifferential identifies with the negatives of the first -- partial subgradients of `K` at `(u, v)`. Rewriting the support-function formula from Theorem -- 23.4 for this negated slice gives the desired infimum formula. -- Route correction: the textbook all-finite-points statement is false on the branch -- `partialSubdifferentialInFirstVariable K u v = ∅`, because the zero-direction derivative -- infimum is `0` while the pairing infimum is `sInf ∅ = ⊤`. The formalized theorem therefore -- records the valid nonempty-bounded branch isolated by Theorem 23.4. lemma helperForText_35_6_10_exists_direction_with_botDerivative_and_topPairing_of_empty_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartialEmpty : partialSubdifferentialInFirstVariable K u v = ∅) : ∃ u' : Fin m → ℝ, sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v u' 0 L} = (⊥ : EReal) ∧ sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * u' i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v) = (⊤ : EReal)

Lean source

Helper for Text 35.6.10: if ∂₁ K(u, v) is empty, then the textbook formula fails in a first direction where the directional-derivative infimum is ⊥ while the pairing infimum is still ⊤.

open scoped Pointwise -- Proof sketch: apply Theorem 23.4 to the proper convex slice `u₁ ↦ -K u₁ v` at the finite -- point `u`. The directional derivative of this slice in direction `u'` is -- `-K'(u, v; u', 0)`, while its subdifferential identifies with the negatives of the first -- partial subgradients of `K` at `(u, v)`. Rewriting the support-function formula from Theorem -- 23.4 for this negated slice gives the desired infimum formula. -- Route correction: the textbook all-finite-points statement is false on the branch -- `partialSubdifferentialInFirstVariable K u v = ∅`, because the zero-direction derivative -- infimum is `0` while the pairing infimum is `sInf ∅ = ⊤`. The formalized theorem therefore -- records the valid nonempty-bounded branch isolated by Theorem 23.4. lemma helperForText_35_6_10_counterexampleDirection_of_empty_partialFirst_via_infiniteDerivative {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartialEmpty : partialSubdifferentialInFirstVariable K u v = ∅) : ∃ u' : Fin m → ℝ, sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v u' 0 L} ≠ sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * u' i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.10: if the textbook formula held for every first direction at a finite base point, then the first partial subdifferential would have to be nonempty.

open scoped Pointwise -- Proof sketch: apply Theorem 23.4 to the proper convex slice `u₁ ↦ -K u₁ v` at the finite -- point `u`. The directional derivative of this slice in direction `u'` is -- `-K'(u, v; u', 0)`, while its subdifferential identifies with the negatives of the first -- partial subgradients of `K` at `(u, v)`. Rewriting the support-function formula from Theorem -- 23.4 for this negated slice gives the desired infimum formula. -- Route correction: the textbook all-finite-points statement is false on the branch -- `partialSubdifferentialInFirstVariable K u v = ∅`, because the zero-direction derivative -- infimum is `0` while the pairing infimum is `sInf ∅ = ⊤`. The formalized theorem therefore -- records the valid nonempty-bounded branch isolated by Theorem 23.4. lemma helperForText_35_6_10_allDirectionsFormula_forces_nonempty_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hFormula : ∀ u' : Fin m → ℝ, sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v u' 0 L} = sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * u' i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v)) : Set.Nonempty (partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.10: once the remaining nonempty-unbounded branch is excluded, any all-directions instance of the textbook formula already forces the corrected nonempty-bounded branch hypothesis.

open scoped Pointwise -- Proof sketch: apply Theorem 23.4 to the proper convex slice `u₁ ↦ -K u₁ v` at the finite -- point `u`. The directional derivative of this slice in direction `u'` is -- `-K'(u, v; u', 0)`, while its subdifferential identifies with the negatives of the first -- partial subgradients of `K` at `(u, v)`. Rewriting the support-function formula from Theorem -- 23.4 for this negated slice gives the desired infimum formula. -- Route correction: the textbook all-finite-points statement is false on the branch -- `partialSubdifferentialInFirstVariable K u v = ∅`, because the zero-direction derivative -- infimum is `0` while the pairing infimum is `sInf ∅ = ⊤`. The formalized theorem therefore -- records the valid nonempty-bounded branch isolated by Theorem 23.4. lemma helperForText_35_6_10_allDirectionsFormula_forces_goodBranch_of_no_nonempty_unbounded {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hNoNonemptyUnbounded : ¬ (Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ ¬ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v))) (hFormula : ∀ u' : Fin m → ℝ, sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v u' 0 L} = sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * u' i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v)) : Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.10: on the bad branch, if the remaining nonempty-unbounded alternative is also excluded, then the full all-directions formula already fails.

open scoped Pointwise -- Proof sketch: apply Theorem 23.4 to the proper convex slice `u₁ ↦ -K u₁ v` at the finite -- point `u`. The directional derivative of this slice in direction `u'` is -- `-K'(u, v; u', 0)`, while its subdifferential identifies with the negatives of the first -- partial subgradients of `K` at `(u, v)`. Rewriting the support-function formula from Theorem -- 23.4 for this negated slice gives the desired infimum formula. -- Route correction: the textbook all-finite-points statement is false on the branch -- `partialSubdifferentialInFirstVariable K u v = ∅`, because the zero-direction derivative -- infimum is `0` while the pairing infimum is `sInf ∅ = ⊤`. The formalized theorem therefore -- records the valid nonempty-bounded branch isolated by Theorem 23.4. lemma helperForText_35_6_10_badBranch_without_nonempty_unbounded_forces_formula_failure {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hBad : ¬ (Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v))) (hNoRemaining : ¬ (Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ ¬ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v))) : ¬ ∀ u' : Fin m → ℝ, sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v u' 0 L} = sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * u' i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.10: if the full all-directions formula held on the bad branch, the bad branch would necessarily collapse to the nonempty-unbounded alternative.

open scoped Pointwise -- Proof sketch: apply Theorem 23.4 to the proper convex slice `u₁ ↦ -K u₁ v` at the finite -- point `u`. The directional derivative of this slice in direction `u'` is -- `-K'(u, v; u', 0)`, while its subdifferential identifies with the negatives of the first -- partial subgradients of `K` at `(u, v)`. Rewriting the support-function formula from Theorem -- 23.4 for this negated slice gives the desired infimum formula. -- Route correction: the textbook all-finite-points statement is false on the branch -- `partialSubdifferentialInFirstVariable K u v = ∅`, because the zero-direction derivative -- infimum is `0` while the pairing infimum is `sInf ∅ = ⊤`. The formalized theorem therefore -- records the valid nonempty-bounded branch isolated by Theorem 23.4. lemma helperForText_35_6_10_badBranch_allDirectionsFormula_forces_nonempty_unbounded {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hBad : ¬ (Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v))) (hFormula : ∀ u' : Fin m → ℝ, sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v u' 0 L} = sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * u' i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v)) : Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ ¬ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.10: on the bad branch, either the empty-partial obstruction already produces a concrete counterexample direction to the textbook formula, or the only remaining case is the nonempty-unbounded branch.

open scoped Pointwise -- Proof sketch: apply Theorem 23.4 to the proper convex slice `u₁ ↦ -K u₁ v` at the finite -- point `u`. The directional derivative of this slice in direction `u'` is -- `-K'(u, v; u', 0)`, while its subdifferential identifies with the negatives of the first -- partial subgradients of `K` at `(u, v)`. Rewriting the support-function formula from Theorem -- 23.4 for this negated slice gives the desired infimum formula. -- Route correction: the textbook all-finite-points statement is false on the branch -- `partialSubdifferentialInFirstVariable K u v = ∅`, because the zero-direction derivative -- infimum is `0` while the pairing infimum is `sInf ∅ = ⊤`. The formalized theorem therefore -- records the valid nonempty-bounded branch isolated by Theorem 23.4. lemma helperForText_35_6_10_badBranch_counterexampleDirection_or_nonempty_unbounded {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hBad : ¬ (Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v))) : (∃ u' : Fin m → ℝ, sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v u' 0 L} ≠ sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * u' i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v)) ∨ (Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ ¬ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v))

§35 — Continuity and Differentiability of Saddle-Functions — Part 13

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Text 35.6.10: in the branch where the first partial subdifferential ∂₁ K(u, v) is nonempty and bounded, the first-variable directional derivative is the infimum of the pairings ⟪u*, u'⟫ over u* ∈ ∂₁ K(u, v).

open scoped Pointwise theorem section35_text35_6_10 {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hFirstPartial : Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v)) (u' : Fin m → ℝ) : sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v u' 0 L} = sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * u' i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v)

Lean source

Text 35.6.11: in the good branch where the second partial subdifferential at (u, v) is nonempty and bounded, the directional derivative K'(u, v; 0, v') equals the support formula sup {⟪v*, v'⟫ | v* ∈ ∂₂ K(u, v)} for every direction v'. This is the branch supplied by the Chapter 23 route already developed in the file.

open scoped Pointwise theorem section35_text35_6_11 {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartial : Set.Nonempty (partialSubdifferentialInSecondVariable K u v)) (hpartialBdd : Bornology.IsBounded (partialSubdifferentialInSecondVariable K u v)) (v' : Fin n → ℝ) : sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v 0 v' L} = sSup ((fun vStar : Fin n → ℝ => (((∑ i : Fin n, vStar i * v' i) : ℝ) : EReal)) '' partialSubdifferentialInSecondVariable K u v)

Lean source

The concave subdifferential of a real-valued saddle kernel in the first variable on C at (u, v).

open scoped Pointwise def realPartialSubdifferentialInFirstVariableOn {m n : ℕ} (C : Set (Fin m → ℝ)) (K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) (u : Fin m → ℝ) (v : Fin n → ℝ) : Set (Fin m → ℝ) := {uStar | ∀ u' ∈ C, K u' v ≤ K u v + ∑ i : Fin m, uStar i * (u' i - u i)}

Lean source

The convex subdifferential of a real-valued saddle kernel in the second variable on D at (u, v).

open scoped Pointwise def realPartialSubdifferentialInSecondVariableOn {m n : ℕ} (D : Set (Fin n → ℝ)) (K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) (u : Fin m → ℝ) (v : Fin n → ℝ) : Set (Fin n → ℝ) := {vStar | ∀ v' ∈ D, K u v' ≥ K u v + ∑ i : Fin n, vStar i * (v' i - v i)}

Lean source

The bilinear pairing of a product subgradient (u*, v*) with a direction (u', v').

open scoped Pointwise def saddleSubgradientPairing {m n : ℕ} (uStar : Fin m → ℝ) (uDir : Fin m → ℝ) (vStar : Fin n → ℝ) (vDir : Fin n → ℝ) : ℝ := (∑ i : Fin m, uStar i * uDir i) + ∑ j : Fin n, vStar j * vDir j

Lean source

The first-variable directional derivative of a real-valued saddle kernel at (u, v) in the direction u', packaged as the infimum of the real numbers that realize this derivative.

open scoped Pointwise noncomputable def realFirstVariableDirectionalDerivativeValue {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) (u : Fin m → ℝ) (v : Fin n → ℝ) (u' : Fin m → ℝ) : ℝ := sInf {L : ℝ | HasRealSaddleDirectionalDerivativeAt K u v u' 0 L}

Lean source

The second-variable directional derivative of a real-valued saddle kernel at (u, v) in the direction v', packaged as the infimum of the real numbers that realize this derivative.

open scoped Pointwise noncomputable def realSecondVariableDirectionalDerivativeValue {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) (u : Fin m → ℝ) (v : Fin n → ℝ) (v' : Fin n → ℝ) : ℝ := sInf {L : ℝ | HasRealSaddleDirectionalDerivativeAt K u v 0 v' L}

Lean source

The saddle subdifferential of a real-valued saddle kernel on C × D at (u, v), written as the product of the concave and convex partial subdifferentials.

open scoped Pointwise def realSaddleSubdifferentialOn {m n : ℕ} (C : Set (Fin m → ℝ)) (D : Set (Fin n → ℝ)) (K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) (u : Fin m → ℝ) (v : Fin n → ℝ) : Set ((Fin m → ℝ) × (Fin n → ℝ)) := realPartialSubdifferentialInFirstVariableOn C K u v ×ˢ realPartialSubdifferentialInSecondVariableOn D K u v

Lean source

The closed Euclidean ball of radius ε in ℝ^m × ℝ^n, written in split coordinates.

open scoped Pointwise def splitEuclideanClosedBall {m n : ℕ} (ε : ℝ) : Set ((Fin m → ℝ) × (Fin n → ℝ)) := {p | (∑ i : Fin m, p.1 i ^ (2 : ℕ)) + ∑ j : Fin n, p.2 j ^ (2 : ℕ) ≤ ε ^ (2 : ℕ)}

Lean source

Helper for Text 35.6.10: once ∂₁ K(u, v) is nonempty and bounded, the reflected-slice Theorem 23.4 route yields the exact infimum formula claimed in the text.

open scoped Pointwise lemma helperForText_35_6_10_formula_of_nonempty_bounded_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartial : Set.Nonempty (partialSubdifferentialInFirstVariable K u v)) (hpartialBdd : Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v)) (u' : Fin m → ℝ) : sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v u' 0 L} = sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * u' i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.10: under the missing nonempty-bounded hypothesis on ∂₁ K(u, v), the textbook infimum formula does hold for every first direction.

open scoped Pointwise lemma helperForText_35_6_10_allDirectionsFormula_of_nonempty_bounded_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartial : Set.Nonempty (partialSubdifferentialInFirstVariable K u v)) (hpartialBdd : Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v)) : ∀ u' : Fin m → ℝ, sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v u' 0 L} = sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * u' i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.10: once the nonempty-unbounded remainder is excluded, the full all-directions textbook formula is equivalent to the corrected nonempty-bounded branch hypothesis on ∂₁ K(u, v).

open scoped Pointwise lemma helperForText_35_6_10_allDirectionsFormula_iff_goodBranch_of_no_nonempty_unbounded {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hNoNonemptyUnbounded : ¬ (Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ ¬ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v))) : (∀ u' : Fin m → ℝ, sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v u' 0 L} = sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * u' i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v)) ↔ (Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v))

Lean source

Helper for Text 35.6.10: the EReal-valued pairing map in the theorem statement can be rewritten using dotProduct, matching the usual ⟪u*, u'⟫ notation.

open scoped Pointwise lemma helperForText_35_6_10_sInf_pairingImage_eq_sInf_dotProductImage {m : ℕ} (C : Set (Fin m → ℝ)) (u' : Fin m → ℝ) : sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * u' i) : ℝ) : EReal)) '' C) = sInf ((fun uStar : Fin m → ℝ => (((dotProduct uStar u' : ℝ)) : EReal)) '' C)

Lean source

Helper for Text 35.6.10: restatement of section35_text35_6_10 using dotProduct on the right-hand side, for closer alignment with the textbook pairing notation.

open scoped Pointwise lemma helperForText_35_6_10_dotProductFormula {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hFirstPartial : Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v)) (u' : Fin m → ℝ) : sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v u' 0 L} = sInf ((fun uStar : Fin m → ℝ => (((dotProduct uStar u' : ℝ)) : EReal)) '' partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.10: proposition packaging the corrected all-directions formula on the good branch where ∂₁ K(u, v) is nonempty and bounded. This is the target boundary actually supported by the Lean development.

open scoped Pointwise abbrev helperForText_35_6_10_originalAllDirectionsFormula {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (v : Fin n → ℝ) : Prop := Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v) ∧ ∀ u' : Fin m → ℝ, sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v u' 0 L} = sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * u' i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.10: if the packaged original textbook formula held at a finite base point, then specializing to the zero direction forces the corresponding pairing infimum to be 0.

open scoped Pointwise lemma helperForText_35_6_10_originalAllDirectionsFormula_forces_zeroDirection_pairingInf_eq_zero {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hFormula : helperForText_35_6_10_originalAllDirectionsFormula (K := K) u v) : sInf ((fun uStar : Fin m → ℝ => (((∑ i : Fin m, uStar i * (0 : Fin m → ℝ) i) : ℝ) : EReal)) '' partialSubdifferentialInFirstVariable K u v) = 0

Lean source

Helper for Text 35.6.10: the original unrestricted textbook proposition already fails on the empty-first-partial branch, so no proof can exist there without extra hypotheses.

open scoped Pointwise lemma helperForText_35_6_10_originalAllDirectionsFormula_false_of_empty_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartialEmpty : partialSubdifferentialInFirstVariable K u v = ∅) : ¬ helperForText_35_6_10_originalAllDirectionsFormula (K := K) u v

Lean source

Helper for Text 35.6.10: one finite base point with empty first partial subdifferential already refutes the unrestricted textbook schema for the whole kernel.

open scoped Pointwise lemma helperForText_35_6_10_exists_empty_partialFirst_forces_global_originalStatement_failure {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hCounterexample : ∃ u : Fin m → ℝ, ∃ v : Fin n → ℝ, (K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) ∧ partialSubdifferentialInFirstVariable K u v = ∅) : ¬ ∀ u : Fin m → ℝ, ∀ v : Fin n → ℝ, (K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) → helperForText_35_6_10_originalAllDirectionsFormula (K := K) u v

Lean source

Helper for Text 35.6.10: once the missing nonempty-bounded hypothesis on ∂₁ K(u, v) is supplied, the corrected theorem recovers the original all-directions textbook proposition.

open scoped Pointwise lemma helperForText_35_6_10_originalAllDirectionsFormula_of_nonempty_bounded_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hFirstPartial : Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v)) : helperForText_35_6_10_originalAllDirectionsFormula (K := K) u v

Lean source

Helper for Text 35.6.10: if the packaged original unrestricted textbook formula holds at a finite base point, then the first partial subdifferential cannot be empty.

open scoped Pointwise lemma helperForText_35_6_10_originalAllDirectionsFormula_forces_nonempty_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hFormula : helperForText_35_6_10_originalAllDirectionsFormula (K := K) u v) : Set.Nonempty (partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.10: once the nonempty-unbounded remainder is excluded, the packaged original textbook formula already forces the corrected nonempty-bounded branch.

open scoped Pointwise lemma helperForText_35_6_10_originalAllDirectionsFormula_forces_goodBranch_of_no_nonempty_unbounded {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hNoNonemptyUnbounded : ¬ (Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ ¬ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v))) (hFormula : helperForText_35_6_10_originalAllDirectionsFormula (K := K) u v) : Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.10: once the nonempty-unbounded remainder is excluded, the packaged original textbook proposition is equivalent to the corrected nonempty-bounded branch hypothesis on ∂₁ K(u, v).

open scoped Pointwise lemma helperForText_35_6_10_originalAllDirectionsFormula_iff_goodBranch_of_no_nonempty_unbounded {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hNoNonemptyUnbounded : ¬ (Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ ¬ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v))) : helperForText_35_6_10_originalAllDirectionsFormula (K := K) u v ↔ (Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v))

Lean source

Helper for Text 35.6.10: if the original unrestricted textbook schema held globally, then every finite base point would have a nonempty first partial subdifferential.

open scoped Pointwise lemma helperForText_35_6_10_globalOriginalStatement_forces_nonempty_partialFirst {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hOriginalSchema : ∀ u : Fin m → ℝ, ∀ v : Fin n → ℝ, (K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) → helperForText_35_6_10_originalAllDirectionsFormula (K := K) u v) : ∀ u : Fin m → ℝ, ∀ v : Fin n → ℝ, (K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) → Set.Nonempty (partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.10: if the original unrestricted textbook schema held globally and the nonempty-unbounded remainder were excluded pointwise, then every finite base point would satisfy the corrected nonempty-bounded branch hypothesis.

open scoped Pointwise lemma helperForText_35_6_10_globalOriginalStatement_forces_goodBranch_of_pointwise_no_nonempty_unbounded {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hNoNonemptyUnbounded : ∀ u : Fin m → ℝ, ∀ v : Fin n → ℝ, (K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) → ¬ (Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ ¬ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v))) (hOriginalSchema : ∀ u : Fin m → ℝ, ∀ v : Fin n → ℝ, (K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) → helperForText_35_6_10_originalAllDirectionsFormula (K := K) u v) : ∀ u : Fin m → ℝ, ∀ v : Fin n → ℝ, (K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) → Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Text 35.6.11: expand supportFunctionOfSet into the explicit supremum of pairings over ∂₂ K(u, v).

open scoped Pointwise lemma helperForText_35_6_11_supportEq_sSup_pairingImage {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (v : Fin n → ℝ) (v' : Fin n → ℝ) : supportFunctionOfSet (partialSubdifferentialInSecondVariable K u v) v' = sSup ((fun vStar : Fin n → ℝ => (((∑ i : Fin n, vStar i * v' i) : ℝ) : EReal)) '' partialSubdifferentialInSecondVariable K u v)

Lean source

Helper for Text 35.6.11: once ∂₂ K(u, v) is nonempty and bounded, the direct Theorem 23.4 route on the convex slice v₁ ↦ K u v₁ yields the supremum formula from the text.

open scoped Pointwise lemma helperForText_35_6_11_formula_of_nonempty_bounded_partialSecond {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartial : Set.Nonempty (partialSubdifferentialInSecondVariable K u v)) (hpartialBdd : Bornology.IsBounded (partialSubdifferentialInSecondVariable K u v)) (v' : Fin n → ℝ) : sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v 0 v' L} = sSup ((fun vStar : Fin n → ℝ => (((∑ i : Fin n, vStar i * v' i) : ℝ) : EReal)) '' partialSubdifferentialInSecondVariable K u v)

Lean source

Helper for Text 35.6.11: the pointwise supremum formula forces ∂₂ K(u, v) to be nonempty, since emptiness yields a direction where the saddle directional derivative takes the value ⊤.

open scoped Pointwise lemma helperForText_35_6_11_nonempty_partialSecond_required {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hFormula : ∀ v' : Fin n → ℝ, sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v 0 v' L} = sSup ((fun vStar : Fin n → ℝ => (((∑ i : Fin n, vStar i * v' i) : ℝ) : EReal)) '' partialSubdifferentialInSecondVariable K u v)) : Set.Nonempty (partialSubdifferentialInSecondVariable K u v)

Lean source

Helper for Text 35.6.11: if ∂₂ K(u, v) = ∅, then there exists a direction v' where the claimed supremum formula fails (the left-hand side becomes ⊤ while the right-hand side is sSup ∅ = ⊥).

open scoped Pointwise lemma helperForText_35_6_11_exists_counterexample_of_empty_partialSecond {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartialEmpty : partialSubdifferentialInSecondVariable K u v = ∅) : ∃ v' : Fin n → ℝ, sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v 0 v' L} ≠ sSup ((fun vStar : Fin n → ℝ => (((∑ i : Fin n, vStar i * v' i) : ℝ) : EReal)) '' partialSubdifferentialInSecondVariable K u v)

Lean source

Helper for Text 35.6.11: if ∂₂ K(u, v) = ∅, then there exists a direction v' where the left-hand side equals ⊤ while the right-hand side is sSup ∅ = ⊥.

open scoped Pointwise lemma helperForText_35_6_11_exists_top_lhs_and_bot_rhs_of_empty_partialSecond {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartialEmpty : partialSubdifferentialInSecondVariable K u v = ∅) : ∃ v' : Fin n → ℝ, sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v 0 v' L} = (⊤ : EReal) ∧ sSup ((fun vStar : Fin n → ℝ => (((∑ i : Fin n, vStar i * v' i) : ℝ) : EReal)) '' partialSubdifferentialInSecondVariable K u v) = (⊥ : EReal)

Lean source

Helper for Text 35.6.11: if the base point v lies in the interior of the effective domain of the convex slice v₁ ↦ K u v₁, then the supremum formula from Text 35.6.11 holds for every direction v'.

open scoped Pointwise lemma helperForText_35_6_11_formula_of_mem_interior_effectiveDomain_secondSlice {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hSaddle : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hvInt : v ∈ interior (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (K u))) (v' : Fin n → ℝ) : sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v 0 v' L} = sSup ((fun vStar : Fin n → ℝ => (((∑ i : Fin n, vStar i * v' i) : ℝ) : EReal)) '' partialSubdifferentialInSecondVariable K u v)

Lean source

Helper for Text 35.6.12: coerce a real-valued bifunction to an EReal-valued bifunction.

open scoped Pointwise def helperForText_35_6_12_erealOfRealKernel {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u v => ((K u v : ℝ) : EReal)

§35 — Continuity and Differentiability of Saddle-Functions — Part 14

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Helper for Theorem 35.7: an ambient-open convex set is relatively open convex (in the affine-span topology) in the sense of Section 35.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_7_isRelativelyOpenConvex_of_isOpen {E : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E] {s : Set E} (hsConv : Convex ℝ s) (hsOpen : IsOpen s) : IsRelativelyOpenConvex s

Lean source

Helper for Theorem 35.7: relative openness in the affine-span topology is preserved under a continuous linear equivalence.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_7_isRelativelyOpenConvex_image_continuousLinearEquiv {E F : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E] [NormedAddCommGroup F] [NormedSpace ℝ F] (e : E ≃L[ℝ] F) {s : Set E} (hs : IsRelativelyOpenConvex s) : IsRelativelyOpenConvex (e '' s)

Lean source

Helper for Theorem 35.7: from pointwise convergence at every point, the family is pointwise bounded on C × D in the sense required by Theorem 35.2.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_7_pointwiseBoundedFamilyOn_of_pointwiseTendsto {m n : ℕ} {C : Set (EuclideanSpace ℝ (Fin m))} {D : Set (EuclideanSpace ℝ (Fin n))} {K : EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → ℝ} {KSeq : ℕ → EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → ℝ} (hpoint : ∀ u₀ ∈ C, ∀ v₀ ∈ D, Filter.Tendsto (fun i : ℕ => KSeq i u₀ v₀) Filter.atTop (nhds (K u₀ v₀))) : Function.PointwiseBoundedFamilyOn (fun i : ℕ => Function.uncurry (KSeq i)) (C ×ˢ D)

Lean source

Helper for Theorem 35.7: the inverse of a product continuous linear equivalence carries a product set to the product of the inverse images on each factor.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_7_prodCongr_symm_image_prod {E E' F F' : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E] [NormedAddCommGroup E'] [NormedSpace ℝ E'] [NormedAddCommGroup F] [NormedSpace ℝ F] [NormedAddCommGroup F'] [NormedSpace ℝ F'] (eE : E ≃L[ℝ] E') (eF : F ≃L[ℝ] F') (A : Set E') (B : Set F') : (ContinuousLinearEquiv.prodCongr eE eF).symm '' (A ×ˢ B) = (eE.symm '' A) ×ˢ (eF.symm '' B)

Lean source

Helper for Theorem 35.7: Pi-type wrapper around Theorem 35.2. This transports the statement from EuclideanSpace ℝ (Fin k) to the Pi-model Fin k → ℝ via the canonical linear isometry.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_7_section35_theorem35_2_on_pi {I : Type*} {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : I → (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC : IsRelativelyOpenConvex C) (hD : IsRelativelyOpenConvex D) (hK : ∀ i, IsRealConcaveConvexOn C D (K i)) (hWitness : ∃ C' : Set (Fin m → ℝ), ∃ D' : Set (Fin n → ℝ), C' ⊆ C ∧ D' ⊆ D ∧ C ×ˢ D ⊆ convexHull ℝ (closure (C' ×ˢ D')) ∧ Function.PointwiseBoundedFamilyOn (fun i => Function.uncurry (K i)) (C' ×ˢ D')) : ∀ S : Set ((Fin m → ℝ) × (Fin n → ℝ)), S ⊆ C ×ˢ D → IsClosed S → Bornology.IsBounded S → Function.UniformlyBoundedFamilyOn (fun i => Function.uncurry (K i)) S ∧ Function.EquiLipschitzFamilyOn (fun i => Function.uncurry (K i)) S

Lean source

Helper for Theorem 35.7: extend a real convex function on a convex set by ⊤ outside the set, obtaining a global EReal convex function (and recording finiteness on the set).

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_7_convexFunction_ite_top_extension_of_convexOn {n : ℕ} {s : Set (Fin n → ℝ)} {f : (Fin n → ℝ) → ℝ} (hf : ConvexOn ℝ s f) : let F : (Fin n → ℝ) → EReal := fun x => if x ∈ s then ((f x : ℝ) : EReal) else (⊤ : EReal) ConvexFunction F ∧ (∀ x ∈ s, F x ≠ (⊤ : EReal) ∧ F x ≠ (⊥ : EReal))

Lean source

Helper for Theorem 35.7: the moving-slice pointwise convergence needed to apply the Chapter 24 pointwise-limit theorem to the one-variable convex slices.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_7_pointwiseTendsto_movingSlices {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {KSeq : ℕ → (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hKSeq : ∀ i : ℕ, IsRealConcaveConvexOn C D (KSeq i)) (hpoint : ∀ u₀ ∈ C, ∀ v₀ ∈ D, Filter.Tendsto (fun i : ℕ => KSeq i u₀ v₀) Filter.atTop (nhds (K u₀ v₀))) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) (uSeq : ℕ → Fin m → ℝ) (vSeq : ℕ → Fin n → ℝ) (huSeq : ∀ i : ℕ, uSeq i ∈ C) (hvSeq : ∀ i : ℕ, vSeq i ∈ D) (huSeq_tendsto : Filter.Tendsto uSeq Filter.atTop (nhds u)) (hvSeq_tendsto : Filter.Tendsto vSeq Filter.atTop (nhds v)) : (∀ u₀ ∈ C, Filter.Tendsto (fun i : ℕ => KSeq i u₀ (vSeq i)) Filter.atTop (nhds (K u₀ v))) ∧ (∀ v₀ ∈ D, Filter.Tendsto (fun i : ℕ => KSeq i (uSeq i) v₀) Filter.atTop (nhds (K u v₀)))

Lean source

Helper for Theorem 35.7: convert a limsup inequality under negation into a liminf inequality for EReal sequences.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_7_ereal_liminf_of_limsup_neg {a : ℕ → EReal} {b : EReal} (h : Filter.limsup (fun i : ℕ => -a i) Filter.atTop ≤ -b) : b ≤ Filter.liminf a Filter.atTop

Lean source

Helper for Theorem 35.7: bridge the real directional-derivative packaging realFirstVariableDirectionalDerivativeValue/realSecondVariableDirectionalDerivativeValue to the Chapter 23/24 upper directional derivative upperDirectionalDerivativeAt of the corresponding EReal convex extensions.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_7_realDirectionalDerivativeValue_bridges {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsRealConcaveConvexOn C D K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) (u' : Fin m → ℝ) (v' : Fin n → ℝ) : let f : (Fin m → ℝ) → EReal := fun x => if x ∈ C then ((-(K x v) : ℝ) : EReal) else (⊤ : EReal) let g : (Fin n → ℝ) → EReal := fun y => if y ∈ D then ((K u y : ℝ) : EReal) else (⊤ : EReal) (((realFirstVariableDirectionalDerivativeValue K u v u' : ℝ) : EReal) = -upperDirectionalDerivativeAt f u u') ∧ (((realSecondVariableDirectionalDerivativeValue K u v v' : ℝ) : EReal) = upperDirectionalDerivativeAt g v v')

§35 — Continuity and Differentiability of Saddle-Functions — Part 15

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Theorem 35.7: let K be a concave-convex real-valued function on an open convex product C × D ⊆ ℝ^m × ℝ^n, and let K₁, K₂, ... be concave-convex real-valued functions on C × D converging pointwise to K. If (uᵢ, vᵢ) ∈ C × D converges to (u, v), then for every u' ∈ ℝ^m the first-variable directional derivatives satisfy liminf Kᵢ'(uᵢ, vᵢ; u', 0) ≥ K'(u, v; u', 0), for every v' ∈ ℝ^n the second-variable directional derivatives satisfy limsup Kᵢ'(uᵢ, vᵢ; 0, v') ≤ K'(u, v; 0, v'), and for every ε > 0 the saddle subdifferentials eventually satisfy ∂ Kᵢ (uᵢ, vᵢ) ⊆ ∂ K (u, v) + ε B, where B is the Euclidean unit ball in ℝ^(m+n).

open scoped Pointwise open scoped Topology theorem section35_theorem35_7 {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {KSeq : ℕ → (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsRealConcaveConvexOn C D K) (hKSeq : ∀ i : ℕ, IsRealConcaveConvexOn C D (KSeq i)) (hpoint : ∀ u₀ ∈ C, ∀ v₀ ∈ D, Filter.Tendsto (fun i : ℕ => KSeq i u₀ v₀) Filter.atTop (nhds (K u₀ v₀))) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) (uSeq : ℕ → Fin m → ℝ) (vSeq : ℕ → Fin n → ℝ) (huSeq : ∀ i : ℕ, uSeq i ∈ C) (hvSeq : ∀ i : ℕ, vSeq i ∈ D) (huSeq_tendsto : Filter.Tendsto uSeq Filter.atTop (nhds u)) (hvSeq_tendsto : Filter.Tendsto vSeq Filter.atTop (nhds v)) : (∀ u' : Fin m → ℝ, ((realFirstVariableDirectionalDerivativeValue K u v u' : ℝ) : EReal) ≤ Filter.liminf (fun i : ℕ => ((realFirstVariableDirectionalDerivativeValue (KSeq i) (uSeq i) (vSeq i) u' : ℝ) : EReal)) Filter.atTop) ∧ (∀ v' : Fin n → ℝ, Filter.limsup (fun i : ℕ => ((realSecondVariableDirectionalDerivativeValue (KSeq i) (uSeq i) (vSeq i) v' : ℝ) : EReal)) Filter.atTop ≤ ((realSecondVariableDirectionalDerivativeValue K u v v' : ℝ) : EReal)) ∧ ∀ ε : ℝ, 0 < ε → ∃ i0 : ℕ, ∀ i ≥ i0, realSaddleSubdifferentialOn C D (KSeq i) (uSeq i) (vSeq i) ⊆ Set.image2 (fun p q : (Fin m → ℝ) × (Fin n → ℝ) => p + q) (realSaddleSubdifferentialOn C D K u v) (splitEuclideanClosedBall (m := m) (n := n) ε)

Lean source

Helper for Theorem 35.7: bridge the real partial subdifferentials on C and D to the Chapter 23 EReal subdifferential of the one-variable convex extensions used in the proof.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_7_realPartialSubdifferential_bridges {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) : let f : (Fin m → ℝ) → EReal := fun x => if x ∈ C then ((-(K x v) : ℝ) : EReal) else (⊤ : EReal) let g : (Fin n → ℝ) → EReal := fun y => if y ∈ D then ((K u y : ℝ) : EReal) else (⊤ : EReal) (∀ uStar : Fin m → ℝ, uStar ∈ realPartialSubdifferentialInFirstVariableOn C K u v ↔ dotProductEquiv ℝ (Fin m) (-uStar) ∈ subdifferentialAt f u) ∧ (∀ vStar : Fin n → ℝ, vStar ∈ realPartialSubdifferentialInSecondVariableOn D K u v ↔ dotProductEquiv ℝ (Fin n) vStar ∈ subdifferentialAt g v)

Lean source

Helper for Theorem 35.7: combine max-norm bounds on coordinate errors into membership in the split Euclidean closed ball splitEuclideanClosedBall ε (defined by coordinate squares).

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_7_splitBall_combine_errors {m n : ℕ} {ε δ : ℝ} (hδnonneg : 0 ≤ δ) (hδ : ((m + n : ℕ) : ℝ) * δ ^ (2 : ℕ) ≤ ε ^ (2 : ℕ)) {du : Fin m → ℝ} {dv : Fin n → ℝ} (hdu : ‖du‖ ≤ δ) (hdv : ‖dv‖ ≤ δ) : ((du, dv) : (Fin m → ℝ) × (Fin n → ℝ)) ∈ splitEuclideanClosedBall (m := m) (n := n) ε

Lean source

Helper for Corollary 35.7.1: specialize Theorem 35.7 to the constant sequence Kᵢ = K.

open scoped Pointwise open scoped Topology lemma helperForCorollary_35_7_1_constantSequence_asymptotics {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsRealConcaveConvexOn C D K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) (uSeq : ℕ → Fin m → ℝ) (vSeq : ℕ → Fin n → ℝ) (huSeq : ∀ i : ℕ, uSeq i ∈ C) (hvSeq : ∀ i : ℕ, vSeq i ∈ D) (huSeq_tendsto : Filter.Tendsto uSeq Filter.atTop (nhds u)) (hvSeq_tendsto : Filter.Tendsto vSeq Filter.atTop (nhds v)) : (∀ u' : Fin m → ℝ, ((realFirstVariableDirectionalDerivativeValue K u v u' : ℝ) : EReal) ≤ Filter.liminf (fun i : ℕ => ((realFirstVariableDirectionalDerivativeValue K (uSeq i) (vSeq i) u' : ℝ) : EReal)) Filter.atTop) ∧ (∀ v' : Fin n → ℝ, Filter.limsup (fun i : ℕ => ((realSecondVariableDirectionalDerivativeValue K (uSeq i) (vSeq i) v' : ℝ) : EReal)) Filter.atTop ≤ ((realSecondVariableDirectionalDerivativeValue K u v v' : ℝ) : EReal)) ∧ ∀ ε : ℝ, 0 < ε → ∃ i0 : ℕ, ∀ i ≥ i0, realSaddleSubdifferentialOn C D K (uSeq i) (vSeq i) ⊆ Set.image2 (fun p q : (Fin m → ℝ) × (Fin n → ℝ) => p + q) (realSaddleSubdifferentialOn C D K u v) (splitEuclideanClosedBall (m := m) (n := n) ε)

§35 — Continuity and Differentiability of Saddle-Functions — Part 16

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Corollary 35.7.1: let C × D be an open convex set in ℝ^m × ℝ^n, and let K be a concave-convex real-valued function on C × D. Then for each direction u' ∈ ℝ^m, the map (u, v) ↦ K'(u, v; u', 0) is lower semicontinuous on C × D; for each direction v' ∈ ℝ^n, the map (u, v) ↦ K'(u, v; 0, v') is upper semicontinuous on C × D. Moreover, for every (u, v) ∈ C × D and every ε > 0, there exists δ > 0 such that whenever (x, y) ∈ C × D and ((x - u), (y - v)) lies in the closed Euclidean ball of radius δ, one has ∂K(x, y) ⊆ ∂K(u, v) + ε B.

open scoped Pointwise open scoped Topology theorem section35_corollary35_7_1 {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsRealConcaveConvexOn C D K) : (∀ u' : Fin m → ℝ, LowerSemicontinuousOn (fun p : (Fin m → ℝ) × (Fin n → ℝ) => realFirstVariableDirectionalDerivativeValue K p.1 p.2 u') (C ×ˢ D)) ∧ (∀ v' : Fin n → ℝ, UpperSemicontinuousOn (fun p : (Fin m → ℝ) × (Fin n → ℝ) => realSecondVariableDirectionalDerivativeValue K p.1 p.2 v') (C ×ˢ D)) ∧ (∀ ⦃u : Fin m → ℝ⦄, u ∈ C → ∀ ⦃v : Fin n → ℝ⦄, v ∈ D → ∀ ε : ℝ, 0 < ε → ∃ δ : ℝ, 0 < δ ∧ ∀ x : Fin m → ℝ, x ∈ C → ∀ y : Fin n → ℝ, y ∈ D → ((x - u), (y - v)) ∈ splitEuclideanClosedBall (m := m) (n := n) δ → realSaddleSubdifferentialOn C D K x y ⊆ Set.image2 (fun p q : (Fin m → ℝ) × (Fin n → ℝ) => p + q) (realSaddleSubdifferentialOn C D K u v) (splitEuclideanClosedBall (m := m) (n := n) ε))

Lean source

Helper for Corollary 35.7.1: membership in the split Euclidean closed ball controls the Pi-norm of each coordinate component.

open scoped Pointwise open scoped Topology lemma helperForCorollary_35_7_1_coordinateNormBounds_of_mem_splitBall {m n : ℕ} {r : ℝ} (hr : 0 ≤ r) {du : Fin m → ℝ} {dv : Fin n → ℝ} (hmem : ((du, dv) : (Fin m → ℝ) × (Fin n → ℝ)) ∈ splitEuclideanClosedBall (m := m) (n := n) r) : ‖du‖ ≤ r ∧ ‖dv‖ ≤ r

Lean source

Helper for Corollary 35.7.1: the first-variable liminf inequality from Theorem 35.7 implies lower semicontinuity of (u, v) ↦ K'(u, v; u', 0) on C ×ˢ D.

open scoped Pointwise open scoped Topology lemma helperForCorollary_35_7_1_lowerSemicontinuousOn_firstDirectionalDerivative {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsRealConcaveConvexOn C D K) (u' : Fin m → ℝ) : LowerSemicontinuousOn (fun p : (Fin m → ℝ) × (Fin n → ℝ) => realFirstVariableDirectionalDerivativeValue K p.1 p.2 u') (C ×ˢ D)

Lean source

Helper for Corollary 35.7.1: the second-variable limsup inequality from Theorem 35.7 implies upper semicontinuity of (u, v) ↦ K'(u, v; 0, v') on C ×ˢ D.

open scoped Pointwise open scoped Topology lemma helperForCorollary_35_7_1_upperSemicontinuousOn_secondDirectionalDerivative {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsRealConcaveConvexOn C D K) (v' : Fin n → ℝ) : UpperSemicontinuousOn (fun p : (Fin m → ℝ) × (Fin n → ℝ) => realSecondVariableDirectionalDerivativeValue K p.1 p.2 v') (C ×ˢ D)

Lean source

Helper for Corollary 35.7.1: the eventual saddle-subdifferential inclusion from Theorem 35.7 upgrades to a uniform split-ball neighborhood around any base point (u, v) ∈ C ×ˢ D.

open scoped Pointwise open scoped Topology lemma helperForCorollary_35_7_1_local_saddleSubdifferential_subset {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsRealConcaveConvexOn C D K) {u : Fin m → ℝ} (hu : u ∈ C) {v : Fin n → ℝ} (hv : v ∈ D) (ε : ℝ) (hε : 0 < ε) : ∃ δ : ℝ, 0 < δ ∧ ∀ x : Fin m → ℝ, x ∈ C → ∀ y : Fin n → ℝ, y ∈ D → ((x - u), (y - v)) ∈ splitEuclideanClosedBall (m := m) (n := n) δ → realSaddleSubdifferentialOn C D K x y ⊆ Set.image2 (fun p q : (Fin m → ℝ) × (Fin n → ℝ) => p + q) (realSaddleSubdifferentialOn C D K u v) (splitEuclideanClosedBall (m := m) (n := n) ε)

§35 — Continuity and Differentiability of Saddle-Functions — Part 17

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

The packed EReal-valued map on ℝ^(m+n) associated to a saddle kernel K.

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. def packedSaddleKernel {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin (m + n) → ℝ) → EReal := fun z => K (fun i => z (Fin.castAdd n i)) (fun j => z (Fin.natAdd m j))

Lean source

The split gradient pair obtained from a chosen differentiability witness for the packed saddle kernel at (u, v).

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. noncomputable def packedSaddleKernelGradientPairAt {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hDiff : ERealDifferentiableAt (packedSaddleKernel K) (Fin.append u v)) : (Fin m → ℝ) × (Fin n → ℝ) := (fun i => erealGradientAt hDiff (Fin.castAdd n i), fun j => erealGradientAt hDiff (Fin.natAdd m j))

Lean source

An extended-real saddle kernel is finite on a neighborhood of (u, v) when it is finite at every point of some open neighborhood in ℝ^m × ℝ^n.

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. def SaddleKernelFiniteOnNeighborhoodAt {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (u : Fin m → ℝ) (v : Fin n → ℝ) : Prop := ∃ N : Set ((Fin m → ℝ) × (Fin n → ℝ)), IsOpen N ∧ (u, v) ∈ N ∧ ∀ p ∈ N, K p.1 p.2 ≠ (⊤ : EReal) ∧ K p.1 p.2 ≠ (⊥ : EReal)

Lean source

The directional derivative function of a saddle kernel at (u, v) is linear when there is a single pair (u*, v*) whose Euclidean pairing gives every directional derivative.

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. def HasLinearSaddleDirectionalDerivativeAt {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (u : Fin m → ℝ) (v : Fin n → ℝ) : Prop := ∃ uStar : Fin m → ℝ, ∃ vStar : Fin n → ℝ, ∀ u' : Fin m → ℝ, ∀ v' : Fin n → ℝ, IsSaddleDirectionalDerivativeAt K u v u' v' (((((∑ i : Fin m, uStar i * u' i) + ∑ j : Fin n, vStar j * v' j) : ℝ) : EReal))

Lean source

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. lemma helperForTheorem_35_8_finiteRectangle_from_sliceFiniteNeighborhoods {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {C0 : Set (Fin m → ℝ)} {D0 : Set (Fin n → ℝ)} (hK : IsGloballyConcaveConvexERealKernel K) (hC0open : IsOpen C0) (huC0 : u ∈ C0) (hD0open : IsOpen D0) (hvD0 : v ∈ D0) (hFirstFinite : ∀ x ∈ C0, K x v ≠ (⊤ : EReal) ∧ K x v ≠ (⊥ : EReal)) (hSecondFinite : ∀ y ∈ D0, K u y ≠ (⊤ : EReal) ∧ K u y ≠ (⊥ : EReal)) (hCornerFinite : ∀ {x : Fin m → ℝ} {y : Fin n → ℝ}, K x v ≠ (⊤ : EReal) ∧ K x v ≠ (⊥ : EReal) → K (2 • u - x) v ≠ (⊤ : EReal) ∧ K (2 • u - x) v ≠ (⊥ : EReal) → K u y ≠ (⊤ : EReal) ∧ K u y ≠ (⊥ : EReal) → K u (2 • v - y) ≠ (⊤ : EReal) ∧ K u (2 • v - y) ≠ (⊥ : EReal) → K x y ≠ (⊤ : EReal) ∧ K x y ≠ (⊥ : EReal)) : ∃ C : Set (Fin m → ℝ), ∃ D : Set (Fin n → ℝ), IsOpen C ∧ u ∈ C ∧ Convex ℝ C ∧ IsOpen D ∧ v ∈ D ∧ Convex ℝ D ∧ ∀ x ∈ C, ∀ y ∈ D, K x y ≠ (⊤ : EReal) ∧ K x y ≠ (⊥ : EReal)

Lean source

Helper for Theorem 35.8: dot products split over Fin.append.

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. lemma helperForTheorem_35_8_dotProduct_append {m n : ℕ} (u : Fin m → ℝ) (v : Fin n → ℝ) (b : Fin (m + n) → ℝ) : dotProduct (Fin.append u v) b = dotProduct u (fun i => b (Fin.castAdd n i)) + dotProduct v (fun j => b (Fin.natAdd m j))

Lean source

Helper for Theorem 35.8: the packed directional difference quotient along a pure second-variable direction coincides with the one-variable directional difference quotient of the convex slice v' ↦ K u v'.

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. lemma helperForTheorem_35_8_directionalDifferenceQuotient_secondSlice {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (v : Fin n → ℝ) (vDir : Fin n → ℝ) (t : ℝ) : directionalDifferenceQuotientAt (packedSaddleKernel K) (Fin.append u v) (Fin.append (0 : Fin m → ℝ) vDir) t = directionalDifferenceQuotientAt (K u) v vDir t

Lean source

Helper for Theorem 35.8: along the positive ray, differentiability of the packed kernel makes the reflected first-slice quotient eventually equal to the negative packed quotient in the reflected axis direction.

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. lemma helperForTheorem_35_8_directionalDifferenceQuotient_reflectedFirstSlice_eventuallyEq {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (v : Fin n → ℝ) (uDir : Fin m → ℝ) (hDiff : ERealDifferentiableAt (packedSaddleKernel K) (Fin.append u v)) (huDir_ne : uDir ≠ 0) : Filter.EventuallyEq (nhdsWithin (0 : ℝ) (Set.Ioi 0)) (fun t => directionalDifferenceQuotientAt (fun x : Fin m → ℝ => -K (-x) v) (-u) uDir t) (fun t => -directionalDifferenceQuotientAt (packedSaddleKernel K) (Fin.append u v) (Fin.append (-uDir) (0 : Fin n → ℝ)) t)

Lean source

Helper for Theorem 35.8: the packed directional quotient is exactly the saddle directional quotient after splitting a packed direction into its u and v blocks.

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. lemma helperForTheorem_35_8_packedQuotient_eq_saddleQuotient {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (v : Fin n → ℝ) (uDir : Fin m → ℝ) (vDir : Fin n → ℝ) (t : ℝ) : directionalDifferenceQuotientAt (packedSaddleKernel K) (Fin.append u v) (Fin.append uDir vDir) t = saddleDirectionalDifferenceQuotientAt K u v uDir vDir t

Lean source

Helper for Theorem 35.8: from differentiability of the packed saddle kernel, the split gradient pair gives unique membership in each partial subdifferential.

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. lemma helperForTheorem_35_8_forward_unique_partial_subgradients {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hK : IsGloballyConcaveConvexERealKernel K) (hDiff : ERealDifferentiableAt (packedSaddleKernel K) (Fin.append u v)) : let grad : (Fin m → ℝ) × (Fin n → ℝ) := packedSaddleKernelGradientPairAt (K := K) (u := u) (v := v) hDiff (grad.1 ∈ partialSubdifferentialInFirstVariable K u v ∧ ∀ uStar : Fin m → ℝ, uStar ∈ partialSubdifferentialInFirstVariable K u v → uStar = grad.1) ∧ (grad.2 ∈ partialSubdifferentialInSecondVariable K u v ∧ ∀ vStar : Fin n → ℝ, vStar ∈ partialSubdifferentialInSecondVariable K u v → vStar = grad.2)

Lean source

Helper for Theorem 35.8: uniqueness in the product subdifferential implies each partial subdifferential is a singleton.

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. lemma helperForTheorem_35_8_unique_productSubgradient_gives_unique_partials {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (huniq : ∃! g : (Fin m → ℝ) × (Fin n → ℝ), g ∈ productSubdifferentialAt K u v) : ∃ uStar : Fin m → ℝ, ∃ vStar : Fin n → ℝ, partialSubdifferentialInFirstVariable K u v = {uStar} ∧ partialSubdifferentialInSecondVariable K u v = {vStar}

Lean source

Helper for Theorem 35.8: a singleton first partial subdifferential is immediately nonempty and bounded.

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. lemma helperForTheorem_35_8_nonempty_bounded_partialFirst_of_singleton {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {uStar : Fin m → ℝ} (hFirstSingleton : partialSubdifferentialInFirstVariable K u v = {uStar}) : Set.Nonempty (partialSubdifferentialInFirstVariable K u v) ∧ Bornology.IsBounded (partialSubdifferentialInFirstVariable K u v)

Lean source

Helper for Theorem 35.8: a singleton second partial subdifferential is immediately nonempty and bounded.

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. lemma helperForTheorem_35_8_nonempty_bounded_partialSecond_of_singleton {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {vStar : Fin n → ℝ} (hSecondSingleton : partialSubdifferentialInSecondVariable K u v = {vStar}) : Set.Nonempty (partialSubdifferentialInSecondVariable K u v) ∧ Bornology.IsBounded (partialSubdifferentialInSecondVariable K u v)

Lean source

Helper for Theorem 35.8: a singleton first partial subdifferential already fixes the first-axis saddle directional derivative value in every direction.

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. lemma helperForTheorem_35_8_firstAxisDirectionalDerivative_value_of_singleton_partial {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {uStar : Fin m → ℝ} (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hFirstSingleton : partialSubdifferentialInFirstVariable K u v = {uStar}) (uDir : Fin m → ℝ) : sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v uDir 0 L} = ((((∑ i : Fin m, uStar i * uDir i) : ℝ) : EReal))

Lean source

Helper for Theorem 35.8: a singleton second partial subdifferential already fixes the second-axis saddle directional derivative value in every direction.

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. lemma helperForTheorem_35_8_secondAxisDirectionalDerivative_value_of_singleton_partial {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {vStar : Fin n → ℝ} (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hSecondSingleton : partialSubdifferentialInSecondVariable K u v = {vStar}) (vDir : Fin n → ℝ) : sInf {L : EReal | IsSaddleDirectionalDerivativeAt K u v 0 vDir L} = ((((∑ j : Fin n, vStar j * vDir j) : ℝ) : EReal))

Lean source

Helper for Theorem 35.8: singleton partial subdifferentials force differentiability of the reflected first slice and of the honest second slice, so both one-variable slices have interior finite base points.

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. lemma helperForTheorem_35_8_sliceDifferentiabilityWitnesses_of_singleton_partials {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {uStar : Fin m → ℝ} {vStar : Fin n → ℝ} (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hFirstSingleton : partialSubdifferentialInFirstVariable K u v = {uStar}) (hSecondSingleton : partialSubdifferentialInSecondVariable K u v = {vStar}) : ERealDifferentiableAt (fun x : Fin m → ℝ => -K (-x) v) (-u) ∧ ERealDifferentiableAt (K u) v

Lean source

Helper for Theorem 35.8: singleton partial subdifferentials already produce open convex neighborhoods on which the first slice x ↦ K x v and the second slice y ↦ K u y are finite.

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. lemma helperForTheorem_35_8_sliceFiniteNeighborhoods_of_singleton_partials {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {uStar : Fin m → ℝ} {vStar : Fin n → ℝ} (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hFirstSingleton : partialSubdifferentialInFirstVariable K u v = {uStar}) (hSecondSingleton : partialSubdifferentialInSecondVariable K u v = {vStar}) : ∃ C : Set (Fin m → ℝ), ∃ D : Set (Fin n → ℝ), IsOpen C ∧ u ∈ C ∧ Convex ℝ C ∧ (∀ x ∈ C, K x v ≠ (⊤ : EReal) ∧ K x v ≠ (⊥ : EReal)) ∧ IsOpen D ∧ v ∈ D ∧ Convex ℝ D ∧ (∀ y ∈ D, K u y ≠ (⊤ : EReal) ∧ K u y ≠ (⊥ : EReal))

Lean source

Helper for Theorem 35.8: singleton partial subdifferentials should force (u, v) to lie in the interior of the full saddle effective domain.

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. lemma helperForTheorem_35_8_reflection_mem_ball {k : ℕ} {c x : Fin k → ℝ} {ε : ℝ} (hx : x ∈ Metric.ball c ε) : (2 • c - x) ∈ Metric.ball c ε

Lean source

Helper for Theorem 35.8: a point and its reflection across c have midpoint c.

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. lemma helperForTheorem_35_8_midpoint_reflection_eq_center {k : ℕ} (c x : Fin k → ℝ) : (1 / 2 : ℝ) • x + (1 / 2 : ℝ) • (2 • c - x) = c

Lean source

Helper for Theorem 35.8: if a convex slice has value ⊥ at one endpoint of a reflected pair and its midpoint value is not ⊥, then the opposite endpoint must be ⊤.

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. lemma helperForTheorem_35_8_midpointFinite_leftBot_forces_rightTop {k : ℕ} {g : (Fin k → ℝ) → EReal} {x y m : Fin k → ℝ} (hConv : ConvexFunction g) (hMid : (1 / 2 : ℝ) • x + (1 / 2 : ℝ) • y = m) (hmNeBot : g m ≠ (⊥ : EReal)) (hxBot : g x = (⊥ : EReal)) : g y = (⊤ : EReal)

Lean source

Helper for Theorem 35.8: if one corner of the reflected rectangle is ⊤, the alternating convexity/concavity midpoint argument forces the three other corners into the checkerboard pattern ⊥, ⊤, ⊥.

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. lemma helperForTheorem_35_8_topCorner_forces_checkerboard {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} {x : Fin m → ℝ} {y : Fin n → ℝ} (hxvFinite : K x v ≠ (⊤ : EReal) ∧ K x v ≠ (⊥ : EReal)) (hxRefvFinite : K (2 • u - x) v ≠ (⊤ : EReal) ∧ K (2 • u - x) v ≠ (⊥ : EReal)) (huyFinite : K u y ≠ (⊤ : EReal) ∧ K u y ≠ (⊥ : EReal)) (huyRefFinite : K u (2 • v - y) ≠ (⊤ : EReal) ∧ K u (2 • v - y) ≠ (⊥ : EReal)) (hxyTop : K x y = (⊤ : EReal)) : K (2 • u - x) y = (⊥ : EReal) ∧ K (2 • u - x) (2 • v - y) = (⊤ : EReal) ∧ K x (2 • v - y) = (⊥ : EReal)

Lean source

Helper for Theorem 35.8: if one corner of the reflected rectangle is ⊥, the same midpoint propagation yields the opposite checkerboard pattern ⊤, ⊥, ⊤.

open scoped Pointwise open scoped Topology -- Proof sketch: apply the one-variable differentiability criterion from Theorem 25.1 to the -- convex slice `v' ↦ K u v'` and to the convex function `u' ↦ -K u' v`, which encodes -- concavity in the first variable. Finite-point differentiability of the packed `EReal` map on -- `ℝ^(m+n)` identifies a distinguished gradient vector at `(u, v)`, and the saddle inequalities -- show that its split coordinates give the unique saddle subgradient. Conversely, uniqueness of -- the saddle subgradient forces uniqueness of the slice subgradients, from which -- differentiability of the packed map follows. lemma helperForTheorem_35_8_botCorner_forces_checkerboard {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} {x : Fin m → ℝ} {y : Fin n → ℝ} (hxvFinite : K x v ≠ (⊤ : EReal) ∧ K x v ≠ (⊥ : EReal)) (hxRefvFinite : K (2 • u - x) v ≠ (⊤ : EReal) ∧ K (2 • u - x) v ≠ (⊥ : EReal)) (huyFinite : K u y ≠ (⊤ : EReal) ∧ K u y ≠ (⊥ : EReal)) (huyRefFinite : K u (2 • v - y) ≠ (⊤ : EReal) ∧ K u (2 • v - y) ≠ (⊥ : EReal)) (hxyBot : K x y = (⊥ : EReal)) : K (2 • u - x) y = (⊤ : EReal) ∧ K (2 • u - x) (2 • v - y) = (⊥ : EReal) ∧ K x (2 • v - y) = (⊤ : EReal)

§35 — Continuity and Differentiability of Saddle-Functions — Part 18

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Helper for Theorem 35.8: in a convex slice, a ⊥ value at the end of a positive ray from a finite base point propagates to every shorter positive point on that same ray.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_convexSlice_bot_propagates_along_short_ray {k : ℕ} {g : (Fin k → ℝ) → EReal} {c d : Fin k → ℝ} {t : ℝ} (hConv : ConvexFunction g) (hcFinite : g c ≠ (⊤ : EReal) ∧ g c ≠ (⊥ : EReal)) (ht : 0 < t) (ht_le_one : t ≤ 1) (hEndBot : g (c + d) = (⊥ : EReal)) : g (c + t • d) = (⊥ : EReal)

Lean source

Helper for Theorem 35.8: shrinking a reflected checkerboard in the second coordinate preserves the same alternating ⊤/⊥ pattern.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_reflectedSecondSlice_checkerboard_shrinks_along_short_ray {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} {x : Fin m → ℝ} {y : Fin n → ℝ} {t : ℝ} (hxFinite : K x v ≠ (⊤ : EReal) ∧ K x v ≠ (⊥ : EReal)) (hxRefFinite : K (2 • u - x) v ≠ (⊤ : EReal) ∧ K (2 • u - x) v ≠ (⊥ : EReal)) (hChecker : (K x y = (⊤ : EReal) ∧ K (2 • u - x) y = (⊥ : EReal) ∧ K (2 • u - x) (2 • v - y) = (⊤ : EReal) ∧ K x (2 • v - y) = (⊥ : EReal)) ∨ (K x y = (⊥ : EReal) ∧ K (2 • u - x) y = (⊤ : EReal) ∧ K (2 • u - x) (2 • v - y) = (⊥ : EReal) ∧ K x (2 • v - y) = (⊤ : EReal))) (ht : 0 < t) (ht_le_one : t ≤ 1) : let yPlus : Fin n → ℝ := v + t • (y - v) let yMinus : Fin n → ℝ := v - t • (y - v) (K x yPlus = (⊤ : EReal) ∧ K (2 • u - x) yPlus = (⊥ : EReal) ∧ K (2 • u - x) yMinus = (⊤ : EReal) ∧ K x yMinus = (⊥ : EReal)) ∨ (K x yPlus = (⊥ : EReal) ∧ K (2 • u - x) yPlus = (⊤ : EReal) ∧ K (2 • u - x) yMinus = (⊥ : EReal) ∧ K x yMinus = (⊤ : EReal))

Lean source

Helper for Theorem 35.8: shrinking a reflected checkerboard in the first coordinate preserves the same alternating ⊤/⊥ pattern.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_reflectedFirstSlice_checkerboard_shrinks_along_short_ray {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {x : Fin m → ℝ} {yPlus yMinus : Fin n → ℝ} {s : ℝ} (hyPlusFinite : K u yPlus ≠ (⊤ : EReal) ∧ K u yPlus ≠ (⊥ : EReal)) (hyMinusFinite : K u yMinus ≠ (⊤ : EReal) ∧ K u yMinus ≠ (⊥ : EReal)) (hChecker : (K x yPlus = (⊤ : EReal) ∧ K (2 • u - x) yPlus = (⊥ : EReal) ∧ K (2 • u - x) yMinus = (⊤ : EReal) ∧ K x yMinus = (⊥ : EReal)) ∨ (K x yPlus = (⊥ : EReal) ∧ K (2 • u - x) yPlus = (⊤ : EReal) ∧ K (2 • u - x) yMinus = (⊥ : EReal) ∧ K x yMinus = (⊤ : EReal))) (hs : 0 < s) (hs_le_one : s ≤ 1) : let xPlus : Fin m → ℝ := u + s • (x - u) let xMinus : Fin m → ℝ := u - s • (x - u) (K xPlus yPlus = (⊤ : EReal) ∧ K xMinus yPlus = (⊥ : EReal) ∧ K xMinus yMinus = (⊤ : EReal) ∧ K xPlus yMinus = (⊥ : EReal)) ∨ (K xPlus yPlus = (⊥ : EReal) ∧ K xMinus yPlus = (⊤ : EReal) ∧ K xMinus yMinus = (⊥ : EReal) ∧ K xPlus yMinus = (⊤ : EReal))

Lean source

Helper for Theorem 35.8: sufficiently small reflected pairs stay inside the chosen slice neighborhoods around (u, v).

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_small_reflected_pair_mem_slice_neighborhoods {m n : ℕ} {u : Fin m → ℝ} {v : Fin n → ℝ} {x : Fin m → ℝ} {y : Fin n → ℝ} {C0 : Set (Fin m → ℝ)} {D0 : Set (Fin n → ℝ)} (hC0open : IsOpen C0) (huC0 : u ∈ C0) (hD0open : IsOpen D0) (hvD0 : v ∈ D0) : ∃ s t : ℝ, 0 < s ∧ s ≤ 1 ∧ 0 < t ∧ t ≤ 1 ∧ u + s • (x - u) ∈ C0 ∧ u - s • (x - u) ∈ C0 ∧ v + t • (y - v) ∈ D0 ∧ v - t • (y - v) ∈ D0

Lean source

Helper for Theorem 35.8: the independent slice-neighborhood shrink parameters can be replaced by a single common step that keeps both reflected pairs inside the chosen neighborhoods.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_common_small_reflected_pair_mem_slice_neighborhoods {m n : ℕ} {u : Fin m → ℝ} {v : Fin n → ℝ} {x : Fin m → ℝ} {y : Fin n → ℝ} {C0 : Set (Fin m → ℝ)} {D0 : Set (Fin n → ℝ)} (hC0open : IsOpen C0) (huC0 : u ∈ C0) (hD0open : IsOpen D0) (hvD0 : v ∈ D0) : ∃ r : ℝ, 0 < r ∧ r ≤ 1 ∧ u + r • (x - u) ∈ C0 ∧ u - r • (x - u) ∈ C0 ∧ v + r • (y - v) ∈ D0 ∧ v - r • (y - v) ∈ D0

Lean source

Helper for Theorem 35.8: in a convex slice neighborhood, once one point on a ray from the center is known to lie in the set, every shorter point on the same ray lies there as well.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_shorter_ray_mem_of_convex {k : ℕ} {C : Set (Fin k → ℝ)} {u d : Fin k → ℝ} {r s : ℝ} (hCconv : Convex ℝ C) (hu : u ∈ C) (hRay : u + r • d ∈ C) (hr_nonneg : 0 ≤ r) (hs_nonneg : 0 ≤ s) (hs_le : s ≤ r) : u + s • d ∈ C

Lean source

Helper for Theorem 35.8: interior membership in the saddle effective domain yields an open convex rectangle around (u, v) on which K stays finite.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_openConvexFiniteRectangle_of_jointInterior {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hInterior : (u, v) ∈ interior (saddleFunctionEffectiveDomain K)) : ∃ C : Set (Fin m → ℝ), ∃ D : Set (Fin n → ℝ), IsOpen C ∧ u ∈ C ∧ Convex ℝ C ∧ IsOpen D ∧ v ∈ D ∧ Convex ℝ D ∧ ∀ u' ∈ C, ∀ v' ∈ D, K u' v' ≠ (⊤ : EReal) ∧ K u' v' ≠ (⊥ : EReal)

Lean source

Helper for Theorem 35.8: once the reflected checkerboard is shrunk with a common parameter, its first corner is already an explicit mixed saddle secant quotient equal to ⊤ or ⊥.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_checkerboardForcesInfiniteMixedQuotient_afterCommonShrink {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {x : Fin m → ℝ} {y : Fin n → ℝ} {r : ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hr : 0 < r) (hChecker : let xPlus : Fin m → ℝ := u + r • (x - u) let xMinus : Fin m → ℝ := u - r • (x - u) let yPlus : Fin n → ℝ := v + r • (y - v) let yMinus : Fin n → ℝ := v - r • (y - v) (K xPlus yPlus = (⊤ : EReal) ∧ K xMinus yPlus = (⊥ : EReal) ∧ K xMinus yMinus = (⊤ : EReal) ∧ K xPlus yMinus = (⊥ : EReal)) ∨ (K xPlus yPlus = (⊥ : EReal) ∧ K xMinus yPlus = (⊤ : EReal) ∧ K xMinus yMinus = (⊥ : EReal) ∧ K xPlus yMinus = (⊤ : EReal))) : saddleDirectionalDifferenceQuotientAt K u v (x - u) (y - v) r = (⊤ : EReal) ∨ saddleDirectionalDifferenceQuotientAt K u v (x - u) (y - v) r = (⊥ : EReal)

Lean source

Helper for Theorem 35.8: once both the base value and the moved mixed corner are finite, the corresponding mixed directional quotient at a positive step cannot be ⊤ or ⊥.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_mixedQuotient_ne_top_or_bot_of_finiteMovedCorner {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {uDir : Fin m → ℝ} {vDir : Fin n → ℝ} {r : ℝ} (hr : 0 < r) (hBase : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hMoved : K (u + r • uDir) (v + r • vDir) ≠ (⊤ : EReal) ∧ K (u + r • uDir) (v + r • vDir) ≠ (⊥ : EReal)) : saddleDirectionalDifferenceQuotientAt K u v uDir vDir r ≠ (⊤ : EReal) ∧ saddleDirectionalDifferenceQuotientAt K u v uDir vDir r ≠ (⊥ : EReal)

Lean source

Helper for Theorem 35.8: singleton slice data gives a uniform positive radius on which all four reflected axis points along the directions du and dv stay finite.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_small_reflected_axis_finiteness_of_singletonSliceData {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {uStar : Fin m → ℝ} {vStar : Fin n → ℝ} (du : Fin m → ℝ) (dv : Fin n → ℝ) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hFirstSingleton : partialSubdifferentialInFirstVariable K u v = {uStar}) (hSecondSingleton : partialSubdifferentialInSecondVariable K u v = {vStar}) : ∃ ρ : ℝ, 0 < ρ ∧ ∀ t : ℝ, 0 < t → t < ρ → (K (u + t • du) v ≠ (⊤ : EReal) ∧ K (u + t • du) v ≠ (⊥ : EReal)) ∧ (K (u - t • du) v ≠ (⊤ : EReal) ∧ K (u - t • du) v ≠ (⊥ : EReal)) ∧ (K u (v + t • dv) ≠ (⊤ : EReal) ∧ K u (v + t • dv) ≠ (⊥ : EReal)) ∧ (K u (v - t • dv) ≠ (⊤ : EReal) ∧ K u (v - t • dv) ≠ (⊥ : EReal))

§35 — Continuity and Differentiability of Saddle-Functions — Part 19

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Helper for Theorem 35.8: the Chapter 34 coordinate effective-domain conditions are already the separate ≠ ⊥ and ≠ ⊤ statements needed to conclude finiteness at a mixed point.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_finite_of_mem_effectiveDomains {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hK : IsGloballyConcaveConvexERealKernel K) {x : Fin m → ℝ} {y : Fin n → ℝ} (hxDom : x ∈ effectiveDomain₁ K) (hyDom : y ∈ effectiveDomain₂ K) : K x y ≠ (⊤ : EReal) ∧ K x y ≠ (⊥ : EReal)

Lean source

Helper for Theorem 35.8: singleton first/second partial data already shows that the base point lies in dom₁ K × dom₂ K; what remains missing later is only a neighborhood upgrade of these domain memberships.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_base_mem_effectiveDomains_of_singleton_partials {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {uStar : Fin m → ℝ} {vStar : Fin n → ℝ} (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hFirstSingleton : partialSubdifferentialInFirstVariable K u v = {uStar}) (hSecondSingleton : partialSubdifferentialInSecondVariable K u v = {vStar}) : u ∈ effectiveDomain₁ K ∧ v ∈ effectiveDomain₂ K

Lean source

Helper for Theorem 35.8: a pair of coordinate effective-domain memberships is exactly the pair of universal no-⊥ and no-⊤ clauses needed later for the moved first and second slices.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_effectiveDomainPair_iff_universalMovedSliceClauses {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {x : Fin m → ℝ} {y : Fin n → ℝ} : (x ∈ effectiveDomain₁ K ∧ y ∈ effectiveDomain₂ K) ↔ (∀ y' : Fin n → ℝ, K x y' ≠ (⊥ : EReal)) ∧ ∀ x' : Fin m → ℝ, K x' y ≠ (⊤ : EReal)

Lean source

Helper for Theorem 35.8: once (u, v) is already known to lie in dom₁ K × dom₂ K, every translated second slice and every reflected translated first slice already satisfies the convexity and nonempty-effective-domain half of the properness package.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_movedSlices_convex_and_nonemptyEffectiveDomains_of_baseEffectiveDomains {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (du : Fin m → ℝ) (dv : Fin n → ℝ) (hK : IsGloballyConcaveConvexERealKernel K) (huDom : u ∈ effectiveDomain₁ K) (hvDom : v ∈ effectiveDomain₂ K) : ∀ t : ℝ, ConvexFunction (K (u + t • du)) ∧ (convexFunctionEffectiveDomain (K (u + t • du))).Nonempty ∧ ConvexFunction (K (u - t • du)) ∧ (convexFunctionEffectiveDomain (K (u - t • du))).Nonempty ∧ ConvexFunction (fun x : Fin m → ℝ => -K (-x) (v + t • dv)) ∧ (convexFunctionEffectiveDomain (fun x : Fin m → ℝ => -K (-x) (v + t • dv))).Nonempty ∧ ConvexFunction (fun x : Fin m → ℝ => -K (-x) (v - t • dv)) ∧ (convexFunctionEffectiveDomain (fun x : Fin m → ℝ => -K (-x) (v - t • dv))).Nonempty

Lean source

Helper for Theorem 35.8: once a short ray from (u, v) is known to stay inside dom₁ K × dom₂ K, Chapter 34 already turns that domain information into finiteness of the mixed points on the same ray.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_small_mixed_finiteness_of_shortRay_mem_effectiveDomains {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (du : Fin m → ℝ) (dv : Fin n → ℝ) (hK : IsGloballyConcaveConvexERealKernel K) (hShortRayDomains : ∃ ρ : ℝ, 0 < ρ ∧ ∀ t : ℝ, 0 < t → t < ρ → (u + t • du) ∈ effectiveDomain₁ K ∧ (v + t • dv) ∈ effectiveDomain₂ K) : ∃ ρ : ℝ, 0 < ρ ∧ ∀ t : ℝ, 0 < t → t < ρ → K (u + t • du) (v + t • dv) ≠ (⊤ : EReal) ∧ K (u + t • du) (v + t • dv) ≠ (⊥ : EReal)

Lean source

Helper for Theorem 35.8: the already-proved reflected axis finiteness radius can be repackaged as short-ray membership in the one-variable slice effective domains through v and u.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_small_reflected_points_mem_sliceEffectiveDomains_of_singletonSliceData {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {uStar : Fin m → ℝ} {vStar : Fin n → ℝ} (du : Fin m → ℝ) (dv : Fin n → ℝ) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hFirstSingleton : partialSubdifferentialInFirstVariable K u v = {uStar}) (hSecondSingleton : partialSubdifferentialInSecondVariable K u v = {vStar}) : ∃ ρ : ℝ, 0 < ρ ∧ ∀ t : ℝ, 0 < t → t < ρ → (u + t • du) ∈ concaveFunctionEffectiveDomain (fun x => K x v) ∧ (u - t • du) ∈ concaveFunctionEffectiveDomain (fun x => K x v) ∧ (v + t • dv) ∈ convexFunctionEffectiveDomain (K u) ∧ (v - t • dv) ∈ convexFunctionEffectiveDomain (K u)

Lean source

Helper for Theorem 35.8: once the base point already lies in dom₁ K × dom₂ K, the short-ray fixed-slice data supplies concrete finite reference points on each moved slice. What still remains later is only the universal ≠ ⊥ / ≠ ⊤ upgrade required by effectiveDomain₁ and effectiveDomain₂.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_movedSlices_finite_referencePoints_of_baseDomains_and_sliceRayDomains {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (du : Fin m → ℝ) (dv : Fin n → ℝ) (huBaseDom : u ∈ effectiveDomain₁ K) (hvBaseDom : v ∈ effectiveDomain₂ K) (hSliceRayDomains : ∃ ρ : ℝ, 0 < ρ ∧ ∀ t : ℝ, 0 < t → t < ρ → (u + t • du) ∈ concaveFunctionEffectiveDomain (fun x => K x v) ∧ (u - t • du) ∈ concaveFunctionEffectiveDomain (fun x => K x v) ∧ (v + t • dv) ∈ convexFunctionEffectiveDomain (K u) ∧ (v - t • dv) ∈ convexFunctionEffectiveDomain (K u)) : ∃ ρ : ℝ, 0 < ρ ∧ ∀ t : ℝ, 0 < t → t < ρ → (K (u + t • du) v ≠ (⊤ : EReal) ∧ K (u + t • du) v ≠ (⊥ : EReal)) ∧ (K (u - t • du) v ≠ (⊤ : EReal) ∧ K (u - t • du) v ≠ (⊥ : EReal)) ∧ (K u (v + t • dv) ≠ (⊤ : EReal) ∧ K u (v + t • dv) ≠ (⊥ : EReal)) ∧ (K u (v - t • dv) ≠ (⊤ : EReal) ∧ K u (v - t • dv) ≠ (⊥ : EReal)) ∧ v ∈ convexFunctionEffectiveDomain (K (u + t • du)) ∧ v ∈ convexFunctionEffectiveDomain (K (u - t • du)) ∧ (-u) ∈ convexFunctionEffectiveDomain (fun x : Fin m → ℝ => -K (-x) (v + t • dv)) ∧ (-u) ∈ convexFunctionEffectiveDomain (fun x : Fin m → ℝ => -K (-x) (v - t • dv))

Lean source

Helper for Theorem 35.8: singleton partial data already packages the positive-ray moved-slice convexity, nonempty effective domains, and concrete finite reference points that are available before the missing properness upgrade.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_positiveRayMovedSliceWitnesses_of_singleton_partials {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {uStar : Fin m → ℝ} {vStar : Fin n → ℝ} (du : Fin m → ℝ) (dv : Fin n → ℝ) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hFirstSingleton : partialSubdifferentialInFirstVariable K u v = {uStar}) (hSecondSingleton : partialSubdifferentialInSecondVariable K u v = {vStar}) : ∃ ρ : ℝ, 0 < ρ ∧ ∀ t : ℝ, 0 < t → t < ρ → ConvexFunction (K (u + t • du)) ∧ (convexFunctionEffectiveDomain (K (u + t • du))).Nonempty ∧ ConvexFunction (fun x : Fin m → ℝ => -K (-x) (v + t • dv)) ∧ (convexFunctionEffectiveDomain (fun x : Fin m → ℝ => -K (-x) (v + t • dv))).Nonempty ∧ (K (u + t • du) v ≠ (⊤ : EReal) ∧ K (u + t • du) v ≠ (⊥ : EReal)) ∧ (K u (v + t • dv) ≠ (⊤ : EReal) ∧ K u (v + t • dv) ≠ (⊥ : EReal)) ∧ v ∈ convexFunctionEffectiveDomain (K (u + t • du)) ∧ (-u) ∈ convexFunctionEffectiveDomain (fun x : Fin m → ℝ => -K (-x) (v + t • dv))

Lean source

Helper for Theorem 35.8: singleton first/second partial data should keep sufficiently small mixed secants on any fixed ray from (u, v) inside the finite part of EReal.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_small_mixed_secants_finite_of_universalMovedSliceClauses {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (du : Fin m → ℝ) (dv : Fin n → ℝ) (hK : IsGloballyConcaveConvexERealKernel K) (hUniversalMovedSlices : ∃ ρ : ℝ, 0 < ρ ∧ ∀ t : ℝ, 0 < t → t < ρ → (∀ y : Fin n → ℝ, K (u + t • du) y ≠ (⊥ : EReal)) ∧ ∀ x : Fin m → ℝ, K x (v + t • dv) ≠ (⊤ : EReal)) : ∃ ρ : ℝ, 0 < ρ ∧ ∀ t : ℝ, 0 < t → t < ρ → K (u + t • du) (v + t • dv) ≠ (⊤ : EReal) ∧ K (u + t • du) (v + t • dv) ≠ (⊥ : EReal)

Lean source

Helper for Theorem 35.8: singleton first/second partial data should force a uniform short-ray universal no-⊥ / no-⊤ statement for the moved slices through (u, v).

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_uniformMovedSlices_noBot_noTop_of_preparedMovedSliceInterior {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (du : Fin m → ℝ) (dv : Fin n → ℝ) (hK : IsGloballyConcaveConvexERealKernel K) (hPreparedMovedSlices : ∃ ρ : ℝ, 0 < ρ ∧ ∀ t : ℝ, 0 < t → t < ρ → ConvexFunction (K (u + t • du)) ∧ (convexFunctionEffectiveDomain (K (u + t • du))).Nonempty ∧ ConvexFunction (fun x : Fin m → ℝ => -K (-x) (v + t • dv)) ∧ (convexFunctionEffectiveDomain (fun x : Fin m → ℝ => -K (-x) (v + t • dv))).Nonempty ∧ (K (u + t • du) v ≠ (⊤ : EReal) ∧ K (u + t • du) v ≠ (⊥ : EReal)) ∧ (K u (v + t • dv) ≠ (⊤ : EReal) ∧ K u (v + t • dv) ≠ (⊥ : EReal)) ∧ v ∈ convexFunctionEffectiveDomain (K (u + t • du)) ∧ (-u) ∈ convexFunctionEffectiveDomain (fun x : Fin m → ℝ => -K (-x) (v + t • dv))) (hMovedInterior : ∃ ρ : ℝ, 0 < ρ ∧ ∀ t : ℝ, 0 < t → t < ρ → v ∈ interior (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (K (u + t • du))) ∧ (-u) ∈ interior (effectiveDomain (Set.univ : Set (Fin m → ℝ)) (fun x : Fin m → ℝ => -K (-x) (v + t • dv)))) : ∃ ρ : ℝ, 0 < ρ ∧ ∀ t : ℝ, 0 < t → t < ρ → (∀ y : Fin n → ℝ, K (u + t • du) y ≠ (⊥ : EReal)) ∧ ∀ x : Fin m → ℝ, K x (v + t • dv) ≠ (⊤ : EReal)

Lean source

Helper for Theorem 35.8: every open neighborhood of a point contains a smaller ball that is stable under reflection through that point.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_reflectionStableBall_subset_of_open {k : ℕ} {c : Fin k → ℝ} {S : Set (Fin k → ℝ)} (hSopen : IsOpen S) (hcS : c ∈ S) : ∃ ε : ℝ, 0 < ε ∧ Metric.ball c ε ⊆ S ∧ ∀ z : Fin k → ℝ, z ∈ Metric.ball c ε → 2 • c - z ∈ S

Lean source

Helper for Theorem 35.8: once a positive-radius ball lies in the effective domain, its center is an interior point of that effective domain.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_memInterior_effectiveDomain_of_ball_subset {k : ℕ} {f : (Fin k → ℝ) → EReal} {c : Fin k → ℝ} {ε : ℝ} (hε : 0 < ε) (hBall : Metric.ball c ε ⊆ effectiveDomain (Set.univ : Set (Fin k → ℝ)) f) : c ∈ interior (effectiveDomain (Set.univ : Set (Fin k → ℝ)) f)

Lean source

Helper for Theorem 35.8: pointwise exclusion of ⊤ on a ball is exactly the ballwise effective-domain inclusion needed for a moved second slice.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_ball_subset_effectiveDomain_of_pointwise_neTop {k : ℕ} {f : (Fin k → ℝ) → EReal} {c : Fin k → ℝ} {ε : ℝ} (hNoTop : ∀ y : Fin k → ℝ, y ∈ Metric.ball c ε → f y ≠ (⊤ : EReal)) : Metric.ball c ε ⊆ effectiveDomain (Set.univ : Set (Fin k → ℝ)) f

Lean source

Helper for Theorem 35.8: pointwise exclusion of ⊥ on a ball around u is exactly the ballwise effective-domain inclusion for the reflected moved first slice around -u.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_reflectedBall_subset_effectiveDomain_of_pointwise_neBot {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {y : Fin n → ℝ} {ε : ℝ} (hNoBot : ∀ x : Fin m → ℝ, x ∈ Metric.ball u ε → K x y ≠ (⊥ : EReal)) : Metric.ball (-u) ε ⊆ effectiveDomain (Set.univ : Set (Fin m → ℝ)) (fun x : Fin m → ℝ => -K (-x) y)

Lean source

Helper for Theorem 35.8: an alternating reflected checkerboard already makes the mixed saddle quotient at unit step equal to ⊤ or ⊥.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_checkerboardForcesInfiniteMixedQuotient_atUnitStep {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {x : Fin m → ℝ} {y : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hChecker : (K x y = (⊤ : EReal) ∧ K (2 • u - x) y = (⊥ : EReal) ∧ K (2 • u - x) (2 • v - y) = (⊤ : EReal) ∧ K x (2 • v - y) = (⊥ : EReal)) ∨ (K x y = (⊥ : EReal) ∧ K (2 • u - x) y = (⊤ : EReal) ∧ K (2 • u - x) (2 • v - y) = (⊥ : EReal) ∧ K x (2 • v - y) = (⊤ : EReal))) : saddleDirectionalDifferenceQuotientAt K u v (x - u) (y - v) 1 = (⊤ : EReal) ∨ saddleDirectionalDifferenceQuotientAt K u v (x - u) (y - v) 1 = (⊥ : EReal)

Lean source

Helper for Theorem 35.8: a bad fixed-time checkerboard on the second-coordinate reflection pair already forces the mixed quotient along (t • du, y - v) to be infinite at unit step.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_fixedTimeSecondBall_checkerboardForcesInfiniteMixedQuotient {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (du : Fin m → ℝ) {t : ℝ} {y : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hChecker : (K (u + t • du) y = (⊤ : EReal) ∧ K (u - t • du) y = (⊥ : EReal) ∧ K (u - t • du) (2 • v - y) = (⊤ : EReal) ∧ K (u + t • du) (2 • v - y) = (⊥ : EReal)) ∨ (K (u + t • du) y = (⊥ : EReal) ∧ K (u - t • du) y = (⊤ : EReal) ∧ K (u - t • du) (2 • v - y) = (⊥ : EReal) ∧ K (u + t • du) (2 • v - y) = (⊤ : EReal))) : saddleDirectionalDifferenceQuotientAt K u v (t • du) (y - v) 1 = (⊤ : EReal) ∨ saddleDirectionalDifferenceQuotientAt K u v (t • du) (y - v) 1 = (⊥ : EReal)

Lean source

Helper for Theorem 35.8: a bad fixed-time checkerboard on the first-coordinate reflection pair already forces the mixed quotient along (x - u, t • dv) to be infinite at unit step.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_fixedTimeFirstBall_checkerboardForcesInfiniteMixedQuotient {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (dv : Fin n → ℝ) {t : ℝ} {x : Fin m → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hChecker : (K x (v + t • dv) = (⊤ : EReal) ∧ K (2 • u - x) (v + t • dv) = (⊥ : EReal) ∧ K (2 • u - x) (v - t • dv) = (⊤ : EReal) ∧ K x (v - t • dv) = (⊥ : EReal)) ∨ (K x (v + t • dv) = (⊥ : EReal) ∧ K (2 • u - x) (v + t • dv) = (⊤ : EReal) ∧ K (2 • u - x) (v - t • dv) = (⊥ : EReal) ∧ K x (v - t • dv) = (⊤ : EReal))) : saddleDirectionalDifferenceQuotientAt K u v (x - u) (t • dv) 1 = (⊤ : EReal) ∨ saddleDirectionalDifferenceQuotientAt K u v (x - u) (t • dv) 1 = (⊥ : EReal)

Lean source

Helper for Theorem 35.8: once the four reflected axis values are finite, any extreme value at the mixed corner K x y already normalizes to one of the two reflected checkerboard branches.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_extremeMixedCorner_forces_reflectedCheckerboard {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {x : Fin m → ℝ} {y : Fin n → ℝ} (hK : IsGloballyConcaveConvexERealKernel K) (hxFinite : K x v ≠ (⊤ : EReal) ∧ K x v ≠ (⊥ : EReal)) (hxRefFinite : K (2 • u - x) v ≠ (⊤ : EReal) ∧ K (2 • u - x) v ≠ (⊥ : EReal)) (hyFinite : K u y ≠ (⊤ : EReal) ∧ K u y ≠ (⊥ : EReal)) (hyRefFinite : K u (2 • v - y) ≠ (⊤ : EReal) ∧ K u (2 • v - y) ≠ (⊥ : EReal)) (hxyExtreme : K x y = (⊤ : EReal) ∨ K x y = (⊥ : EReal)) : (K x y = (⊤ : EReal) ∧ K (2 • u - x) y = (⊥ : EReal) ∧ K (2 • u - x) (2 • v - y) = (⊤ : EReal) ∧ K x (2 • v - y) = (⊥ : EReal)) ∨ (K x y = (⊥ : EReal) ∧ K (2 • u - x) y = (⊤ : EReal) ∧ K (2 • u - x) (2 • v - y) = (⊥ : EReal) ∧ K x (2 • v - y) = (⊤ : EReal))

§35 — Continuity and Differentiability of Saddle-Functions — Part 21

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_linear_saddleDirectionalDerivative_of_singleton_partials {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {uStar : Fin m → ℝ} {vStar : Fin n → ℝ} (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hFiniteRect : ∃ C : Set (Fin m → ℝ), ∃ D : Set (Fin n → ℝ), IsOpen C ∧ u ∈ C ∧ Convex ℝ C ∧ IsOpen D ∧ v ∈ D ∧ Convex ℝ D ∧ ∀ u' ∈ C, ∀ v' ∈ D, K u' v' ≠ (⊤ : EReal) ∧ K u' v' ≠ (⊥ : EReal)) (hFirstSingleton : partialSubdifferentialInFirstVariable K u v = {uStar}) (hSecondSingleton : partialSubdifferentialInSecondVariable K u v = {vStar}) : ∀ u' v', IsSaddleDirectionalDerivativeAt K u v u' v' (((((∑ i : Fin m, uStar i * u' i) + ∑ j : Fin n, vStar j * v' j) : ℝ) : EReal))

Lean source

Helper for Theorem 35.8: a finite open rectangle immediately gives interior membership in the saddle effective domain.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_memInterior_saddleFunctionEffectiveDomain_of_finiteRectangle {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hCopen : IsOpen C) (huC : u ∈ C) (hDopen : IsOpen D) (hvD : v ∈ D) (hFinite : ∀ u' ∈ C, ∀ v' ∈ D, K u' v' ≠ (⊤ : EReal) ∧ K u' v' ≠ (⊥ : EReal)) : (u, v) ∈ interior (saddleFunctionEffectiveDomain K)

Lean source

Helper for Theorem 35.8: interior membership in the saddle effective domain yields the finite open convex rectangle required by Theorem 35.6.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_finiteRectangle_of_memInterior_effectiveDomain {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hInterior : (u, v) ∈ interior (saddleFunctionEffectiveDomain K)) : ∃ C : Set (Fin m → ℝ), ∃ D : Set (Fin n → ℝ), IsOpen C ∧ u ∈ C ∧ Convex ℝ C ∧ IsOpen D ∧ v ∈ D ∧ Convex ℝ D ∧ ∀ u' ∈ C, ∀ v' ∈ D, K u' v' ≠ (⊤ : EReal) ∧ K u' v' ≠ (⊥ : EReal)

Lean source

Helper for Theorem 35.8: on any finite open convex rectangle around (u, v), the real-valued kernel obtained by taking toReal has saddle subdifferential exactly {(uStar, vStar)} once the global singleton partials and the linear mixed directional-derivative formula are available.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_localRealSingletonSubgradient_onRectangle {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {uStar : Fin m → ℝ} {vStar : Fin n → ℝ} (hK : IsGloballyConcaveConvexERealKernel K) (hFiniteRect : ∃ C : Set (Fin m → ℝ), ∃ D : Set (Fin n → ℝ), IsOpen C ∧ u ∈ C ∧ Convex ℝ C ∧ IsOpen D ∧ v ∈ D ∧ Convex ℝ D ∧ ∀ u' ∈ C, ∀ v' ∈ D, K u' v' ≠ (⊤ : EReal) ∧ K u' v' ≠ (⊥ : EReal)) (hFirstSingleton : partialSubdifferentialInFirstVariable K u v = {uStar}) (hSecondSingleton : partialSubdifferentialInSecondVariable K u v = {vStar}) (hLinearDir : ∀ u' v', IsSaddleDirectionalDerivativeAt K u v u' v' (((((∑ i : Fin m, uStar i * u' i) + ∑ j : Fin n, vStar j * v' j) : ℝ) : EReal))) : ∃ C : Set (Fin m → ℝ), ∃ D : Set (Fin n → ℝ), IsOpen C ∧ u ∈ C ∧ Convex ℝ C ∧ IsOpen D ∧ v ∈ D ∧ Convex ℝ D ∧ (∀ u' ∈ C, ∀ v' ∈ D, K u' v' ≠ (⊤ : EReal) ∧ K u' v' ≠ (⊥ : EReal)) ∧ let Kloc : (Fin m → ℝ) → (Fin n → ℝ) → ℝ := fun x y => (K x y).toReal IsRealConcaveConvexOn C D Kloc ∧ realSaddleSubdifferentialOn C D Kloc u v = {(uStar, vStar)}

Lean source

Helper for Theorem 35.8: the ℓ¹/ℓ∞ estimate controls the absolute value of a dot product.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_abs_dotProduct_le_l1Norm_mul_norm {k : ℕ} (a b : Fin k → ℝ) :

Lean source

Helper for Theorem 35.8: the first block of a packed vector has norm bounded by the packed norm.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_norm_le_norm_append_left {m n : ℕ} (a : Fin m → ℝ) (b : Fin n → ℝ) : ‖a‖ ≤ ‖Fin.append a b‖

Lean source

Helper for Theorem 35.8: the second block of a packed vector has norm bounded by the packed norm.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_norm_le_norm_append_right {m n : ℕ} (a : Fin m → ℝ) (b : Fin n → ℝ) : ‖b‖ ≤ ‖Fin.append a b‖

Lean source

Helper for Theorem 35.8: on Fin k → ℝ, the ℓ¹ norm is bounded by the dimension times the sup norm.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_l1Norm_le_card_mul_norm {k : ℕ} (x : Fin k → ℝ) : l1Norm x ≤ (k : ℝ) * ‖x‖

Lean source

Helper for Theorem 35.8: every interior point of the finite open rectangle has a nonempty real saddle subdifferential.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_nonempty_realSaddleSubdifferentialOn_of_mem_openRectangle {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hCopen : IsOpen C) (hDopen : IsOpen D) (hCconv : Convex ℝ C) (hDconv : Convex ℝ D) (hK : IsRealConcaveConvexOn C D K) {x : Fin m → ℝ} {y : Fin n → ℝ} (hx : x ∈ C) (hy : y ∈ D) : Set.Nonempty (realSaddleSubdifferentialOn C D K x y)

Lean source

Helper for Theorem 35.8: on a finite open rectangle, Corollary 35.7.1 upgrades a singleton base saddle subgradient into split-ball control of every nearby saddle subgradient.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_nearbyRealSubgradient_close_to_singleton {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {Kloc : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {u : Fin m → ℝ} {v : Fin n → ℝ} {uStar : Fin m → ℝ} {vStar : Fin n → ℝ} (hCopen : IsOpen C) (huC : u ∈ C) (hCconv : Convex ℝ C) (hDopen : IsOpen D) (hvD : v ∈ D) (hDconv : Convex ℝ D) (hRealCC : IsRealConcaveConvexOn C D Kloc) (hLocal : realSaddleSubdifferentialOn C D Kloc u v = {(uStar, vStar)}) : ∀ η : ℝ, 0 < η → ∃ δ : ℝ, 0 < δ ∧ ∀ x : Fin m → ℝ, x ∈ C → ∀ y : Fin n → ℝ, y ∈ D → ((x - u), (y - v)) ∈ splitEuclideanClosedBall (m := m) (n := n) δ → ∀ ⦃p : Fin m → ℝ⦄ ⦃q : Fin n → ℝ⦄, (p, q) ∈ realSaddleSubdifferentialOn C D Kloc x y → ((p - uStar), (q - vStar)) ∈ splitEuclideanClosedBall (m := m) (n := n) η

Lean source

Helper for Theorem 35.8: once a nearby local saddle subgradient (p, q) is chosen, the subgradient inequalities sandwich the packed real remainder by explicit ℓ¹/ℓ∞ errors.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_packedRealErrorBound_of_nearbySingletonSubgradients {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {Kloc : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {u : Fin m → ℝ} {v : Fin n → ℝ} {uStar : Fin m → ℝ} {vStar : Fin n → ℝ} (huC : u ∈ C) (hvD : v ∈ D) (hBase : realSaddleSubdifferentialOn C D Kloc u v = {(uStar, vStar)}) {x : Fin m → ℝ} {y : Fin n → ℝ} (hxC : x ∈ C) (hyD : y ∈ D) {p : Fin m → ℝ} {q : Fin n → ℝ} (hNear : (p, q) ∈ realSaddleSubdifferentialOn C D Kloc x y) :

§35 — Continuity and Differentiability of Saddle-Functions — Part 22

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Theorem 35.8, with the local-finiteness qualification made explicit for extended-real kernels. Differentiability gives a unique saddle subgradient. Conversely, uniqueness gives differentiability once K is finite on an open convex rectangle around the base point. The qualification is necessary for the present extended-real API: singleton coordinate subdifferentials alone do not exclude a separately convex-concave ⊤/⊥ checkerboard away from the two coordinate axes.

open scoped Pointwise open scoped Topology theorem section35_theorem35_8 {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) : ((hDiff : ERealDifferentiableAt (packedSaddleKernel K) (Fin.append u v)) → let grad : (Fin m → ℝ) × (Fin n → ℝ) := packedSaddleKernelGradientPairAt (K := K) (u := u) (v := v) hDiff grad ∈ productSubdifferentialAt K u v ∧ ∀ g : (Fin m → ℝ) × (Fin n → ℝ), g ∈ productSubdifferentialAt K u v → g = grad) ∧ ((hFiniteRect : ∃ C : Set (Fin m → ℝ), ∃ D : Set (Fin n → ℝ), IsOpen C ∧ u ∈ C ∧ Convex ℝ C ∧ IsOpen D ∧ v ∈ D ∧ Convex ℝ D ∧ ∀ u' ∈ C, ∀ v' ∈ D, K u' v' ≠ (⊤ : EReal) ∧ K u' v' ≠ (⊥ : EReal)) → (∃! g : (Fin m → ℝ) × (Fin n → ℝ), g ∈ productSubdifferentialAt K u v) → ERealDifferentiableAt (packedSaddleKernel K) (Fin.append u v))

Lean source

Helper for Theorem 35.8: once the local real rectangle kernel has singleton saddle subdifferential at (u, v), the remaining analytic task is to turn that local singleton control into differentiability of the packed real map on the rectangle.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_puncturedPackedRealRemainder_small {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {uStar : Fin m → ℝ} {vStar : Fin n → ℝ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hCopen : IsOpen C) (huC : u ∈ C) (hCconv : Convex ℝ C) (hDopen : IsOpen D) (hvD : v ∈ D) (hDconv : Convex ℝ D) (hFiniteCD : ∀ x ∈ C, ∀ y ∈ D, K x y ≠ (⊤ : EReal) ∧ K x y ≠ (⊥ : EReal)) (hLocal : let Kloc : (Fin m → ℝ) → (Fin n → ℝ) → ℝ := fun x y => (K x y).toReal IsRealConcaveConvexOn C D Kloc ∧ realSaddleSubdifferentialOn C D Kloc u v = {(uStar, vStar)}) : let Kloc : (Fin m → ℝ) → (Fin n → ℝ) → ℝ := fun x y => (K x y).toReal let fLoc : (Fin (m + n) → ℝ) → ℝ := fun z => Kloc (fun i => z (Fin.castAdd n i)) (fun j => z (Fin.natAdd m j)) let L : (Fin (m + n) → ℝ) →L[ℝ] ℝ := helperForCorollary_25_5_1_dotProductContinuousLinearMap (Fin.append uStar vStar) ∀ ε : ℝ, 0 < ε → ∀ᶠ z in 𝓝[≠] (Fin.append u v), ‖z - Fin.append u v‖⁻¹ * ‖fLoc z - fLoc (Fin.append u v) - L (z - Fin.append u v)‖ ≤ ε

Lean source

Helper for Theorem 35.8: once the local real rectangle kernel has singleton saddle subdifferential at (u, v), the remaining analytic task is to turn that local singleton control into differentiability of the packed real map on the rectangle.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_packedRealDifferentiable_of_localSingletonSubgradient {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {uStar : Fin m → ℝ} {vStar : Fin n → ℝ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hCopen : IsOpen C) (huC : u ∈ C) (hCconv : Convex ℝ C) (hDopen : IsOpen D) (hvD : v ∈ D) (hDconv : Convex ℝ D) (hFiniteCD : ∀ x ∈ C, ∀ y ∈ D, K x y ≠ (⊤ : EReal) ∧ K x y ≠ (⊥ : EReal)) (hLocal : let Kloc : (Fin m → ℝ) → (Fin n → ℝ) → ℝ := fun x y => (K x y).toReal IsRealConcaveConvexOn C D Kloc ∧ realSaddleSubdifferentialOn C D Kloc u v = {(uStar, vStar)}) : let Kloc : (Fin m → ℝ) → (Fin n → ℝ) → ℝ := fun x y => (K x y).toReal let fLoc : (Fin (m + n) → ℝ) → ℝ := fun z => Kloc (fun i => z (Fin.castAdd n i)) (fun j => z (Fin.natAdd m j)) DifferentiableAt ℝ fLoc (Fin.append u v)

Lean source

Helper for Theorem 35.8: a real Fréchet differentiability witness can be coerced to an everywhere-finite EReal differentiability witness by taking the trivial Set.univ extension.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_ERealDifferentiableAt_coe_of_realDifferentiableAt {k : ℕ} {f : (Fin k → ℝ) → ℝ} {x : Fin k → ℝ} (hdiff : DifferentiableAt ℝ f x) : ERealDifferentiableAt (fun z => ((f z : ℝ) : EReal)) x

Lean source

Helper for Theorem 35.8: EReal differentiability transfers across equality on an open neighborhood once the target function is finite on that neighborhood.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_ERealDifferentiableAt_of_eqOn_open {k : ℕ} {f g : (Fin k → ℝ) → EReal} {x : Fin k → ℝ} {W : Set (Fin k → ℝ)} (hWopen : IsOpen W) (hxW : x ∈ W) (hEqOn : ∀ z ∈ W, f z = g z) (hFiniteOn : ∀ z ∈ W, f z ≠ (⊤ : EReal) ∧ f z ≠ (⊥ : EReal)) (hg : ERealDifferentiableAt g x) : ERealDifferentiableAt f x

Lean source

Helper for Theorem 35.8: after the packed real map is differentiable on a finite open rectangle, the Chapter 25 +∞ extension principle lifts that local real differentiability back to EReal differentiability of the packed saddle kernel.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_localPackedExtension_of_realRectangleKernel {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hCopen : IsOpen C) (huC : u ∈ C) (hDopen : IsOpen D) (hvD : v ∈ D) (hFiniteCD : ∀ x ∈ C, ∀ y ∈ D, K x y ≠ (⊤ : EReal) ∧ K x y ≠ (⊥ : EReal)) (hPackedDiff : let Kloc : (Fin m → ℝ) → (Fin n → ℝ) → ℝ := fun x y => (K x y).toReal let fLoc : (Fin (m + n) → ℝ) → ℝ := fun z => Kloc (fun i => z (Fin.castAdd n i)) (fun j => z (Fin.natAdd m j)) DifferentiableAt ℝ fLoc (Fin.append u v)) : ERealDifferentiableAt (packedSaddleKernel K) (Fin.append u v)

Lean source

Helper for Theorem 35.8: a finite local rectangle together with a linear mixed saddle directional derivative should package into packed EReal differentiability.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_8_packedDifferentiable_of_linear_saddleDirectionalDerivative {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {uStar : Fin m → ℝ} {vStar : Fin n → ℝ} (hK : IsGloballyConcaveConvexERealKernel K) (hFiniteRect : ∃ C : Set (Fin m → ℝ), ∃ D : Set (Fin n → ℝ), IsOpen C ∧ u ∈ C ∧ Convex ℝ C ∧ IsOpen D ∧ v ∈ D ∧ Convex ℝ D ∧ ∀ u' ∈ C, ∀ v' ∈ D, K u' v' ≠ (⊤ : EReal) ∧ K u' v' ≠ (⊥ : EReal)) (hFirstSingleton : partialSubdifferentialInFirstVariable K u v = {uStar}) (hSecondSingleton : partialSubdifferentialInSecondVariable K u v = {vStar}) (hDir : ∀ u' v', IsSaddleDirectionalDerivativeAt K u v u' v' (((((∑ i : Fin m, uStar i * u' i) + ∑ j : Fin n, vStar j * v' j) : ℝ) : EReal))) : ERealDifferentiableAt (packedSaddleKernel K) (Fin.append u v)

Lean source

Helper for Corollary 35.8.1: a finite neighborhood around (u, v) contains an open convex product rectangle on which K stays finite.

open scoped Pointwise open scoped Topology lemma helperForCorollary_35_8_1_finiteRectangle_of_neighborhood {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hNeighborhood : SaddleKernelFiniteOnNeighborhoodAt K u v) : ∃ C : Set (Fin m → ℝ), ∃ D : Set (Fin n → ℝ), IsOpen C ∧ u ∈ C ∧ Convex ℝ C ∧ IsOpen D ∧ v ∈ D ∧ Convex ℝ D ∧ ∀ u' ∈ C, ∀ v' ∈ D, K u' v' ≠ (⊤ : EReal) ∧ K u' v' ≠ (⊥ : EReal)

Lean source

Helper for Corollary 35.8.1: first-variable partial subgradients are exactly the subgradients of the convex slice x ↦ -K x v with the expected sign change.

open scoped Pointwise open scoped Topology lemma helperForCorollary_35_8_1_negFirstSliceSubgradient_iff_partialFirstMem {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (uStar : Fin m → ℝ) : IsSubgradientAt (fun x : Fin m → ℝ => -K x v) u (dotProductEquiv ℝ (Fin m) (-uStar)) ↔ uStar ∈ partialSubdifferentialInFirstVariable K u v

Lean source

Helper for Corollary 35.8.1: on a finite rectangle, a linear saddle directional derivative forces both partial subdifferentials to be singletons.

open scoped Pointwise open scoped Topology lemma helperForCorollary_35_8_1_singletonPartials_of_linearSaddleDirectionalDerivative {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} {uStar : Fin m → ℝ} {vStar : Fin n → ℝ} (hK : IsGloballyConcaveConvexERealKernel K) (hFiniteRect : ∃ C : Set (Fin m → ℝ), ∃ D : Set (Fin n → ℝ), IsOpen C ∧ u ∈ C ∧ Convex ℝ C ∧ IsOpen D ∧ v ∈ D ∧ Convex ℝ D ∧ ∀ u' ∈ C, ∀ v' ∈ D, K u' v' ≠ (⊤ : EReal) ∧ K u' v' ≠ (⊥ : EReal)) (hDir : ∀ u' v', IsSaddleDirectionalDerivativeAt K u v u' v' (((((∑ i : Fin m, uStar i * u' i) + ∑ j : Fin n, vStar j * v' j) : ℝ) : EReal))) : partialSubdifferentialInFirstVariable K u v = {uStar} ∧ partialSubdifferentialInSecondVariable K u v = {vStar}

Lean source

Helper for Corollary 35.8.1: the reflected first-slice quotient is exactly the negative of the packed quotient in the matching first-block direction.

open scoped Pointwise open scoped Topology lemma helperForCorollary_35_8_1_reflectedFirstSliceQuotient_eq_negPackedQuotient {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (uDir : Fin m → ℝ) (t : ℝ) : directionalDifferenceQuotientAt (fun x : Fin m → ℝ => -K (-x) v) (-u) uDir t = -directionalDifferenceQuotientAt (packedSaddleKernel K) (Fin.append u v) (Fin.append (-uDir) (0 : Fin n → ℝ)) t

Lean source

Helper for Corollary 35.8.1: along the positive ith first-variable basis vector, the reflected slice quotient equals the negative packed quotient along the negative packed basis.

open scoped Pointwise open scoped Topology lemma helperForCorollary_35_8_1_reflectedFirstBasisQuotient_eq_negPackedNegativeBasisQuotient {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (i : Fin m) (t : ℝ) : directionalDifferenceQuotientAt (fun x : Fin m → ℝ => -K (-x) v) (-u) (Pi.single i (1 : ℝ)) t = -directionalDifferenceQuotientAt (packedSaddleKernel K) (Fin.append u v) (-(Pi.single (Fin.castAdd n i) (1 : ℝ) : Fin (m + n) → ℝ)) t

Lean source

Helper for Corollary 35.8.1: along the negative ith first-variable basis vector, the reflected slice quotient equals the negative packed quotient along the positive packed basis.

open scoped Pointwise open scoped Topology lemma helperForCorollary_35_8_1_reflectedNegativeBasisQuotient_eq_negPackedPositiveBasisQuotient {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (i : Fin m) (t : ℝ) : directionalDifferenceQuotientAt (fun x : Fin m → ℝ => -K (-x) v) (-u) (-(Pi.single i (1 : ℝ) : Fin m → ℝ)) t = -directionalDifferenceQuotientAt (packedSaddleKernel K) (Fin.append u v) (Pi.single (Fin.castAdd n i) (1 : ℝ)) t

§35 — Continuity and Differentiability of Saddle-Functions — Part 23

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Differentiability of the packed extended-real map supplies the local finiteness qualification used in the corrected form of Theorem 35.8.

open scoped Pointwise open scoped Topology lemma helperForCorollary_35_8_1_finiteNeighborhood_of_packedDifferentiable {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hDiff : ERealDifferentiableAt (packedSaddleKernel K) (Fin.append u v)) : SaddleKernelFiniteOnNeighborhoodAt K u v

Lean source

Corollary 35.8.1, qualified extended-real form: differentiability is equivalent to local finiteness together with a linear saddle directional derivative. The unqualified claim that finite two-sided coordinate derivatives alone imply local finiteness is omitted: the same off-axis ⊤/⊥ checkerboard gives a counterexample.

open scoped Pointwise open scoped Topology theorem section35_corollary35_8_1 {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) : ERealDifferentiableAt (packedSaddleKernel K) (Fin.append u v) ↔ SaddleKernelFiniteOnNeighborhoodAt K u v ∧ HasLinearSaddleDirectionalDerivativeAt K u v

Lean source

The packed real-valued map on ℝ^(m+n) associated to a saddle kernel K.

open scoped Pointwise open scoped Topology def packedRealSaddleKernel {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) : (Fin (m + n) → ℝ) → ℝ := fun z => K (fun i => z (Fin.castAdd n i)) (fun j => z (Fin.natAdd m j))

Lean source

The split Fréchet-derivative vector of the packed real saddle kernel at (u, v).

open scoped Pointwise open scoped Topology noncomputable def packedRealSaddleKernelGradientPair {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) (u : Fin m → ℝ) (v : Fin n → ℝ) : (Fin m → ℝ) × (Fin n → ℝ) := (fun i => fderiv ℝ (packedRealSaddleKernel K) (Fin.append u v) (Pi.single (Fin.castAdd n i) 1), fun j => fderiv ℝ (packedRealSaddleKernel K) (Fin.append u v) (Pi.single (Fin.natAdd m j) 1))

Lean source

Helper for Corollary 35.8.1: the packed quotient along a second-block basis vector is exactly the quotient of the honest second slice y ↦ K u y along the corresponding coordinate basis.

open scoped Pointwise open scoped Topology lemma helperForCorollary_35_8_1_packedSecondBasisQuotient_eq_secondSliceBasisQuotient {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (v : Fin n → ℝ) (j : Fin n) (t : ℝ) : directionalDifferenceQuotientAt (packedSaddleKernel K) (Fin.append u v) (Pi.single (Fin.natAdd m j) (1 : ℝ)) t = directionalDifferenceQuotientAt (K u) v (Pi.single j (1 : ℝ)) t

Lean source

Helper for Corollary 35.8.1: a finite packed first-block coordinate partial transfers to the corresponding coordinate partial of the reflected convex slice x ↦ -K (-x) v.

open scoped Pointwise open scoped Topology lemma helperForCorollary_35_8_1_reflectedFirstSliceCoordinatePartial_of_packedFirstCoordinatePartial {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (i : Fin m) (L : ℝ) (hpartial : HasCoordinatePartialDerivativeAt (packedSaddleKernel K) (Fin.append u v) (Fin.castAdd n i) (L : EReal)) : HasCoordinatePartialDerivativeAt (fun x : Fin m → ℝ => -K (-x) v) (-u) i (L : EReal)

Lean source

Helper for Corollary 35.8.1: finite packed coordinate partials induce coordinate partials on the reflected first slice and the ordinary second slice needed for the Chapter 25 slice theorem.

open scoped Pointwise open scoped Topology lemma helperForCorollary_35_8_1_sliceCoordinatePartials_of_packedCoordinatePartials {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartials : ∀ j : Fin (m + n), ∃ L : ℝ, HasCoordinatePartialDerivativeAt (packedSaddleKernel K) (Fin.append u v) j (L : EReal)) : (∀ i : Fin m, ∃ L : ℝ, HasCoordinatePartialDerivativeAt (fun x : Fin m → ℝ => -K (-x) v) (-u) i (L : EReal)) ∧ (∀ j : Fin n, ∃ L : ℝ, HasCoordinatePartialDerivativeAt (K u) v j (L : EReal))

Lean source

Helper for Corollary 35.8.1: if all packed coordinate partials exist and are finite, then the two slice subdifferentials are singletons, hence so are the saddle partial subdifferentials.

open scoped Pointwise open scoped Topology lemma helperForCorollary_35_8_1_singletonPartials_of_coordinatePartials {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → EReal} {u : Fin m → ℝ} {v : Fin n → ℝ} (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hpartials : ∀ j : Fin (m + n), ∃ L : ℝ, HasCoordinatePartialDerivativeAt (packedSaddleKernel K) (Fin.append u v) j (L : EReal)) : ∃ uStar : Fin m → ℝ, ∃ vStar : Fin n → ℝ, partialSubdifferentialInFirstVariable K u v = {uStar} ∧ partialSubdifferentialInSecondVariable K u v = {vStar}

Lean source

Helper for Theorem 35.9: differentiability of the honest product map Function.uncurry K is equivalent to differentiability of the packed map under Fin.append.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_9_uncurriedDifferentiableAt_iff_packedDifferentiableAt {m n : ℕ} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {u : Fin m → ℝ} {v : Fin n → ℝ} : DifferentiableAt ℝ (Function.uncurry K) (u, v) ↔ DifferentiableAt ℝ (packedRealSaddleKernel K) (Fin.append u v)

Lean source

Helper for Theorem 35.9: on a ball whose doubled closed ball stays inside C × D, the exceptional set where the packed map fails to be differentiable has measure zero.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_9_nullExceptionalSet_onBall {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsRealConcaveConvexOn C D K) (c : (Fin m → ℝ) × (Fin n → ℝ)) {r : ℝ} (hr : 0 < r) (hclosedSub : Metric.closedBall c (2 * r) ⊆ C ×ˢ D) : MeasureTheory.volume (Metric.ball c r \ {p | p ∈ C ×ˢ D ∧ DifferentiableAt ℝ (packedRealSaddleKernel K) (Fin.append p.1 p.2)}) = 0

Lean source

Helper for Theorem 35.9: at a packed differentiability point in C × D, the real saddle subdifferential is the singleton determined by the split packed gradient.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_9_realSaddleSubdifferential_eq_singleton_of_mem_E {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsRealConcaveConvexOn C D K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) (hdiff : DifferentiableAt ℝ (packedRealSaddleKernel K) (Fin.append u v)) : realSaddleSubdifferentialOn C D K u v = {packedRealSaddleKernelGradientPair K u v}

§35 — Continuity and Differentiability of Saddle-Functions — Part 24

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Theorem 35.9: let C × D be an open convex set in ℝ^m × ℝ^n, and let K be a concave-convex real-valued function on C × D. If E is the subset of C × D where K is differentiable, then E is dense in C × D, the complement (C × D) \ E has Lebesgue measure zero, and the gradient mapping is continuous on E. The differentiability and gradient are expressed below via the packed map z ↦ K(z₁, z₂) on ℝ^(m+n), which is equivalent to differentiability of K on the product space.

open scoped Pointwise open scoped Topology theorem section35_theorem35_9 {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsRealConcaveConvexOn C D K) : let E : Set ((Fin m → ℝ) × (Fin n → ℝ)) := {p | p ∈ C ×ˢ D ∧ DifferentiableAt ℝ (packedRealSaddleKernel K) (Fin.append p.1 p.2)} C ×ˢ D ⊆ closure E ∧ MeasureTheory.volume ((C ×ˢ D) \ E) = 0 ∧ Continuous (fun p : {p // p ∈ E} => packedRealSaddleKernelGradientPair K p.1.1 p.1.2)

Lean source

Theorem 35.10: let C × D be an open convex set in ℝ^m × ℝ^n, let K be a finite differentiable concave-convex function on C × D, and let K₁, K₂, ... be finite differentiable concave-convex functions on C × D converging pointwise to K. Then the split gradient maps ∇Kᵢ(u, v) converge pointwise to ∇K(u, v) for every (u, v) ∈ C × D, and in fact converge uniformly on every closed bounded subset of C × D.

open scoped Pointwise open scoped Topology theorem section35_theorem35_10 {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {KSeq : ℕ → (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsRealConcaveConvexOn C D K) (hK_diff : ∀ u ∈ C, ∀ v ∈ D, DifferentiableAt ℝ (packedRealSaddleKernel K) (Fin.append u v)) (hKSeq : ∀ i, IsRealConcaveConvexOn C D (KSeq i)) (hKSeq_diff : ∀ i, ∀ u ∈ C, ∀ v ∈ D, DifferentiableAt ℝ (packedRealSaddleKernel (KSeq i)) (Fin.append u v)) (hpointwise : ∀ u ∈ C, ∀ v ∈ D, Filter.Tendsto (fun i : ℕ => KSeq i u v) Filter.atTop (nhds (K u v))) : (∀ u ∈ C, ∀ v ∈ D, Filter.Tendsto (fun i : ℕ => packedRealSaddleKernelGradientPair (KSeq i) u v) Filter.atTop (nhds (packedRealSaddleKernelGradientPair K u v))) ∧ ∀ S : Set ((Fin m → ℝ) × (Fin n → ℝ)), S ⊆ C ×ˢ D → IsClosed S → Bornology.IsBounded S → TendstoUniformlyOn (fun i p => packedRealSaddleKernelGradientPair (KSeq i) p.1 p.2) (fun p => packedRealSaddleKernelGradientPair K p.1 p.2) Filter.atTop S

Lean source

Text 35.10.1: let C × D be a nonempty open convex subset of ℝ^m × ℝ^n, let K be a finite differentiable concave-convex function on C × D, and let K₁, K₂, ... be finite differentiable concave-convex functions on C × D. If there exist dense subsets C' ⊆ C and D' ⊆ D such that (A) Kᵢ(u, v) → K(u, v) for every (u, v) ∈ C' × D', then (B) Kᵢ(u, v) → K(u, v) for every (u, v) ∈ C × D. Consequently, the conclusion of Theorem 35.10 holds: the split gradient maps converge pointwise on C × D and uniformly on each closed bounded subset of C × D.

open scoped Pointwise open scoped Topology theorem section35_text35_10_1 {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {KSeq : ℕ → (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hNonempty : (C ×ˢ D).Nonempty) (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsRealConcaveConvexOn C D K) (hK_diff : ∀ u ∈ C, ∀ v ∈ D, DifferentiableAt ℝ (packedRealSaddleKernel K) (Fin.append u v)) (hKSeq : ∀ i, IsRealConcaveConvexOn C D (KSeq i)) (hKSeq_diff : ∀ i, ∀ u ∈ C, ∀ v ∈ D, DifferentiableAt ℝ (packedRealSaddleKernel (KSeq i)) (Fin.append u v)) (hDense : ∃ C' : Set (Fin m → ℝ), ∃ D' : Set (Fin n → ℝ), C' ⊆ C ∧ D' ⊆ D ∧ C ⊆ closure C' ∧ D ⊆ closure D' ∧ ∀ u ∈ C', ∀ v ∈ D', Filter.Tendsto (fun i : ℕ => KSeq i u v) Filter.atTop (nhds (K u v))) : (∀ u ∈ C, ∀ v ∈ D, Filter.Tendsto (fun i : ℕ => KSeq i u v) Filter.atTop (nhds (K u v))) ∧ (∀ u ∈ C, ∀ v ∈ D, Filter.Tendsto (fun i : ℕ => packedRealSaddleKernelGradientPair (KSeq i) u v) Filter.atTop (nhds (packedRealSaddleKernelGradientPair K u v))) ∧ ∀ S : Set ((Fin m → ℝ) × (Fin n → ℝ)), S ⊆ C ×ˢ D → IsClosed S → Bornology.IsBounded S → TendstoUniformlyOn (fun i p => packedRealSaddleKernelGradientPair (KSeq i) p.1 p.2) (fun p => packedRealSaddleKernelGradientPair K p.1 p.2) Filter.atTop S

Lean source

Helper for Theorem 35.9: the packed gradient pair varies continuously on the differentiability locus because nearby saddle subgradients stay close to the singleton base subgradient.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_9_gradientPair_continuousOn_E {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsRealConcaveConvexOn C D K) : let E : Set ((Fin m → ℝ) × (Fin n → ℝ)) := {p | p ∈ C ×ˢ D ∧ DifferentiableAt ℝ (packedRealSaddleKernel K) (Fin.append p.1 p.2)} Continuous (fun p : {p // p ∈ E} => packedRealSaddleKernelGradientPair K p.1.1 p.1.2)

Lean source

Helper for Theorem 35.10: since K is differentiable at every point of C × D, Theorem 35.9 identifies the packed gradient pair as a continuous map on the whole product domain.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_10_limitGradient_continuousOn_product {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsRealConcaveConvexOn C D K) (hK_diff : ∀ u ∈ C, ∀ v ∈ D, DifferentiableAt ℝ (packedRealSaddleKernel K) (Fin.append u v)) : ContinuousOn (fun p : (Fin m → ℝ) × (Fin n → ℝ) => packedRealSaddleKernelGradientPair K p.1 p.2) (C ×ˢ D)

Lean source

Helper for Theorem 35.10: Theorem 35.7 turns moving-point convergence of kernels into convergence of the corresponding packed gradient pairs.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_10_moving_gradientPair_tendsto {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {KSeq : ℕ → (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsRealConcaveConvexOn C D K) (hK_diff : ∀ u ∈ C, ∀ v ∈ D, DifferentiableAt ℝ (packedRealSaddleKernel K) (Fin.append u v)) (hKSeq : ∀ i, IsRealConcaveConvexOn C D (KSeq i)) (hKSeq_diff : ∀ i, ∀ u ∈ C, ∀ v ∈ D, DifferentiableAt ℝ (packedRealSaddleKernel (KSeq i)) (Fin.append u v)) (hpointwise : ∀ u ∈ C, ∀ v ∈ D, Filter.Tendsto (fun i : ℕ => KSeq i u v) Filter.atTop (nhds (K u v))) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) (uSeq : ℕ → Fin m → ℝ) (vSeq : ℕ → Fin n → ℝ) (huSeq : ∀ i : ℕ, uSeq i ∈ C) (hvSeq : ∀ i : ℕ, vSeq i ∈ D) (huSeq_tendsto : Filter.Tendsto uSeq Filter.atTop (nhds u)) (hvSeq_tendsto : Filter.Tendsto vSeq Filter.atTop (nhds v)) : Filter.Tendsto (fun i : ℕ => packedRealSaddleKernelGradientPair (KSeq i) (uSeq i) (vSeq i)) Filter.atTop (nhds (packedRealSaddleKernelGradientPair K u v))

Lean source

Helper for Theorem 35.10: the fixed-point convergence statement is the constant-sequence specialization of the moving-point gradient-pair convergence lemma.

open scoped Pointwise open scoped Topology lemma helperForTheorem_35_10_pointwise_gradientPair_tendsto {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {KSeq : ℕ → (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsRealConcaveConvexOn C D K) (hK_diff : ∀ u ∈ C, ∀ v ∈ D, DifferentiableAt ℝ (packedRealSaddleKernel K) (Fin.append u v)) (hKSeq : ∀ i, IsRealConcaveConvexOn C D (KSeq i)) (hKSeq_diff : ∀ i, ∀ u ∈ C, ∀ v ∈ D, DifferentiableAt ℝ (packedRealSaddleKernel (KSeq i)) (Fin.append u v)) (hpointwise : ∀ u ∈ C, ∀ v ∈ D, Filter.Tendsto (fun i : ℕ => KSeq i u v) Filter.atTop (nhds (K u v))) {u : Fin m → ℝ} {v : Fin n → ℝ} (hu : u ∈ C) (hv : v ∈ D) : Filter.Tendsto (fun i : ℕ => packedRealSaddleKernelGradientPair (KSeq i) u v) Filter.atTop (nhds (packedRealSaddleKernelGradientPair K u v))

Lean source

Helper for Text 35.10.1: repackage the dense convergence hypothesis into the existential limit format required by Theorem 35.4.

open scoped Pointwise open scoped Topology lemma helperForText_35_10_1_denseWitness_existsLimits {E F : Type*} [TopologicalSpace E] [TopologicalSpace F] {C : Set E} {D : Set F} {K : E → F → ℝ} {KSeq : ℕ → E → F → ℝ} (hDense : ∃ C' : Set E, ∃ D' : Set F, C' ⊆ C ∧ D' ⊆ D ∧ C ⊆ closure C' ∧ D ⊆ closure D' ∧ ∀ u ∈ C', ∀ v ∈ D', Filter.Tendsto (fun i : ℕ => KSeq i u v) Filter.atTop (nhds (K u v))) : ∃ C' : Set E, ∃ D' : Set F, C' ⊆ C ∧ D' ⊆ D ∧ C ⊆ closure C' ∧ D ⊆ closure D' ∧ ∀ u ∈ C', ∀ v ∈ D', ∃ l : ℝ, Filter.Tendsto (fun i : ℕ => KSeq i u v) Filter.atTop (nhds l)

Lean source

Helper for Text 35.10.1: two finite concave-convex kernels that agree on a dense product subset of an open convex rectangle agree on the whole rectangle.

open scoped Pointwise open scoped Topology lemma helperForText_35_10_1_eqOn_product_of_denseEq {m n : ℕ} {C : Set (EuclideanSpace ℝ (Fin m))} {D : Set (EuclideanSpace ℝ (Fin n))} {K L : EuclideanSpace ℝ (Fin m) → EuclideanSpace ℝ (Fin n) → ℝ} (hC : IsRelativelyOpenConvex C) (hD : IsRelativelyOpenConvex D) (hK : IsRealConcaveConvexOn C D K) (hL : IsRealConcaveConvexOn C D L) {C' : Set (EuclideanSpace ℝ (Fin m))} {D' : Set (EuclideanSpace ℝ (Fin n))} (hC'sub : C' ⊆ C) (hD'sub : D' ⊆ D) (hCclosure : C ⊆ closure C') (hDclosure : D ⊆ closure D') (hEqDense : ∀ u ∈ C', ∀ v ∈ D', L u v = K u v) : ∀ u ∈ C, ∀ v ∈ D, L u v = K u v

Lean source

Helper for Text 35.10.1: the dense convergence hypothesis already forces pointwise convergence of Kᵢ to the prescribed kernel K on all of C × D.

open scoped Pointwise open scoped Topology lemma helperForText_35_10_1_pointwiseTendsto_to_prescribedKernel {m n : ℕ} {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} {K : (Fin m → ℝ) → (Fin n → ℝ) → ℝ} {KSeq : ℕ → (Fin m → ℝ) → (Fin n → ℝ) → ℝ} (hC_open : IsOpen C) (hD_open : IsOpen D) (hC_conv : Convex ℝ C) (hD_conv : Convex ℝ D) (hK : IsRealConcaveConvexOn C D K) (hKSeq : ∀ i, IsRealConcaveConvexOn C D (KSeq i)) (hDense : ∃ C' : Set (Fin m → ℝ), ∃ D' : Set (Fin n → ℝ), C' ⊆ C ∧ D' ⊆ D ∧ C ⊆ closure C' ∧ D ⊆ closure D' ∧ ∀ u ∈ C', ∀ v ∈ D', Filter.Tendsto (fun i : ℕ => KSeq i u v) Filter.atTop (nhds (K u v))) : ∀ u ∈ C, ∀ v ∈ D, Filter.Tendsto (fun i : ℕ => KSeq i u v) Filter.atTop (nhds (K u v))

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

§36 — Minimax Problems — Part 1

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

The value α := sup_{u ∈ C} inf_{v ∈ D} K(u,v) (the maximin value) for an extended-real function K.

variable {C D : Type*} noncomputable def maximinValue (K : C → D → EReal) : EReal := iSup fun u : C => iInf fun v : D => K u v

Lean source

The value β := inf_{v ∈ D} sup_{u ∈ C} K(u,v) (the minimax value) for an extended-real function K.

variable {C D : Type*} noncomputable def minimaxValue (K : C → D → EReal) : EReal := iInf fun v : D => iSup fun u : C => K u v

Lean source

Definition 36.0.1: For nonempty C and D and K : C × D → [-∞, +∞], set α := sup_{u ∈ C} inf_{v ∈ D} K(u,v) and β := inf_{v ∈ D} sup_{u ∈ C} K(u,v); if α = β, then this common value is called the minimax value (or saddle-value) of K.

variable {C D : Type*} def IsMinimaxValue (K : C → D → EReal) (x : EReal) : Prop := Nonempty C ∧ Nonempty D ∧ maximinValue (C := C) (D := D) K = minimaxValue (C := C) (D := D) K ∧ x = maximinValue (C := C) (D := D) K -- Proof sketch: Use the order-theoretic inequality `⨆ u, ⨅ v, K u v ≤ ⨅ v, ⨆ u, K u v` -- in the complete lattice `EReal`, which follows by comparing both sides against the family -- of bounds `⨅ v, K u v ≤ ⨆ u, K u v` for each `u` and `v`.

Lean source

Lemma 36.1: If K : C × D → [-∞, +∞] is defined on a nonempty product set, then sup_u inf_v K(u,v) ≤ inf_v sup_u K(u,v).

variable {C D : Type*} lemma maximinValue_le_minimaxValue (K : C → D → EReal) (hC : Nonempty C) (hD : Nonempty D) : maximinValue (C := C) (D := D) K ≤ minimaxValue (C := C) (D := D) K

Lean source

Definition 36.1.1: Let C and D be nonempty sets and K : C × D → [-∞, +∞]. A pair (u₀, v₀) ∈ C × D is a saddle point of K (maximization in u ∈ C and minimization in v ∈ D) if K u v₀ ≤ K u₀ v₀ ≤ K u₀ v for all u : C and v : D. Equivalently, v₀ minimizes v ↦ K u₀ v on D and u₀ maximizes u ↦ K u v₀ on C, and if the extrema are attained then K u₀ v₀ equals both the min and the max.

variable {C D : Type*} def IsSaddlePoint (K : C → D → EReal) (u₀ : C) (v₀ : D) : Prop := (∀ u : C, K u v₀ ≤ K u₀ v₀) ∧ (∀ v : D, K u₀ v₀ ≤ K u₀ v) -- Proof sketch: Unfold `IsSaddlePoint` and compare `K u₀ v₀` to the families -- `iInf (K u ·)` and `iSup (K · v)` using monotonicity of `iSup`/`iInf`. For the -- “only if” direction, show `maximinValue K = iInf (K u₀ ·)` and -- `minimaxValue K = iSup (K · v₀)` by bounding each extremum above and below by -- `K u₀ v₀`; for the converse, use the attainment equalities plus -- `iInf_le`/`le_iSup` to recover the saddle inequalities.

Lean source

Lemma 36.2: Let K : C × D → [-∞, +∞] be an extended-real function on a nonempty product C × D. A point (u₀, v₀) is a saddle-point of K (maximizing over C and minimizing over D) if and only if the supremum in sup_u inf_v K(u,v) is attained at u₀, the infimum in inf_v sup_u K(u,v) is attained at v₀, and these two extrema are equal. If (u₀, v₀) is a saddle-point, the saddle-value of K is K u₀ v₀.

variable {C D : Type*} lemma isSaddlePoint_iff_maximinValue_eq_minimaxValue_and_attained (K : C → D → EReal) (u₀ : C) (v₀ : D) : IsSaddlePoint (C := C) (D := D) K u₀ v₀ ↔ (maximinValue (C := C) (D := D) K = iInf fun v : D => K u₀ v) ∧ (minimaxValue (C := C) (D := D) K = iSup fun u : C => K u v₀) ∧ maximinValue (C := C) (D := D) K = minimaxValue (C := C) (D := D) K ∧ K u₀ v₀ = maximinValue (C := C) (D := D) K

Lean source

A ±∞-extension of a real-valued function K defined on a nonempty product C × D of subsets of ambient spaces: it agrees with K on C × D, equals +∞ on C × (Dᶜ), equals -∞ on (Cᶜ) × D, and takes a chosen fallback value outside C ∪ D.

variable {C D : Type*} noncomputable def pmInfExtension {U V : Type*} (C : Set U) (D : Set V) (K : {u // u ∈ C} → {v // v ∈ D} → ℝ) (z : EReal := 0) : U → V → EReal := fun u v => if hu : u ∈ C then if hv : v ∈ D then (K ⟨u, hu⟩ ⟨v, hv⟩ : EReal) else ⊤ else if v ∈ D then ⊥ else z

Lean source

A ±∞-extension of a real-valued function K on C × D where the value on (Cᶜ) × (Dᶜ) is allowed to be an arbitrary function Z of (u,v).

variable {C D : Type*} noncomputable def pmInfExtensionWith {U V : Type*} (C : Set U) (D : Set V) (K : {u // u ∈ C} → {v // v ∈ D} → ℝ) (Z : U → V → EReal) : U → V → EReal

Lean source

Proposition 36.2.1: (±∞-extension from C × D to ℝ^m × ℝ^n.) Let C ⊆ ℝ^m and D ⊆ ℝ^n be nonempty and let K : C × D → ℝ. Define an extension K̄ : ℝ^m × ℝ^n → [-∞, +∞] by setting K̄ = K on C × D, K̄ = +∞ on C × (Dᶜ), K̄ = -∞ on (Cᶜ) × D, and choosing an arbitrary value (depending on (u,v)) on (Cᶜ) × (Dᶜ). Then: (1) inf_v K̄(u,v) = inf_{v ∈ D} K(u,v) < +∞, with value -∞ when u ∉ C; (2) sup_u K̄(u,v) = sup_{u ∈ C} K(u,v) > -∞, with value +∞ when v ∉ D; (3) consequently the maximin/minimax values computed on ℝ^m × ℝ^n coincide with those computed on C × D, and K̄ has a saddle-value (resp. saddle-point) w.r.t. ℝ^m × ℝ^n iff K has one w.r.t. C × D (with coinciding values/points).

variable {C D : Type*} theorem pmInfExtension_inf_sup_and_saddle_equivalences {m n : ℕ} (C : Set (Fin m → ℝ)) (D : Set (Fin n → ℝ)) (hC : C.Nonempty) (hD : D.Nonempty) (K : {u // u ∈ C} → {v // v ∈ D} → ℝ) (Z : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (∀ u : Fin m → ℝ, (iInf fun v : Fin n → ℝ => pmInfExtensionWith (C := C) (D := D) K (Z := Z) u v) = (if hu : u ∈ C then iInf fun v : {v // v ∈ D} => (K ⟨u, hu⟩ v : EReal) else ⊥) ∧ (iInf fun v : Fin n → ℝ => pmInfExtensionWith (C := C) (D := D) K (Z := Z) u v) < (⊤ : EReal) ∧ (u ∉ C → (iInf fun v : Fin n → ℝ => pmInfExtensionWith (C := C) (D := D) K (Z := Z) u v) = ⊥)) ∧ (∀ v : Fin n → ℝ, (iSup fun u : Fin m → ℝ => pmInfExtensionWith (C := C) (D := D) K (Z := Z) u v) = (if hv : v ∈ D then iSup fun u : {u // u ∈ C} => (K u ⟨v, hv⟩ : EReal) else ⊤) ∧ (⊥ : EReal) < (iSup fun u : Fin m → ℝ => pmInfExtensionWith (C := C) (D := D) K (Z := Z) u v) ∧ (v ∉ D → (iSup fun u : Fin m → ℝ => pmInfExtensionWith (C := C) (D := D) K (Z := Z) u v) = ⊤)) ∧ (iSup fun u : Fin m → ℝ => iInf fun v : Fin n → ℝ => pmInfExtensionWith (C := C) (D := D) K (Z := Z) u v) = (iSup fun u : {u // u ∈ C} => iInf fun v : {v // v ∈ D} => (K u v : EReal)) ∧ (iInf fun v : Fin n → ℝ => iSup fun u : Fin m → ℝ => pmInfExtensionWith (C := C) (D := D) K (Z := Z) u v) = (iInf fun v : {v // v ∈ D} => iSup fun u : {u // u ∈ C} => (K u v : EReal)) ∧ (∀ x : EReal, IsMinimaxValue (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) (pmInfExtensionWith (C := C) (D := D) K (Z := Z)) x ↔ IsMinimaxValue (C := {u // u ∈ C}) (D := {v // v ∈ D}) (fun u v => (K u v : EReal)) x) ∧ (∀ (u₀ : Fin m → ℝ) (v₀ : Fin n → ℝ), IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) (pmInfExtensionWith (C := C) (D := D) K (Z := Z)) u₀ v₀ ↔ ∃ (hu₀ : u₀ ∈ C) (hv₀ : v₀ ∈ D), IsSaddlePoint (C := {u // u ∈ C}) (D := {v // v ∈ D}) (fun u v => (K u v : EReal)) ⟨u₀, hu₀⟩ ⟨v₀, hv₀⟩)

Lean source

The first domain dom₁ K of an extended-real saddle function K(u,v), defined as the set of points u where inf_v K(u,v) > -∞.

variable {C D : Type*} noncomputable def saddleDom1 {U V : Type*} (K : U → V → EReal) : Set U := {u | (⊥ : EReal) < iInf fun v : V => K u v}

Lean source

The second domain dom₂ K of an extended-real saddle function K(u,v), defined as the set of points v where sup_u K(u,v) < +∞.

variable {C D : Type*} noncomputable def saddleDom2 {U V : Type*} (K : U → V → EReal) : Set V := {v | (iSup fun u : U => K u v) < (⊤ : EReal)}

Lean source

A bundled hypothesis package expressing the *domain boundary behavior* expected of an extended-real saddle function K used in minimax problems: the projected domains dom₁ K and dom₂ K are nonempty, K(u,v) = -∞ on (dom₁ K)ᶜ × dom₂ K, and K(u,v) = +∞ on dom₁ K × (dom₂ K)ᶜ.

variable {C D : Type*} def HasSaddleDomBoundaryBehavior {U V : Type*} (K : U → V → EReal) : Prop := (saddleDom1 (U := U) (V := V) K).Nonempty ∧ (saddleDom2 (U := U) (V := V) K).Nonempty ∧ (∀ u : U, ∀ v : V, u ∉ saddleDom1 (U := U) (V := V) K → v ∈ saddleDom2 (U := U) (V := V) K → K u v = (⊥ : EReal)) ∧ (∀ u : U, ∀ v : V, u ∈ saddleDom1 (U := U) (V := V) K → v ∉ saddleDom2 (U := U) (V := V) K → K u v = (⊤ : EReal))

Lean source

The core concavity/convexity/closedness/properness content of the textbook hypothesis “closed proper concave-convex”, separated from the ±∞ boundary-value behavior.

variable {C D : Type*} def IsClosedProperConcaveConvexCore {U V : Type*} [AddCommGroup U] [Module ℝ U] [TopologicalSpace U] [AddCommGroup V] [Module ℝ V] [TopologicalSpace V] (K : U → V → EReal) : Prop

Lean source

The book-level hypothesis that an extended-real function K(u,v) is a *closed proper concave-convex saddle function* (Rockafellar, Theorem 36.3 context). This predicate is intentionally separated from HasSaddleDomBoundaryBehavior: the latter captures the ±∞ boundary values on complements of dom₁ K and dom₂ K, while this predicate is meant to also include the concavity/convexity and closedness/properness content of the textbook notion.

variable {C D : Type*} def IsClosedProperConcaveConvex {U V : Type*} [AddCommGroup U] [Module ℝ U] [TopologicalSpace U] [AddCommGroup V] [Module ℝ V] [TopologicalSpace V] (K : U → V → EReal) : Prop := HasSaddleDomBoundaryBehavior (U := U) (V := V) K ∧ IsClosedProperConcaveConvexCore (U := U) (V := V) K

Lean source

Helper for Lemma 36.2: at a saddle point, the v-infimum along the saddle row is the center value K u₀ v₀.

variable {C D : Type*} lemma helperForLemma_36_2_rowInf_eq_centerValue (K : C → D → EReal) (u₀ : C) (v₀ : D) (hs : IsSaddlePoint (C := C) (D := D) K u₀ v₀) : iInf (fun v : D => K u₀ v) = K u₀ v₀

Lean source

Helper for Lemma 36.2: at a saddle point, the u-supremum along the saddle column is the center value K u₀ v₀.

variable {C D : Type*} lemma helperForLemma_36_2_colSup_eq_centerValue (K : C → D → EReal) (u₀ : C) (v₀ : D) (hs : IsSaddlePoint (C := C) (D := D) K u₀ v₀) : iSup (fun u : C => K u v₀) = K u₀ v₀

Lean source

Helper for Lemma 36.2: at a saddle point, the maximin value is attained at the saddle row u₀.

variable {C D : Type*} lemma helperForLemma_36_2_maximin_eq_attainedRowInf (K : C → D → EReal) (u₀ : C) (v₀ : D) (hs : IsSaddlePoint (C := C) (D := D) K u₀ v₀) : maximinValue (C := C) (D := D) K = iInf (fun v : D => K u₀ v)

Lean source

Helper for Lemma 36.2: at a saddle point, the minimax value is attained at the saddle column v₀.

variable {C D : Type*} lemma helperForLemma_36_2_minimax_eq_attainedColSup (K : C → D → EReal) (u₀ : C) (v₀ : D) (hs : IsSaddlePoint (C := C) (D := D) K u₀ v₀) : minimaxValue (C := C) (D := D) K = iSup (fun u : C => K u v₀)

Lean source

Helper for Proposition 36.2.1: when u ∈ C, the ambient row infimum of the ±∞-extension agrees with the subtype row infimum over D.

variable {C D : Type*} lemma helperForProposition_36_2_1_rowInf_eq_subtype_of_mem {m n : ℕ} (C : Set (Fin m → ℝ)) (D : Set (Fin n → ℝ)) (K : {u // u ∈ C} → {v // v ∈ D} → ℝ) (Z : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (u : Fin m → ℝ) (hu : u ∈ C) : (iInf fun v : Fin n → ℝ => pmInfExtensionWith (C := C) (D := D) K (Z := Z) u v) = iInf fun v : {v // v ∈ D} => (K ⟨u, hu⟩ v : EReal)

Lean source

Helper for Proposition 36.2.1: when u ∉ C, the ambient row infimum of the ±∞-extension is forced to be -∞.

variable {C D : Type*} lemma helperForProposition_36_2_1_rowInf_eq_bot_of_not_mem {m n : ℕ} (C : Set (Fin m → ℝ)) (D : Set (Fin n → ℝ)) (hD : D.Nonempty) (K : {u // u ∈ C} → {v // v ∈ D} → ℝ) (Z : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (u : Fin m → ℝ) (hu : u ∉ C) : (iInf fun v : Fin n → ℝ => pmInfExtensionWith (C := C) (D := D) K (Z := Z) u v) = (⊥ : EReal)

Lean source

Helper for Proposition 36.2.1: package the row-infimum formula, its finiteness from above, and the u ∉ C boundary behavior.

variable {C D : Type*} lemma helperForProposition_36_2_1_rowInf_package {m n : ℕ} (C : Set (Fin m → ℝ)) (D : Set (Fin n → ℝ)) (hD : D.Nonempty) (K : {u // u ∈ C} → {v // v ∈ D} → ℝ) (Z : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (u : Fin m → ℝ) : (iInf fun v : Fin n → ℝ => pmInfExtensionWith (C := C) (D := D) K (Z := Z) u v) = (if hu : u ∈ C then iInf fun v : {v // v ∈ D} => (K ⟨u, hu⟩ v : EReal) else ⊥) ∧ (iInf fun v : Fin n → ℝ => pmInfExtensionWith (C := C) (D := D) K (Z := Z) u v) < (⊤ : EReal) ∧ (u ∉ C → (iInf fun v : Fin n → ℝ => pmInfExtensionWith (C := C) (D := D) K (Z := Z) u v) = ⊥)

Lean source

Helper for Proposition 36.2.1: when v ∈ D, the ambient column supremum of the ±∞-extension agrees with the subtype column supremum over C.

variable {C D : Type*} lemma helperForProposition_36_2_1_colSup_eq_subtype_of_mem {m n : ℕ} (C : Set (Fin m → ℝ)) (D : Set (Fin n → ℝ)) (K : {u // u ∈ C} → {v // v ∈ D} → ℝ) (Z : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (v : Fin n → ℝ) (hv : v ∈ D) : (iSup fun u : Fin m → ℝ => pmInfExtensionWith (C := C) (D := D) K (Z := Z) u v) = iSup fun u : {u // u ∈ C} => (K u ⟨v, hv⟩ : EReal)

Lean source

Helper for Proposition 36.2.1: when v ∉ D, the ambient column supremum of the ±∞-extension is forced to be +∞.

variable {C D : Type*} lemma helperForProposition_36_2_1_colSup_eq_top_of_not_mem {m n : ℕ} (C : Set (Fin m → ℝ)) (D : Set (Fin n → ℝ)) (hC : C.Nonempty) (K : {u // u ∈ C} → {v // v ∈ D} → ℝ) (Z : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (v : Fin n → ℝ) (hv : v ∉ D) : (iSup fun u : Fin m → ℝ => pmInfExtensionWith (C := C) (D := D) K (Z := Z) u v) = (⊤ : EReal)

Lean source

Helper for Proposition 36.2.1: package the column-supremum formula, its finiteness from below, and the v ∉ D boundary behavior.

variable {C D : Type*} lemma helperForProposition_36_2_1_colSup_package {m n : ℕ} (C : Set (Fin m → ℝ)) (D : Set (Fin n → ℝ)) (hC : C.Nonempty) (K : {u // u ∈ C} → {v // v ∈ D} → ℝ) (Z : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (v : Fin n → ℝ) : (iSup fun u : Fin m → ℝ => pmInfExtensionWith (C := C) (D := D) K (Z := Z) u v) = (if hv : v ∈ D then iSup fun u : {u // u ∈ C} => (K u ⟨v, hv⟩ : EReal) else ⊤) ∧ (⊥ : EReal) < (iSup fun u : Fin m → ℝ => pmInfExtensionWith (C := C) (D := D) K (Z := Z) u v) ∧ (v ∉ D → (iSup fun u : Fin m → ℝ => pmInfExtensionWith (C := C) (D := D) K (Z := Z) u v) = ⊤)

Lean source

Helper for Proposition 36.2.1: the outer u-supremum of the ambient row infima coincides with the subtype maximin expression over C × D.

variable {C D : Type*} lemma helperForProposition_36_2_1_maximin_eq_subtype {m n : ℕ} (C : Set (Fin m → ℝ)) (D : Set (Fin n → ℝ)) (hD : D.Nonempty) (K : {u // u ∈ C} → {v // v ∈ D} → ℝ) (Z : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (iSup fun u : Fin m → ℝ => iInf fun v : Fin n → ℝ => pmInfExtensionWith (C := C) (D := D) K (Z := Z) u v) = (iSup fun u : {u // u ∈ C} => iInf fun v : {v // v ∈ D} => (K u v : EReal))

Lean source

Helper for Proposition 36.2.1: the outer v-infimum of the ambient column suprema coincides with the subtype minimax expression over C × D.

variable {C D : Type*} lemma helperForProposition_36_2_1_minimax_eq_subtype {m n : ℕ} (C : Set (Fin m → ℝ)) (D : Set (Fin n → ℝ)) (hC : C.Nonempty) (K : {u // u ∈ C} → {v // v ∈ D} → ℝ) (Z : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (iInf fun v : Fin n → ℝ => iSup fun u : Fin m → ℝ => pmInfExtensionWith (C := C) (D := D) K (Z := Z) u v) = (iInf fun v : {v // v ∈ D} => iSup fun u : {u // u ∈ C} => (K u v : EReal))

Lean source

Helper for Proposition 36.2.1: an ambient saddle point for the ±∞-extension is exactly a subtype saddle point on C × D, and the ambient point is forced into C × D.

variable {C D : Type*} lemma helperForProposition_36_2_1_saddlePoint_transport {m n : ℕ} (C : Set (Fin m → ℝ)) (D : Set (Fin n → ℝ)) (hC : C.Nonempty) (hD : D.Nonempty) (K : {u // u ∈ C} → {v // v ∈ D} → ℝ) (Z : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (u₀ : Fin m → ℝ) (v₀ : Fin n → ℝ) : IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) (pmInfExtensionWith (C := C) (D := D) K (Z := Z)) u₀ v₀ ↔ ∃ (hu₀ : u₀ ∈ C) (hv₀ : v₀ ∈ D), IsSaddlePoint (C := {u // u ∈ C}) (D := {v // v ∈ D}) (fun u v => (K u v : EReal)) ⟨u₀, hu₀⟩ ⟨v₀, hv₀⟩

Lean source

Helper for Theorem 36.3: on a row indexed by u ∈ dom₁ K, restricting the infimum to dom₂ K does not change its value.

variable {C D : Type*} lemma helperForTheorem_36_3_rowInf_eq_subtype_of_memSaddleDom1 {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hbdry : HasSaddleDomBoundaryBehavior K) (u : Fin m → ℝ) (hu : u ∈ saddleDom1 K) : iInf (fun v : Fin n → ℝ => K u v) = iInf (fun v : {v // v ∈ saddleDom2 K} => K u v)

Lean source

Helper for Theorem 36.3: outside dom₁ K, the row infimum is exactly -∞.

variable {C D : Type*} lemma helperForTheorem_36_3_rowInf_eq_bot_of_not_memSaddleDom1 {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (u : Fin m → ℝ) (hu : u ∉ saddleDom1 K) : iInf (fun v : Fin n → ℝ => K u v) = (⊥ : EReal)

Lean source

Helper for Theorem 36.3: on a column indexed by v ∈ dom₂ K, restricting the supremum to dom₁ K does not change its value.

variable {C D : Type*} lemma helperForTheorem_36_3_colSup_eq_subtype_of_memSaddleDom2 {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hbdry : HasSaddleDomBoundaryBehavior K) (v : Fin n → ℝ) (hv : v ∈ saddleDom2 K) : iSup (fun u : Fin m → ℝ => K u v) = iSup (fun u : {u // u ∈ saddleDom1 K} => K u v)

Lean source

Helper for Theorem 36.3: outside dom₂ K, the column supremum is exactly +∞.

variable {C D : Type*} lemma helperForTheorem_36_3_colSup_eq_top_of_not_memSaddleDom2 {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (v : Fin n → ℝ) (hv : v ∉ saddleDom2 K) : iSup (fun u : Fin m → ℝ => K u v) = (⊤ : EReal)

Lean source

Helper for Theorem 36.3: restricting the maximin expression from ℝ^m × ℝ^n to dom₁ K × dom₂ K does not change its value.

variable {C D : Type*} lemma helperForTheorem_36_3_maximin_eq_restrictedToSaddleDom {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hbdry : HasSaddleDomBoundaryBehavior K) : maximinValue (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K = maximinValue (C := {u // u ∈ saddleDom1 K}) (D := {v // v ∈ saddleDom2 K}) (fun u v => K u v)

Lean source

Helper for Theorem 36.3: restricting the minimax expression from ℝ^m × ℝ^n to dom₁ K × dom₂ K does not change its value.

variable {C D : Type*} lemma helperForTheorem_36_3_minimax_eq_restrictedToSaddleDom {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hbdry : HasSaddleDomBoundaryBehavior K) : minimaxValue (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K = minimaxValue (C := {u // u ∈ saddleDom1 K}) (D := {v // v ∈ saddleDom2 K}) (fun u v => K u v)

§36 — Minimax Problems — Part 2

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Theorem 36.3: Let K be a closed proper concave-convex function on ℝ^m × ℝ^n, and let C = dom₁ K and D = dom₂ K. Then restricting the maximin and minimax expressions to C × D does not change their values, and the saddle-values and saddle-points of K with respect to ℝ^m × ℝ^n agree with those with respect to C × D.

theorem maximin_minimax_and_saddle_restrict_to_saddleDom {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hK : IsClosedProperConcaveConvex K) : maximinValue (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K = maximinValue (C := {u // u ∈ saddleDom1 K}) (D := {v // v ∈ saddleDom2 K}) (fun u v => K u v) ∧ minimaxValue (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K = minimaxValue (C := {u // u ∈ saddleDom1 K}) (D := {v // v ∈ saddleDom2 K}) (fun u v => K u v) ∧ (∀ x : EReal, IsMinimaxValue (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K x ↔ IsMinimaxValue (C := {u // u ∈ saddleDom1 K}) (D := {v // v ∈ saddleDom2 K}) (fun u v => K u v) x) ∧ (∀ (u₀ : Fin m → ℝ) (v₀ : Fin n → ℝ), IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K u₀ v₀ ↔ ∃ (hu₀ : u₀ ∈ saddleDom1 K) (hv₀ : v₀ ∈ saddleDom2 K), IsSaddlePoint (C := {u // u ∈ saddleDom1 K}) (D := {v // v ∈ saddleDom2 K}) (fun u v => K u v) ⟨u₀, hu₀⟩ ⟨v₀, hv₀⟩)

Lean source

The domain dom K of an extended-real saddle function K(u,v), defined as the product dom₁ K × dom₂ K.

noncomputable def saddleDom {U V : Type*} (K : U → V → EReal) : Set (U × V) := saddleDom1 (U := U) (V := V) K ×ˢ saddleDom2 (U := U) (V := V) K

Lean source

Corollary 36.3.1: Let K be a closed proper saddle-function on ℝ^m × ℝ^n. If K has a saddle-point, then this saddle-point lies in dom K, and the saddle-value of K is finite.

theorem saddlePoint_mem_dom_and_saddleValue_finite {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hK : IsClosedProperConcaveConvex K) (hs : ∃ u₀ v₀, IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K u₀ v₀) : (∀ {u₀ : Fin m → ℝ} {v₀ : Fin n → ℝ}, IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K u₀ v₀ → (u₀, v₀) ∈ saddleDom (U := (Fin m → ℝ)) (V := (Fin n → ℝ)) K ∧ K u₀ v₀ ≠ (⊥ : EReal) ∧ K u₀ v₀ ≠ (⊤ : EReal)) ∧ (∃ x : EReal, IsMinimaxValue (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K x ∧ x ≠ (⊥ : EReal) ∧ x ≠ (⊤ : EReal))

Lean source

Two extended-real saddle functions K and L are *equivalent* if they have the same projected domains dom₁ and dom₂, satisfy the ±∞ boundary-value behavior on these domains, and agree on dom₁ × dom₂.

def EquivalentSaddleFunctions {U V : Type*} (K L : U → V → EReal) : Prop := HasSaddleDomBoundaryBehavior (U := U) (V := V) K ∧ HasSaddleDomBoundaryBehavior (U := U) (V := V) L ∧ saddleDom1 (U := U) (V := V) K = saddleDom1 (U := U) (V := V) L ∧ saddleDom2 (U := U) (V := V) K = saddleDom2 (U := U) (V := V) L ∧ ∀ u : U, ∀ v : V, u ∈ saddleDom1 (U := U) (V := V) K → v ∈ saddleDom2 (U := U) (V := V) K → K u v = L u v

Lean source

Theorem 36.4: Equivalent saddle-functions on ℝ^m × ℝ^n have the same saddle-value and saddle-points (if any).

theorem equivalentSaddleFunctions_saddleValue_and_saddlePoints {m n : ℕ} {K L : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hKL : EquivalentSaddleFunctions (U := (Fin m → ℝ)) (V := (Fin n → ℝ)) K L) : (∀ x : EReal, IsMinimaxValue (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K x ↔ IsMinimaxValue (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) L x) ∧ (∀ (u₀ : Fin m → ℝ) (v₀ : Fin n → ℝ), IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K u₀ v₀ ↔ IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) L u₀ v₀)

Lean source

Helper for Theorem 36.3: ambient saddle points are exactly subtype saddle points on dom₁ K × dom₂ K, and the ambient point is forced into the projected domains.

lemma helperForTheorem_36_3_saddlePoint_transport {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hbdry : HasSaddleDomBoundaryBehavior K) (u₀ : Fin m → ℝ) (v₀ : Fin n → ℝ) : IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K u₀ v₀ ↔ ∃ (hu₀ : u₀ ∈ saddleDom1 K) (hv₀ : v₀ ∈ saddleDom2 K), IsSaddlePoint (C := {u // u ∈ saddleDom1 K}) (D := {v // v ∈ saddleDom2 K}) (fun u v => K u v) ⟨u₀, hu₀⟩ ⟨v₀, hv₀⟩

Lean source

Helper for Corollary 36.3.1: every saddle point lies in dom K, and its center value is neither -∞ nor +∞.

lemma helperForCorollary_36_3_1_saddlePoint_mem_dom_and_finiteCenter {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hK : IsClosedProperConcaveConvex K) {u₀ : Fin m → ℝ} {v₀ : Fin n → ℝ} (hsuv : IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K u₀ v₀) : (u₀, v₀) ∈ saddleDom (U := (Fin m → ℝ)) (V := (Fin n → ℝ)) K ∧ K u₀ v₀ ≠ (⊥ : EReal) ∧ K u₀ v₀ ≠ (⊤ : EReal)

Lean source

Helper for Corollary 36.3.1: a saddle point furnishes a finite minimax value.

lemma helperForCorollary_36_3_1_exists_finite_minimaxValue_from_saddlePoint {m n : ℕ} (K : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hK : IsClosedProperConcaveConvex K) (hs : ∃ u₀ v₀, IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K u₀ v₀) : ∃ x : EReal, IsMinimaxValue (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K x ∧ x ≠ (⊥ : EReal) ∧ x ≠ (⊤ : EReal)

Lean source

Helper for Theorem 36.4: on the common saddle domain, equivalent saddle-functions have the same restricted kernel.

lemma helperForTheorem_36_4_restrictedKernel_eq_on_commonSaddleDom {m n : ℕ} {K L : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hAgree : ∀ u : Fin m → ℝ, ∀ v : Fin n → ℝ, u ∈ saddleDom1 K → v ∈ saddleDom2 K → K u v = L u v) : (fun u : {u // u ∈ saddleDom1 K} => fun v : {v // v ∈ saddleDom2 K} => K u v) = (fun u : {u // u ∈ saddleDom1 K} => fun v : {v // v ∈ saddleDom2 K} => L u v)

Lean source

Helper for Theorem 36.4: for a saddle-function with the standard boundary behavior, the ambient minimax-value predicate is equivalent to the minimax-value predicate on dom₁ × dom₂.

lemma helperForTheorem_36_4_isMinimaxValue_iff_restrictedToSaddleDom {m n : ℕ} (M : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hM : HasSaddleDomBoundaryBehavior M) (x : EReal) : IsMinimaxValue (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) M x ↔ IsMinimaxValue (C := {u // u ∈ saddleDom1 M}) (D := {v // v ∈ saddleDom2 M}) (fun u v => M u v) x

Lean source

Helper for Theorem 36.4: on a fixed common subtype domain, equal restricted kernels have the same saddle-point predicate.

lemma helperForTheorem_36_4_restrictedSaddlePoint_iff_of_restrictedKernel_eq {m n : ℕ} {K L : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hEq : (fun u : {u // u ∈ saddleDom1 K} => fun v : {v // v ∈ saddleDom2 K} => K u v) = (fun u : {u // u ∈ saddleDom1 K} => fun v : {v // v ∈ saddleDom2 K} => L u v)) : ∀ (u₀ : {u // u ∈ saddleDom1 K}) (v₀ : {v // v ∈ saddleDom2 K}), IsSaddlePoint (C := {u // u ∈ saddleDom1 K}) (D := {v // v ∈ saddleDom2 K}) (fun u v => K u v) u₀ v₀ ↔ IsSaddlePoint (C := {u // u ∈ saddleDom1 K}) (D := {v // v ∈ saddleDom2 K}) (fun u v => L u v) u₀ v₀

Lean source

Helper for Theorem 36.4: changing only the subtype domain by equal sets does not change the restricted minimax-value predicate.

lemma helperForTheorem_36_4_restrictedIsMinimaxValue_iff_of_domain_eq {m n : ℕ} (M : (Fin m → ℝ) → (Fin n → ℝ) → EReal) {S S' : Set (Fin m → ℝ)} {T T' : Set (Fin n → ℝ)} (hS : S = S') (hT : T = T') (x : EReal) : IsMinimaxValue (C := {u // u ∈ S}) (D := {v // v ∈ T}) (fun u v => M u v) x ↔ IsMinimaxValue (C := {u // u ∈ S'}) (D := {v // v ∈ T'}) (fun u v => M u v) x

Lean source

Helper for Theorem 36.4: changing only the subtype domain by equal sets does not change the restricted existential saddle-point transport statement.

lemma helperForTheorem_36_4_restrictedSaddlePointExists_iff_of_domain_eq {m n : ℕ} (M : (Fin m → ℝ) → (Fin n → ℝ) → EReal) {S S' : Set (Fin m → ℝ)} {T T' : Set (Fin n → ℝ)} (hS : S = S') (hT : T = T') (u₀ : Fin m → ℝ) (v₀ : Fin n → ℝ) : (∃ (hu₀ : u₀ ∈ S) (hv₀ : v₀ ∈ T), IsSaddlePoint (C := {u // u ∈ S}) (D := {v // v ∈ T}) (fun u v => M u v) ⟨u₀, hu₀⟩ ⟨v₀, hv₀⟩) ↔ (∃ (hu₀ : u₀ ∈ S') (hv₀ : v₀ ∈ T'), IsSaddlePoint (C := {u // u ∈ S'}) (D := {v // v ∈ T'}) (fun u v => M u v) ⟨u₀, hu₀⟩ ⟨v₀, hv₀⟩)

§36 — Minimax Problems — Part 3

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

The hypograph {((u,x),t) | (t : EReal) ≤ F u x} of a bifunction F, viewed as a subset of (ℝ^m × ℝ^n) × ℝ.

def bifunctionHypograph {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Set (((Fin m → ℝ) × (Fin n → ℝ)) × ℝ) := {p | (p.2 : EReal) ≤ F p.1.1 p.1.2}

Lean source

The epigraph {((u,x),t) | F u x ≤ (t : EReal)} of a bifunction F, viewed as a subset of (ℝ^m × ℝ^n) × ℝ.

def bifunctionEpigraph {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Set (((Fin m → ℝ) × (Fin n → ℝ)) × ℝ) := {p | F p.1.1 p.1.2 ≤ (p.2 : EReal)}

Lean source

A concave bifunction, modeled by convexity of its hypograph in (ℝ^m × ℝ^n) × ℝ.

def IsHypographConvexConcaveBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := Convex ℝ (bifunctionHypograph (m := m) (n := n) F)

Lean source

A convex bifunction, modeled by convexity of its epigraph in (ℝ^m × ℝ^n) × ℝ.

def IsEpigraphConvexBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := Convex ℝ (bifunctionEpigraph (m := m) (n := n) F)

Lean source

Closedness for a convex bifunction, modeled as topological closedness of its epigraph. This is a section-local epigraph notion, distinct from the canonical closed convex bifunction predicate used elsewhere in Chapter 7.

def IsEpigraphClosedConvexBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := IsClosed (bifunctionEpigraph (m := m) (n := n) F)

Lean source

Closedness for a concave bifunction, modeled as topological closedness of its hypograph.

def IsHypographClosedConcaveBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := IsClosed (bifunctionHypograph (m := m) (n := n) F)

Lean source

Properness for a convex bifunction: it never takes the value -∞ and is not identically +∞.

def IsEpigraphProperConvexBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := (∃ u x, F u x < (⊤ : EReal)) ∧ ∀ u x, F u x ≠ (⊥ : EReal)

Lean source

Properness for a concave bifunction: it never takes the value +∞ and is not identically -∞.

def IsHypographProperConcaveBifunction {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := (∃ u x, (⊥ : EReal) < F u x) ∧ ∀ u x, F u x ≠ (⊤ : EReal)

Lean source

Proposition 36.4.2: Let F be a convex (resp. concave) bifunction from ℝ^m to ℝ^n, and define its inverse bifunction F_* by (F_* x)(u) = -(F u)(x). Then F_* is concave (resp. convex). Moreover, within the convex/concave classes, the mapping F ↦ F_* preserves closedness and properness, and it is involutive: (F_*)_* = F.

theorem bifunctionInverse_convex_concave_closed_proper_involutive {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (IsEpigraphConvexBifunction (m := m) (n := n) F → IsHypographConvexConcaveBifunction (m := n) (n := m) (bifunctionInverse F)) ∧ (IsHypographConvexConcaveBifunction (m := m) (n := n) F → IsEpigraphConvexBifunction (m := n) (n := m) (bifunctionInverse F)) ∧ (IsEpigraphClosedConvexBifunction (m := m) (n := n) F → IsHypographClosedConcaveBifunction (m := n) (n := m) (bifunctionInverse F)) ∧ (IsHypographClosedConcaveBifunction (m := m) (n := n) F → IsEpigraphClosedConvexBifunction (m := n) (n := m) (bifunctionInverse F)) ∧ (IsEpigraphProperConvexBifunction (m := m) (n := n) F → IsHypographProperConcaveBifunction (m := n) (n := m) (bifunctionInverse F)) ∧ (IsHypographProperConcaveBifunction (m := m) (n := n) F → IsEpigraphProperConvexBifunction (m := n) (n := m) (bifunctionInverse F)) ∧ bifunctionInverse (bifunctionInverse F) = F

Lean source

The Euclidean adjoint bifunction F^* of F, modeled as a Fenchel-type conjugation with respect to the Euclidean dot products on ℝ^m and ℝ^n, and with the argument order swapped.

noncomputable def bifunctionEuclideanAdjoint {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin n → ℝ) → (Fin m → ℝ) → EReal := fun x u => iSup fun u' : Fin m → ℝ => iSup fun x' : Fin n → ℝ => ((finDot (n := n) x x' + finDot (n := m) u u' : ℝ) : EReal) + (-F u' x')

Lean source

Proposition 36.4.3: for a convex (respectively concave) bifunction, inversion commutes with the branch-appropriate Chapter 6 adjoint. The canonical proof that the inverse has the opposite orientation is supplied explicitly in each branch, so the convex branch uses the concave adjoint of F_*, while the concave branch uses the convex adjoint of F_*.

theorem bifunctionInverse_adjoint_commutes {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (∀ (hFconv : ConvexBifunction F) (hInvConc : ConcaveBifunction (bifunctionInverse F)), adjointOfConcaveBifunction ⟨bifunctionInverse F, hInvConc⟩ = bifunctionInverse (adjointOfConvexBifunction ⟨F, hFconv⟩)) ∧ (∀ (hFconc : ConcaveBifunction F) (hInvConv : ConvexBifunction (bifunctionInverse F)), adjointOfConvexBifunction ⟨bifunctionInverse F, hInvConv⟩ = bifunctionInverse (adjointOfConcaveBifunction ⟨F, hFconc⟩))

Lean source

Helper for Proposition 36.4.2: swap the two vector coordinates and negate the scalar coordinate when transporting epi/hypographs across bifunction inversion.

def helperForProposition_36_4_2_swapNegate {m n : ℕ} : (((Fin n → ℝ) × (Fin m → ℝ)) × ℝ) → (((Fin m → ℝ) × (Fin n → ℝ)) × ℝ) := fun p => ((p.1.2, p.1.1), -p.2)

Lean source

Helper for Proposition 36.4.2: the swap-negate transport map is linear, so convexity pulls back along it.

lemma helperForProposition_36_4_2_swapNegate_isLinear {m n : ℕ} : IsLinearMap ℝ (helperForProposition_36_4_2_swapNegate (m := m) (n := n))

Lean source

Helper for Proposition 36.4.2: the swap-negate transport map is continuous, so closedness pulls back along it.

lemma helperForProposition_36_4_2_swapNegate_continuous {m n : ℕ} : Continuous (helperForProposition_36_4_2_swapNegate (m := m) (n := n))

Lean source

Helper for Proposition 36.4.2: the hypograph of the inverse bifunction is the preimage of the original epigraph under the swap-negate transport.

lemma helperForProposition_36_4_2_inverseHypograph_preimageEpigraph {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : bifunctionHypograph (m := n) (n := m) (bifunctionInverse F) = helperForProposition_36_4_2_swapNegate (m := m) (n := n) ⁻¹' bifunctionEpigraph (m := m) (n := n) F

Lean source

Helper for Proposition 36.4.2: the epigraph of the inverse bifunction is the preimage of the original hypograph under the same swap-negate transport.

lemma helperForProposition_36_4_2_inverseEpigraph_preimageHypograph {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : bifunctionEpigraph (m := n) (n := m) (bifunctionInverse F) = helperForProposition_36_4_2_swapNegate (m := m) (n := n) ⁻¹' bifunctionHypograph (m := m) (n := n) F

Lean source

Helper for Proposition 36.4.2: proper convexity transfers to proper concavity after swapping the arguments and negating the EReal value.

lemma helperForProposition_36_4_2_inverseProperConvex_to_Concave {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsEpigraphProperConvexBifunction (m := m) (n := n) F) : IsHypographProperConcaveBifunction (m := n) (n := m) (bifunctionInverse F)

Lean source

Helper for Proposition 36.4.2: proper concavity transfers to proper convexity by the same swap-and-negate argument.

lemma helperForProposition_36_4_2_inverseProperConcave_to_Convex {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsHypographProperConcaveBifunction (m := m) (n := n) F) : IsEpigraphProperConvexBifunction (m := n) (n := m) (bifunctionInverse F)

Lean source

Helper for Proposition 36.4.3: projection onto the scalar coordinate is linear, so half-space convexity can be pulled back to bifunction epi/hypographs.

lemma helperForProposition_36_4_3_scalarProjection_isLinear {m n : ℕ} : IsLinearMap ℝ (fun p : (((Fin m → ℝ) × (Fin n → ℝ)) × ℝ) => p.2)

Lean source

Helper for Proposition 36.4.3: the constant-zero bifunction has a convex epigraph.

lemma helperForProposition_36_4_3_constantZero_isEpigraphConvex {m n : ℕ} : IsEpigraphConvexBifunction (m := m) (n := n) (fun _ _ => (0 : EReal))

Lean source

Helper for Proposition 36.4.3: the constant-zero bifunction has a convex hypograph.

lemma helperForProposition_36_4_3_constantZero_isHypographConcave {m n : ℕ} : IsHypographConvexConcaveBifunction (m := m) (n := n) (fun _ _ => (0 : EReal))

Lean source

Helper for Proposition 36.4.3: the unique vector in Fin 1 → ℝ whose only coordinate is 1.

def helperForProposition_36_4_3_oneVector : Fin 1 → ℝ := fun _ => 1

Lean source

Helper for Proposition 36.4.3: the zero vector in Fin 1 → ℝ.

def helperForProposition_36_4_3_zeroVector : Fin 1 → ℝ := fun _ => 0

Lean source

Helper for Proposition 36.4.3: evaluating the left composite at (1, 0) on the dimension-one constant-zero bifunction gives a value at least 1.

lemma helperForProposition_36_4_3_constantZero_leftComposite_ge_real (r : ℝ) : ((r : ℝ) : EReal) ≤ bifunctionEuclideanAdjoint (m := 1) (n := 1) (bifunctionInverse (fun _ _ => (0 : EReal))) helperForProposition_36_4_3_oneVector helperForProposition_36_4_3_zeroVector

Lean source

Helper for Proposition 36.4.3: the inner adjoint term on the right composite dominates every real number at the swapped witness (0, 1).

lemma helperForProposition_36_4_3_constantZero_innerAdjoint_ge_real (r : ℝ) : ((r : ℝ) : EReal) ≤ bifunctionEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal)) helperForProposition_36_4_3_zeroVector helperForProposition_36_4_3_oneVector

Lean source

Helper for Proposition 36.4.3: after applying the inverse, the right composite at (1, 0) is bounded above by -r for every real number r.

lemma helperForProposition_36_4_3_constantZero_rightComposite_le_neg_real (r : ℝ) : bifunctionInverse (bifunctionEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal))) helperForProposition_36_4_3_oneVector helperForProposition_36_4_3_zeroVector ≤ ((-r : ℝ) : EReal)

Lean source

Helper for Proposition 36.4.3: evaluating the left composite at (1, 0) on the dimension-one constant-zero bifunction gives a value at least 1.

lemma helperForProposition_36_4_3_constantZero_leftComposite_ge_one : (1 : EReal) ≤ bifunctionEuclideanAdjoint (m := 1) (n := 1) (bifunctionInverse (fun _ _ => (0 : EReal))) helperForProposition_36_4_3_oneVector helperForProposition_36_4_3_zeroVector

Lean source

Helper for Proposition 36.4.3: the inner adjoint term on the right composite is also at least 1 at the swapped point (0, 1).

lemma helperForProposition_36_4_3_constantZero_innerAdjoint_ge_one : (1 : EReal) ≤ bifunctionEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal)) helperForProposition_36_4_3_zeroVector helperForProposition_36_4_3_oneVector

Lean source

Helper for Proposition 36.4.3: after applying the inverse, the right composite at (1, 0) is bounded above by -1.

lemma helperForProposition_36_4_3_constantZero_rightComposite_le_negOne : bifunctionInverse (bifunctionEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal))) helperForProposition_36_4_3_oneVector helperForProposition_36_4_3_zeroVector ≤ (-1 : EReal)

Lean source

Helper for Proposition 36.4.3: the embedded real numbers still satisfy 1 ≰ -1 inside EReal.

lemma helperForProposition_36_4_3_one_not_le_negOne : ¬ ((1 : EReal) ≤ (-1 : EReal))

Lean source

Helper for Proposition 36.4.3: at the witness (1, 0), the left composite cannot lie below the right composite for the dimension-one constant-zero bifunction.

lemma helperForProposition_36_4_3_constantZero_leftComposite_not_le_rightComposite : ¬ (bifunctionEuclideanAdjoint (m := 1) (n := 1) (bifunctionInverse (fun _ _ => (0 : EReal))) helperForProposition_36_4_3_oneVector helperForProposition_36_4_3_zeroVector ≤ bifunctionInverse (bifunctionEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal))) helperForProposition_36_4_3_oneVector helperForProposition_36_4_3_zeroVector)

Lean source

Helper for Proposition 36.4.3: the two composite bifunctions already differ pointwise at the explicit witness (1, 0) for the dimension-one constant-zero bifunction.

lemma helperForProposition_36_4_3_constantZero_pointwiseValues_ne : bifunctionEuclideanAdjoint (m := 1) (n := 1) (bifunctionInverse (fun _ _ => (0 : EReal))) helperForProposition_36_4_3_oneVector helperForProposition_36_4_3_zeroVector ≠ bifunctionInverse (bifunctionEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal))) helperForProposition_36_4_3_oneVector helperForProposition_36_4_3_zeroVector

Lean source

Helper for Proposition 36.4.3: with the current adjoint definition, the constant-zero bifunction in dimension 1 does not satisfy the claimed commutation law.

lemma helperForProposition_36_4_3_constantZero_commutation_fails : bifunctionEuclideanAdjoint (m := 1) (n := 1) (bifunctionInverse (fun _ _ => (0 : EReal))) ≠ bifunctionInverse (bifunctionEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal)))

Lean source

Helper for Proposition 36.4.3: the specialized statement already fails on the constant-zero bifunction, so the current target cannot be proved without changing the local setup.

lemma helperForProposition_36_4_3_constantZero_convexBranch_fails : ¬ (IsEpigraphConvexBifunction (m := 1) (n := 1) (fun _ _ => (0 : EReal)) → bifunctionEuclideanAdjoint (m := 1) (n := 1) (bifunctionInverse (fun _ _ => (0 : EReal))) = bifunctionInverse (bifunctionEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal))))

Lean source

Helper for Proposition 36.4.3: the concave branch of the displayed implication also fails on the constant-zero bifunction in dimension 1.

lemma helperForProposition_36_4_3_constantZero_concaveBranch_fails : ¬ (IsHypographConvexConcaveBifunction (m := 1) (n := 1) (fun _ _ => (0 : EReal)) → bifunctionEuclideanAdjoint (m := 1) (n := 1) (bifunctionInverse (fun _ _ => (0 : EReal))) = bifunctionInverse (bifunctionEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal))))

Lean source

Helper for Proposition 36.4.3: the specialized statement already fails on the constant-zero bifunction, so the current target cannot be proved without changing the local setup.

lemma helperForProposition_36_4_3_constantZero_target_fails : ¬ ((IsEpigraphConvexBifunction (m := 1) (n := 1) (fun _ _ => (0 : EReal)) → bifunctionEuclideanAdjoint (m := 1) (n := 1) (bifunctionInverse (fun _ _ => (0 : EReal))) = bifunctionInverse (bifunctionEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal)))) ∧ (IsHypographConvexConcaveBifunction (m := 1) (n := 1) (fun _ _ => (0 : EReal)) → bifunctionEuclideanAdjoint (m := 1) (n := 1) (bifunctionInverse (fun _ _ => (0 : EReal))) = bifunctionInverse (bifunctionEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal)))))

Lean source

Helper for Proposition 36.4.3: any universal proof of the displayed commutation law would specialize to the already disproved constant-zero instance in dimension 1.

lemma helperForProposition_36_4_3_universalTarget_fails : ¬ (∀ {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal), (IsEpigraphConvexBifunction (m := m) (n := n) F → bifunctionEuclideanAdjoint (m := n) (n := m) (bifunctionInverse F) = bifunctionInverse (bifunctionEuclideanAdjoint (m := m) (n := n) F)) ∧ (IsHypographConvexConcaveBifunction (m := m) (n := n) F → bifunctionEuclideanAdjoint (m := n) (n := m) (bifunctionInverse F) = bifunctionInverse (bifunctionEuclideanAdjoint (m := m) (n := n) F)))

Lean source

Helper for Proposition 36.4.3: the exact universal theorem signature is empty under the current section-local adjoint convention.

lemma helperForProposition_36_4_3_exactTheoremSignature_isEmpty : IsEmpty (∀ {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal), (IsEpigraphConvexBifunction (m := m) (n := n) F → bifunctionEuclideanAdjoint (m := n) (n := m) (bifunctionInverse F) = bifunctionInverse (bifunctionEuclideanAdjoint (m := m) (n := n) F)) ∧ (IsHypographConvexConcaveBifunction (m := m) (n := n) F → bifunctionEuclideanAdjoint (m := n) (n := m) (bifunctionInverse F) = bifunctionInverse (bifunctionEuclideanAdjoint (m := m) (n := n) F)))

§36 — Minimax Problems — Part 4

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

The bifunction commonly denoted F_*^*, obtained by composing inverse (_*) and adjoint (^*): F_*^* := (F_*)^*.

noncomputable def bifunctionInverseEuclideanAdjoint {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin m → ℝ) → (Fin n → ℝ) → EReal := bifunctionEuclideanAdjoint (m := n) (n := m) (bifunctionInverse F)

Lean source

The biconjugate F^{**} of a bifunction F, defined as applying the adjoint operation twice.

noncomputable def bifunctionEuclideanBiconjugate {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin m → ℝ) → (Fin n → ℝ) → EReal := bifunctionEuclideanAdjoint (m := n) (n := m) (bifunctionEuclideanAdjoint (m := m) (n := n) F)

Lean source

The textbook Euclidean adjoint on Fin k → ℝ, matching the Chapter 8 infimum convention F^*(x^*,u^*) = inf_{u,x} (⟨x^*,x⟩ - ⟨u^*,u⟩ + F(u,x)).

noncomputable def bifunctionEuclideanAdjointTextbook {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin n → ℝ) → (Fin m → ℝ) → EReal := fun xStar uStar => iInf fun u : Fin m → ℝ => iInf fun x : Fin n → ℝ => (((finDot (n := n) xStar x - finDot (n := m) uStar u : ℝ)) : EReal) + F u x

Lean source

The textbook object F_*^* = (F_*)^* formed from the Euclidean adjoint above.

noncomputable def bifunctionInverseEuclideanAdjointTextbook {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin m → ℝ) → (Fin n → ℝ) → EReal := bifunctionEuclideanAdjointTextbook (m := n) (n := m) (bifunctionInverse F)

Lean source

The textbook Euclidean biconjugate F^{**}.

noncomputable def bifunctionEuclideanBiconjugateTextbook {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin m → ℝ) → (Fin n → ℝ) → EReal := bifunctionEuclideanAdjointTextbook (m := n) (n := m) (bifunctionEuclideanAdjointTextbook (m := m) (n := n) F)

Lean source

Proposition 36.4.4: let F be a canonical convex bifunction and explicitly supply the canonical concave orientation of its inverse F_*. Then F_*^* = adjointOfConcaveBifunction F_* is convex. Moreover its inverse has a canonical concave orientation for which the next adjoint is the canonical convex bifunction closure of F; this is the packaged form of (F_*^*)_*^* = F^{**} = cl F.

theorem bifunctionInverseAdjoint_convex_and_iterate_eq_biconjugate_and_closure {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hFconv : ConvexBifunction F) (hInvConc : ConcaveBifunction (bifunctionInverse F)) : let H := adjointOfConcaveBifunction ⟨bifunctionInverse F, hInvConc⟩ ConvexBifunction H ∧ ∃ hHInvConc : ConcaveBifunction (bifunctionInverse H), adjointOfConcaveBifunction ⟨bifunctionInverse H, hHInvConc⟩ = convexBifunctionClosure F

Lean source

Helper for Proposition 36.4.4: the first _*^* iterate of the dimension-one constant-zero bifunction is already ⊤ at the witness (1, 0).

lemma helperForProposition_36_4_4_constantZero_firstIterate_topAtWitness : bifunctionInverseEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal)) helperForProposition_36_4_3_oneVector helperForProposition_36_4_3_zeroVector = (⊤ : EReal)

Lean source

Helper for Proposition 36.4.4: the second _*^* iterate of the dimension-one constant-zero bifunction is also ⊤ at the same witness (1, 0).

lemma helperForProposition_36_4_4_constantZero_secondIterate_topAtWitness : bifunctionInverseEuclideanAdjoint (m := 1) (n := 1) (bifunctionInverseEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal))) helperForProposition_36_4_3_oneVector helperForProposition_36_4_3_zeroVector = (⊤ : EReal)

Lean source

Helper for Proposition 36.4.4: the Euclidean adjoint of the dimension-one constant-zero bifunction vanishes at (0, 0).

lemma helperForProposition_36_4_4_constantZero_adjoint_atZero_eq_zero : bifunctionEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal)) helperForProposition_36_4_3_zeroVector helperForProposition_36_4_3_zeroVector = (0 : EReal)

Lean source

Helper for Proposition 36.4.4: every outer pairing term at the witness (1, 0) is dominated by the Euclidean adjoint of the dimension-one constant-zero bifunction.

lemma helperForProposition_36_4_4_constantZero_outerPairing_le_adjoint (u x : Fin 1 → ℝ) : (((finDot (n := 1) helperForProposition_36_4_3_oneVector x + finDot (n := 1) helperForProposition_36_4_3_zeroVector u : ℝ)) : EReal) ≤ bifunctionEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal)) u x

Lean source

Helper for Proposition 36.4.4: the lower-semicontinuous hull of the dimension-one constant-zero bifunction is still the constant-zero function.

lemma helperForProposition_36_4_4_constantZero_lowerSemicontinuousHull_eq_zero : erealLowerSemicontinuousHull (fun _ : (Fin 1 → ℝ) × (Fin 1 → ℝ) => (0 : EReal)) = fun _ => (0 : EReal)

Lean source

Helper for Proposition 36.4.4: at the witness (1, 0), both the biconjugate and the closure of the dimension-one constant-zero bifunction equal 0.

lemma helperForProposition_36_4_4_constantZero_biconjugate_and_closure_eq_zero : bifunctionEuclideanBiconjugate (m := 1) (n := 1) (fun _ _ => (0 : EReal)) helperForProposition_36_4_3_oneVector helperForProposition_36_4_3_zeroVector = (0 : EReal) ∧ bifunctionClosure (fun _ _ : Fin 1 → ℝ => (0 : EReal)) helperForProposition_36_4_3_oneVector helperForProposition_36_4_3_zeroVector = (0 : EReal)

Lean source

Helper for Proposition 36.4.4: evaluating the specialized equality branch at the witness (1, 0) forces the impossible identity ⊤ = 0.

lemma helperForProposition_36_4_4_constantZero_witness_forces_top_eq_zero (hIterateEq : bifunctionInverseEuclideanAdjoint (m := 1) (n := 1) (bifunctionInverseEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal))) = bifunctionEuclideanBiconjugate (m := 1) (n := 1) (fun _ _ => (0 : EReal))) (hClosureEq : bifunctionEuclideanBiconjugate (m := 1) (n := 1) (fun _ _ => (0 : EReal)) = bifunctionClosure (fun _ _ : Fin 1 → ℝ => (0 : EReal))) : (⊤ : EReal) = (0 : EReal)

Lean source

Helper for Proposition 36.4.4: the equality branch of the dimension-one constant-zero specialization already fails at the explicit witness (1, 0).

lemma helperForProposition_36_4_4_constantZero_equalityBranch_fails : ¬ (bifunctionInverseEuclideanAdjoint (m := 1) (n := 1) (bifunctionInverseEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal))) = bifunctionEuclideanBiconjugate (m := 1) (n := 1) (fun _ _ => (0 : EReal)) ∧ bifunctionEuclideanBiconjugate (m := 1) (n := 1) (fun _ _ => (0 : EReal)) = bifunctionClosure (fun _ _ : Fin 1 → ℝ => (0 : EReal)))

Lean source

Helper for Proposition 36.4.4: even the conclusion of the dimension-one constant-zero specialization is empty, because its equality branch already forces ⊤ = 0.

lemma helperForProposition_36_4_4_constantZero_conclusion_isEmpty : IsEmpty (IsEpigraphConvexBifunction (m := 1) (n := 1) (bifunctionInverseEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal))) ∧ bifunctionInverseEuclideanAdjoint (m := 1) (n := 1) (bifunctionInverseEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal))) = bifunctionEuclideanBiconjugate (m := 1) (n := 1) (fun _ _ => (0 : EReal)) ∧ bifunctionEuclideanBiconjugate (m := 1) (n := 1) (fun _ _ => (0 : EReal)) = bifunctionClosure (fun _ _ : Fin 1 → ℝ => (0 : EReal)))

Lean source

Helper for Proposition 36.4.4: the specialized target statement already fails on the dimension-one constant-zero bifunction, so the local theorem cannot be proved as written.

lemma helperForProposition_36_4_4_constantZero_target_fails : ¬ (IsEpigraphConvexBifunction (m := 1) (n := 1) (fun _ _ => (0 : EReal)) → IsEpigraphConvexBifunction (m := 1) (n := 1) (bifunctionInverseEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal))) ∧ bifunctionInverseEuclideanAdjoint (m := 1) (n := 1) (bifunctionInverseEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal))) = bifunctionEuclideanBiconjugate (m := 1) (n := 1) (fun _ _ => (0 : EReal)) ∧ bifunctionEuclideanBiconjugate (m := 1) (n := 1) (fun _ _ => (0 : EReal)) = bifunctionClosure (fun _ _ : Fin 1 → ℝ => (0 : EReal)))

Lean source

Helper for Proposition 36.4.4: the specialized theorem type for the dimension-one constant-zero bifunction is empty.

lemma helperForProposition_36_4_4_constantZero_targetType_isEmpty : IsEmpty (IsEpigraphConvexBifunction (m := 1) (n := 1) (fun _ _ => (0 : EReal)) → IsEpigraphConvexBifunction (m := 1) (n := 1) (bifunctionInverseEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal))) ∧ bifunctionInverseEuclideanAdjoint (m := 1) (n := 1) (bifunctionInverseEuclideanAdjoint (m := 1) (n := 1) (fun _ _ => (0 : EReal))) = bifunctionEuclideanBiconjugate (m := 1) (n := 1) (fun _ _ => (0 : EReal)) ∧ bifunctionEuclideanBiconjugate (m := 1) (n := 1) (fun _ _ => (0 : EReal)) = bifunctionClosure (fun _ _ : Fin 1 → ℝ => (0 : EReal)))

Lean source

Helper for Proposition 36.4.4: any universal proof of the displayed theorem would specialize to the already disproved dimension-one constant-zero instance.

lemma helperForProposition_36_4_4_universalTarget_fails : ¬ (∀ {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal), IsEpigraphConvexBifunction (m := m) (n := n) F → IsEpigraphConvexBifunction (m := m) (n := n) (bifunctionInverseEuclideanAdjoint (m := m) (n := n) F) ∧ bifunctionInverseEuclideanAdjoint (m := m) (n := n) (bifunctionInverseEuclideanAdjoint (m := m) (n := n) F) = bifunctionEuclideanBiconjugate (m := m) (n := n) F ∧ bifunctionEuclideanBiconjugate (m := m) (n := n) F = bifunctionClosure F)

Lean source

Helper for Proposition 36.4.4: the exact universal theorem type is empty under the current section-local _*^* convention.

lemma helperForProposition_36_4_4_targetType_isEmpty : IsEmpty (∀ {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal), IsEpigraphConvexBifunction (m := m) (n := n) F → IsEpigraphConvexBifunction (m := m) (n := n) (bifunctionInverseEuclideanAdjoint (m := m) (n := n) F) ∧ bifunctionInverseEuclideanAdjoint (m := m) (n := n) (bifunctionInverseEuclideanAdjoint (m := m) (n := n) F) = bifunctionEuclideanBiconjugate (m := m) (n := n) F ∧ bifunctionEuclideanBiconjugate (m := m) (n := n) F = bifunctionClosure F)

Lean source

Helper for Proposition 36.4.4: the exact theorem signature with the convexity premise written as an explicit argument is already empty under the current local _*^* convention.

lemma helperForProposition_36_4_4_exactTheoremSignature_isEmpty : IsEmpty (∀ {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsEpigraphConvexBifunction (m := m) (n := n) F), IsEpigraphConvexBifunction (m := m) (n := n) (bifunctionInverseEuclideanAdjoint (m := m) (n := n) F) ∧ bifunctionInverseEuclideanAdjoint (m := m) (n := n) (bifunctionInverseEuclideanAdjoint (m := m) (n := n) F) = bifunctionEuclideanBiconjugate (m := m) (n := n) F ∧ bifunctionEuclideanBiconjugate (m := m) (n := n) F = bifunctionClosure F)

Lean source

Helper for Proposition 36.4.4: the textbook _*^* iterate of the dimension-one constant-zero bifunction vanishes at the origin (0, 0).

lemma helperForProposition_36_4_4_textbookZeroIterate_atOrigin_eq_zero : bifunctionInverseEuclideanAdjointTextbook (m := 1) (n := 1) (fun _ _ => (0 : EReal)) helperForProposition_36_4_3_zeroVector helperForProposition_36_4_3_zeroVector = (0 : EReal)

Lean source

Helper for Proposition 36.4.4: the textbook _*^* iterate of the dimension-one constant-zero bifunction already equals ⊥ at the witness (1, 0).

lemma helperForProposition_36_4_4_textbookZeroIterate_atOneZero_eq_bot : bifunctionInverseEuclideanAdjointTextbook (m := 1) (n := 1) (fun _ _ => (0 : EReal)) helperForProposition_36_4_3_oneVector helperForProposition_36_4_3_zeroVector = (⊥ : EReal)

Lean source

Helper for Proposition 36.4.4: the textbook _*^* iterate of the dimension-one constant-zero bifunction also equals ⊥ at the witness (-1, 0).

lemma helperForProposition_36_4_4_textbookZeroIterate_atNegOneZero_eq_bot : bifunctionInverseEuclideanAdjointTextbook (m := 1) (n := 1) (fun _ _ => (0 : EReal)) (fun _ => (-1 : ℝ)) helperForProposition_36_4_3_zeroVector = (⊥ : EReal)

Lean source

Helper for Proposition 36.4.4: the textbook _*^* iterate of the dimension-one constant-zero bifunction does not have a convex epigraph.

lemma helperForProposition_36_4_4_textbookZeroIterate_epigraph_not_convex : ¬ IsEpigraphConvexBifunction (m := 1) (n := 1) (bifunctionInverseEuclideanAdjointTextbook (m := 1) (n := 1) (fun _ _ => (0 : EReal)))

Lean source

Helper for Proposition 36.4.4: any implication-shaped universal proof of the textbook _*^* theorem would already contradict the dimension-one constant-zero specialization.

lemma helperForProposition_36_4_4_textbookUniversalTarget_fails (hUniversal : ∀ {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal), IsEpigraphConvexBifunction (m := m) (n := n) F → IsEpigraphConvexBifunction (m := m) (n := n) (bifunctionInverseEuclideanAdjointTextbook (m := m) (n := n) F) ∧ bifunctionInverseEuclideanAdjointTextbook (m := m) (n := n) (bifunctionInverseEuclideanAdjointTextbook (m := m) (n := n) F) = bifunctionEuclideanBiconjugateTextbook (m := m) (n := n) F ∧ bifunctionEuclideanBiconjugateTextbook (m := m) (n := n) F = bifunctionClosure F) : False

Lean source

Helper for Proposition 36.4.4: the implication-shaped textbook theorem type is empty under the current section-local _*^* convention.

lemma helperForProposition_36_4_4_textbookTargetType_isEmpty : IsEmpty (∀ {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal), IsEpigraphConvexBifunction (m := m) (n := n) F → IsEpigraphConvexBifunction (m := m) (n := n) (bifunctionInverseEuclideanAdjointTextbook (m := m) (n := n) F) ∧ bifunctionInverseEuclideanAdjointTextbook (m := m) (n := n) (bifunctionInverseEuclideanAdjointTextbook (m := m) (n := n) F) = bifunctionEuclideanBiconjugateTextbook (m := m) (n := n) F ∧ bifunctionEuclideanBiconjugateTextbook (m := m) (n := n) F = bifunctionClosure F)

Lean source

Helper for Proposition 36.4.4: the exact theorem signature with the convexity premise written as an explicit argument is already empty under the current textbook _*^* convention.

lemma helperForProposition_36_4_4_textbookExactTheoremSignature_isEmpty : IsEmpty (∀ {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsEpigraphConvexBifunction (m := m) (n := n) F), IsEpigraphConvexBifunction (m := m) (n := n) (bifunctionInverseEuclideanAdjointTextbook (m := m) (n := n) F) ∧ bifunctionInverseEuclideanAdjointTextbook (m := m) (n := n) (bifunctionInverseEuclideanAdjointTextbook (m := m) (n := n) F) = bifunctionEuclideanBiconjugateTextbook (m := m) (n := n) F ∧ bifunctionEuclideanBiconjugateTextbook (m := m) (n := n) F = bifunctionClosure F)

§36 — Minimax Problems — Part 5

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Definition 36.4.5: Let F be a convex bifunction from ℝ^m to ℝ^n. The *Lagrangian* of the associated convex program is the function L : ℝ^m × ℝ^n → [-∞, +∞] defined by L(u*, x) := inf_{u ∈ ℝ^m} (⟨u*, u⟩ + (F u) x).

noncomputable def bifunctionLagrangian {m n : ℕ} (F : {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal // IsEpigraphConvexBifunction (m := m) (n := n) F}) : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun uStar x => iInf fun u : Fin m → ℝ => ((finDot (n := m) uStar u : ℝ) : EReal) + F.1 u x

Lean source

The infimal pairing ⟪uStar, g⟫ := inf_u (⟨uStar, u⟩ - g u) used in the book for an extended-real function g : ℝ^m → [-∞, +∞].

noncomputable def infPairing {m : ℕ} (uStar : Fin m → ℝ) (g : (Fin m → ℝ) → EReal) : EReal := iInf fun u : Fin m → ℝ => ((finDot (n := m) uStar u : ℝ) : EReal) + (-g u)

Lean source

Concavity in the first argument uStar (pointwise in x) for an extended-real bifunction L(uStar, x).

def IsConcaveInFirst {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := ∀ x : Fin n → ℝ, Convex ℝ {p : (Fin m → ℝ) × ℝ | (p.2 : EReal) ≤ L p.1 x}

Lean source

Convexity in the second argument x (pointwise in uStar) for an extended-real bifunction L(uStar, x).

def IsConvexInSecond {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := ∀ uStar : Fin m → ℝ, Convex ℝ {p : (Fin n → ℝ) × ℝ | L uStar p.1 ≤ (p.2 : EReal)}

Lean source

Proposition 36.4.6: Let F be a convex bifunction from ℝ^m to ℝ^n with inverse F_*, and let L be defined by Definition 36.4.5. Then L(uStar, x) = inf_u (⟨uStar, u⟩ - (F_* x)(u)) = ⟪uStar, F_* x⟫. In particular, L is concave in uStar and convex in x.

theorem bifunctionLagrangian_eq_infPairing_inverse_and_concave_convex {m n : ℕ} (F : {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal // IsEpigraphConvexBifunction (m := m) (n := n) F}) : (∀ uStar : Fin m → ℝ, ∀ x : Fin n → ℝ, bifunctionLagrangian (m := m) (n := n) F uStar x = iInf fun u : Fin m → ℝ => ((finDot (n := m) uStar u : ℝ) : EReal) + (-bifunctionInverse F.1 x u)) ∧ (∀ uStar : Fin m → ℝ, ∀ x : Fin n → ℝ, bifunctionLagrangian (m := m) (n := n) F uStar x = infPairing (m := m) uStar (bifunctionInverse F.1 x)) ∧ IsConcaveInFirst (m := m) (n := n) (bifunctionLagrangian (m := m) (n := n) F) ∧ IsConvexInSecond (m := m) (n := n) (bifunctionLagrangian (m := m) (n := n) F)

Lean source

Upper closedness in the first argument uStar (pointwise in x) for an extended-real function L(uStar, x): for each x, the hypograph {(uStar,t) | t ≤ L(uStar,x)} is closed.

def IsUpperClosedInFirst {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := ∀ x : Fin n → ℝ, IsClosed {p : (Fin m → ℝ) × ℝ | (p.2 : EReal) ≤ L p.1 x}

Lean source

Lower closedness in the second argument x (pointwise in uStar) for an extended-real function L(uStar, x): for each uStar, the epigraph {(x,t) | L(uStar,x) ≤ t} is closed.

def IsLowerClosedInSecond {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := ∀ uStar : Fin m → ℝ, IsClosed {p : (Fin n → ℝ) × ℝ | L uStar p.1 ≤ (p.2 : EReal)}

Lean source

A bundled predicate expressing that L(uStar, x) is *upper closed concave-convex* on ℝ^m × ℝ^n: it is upper closed and concave in uStar, and lower closed and convex in x.

def IsUpperClosedConcaveConvex {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := IsUpperClosedInFirst (m := m) (n := n) L ∧ IsLowerClosedInSecond (m := m) (n := n) L ∧ IsConcaveInFirst (m := m) (n := n) L ∧ IsConvexInSecond (m := m) (n := n) L

Lean source

Helper for Proposition 36.4.6: finDot is linear in its first argument.

lemma helperForProposition_36_4_6_finDot_smul_add_left {m : ℕ} (uStar₁ uStar₂ u : Fin m → ℝ) (a b : ℝ) : finDot (n := m) (a • uStar₁ + b • uStar₂) u = a * finDot (n := m) uStar₁ u + b * finDot (n := m) uStar₂ u

Lean source

Helper for Proposition 36.4.6: finDot is linear in its second argument.

lemma helperForProposition_36_4_6_finDot_smul_add_right {m : ℕ} (uStar u₁ u₂ : Fin m → ℝ) (a b : ℝ) : finDot (n := m) uStar (a • u₁ + b • u₂) = a * finDot (n := m) uStar u₁ + b * finDot (n := m) uStar u₂

Lean source

Helper for Proposition 36.4.6: the local finDot notation agrees with the standard dotProduct.

lemma helperForProposition_36_4_6_finDot_eq_dotProduct {m : ℕ} (uStar u : Fin m → ℝ) : finDot (n := m) uStar u = dotProduct uStar u

Lean source

Helper for Proposition 36.4.6: the local infimal pairing is exactly the chapter's concave conjugate.

lemma helperForProposition_36_4_6_infPairing_eq_concaveConjugate {m : ℕ} (uStar : Fin m → ℝ) (g : (Fin m → ℝ) → EReal) : infPairing (m := m) uStar g = concaveConjugate g uStar

Lean source

Helper for Proposition 36.4.6: negate the scalar coordinate while keeping the vector coordinate fixed.

def helperForProposition_36_4_6_negateScalarSecond {m : ℕ} : ((Fin m → ℝ) × ℝ) → ((Fin m → ℝ) × ℝ) := fun p => (p.1, -p.2)

Lean source

Helper for Proposition 36.4.6: negating the scalar coordinate is linear on (ℝ^m) × ℝ.

lemma helperForProposition_36_4_6_negateScalarSecond_isLinear {m : ℕ} : IsLinearMap ℝ (helperForProposition_36_4_6_negateScalarSecond (m := m))

Lean source

Helper for Proposition 36.4.6: the hypograph of infPairing is the preimage of the epigraph of its negation under scalar negation.

lemma helperForProposition_36_4_6_infPairing_hypograph_preimage_negEpigraph {m : ℕ} (g : (Fin m → ℝ) → EReal) : {p : (Fin m → ℝ) × ℝ | (p.2 : EReal) ≤ infPairing (m := m) p.1 g} = helperForProposition_36_4_6_negateScalarSecond (m := m) ⁻¹' epigraph (Set.univ : Set (Fin m → ℝ)) (fun uStar => -(infPairing (m := m) uStar g))

Lean source

Helper for Proposition 36.4.6: for every function g, the map uStar ↦ infPairing uStar g is concave.

lemma helperForProposition_36_4_6_infPairing_isConcave {m : ℕ} (g : (Fin m → ℝ) → EReal) : Convex ℝ {p : (Fin m → ℝ) × ℝ | (p.2 : EReal) ≤ infPairing (m := m) p.1 g}

Lean source

Helper for Proposition 36.4.6: the graph function of a section-local convex bifunction is a convex function on the product space.

lemma helperForProposition_36_4_6_graphFunction_isConvexFunction {m n : ℕ} (F : {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal // IsEpigraphConvexBifunction (m := m) (n := n) F}) : ConvexFunction (graphFunctionOfBifunction F.1)

Lean source

Helper for Proposition 36.4.6: the tilted graph function used for the fixed-uStar sections is convex.

lemma helperForProposition_36_4_6_tiltedGraph_isConvexFunction {m n : ℕ} (F : {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal // IsEpigraphConvexBifunction (m := m) (n := n) F}) (uStar : Fin m → ℝ) : ConvexFunction (fun z : Fin (m + n) → ℝ => ((finDot (n := m) uStar (projXLinearMap (n := m) (m := n) z) : ℝ) : EReal) + graphFunctionOfBifunction F.1 z)

Lean source

Helper for Proposition 36.4.6: for fixed uStar, the x-section of the Lagrangian is the fiber infimum of the tilted graph under projection to the x-coordinates.

lemma helperForProposition_36_4_6_secondSection_eq_projectionInf {m n : ℕ} (F : {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal // IsEpigraphConvexBifunction (m := m) (n := n) F}) (uStar : Fin m → ℝ) : (fun x : Fin n → ℝ => bifunctionLagrangian (m := m) (n := n) F uStar x) = imageUnderLinearMap (projLamLinearMap (n := m) (m := n)) (fun z : Fin (m + n) → ℝ => ((finDot (n := m) uStar (projXLinearMap (n := m) (m := n) z) : ℝ) : EReal) + graphFunctionOfBifunction F.1 z)

Lean source

Helper for Theorem 36.5: a Lagrangian already has the concave-convex orientation from Proposition 36.4.6.

lemma helperForTheorem_36_5_lagrangian_has_concaveConvex_orientation {m n : ℕ} (F : {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal // IsEpigraphConvexBifunction (m := m) (n := n) F}) : IsConcaveInFirst (m := m) (n := n) (bifunctionLagrangian (m := m) (n := n) F) ∧ IsConvexInSecond (m := m) (n := n) (bifunctionLagrangian (m := m) (n := n) F)

Lean source

Helper for Theorem 36.5: every one-variable infimal pairing section is concave-closed.

lemma helperForTheorem_36_5_infPairing_isFunctionConcaveClosed {m : ℕ} (g : (Fin m → ℝ) → EReal) : IsFunctionConcaveClosed (fun uStar : Fin m → ℝ => infPairing (m := m) uStar g)

Lean source

Helper for Theorem 36.5: upper/lower closedness of an upper closed concave-convex bifunction already gives the Section 33 one-variable closedness data on every slice.

lemma helperForTheorem_36_5_upperClosedSlices_to_functionClosedData {m n : ℕ} {L : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hL : IsUpperClosedConcaveConvex (m := m) (n := n) L) : (∀ x : Fin n → ℝ, IsFunctionConcaveClosed (fun uStar : Fin m → ℝ => L uStar x)) ∧ (∀ uStar : Fin m → ℝ, IsFunctionConvexClosed (fun x : Fin n → ℝ => L uStar x))

Lean source

Helper for Theorem 36.5: upper closedness together with first-variable concavity upgrades each fixed-x slice to a closed concave function in the Chapter 6 sense.

lemma helperForTheorem_36_5_sliceClosedConcave_of_upperClosedConcave {m n : ℕ} {L : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hUpper : IsUpperClosedInFirst (m := m) (n := n) L) (hConc : IsConcaveInFirst (m := m) (n := n) L) : ∀ x : Fin n → ℝ, ClosedConcaveFunction (fun uStar : Fin m → ℝ => L uStar x)

Lean source

Helper for Theorem 36.5: the conjugate reconstruction F(u, x) = sup_{u*} (L(u*, x) - ⟨u*, u⟩) is convex in the epigraph sense whenever the second sections of L are convex.

lemma helperForTheorem_36_5_reconstructedBifunction_isEpigraphConvex {m n : ℕ} {L : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hConv : IsConvexInSecond (m := m) (n := n) L) : IsEpigraphConvexBifunction (m := m) (n := n) (fun u : Fin m → ℝ => fun x : Fin n → ℝ => iSup fun uStar : Fin m → ℝ => L uStar x + (-((finDot (n := m) uStar u : ℝ) : EReal)))

Lean source

Helper for Theorem 36.5: the epigraph of the reconstructed conjugate bifunction is the intersection of the closed fixed-uStar slice epigraphs after shifting the height coordinate by ⟨uStar, u⟩.

lemma helperForTheorem_36_5_reconstructedBifunction_epigraph_eq_iInter_preimages {m n : ℕ} {L : (Fin m → ℝ) → (Fin n → ℝ) → EReal} : bifunctionEpigraph (m := m) (n := n) (fun u : Fin m → ℝ => fun x : Fin n → ℝ => iSup fun uStar : Fin m → ℝ => L uStar x + (-((finDot (n := m) uStar u : ℝ) : EReal))) = ⋂ uStar : Fin m → ℝ, (fun p : ((Fin m → ℝ) × (Fin n → ℝ)) × ℝ => (p.1.2, p.2 + finDot (n := m) uStar p.1.1)) ⁻¹' {q : (Fin n → ℝ) × ℝ | L uStar q.1 ≤ (q.2 : EReal)}

Lean source

Helper for Theorem 36.5: the same conjugate reconstruction has closed epigraph whenever the second sections of L are lower closed.

lemma helperForTheorem_36_5_reconstructedBifunction_isEpigraphClosed {m n : ℕ} {L : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hLower : IsLowerClosedInSecond (m := m) (n := n) L) : IsEpigraphClosedConvexBifunction (m := m) (n := n) (fun u : Fin m → ℝ => fun x : Fin n → ℝ => iSup fun uStar : Fin m → ℝ => L uStar x + (-((finDot (n := m) uStar u : ℝ) : EReal)))

Lean source

Helper for Theorem 36.5: combining the previous two facts packages the reconstructed conjugate bifunction as a closed convex bifunction.

lemma helperForTheorem_36_5_reconstructedBifunction_closed_convex {m n : ℕ} {L : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hLower : IsLowerClosedInSecond (m := m) (n := n) L) (hConv : IsConvexInSecond (m := m) (n := n) L) : IsEpigraphConvexBifunction (m := m) (n := n) (fun u : Fin m → ℝ => fun x : Fin n → ℝ => iSup fun uStar : Fin m → ℝ => L uStar x + (-((finDot (n := m) uStar u : ℝ) : EReal))) ∧ IsEpigraphClosedConvexBifunction (m := m) (n := n) (fun u : Fin m → ℝ => fun x : Fin n → ℝ => iSup fun uStar : Fin m → ℝ => L uStar x + (-((finDot (n := m) uStar u : ℝ) : EReal)))

Lean source

Helper for Theorem 36.5: once each fixed-x slice of L is closed concave, the Lagrangian of the reconstructed bifunction recovers L pointwise by the concave biconjugation theorem.

lemma helperForTheorem_36_5_reconstructedBifunction_recovers_L {m n : ℕ} {L : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hClosedSlices : ∀ x : Fin n → ℝ, ClosedConcaveFunction (fun uStar : Fin m → ℝ => L uStar x)) (hProperSlices : ∀ x : Fin n → ℝ, ProperConcaveERealFunction (fun uStar : Fin m → ℝ => L uStar x)) (hFconv : IsEpigraphConvexBifunction (m := m) (n := n) (fun u : Fin m → ℝ => fun x : Fin n → ℝ => iSup fun uStar : Fin m → ℝ => L uStar x + (-((finDot (n := m) uStar u : ℝ) : EReal)))) : ∀ uStar : Fin m → ℝ, ∀ x : Fin n → ℝ, bifunctionLagrangian (m := m) (n := n) ⟨(fun u : Fin m → ℝ => fun x : Fin n → ℝ => iSup fun vStar : Fin m → ℝ => L vStar x + (-((finDot (n := m) vStar u : ℝ) : EReal))), hFconv⟩ uStar x = L uStar x

§36 — Minimax Problems — Part 6

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Theorem 36.5, with the necessary projection-closedness qualification: for a bifunction L that is lower closed in its second variable, being the Lagrangian of a convex program associated with a closed convex bifunction is equivalent to being upper closed concave-convex, provided its fixed-x concave slices are proper.

theorem lagrangian_iff_upperClosedConcaveConvex {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hLowerL : IsLowerClosedInSecond (m := m) (n := n) L) (hProperSlices : ∀ x : Fin n → ℝ, ProperConcaveERealFunction (fun uStar : Fin m → ℝ => L uStar x)) : (∃ (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hFconv : IsEpigraphConvexBifunction (m := m) (n := n) F), IsEpigraphClosedConvexBifunction (m := m) (n := n) F ∧ L = bifunctionLagrangian (m := m) (n := n) ⟨F, hFconv⟩) ↔ IsUpperClosedConcaveConvex (m := m) (n := n) L

Lean source

The bifunction F : ℝ^m → ℝ^n → [-∞, +∞] obtained from a Lagrangian-type bifunction L(uStar, x) by taking the pointwise Fenchel conjugate in the first variable: (F u) x = sup_{uStar} (L(uStar, x) - ⟨uStar, u⟩).

noncomputable def lagrangianToBifunction {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin m → ℝ) → (Fin n → ℝ) → EReal := fun u x => iSup fun uStar : Fin m → ℝ => L uStar x + (-((finDot (n := m) uStar u : ℝ) : EReal))

Lean source

The primal objective function associated to a Lagrangian-type bifunction L(uStar, x): φ(x) = sup_{uStar} L(uStar, x).

noncomputable def primalObjectiveOfLagrangian {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin n → ℝ) → EReal := fun x => iSup fun uStar : Fin m → ℝ => L uStar x

Lean source

The dual objective function associated to a Lagrangian-type bifunction L(uStar, x): ψ(uStar) = inf_x L(uStar, x).

noncomputable def dualObjectiveOfLagrangian {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (Fin m → ℝ) → EReal := fun uStar => iInf fun x : Fin n → ℝ => L uStar x

Lean source

Proposition 36.5.1: Let L be an upper closed concave-convex function on ℝ^m × ℝ^n. Define a (closed convex) bifunction F by (F u) x = sup_{uStar} (L(uStar, x) - ⟨uStar, u⟩). Then: (1) the primal objective function is φ(x) = sup_{uStar} L(uStar, x) and the primal optimal value is inf_x sup_{uStar} L(uStar, x); (2) the textbook Euclidean adjoint of F, evaluated at the sign-corrected dual pair, satisfies F^*(-xStar, -uStar) = inf_x (L(uStar, x) - ⟨xStar, x⟩); (3) the dual objective function is ψ(uStar) = inf_x L(uStar, x) and the dual optimal value is sup_{uStar} inf_x L(uStar, x).

theorem lagrangian_primal_dual_objective_and_adjoint_formulas {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hL : IsUpperClosedConcaveConvex (m := m) (n := n) L) (hProperSlices : ∀ x : Fin n → ℝ, ProperConcaveERealFunction (fun uStar : Fin m → ℝ => L uStar x)) : let F := lagrangianToBifunction (m := m) (n := n) L IsEpigraphConvexBifunction (m := m) (n := n) F ∧ IsEpigraphClosedConvexBifunction (m := m) (n := n) F ∧ (∀ x : Fin n → ℝ, primalObjectiveOfLagrangian (m := m) (n := n) L x = iSup fun uStar : Fin m → ℝ => L uStar x) ∧ (minimaxValue (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) L = iInf fun x : Fin n → ℝ => primalObjectiveOfLagrangian (m := m) (n := n) L x) ∧ (∀ xStar : Fin n → ℝ, ∀ uStar : Fin m → ℝ, bifunctionEuclideanAdjointTextbook (m := m) (n := n) F (-xStar) (-uStar) = iInf fun x : Fin n → ℝ => L uStar x + (-((finDot (n := n) xStar x : ℝ) : EReal))) ∧ (∀ uStar : Fin m → ℝ, dualObjectiveOfLagrangian (m := m) (n := n) L uStar = iInf fun x : Fin n → ℝ => L uStar x) ∧ (maximinValue (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) L = iSup fun uStar : Fin m → ℝ => dualObjectiveOfLagrangian (m := m) (n := n) L uStar)

Lean source

A dual vector xStar is a subgradient of an EReal-valued function f at x if for every z one has f z ≥ f x + ⟪xStar, z - x⟫, expressed via evaluation of xStar on z - x.

def IsSubgradientAtEReal {n : ℕ} (f : (Fin n → ℝ) → EReal) (x : Fin n → ℝ) (xStar : Module.Dual ℝ (Fin n → ℝ)) : Prop := ∀ z, f z ≥ f x + ((xStar (z - x) : ℝ) : EReal)

Lean source

The subdifferential of an EReal-valued function f at x, defined as the set of all subgradients at x.

def subdifferentialAtEReal {n : ℕ} (f : (Fin n → ℝ) → EReal) (x : Fin n → ℝ) : Set (Module.Dual ℝ (Fin n → ℝ)) := {g | IsSubgradientAtEReal f x g}

Lean source

The partial subdifferential ∂₁ L(uStar, x) in the first variable for a concave-convex function L(uStar, x), modeled as the (convex) subdifferential of uStar ↦ -L(uStar, x).

def lagrangianPartialSubdifferentialInFirst {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (uStar : Fin m → ℝ) (x : Fin n → ℝ) : Set (Module.Dual ℝ (Fin m → ℝ)) := subdifferentialAtEReal (n := m) (fun uStar' => -L uStar' x) uStar

Lean source

The partial subdifferential ∂₂ L(uStar, x) in the second variable for a concave-convex function L(uStar, x), modeled as the (convex) subdifferential of x ↦ L(uStar, x).

def lagrangianPartialSubdifferentialInSecond {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (uStar : Fin m → ℝ) (x : Fin n → ℝ) : Set (Module.Dual ℝ (Fin n → ℝ)) := subdifferentialAtEReal (n := n) (fun x' => L uStar x') x

Lean source

The product subdifferential ∂ L(uStar, x) for a concave-convex function L, modeled as ∂₁ L(uStar, x) × ∂₂ L(uStar, x).

def lagrangianSaddleSubdifferential {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (uStar : Fin m → ℝ) (x : Fin n → ℝ) : Set (Module.Dual ℝ (Fin m → ℝ) × Module.Dual ℝ (Fin n → ℝ)) := lagrangianPartialSubdifferentialInFirst (m := m) (n := n) L uStar x ×ˢ lagrangianPartialSubdifferentialInSecond (m := m) (n := n) L uStar x

Lean source

The Kuhn--Tucker condition for a concave-convex function L at the point (uStar, x), namely the inclusion (0, 0) ∈ ∂ L(uStar, x).

def IsKuhnTuckerCondition {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (uStar : Fin m → ℝ) (x : Fin n → ℝ) : Prop

Lean source

Proposition 36.5.2: Let L be a concave-convex function on ℝ^m × ℝ^n. A pair (uStar0, x0) is a saddle-point of L, i.e. L uStar x0 ≤ L uStar0 x0 ≤ L uStar0 x for all uStar and x, if and only if 0 ∈ ∂₂ L(uStar0, x0) and 0 ∈ ∂₁ L(uStar0, x0) (with ∂₁ modeled as the subdifferential of -L in the first variable). Equivalently, `(0, 0) ∈ ∂ L(uStar0, x0) = ∂₁ L(uStar0, x0) × ∂₂ L(uStar0, x0)`, and this inclusion is called the Kuhn--Tucker condition.

theorem saddlePoint_iff_zero_mem_partialSubdifferentials {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hL : IsConcaveInFirst (m := m) (n := n) L ∧ IsConvexInSecond (m := m) (n := n) L) (uStar0 : Fin m → ℝ) (x0 : Fin n → ℝ) : IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) L uStar0 x0 ↔ (0 ∈ lagrangianPartialSubdifferentialInSecond (m := m) (n := n) L uStar0 x0 ∧ 0 ∈ lagrangianPartialSubdifferentialInFirst (m := m) (n := n) L uStar0 x0)

Lean source

The associated convex program (P) of a convex bifunction F is *strongly consistent* in the sense assumed in Rockafellar's Theorem 36.6, modeled here by attainment of a finite primal optimal value for the program induced by F.

def IsStronglyConsistentAssociatedProgram {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := ∃ hF : IsEpigraphConvexBifunction (m := m) (n := n) F, ∃ x : Fin n → ℝ, let L := bifunctionLagrangian (m := m) (n := n) ⟨F, hF⟩ let φ := primalObjectiveOfLagrangian (m := m) (n := n) L φ x = iInf (fun x' : Fin n → ℝ => φ x') ∧ φ x ≠ (⊤ : EReal) ∧ φ x ≠ (⊥ : EReal)

Lean source

The associated convex program (P) of a convex bifunction F is *strictly consistent* in the sense assumed in Rockafellar's Theorem 36.6, modeled here by a Slater-type finite point for the Lagrangian slices of the program induced by F.

def IsStrictlyConsistentAssociatedProgram {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := ∃ hF : IsEpigraphConvexBifunction (m := m) (n := n) F, ∃ x : Fin n → ℝ, let L := bifunctionLagrangian (m := m) (n := n) ⟨F, hF⟩ let φ := primalObjectiveOfLagrangian (m := m) (n := n) L (∀ uStar : Fin m → ℝ, L uStar x ≠ (⊤ : EReal) ∧ L uStar x ≠ (⊥ : EReal)) ∧ φ x ≠ (⊤ : EReal) ∧ φ x ≠ (⊥ : EReal)

Lean source

The associated convex program (P) of a convex bifunction F is *polyhedral* in the sense used in Rockafellar's Theorem 36.6, modeled here by polyhedrality of the induced primal objective function.

def IsPolyhedralAssociatedProgram {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop

Lean source

The associated convex program (P) of a convex bifunction F is *consistent* in the sense used in Rockafellar's Theorem 36.6, modeled here by existence of a primal point with finite objective value for the program induced by F.

def IsConsistentAssociatedProgram {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := ∃ hF : IsEpigraphConvexBifunction (m := m) (n := n) F, ∃ x : Fin n → ℝ, let φ := primalObjectiveOfLagrangian (m := m) (n := n) (bifunctionLagrangian (m := m) (n := n) ⟨F, hF⟩) φ x ≠ (⊤ : EReal) ∧ φ x ≠ (⊥ : EReal)

Lean source

A predicate expressing Rockafellar's qualification hypotheses in Theorem 36.6 for the convex program (P) associated with a closed proper convex bifunction F: namely, that (P) is strongly consistent, or strictly consistent, or polyhedral and consistent.

def AssociatedProgramQualifiesForKuhnTucker {m n : ℕ} (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Prop := IsStronglyConsistentAssociatedProgram (m := m) (n := n) F ∨ IsStrictlyConsistentAssociatedProgram (m := m) (n := n) F ∨ (IsPolyhedralAssociatedProgram (m := m) (n := n) F ∧ IsConsistentAssociatedProgram (m := m) (n := n) F)

Lean source

A point x is a primal optimal solution for a Lagrangian L(uStar, x) if it attains the primal value inf_x sup_{uStar} L(uStar, x).

def IsPrimalOptimalSolutionOfLagrangian {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (x : Fin n → ℝ) : Prop := primalObjectiveOfLagrangian (m := m) (n := n) L x = iInf fun x' : Fin n → ℝ => primalObjectiveOfLagrangian (m := m) (n := n) L x'

Lean source

A point x is an optimal solution of the convex program (P) associated with a convex bifunction F, expressed via the Lagrangian bifunctionLagrangian F.

def IsOptimalSolutionOfAssociatedProgram {m n : ℕ} (F : {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal // IsEpigraphConvexBifunction (m := m) (n := n) F}) (x : Fin n → ℝ) : Prop := IsPrimalOptimalSolutionOfLagrangian (m := m) (n := n) (bifunctionLagrangian (m := m) (n := n) F) x

Lean source

A Kuhn--Tucker vector uStar for the convex program associated with F at the primal point x, meaning (0,0) ∈ ∂ L(uStar, x) for the Lagrangian L = bifunctionLagrangian F.

def IsKuhnTuckerVectorOfAssociatedProgram {m n : ℕ} (F : {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal // IsEpigraphConvexBifunction (m := m) (n := n) F}) (x : Fin n → ℝ) (uStar : Fin m → ℝ) : Prop

Lean source

Theorem 36.6: Let (P) be the convex program (Section 28) associated with a closed proper convex bifunction F : ℝ^m → ℝ^n → [-∞, +∞]. Assume (P) is strongly (or strictly) consistent, or that (P) is polyhedral and consistent. With the explicit additional qualification that the dual value is attained with no gap, a vector x̃ ∈ ℝ^n is an optimal solution of (P) if and only if there exists ũ* ∈ ℝ^m such that (0,0) ∈ ∂ L(ũ*, x̃), where L is the Lagrangian of (P). The vectors ũ* satisfying this condition are exactly the Kuhn--Tucker vectors for (P).

theorem optimalSolution_iff_exists_kuhnTuckerVector {m n : ℕ} (F : {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal // IsEpigraphConvexBifunction (m := m) (n := n) F}) (hFclosed : IsEpigraphClosedConvexBifunction (m := m) (n := n) F.1) (hFproper : IsEpigraphProperConvexBifunction (m := m) (n := n) F.1) (hQual : AssociatedProgramQualifiesForKuhnTucker (m := m) (n := n) F.1) (hDualAttain : ∃ uStar : Fin m → ℝ, dualObjectiveOfLagrangian (m := m) (n := n) (bifunctionLagrangian (m := m) (n := n) F) uStar = iInf fun x : Fin n → ℝ => primalObjectiveOfLagrangian (m := m) (n := n) (bifunctionLagrangian (m := m) (n := n) F) x) (xTilde : Fin n → ℝ) : IsOptimalSolutionOfAssociatedProgram (m := m) (n := n) F xTilde ↔ ∃ uStarTilde : Fin m → ℝ, IsKuhnTuckerVectorOfAssociatedProgram (m := m) (n := n) F xTilde uStarTilde

Lean source

Helper for Proposition 36.5.1: the dimension-one constant-zero kernel is an admissible upper closed concave-convex Lagrangian.

lemma helperForProposition_36_5_1_constantZero_isUpperClosedConcaveConvex : IsUpperClosedConcaveConvex (m := 1) (n := 1) (fun _ _ => (0 : EReal))

Lean source

Helper for Proposition 36.5.1: the bifunction reconstructed from the dimension-one constant-zero kernel takes the value ⊤ at the witness (u, x) = (1, 0).

lemma helperForProposition_36_5_1_constantZero_lagrangianAtOneZero_eq_top : lagrangianToBifunction (m := 1) (n := 1) (fun _ _ => (0 : EReal)) helperForProposition_36_4_3_oneVector helperForProposition_36_4_3_zeroVector = (⊤ : EReal)

Lean source

Helper for Proposition 36.5.1: after applying the repository's swap-and-negate inverse, the same dimension-one witness yields the value ⊥.

lemma helperForProposition_36_5_1_constantZero_inverseAtZeroOne_eq_bot : bifunctionInverse (lagrangianToBifunction (m := 1) (n := 1) (fun _ _ => (0 : EReal))) helperForProposition_36_4_3_zeroVector helperForProposition_36_4_3_oneVector = (⊥ : EReal)

Lean source

Helper for Proposition 36.5.1: the displayed clause-(2) right-hand side evaluates to 0 at the witness xStar = 0, uStar = 1 for the dimension-one constant-zero kernel.

lemma helperForProposition_36_5_1_constantZero_displayedRightSideAtZeroOne_eq_zero : (iInf fun x : Fin 1 → ℝ => (fun _ _ => (0 : EReal)) helperForProposition_36_4_3_oneVector x + (-((finDot (n := 1) helperForProposition_36_4_3_zeroVector x : ℝ) : EReal))) = (0 : EReal)

Lean source

Helper for Proposition 36.5.1: the textbook adjoint operator from section36_part4 does realize the displayed clause-(2) value 0 at the constant-zero witness xStar = 0, uStar = 1.

lemma helperForProposition_36_5_1_constantZero_textbookAdjointAtZeroOne_eq_zero : bifunctionEuclideanAdjointTextbook (m := 1) (n := 1) (lagrangianToBifunction (m := 1) (n := 1) (fun _ _ => (0 : EReal))) helperForProposition_36_4_3_zeroVector helperForProposition_36_4_3_oneVector = (0 : EReal)

Lean source

Helper for Proposition 36.5.1: at the explicit constant-zero witness, the displayed clause-(2) right-hand side agrees with the textbook Euclidean adjoint from section36_part4.

lemma helperForProposition_36_5_1_constantZero_displayedRightSide_eq_textbookAdjointAtWitness : (iInf fun x : Fin 1 → ℝ => (fun _ _ => (0 : EReal)) helperForProposition_36_4_3_oneVector x + (-((finDot (n := 1) helperForProposition_36_4_3_zeroVector x : ℝ) : EReal))) = bifunctionEuclideanAdjointTextbook (m := 1) (n := 1) (lagrangianToBifunction (m := 1) (n := 1) (fun _ _ => (0 : EReal))) helperForProposition_36_4_3_zeroVector helperForProposition_36_4_3_oneVector

Lean source

Helper for Proposition 36.5.1: at the constant-zero witness, the theorem's current left-hand side bifunctionInverse F xStar uStar differs from the textbook adjoint bifunctionEuclideanAdjointTextbook F xStar uStar.

lemma helperForProposition_36_5_1_constantZero_inverseDiffersFromTextbookAdjointAtWitness : bifunctionInverse (lagrangianToBifunction (m := 1) (n := 1) (fun _ _ => (0 : EReal))) helperForProposition_36_4_3_zeroVector helperForProposition_36_4_3_oneVector ≠ bifunctionEuclideanAdjointTextbook (m := 1) (n := 1) (lagrangianToBifunction (m := 1) (n := 1) (fun _ _ => (0 : EReal))) helperForProposition_36_4_3_zeroVector helperForProposition_36_4_3_oneVector

Lean source

Helper for Proposition 36.5.1: any hypothetical clause-(2) equality at the explicit dimension-one constant-zero witness would identify the repository inverse with the textbook adjoint at that witness.

lemma helperForProposition_36_5_1_constantZero_clauseTwoAtWitness_forces_inverse_eq_textbookAdjoint (hClauseTwoAtWitness : bifunctionInverse (lagrangianToBifunction (m := 1) (n := 1) (fun _ _ => (0 : EReal))) helperForProposition_36_4_3_zeroVector helperForProposition_36_4_3_oneVector = (iInf fun x : Fin 1 → ℝ => (fun _ _ => (0 : EReal)) helperForProposition_36_4_3_oneVector x + (-((finDot (n := 1) helperForProposition_36_4_3_zeroVector x : ℝ) : EReal)))) : bifunctionInverse (lagrangianToBifunction (m := 1) (n := 1) (fun _ _ => (0 : EReal))) helperForProposition_36_4_3_zeroVector helperForProposition_36_4_3_oneVector = bifunctionEuclideanAdjointTextbook (m := 1) (n := 1) (lagrangianToBifunction (m := 1) (n := 1) (fun _ _ => (0 : EReal))) helperForProposition_36_4_3_zeroVector helperForProposition_36_4_3_oneVector

Lean source

Helper for Proposition 36.5.1: any hypothetical clause-(2) equality at the explicit dimension-one constant-zero witness forces the impossible identity ⊥ = 0.

lemma helperForProposition_36_5_1_constantZero_clauseTwoAtWitness_forces_bot_eq_zero (hClauseTwoAtWitness : bifunctionInverse (lagrangianToBifunction (m := 1) (n := 1) (fun _ _ => (0 : EReal))) helperForProposition_36_4_3_zeroVector helperForProposition_36_4_3_oneVector = (iInf fun x : Fin 1 → ℝ => (fun _ _ => (0 : EReal)) helperForProposition_36_4_3_oneVector x + (-((finDot (n := 1) helperForProposition_36_4_3_zeroVector x : ℝ) : EReal)))) : (⊥ : EReal) = (0 : EReal)

Lean source

Helper for Proposition 36.5.1: the clause-(2) identity already fails at the explicit witness xStar = 0, uStar = 1 for the dimension-one constant-zero kernel.

lemma helperForProposition_36_5_1_constantZero_clauseTwoFailsAtZeroOne : bifunctionInverse (lagrangianToBifunction (m := 1) (n := 1) (fun _ _ => (0 : EReal))) helperForProposition_36_4_3_zeroVector helperForProposition_36_4_3_oneVector ≠ (iInf fun x : Fin 1 → ℝ => (fun _ _ => (0 : EReal)) helperForProposition_36_4_3_oneVector x + (-((finDot (n := 1) helperForProposition_36_4_3_zeroVector x : ℝ) : EReal)))

Lean source

Helper for Proposition 36.5.1: the specialized clause-(2) cannot hold universally in the dimension-one constant-zero example.

lemma helperForProposition_36_5_1_constantZero_forallClauseTwoFails : ¬ (∀ xStar : Fin 1 → ℝ, ∀ uStar : Fin 1 → ℝ, bifunctionInverse (lagrangianToBifunction (m := 1) (n := 1) (fun _ _ => (0 : EReal))) xStar uStar = iInf fun x : Fin 1 → ℝ => (fun _ _ => (0 : EReal)) uStar x + (-((finDot (n := 1) xStar x : ℝ) : EReal)))

Lean source

Helper for Proposition 36.5.1: the specialized conclusion for the dimension-one constant-zero kernel is empty because clause (2) equates ⊥ with 0 at the explicit witness.

lemma helperForProposition_36_5_1_constantZero_conclusion_empty : ¬ (let F := lagrangianToBifunction (m := 1) (n := 1) (fun _ _ => (0 : EReal)) IsEpigraphConvexBifunction (m := 1) (n := 1) F ∧ IsEpigraphClosedConvexBifunction (m := 1) (n := 1) F ∧ (∀ x : Fin 1 → ℝ, primalObjectiveOfLagrangian (m := 1) (n := 1) (fun _ _ => (0 : EReal)) x = iSup fun _ : Fin 1 → ℝ => (0 : EReal)) ∧ (minimaxValue (C := (Fin 1 → ℝ)) (D := (Fin 1 → ℝ)) (fun _ _ => (0 : EReal)) = iInf fun x : Fin 1 → ℝ => primalObjectiveOfLagrangian (m := 1) (n := 1) (fun _ _ => (0 : EReal)) x) ∧ (∀ xStar : Fin 1 → ℝ, ∀ uStar : Fin 1 → ℝ, bifunctionInverse F xStar uStar = iInf fun x : Fin 1 → ℝ => (fun _ _ => (0 : EReal)) uStar x + (-((finDot (n := 1) xStar x : ℝ) : EReal))) ∧ (∀ uStar : Fin 1 → ℝ, dualObjectiveOfLagrangian (m := 1) (n := 1) (fun _ _ => (0 : EReal)) uStar = iInf fun _ : Fin 1 → ℝ => (0 : EReal)) ∧ (maximinValue (C := (Fin 1 → ℝ)) (D := (Fin 1 → ℝ)) (fun _ _ => (0 : EReal)) = iSup fun uStar : Fin 1 → ℝ => dualObjectiveOfLagrangian (m := 1) (n := 1) (fun _ _ => (0 : EReal)) uStar))

Lean source

Helper for Proposition 36.5.1: the current theorem statement has a concrete counterexample, namely the dimension-one constant-zero Lagrangian.

lemma helperForProposition_36_5_1_hasCounterexample : ∃ L : (Fin 1 → ℝ) → (Fin 1 → ℝ) → EReal, IsUpperClosedConcaveConvex (m := 1) (n := 1) L ∧ ¬ (let F := lagrangianToBifunction (m := 1) (n := 1) L IsEpigraphConvexBifunction (m := 1) (n := 1) F ∧ IsEpigraphClosedConvexBifunction (m := 1) (n := 1) F ∧ (∀ x : Fin 1 → ℝ, primalObjectiveOfLagrangian (m := 1) (n := 1) L x = iSup fun uStar : Fin 1 → ℝ => L uStar x) ∧ (minimaxValue (C := (Fin 1 → ℝ)) (D := (Fin 1 → ℝ)) L = iInf fun x : Fin 1 → ℝ => primalObjectiveOfLagrangian (m := 1) (n := 1) L x) ∧ (∀ xStar : Fin 1 → ℝ, ∀ uStar : Fin 1 → ℝ, bifunctionInverse F xStar uStar = iInf fun x : Fin 1 → ℝ => L uStar x + (-((finDot (n := 1) xStar x : ℝ) : EReal))) ∧ (∀ uStar : Fin 1 → ℝ, dualObjectiveOfLagrangian (m := 1) (n := 1) L uStar = iInf fun x : Fin 1 → ℝ => L uStar x) ∧ (maximinValue (C := (Fin 1 → ℝ)) (D := (Fin 1 → ℝ)) L = iSup fun uStar : Fin 1 → ℝ => dualObjectiveOfLagrangian (m := 1) (n := 1) L uStar))

Lean source

Helper for Proposition 36.5.1: any hypothetical proof of the full generic theorem statement specializes to the explicit dimension-one constant-zero witness.

lemma helperForProposition_36_5_1_specializeGenericStatementToConstantZero (hGeneric : ∀ {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (_hL : IsUpperClosedConcaveConvex (m := m) (n := n) L), let F := lagrangianToBifunction (m := m) (n := n) L IsEpigraphConvexBifunction (m := m) (n := n) F ∧ IsEpigraphClosedConvexBifunction (m := m) (n := n) F ∧ (∀ x : Fin n → ℝ, primalObjectiveOfLagrangian (m := m) (n := n) L x = iSup fun uStar : Fin m → ℝ => L uStar x) ∧ (minimaxValue (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) L = iInf fun x : Fin n → ℝ => primalObjectiveOfLagrangian (m := m) (n := n) L x) ∧ (∀ xStar : Fin n → ℝ, ∀ uStar : Fin m → ℝ, bifunctionInverse F xStar uStar = iInf fun x : Fin n → ℝ => L uStar x + (-((finDot (n := n) xStar x : ℝ) : EReal))) ∧ (∀ uStar : Fin m → ℝ, dualObjectiveOfLagrangian (m := m) (n := n) L uStar = iInf fun x : Fin n → ℝ => L uStar x) ∧ (maximinValue (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) L = iSup fun uStar : Fin m → ℝ => dualObjectiveOfLagrangian (m := m) (n := n) L uStar)) : let F := lagrangianToBifunction (m := 1) (n := 1) (fun _ _ => (0 : EReal)) IsEpigraphConvexBifunction (m := 1) (n := 1) F ∧ IsEpigraphClosedConvexBifunction (m := 1) (n := 1) F ∧ (∀ x : Fin 1 → ℝ, primalObjectiveOfLagrangian (m := 1) (n := 1) (fun _ _ => (0 : EReal)) x = iSup fun _ : Fin 1 → ℝ => (0 : EReal)) ∧ (minimaxValue (C := (Fin 1 → ℝ)) (D := (Fin 1 → ℝ)) (fun _ _ => (0 : EReal)) = iInf fun x : Fin 1 → ℝ => primalObjectiveOfLagrangian (m := 1) (n := 1) (fun _ _ => (0 : EReal)) x) ∧ (∀ xStar : Fin 1 → ℝ, ∀ uStar : Fin 1 → ℝ, bifunctionInverse F xStar uStar = iInf fun x : Fin 1 → ℝ => (fun _ _ => (0 : EReal)) uStar x + (-((finDot (n := 1) xStar x : ℝ) : EReal))) ∧ (∀ uStar : Fin 1 → ℝ, dualObjectiveOfLagrangian (m := 1) (n := 1) (fun _ _ => (0 : EReal)) uStar = iInf fun _ : Fin 1 → ℝ => (0 : EReal)) ∧ (maximinValue (C := (Fin 1 → ℝ)) (D := (Fin 1 → ℝ)) (fun _ _ => (0 : EReal)) = iSup fun uStar : Fin 1 → ℝ => dualObjectiveOfLagrangian (m := 1) (n := 1) (fun _ _ => (0 : EReal)) uStar)

Lean source

Helper for Proposition 36.5.1: the full generic theorem statement is false, because specializing it to the dimension-one constant-zero kernel recovers the explicit contradiction proved above.

lemma helperForProposition_36_5_1_targetStatement_isFalse : ¬ (∀ {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (_hL : IsUpperClosedConcaveConvex (m := m) (n := n) L), let F := lagrangianToBifunction (m := m) (n := n) L IsEpigraphConvexBifunction (m := m) (n := n) F ∧ IsEpigraphClosedConvexBifunction (m := m) (n := n) F ∧ (∀ x : Fin n → ℝ, primalObjectiveOfLagrangian (m := m) (n := n) L x = iSup fun uStar : Fin m → ℝ => L uStar x) ∧ (minimaxValue (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) L = iInf fun x : Fin n → ℝ => primalObjectiveOfLagrangian (m := m) (n := n) L x) ∧ (∀ xStar : Fin n → ℝ, ∀ uStar : Fin m → ℝ, bifunctionInverse F xStar uStar = iInf fun x : Fin n → ℝ => L uStar x + (-((finDot (n := n) xStar x : ℝ) : EReal))) ∧ (∀ uStar : Fin m → ℝ, dualObjectiveOfLagrangian (m := m) (n := n) L uStar = iInf fun x : Fin n → ℝ => L uStar x) ∧ (maximinValue (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) L = iSup fun uStar : Fin m → ℝ => dualObjectiveOfLagrangian (m := m) (n := n) L uStar))

Lean source

Helper for Proposition 36.5.2: zero belongs to the second partial subdifferential exactly when the fixed-uStar0 slice is minimized at x0.

lemma helperForProposition_36_5_2_zero_mem_second_partialSubdifferential_iff {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (uStar0 : Fin m → ℝ) (x0 : Fin n → ℝ) : 0 ∈ lagrangianPartialSubdifferentialInSecond (m := m) (n := n) L uStar0 x0 ↔ ∀ x : Fin n → ℝ, L uStar0 x0 ≤ L uStar0 x

Lean source

Helper for Proposition 36.5.2: the reflected first-slice inequality for -L is equivalent to the original maximizing inequality for L.

lemma helperForProposition_36_5_2_reflected_firstSlice_inequality_iff {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (uStar0 : Fin m → ℝ) (x0 : Fin n → ℝ) : (∀ uStar : Fin m → ℝ, -L uStar0 x0 ≤ -L uStar x0) ↔ ∀ uStar : Fin m → ℝ, L uStar x0 ≤ L uStar0 x0

Lean source

Helper for Proposition 36.5.2: zero belongs to the first partial subdifferential exactly when the fixed-x0 slice is maximized at uStar0.

lemma helperForProposition_36_5_2_zero_mem_first_partialSubdifferential_iff {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (uStar0 : Fin m → ℝ) (x0 : Fin n → ℝ) : 0 ∈ lagrangianPartialSubdifferentialInFirst (m := m) (n := n) L uStar0 x0 ↔ ∀ uStar : Fin m → ℝ, L uStar x0 ≤ L uStar0 x0

Lean source

Helper for Proposition 36.5.2: the saddle-point inequalities are exactly the two one-variable optimality conditions, ordered to match the partial subdifferentials.

lemma helperForProposition_36_5_2_saddle_iff_split_zero_partialSubdifferentials {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (uStar0 : Fin m → ℝ) (x0 : Fin n → ℝ) : IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) L uStar0 x0 ↔ ((∀ x : Fin n → ℝ, L uStar0 x0 ≤ L uStar0 x) ∧ (∀ uStar : Fin m → ℝ, L uStar x0 ≤ L uStar0 x0))

Lean source

Helper for Proposition 36.5.2: the Kuhn--Tucker condition is exactly the pair of zero partial-subdifferential memberships appearing in the saddle-point characterization.

lemma helperForProposition_36_5_2_kuhnTuckerCondition_iff {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (uStar0 : Fin m → ℝ) (x0 : Fin n → ℝ) : IsKuhnTuckerCondition (m := m) (n := n) L uStar0 x0 ↔ (0 ∈ lagrangianPartialSubdifferentialInSecond (m := m) (n := n) L uStar0 x0 ∧ 0 ∈ lagrangianPartialSubdifferentialInFirst (m := m) (n := n) L uStar0 x0)

Lean source

Helper for Proposition 36.5.2: after packaging the two partial-subdifferential conditions into the product subdifferential, saddle points are exactly Kuhn--Tucker points.

lemma helperForProposition_36_5_2_saddlePoint_iff_kuhnTuckerCondition {m n : ℕ} (L : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hL : IsConcaveInFirst (m := m) (n := n) L ∧ IsConvexInSecond (m := m) (n := n) L) (uStar0 : Fin m → ℝ) (x0 : Fin n → ℝ) : IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) L uStar0 x0 ↔ IsKuhnTuckerCondition (m := m) (n := n) L uStar0 x0

§37 — Conjugate Saddle-Functions and Minimax Theorems — Part 1

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Theorem 37.1: the inf_u sup_x* value of the affine tilt of a saddle kernel K.

variable {m n : ℕ} noncomputable def theorem37ValueInfSup (K : SaddleFunction m n) (uStar : Fin m → ℝ) (x : Fin n → ℝ) : EReal := iInf fun u : Fin m → ℝ => iSup fun xStar : Fin n → ℝ => (((finDot u uStar + finDot x xStar : ℝ) : EReal) - K u xStar)

Lean source

Theorem 37.1: the sup_x* inf_u value of the affine tilt of a saddle kernel K.

variable {m n : ℕ} noncomputable def theorem37ValueSupInf (K : SaddleFunction m n) (uStar : Fin m → ℝ) (x : Fin n → ℝ) : EReal := iSup fun xStar : Fin n → ℝ => iInf fun u : Fin m → ℝ => (((finDot u uStar + finDot x xStar : ℝ) : EReal) - K u xStar)

Lean source

Proposition 37.1.2: the lower conjugate never exceeds the upper conjugate. In the present formalization these are theorem37ValueSupInf and theorem37ValueInfSup, respectively.

variable {m n : ℕ} theorem lowerConjugate_le_upperConjugate (K : SaddleFunction m n) (uStar : Fin m → ℝ) (x : Fin n → ℝ) : theorem37ValueSupInf K uStar x ≤ theorem37ValueInfSup K uStar x

Lean source

Theorem 37.1: let F be a closed convex bifunction, let K ∈ Ω(F), and let KStar ∈ Ω(F_*). Then the two affine-tilt minimax values attached to K recover the inverse pairing of F and the inverse-adjoint pairing of F, while the corresponding values attached to KStar recover the original pairings of F.

variable {m n : ℕ} theorem section37_theorem37_1 (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (K : SaddleFunction m n) (hK : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) (KStar : SaddleFunction m n) (hFStar : IsClosedConvexBifunction (bifunctionInverse (section34ConcaveBifunctionAdjoint F))) (hKStar : KStar ∈ EquivalenceClassGeneratedByConvexBifunction ⟨bifunctionInverse (section34ConcaveBifunctionAdjoint F), hFStar⟩) (hGlobal : Section34Theorem34_2GlobalQualification m n) : (∀ uStar x, theorem37ValueInfSup K uStar x = infPairing uStar (bifunctionInverse F x)) ∧ (∀ uStar x, theorem37ValueSupInf K uStar x = convexBifunctionPairing (bifunctionInverse (section34ConcaveBifunctionAdjoint F)) uStar x) ∧ (∀ u xStar, theorem37ValueInfSup KStar u xStar = convexBifunctionPairing F u xStar) ∧ ∀ u xStar, theorem37ValueSupInf KStar u xStar = convexBifunctionPairing F u xStar

Lean source

Helper for Proposition 37.1.2: the affine-tilted kernel whose maximin and minimax values recover the lower and upper conjugates.

variable {m n : ℕ} noncomputable def helperForProposition_37_1_2_affineTiltKernel (K : SaddleFunction m n) (uStar : Fin m → ℝ) (x : Fin n → ℝ) : (Fin n → ℝ) → (Fin m → ℝ) → EReal := fun xStar u => (((finDot u uStar + finDot x xStar : ℝ) : EReal) - K u xStar)

Lean source

Helper for Proposition 37.1.2: the lower conjugate is the maximin value of the affine-tilted kernel.

variable {m n : ℕ} lemma helperForProposition_37_1_2_affineTiltKernel_maximin_eq (K : SaddleFunction m n) (uStar : Fin m → ℝ) (x : Fin n → ℝ) : maximinValue (C := Fin n → ℝ) (D := Fin m → ℝ) (helperForProposition_37_1_2_affineTiltKernel K uStar x) = theorem37ValueSupInf K uStar x

Lean source

Helper for Proposition 37.1.2: the upper conjugate is the minimax value of the affine-tilted kernel.

variable {m n : ℕ} lemma helperForProposition_37_1_2_affineTiltKernel_minimax_eq (K : SaddleFunction m n) (uStar : Fin m → ℝ) (x : Fin n → ℝ) : minimaxValue (C := Fin n → ℝ) (D := Fin m → ℝ) (helperForProposition_37_1_2_affineTiltKernel K uStar x) = theorem37ValueInfSup K uStar x

Lean source

Helper for Corollary 37.1.1: a closed convex bifunction is graph-convex once its sectionwise closure is identified with the original section values.

variable {m n : ℕ} lemma helperForCorollary_37_1_1_closedConvex_isGraphConvex (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) : IsGraphConvexBifunction F

Lean source

Helper for Corollary 37.1.1: the original pairing slice is already fixed by the one-variable convex closure operator.

variable {m n : ℕ} lemma helperForCorollary_37_1_1_originalKernelSlice_functionConvexClosure_eq_self (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (u : Fin m → ℝ) (xStar : Fin n → ℝ) : functionConvexClosure (fun y => convexBifunctionPairing F u y) xStar = convexBifunctionPairing F u xStar

Lean source

Helper for Corollary 37.1.1: Theorem 34.2 identifies each fixed xStar slice of the concave adjoint with the original convex pairing section.

variable {m n : ℕ} lemma helperForCorollary_37_1_1_infPairing_concaveAdjointSlice_eq_originalPairing (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (hGlobal : Section34Theorem34_2GlobalQualification m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) : infPairing u (section34ConcaveBifunctionAdjoint F xStar) = convexBifunctionPairing F u xStar

Lean source

Helper for Corollary 37.1.1: the first adjoint-inverse iterate at a fixed x is the negative of the convex conjugate of the original pairing slice.

variable {m n : ℕ} lemma helperForCorollary_37_1_1_firstIterateSlice_concaveConjugate_eq_neg_sliceConjugate (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (hGlobal : Section34Theorem34_2GlobalQualification m n) (u : Fin m → ℝ) (x : Fin n → ℝ) : concaveConjugate (fun u' => convexBifunctionPairing (bifunctionInverse (section34ConcaveBifunctionAdjoint F)) u' x) u = -convexConjugate (fun y => convexBifunctionPairing F u y) x

Lean source

Helper for Corollary 37.1.1: unfolding the double adjoint-inverse slice produces the ordinary slice biconjugate of the original pairing section.

variable {m n : ℕ} lemma helperForCorollary_37_1_1_doubleAdjointInverseSlice_eq_negConcaveConjugateConjugate (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (hGlobal : Section34Theorem34_2GlobalQualification m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) : convexBifunctionPairing (bifunctionInverse (section34ConcaveBifunctionAdjoint (bifunctionInverse (section34ConcaveBifunctionAdjoint F)))) u xStar = convexConjugate (convexConjugate (fun y => convexBifunctionPairing F u y)) xStar

Lean source

Helper for Corollary 37.1.1: the biconjugate of the original pairing slice agrees with its one-variable convex closure.

variable {m n : ℕ} lemma helperForCorollary_37_1_1_originalKernelSlice_biconjugate_eq_functionConvexClosure (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (u : Fin m → ℝ) (xStar : Fin n → ℝ) : convexConjugate (convexConjugate (fun y => convexBifunctionPairing F u y)) xStar = functionConvexClosure (fun y => convexBifunctionPairing F u y) xStar

Lean source

Helper for Corollary 37.1.1: the upper recovery kernel obtained from F_* should coincide with the convex closure of the original slice y ↦ ⟪F u, y⟫. This is the remaining Lean blocker inside Theorem 37.1: after unfolding, the left-hand side is a nested convexConjugate/concaveConjugate expression for the double adjoint-inverse transform, while the right-hand side is the one-variable convex closure of the original slice. The intended route is to identify that nested transform with the Fenchel biconjugate of the slice and then invoke helperForTheorem33_1_biconjugate_eq_functionConvexClosure_of_convex.

variable {m n : ℕ} lemma helperForCorollary_37_1_1_upperRecoveredKernel_eq_functionConvexClosure (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (hGlobal : Section34Theorem34_2GlobalQualification m n) (u : Fin m → ℝ) (xStar : Fin n → ℝ) : convexBifunctionPairing (bifunctionInverse (section34ConcaveBifunctionAdjoint (bifunctionInverse (section34ConcaveBifunctionAdjoint F)))) u xStar = functionConvexClosure (fun y => convexBifunctionPairing F u y) xStar

Lean source

Helper for Corollary 37.1.2: a closed proper saddle-function admits a proper closed convex representative in the Section 34 generated class.

variable {m n : ℕ} lemma helperForCorollary_37_1_2_closedProperRepresentative (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) : ∃ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, ∃ hF : IsClosedConvexBifunction F, IsProperConvexBifunction F ∧ K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩

Lean source

Helper for Corollary 37.1.2: once the dual bifunction F_* is known to be closed convex, Theorem 37.1 rewrites the lower conjugate of K as the canonical dual pairing.

variable {m n : ℕ} lemma helperForCorollary_37_1_2_lowerConjugate_eq_dualLowerKernel (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (K : SaddleFunction m n) (hK : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) (hFStar : IsClosedConvexBifunction (bifunctionInverse (section34ConcaveBifunctionAdjoint F))) (hGlobal : Section34Theorem34_2GlobalQualification m n) : ∀ uStar x, theorem37ValueSupInf K uStar x = convexBifunctionPairing (bifunctionInverse (section34ConcaveBifunctionAdjoint F)) uStar x

Lean source

Helper for Corollary 37.1.2: the explicit dual inverse F_* is already proper on the convex side once the original representative F is closed proper convex.

variable {m n : ℕ} lemma helperForCorollary_37_1_2_dualAdjointInverse_isProperConvex (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (hFproper : IsProperConvexBifunction F) : IsProperConvexBifunction (bifunctionInverse (section34ConcaveBifunctionAdjoint F))

§37 — Conjugate Saddle-Functions and Minimax Theorems — Part 2

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Corollary 37.1.2: the lower and upper conjugates of a closed proper saddle-function admit a common nonempty convex effective domain on which the usual closure and relative-interior conclusions of Section 34 are expected to hold.

variable {m n : ℕ} theorem corollary37_1_2_lower_upper_conjugates_structure (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) : let KLowerStar : SaddleFunction m n := fun uStar x => theorem37ValueSupInf K uStar x let KUpperStar : SaddleFunction m n := fun uStar x => theorem37ValueInfSup K uStar x ∃ CStar : Set (Fin m → ℝ), ∃ DStar : Set (Fin n → ℝ), CStar.Nonempty ∧ DStar.Nonempty ∧ Convex ℝ CStar ∧ Convex ℝ DStar ∧ effectiveDomain₁ KLowerStar = CStar ∧ effectiveDomain₂ KLowerStar = DStar ∧ effectiveDomain₁ KUpperStar = CStar ∧ effectiveDomain₂ KUpperStar = DStar ∧ partialClosure₁ KLowerStar = KUpperStar ∧ partialClosure₂ KUpperStar = KLowerStar ∧ ∀ uStar x, uStar ∈ intrinsicInterior ℝ CStar ∨ x ∈ intrinsicInterior ℝ DStar → KLowerStar uStar x = KUpperStar uStar x

Lean source

Proposition 37.1.3: the minimax and maximin values of K are the negatives of the lower and upper conjugates at the origin, and these two conjugates share the same product effective domain C* × D*.

variable {m n : ℕ} theorem proposition37_1_3_minimax_maximin_and_commonDomain (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) : ∃ CStar : Set (Fin m → ℝ), ∃ DStar : Set (Fin n → ℝ), minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K = -theorem37ValueSupInf K 0 0 ∧ maximinValue (C := Fin m → ℝ) (D := Fin n → ℝ) K = -theorem37ValueInfSup K 0 0 ∧ saddleEffectiveDomain (fun uStar x => theorem37ValueSupInf K uStar x) = CStar ×ˢ DStar ∧ saddleEffectiveDomain (fun uStar x => theorem37ValueInfSup K uStar x) = CStar ×ˢ DStar

Lean source

Helper for Corollary 37.1.2: Theorem 34.2 already upgrades the Section 34 adjoint to a graph-function concave-closed bifunction.

variable {m n : ℕ} lemma helperForCorollary_37_1_2_section34Adjoint_graphConcaveClosed (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (hGlobal : Section34Theorem34_2GlobalQualification m n) : IsFunctionConcaveClosed (graphFunctionOfBifunction (section34ConcaveBifunctionAdjoint F))

Lean source

Helper for Corollary 37.1.2: the inverse graph is obtained by swapping the two coordinate blocks before evaluating the negated graph of the original bifunction.

variable {m n : ℕ} def helperForCorollary_37_1_2_swapBlock (z : Fin (m + n) → ℝ) : Fin (n + m) → ℝ := Fin.append (fun i => z (Fin.natAdd m i)) (fun j => z (Fin.castAdd n j))

Lean source

Helper for Corollary 37.1.2: negating a Rockafellar concave bifunction produces a Rockafellar convex bifunction on the same ambient product space.

variable {m n : ℕ} lemma helperForCorollary_37_1_2_negatedConcaveBifunction_isRockafellarConvex (G : (Fin n → ℝ) → (Fin m → ℝ) → EReal) (hRockG : IsRockafellarConcaveBifunction G) (hNoTopG : HasNoTopValuesBifunction G) : IsRockafellarConvexBifunction (fun x u => -G x u)

Lean source

Helper for Corollary 37.1.2: negating a graph-function-closed concave bifunction turns the graph function into a convex-closed one.

variable {m n : ℕ} lemma helperForCorollary_37_1_2_negatedGraph_isFunctionConvexClosed (G : (Fin n → ℝ) → (Fin m → ℝ) → EReal) (hGraphClosedG : IsFunctionConcaveClosed (graphFunctionOfBifunction G)) : IsFunctionConvexClosed (graphFunctionOfBifunction (fun x u => -G x u))

Lean source

Helper for Corollary 37.1.2: the graph of the inverse bifunction is exactly the graph of the negated original bifunction precomposed with the block-swap map.

variable {m n : ℕ} lemma helperForCorollary_37_1_2_inverse_graphFunction_eq_blockSwap_negatedGraph (G : (Fin n → ℝ) → (Fin m → ℝ) → EReal) : graphFunctionOfBifunction (bifunctionInverse G) = fun z => graphFunctionOfBifunction (fun x u => -G x u) (helperForCorollary_37_1_2_swapBlock (m := m) (n := n) z)

Lean source

Helper for Corollary 37.1.2: precomposing a convex graph function with the block-swap map preserves convexity.

variable {m n : ℕ} lemma helperForCorollary_37_1_2_blockSwap_precomp_preserves_graphConvex {f : (Fin (n + m) → ℝ) → EReal} (hf : IsERealConvexOn (Set.univ : Set (Fin (n + m) → ℝ)) f) : IsERealConvexOn (Set.univ : Set (Fin (m + n) → ℝ)) (fun z => f (helperForCorollary_37_1_2_swapBlock (m := m) (n := n) z))

Lean source

Helper for Corollary 37.1.2: once a concave bifunction is graph-function closed, its inverse only still needs the swap-negate graph-closure transport to become a closed convex bifunction.

variable {m n : ℕ} lemma helperForCorollary_37_1_2_inverseClosedConcave_to_closedConvex (G : (Fin n → ℝ) → (Fin m → ℝ) → EReal) (hRockG : IsRockafellarConcaveBifunction G) (hNoTopG : HasNoTopValuesBifunction G) (hGraphClosedG : IsFunctionConcaveClosed (graphFunctionOfBifunction G)) : IsClosedConvexBifunction (bifunctionInverse G)

Lean source

Helper for Corollary 37.1.2: the dual inverse F_* should inherit image-closed convexity from the image-closed concavity of the adjoint F^*.

variable {m n : ℕ} lemma helperForCorollary_37_1_2_dualAdjointInverse_isClosedConvex (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (hGlobal : Section34Theorem34_2GlobalQualification m n) : IsClosedConvexBifunction (bifunctionInverse (section34ConcaveBifunctionAdjoint F))

Lean source

Helper for Corollary 37.1.2: the convex pairing of the dual inverse F_* is exactly the infPairing of the original inverse slice.

variable {m n : ℕ} lemma helperForCorollary_37_1_2_dualInversePairing_eq_originalInverseInfPairing (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (hFStar : IsClosedConvexBifunction (bifunctionInverse (section34ConcaveBifunctionAdjoint F))) (hGlobal : Section34Theorem34_2GlobalQualification m n) (uStar : Fin m → ℝ) (x : Fin n → ℝ) : let FStar := bifunctionInverse (section34ConcaveBifunctionAdjoint F) infPairing uStar (bifunctionInverse F x) = convexBifunctionPairing FStar uStar x

Lean source

Helper for Corollary 37.1.2: once the dual inverse F_* is known to be closed convex, the upper conjugate of K should be the Section 34 upper pairing attached to F_*.

variable {m n : ℕ} lemma helperForCorollary_37_1_2_upperConjugate_eq_dualUpperKernel (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (K : SaddleFunction m n) (hK : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) (hFStar : IsClosedConvexBifunction (bifunctionInverse (section34ConcaveBifunctionAdjoint F))) (hGlobal : Section34Theorem34_2GlobalQualification m n) : ∀ uStar x, theorem37ValueInfSup K uStar x = section34ConcaveBifunctionAdjointPairing (bifunctionInverse (section34ConcaveBifunctionAdjoint F)) uStar x

Lean source

Helper for Proposition 37.1.3: a point in the saddle effective domain of the canonical dual kernel projects to points in the corresponding first and second effective domains.

variable {m n : ℕ} lemma helperForProposition_37_1_3_kernelDomainNonempty_of_saddleDomain (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : (saddleEffectiveDomain (convexBifunctionClosedKernel F)).Nonempty → (convexBifunctionDomain F).Nonempty ∧ (convexBifunctionKernelAdjointDomain F).Nonempty

Lean source

Helper for Proposition 37.1.3: the zero affine tilt in the upper conjugate leaves only the plain negated kernel inside the inf_u sup_v expression.

variable {m n : ℕ} lemma helperForProposition_37_1_3_zeroTiltUpperConjugate (K : SaddleFunction m n) : theorem37ValueInfSup K 0 0 = iInf (fun u : Fin m → ℝ => iSup (fun x : Fin n → ℝ => -K u x))

Lean source

Helper for Proposition 37.1.3: the zero affine tilt in the lower conjugate leaves only the plain negated kernel inside the sup_v inf_u expression.

variable {m n : ℕ} lemma helperForProposition_37_1_3_zeroTiltLowerConjugate (K : SaddleFunction m n) : theorem37ValueSupInf K 0 0 = iSup (fun x : Fin n → ℝ => iInf (fun u : Fin m → ℝ => -K u x))

Lean source

Helper for Proposition 37.1.3: evaluating the lower conjugate at the origin recovers the plain inf_v sup_u value of K.

variable {m n : ℕ} lemma helperForProposition_37_1_3_zeroLowerConjugateValue (K : SaddleFunction m n) : -theorem37ValueSupInf K 0 0 = iInf (fun x : Fin n → ℝ => iSup (fun u : Fin m → ℝ => K u x))

Lean source

Helper for Proposition 37.1.3: evaluating the upper conjugate at the origin recovers the plain sup_u inf_v value of K.

variable {m n : ℕ} lemma helperForProposition_37_1_3_zeroUpperConjugateValue (K : SaddleFunction m n) : -theorem37ValueInfSup K 0 0 = iSup (fun u : Fin m → ℝ => iInf (fun x : Fin n → ℝ => K u x))

Lean source

Helper for Proposition 37.1.3: rewriting the lower zero-tilt identity isolates the origin conjugate value itself as the negative minimax value.

variable {m n : ℕ} lemma helperForProposition_37_1_3_zeroLowerConjugate_eq_neg_minimaxValue (K : SaddleFunction m n) : theorem37ValueSupInf K 0 0 = - minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K

Lean source

Helper for Proposition 37.1.3: rewriting the upper zero-tilt identity isolates the origin conjugate value itself as the negative maximin value.

variable {m n : ℕ} lemma helperForProposition_37_1_3_zeroUpperConjugate_eq_neg_maximinValue (K : SaddleFunction m n) : theorem37ValueInfSup K 0 0 = - maximinValue (C := Fin m → ℝ) (D := Fin n → ℝ) K

Lean source

Helper for Proposition 37.1.3: Corollary 37.1.2 already provides a common product effective domain for the lower and upper conjugates.

variable {m n : ℕ} lemma helperForProposition_37_1_3_commonSaddleEffectiveDomain (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) : ∃ CStar : Set (Fin m → ℝ), ∃ DStar : Set (Fin n → ℝ), saddleEffectiveDomain (fun uStar x => theorem37ValueSupInf K uStar x) = CStar ×ˢ DStar ∧ saddleEffectiveDomain (fun uStar x => theorem37ValueInfSup K uStar x) = CStar ×ˢ DStar

Lean source

Helper for Proposition 37.1.3: the lower and upper conjugates therefore share the same effective domain.

variable {m n : ℕ} lemma helperForProposition_37_1_3_lowerUpperSaddleEffectiveDomain_eq (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) : saddleEffectiveDomain (fun uStar x => theorem37ValueSupInf K uStar x) = saddleEffectiveDomain (fun uStar x => theorem37ValueInfSup K uStar x)

Lean source

Helper for Proposition 37.1.3: the two zero-tilt conjugate evaluations already package the textbook minimax and maximin identities.

variable {m n : ℕ} lemma helperForProposition_37_1_3_originValueEqualities (K : SaddleFunction m n) : minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K = -theorem37ValueSupInf K 0 0 ∧ maximinValue (C := Fin m → ℝ) (D := Fin n → ℝ) K = -theorem37ValueInfSup K 0 0

§37 — Conjugate Saddle-Functions and Minimax Theorems — Part 3

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Corollary 37.1.3: if the origin lies in the relative interior of either common effective domain of the two Section 37 conjugates, then the minimax and maximin values of K coincide; if it lies in both, then this common saddle-value is finite.

variable {m n : ℕ} theorem corollary37_1_3_origin_relativeInterior_yields_minimax (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) : let KLowerStar : SaddleFunction m n := fun uStar x => theorem37ValueSupInf K uStar x let CStar : Set (Fin m → ℝ) := effectiveDomain₁ KLowerStar let DStar : Set (Fin n → ℝ) := effectiveDomain₂ KLowerStar (((0 : Fin m → ℝ) ∈ intrinsicInterior ℝ CStar) ∨ ((0 : Fin n → ℝ) ∈ intrinsicInterior ℝ DStar)) → minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K = maximinValue (C := Fin m → ℝ) (D := Fin n → ℝ) K ∧ ((((0 : Fin m → ℝ) ∈ intrinsicInterior ℝ CStar) ∧ ((0 : Fin n → ℝ) ∈ intrinsicInterior ℝ DStar)) → minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K ≠ (⊤ : EReal) ∧ minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K ≠ (⊥ : EReal))

Lean source

Helper for Corollary 37.1.3: at zero tilt, the upper Section 37 conjugate is exactly the negative of the textbook minimax value.

variable {m n : ℕ} lemma helperForCorollary_37_1_3_zeroTilt_upperConjugate_eq_neg_minimaxValue (K : SaddleFunction m n) : theorem37ValueSupInf K 0 0 = - minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K

Lean source

Helper for Corollary 37.1.3: at zero tilt, the lower Section 37 conjugate is exactly the negative of the textbook maximin value.

variable {m n : ℕ} lemma helperForCorollary_37_1_3_zeroTilt_lowerConjugate_eq_neg_maximinValue (K : SaddleFunction m n) : theorem37ValueInfSup K 0 0 = - maximinValue (C := Fin m → ℝ) (D := Fin n → ℝ) K

Lean source

Helper for Corollary 37.1.3: equality of the two zero-tilt Section 37 conjugates already forces the textbook minimax identity.

variable {m n : ℕ} lemma helperForCorollary_37_1_3_originConjugateAgreement_implies_minimax (K : SaddleFunction m n) (hEq0 : theorem37ValueSupInf K 0 0 = theorem37ValueInfSup K 0 0) : minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K = maximinValue (C := Fin m → ℝ) (D := Fin n → ℝ) K

Lean source

Helper for Corollary 37.1.3: Corollary 37.1.2 identifies a common relative-interior point of the two zero-tilt effective domains with equality of the two conjugate values at the origin.

variable {m n : ℕ} lemma helperForCorollary_37_1_3_zeroConjugates_agree_of_originInCommonRelativeInterior (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) (hOrigin : (0 : Fin m → ℝ) ∈ intrinsicInterior ℝ (effectiveDomain₁ (fun uStar x => theorem37ValueSupInf K uStar x)) ∨ (0 : Fin n → ℝ) ∈ intrinsicInterior ℝ (effectiveDomain₂ (fun uStar x => theorem37ValueSupInf K uStar x))) : theorem37ValueSupInf K 0 0 = theorem37ValueInfSup K 0 0

Lean source

Helper for Corollary 37.1.3: origin membership in the two coordinate effective domains is exactly the pair of strict lower and upper bounds on the zero-tilt saddle value.

variable {m n : ℕ} lemma helperForCorollary_37_1_3_originEffectiveDomain_membership_gives_zeroTilt_strictBounds (K : SaddleFunction m n) (hMem1 : (0 : Fin m → ℝ) ∈ effectiveDomain₁ (fun uStar x => theorem37ValueSupInf K uStar x)) (hMem2 : (0 : Fin n → ℝ) ∈ effectiveDomain₂ (fun uStar x => theorem37ValueSupInf K uStar x)) : (⊥ : EReal) < theorem37ValueSupInf K 0 0 ∧ theorem37ValueSupInf K 0 0 < (⊤ : EReal)

Lean source

Helper for Corollary 37.1.3: once the zero-tilt saddle value is squeezed strictly between ⊥ and ⊤, it is finite in the extended-real sense.

variable {m n : ℕ} lemma helperForCorollary_37_1_3_zeroTilt_strictBounds_imply_finiteness (K : SaddleFunction m n) (hBounds : (⊥ : EReal) < theorem37ValueSupInf K 0 0 ∧ theorem37ValueSupInf K 0 0 < (⊤ : EReal)) : theorem37ValueSupInf K 0 0 ≠ (⊤ : EReal) ∧ theorem37ValueSupInf K 0 0 ≠ (⊥ : EReal)

Lean source

Helper for Corollary 37.1.3: if the origin lies in both common relative interiors, then the zero-tilt saddle value avoids both ⊤ and ⊥.

variable {m n : ℕ} lemma helperForCorollary_37_1_3_zeroSaddleValue_finite_of_originInBothRelativeInteriors (K : SaddleFunction m n) (hOrigin1 : (0 : Fin m → ℝ) ∈ intrinsicInterior ℝ (effectiveDomain₁ (fun uStar x => theorem37ValueSupInf K uStar x))) (hOrigin2 : (0 : Fin n → ℝ) ∈ intrinsicInterior ℝ (effectiveDomain₂ (fun uStar x => theorem37ValueSupInf K uStar x))) : theorem37ValueSupInf K 0 0 ≠ (⊤ : EReal) ∧ theorem37ValueSupInf K 0 0 ≠ (⊥ : EReal)

Lean source

Helper for Corollary 37.1.3: finiteness of the zero-tilt upper conjugate transports to the textbook minimax value because Proposition 37.1.3 identifies that conjugate with its negation.

variable {m n : ℕ} lemma helperForCorollary_37_1_3_zeroTiltUpperConjugate_finite_implies_minimaxFinite (K : SaddleFunction m n) (hFiniteZero : theorem37ValueSupInf K 0 0 ≠ (⊤ : EReal) ∧ theorem37ValueSupInf K 0 0 ≠ (⊥ : EReal)) : minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K ≠ (⊤ : EReal) ∧ minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K ≠ (⊥ : EReal)

§37 — Conjugate Saddle-Functions and Minimax Theorems — Part 4

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Helper for Theorem 37.2: Corollary 37.1.2 already packages a common pair of coordinate effective domains for the two Section 37 conjugates.

variable {m n : ℕ} lemma helperForTheorem_37_2_commonEffectiveDomains (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) : ∃ CStar : Set (Fin m → ℝ), ∃ DStar : Set (Fin n → ℝ), CStar.Nonempty ∧ DStar.Nonempty ∧ Convex ℝ CStar ∧ Convex ℝ DStar ∧ effectiveDomain₁ (fun uStar x => theorem37ValueSupInf K uStar x) = CStar ∧ effectiveDomain₂ (fun uStar x => theorem37ValueSupInf K uStar x) = DStar ∧ effectiveDomain₁ (fun uStar x => theorem37ValueInfSup K uStar x) = CStar ∧ effectiveDomain₂ (fun uStar x => theorem37ValueInfSup K uStar x) = DStar

Lean source

Helper for Theorem 37.2: the common effective domains can be chosen canonically as the two coordinate effective domains of the lower Section 37 conjugate.

variable {m n : ℕ} lemma helperForTheorem_37_2_canonicalCommonEffectiveDomains (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) : let CStar : Set (Fin m → ℝ) := effectiveDomain₁ (fun uStar x => theorem37ValueSupInf K uStar x) let DStar : Set (Fin n → ℝ) := effectiveDomain₂ (fun uStar x => theorem37ValueSupInf K uStar x) CStar.Nonempty ∧ DStar.Nonempty ∧ Convex ℝ CStar ∧ Convex ℝ DStar ∧ effectiveDomain₁ (fun uStar x => theorem37ValueInfSup K uStar x) = CStar ∧ effectiveDomain₂ (fun uStar x => theorem37ValueInfSup K uStar x) = DStar

Lean source

Helper for Theorem 37.2: Jensen convexity on Set.univ upgrades directly to convexity of the epigraph, hence to ConvexFunctionOn Set.univ.

variable {m n : ℕ} lemma helperForTheorem_37_2_convexFunctionOn_univ_of_IsERealConvexOn {k : ℕ} {f : (Fin k → ℝ) → EReal} (hf : IsERealConvexOn (Set.univ : Set (Fin k → ℝ)) f) : ConvexFunctionOn (Set.univ : Set (Fin k → ℝ)) f

Lean source

Helper for Theorem 37.2: properness of the saddle function already forces both primal effective coordinate domains to be nonempty.

variable {m n : ℕ} lemma helperForTheorem_37_2_effectiveDomains_nonempty (K : SaddleFunction m n) (hKproper : IsProperSaddleFunction K) : (effectiveDomain₁ K).Nonempty ∧ (effectiveDomain₂ K).Nonempty

Lean source

Helper for Theorem 37.2: the original effective domains C and D of a closed proper concave-convex saddle-function have nonempty relative interiors.

variable {m n : ℕ} lemma helperForTheorem_37_2_intrinsicInterior_nonempty (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) : let C : Set (Fin m → ℝ) := effectiveDomain₁ K let D : Set (Fin n → ℝ) := effectiveDomain₂ K (intrinsicInterior ℝ C).Nonempty ∧ (intrinsicInterior ℝ D).Nonempty

Lean source

Helper for Theorem 37.2: every u in ri C yields a closed proper convex slice K(u, ·) with effective domain exactly D.

variable {m n : ℕ} lemma helperForTheorem_37_2_convexSlice_on_intrinsicInterior (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) {u : Fin m → ℝ} (hu : u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)) : IsProperClosedConvexFunctionWithDomain (K u) (effectiveDomain₂ K)

Lean source

Helper for Theorem 37.2: once u ∈ ri C, the unrestricted effective domain of the convex slice K(u, ·) is exactly the second coordinate domain D.

variable {m n : ℕ} lemma helperForTheorem_37_2_convexSlice_effectiveDomain_eq (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) {u : Fin m → ℝ} (hu : u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)) : effectiveDomain (Set.univ : Set (Fin n → ℝ)) (K u) = effectiveDomain₂ K

Lean source

Helper for Theorem 37.2: every interior convex slice K(u, ·) has a finite point, so its ambient-space effective domain is nonempty.

variable {m n : ℕ} lemma helperForTheorem_37_2_convexSlice_effectiveDomain_nonempty (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) {u : Fin m → ℝ} (hu : u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)) : (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (K u)).Nonempty

§37 — Conjugate Saddle-Functions and Minimax Theorems — Part 5

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Boundary-generation qualification for Theorem 37.2: after choosing any closed convex representative of K, the support supremum over all first-coordinate slices is already generated by slices whose first coordinate lies in the relative interior of the common primal domain.

variable {m n : ℕ} def Section37SecondDualBoundaryGeneration (K : SaddleFunction m n) : Prop := ∀ (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F), K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩ → ∀ w : Fin n → ℝ, iSup (fun u : Fin m → ℝ => supportFunctionEReal (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (F u)) w) = iSup (fun u : {u // u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)} => supportFunctionEReal (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (F u.1)) w)

Lean source

Helper for Theorem 37.2: negating a concave section produces the convex slice needed for the second support-function formula.

variable {m n : ℕ} lemma helperForTheorem_37_2_negatedFirstSlice_on_intrinsicInterior (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) {v : Fin n → ℝ} (hv : v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)) : IsProperClosedConvexFunctionWithDomain (fun u => -K u v) (effectiveDomain₁ K)

Lean source

Helper for Theorem 37.2: a proper saddle-function has a point in its first effective domain.

variable {m n : ℕ} lemma helperForTheorem_37_2_firstDomain_nonempty (K : SaddleFunction m n) (hKproper : IsProperSaddleFunction K) : (effectiveDomain₁ K).Nonempty

Lean source

Helper for Theorem 37.2: a proper saddle-function has a point in its second effective domain.

variable {m n : ℕ} lemma helperForTheorem_37_2_secondDomain_nonempty (K : SaddleFunction m n) (hKproper : IsProperSaddleFunction K) : (effectiveDomain₂ K).Nonempty

Lean source

Helper for Theorem 37.2: every relative-interior slice K(u, ·) is proper as a convex function on the full ambient space.

variable {m n : ℕ} lemma helperForTheorem_37_2_convexSlice_properOn_univ (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) {u : Fin m → ℝ} (hu : u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)) : ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (K u)

Lean source

Helper for Theorem 37.2: each interior slice-conjugate domain is nonempty.

variable {m n : ℕ} lemma helperForTheorem_37_2_sliceConjugateDomain_nonempty (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) {u : Fin m → ℝ} (hu : u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)) : (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (fenchelConjugate n (K u))).Nonempty

Lean source

Helper for Theorem 37.2: each interior slice-conjugate domain is convex.

variable {m n : ℕ} lemma helperForTheorem_37_2_sliceConjugateDomain_convex (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) {u : Fin m → ℝ} (hu : u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)) : Convex ℝ (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (fenchelConjugate n (K u)))

Lean source

Helper for Theorem 37.2: every relative-interior first slice u ↦ -K(u, v) is proper as a convex function on the full ambient space.

variable {m n : ℕ} lemma helperForTheorem_37_2_negatedFirstSlice_properOn_univ (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) {v : Fin n → ℝ} (hv : v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)) : ProperConvexFunctionOn (Set.univ : Set (Fin m → ℝ)) (fun u => -K u v)

Lean source

Helper for Theorem 37.2: each sign-twisted first-slice conjugate domain is nonempty.

variable {m n : ℕ} lemma helperForTheorem_37_2_negatedFirstSliceConjugateDomain_nonempty (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) {v : Fin n → ℝ} (hv : v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)) : (effectiveDomain (Set.univ : Set (Fin m → ℝ)) (fenchelConjugate m (fun u => -K u v))).Nonempty

Lean source

Helper for Theorem 37.2: each sign-twisted first-slice conjugate domain is convex.

variable {m n : ℕ} lemma helperForTheorem_37_2_negatedFirstSliceConjugateDomain_convex (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) {v : Fin n → ℝ} (hv : v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)) : Convex ℝ (effectiveDomain (Set.univ : Set (Fin m → ℝ)) (fenchelConjugate m (fun u => -K u v)))

Lean source

Helper for Theorem 37.2: any point where the representative slice F u is finite already lies in the common second effective domain of the upper Section 37 conjugate.

variable {m n : ℕ} lemma helperForTheorem_37_2_representativeSliceDomain_subset_upperSecondDomain (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (K : SaddleFunction m n) (hK : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) (hFStar : IsClosedConvexBifunction (bifunctionInverse (section34ConcaveBifunctionAdjoint F))) (hGlobal : Section34Theorem34_2GlobalQualification m n) {u : Fin m → ℝ} : effectiveDomain (Set.univ : Set (Fin n → ℝ)) (F u) ⊆ effectiveDomain₂ (fun uStar x => theorem37ValueInfSup K uStar x)

Lean source

Helper for Theorem 37.2: the first effective domain of the canonical Section 34 kernel is just the ordinary finite-parameter domain of the underlying convex bifunction.

variable {m n : ℕ} lemma helperForTheorem_37_2_firstKernelDomain_eq_convexBifunctionEffectiveDomain (G : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : effectiveDomain₁ (convexBifunctionClosedKernel G) = convexBifunctionEffectiveDomain G

Lean source

Helper for Theorem 37.2: the second effective domain of the canonical Section 34 kernel is the ordinary effective domain of the Section 34 adjoint bifunction.

variable {m n : ℕ} lemma helperForTheorem_37_2_secondKernelDomain_eq_concaveAdjointEffectiveDomain (G : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hG : IsClosedConvexBifunction G) (hQ : Section34Theorem34_2Qualification G) : effectiveDomain₂ (convexBifunctionClosedKernel G) = concaveBifunctionEffectiveDomain (section34ConcaveBifunctionAdjoint G)

Lean source

Helper for Theorem 37.2: the recovered dual inverse has adjoint equal to the inverse of the original representative.

variable {m n : ℕ} lemma helperForTheorem_37_2_dualInverseAdjoint_eq_originalInverse (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (hFStar : IsClosedConvexBifunction (bifunctionInverse (section34ConcaveBifunctionAdjoint F))) (hGlobal : Section34Theorem34_2GlobalQualification m n) : section34ConcaveBifunctionAdjoint (bifunctionInverse (section34ConcaveBifunctionAdjoint F)) = bifunctionInverse F

Lean source

Helper for Theorem 37.2: precomposing a function with negation transports the effective domain by the same negation map.

variable {m n : ℕ} lemma helperForTheorem_37_2_negPrecompose_effectiveDomain (h : (Fin m → ℝ) → EReal) : effectiveDomain (Set.univ : Set (Fin m → ℝ)) (fun uStar => h (-uStar)) = Neg.neg ⁻¹' effectiveDomain (Set.univ : Set (Fin m → ℝ)) h

Lean source

Helper for Theorem 37.2: negating the underlying set transports the support function by negating its argument.

variable {m n : ℕ} lemma helperForTheorem_37_2_supportFunctionEReal_negPreimage {k : ℕ} (C : Set (Fin k → ℝ)) (z : Fin k → ℝ) : supportFunctionEReal (Neg.neg ⁻¹' C) z = supportFunctionEReal C (-z)

Lean source

Helper for Theorem 37.2: support functions are monotone under inclusion of the underlying sets.

variable {m n : ℕ} lemma helperForTheorem_37_2_supportFunctionEReal_mono_of_subset {k : ℕ} {A B : Set (Fin k → ℝ)} (hAB : A ⊆ B) (y : Fin k → ℝ) : supportFunctionEReal A y ≤ supportFunctionEReal B y

Lean source

Helper for Theorem 37.2: each interior slice-conjugate domain is contained in the common second dual domain D*.

variable {m n : ℕ} lemma helperForTheorem_37_2_sliceConjugateDomain_subset_commonSecondDomain (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) {u : Fin m → ℝ} (_hu : u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)) : let DStar : Set (Fin n → ℝ) := effectiveDomain₂ (fun uStar x => theorem37ValueSupInf K uStar x) effectiveDomain (Set.univ : Set (Fin n → ℝ)) (fenchelConjugate n (K u)) ⊆ DStar

Lean source

Helper for Theorem 37.2: every relative-interior second slice has a nonempty convex Fenchel-domain.

variable {m n : ℕ} lemma helperForTheorem_37_2_sliceConjugateDomain_convex_nonempty (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) {u : Fin m → ℝ} (hu : u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)) : Convex ℝ (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (fenchelConjugate n (K u))) ∧ (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (fenchelConjugate n (K u))).Nonempty

Lean source

Helper for Theorem 37.2: every relative-interior first slice, after the sign flip used in the textbook formula, has a nonempty convex Fenchel-domain.

variable {m n : ℕ} lemma helperForTheorem_37_2_negatedSliceConjugateDomain_convex_nonempty (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) {v : Fin n → ℝ} (hv : v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)) : Convex ℝ (effectiveDomain (Set.univ : Set (Fin m → ℝ)) (fenchelConjugate m (fun u => -K u v))) ∧ (effectiveDomain (Set.univ : Set (Fin m → ℝ)) (fenchelConjugate m (fun u => -K u v))).Nonempty

Lean source

Helper for Theorem 37.2: every relative-interior second slice contributes a lower bound for the support of D*.

variable {m n : ℕ} lemma helperForTheorem_37_2_fixedSliceSup_le_secondDualSupport (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) {u : Fin m → ℝ} (hu : u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)) (w : Fin n → ℝ) : let D : Set (Fin n → ℝ) := effectiveDomain₂ K let DStar : Set (Fin n → ℝ) := effectiveDomain₂ (fun uStar x => theorem37ValueSupInf K uStar x) (⨆ v : {v // v ∈ D}, K u (v.1 + w) - K u v.1) ≤ supportFunctionEReal DStar w

Lean source

Helper for Theorem 37.2: the full relative-interior supremum is always bounded above by the support of D*.

variable {m n : ℕ} lemma helperForTheorem_37_2_iSupInteriorSlices_le_secondDualSupport (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) (w : Fin n → ℝ) : iSup (fun u : {u // u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)} => iSup (fun v : {v // v ∈ effectiveDomain₂ K} => K u.1 (v.1 + w) - K u.1 v.1)) ≤ supportFunctionEReal (effectiveDomain₂ (fun uStar x => theorem37ValueSupInf K uStar x)) w

Lean source

Helper for Theorem 37.2: for any closed convex representative F of K, the common second dual domain D* is exactly the union of the effective domains of the slices F u.

variable {m n : ℕ} lemma helperForTheorem_37_2_secondDualDomain_eq_iUnion_representativeSliceDomains (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (K : SaddleFunction m n) (hK : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) (hGlobal : Section34Theorem34_2GlobalQualification m n) : effectiveDomain₂ (fun uStar x => theorem37ValueSupInf K uStar x) = ⋃ u : Fin m → ℝ, effectiveDomain (Set.univ : Set (Fin n → ℝ)) (F u)

Lean source

Helper for Theorem 37.2: the support function of a union is the supremum of the support functions of the members.

variable {m n : ℕ} lemma helperForTheorem_37_2_supportFunctionEReal_iUnion (C : (Fin m → ℝ) → Set (Fin n → ℝ)) (w : Fin n → ℝ) : supportFunctionEReal (⋃ u : Fin m → ℝ, C u) w = iSup (fun u : Fin m → ℝ => supportFunctionEReal (C u) w)

Lean source

Helper for Theorem 37.2: the support of D* is the supremum of the support functions of the representative slice domains.

variable {m n : ℕ} lemma helperForTheorem_37_2_secondDualSupport_eq_iSupRepresentativeSliceSupports (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (K : SaddleFunction m n) (hK : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) (hGlobal : Section34Theorem34_2GlobalQualification m n) (w : Fin n → ℝ) : supportFunctionEReal (effectiveDomain₂ (fun uStar x => theorem37ValueSupInf K uStar x)) w = iSup (fun u : Fin m → ℝ => supportFunctionEReal (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (F u)) w)

Lean source

Helper for Theorem 37.2: on relative-interior first slices, the support of the recovered representative domain is exactly the textbook supremum over v ∈ D.

variable {m n : ℕ} lemma helperForTheorem_37_2_representativeInteriorSliceSupport_eq_primalSliceSup (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) (hF : IsClosedConvexBifunction F) (K : SaddleFunction m n) (hK : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) {u : Fin m → ℝ} (hu : u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)) (w : Fin n → ℝ) : supportFunctionEReal (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (F u)) w = iSup (fun v : {v // v ∈ effectiveDomain₂ K} => K u (v.1 + w) - K u v.1)

Lean source

Helper for Theorem 37.2: under explicit boundary generation, the support function of the second common dual domain is the supremum of the second-family recession expressions over the relative interior of C.

variable {m n : ℕ} lemma helperForTheorem_37_2_secondDualSupport_formula (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) (hBoundary : Section37SecondDualBoundaryGeneration K) (w : Fin n → ℝ) : supportFunctionEReal (effectiveDomain₂ (fun uStar x => theorem37ValueSupInf K uStar x)) w = iSup (fun u : {u // u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)} => iSup (fun v : {v // v ∈ effectiveDomain₂ K} => K u.1 (v.1 + w) - K u.1 v.1))

§37 — Conjugate Saddle-Functions and Minimax Theorems — Part 6

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Explicit regularity data for Theorem 37.2. The no-⊤ field is exactly what makes the swap-negation a concave-convex saddle function. The two Section 34 and boundary-generation fields support the second-dual formula for K and, separately, for its inverse.

variable {m n : ℕ} structure Section37Theorem37_2Qualification (K : SaddleFunction m n) : Prop where noTop : HasNoTopValuesBifunction K primalGlobal : Section34Theorem34_2GlobalQualification m n primalBoundary : Section37SecondDualBoundaryGeneration K inverseGlobal : Section34Theorem34_2GlobalQualification n m inverseBoundary : Section37SecondDualBoundaryGeneration (bifunctionInverse K)

Lean source

variable {m n : ℕ} theorem section37_theorem37_2 (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) : let C : Set (Fin m → ℝ) := effectiveDomain₁ K let D : Set (Fin n → ℝ) := effectiveDomain₂ K let CStar : Set (Fin m → ℝ) := effectiveDomain₁ (fun uStar x => theorem37ValueSupInf K uStar x) let DStar : Set (Fin n → ℝ) := effectiveDomain₂ (fun uStar x => theorem37ValueSupInf K uStar x) ∀ w z, (supportFunctionEReal DStar w = iSup (fun u : {u // u ∈ intrinsicInterior ℝ C} => iSup (fun v : {v // v ∈ D} => K u.1 (v.1 + w) - K u.1 v.1))) ∧ (-supportFunctionEReal CStar (-z) = iInf (fun v : {v // v ∈ intrinsicInterior ℝ D} => iInf (fun u : {u // u ∈ C} => K (u.1 + z) v.1 - K u.1 v.1)))

Lean source

Corollary 37.2.1: the origin lies in the interior of D* exactly when the convex slices K(u, ·) for u ∈ ri C have no common recession direction, and similarly for C* and the convex slices -K(·, v) for v ∈ ri D.

variable {m n : ℕ} theorem corollary37_2_1_origin_interior_iff_no_common_recession_direction (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) : let C : Set (Fin m → ℝ) := effectiveDomain₁ K let D : Set (Fin n → ℝ) := effectiveDomain₂ K let CStar : Set (Fin m → ℝ) := effectiveDomain₁ (fun uStar x => theorem37ValueSupInf K uStar x) let DStar : Set (Fin n → ℝ) := effectiveDomain₂ (fun uStar x => theorem37ValueSupInf K uStar x) ((0 : Fin n → ℝ) ∈ interior DStar ↔ ∀ w : Fin n → ℝ, w ≠ 0 → ∃ u : {u // u ∈ intrinsicInterior ℝ C}, ¬ IsRecessionDirection (K u.1) w) ∧ ((0 : Fin m → ℝ) ∈ interior CStar ↔ ∀ z : Fin m → ℝ, z ≠ 0 → ∃ v : {v // v ∈ intrinsicInterior ℝ D}, ¬ IsRecessionDirection (fun u => -K u v.1) z)

Lean source

Helper for Theorem 37.2: a concave EReal section becomes convex after negation once the original section is known to avoid ⊤.

variable {m n : ℕ} lemma helperForTheorem_37_2_concaveNegation_isConvex_of_noTop {k : ℕ} {f : (Fin k → ℝ) → EReal} (hConc : IsERealConcaveOn (Set.univ : Set (Fin k → ℝ)) f) (hNoTop : ∀ x, f x ≠ ⊤) : IsERealConvexOn (Set.univ : Set (Fin k → ℝ)) (fun x => -f x)

Lean source

Helper for Corollary 37.2.1: the raw outside-domain inverse model u ↦ -K(u,v) need not be globally convex, so the remaining inverse-slice branch cannot be closed by a blanket negation transport.

variable {m n : ℕ} lemma helperForCorollary_37_2_1_not_convex_of_bot_top_top_sample {f : (Fin 1 → ℝ) → EReal} {x y z : Fin 1 → ℝ} (hx : f x = (⊥ : EReal)) (hy : f y = (⊤ : EReal)) (hz : f z = (⊤ : EReal)) (hzCombo : ((1 / 3 : ℝ) • x + (2 / 3 : ℝ) • y) = z) : ¬ IsERealConvexOn (Set.univ : Set (Fin 1 → ℝ)) f

Lean source

Helper for Corollary 37.2.1: the raw outside-domain inverse model u ↦ -K(u,v) need not be globally convex, so the remaining inverse-slice branch cannot be closed by a blanket negation transport.

variable {m n : ℕ} lemma helperForCorollary_37_2_1_halfspace_bot_top_not_convex : ¬ IsERealConvexOn (Set.univ : Set (Fin 1 → ℝ)) (fun u : Fin 1 → ℝ => if 0 ≤ u 0 then (⊥ : EReal) else (⊤ : EReal))

Lean source

Helper for Theorem 37.2: swapping variables and negating preserves the concave-convex orientation when the original saddle function avoids ⊤.

variable {m n : ℕ} lemma helperForTheorem_37_2_inverseConcaveConvex (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (_hKproper : IsProperSaddleFunction K) (hKNoTop : HasNoTopValuesBifunction K) : IsConcaveConvex (bifunctionInverse K)

Lean source

Helper for Theorem 37.2: the first partial closure of the inverse saddle-function is the inverse of the second partial closure of the original saddle-function.

variable {m n : ℕ} lemma helperForTheorem_37_2_inversePartialClosureFirst_eq_inversePartialClosureSecond (K : SaddleFunction m n) : partialClosure₁ (bifunctionInverse K) = bifunctionInverse (partialClosure₂ K)

Lean source

Helper for Theorem 37.2: the second partial closure of the inverse saddle-function is the inverse of the first partial closure of the original saddle-function.

variable {m n : ℕ} lemma helperForTheorem_37_2_inversePartialClosureSecond_eq_inversePartialClosureFirst (K : SaddleFunction m n) : partialClosure₂ (bifunctionInverse K) = bifunctionInverse (partialClosure₁ K)

Lean source

Helper for Theorem 37.2: closedness is preserved by the inverse saddle transform.

variable {m n : ℕ} lemma helperForTheorem_37_2_inverseClosed (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hKNoTop : HasNoTopValuesBifunction K) : IsClosedSaddleFunction (bifunctionInverse K)

Lean source

Helper for Theorem 37.2: saddle-equivalence is stable under the inverse saddle transform.

variable {m n : ℕ} lemma helperForTheorem_37_2_inverseSaddleEquivalent {K L : SaddleFunction m n} (hKL : saddleEquivalent K L) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hKNoTop : HasNoTopValuesBifunction K) (hLclosed : IsClosedSaddleFunction L) (hLproper : IsProperSaddleFunction L) (hLNoTop : HasNoTopValuesBifunction L) : saddleEquivalent (bifunctionInverse K) (bifunctionInverse L)

Lean source

Helper for Theorem 37.2: in the inverse saddle function, the first effective domain is exactly the second effective domain of the original saddle function.

variable {m n : ℕ} lemma helperForTheorem_37_2_inverseFirstDomain_eq_originalSecondDomain (K : SaddleFunction m n) : effectiveDomain₁ (bifunctionInverse K) = effectiveDomain₂ K

Lean source

Helper for Theorem 37.2: in the inverse saddle function, the second effective domain is exactly the first effective domain of the original saddle function.

variable {m n : ℕ} lemma helperForTheorem_37_2_inverseSecondDomain_eq_originalFirstDomain (K : SaddleFunction m n) : effectiveDomain₂ (bifunctionInverse K) = effectiveDomain₁ K

Lean source

Helper for Theorem 37.2: properness is preserved when passing to the inverse saddle function.

variable {m n : ℕ} lemma helperForTheorem_37_2_inverseProper (K : SaddleFunction m n) (hKproper : IsProperSaddleFunction K) : IsProperSaddleFunction (bifunctionInverse K)

Lean source

Helper for Theorem 37.2: the lower Section 37 conjugate of the inverse saddle-function is the negative of the upper Section 37 conjugate of K after negating both dual variables.

variable {m n : ℕ} lemma helperForTheorem_37_2_inverseLowerConjugate_pointwiseRewrite (K : SaddleFunction m n) (uStar : Fin n → ℝ) (x : Fin m → ℝ) : theorem37ValueSupInf (bifunctionInverse K) uStar x = -theorem37ValueInfSup K (-x) (-uStar)

Lean source

Helper for Theorem 37.2: the slice supremum appearing in the second-variable formula for the inverse saddle function is exactly the negative of the first-variable iInf expression from the textbook formula for K.

variable {m n : ℕ} lemma helperForTheorem_37_2_inverseSliceSup_eq_negatedFirstDualInf (K : SaddleFunction m n) (z : Fin m → ℝ) : iSup (fun v : {v // v ∈ intrinsicInterior ℝ (effectiveDomain₁ (bifunctionInverse K))} => iSup (fun u : {u // u ∈ effectiveDomain₂ (bifunctionInverse K)} => bifunctionInverse K v.1 (u.1 + z) - bifunctionInverse K v.1 u.1)) = -iInf (fun v : {v // v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)} => iInf (fun u : {u // u ∈ effectiveDomain₁ K} => K (u.1 + z) v.1 - K u.1 v.1))

Lean source

Helper for Theorem 37.2: the common dual second domain of the inverse saddle-function is the negated preimage of the common dual first domain of the original saddle-function.

variable {m n : ℕ} lemma helperForTheorem_37_2_inverseSecondDualDomain_eq_originalFirstDualDomain (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) : effectiveDomain₂ (fun uStar x => theorem37ValueSupInf (bifunctionInverse K) uStar x) = Neg.neg ⁻¹' effectiveDomain₁ (fun uStar x => theorem37ValueSupInf K uStar x)

Lean source

Helper for Theorem 37.2: the inverse-domain transport proves the textbook sign-correct first dual support formula with argument -z.

variable {m n : ℕ} lemma helperForTheorem_37_2_firstDualSupport_formula_signCorrected (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) (z : Fin m → ℝ) : -supportFunctionEReal (effectiveDomain₁ (fun uStar x => theorem37ValueSupInf K uStar x)) (-z) = iInf (fun v : {v // v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)} => iInf (fun u : {u // u ∈ effectiveDomain₁ K} => K (u.1 + z) v.1 - K u.1 v.1))

Lean source

Helper for Theorem 37.2: for each v ∈ ri D, the support of the Fenchel-domain of the negated first slice is exactly its recession function.

variable {m n : ℕ} lemma helperForTheorem_37_2_negatedFirstInteriorSliceSupport_eq_recession (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) {v : Fin n → ℝ} (hv : v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)) (z : Fin m → ℝ) : supportFunctionEReal (effectiveDomain (Set.univ : Set (Fin m → ℝ)) (fenchelConjugate m (fun u => -K u v))) z = recessionFunction (fun u => -K u v) z

Lean source

Helper for Theorem 37.2: for each v ∈ ri D, the first-family infimum is the negative of the recession function of the convex slice -K(·, v).

variable {m n : ℕ} lemma helperForTheorem_37_2_firstSliceInf_eq_neg_recessionFunction_preTheorem (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) (v : {v // v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)}) (z : Fin m → ℝ) : iInf (fun u : {u // u ∈ effectiveDomain₁ K} => K (u.1 + z) v.1 - K u.1 v.1) = -recessionFunction (fun u => -K u v.1) z

Lean source

Helper for Theorem 37.2: this compatibility wrapper records the sign-correct first dual support formula with support argument -z.

variable {m n : ℕ} lemma helperForTheorem_37_2_firstDualSupport_formula (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) (z : Fin m → ℝ) : -supportFunctionEReal (effectiveDomain₁ (fun uStar x => theorem37ValueSupInf K uStar x)) (-z) = iInf (fun v : {v // v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)} => iInf (fun u : {u // u ∈ effectiveDomain₁ K} => K (u.1 + z) v.1 - K u.1 v.1))

Lean source

Helper for Corollary 37.2.1: every second-family increment from Theorem 37.2 vanishes at the zero recession direction.

variable {m n : ℕ} lemma helperForCorollary_37_2_1_secondFamilyIncrementAtZero (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) (u : {u // u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)}) (v : {v // v ∈ effectiveDomain₂ K}) : K u.1 (v.1 + (0 : Fin n → ℝ)) - K u.1 v.1 = 0

Lean source

Helper for Corollary 37.2.1: every first-family increment from Theorem 37.2 vanishes at the zero recession direction.

variable {m n : ℕ} lemma helperForCorollary_37_2_1_firstFamilyIncrementAtZero (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) (v : {v // v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)}) (u : {u // u ∈ effectiveDomain₁ K}) : K (u.1 + (0 : Fin m → ℝ)) v.1 - K u.1 v.1 = 0

Lean source

Helper for Corollary 37.2.1: the second-family recession expression from Theorem 37.2 specializes to 0 at the zero direction.

variable {m n : ℕ} lemma helperForCorollary_37_2_1_secondFamilySupAtZero (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) : iSup (fun u : {u // u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)} => iSup (fun v : {v // v ∈ effectiveDomain₂ K} => K u.1 (v.1 + (0 : Fin n → ℝ)) - K u.1 v.1)) = 0

Lean source

Helper for Corollary 37.2.1: the first-family recession expression from Theorem 37.2 specializes to 0 at the zero direction.

variable {m n : ℕ} lemma helperForCorollary_37_2_1_firstFamilyInfAtZero (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) : iInf (fun v : {v // v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)} => iInf (fun u : {u // u ∈ effectiveDomain₁ K} => K (u.1 + (0 : Fin m → ℝ)) v.1 - K u.1 v.1)) = 0

Lean source

Helper for Corollary 37.2.1: for each u ∈ ri C, the second-family slice supremum from Theorem 37.2 is exactly the recession function of K(u, ·).

variable {m n : ℕ} lemma helperForCorollary_37_2_1_secondSliceSup_eq_recessionFunction (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) (u : {u // u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)}) (w : Fin n → ℝ) : iSup (fun v : {v // v ∈ effectiveDomain₂ K} => K u.1 (v.1 + w) - K u.1 v.1) = recessionFunction (K u.1) w

Lean source

Helper for Corollary 37.2.1: the support function of D* is the supremum of the recession functions of the primal slices K(u, ·) over u ∈ ri C.

variable {m n : ℕ} lemma helperForCorollary_37_2_1_secondSupport_eq_iSup_sliceRecession (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) (w : Fin n → ℝ) : supportFunctionEReal (effectiveDomain₂ (fun uStar x => theorem37ValueSupInf K uStar x)) w = iSup (fun u : {u // u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)} => recessionFunction (K u.1) w)

Lean source

Helper for Corollary 37.2.1: for each v ∈ ri D, the first-family infimum from Theorem 37.2 is the negative of the recession function of the convex slice -K(·, v).

variable {m n : ℕ} lemma helperForCorollary_37_2_1_firstSliceInf_eq_neg_recessionFunction (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) (v : {v // v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)}) (z : Fin m → ℝ) : iInf (fun u : {u // u ∈ effectiveDomain₁ K} => K (u.1 + z) v.1 - K u.1 v.1) = -recessionFunction (fun u => -K u v.1) z

Lean source

Helper for Corollary 37.2.1: the sign-correct first support identity rewrites support C* (-z) as the supremum of the recession functions of the convex slices -K(·, v) over v ∈ ri D.

variable {m n : ℕ} lemma helperForCorollary_37_2_1_firstSupport_negArg_eq_iSup_negatedFirstSliceRecession (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) (z : Fin m → ℝ) : supportFunctionEReal (effectiveDomain₁ (fun uStar x => theorem37ValueSupInf K uStar x)) (-z) = iSup (fun v : {v // v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)} => recessionFunction (fun u => -K u v.1) z)

Lean source

Helper for Corollary 37.2.1: positivity of an indexed supremum is equivalent to the existence of a positive slice.

variable {m n : ℕ} lemma helperForCorollary_37_2_1_iSup_gt_zero_iff_exists_slice_gt_zero {ι : Type*} (a : ι → EReal) : iSup a > (0 : EReal) ↔ ∃ i, a i > (0 : EReal)

§37 — Conjugate Saddle-Functions and Minimax Theorems — Part 7

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Theorem 37.3: either recession-direction condition from Corollary 37.2.1 implies existence of the saddle-value of K, and if both hold then this common value is finite.

variable {m n : ℕ} theorem section37_theorem37_3 (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) : let C : Set (Fin m → ℝ) := effectiveDomain₁ K let D : Set (Fin n → ℝ) := effectiveDomain₂ K ((∀ w : Fin n → ℝ, w ≠ 0 → ∃ u : {u // u ∈ intrinsicInterior ℝ C}, ¬ IsRecessionDirection (K u.1) w) ∨ (∀ z : Fin m → ℝ, z ≠ 0 → ∃ v : {v // v ∈ intrinsicInterior ℝ D}, ¬ IsRecessionDirection (fun u => -K u v.1) z)) → minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K = maximinValue (C := Fin m → ℝ) (D := Fin n → ℝ) K ∧ (((∀ w : Fin n → ℝ, w ≠ 0 → ∃ u : {u // u ∈ intrinsicInterior ℝ C}, ¬ IsRecessionDirection (K u.1) w) ∧ (∀ z : Fin m → ℝ, z ≠ 0 → ∃ v : {v // v ∈ intrinsicInterior ℝ D}, ¬ IsRecessionDirection (fun u => -K u v.1) z)) → minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K ≠ (⊤ : EReal) ∧ minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K ≠ (⊥ : EReal))

Lean source

Helper for Theorem 37.3: condition (a) transports the origin from the ordinary interior of D* to the intrinsic interior needed by Corollary 37.1.3.

variable {m n : ℕ} lemma helperForTheorem_37_3_origin_mem_intrinsicInterior_secondDual_of_noCommonSecondRecession (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) (hNoCommonSecond : ∀ w : Fin n → ℝ, w ≠ 0 → ∃ u : {u // u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)}, ¬ IsRecessionDirection (K u.1) w) : (0 : Fin n → ℝ) ∈ intrinsicInterior ℝ (effectiveDomain₂ (fun uStar x => theorem37ValueSupInf K uStar x))

Lean source

Helper for Theorem 37.3: condition (b) transports the origin from the ordinary interior of C* to the intrinsic interior needed by Corollary 37.1.3.

variable {m n : ℕ} lemma helperForTheorem_37_3_origin_mem_intrinsicInterior_firstDual_of_noCommonFirstRecession (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) (hNoCommonFirst : ∀ z : Fin m → ℝ, z ≠ 0 → ∃ v : {v // v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)}, ¬ IsRecessionDirection (fun u => -K u v.1) z) : (0 : Fin m → ℝ) ∈ intrinsicInterior ℝ (effectiveDomain₁ (fun uStar x => theorem37ValueSupInf K uStar x))

Lean source

Helper for Theorem 37.3: when both recession-direction conditions hold, Corollary 37.1.3 gives finiteness of the common saddle-value.

variable {m n : ℕ} lemma helperForTheorem_37_3_finiteSaddleValue_of_bothNoCommonRecessionConditions (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) (hNoCommonSecond : ∀ w : Fin n → ℝ, w ≠ 0 → ∃ u : {u // u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)}, ¬ IsRecessionDirection (K u.1) w) (hNoCommonFirst : ∀ z : Fin m → ℝ, z ≠ 0 → ∃ v : {v // v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)}, ¬ IsRecessionDirection (fun u => -K u v.1) z) : minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K ≠ (⊤ : EReal) ∧ minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K ≠ (⊥ : EReal)

§37 — Conjugate Saddle-Functions and Minimax Theorems — Part 8

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Theorem 37.4: a pair (uStar, vStar) lies in ∂K(u, v) exactly when the affine tilt K - ⟨\cdot,u^*⟩ - ⟨\cdot,v^*⟩ has (u, v) as a saddle point; for closed proper K one has ri (dom K) ⊆ dom ∂K ⊆ dom K.

variable {m n : ℕ} theorem section37_theorem37_4 (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) : (∀ u v uStar vStar, (uStar, vStar) ∈ productSubdifferentialAt K u v ↔ IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) (helperForTheorem_37_4_affineTiltKernel K uStar vStar) u v) ∧ saddleKernelDomain K ⊆ {p | Set.Nonempty (productSubdifferentialAt K p.1 p.2)} ∧ {p | Set.Nonempty (productSubdifferentialAt K p.1 p.2)} ⊆ saddleEffectiveDomain K

Lean source

Corollary 37.4.1: equivalent saddle-functions have the same product subdifferential, and their values agree on the common domain where this product subdifferential is nonempty.

variable {m n : ℕ} theorem corollary37_4_1_equivalentSaddleFunctions_have_same_productSubdifferential {K L : SaddleFunction m n} (hKL : EquivalentSaddleFunctions K L) : (∀ u v, productSubdifferentialAt K u v = productSubdifferentialAt L u v) ∧ ({p : (Fin m → ℝ) × (Fin n → ℝ) | Set.Nonempty (productSubdifferentialAt K p.1 p.2)} = {p : (Fin m → ℝ) × (Fin n → ℝ) | Set.Nonempty (productSubdifferentialAt L p.1 p.2)}) ∧ (∀ u v, Set.Nonempty (productSubdifferentialAt K u v) → K u v = L u v)

Lean source

Explicit infinite-value qualification needed by the canonical Section 34 witness route used to recover a graph-closed convex representative.

variable {m n : ℕ} def Section37ClosedRepresentativeQualification (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) : Prop := HasNoBotValuesBifunction K ∧ HasNoTopOrBotValuesBifunction (lowerClosureConcaveConvex K hKclosed.1.1)

Lean source

Helper for Theorem 37.4: this is the affine tilt K - ⟨\cdot,u^*⟩ - ⟨\cdot,v^*⟩ whose saddle points characterize productSubdifferentialAt K u v.

variable {m n : ℕ} noncomputable def helperForTheorem_37_4_affineTiltKernel (K : SaddleFunction m n) (uStar : Fin m → ℝ) (vStar : Fin n → ℝ) : SaddleFunction m n := fun u v => K u v - (((finDot u uStar : ℝ) : EReal)) - (((finDot v vStar : ℝ) : EReal))

Lean source

Helper for Theorem 37.4: the first-variable partial increment is the difference of the two corresponding dot products.

variable {m n : ℕ} lemma helperForTheorem_37_4_coe_firstPartialIncrement_eq_finDot_sub (u : Fin m → ℝ) (u' : Fin m → ℝ) (uStar : Fin m → ℝ) : (((∑ i : Fin m, uStar i * (u' i - u i) : ℝ)) : EReal) = (((finDot u' uStar - finDot u uStar : ℝ)) : EReal)

Lean source

Helper for Theorem 37.4: the second-variable partial increment is the difference of the two corresponding dot products.

variable {m n : ℕ} lemma helperForTheorem_37_4_coe_secondPartialIncrement_eq_finDot_sub (v : Fin n → ℝ) (v' : Fin n → ℝ) (vStar : Fin n → ℝ) : (((∑ i : Fin n, vStar i * (v' i - v i) : ℝ)) : EReal) = (((finDot v' vStar - finDot v vStar : ℝ)) : EReal)

Lean source

Helper for Theorem 37.4: on Fin k → ℝ, intrinsic-interior points are also points of the transported Euclidean relative interior.

variable {m n : ℕ} lemma helperForTheorem_37_4_mem_euclideanRelativeInterior_fin_of_mem_intrinsicInterior {k : ℕ} {C : Set (Fin k → ℝ)} {x : Fin k → ℝ} (hx : x ∈ intrinsicInterior ℝ C) : x ∈ euclideanRelativeInterior_fin k C

Lean source

Helper for Theorem 37.4: the EReal sum of coordinatewise affine products is the coercion of the corresponding real sum.

variable {m n : ℕ} lemma helperForTheorem_37_4_sumERealProducts_eq_coe_sum {k : ℕ} (a x y : Fin k → ℝ) : (∑ i : Fin k, (((a i : ℝ) : EReal) * ((((x i - y i : ℝ)) : EReal)))) = (((∑ i : Fin k, a i * (x i - y i) : ℝ)) : EReal)

Lean source

Helper for Theorem 37.4: the coordinatewise EReal products that occur after unfolding the partial subdifferentials sum to the same coerced real affine increment.

variable {m n : ℕ} lemma helperForTheorem_37_4_sumERealProducts_subtractedCoordinates_eq_coe_sum {k : ℕ} (a x y : Fin k → ℝ) : (∑ i : Fin k, (((a i : ℝ) : EReal) * (((x i : ℝ) : EReal) - (((y i : ℝ) : EReal))))) = (((∑ i : Fin k, a i * (x i - y i) : ℝ)) : EReal)

Lean source

Helper for Theorem 37.4: membership in the product subdifferential is exactly the saddle-point condition for the affine tilt by (uStar, vStar).

variable {m n : ℕ} lemma helperForTheorem_37_4_mem_productSubdifferential_iff_saddlePoint_affineTilt (K : SaddleFunction m n) (u : Fin m → ℝ) (v : Fin n → ℝ) (uStar : Fin m → ℝ) (vStar : Fin n → ℝ) : (uStar, vStar) ∈ productSubdifferentialAt K u v ↔ IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) (helperForTheorem_37_4_affineTiltKernel K uStar vStar) u v

Lean source

Helper for Theorem 37.4: a closed proper saddle-function has nonempty product subdifferential at every point of ri (dom K).

variable {m n : ℕ} lemma helperForTheorem_37_4_nonempty_productSubdifferential_of_mem_saddleKernelDomain (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) {u : Fin m → ℝ} {v : Fin n → ℝ} (huv : (u, v) ∈ saddleKernelDomain K) : Set.Nonempty (productSubdifferentialAt K u v)

Lean source

Helper for Theorem 37.4: for a proper saddle-function, any nonempty product subdifferential can only occur on dom K = dom₁ K × dom₂ K.

variable {m n : ℕ} lemma helperForTheorem_37_4_mem_saddleEffectiveDomain_of_nonempty_productSubdifferential (K : SaddleFunction m n) (hKproper : IsProperSaddleFunction K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hSub : Set.Nonempty (productSubdifferentialAt K u v)) : (u, v) ∈ saddleEffectiveDomain K

Lean source

Helper for Corollary 37.4.1: equivalence of saddle-functions is symmetric.

variable {m n : ℕ} lemma helperForCorollary_37_4_1_equivalentSaddleFunctions_symm {K L : SaddleFunction m n} (hKL : EquivalentSaddleFunctions K L) : EquivalentSaddleFunctions L K

Lean source

Helper for Corollary 37.4.1: a saddle point of the affine tilt can occur only at a point of the original common saddle domain.

variable {m n : ℕ} lemma helperForCorollary_37_4_1_mem_originalSaddleDom_of_isSaddlePoint_affineTilt {K L : SaddleFunction m n} (hKL : EquivalentSaddleFunctions K L) {u : Fin m → ℝ} {v : Fin n → ℝ} {uStar : Fin m → ℝ} {vStar : Fin n → ℝ} (hSaddle : IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) (helperForTheorem_37_4_affineTiltKernel K uStar vStar) u v) : u ∈ saddleDom1 K ∧ v ∈ saddleDom2 K

Lean source

Helper for Corollary 37.4.1: once the second coordinate lies in the common saddle domain, the affine tilts of equivalent saddle-functions agree on the whole corresponding row.

variable {m n : ℕ} lemma helperForCorollary_37_4_1_affineTilt_eq_on_row_of_mem_saddleDom2 {K L : SaddleFunction m n} {uStar : Fin m → ℝ} {vStar : Fin n → ℝ} (hKL : EquivalentSaddleFunctions K L) {v : Fin n → ℝ} (hv : v ∈ saddleDom2 K) : ∀ u : Fin m → ℝ, helperForTheorem_37_4_affineTiltKernel K uStar vStar u v = helperForTheorem_37_4_affineTiltKernel L uStar vStar u v

Lean source

Helper for Corollary 37.4.1: once the first coordinate lies in the common saddle domain, the affine tilts of equivalent saddle-functions agree on the whole corresponding column.

variable {m n : ℕ} lemma helperForCorollary_37_4_1_affineTilt_eq_on_col_of_mem_saddleDom1 {K L : SaddleFunction m n} {uStar : Fin m → ℝ} {vStar : Fin n → ℝ} (hKL : EquivalentSaddleFunctions K L) {u : Fin m → ℝ} (hu : u ∈ saddleDom1 K) : ∀ v : Fin n → ℝ, helperForTheorem_37_4_affineTiltKernel K uStar vStar u v = helperForTheorem_37_4_affineTiltKernel L uStar vStar u v

Lean source

Helper for Corollary 37.4.1: if one affine tilt has a saddle point at (u,v), the corresponding affine tilt of an equivalent saddle-function has the same saddle point.

variable {m n : ℕ} lemma helperForCorollary_37_4_1_saddlePoint_affineTilt_transport {K L : SaddleFunction m n} (hKL : EquivalentSaddleFunctions K L) {u : Fin m → ℝ} {v : Fin n → ℝ} {uStar : Fin m → ℝ} {vStar : Fin n → ℝ} (hSaddle : IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) (helperForTheorem_37_4_affineTiltKernel K uStar vStar) u v) : IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) (helperForTheorem_37_4_affineTiltKernel L uStar vStar) u v

Lean source

Helper for Corollary 37.4.1: the affine-tilt saddle-point predicate is identical for equivalent saddle-functions.

variable {m n : ℕ} lemma helperForCorollary_37_4_1_saddlePoint_affineTilt_iff_of_equivalentSaddleFunctions {K L : SaddleFunction m n} (hKL : EquivalentSaddleFunctions K L) (u : Fin m → ℝ) (v : Fin n → ℝ) (uStar : Fin m → ℝ) (vStar : Fin n → ℝ) : IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) (helperForTheorem_37_4_affineTiltKernel K uStar vStar) u v ↔ IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) (helperForTheorem_37_4_affineTiltKernel L uStar vStar) u v

Lean source

Helper for Corollary 37.4.1: equivalent saddle-functions have the same product subdifferential at every point.

variable {m n : ℕ} lemma helperForCorollary_37_4_1_productSubdifferential_eq_of_equivalentSaddleFunctions {K L : SaddleFunction m n} (hKL : EquivalentSaddleFunctions K L) : ∀ u v, productSubdifferentialAt K u v = productSubdifferentialAt L u v

Lean source

Helper for Corollary 37.4.1: on every point where the product subdifferential is nonempty, equivalent saddle-functions already agree in value.

variable {m n : ℕ} lemma helperForCorollary_37_4_1_value_eq_on_productSubdifferentialDomain {K L : SaddleFunction m n} (hKL : EquivalentSaddleFunctions K L) : ∀ u v, Set.Nonempty (productSubdifferentialAt K u v) → K u v = L u v

Lean source

Helper for Corollary 37.5.1: the graph of the product subdifferential of K, written in the four-block coordinates (u, v, uStar, vStar).

variable {m n : ℕ} def helperForCorollary_37_5_1_productSubdifferentialGraph (K : SaddleFunction m n) : Set (((Fin m → ℝ) × (Fin n → ℝ)) × ((Fin m → ℝ) × (Fin n → ℝ))) := {p | p.2 ∈ productSubdifferentialAt K p.1.1 p.1.2}

Lean source

Helper for Corollary 37.5.1: the textbook map sending a graph point (u, v, uStar, vStar) to (u - uStar, v + vStar).

variable {m n : ℕ} def helperForCorollary_37_5_1_bookMap : (((Fin m → ℝ) × (Fin n → ℝ)) × ((Fin m → ℝ) × (Fin n → ℝ))) → ((Fin m → ℝ) × (Fin n → ℝ)) := fun p => (p.1.1 - p.2.1, p.1.2 + p.2.2)

Lean source

Helper for Corollary 37.5.1: after packing the primal variables with Fin.append, the textbook map becomes the packed addition map with the first dual block sign-twisted.

variable {m n : ℕ} lemma helperForCorollary_37_5_1_appendHomeomorph_bookMap_eq_packedAddition (u : Fin m → ℝ) (v : Fin n → ℝ) (uStar : Fin m → ℝ) (vStar : Fin n → ℝ) : (Fin.appendHomeomorph (X := ℝ) m n) (helperForCorollary_37_5_1_bookMap (((u, v), (uStar, vStar)))) = Fin.append u v + Fin.append (-uStar) vStar

Lean source

Helper for Corollary 37.5.1: unpacking the packed addition map recovers exactly the textbook map (u - uStar, v + vStar).

variable {m n : ℕ} lemma helperForCorollary_37_5_1_bookMap_eq_appendHomeomorph_symm_packedAddition (u : Fin m → ℝ) (v : Fin n → ℝ) (uStar : Fin m → ℝ) (vStar : Fin n → ℝ) : helperForCorollary_37_5_1_bookMap (((u, v), (uStar, vStar))) = (Fin.appendHomeomorph (X := ℝ) m n).symm (Fin.append u v + Fin.append (-uStar) vStar)

Lean source

Helper for Corollary 37.5.1: package the four-block graph coordinates (u, v, uStar, vStar) into the corrected packed coordinates ((u, vStar), (-uStar, v)).

variable {m n : ℕ} def helperForCorollary_37_5_1_packGraphCoordinates : (((Fin m → ℝ) × (Fin n → ℝ)) × ((Fin m → ℝ) × (Fin n → ℝ))) → ((Fin (m + n) → ℝ) × (Fin (m + n) → ℝ)) := fun p => (Fin.append p.1.1 p.2.2, Fin.append (-p.2.1) p.1.2)

Lean source

Helper for Corollary 37.5.1: this is the ordinary subdifferential graph of the packed graph function attached to a convex bifunction.

variable {m n : ℕ} def helperForCorollary_37_5_1_packedSubdifferentialGraph (F : (Fin m → ℝ) → (Fin n → ℝ) → EReal) : Set ((Fin (m + n) → ℝ) × (Fin (m + n) → ℝ)) := {q | dotProductEquiv ℝ (Fin (m + n)) q.2 ∈ subdifferentialAt (graphFunctionOfBifunction F) q.1}

Lean source

Helper for Corollary 37.5.1: the packed coordinate swap/sign map is continuous, so closedness can be transported by preimages once the graph-bridge is known.

variable {m n : ℕ} lemma helperForCorollary_37_5_1_packGraphCoordinates_continuous : Continuous (helperForCorollary_37_5_1_packGraphCoordinates (m := m) (n := n))

Lean source

Helper for Corollary 37.5.1: a closed proper convex bifunction gives a closed proper packed convex graph function on ℝ^(m+n).

variable {m n : ℕ} lemma helperForCorollary_37_5_1_graphFunction_properConvex {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : IsClosedConvexBifunction F) (hFproper : IsProperConvexBifunction F) : ProperConvexFunctionOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F)

Lean source

Helper for Corollary 37.5.1: a closed proper saddle-function admits a representative that is closed in the Chapter 6 graph-function sense as well as proper in the Section 34 image-closed sense.

variable {m n : ℕ} lemma helperForCorollary_37_5_1_closedProperRepresentativeWithClosedWitness (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hRepresentative : Section37ClosedRepresentativeQualification K hKclosed) (hGlobal : Section34Theorem34_2GlobalQualification m n) : ∃ F : (Fin m → ℝ) → (Fin n → ℝ) → EReal, ∃ hF : IsClosedConvexBifunction F, ClosedConvexBifunction F ∧ IsProperConvexBifunction F ∧ K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩

Lean source

Helper for Corollary 37.5.1: a closed proper convex bifunction gives a closed proper packed convex graph function on ℝ^(m+n).

variable {m n : ℕ} lemma helperForCorollary_37_5_1_graphFunction_closedProperConvex {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : IsClosedConvexBifunction F) (hClosed : ClosedConvexBifunction F) (hFproper : IsProperConvexBifunction F) : ClosedConvexFunction (graphFunctionOfBifunction F) ∧ ProperConvexFunctionOn (Set.univ : Set (Fin (m + n) → ℝ)) (graphFunctionOfBifunction F)

Lean source

Helper for Corollary 37.5.1: generated-class membership is exactly saddle-equivalence with the canonical pairing kernel of the representing convex bifunction.

variable {m n : ℕ} lemma helperForCorollary_37_5_1_generatedClass_gives_saddleEquivalent (K : SaddleFunction m n) {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : IsClosedConvexBifunction F) (hKGenerated : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) : saddleEquivalent K (convexBifunctionPairing F)

Lean source

Helper for Corollary 37.5.1: the corrected packing map is an ambient homeomorphism before restricting to either graph.

variable {m n : ℕ} lemma helperForCorollary_37_5_1_packGraphCoordinates_homeomorph : ∃ e : (((Fin m → ℝ) × (Fin n → ℝ)) × ((Fin m → ℝ) × (Fin n → ℝ))) ≃ₜ ((Fin (m + n) → ℝ) × (Fin (m + n) → ℝ)), ∀ p, e p = helperForCorollary_37_5_1_packGraphCoordinates (m := m) (n := n) p

Lean source

Helper for Corollary 37.5.1: the packed dual pairing with ((u', x') - (u, vStar), (-uStar, v)) is exactly the split affine term ⟪x', v⟫ - ⟪vStar, v⟫ - (⟪u', uStar⟫ - ⟪u, uStar⟫).

variable {m n : ℕ} lemma helperForCorollary_37_5_1_packedDotIncrement_eq_splitAffineTerm (u : Fin m → ℝ) (v : Fin n → ℝ) (uStar : Fin m → ℝ) (vStar : Fin n → ℝ) (u' : Fin m → ℝ) (x' : Fin n → ℝ) : (((dotProductEquiv ℝ (Fin (m + n)) (Fin.append (-uStar) v) (Fin.append u' x' - Fin.append u vStar) : ℝ)) : EReal) = (((dotProduct x' v : ℝ) : EReal) - ((dotProduct vStar v : ℝ) : EReal)) - (((dotProduct u' uStar : ℝ) : EReal) - ((dotProduct u uStar : ℝ) : EReal))

§37 — Conjugate Saddle-Functions and Minimax Theorems — Part 9

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Helper for Corollary 37.6.1: a bounded closed convex set in finite-dimensional Euclidean space has no nonzero recession direction.

variable {m n : ℕ} lemma helperForCorollary_37_6_1_nonzero_not_mem_recessionCone_of_bounded_closed_convex {k : ℕ} {C : Set (EuclideanSpace ℝ (Fin k))} (hCne : C.Nonempty) (hCclosed : IsClosed C) (hCconv : Convex ℝ C) (hCbdd : Bornology.IsBounded C) {w : EuclideanSpace ℝ (Fin k)} (hw : w ≠ 0) : w ∉ Set.recessionCone C

Lean source

Helper for Corollary 37.6.1: the same bounded-recession obstruction holds on the coordinate space Fin k → ℝ, which is the form used by effective domains in Section 37.

variable {m n : ℕ} lemma helperForCorollary_37_6_1_nonzero_not_mem_recessionCone_of_bounded_closed_convex_fin {k : ℕ} {C : Set (Fin k → ℝ)} (hCne : C.Nonempty) (hCclosed : IsClosed C) (hCconv : Convex ℝ C) (hCbdd : Bornology.IsBounded C) {w : Fin k → ℝ} (hw : w ≠ 0) : w ∉ Set.recessionCone C

Lean source

Helper for Corollary 37.6.1: bounded closed convex coordinate domains have only the trivial recession direction in each coordinate block.

variable {m n : ℕ} lemma helperForCorollary_37_6_1_nonzero_not_mem_recessionCones_of_bounded_closed_convex_domains {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hCne : C.Nonempty) (hCclosed : IsClosed C) (hCconv : Convex ℝ C) (hCbdd : Bornology.IsBounded C) (hDne : D.Nonempty) (hDclosed : IsClosed D) (hDconv : Convex ℝ D) (hDbdd : Bornology.IsBounded D) : (∀ z : Fin m → ℝ, z ≠ 0 → z ∉ Set.recessionCone C) ∧ (∀ w : Fin n → ℝ, w ≠ 0 → w ∉ Set.recessionCone D)

Lean source

Helper for Corollary 37.6.1: closed proper saddle-functions admit subtype witnesses in the intrinsic interiors of both effective coordinate domains, matching the witness types used later by Theorems 37.3 and 37.6.

variable {m n : ℕ} lemma helperForCorollary_37_6_1_intrinsicInteriorSubtype_nonempty (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) : Nonempty {u // u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)} ∧ Nonempty {v // v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)}

Lean source

Helper for Corollary 37.6.1: a bounded closed proper convex function has no nonzero recession directions, because any nonzero recession direction would already be a recession direction of a nonempty bounded real sublevel set.

variable {m n : ℕ} lemma helperForCorollary_37_6_1_nonzero_not_isRecessionDirection_of_bounded_closedProperConvex {k : ℕ} {g : (Fin k → ℝ) → EReal} {D : Set (Fin k → ℝ)} (hg : IsProperClosedConvexFunctionWithDomain g D) (hDne : D.Nonempty) (hDbdd : Bornology.IsBounded D) {w : Fin k → ℝ} (hw : w ≠ 0) : ¬ IsRecessionDirection g w

Lean source

Helper for Corollary 37.6.1: if the second effective domain is bounded, then no nonzero vector can be a recession direction of any relative-interior second slice K(u, ·).

variable {m n : ℕ} lemma helperForCorollary_37_6_1_nonzero_not_isRecessionDirection_secondSlice_of_bounded_domain (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) (hDbdd : Bornology.IsBounded (effectiveDomain₂ K)) (u : {u // u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)}) {w : Fin n → ℝ} (hw : w ≠ 0) : ¬ IsRecessionDirection (K u.1) w

Lean source

Helper for Corollary 37.6.1: if the first effective domain is bounded, then no nonzero vector can be a recession direction of any relative-interior negated first slice u ↦ -K(u, v).

variable {m n : ℕ} lemma helperForCorollary_37_6_1_nonzero_not_isRecessionDirection_negatedFirstSlice_of_bounded_domain (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) (hCbdd : Bornology.IsBounded (effectiveDomain₁ K)) (v : {v // v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)}) {z : Fin m → ℝ} (hz : z ≠ 0) : ¬ IsRecessionDirection (fun u => -K u v.1) z

Lean source

Helper for Corollary 37.6.1: bounded effective coordinate domains provide exactly the two no-common-recession hypotheses required later by Theorems 37.3 and 37.6.

variable {m n : ℕ} lemma helperForCorollary_37_6_1_bounded_effectiveDomains_yield_bothNoCommonRecessionConditions (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) (hCbdd : Bornology.IsBounded (effectiveDomain₁ K)) (hDbdd : Bornology.IsBounded (effectiveDomain₂ K)) : (∀ w : Fin n → ℝ, w ≠ 0 → ∃ u : {u // u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)}, ¬ IsRecessionDirection (K u.1) w) ∧ (∀ z : Fin m → ℝ, z ≠ 0 → ∃ v : {v // v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)}, ¬ IsRecessionDirection (fun u => -K u v.1) z)

Lean source

Helper for Corollary 37.6.1: bounded effective coordinate domains already imply the finite common saddle-value conclusion from the earlier Section 37.3 minimax theorem.

variable {m n : ℕ} lemma helperForCorollary_37_6_1_finiteSaddleValue_of_bounded_effectiveDomains (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) (hCbdd : Bornology.IsBounded (effectiveDomain₁ K)) (hDbdd : Bornology.IsBounded (effectiveDomain₂ K)) : minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K = maximinValue (C := Fin m → ℝ) (D := Fin n → ℝ) K ∧ minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K ≠ (⊤ : EReal) ∧ minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K ≠ (⊥ : EReal)

Lean source

Helper for Corollary 37.6.1: once bounded effective domains force a finite common extended-real saddle value, both minimax and maximin are recovered by coercing the minimax value's real part back into EReal.

variable {m n : ℕ} lemma helperForCorollary_37_6_1_commonSaddleValue_eq_minimax_toReal_of_bounded_effectiveDomains (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) (hCbdd : Bornology.IsBounded (effectiveDomain₁ K)) (hDbdd : Bornology.IsBounded (effectiveDomain₂ K)) : minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K = ((((minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K).toReal : ℝ)) : EReal) ∧ maximinValue (C := Fin m → ℝ) (D := Fin n → ℝ) K = ((((minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K).toReal : ℝ)) : EReal)

Lean source

Helper for Corollary 37.6.1: bounded effective coordinate domains therefore yield a genuine real number representing the common minimax-maximin value.

variable {m n : ℕ} lemma helperForCorollary_37_6_1_exists_real_commonSaddleValue_of_bounded_effectiveDomains (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) (hCbdd : Bornology.IsBounded (effectiveDomain₁ K)) (hDbdd : Bornology.IsBounded (effectiveDomain₂ K)) : ∃ α : ℝ, minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K = (α : EReal) ∧ maximinValue (C := Fin m → ℝ) (D := Fin n → ℝ) K = (α : EReal)

Lean source

Helper for Corollary 37.6.1: this part file already packages the bounded-domain recession obstructions together with the finite real common saddle-value conclusion, leaving only the later Section 37.6 saddle-point existence bridge to be supplied elsewhere.

variable {m n : ℕ} lemma helperForCorollary_37_6_1_bounded_effectiveDomains_package_part9_consequences (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) (hCbdd : Bornology.IsBounded (effectiveDomain₁ K)) (hDbdd : Bornology.IsBounded (effectiveDomain₂ K)) : (∀ w : Fin n → ℝ, w ≠ 0 → ∃ u : {u // u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)}, ¬ IsRecessionDirection (K u.1) w) ∧ (∀ z : Fin m → ℝ, z ≠ 0 → ∃ v : {v // v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)}, ¬ IsRecessionDirection (fun u => -K u v.1) z) ∧ ∃ α : ℝ, minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K = (α : EReal) ∧ maximinValue (C := Fin m → ℝ) (D := Fin n → ℝ) K = (α : EReal)

Lean source

Helper for Corollary 37.6.2: bounded effective coordinate domains already provide the exact two recession hypotheses required by the later Section 37.6 saddle-point existence theorem.

variable {m n : ℕ} lemma helperForCorollary_37_6_2_bounded_effectiveDomains_yield_theorem37_6_hypotheses (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) (hCbdd : Bornology.IsBounded (effectiveDomain₁ K)) (hDbdd : Bornology.IsBounded (effectiveDomain₂ K)) : (∀ w : Fin n → ℝ, w ≠ 0 → ∃ u : {u // u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)}, ¬ IsRecessionDirection (K u.1) w) ∧ (∀ z : Fin m → ℝ, z ≠ 0 → ∃ v : {v // v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)}, ¬ IsRecessionDirection (fun u => -K u v.1) z)

Lean source

Helper for Corollary 37.6.2: bounded effective coordinate domains also provide a real common minimax-maximin value, so the eventual corollary proof only needs the later existence bridge.

variable {m n : ℕ} lemma helperForCorollary_37_6_2_exists_real_commonSaddleValue_of_bounded_effectiveDomains (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) (hCbdd : Bornology.IsBounded (effectiveDomain₁ K)) (hDbdd : Bornology.IsBounded (effectiveDomain₂ K)) : ∃ α : ℝ, minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K = (α : EReal) ∧ maximinValue (C := Fin m → ℝ) (D := Fin n → ℝ) K = (α : EReal)

Lean source

Helper for Corollary 37.6.2: this part file already packages every bounded-domain consequence needed before the later Section 37.6 saddle-point existence bridge can be invoked.

variable {m n : ℕ} lemma helperForCorollary_37_6_2_bounded_effectiveDomains_package_part9_consequences (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) (hCbdd : Bornology.IsBounded (effectiveDomain₁ K)) (hDbdd : Bornology.IsBounded (effectiveDomain₂ K)) : (∀ w : Fin n → ℝ, w ≠ 0 → ∃ u : {u // u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)}, ¬ IsRecessionDirection (K u.1) w) ∧ (∀ z : Fin m → ℝ, z ≠ 0 → ∃ v : {v // v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)}, ¬ IsRecessionDirection (fun u => -K u v.1) z) ∧ ∃ α : ℝ, minimaxValue (C := Fin m → ℝ) (D := Fin n → ℝ) K = (α : EReal) ∧ maximinValue (C := Fin m → ℝ) (D := Fin n → ℝ) K = (α : EReal)

Lean source

Helper for Corollary 37.5.1: the packed ordinary subgradient condition is equivalent to the explicit tilted-fiber lower bound on F u' x' - ⟪x', v⟫ + ⟪u', uStar⟫.

variable {m n : ℕ} lemma helperForCorollary_37_5_1_packedSubgradientMem_iff_tiltedFiberLowerBound {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (v : Fin n → ℝ) (uStar : Fin m → ℝ) (vStar : Fin n → ℝ) : dotProductEquiv ℝ (Fin (m + n)) (Fin.append (-uStar) v) ∈ subdifferentialAt (graphFunctionOfBifunction F) (Fin.append u vStar) ↔ ∀ u' x', F u' x' - ((dotProduct x' v : ℝ) : EReal) + ((dotProduct u' uStar : ℝ) : EReal) ≥ F u vStar - ((dotProduct vStar v : ℝ) : EReal) + ((dotProduct u uStar : ℝ) : EReal)

Lean source

Helper for Corollary 37.5.1: second-partial membership for the canonical pairing is exactly attainment of the Fenchel kernel at the dual point vStar.

variable {m n : ℕ} lemma helperForCorollary_37_5_1_secondPartialMem_iff_pairingAttainment {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : IsClosedConvexBifunction F) (u : Fin m → ℝ) (v : Fin n → ℝ) (vStar : Fin n → ℝ) : vStar ∈ partialSubdifferentialInSecondVariable (convexBifunctionPairing F) u v ↔ convexBifunctionPairing F u v = (((dotProduct vStar v : ℝ) : EReal) - F u vStar)

Lean source

Helper for Corollary 37.5.1: re-adding the canceled affine term ⟪u',uStar⟫ collapses the intermediate correction ⟪u,uStar⟫ - ⟪u',uStar⟫ back to the baseline affine anchor.

variable {m n : ℕ} lemma helperForCorollary_37_5_1_readding_uStar_cancels_affineCorrection {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (v : Fin n → ℝ) (uStar : Fin m → ℝ) (vStar : Fin n → ℝ) (u' : Fin m → ℝ) : (F u vStar - ((dotProduct vStar v : ℝ) : EReal) + (((dotProduct u uStar - dotProduct u' uStar : ℝ)) : EReal)) + ((dotProduct u' uStar : ℝ) : EReal) = F u vStar - ((dotProduct vStar v : ℝ) : EReal) + ((dotProduct u uStar : ℝ) : EReal)

Lean source

Helper for Corollary 37.5.1: negating a single endpoint kernel swaps the primal and dual terms into the tilted-fiber order.

variable {m n : ℕ} lemma helperForCorollary_37_5_1_negatedEndpointKernel_eq_tiltedEndpoint {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u' : Fin m → ℝ) (x' : Fin n → ℝ) (v : Fin n → ℝ) : -((((dotProduct x' v : ℝ) : EReal) - F u' x')) = F u' x' - ((dotProduct x' v : ℝ) : EReal)

Lean source

Helper for Corollary 37.5.1: this exact one-point EReal normalization rewrites the tilted fiber lower bound into the affine kernel bound used by the pairing supremum, and conversely.

variable {m n : ℕ} lemma helperForCorollary_37_5_1_exactERealKernelTiltNormalization {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (v : Fin n → ℝ) (uStar : Fin m → ℝ) (vStar : Fin n → ℝ) (u' : Fin m → ℝ) (x' : Fin n → ℝ) : ((((dotProduct x' v : ℝ) : EReal) - F u' x') ≤ (((dotProduct vStar v : ℝ) : EReal) - F u vStar) + (((dotProduct u' uStar - dotProduct u uStar : ℝ)) : EReal)) ↔ F u' x' - ((dotProduct x' v : ℝ) : EReal) + ((dotProduct u' uStar : ℝ) : EReal) ≥ F u vStar - ((dotProduct vStar v : ℝ) : EReal) + ((dotProduct u uStar : ℝ) : EReal)

Lean source

Helper for Corollary 37.5.1: once the packed tilted-fiber lower bound has fixed the center value of the pairing, taking the supremum over each fiber recovers the first partial subdifferential.

variable {m n : ℕ} lemma helperForCorollary_37_5_1_tiltedFiberLowerBound_implies_pairingFirstPartial {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (u : Fin m → ℝ) (v : Fin n → ℝ) (uStar : Fin m → ℝ) (vStar : Fin n → ℝ) (hCenter : convexBifunctionPairing F u v = (((dotProduct vStar v : ℝ) : EReal) - F u vStar)) (hLower : ∀ u' x', F u' x' - ((dotProduct x' v : ℝ) : EReal) + ((dotProduct u' uStar : ℝ) : EReal) ≥ F u vStar - ((dotProduct vStar v : ℝ) : EReal) + ((dotProduct u uStar : ℝ) : EReal)) : uStar ∈ partialSubdifferentialInFirstVariable (convexBifunctionPairing F) u v

§37 — Conjugate Saddle-Functions and Minimax Theorems — Part 10

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Corollary 37.5.1: if K is a closed proper saddle-function, the graph of ∂K is closed, and the textbook map (u, v, uStar, vStar) ↦ (u - uStar, v + vStar) is a homeomorphism from that graph onto ℝ^m × ℝ^n.

variable {m n : ℕ} theorem corollary37_5_1_productSubdifferentialGraph_closed_homeomorphic (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hRepresentative : Section37ClosedRepresentativeQualification K hKclosed) (hGlobal : Section34Theorem34_2GlobalQualification m n) : IsClosed (helperForCorollary_37_5_1_productSubdifferentialGraph (m := m) (n := n) K) ∧ ∃ h : {p // p ∈ helperForCorollary_37_5_1_productSubdifferentialGraph (m := m) (n := n) K} ≃ₜ ((Fin m → ℝ) × (Fin n → ℝ)), ∀ p, h p = helperForCorollary_37_5_1_bookMap p.1

Lean source

Helper for Corollary 37.5.1: for the canonical pairing saddle-function attached to F, the four-block graph point ((u, v), (uStar, vStar)) is exactly the packed ordinary subgradient point ((u, vStar), (-uStar, v)) of the graph function of F.

variable {m n : ℕ} lemma helperForCorollary_37_5_1_pairingGraphPoint_iff_packedSubdifferentialGraphPoint {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : IsClosedConvexBifunction F) (u : Fin m → ℝ) (v : Fin n → ℝ) (uStar : Fin m → ℝ) (vStar : Fin n → ℝ) : (((u, v), (uStar, vStar)) ∈ helperForCorollary_37_5_1_productSubdifferentialGraph (m := m) (n := n) (convexBifunctionPairing F)) ↔ (helperForCorollary_37_5_1_packGraphCoordinates (m := m) (n := n) (((u, v), (uStar, vStar))) ∈ helperForCorollary_37_5_1_packedSubdifferentialGraph (m := m) (n := n) F)

Lean source

Helper for Corollary 37.5.1: the canonical pairing representative of a closed convex bifunction is already fixed by both mixed partial closures.

variable {m n : ℕ} lemma helperForCorollary_37_5_1_pairing_fixed_by_partialClosures {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : IsClosedConvexBifunction F) (hQ : Section34Theorem34_2Qualification F) : partialClosure₁ (convexBifunctionPairing F) = convexBifunctionPairing F ∧ partialClosure₂ (convexBifunctionPairing F) = convexBifunctionPairing F

Lean source

Helper for Corollary 37.5.1: generated-class membership forces the given closed saddle kernel to coincide pointwise with the canonical pairing representative.

variable {m n : ℕ} lemma helperForCorollary_37_5_1_generatedRepresentative_eq_pairing (K : SaddleFunction m n) {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : IsClosedConvexBifunction F) (hQ : Section34Theorem34_2Qualification F) (hKGenerated : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) : K = convexBifunctionPairing F

Lean source

Helper for Corollary 37.5.1: once K is represented by a closed proper convex bifunction F, the graph of ∂K should match the ordinary subdifferential graph of graphFunctionOfBifunction F under the corrected swap/sign packed coordinates.

variable {m n : ℕ} lemma helperForCorollary_37_5_1_originalGraphPoint_iff_packedSubdifferentialGraphPoint (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : IsClosedConvexBifunction F) (hQ : Section34Theorem34_2Qualification F) (hFproper : IsProperConvexBifunction F) (hKGenerated : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) (u : Fin m → ℝ) (v : Fin n → ℝ) (uStar : Fin m → ℝ) (vStar : Fin n → ℝ) : (((u, v), (uStar, vStar)) ∈ helperForCorollary_37_5_1_productSubdifferentialGraph (m := m) (n := n) K) ↔ (helperForCorollary_37_5_1_packGraphCoordinates (m := m) (n := n) (((u, v), (uStar, vStar))) ∈ helperForCorollary_37_5_1_packedSubdifferentialGraph (m := m) (n := n) F)

Lean source

Helper for Corollary 37.5.1: the ambient packing homeomorphism restricts to a homeomorphism between the original four-block graph of ∂K and the packed ordinary subdifferential graph of graphFunctionOfBifunction F.

variable {m n : ℕ} lemma helperForCorollary_37_5_1_restrictedPackHomeomorph (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : IsClosedConvexBifunction F) (hQ : Section34Theorem34_2Qualification F) (hFproper : IsProperConvexBifunction F) (hKGenerated : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) : ∃ e : {p // p ∈ helperForCorollary_37_5_1_productSubdifferentialGraph (m := m) (n := n) K} ≃ₜ {q // q ∈ helperForCorollary_37_5_1_packedSubdifferentialGraph (m := m) (n := n) F}, ∀ p, (e p).1 = helperForCorollary_37_5_1_packGraphCoordinates (m := m) (n := n) p.1

Lean source

Helper for Corollary 37.5.3: the origin fiber of any product subdifferential is a closed convex subset of ℝ^m × ℝ^n.

variable {m n : ℕ} lemma helperForCorollary_37_5_3_origin_fiber_closed_convex (KStar : SaddleFunction m n) : IsClosed (productSubdifferentialAt KStar (0 : Fin m → ℝ) (0 : Fin n → ℝ)) ∧ Convex ℝ (productSubdifferentialAt KStar (0 : Fin m → ℝ) (0 : Fin n → ℝ))

§37 — Conjugate Saddle-Functions and Minimax Theorems — Part 11

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Theorem 37.6: a closed proper concave-convex kernel satisfying both recession hypotheses from Theorem 37.3 has a saddle point.

variable {m n : ℕ} theorem section37_theorem37_6 (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) (hRepresentative : Section37ClosedRepresentativeQualification K hKclosed) (hNoCommonSecond : ∀ w : Fin n → ℝ, w ≠ 0 → ∃ u : {u // u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)}, ¬ IsRecessionDirection (K u.1) w) (hNoCommonFirst : ∀ z : Fin m → ℝ, z ≠ 0 → ∃ v : {v // v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)}, ¬ IsRecessionDirection (fun u => -K u v.1) z) : ∃ u v, IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K u v

Lean source

Helper for Theorem 37.6: the lower conjugate of a closed proper saddle-function is itself a closed proper saddle-function.

variable {m n : ℕ} lemma helperForTheorem_37_6_lowerConjugate_closed_proper (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hGlobal : Section34Theorem34_2GlobalQualification m n) : IsClosedSaddleFunction (fun uStar x => theorem37ValueSupInf K uStar x) ∧ IsProperSaddleFunction (fun uStar x => theorem37ValueSupInf K uStar x)

Lean source

Helper for Theorem 37.6: flip the first m packed coordinates and keep the last n coordinates fixed.

variable {m n : ℕ} def helperForTheorem_37_6_flipFirstPackedBlock : (Fin (m + n) → ℝ) → (Fin (m + n) → ℝ) := fun z => Fin.append (fun i : Fin m => -z (Fin.castAdd n i)) (fun j : Fin n => z (Fin.natAdd m j))

Lean source

Helper for Theorem 37.6: on a split packed point, the flip simply negates the first block.

variable {m n : ℕ} lemma helperForTheorem_37_6_flipFirstPackedBlock_append (u : Fin m → ℝ) (v : Fin n → ℝ) : helperForTheorem_37_6_flipFirstPackedBlock (m := m) (n := n) (Fin.append u v) = Fin.append (-u) v

Lean source

Helper for Theorem 37.6: flipping the first packed block is an involution.

variable {m n : ℕ} lemma helperForTheorem_37_6_flipFirstPackedBlock_involutive : Function.Involutive (helperForTheorem_37_6_flipFirstPackedBlock (m := m) (n := n))

Lean source

Helper for Theorem 37.6: the packed first-block flip distributes over subtraction.

variable {m n : ℕ} lemma helperForTheorem_37_6_flipFirstPackedBlock_sub (x y : Fin (m + n) → ℝ) : helperForTheorem_37_6_flipFirstPackedBlock (m := m) (n := n) (x - y) = helperForTheorem_37_6_flipFirstPackedBlock (m := m) (n := n) x - helperForTheorem_37_6_flipFirstPackedBlock (m := m) (n := n) y

Lean source

Helper for Theorem 37.6: the packed first-block flip is self-adjoint for the Euclidean dot product.

variable {m n : ℕ} lemma helperForTheorem_37_6_dotProduct_flipFirstPackedBlock_left (x y : Fin (m + n) → ℝ) : dotProduct (helperForTheorem_37_6_flipFirstPackedBlock (m := m) (n := n) x) y = dotProduct x (helperForTheorem_37_6_flipFirstPackedBlock (m := m) (n := n) y)

Lean source

Helper for Theorem 37.6: precomposing a convex graph function with the first-block flip transports Euclidean subgradients by the same flip on both the base point and the dual vector.

variable {m n : ℕ} lemma helperForTheorem_37_6_euclideanSubgradient_precomp_flipFirstPackedBlock_iff (h : (Fin (m + n) → ℝ) → EReal) (x y : Fin (m + n) → ℝ) : IsEuclideanSubgradientAt (fun z => h (helperForTheorem_37_6_flipFirstPackedBlock (m := m) (n := n) z)) x y ↔ IsEuclideanSubgradientAt h (helperForTheorem_37_6_flipFirstPackedBlock (m := m) (n := n) x) (helperForTheorem_37_6_flipFirstPackedBlock (m := m) (n := n) y)

Lean source

Helper for Theorem 37.6: a pair in the origin fiber of the lower conjugate product subdifferential is exactly a saddle point of the original kernel.

variable {m n : ℕ} lemma helperForTheorem_37_6_origin_productSubdifferentialPair_iff_saddlePoint (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hRepresentative : Section37ClosedRepresentativeQualification K hKclosed) (hGlobal : Section34Theorem34_2GlobalQualification m n) (u : Fin m → ℝ) (v : Fin n → ℝ) : (u, v) ∈ productSubdifferentialAt (fun uStar x => theorem37ValueSupInf K uStar x) (0 : Fin m → ℝ) (0 : Fin n → ℝ) ↔ IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K u v

Lean source

Helper for Theorem 37.6: if both recession-direction hypotheses from Theorem 37.3 hold, then the origin lies in the relative-interior effective domain of the lower conjugate kernel (uStar, x) ↦ theorem37ValueSupInf K uStar x.

variable {m n : ℕ} lemma helperForTheorem_37_6_origin_mem_saddleKernelDomain_of_bothNoCommonRecessionConditions (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) (hNoCommonSecond : ∀ w : Fin n → ℝ, w ≠ 0 → ∃ u : {u // u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)}, ¬ IsRecessionDirection (K u.1) w) (hNoCommonFirst : ∀ z : Fin m → ℝ, z ≠ 0 → ∃ v : {v // v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)}, ¬ IsRecessionDirection (fun u => -K u v.1) z) : ((0 : Fin m → ℝ), (0 : Fin n → ℝ)) ∈ saddleKernelDomain (fun uStar x => theorem37ValueSupInf K uStar x)

Lean source

Helper for Theorem 37.6: once the lower conjugate kernel is known to be closed and proper, the two recession-direction hypotheses force nonemptiness of the origin fiber of its product subdifferential.

variable {m n : ℕ} lemma helperForTheorem_37_6_origin_mem_productSubdifferentialDomain_of_bothNoCommonRecessionConditions (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) (hKStarclosed : IsClosedSaddleFunction (fun uStar x => theorem37ValueSupInf K uStar x)) (hKStarproper : IsProperSaddleFunction (fun uStar x => theorem37ValueSupInf K uStar x)) (hNoCommonSecond : ∀ w : Fin n → ℝ, w ≠ 0 → ∃ u : {u // u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)}, ¬ IsRecessionDirection (K u.1) w) (hNoCommonFirst : ∀ z : Fin m → ℝ, z ≠ 0 → ∃ v : {v // v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)}, ¬ IsRecessionDirection (fun u => -K u v.1) z) : ((0 : Fin m → ℝ), (0 : Fin n → ℝ)) ∈ {p : (Fin m → ℝ) × (Fin n → ℝ) | Set.Nonempty (productSubdifferentialAt (fun uStar x => theorem37ValueSupInf K uStar x) p.1 p.2)}

Lean source

Helper for Theorem 37.6: under the two recession hypotheses, the origin fiber of the lower conjugate product subdifferential is nonempty, closed, and convex.

variable {m n : ℕ} lemma helperForTheorem_37_6_origin_productSubdifferentialFiber_nonempty_closed_convex_of_bothNoCommonRecessionConditions (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) (hKStarclosed : IsClosedSaddleFunction (fun uStar x => theorem37ValueSupInf K uStar x)) (hKStarproper : IsProperSaddleFunction (fun uStar x => theorem37ValueSupInf K uStar x)) (hNoCommonSecond : ∀ w : Fin n → ℝ, w ≠ 0 → ∃ u : {u // u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)}, ¬ IsRecessionDirection (K u.1) w) (hNoCommonFirst : ∀ z : Fin m → ℝ, z ≠ 0 → ∃ v : {v // v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)}, ¬ IsRecessionDirection (fun u => -K u v.1) z) : Set.Nonempty (productSubdifferentialAt (fun uStar x => theorem37ValueSupInf K uStar x) (0 : Fin m → ℝ) (0 : Fin n → ℝ)) ∧ IsClosed (productSubdifferentialAt (fun uStar x => theorem37ValueSupInf K uStar x) (0 : Fin m → ℝ) (0 : Fin n → ℝ)) ∧ Convex ℝ (productSubdifferentialAt (fun uStar x => theorem37ValueSupInf K uStar x) (0 : Fin m → ℝ) (0 : Fin n → ℝ))

Lean source

Helper for Theorem 37.6: the nonempty origin fiber of the lower conjugate product subdifferential contains an explicit pair (u, v).

variable {m n : ℕ} lemma helperForTheorem_37_6_exists_origin_productSubdifferentialPair_of_bothNoCommonRecessionConditions (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hQ : Section37Theorem37_2Qualification K) (hKStarclosed : IsClosedSaddleFunction (fun uStar x => theorem37ValueSupInf K uStar x)) (hKStarproper : IsProperSaddleFunction (fun uStar x => theorem37ValueSupInf K uStar x)) (hNoCommonSecond : ∀ w : Fin n → ℝ, w ≠ 0 → ∃ u : {u // u ∈ intrinsicInterior ℝ (effectiveDomain₁ K)}, ¬ IsRecessionDirection (K u.1) w) (hNoCommonFirst : ∀ z : Fin m → ℝ, z ≠ 0 → ∃ v : {v // v ∈ intrinsicInterior ℝ (effectiveDomain₂ K)}, ¬ IsRecessionDirection (fun u => -K u v.1) z) : ∃ u v, (u, v) ∈ productSubdifferentialAt (fun uStar x => theorem37ValueSupInf K uStar x) (0 : Fin m → ℝ) (0 : Fin n → ℝ)

§37 — Conjugate Saddle-Functions and Minimax Theorems — Part 12

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Corollary 37.6.2: a continuous finite concave-convex kernel on nonempty closed bounded convex sets has a saddle point on C × D.

variable {m n : ℕ} theorem corollary37_6_2_continuousFiniteConcaveConvexOn_closedBoundedConvex_hasSaddlePoint {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hCclosed : IsClosed C) (hDclosed : IsClosed D) (hCconv : Convex ℝ C) (hDconv : Convex ℝ D) (hCbdd : Bornology.IsBounded C) (hDbdd : Bornology.IsBounded D) (hCne : C.Nonempty) (hDne : D.Nonempty) (hStrongClosures : Section34Text34_1_9StrongClosureQualification m n) (hClosedNoBot : ∀ L : SaddleFunction m n, IsClosedSaddleFunction L → HasNoBotValuesBifunction L) (J : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) (hQ : Section37Theorem37_2Qualification (lowerSimpleExtension C D (fun u v => ((J u v : ℝ) : EReal)))) (hRepresentative : ∀ hClosed : IsClosedSaddleFunction (lowerSimpleExtension C D (fun u v => ((J u v : ℝ) : EReal))), Section37ClosedRepresentativeQualification (lowerSimpleExtension C D (fun u v => ((J u v : ℝ) : EReal))) hClosed) (hJcont : ContinuousOn (fun p : (Fin m → ℝ) × (Fin n → ℝ) => J p.1 p.2) (C ×ˢ D)) (hJcc : IsConcaveConvexOn C D (fun u v => ((J u v : ℝ) : EReal))) : ∃ u : C, ∃ v : D, IsSaddlePoint (C := C) (D := D) (fun u v => ((J u.1 v.1 : ℝ) : EReal)) u v

Lean source

Helper for Corollary 37.6.2: when both arguments satisfy the constraints, the ambient lower simple extension agrees with the original real-valued kernel on subtype points.

variable {m n : ℕ} lemma helperForCorollary_37_6_2_lowerSimpleExtensionOfReal_agrees_on_subtypes {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (J : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) (u : C) (v : D) : lowerSimpleExtension C D (fun u v => ((J u v : ℝ) : EReal)) u.1 v.1 = ((J u.1 v.1 : ℝ) : EReal)

Lean source

Helper for Corollary 37.6.2: for a real-valued kernel, the first effective domain of the lower simple extension is exactly the first constraint set C.

variable {m n : ℕ} lemma helperForCorollary_37_6_2_effectiveDomain1_lowerSimpleExtensionOfReal_eq {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hDne : D.Nonempty) (J : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) : effectiveDomain₁ (lowerSimpleExtension C D (fun u v => ((J u v : ℝ) : EReal))) = C

Lean source

Helper for Corollary 37.6.2: for a real-valued kernel, the second effective domain of the lower simple extension is exactly the second constraint set D.

variable {m n : ℕ} lemma helperForCorollary_37_6_2_effectiveDomain2_lowerSimpleExtensionOfReal_eq {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hCne : C.Nonempty) (J : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) : effectiveDomain₂ (lowerSimpleExtension C D (fun u v => ((J u v : ℝ) : EReal))) = D

Lean source

Helper for Corollary 37.6.2: the lower simple extension of a continuous finite concave-convex kernel on C × D is a closed proper saddle-function on the ambient space.

variable {m n : ℕ} lemma helperForCorollary_37_6_2_lowerSimpleExtensionOfReal_closed_proper {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hCclosed : IsClosed C) (hDclosed : IsClosed D) (hCconv : Convex ℝ C) (hDconv : Convex ℝ D) (hCne : C.Nonempty) (hDne : D.Nonempty) (hStrongClosures : Section34Text34_1_9StrongClosureQualification m n) (hClosedNoBot : ∀ L : SaddleFunction m n, IsClosedSaddleFunction L → HasNoBotValuesBifunction L) (J : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) (hJcont : ContinuousOn (fun p : (Fin m → ℝ) × (Fin n → ℝ) => J p.1 p.2) (C ×ˢ D)) (hJcc : IsConcaveConvexOn C D (fun u v => ((J u v : ℝ) : EReal))) : IsClosedSaddleFunction (lowerSimpleExtension C D (fun u v => ((J u v : ℝ) : EReal))) ∧ IsProperSaddleFunction (lowerSimpleExtension C D (fun u v => ((J u v : ℝ) : EReal)))

Lean source

Helper for Corollary 37.6.2: any global saddle point of the lower simple extension of J must lie in C × D.

variable {m n : ℕ} lemma helperForCorollary_37_6_2_mem_constraints_of_isSaddlePoint_lowerSimpleExtensionOfReal {C : Set (Fin m → ℝ)} {D : Set (Fin n → ℝ)} (hCne : C.Nonempty) (hDne : D.Nonempty) (J : (Fin m → ℝ) → (Fin n → ℝ) → ℝ) {u : Fin m → ℝ} {v : Fin n → ℝ} (hS : IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) (lowerSimpleExtension C D (fun u v => ((J u v : ℝ) : EReal))) u v) : u ∈ C ∧ v ∈ D

§37 — Conjugate Saddle-Functions and Minimax Theorems — Part 13

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Corollary 37.5.3: for a closed proper saddle-function K, any closed proper conjugate representative KStar equivalent to the lower Section 37 conjugate has origin product subdifferential equal to the saddle-point set of K.

variable {m n : ℕ} theorem corollary37_5_3_conjugate_origin_productSubdifferential_eq_saddle_points (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hRepresentative : Section37ClosedRepresentativeQualification K hKclosed) (hGlobal : Section34Theorem34_2GlobalQualification m n) (KStar : SaddleFunction m n) (hKStarclosed : IsClosedSaddleFunction KStar) (hKStarproper : IsProperSaddleFunction KStar) (hKStarConj : EquivalentSaddleFunctions KStar (fun uStar x => theorem37ValueSupInf K uStar x)) : productSubdifferentialAt KStar (0 : Fin m → ℝ) (0 : Fin n → ℝ) = {p : (Fin m → ℝ) × (Fin n → ℝ) | IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K p.1 p.2}

Lean source

Helper for Corollary 37.5.3: once the origin fiber of ∂KStar is identified with the saddle-point set of K, origin-membership in dom ∂KStar is equivalent to existence of a saddle point of K.

variable {m n : ℕ} lemma helperForCorollary_37_5_3_origin_domain_iff_exists_saddle_point {K KStar : SaddleFunction m n} (hOriginFiber : productSubdifferentialAt KStar (0 : Fin m → ℝ) (0 : Fin n → ℝ) = {p : (Fin m → ℝ) × (Fin n → ℝ) | IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K p.1 p.2}) : (((0 : Fin m → ℝ), (0 : Fin n → ℝ)) ∈ {p : (Fin m → ℝ) × (Fin n → ℝ) | Set.Nonempty (productSubdifferentialAt KStar p.1 p.2)}) ↔ ∃ u v, IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K u v

Lean source

Helper for Corollary 37.5.3: the ri (dom KStar) inclusion from Theorem 37.4 yields a saddle point of K as soon as the origin fiber of ∂KStar has been identified.

variable {m n : ℕ} lemma helperForCorollary_37_5_3_origin_kernelDomain_implies_exists_saddle_point {K KStar : SaddleFunction m n} (hKStarclosed : IsClosedSaddleFunction KStar) (hKStarproper : IsProperSaddleFunction KStar) (hGlobal : Section34Theorem34_2GlobalQualification m n) (hOriginFiber : productSubdifferentialAt KStar (0 : Fin m → ℝ) (0 : Fin n → ℝ) = {p : (Fin m → ℝ) × (Fin n → ℝ) | IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K p.1 p.2}) : ((0 : Fin m → ℝ), (0 : Fin n → ℝ)) ∈ saddleKernelDomain KStar → ∃ u v, IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K u v

Lean source

Helper for Corollary 37.5.3: flip the first m packed coordinates and keep the last n coordinates fixed.

variable {m n : ℕ} def helperForCorollary_37_5_3_flipFirstPackedBlock : (Fin (m + n) → ℝ) → (Fin (m + n) → ℝ) := fun z => Fin.append (fun i : Fin m => -z (Fin.castAdd n i)) (fun j : Fin n => z (Fin.natAdd m j))

Lean source

Helper for Corollary 37.5.3: on a split packed point, the flip simply negates the first block.

variable {m n : ℕ} lemma helperForCorollary_37_5_3_flipFirstPackedBlock_append (u : Fin m → ℝ) (v : Fin n → ℝ) : helperForCorollary_37_5_3_flipFirstPackedBlock (m := m) (n := n) (Fin.append u v) = Fin.append (-u) v

Lean source

Helper for Corollary 37.5.3: flipping the first packed block is an involution.

variable {m n : ℕ} lemma helperForCorollary_37_5_3_flipFirstPackedBlock_involutive : Function.Involutive (helperForCorollary_37_5_3_flipFirstPackedBlock (m := m) (n := n))

Lean source

Helper for Corollary 37.5.3: the packed first-block flip distributes over subtraction.

variable {m n : ℕ} lemma helperForCorollary_37_5_3_flipFirstPackedBlock_sub (x y : Fin (m + n) → ℝ) : helperForCorollary_37_5_3_flipFirstPackedBlock (m := m) (n := n) (x - y) = helperForCorollary_37_5_3_flipFirstPackedBlock (m := m) (n := n) x - helperForCorollary_37_5_3_flipFirstPackedBlock (m := m) (n := n) y

Lean source

Helper for Corollary 37.5.3: the packed first-block flip is self-adjoint for the Euclidean dot product.

variable {m n : ℕ} lemma helperForCorollary_37_5_3_dotProduct_flipFirstPackedBlock_left (x y : Fin (m + n) → ℝ) : dotProduct (helperForCorollary_37_5_3_flipFirstPackedBlock (m := m) (n := n) x) y = dotProduct x (helperForCorollary_37_5_3_flipFirstPackedBlock (m := m) (n := n) y)

Lean source

Helper for Corollary 37.5.3: precomposing a convex graph function with the first-block flip transports Euclidean subgradients by the same flip on both the base point and the dual vector.

variable {m n : ℕ} lemma helperForCorollary_37_5_3_euclideanSubgradient_precomp_flipFirstPackedBlock_iff (h : (Fin (m + n) → ℝ) → EReal) (x y : Fin (m + n) → ℝ) : IsEuclideanSubgradientAt (fun z => h (helperForCorollary_37_5_3_flipFirstPackedBlock (m := m) (n := n) z)) x y ↔ IsEuclideanSubgradientAt h (helperForCorollary_37_5_3_flipFirstPackedBlock (m := m) (n := n) x) (helperForCorollary_37_5_3_flipFirstPackedBlock (m := m) (n := n) y)

Lean source

Helper for Corollary 37.5.3: the only remaining primal-dual step is the canonical lower Section 37 conjugate case, where KStar = theorem37ValueSupInf K.

variable {m n : ℕ} lemma helperForCorollary_37_5_3_lowerConjugate_origin_productSubdifferential_eq_saddle_points (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hRepresentative : Section37ClosedRepresentativeQualification K hKclosed) (hGlobal : Section34Theorem34_2GlobalQualification m n) : productSubdifferentialAt (fun uStar x => theorem37ValueSupInf K uStar x) (0 : Fin m → ℝ) (0 : Fin n → ℝ) = {p : (Fin m → ℝ) × (Fin n → ℝ) | IsSaddlePoint (C := (Fin m → ℝ)) (D := (Fin n → ℝ)) K p.1 p.2}

§37 — Conjugate Saddle-Functions and Minimax Theorems — Part 14

Lean contributors (project-level attribution): Changyu Zou, Chenyi Li, Guangxuan Pan, Pengfei Hao, Qiming Dai, Shu Miao, Siyuan Shao, Suwu Wu, Wanli Ma, Weiran Shi, Xinyi Guo, Xuran Sun, Yifan Bai, Yijie Wang, Yunfei Zhang, Yunxi Duan, Yuhao Jiang, Zebo Liu, Zhiyan Wang, Zichen Wang, xdedlhqhw.

Source header · Attribution source

Corollary 37.5.2: if K is a closed proper concave-convex saddle-function, then the packed textbook saddle-subgradient map (u, v) ↦ {(-u*, v*) | (u*, v*) ∈ ∂K(u, v)} is maximal monotone.

variable {m n : ℕ} theorem corollary37_5_2_saddle_subgradient_is_maximal_monotone (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hRepresentative : Section37ClosedRepresentativeQualification K hKclosed) (hGlobal : Section34Theorem34_2GlobalQualification m n) : IsMaximalMonotoneEuclideanSetValuedMap (helperForCorollary_37_5_2_packedRho (m := m) (n := n) K)

Lean source

For Corollary 37.5.2: if K is finite and differentiable everywhere, then the packed signed gradient map (u, v) ↦ (-∇₁K(u, v), ∇₂K(u, v)) is maximal monotone.

variable {m n : ℕ} theorem corollary37_5_2_differentiable_gradient_is_maximal_monotone (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) (hRepresentative : Section37ClosedRepresentativeQualification K hKclosed) (hGlobal : Section34Theorem34_2GlobalQualification m n) (hK : IsGloballyConcaveConvexERealKernel K) (hFinite : ∀ u v, K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hDiff : ∀ u v, ERealDifferentiableAt (packedSaddleKernel K) (Fin.append u v)) : IsMaximalMonotoneEuclideanSetValuedMap (helperForCorollary_37_5_2_packedSignedGradient (m := m) (n := n) K hDiff)

Lean source

Helper for Corollary 37.5.2: the packed textbook operator ρ(u,v) = {(-u*, v*) | (u*, v*) ∈ ∂K(u,v)} on ℝ^(m+n).

variable {m n : ℕ} def helperForCorollary_37_5_2_packedRho (K : SaddleFunction m n) : (Fin (m + n) → ℝ) → Set (Fin (m + n) → ℝ) := fun x => let uv := (Fin.appendHomeomorph (X := ℝ) m n).symm x {y | let yBlocks := (Fin.appendHomeomorph (X := ℝ) m n).symm y (-yBlocks.1, yBlocks.2) ∈ productSubdifferentialAt K uv.1 uv.2}

Lean source

Helper for Corollary 37.5.2: on packed graph coordinates, swap the second primal block with the second dual block.

variable {m n : ℕ} def helperForCorollary_37_5_2_swapSecondBlocks : ((Fin (m + n) → ℝ) × (Fin (m + n) → ℝ)) → ((Fin (m + n) → ℝ) × (Fin (m + n) → ℝ)) := fun p => let xBlocks := (Fin.appendHomeomorph (X := ℝ) m n).symm p.1 let yBlocks := (Fin.appendHomeomorph (X := ℝ) m n).symm p.2 (Fin.append xBlocks.1 yBlocks.2, Fin.append yBlocks.1 xBlocks.2)

Lean source

Helper for Corollary 37.5.2: transport a packed set-valued map by the block-swap involution on graph coordinates.

variable {m n : ℕ} def helperForCorollary_37_5_2_swapSecondBlocksTransport (T : (Fin (m + n) → ℝ) → Set (Fin (m + n) → ℝ)) : (Fin (m + n) → ℝ) → Set (Fin (m + n) → ℝ) := fun x => {y | let q := helperForCorollary_37_5_2_swapSecondBlocks (m := m) (n := n) (x, y) q.2 ∈ T q.1}

Lean source

Helper for Corollary 37.5.2: in the everywhere-differentiable case, the packed signed gradient map is the singleton-valued realization of ρ.

variable {m n : ℕ} noncomputable def helperForCorollary_37_5_2_packedSignedGradient (K : SaddleFunction m n) (hDiff : ∀ u v, ERealDifferentiableAt (packedSaddleKernel K) (Fin.append u v)) : (Fin (m + n) → ℝ) → Set (Fin (m + n) → ℝ) := fun x => let uv := (Fin.appendHomeomorph (X := ℝ) m n).symm x let grad := packedSaddleKernelGradientPairAt (K := K) (u := uv.1) (v := uv.2) (hDiff uv.1 uv.2) {Fin.append (-grad.1) grad.2}

Lean source

Helper for Corollary 37.5.2: the block swap is literally involutive on packed graph coordinates.

variable {m n : ℕ} lemma helperForCorollary_37_5_2_swapSecondBlocks_involutive : Function.Involutive (helperForCorollary_37_5_2_swapSecondBlocks (m := m) (n := n))

Lean source

Helper for Corollary 37.5.2: the block swap preserves the monotonicity pairing ⟪x - y, u - v⟫.

variable {m n : ℕ} lemma helperForCorollary_37_5_2_swapSecondBlocks_preserves_pairing (p q : ((Fin (m + n) → ℝ) × (Fin (m + n) → ℝ))) : dotProduct ((helperForCorollary_37_5_2_swapSecondBlocks (m := m) (n := n) p).1 - (helperForCorollary_37_5_2_swapSecondBlocks (m := m) (n := n) q).1) ((helperForCorollary_37_5_2_swapSecondBlocks (m := m) (n := n) p).2 - (helperForCorollary_37_5_2_swapSecondBlocks (m := m) (n := n) q).2) = dotProduct (p.1 - q.1) (p.2 - q.2)

Lean source

Helper for Corollary 37.5.2: the graph of ρ is exactly the block-swapped graph of the ordinary packed subdifferential.

variable {m n : ℕ} lemma helperForCorollary_37_5_2_packedRhoGraphPoint_iff_packedSubdifferentialGraphPoint (K : SaddleFunction m n) (hKclosed : IsClosedSaddleFunction K) (hKproper : IsProperSaddleFunction K) {F : (Fin m → ℝ) → (Fin n → ℝ) → EReal} (hF : IsClosedConvexBifunction F) (hQ : Section34Theorem34_2Qualification F) (hFproper : IsProperConvexBifunction F) (hKGenerated : K ∈ EquivalenceClassGeneratedByConvexBifunction ⟨F, hF⟩) (u : Fin m → ℝ) (v : Fin n → ℝ) (uStar : Fin m → ℝ) (vStar : Fin n → ℝ) : Fin.append (-uStar) vStar ∈ helperForCorollary_37_5_2_packedRho (m := m) (n := n) K (Fin.append u v) ↔ helperForCorollary_37_5_1_packGraphCoordinates (m := m) (n := n) (((u, v), (uStar, vStar))) ∈ helperForCorollary_37_5_1_packedSubdifferentialGraph (m := m) (n := n) F

Lean source

Helper for Corollary 37.5.2: transporting a monotone map by the block swap preserves monotonicity.

variable {m n : ℕ} lemma helperForCorollary_37_5_2_swapSecondBlocks_preserves_monotone (T : (Fin (m + n) → ℝ) → Set (Fin (m + n) → ℝ)) (hT : IsMonotoneEuclideanSetValuedMap T) : IsMonotoneEuclideanSetValuedMap (helperForCorollary_37_5_2_swapSecondBlocksTransport (m := m) (n := n) T)

Lean source

Helper for Corollary 37.5.2: transporting a maximal monotone map by the block swap preserves maximal monotonicity.

variable {m n : ℕ} lemma helperForCorollary_37_5_2_swapSecondBlocks_preserves_maximalMonotone (T : (Fin (m + n) → ℝ) → Set (Fin (m + n) → ℝ)) (hT : IsMaximalMonotoneEuclideanSetValuedMap T) : IsMaximalMonotoneEuclideanSetValuedMap (helperForCorollary_37_5_2_swapSecondBlocksTransport (m := m) (n := n) T)

Lean source

Helper for Corollary 37.5.2: at a finite differentiability point, the ρ fiber is the singleton determined by the signed packed gradient.

variable {m n : ℕ} lemma helperForCorollary_37_5_2_singletonRho_of_differentiablePoint (K : SaddleFunction m n) (hK : IsGloballyConcaveConvexERealKernel K) {u : Fin m → ℝ} {v : Fin n → ℝ} (hFinite : K u v ≠ (⊤ : EReal) ∧ K u v ≠ (⊥ : EReal)) (hDiff : ERealDifferentiableAt (packedSaddleKernel K) (Fin.append u v)) : helperForCorollary_37_5_2_packedRho (m := m) (n := n) K (Fin.append u v) = {Fin.append (-(packedSaddleKernelGradientPairAt (K := K) (u := u) (v := v) hDiff).1) ((packedSaddleKernelGradientPairAt (K := K) (u := u) (v := v) hDiff).2)}