Part III — Duality Correspondences
Proofs omitted. Hover over highlighted terms to see their Lean types.
Reading sections
§11 — Separation Theorems — Part 1
Lean authors: Zichen Wang, Wanli Ma, Suwan Wu, Yunxi Duan, Zaiwen Wen.
Text 11.0.1: Let C₁ and C₂ be non-empty sets in ℝ^n. A hyperplane H is said to separate C₁ and C₂ if C₁ is contained in one of the closed half-spaces associated with H and C₂ lies in the opposite closed half-space.
open scoped Pointwise
def HyperplaneSeparates (n : Nat) (H C₁ C₂ : Set (Fin n → Real)) : Prop :=
C₁.Nonempty ∧
C₂.Nonempty ∧
∃ (b : Fin n → Real) (β : Real),
b ≠ 0 ∧
H = {x : Fin n → Real | x ⬝ᵥ b = β} ∧
((C₁ ⊆ {x : Fin n → Real | x ⬝ᵥ b ≤ β} ∧ C₂ ⊆ {x : Fin n → Real | β ≤ x ⬝ᵥ b}) ∨
(C₂ ⊆ {x : Fin n → Real | x ⬝ᵥ b ≤ β} ∧ C₁ ⊆ {x : Fin n → Real | β ≤ x ⬝ᵥ b}))
Text 11.0.2: A hyperplane H is said to separate C₁ and C₂ properly if C₁ and C₂ are not both actually contained in H itself.
open scoped Pointwise
def HyperplaneSeparatesProperly (n : Nat) (H C₁ C₂ : Set (Fin n → Real)) : Prop :=
HyperplaneSeparates n H C₁ C₂ ∧ ¬ (C₁ ⊆ H ∧ C₂ ⊆ H)
Extract the "properness" clause from HyperplaneSeparatesProperly: the two sets cannot both lie entirely in the separating hyperplane.
open scoped Pointwise
lemma not_both_subset_of_hyperplaneSeparatesProperly {n : Nat}
{H C₁ C₂ : Set (Fin n → Real)} :
HyperplaneSeparatesProperly n H C₁ C₂ → ¬ (C₁ ⊆ H ∧ C₂ ⊆ H)
Pure-Prop helper: from ¬ (A ⊆ H ∧ B ⊆ H) we get A ⊆ H → ¬ B ⊆ H.
open scoped Pointwise
lemma subset_left_imp_not_subset_right {α : Type*} {A B H : Set α}
(hnot : ¬ (A ⊆ H ∧ B ⊆ H)) : (A ⊆ H → ¬ B ⊆ H)
Pure-Prop helper: from ¬ (A ⊆ H ∧ B ⊆ H) we get B ⊆ H → ¬ A ⊆ H.
open scoped Pointwise
lemma subset_right_imp_not_subset_left {α : Type*} {A B H : Set α}
(hnot : ¬ (A ⊆ H ∧ B ⊆ H)) : (B ⊆ H → ¬ A ⊆ H)
Text 11.1.1: Proper separation allows that one (but not both) of the sets be contained in the separating hyperplane itself.
open scoped Pointwise
theorem hyperplaneSeparatesProperly_imp_atMostOne_subset_hyperplane (n : Nat)
(H C₁ C₂ : Set (Fin n → Real)) :
HyperplaneSeparatesProperly n H C₁ C₂ →
(C₁ ⊆ H → ¬ C₂ ⊆ H) ∧ (C₂ ⊆ H → ¬ C₁ ⊆ H)
Text 11.0.3: A hyperplane H is said to separate C₁ and C₂ strongly if there exists ε > 0 such that C₁ + ε B is contained in one of the open half-spaces associated with H and C₂ + ε B is contained in the opposite open half-space, where B is the unit Euclidean ball {x | ‖x‖ ≤ 1}. (Here Cᵢ + ε B consists of points x such that ‖x - y‖ ≤ ε for at least one y ∈ Cᵢ.)
open scoped Pointwise
def HyperplaneSeparatesStrongly (n : Nat) (H C₁ C₂ : Set (Fin n → Real)) : Prop
The set C₁ from Text 11.1.2 (a closed convex "hyperbola region").
open scoped Pointwise
def C11_1_2_C1 : Set (Fin 2 → Real) :=
{x | 0 ≤ x 0 ∧ (1 : Real) ≤ x 0 * x 1}
The set C₂ from Text 11.1.2 (a closed convex ray on the x-axis).
open scoped Pointwise
def C11_1_2_C2 : Set (Fin 2 → Real) :=
{x | 0 ≤ x 0 ∧ x 1 = 0}
Any point in C11_1_2_C1 has strictly positive first coordinate.
open scoped Pointwise
lemma C11_1_2_C1_pos {x : Fin 2 → Real} (hx : x ∈ C11_1_2_C1) : 0 < x 0
Any point in C11_1_2_C1 satisfies the equivalent inequality x0⁻¹ ≤ x1.
open scoped Pointwise
lemma inv_le_coord1_of_mem_C11_1_2_C1 {x : Fin 2 → Real} (hx : x ∈ C11_1_2_C1) :
(x 0)⁻¹ ≤ x 1
The sets C11_1_2_C1 and C11_1_2_C2 are closed and convex and disjoint.
open scoped Pointwise
lemma C11_1_2_closed_convex_disjoint :
IsClosed C11_1_2_C1 ∧
IsClosed C11_1_2_C2 ∧
Convex Real C11_1_2_C1 ∧
Convex Real C11_1_2_C2 ∧ Disjoint C11_1_2_C1 C11_1_2_C2
Any ε-thickening of C11_1_2_C1 and C11_1_2_C2 intersects (hence no strong separation).
open scoped Pointwise
lemma C11_1_2_thickenings_intersect :
∀ ε : Real,
ε > 0 →
let B : Set (Fin 2 → Real) := {x | ‖x‖ ≤ (1 : Real)}
∃ z, z ∈ C11_1_2_C1 + (ε • B) ∧ z ∈ C11_1_2_C2 + (ε • B)
The counterexample sets cannot be separated strongly by any hyperplane.
open scoped Pointwise
lemma C11_1_2_not_hyperplaneSeparatesStrongly :
¬ (∃ H, HyperplaneSeparatesStrongly 2 H C11_1_2_C1 C11_1_2_C2)
Text 11.1.2: Not every pair of disjoint closed convex sets can be separated strongly.
open scoped Pointwise
theorem exists_disjoint_closed_convex_not_hyperplaneSeparatesStrongly :
∃ (n : Nat) (C₁ C₂ : Set (Fin n → Real)),
IsClosed C₁ ∧
IsClosed C₂ ∧
Convex Real C₁ ∧
Convex Real C₂ ∧
Disjoint C₁ C₂ ∧ ¬ (∃ H, HyperplaneSeparatesStrongly n H C₁ C₂)
Text 11.0.4: Strict separation means that C₁ and C₂ are contained in opposing open half-spaces determined by a hyperplane H.
open scoped Pointwise
def HyperplaneSeparatesStrictly (n : Nat) (H C₁ C₂ : Set (Fin n → Real)) : Prop :=
C₁.Nonempty ∧
C₂.Nonempty ∧
∃ (b : Fin n → Real) (β : Real),
b ≠ 0 ∧
H = {x : Fin n → Real | x ⬝ᵥ b = β} ∧
((C₁ ⊆ {x : Fin n → Real | x ⬝ᵥ b < β} ∧ C₂ ⊆ {x : Fin n → Real | β < x ⬝ᵥ b}) ∨
(C₂ ⊆ {x : Fin n → Real | x ⬝ᵥ b < β} ∧ C₁ ⊆ {x : Fin n → Real | β < x ⬝ᵥ b}))
Put a properly separating hyperplane in a consistent "oriented" form: C₁ lies in {x | β ≤ x ⬝ᵥ b} and C₂ lies in {x | x ⬝ᵥ b ≤ β}.
open scoped Pointwise
lemma hyperplaneSeparatesProperly_oriented (n : Nat) (H C₁ C₂ : Set (Fin n → Real)) :
HyperplaneSeparatesProperly n H C₁ C₂ →
∃ (b : Fin n → Real) (β : Real),
b ≠ 0 ∧
H = {x : Fin n → Real | x ⬝ᵥ b = β} ∧
(∀ x ∈ C₁, β ≤ x ⬝ᵥ b) ∧
(∀ x ∈ C₂, x ⬝ᵥ b ≤ β) ∧ ¬ (C₁ ⊆ H ∧ C₂ ⊆ H)
If C₁ lies in {x | β ≤ x ⬝ᵥ b} and C₂ lies in {x | x ⬝ᵥ b ≤ β}, then the extended infimum of x ⬝ᵥ b over C₁ is at least the extended supremum over C₂.
open scoped Pointwise
lemma sInf_ge_sSup_of_pointwise_bounds_EReal {n : Nat} {C₁ C₂ : Set (Fin n → Real)}
(b : Fin n → Real) (β : Real)
(hC₁ : ∀ x ∈ C₁, β ≤ x ⬝ᵥ b) (hC₂ : ∀ x ∈ C₂, x ⬝ᵥ b ≤ β) :
sInf ((fun x : Fin n → Real => ((x ⬝ᵥ b : Real) : EReal)) '' C₁) ≥
sSup ((fun x : Fin n → Real => ((x ⬝ᵥ b : Real) : EReal)) '' C₂)
Properness forces strict separation between the extended sSup on C₁ and extended sInf on C₂ under the same half-space inclusions.
open scoped Pointwise
lemma sSup_gt_sInf_of_not_both_subset_levelset_EReal {n : Nat} {C₁ C₂ : Set (Fin n → Real)}
(hC₁ne : C₁.Nonempty) (hC₂ne : C₂.Nonempty) (b : Fin n → Real) (β : Real)
(hC₁ : ∀ x ∈ C₁, β ≤ x ⬝ᵥ b) (hC₂ : ∀ x ∈ C₂, x ⬝ᵥ b ≤ β)
(hnot : ¬ (C₁ ⊆ {x : Fin n → Real | x ⬝ᵥ b = β} ∧ C₂ ⊆ {x : Fin n → Real | x ⬝ᵥ b = β})) :
sSup ((fun x : Fin n → Real => ((x ⬝ᵥ b : Real) : EReal)) '' C₁) >
sInf ((fun x : Fin n → Real => ((x ⬝ᵥ b : Real) : EReal)) '' C₂)
Build a properly separating hyperplane from EReal inf/sup inequalities (Theorem 11.1).
open scoped Pointwise
lemma exists_hyperplaneSeparatesProperly_of_EReal_inf_sup_conditions (n : Nat)
(C₁ C₂ : Set (Fin n → Real)) (hC₁ : C₁.Nonempty) (hC₂ : C₂.Nonempty) (b : Fin n → Real)
(ha :
sInf ((fun x : Fin n → Real => ((x ⬝ᵥ b : Real) : EReal)) '' C₁) ≥
sSup ((fun x : Fin n → Real => ((x ⬝ᵥ b : Real) : EReal)) '' C₂))
(hb :
sSup ((fun x : Fin n → Real => ((x ⬝ᵥ b : Real) : EReal)) '' C₁) >
sInf ((fun x : Fin n → Real => ((x ⬝ᵥ b : Real) : EReal)) '' C₂)) :
∃ H, HyperplaneSeparatesProperly n H C₁ C₂
Theorem 11.1: Let C₁ and C₂ be non-empty sets in ℝ^n. There exists a hyperplane separating C₁ and C₂ properly if and only if there exists a vector b such that: (a) inf {⟪x, b⟫ | x ∈ C₁} ≥ sup {⟪x, b⟫ | x ∈ C₂}, (b) sup {⟪x, b⟫ | x ∈ C₁} > inf {⟪x, b⟫ | x ∈ C₂}. There exists a hyperplane separating C₁ and C₂ strongly if and only if there exists a vector b such that: (c) inf {⟪x, b⟫ | x ∈ C₁} > sup {⟪x, b⟫ | x ∈ C₂}.
open scoped Pointwise
theorem exists_hyperplaneSeparatesProperly_iff (n : Nat) (C₁ C₂ : Set (Fin n → Real))
(hC₁ : C₁.Nonempty) (hC₂ : C₂.Nonempty) :
(∃ H, HyperplaneSeparatesProperly n H C₁ C₂) ↔
∃ b : Fin n → Real,
sInf ((fun x : Fin n → Real => ((x ⬝ᵥ b : Real) : EReal)) '' C₁) ≥
sSup ((fun x : Fin n → Real => ((x ⬝ᵥ b : Real) : EReal)) '' C₂) ∧
sSup ((fun x : Fin n → Real => ((x ⬝ᵥ b : Real) : EReal)) '' C₁) >
sInf ((fun x : Fin n → Real => ((x ⬝ᵥ b : Real) : EReal)) '' C₂)
Theorem 11.1 (strong separation): Let C₁ and C₂ be non-empty sets in ℝ^n. There exists a hyperplane separating C₁ and C₂ strongly if and only if there exists a vector b such that inf {⟪x, b⟫ | x ∈ C₁} > sup {⟪x, b⟫ | x ∈ C₂}.
open scoped Pointwise
theorem exists_hyperplaneSeparatesStrongly_iff (n : Nat) (C₁ C₂ : Set (Fin n → Real))
(hC₁ : C₁.Nonempty) (hC₂ : C₂.Nonempty) :
(∃ H, HyperplaneSeparatesStrongly n H C₁ C₂) ↔
∃ b : Fin n → Real,
sInf ((fun x : Fin n → Real => ((x ⬝ᵥ b : Real) : EReal)) '' C₁) >
sSup ((fun x : Fin n → Real => ((x ⬝ᵥ b : Real) : EReal)) '' C₂)
The quotient map mkQ collapses an affine subspace to a singleton in the quotient.
open scoped Pointwise
lemma image_mkQ_affineSubspace_eq_singleton {n : Nat} (M : AffineSubspace ℝ (Fin n → ℝ))
{m0 : Fin n → ℝ} (hm0 : m0 ∈ M) :
(M.direction.mkQ '' (M : Set (Fin n → ℝ))) = {M.direction.mkQ m0}
Disjointness forces the distinguished quotient point mkQ m0 (with m0 ∈ M) to lie outside mkQ '' C.
open scoped Pointwise
lemma quotient_point_not_mem_image_of_disjoint {n : Nat} {C : Set (Fin n → ℝ)}
(M : AffineSubspace ℝ (Fin n → ℝ)) {m0 : Fin n → ℝ} (hm0 : m0 ∈ M)
(hCM : Disjoint C (M : Set (Fin n → ℝ))) :
M.direction.mkQ m0 ∉ (M.direction.mkQ '' C)
If x₀ lies outside the affine span of a nonempty set S, there is a continuous linear functional which is constant on affineSpan ℝ S and strictly separates x₀ from S.
open scoped Pointwise
lemma exists_strictSep_point_of_not_mem_affineSpan {E : Type*} [AddCommGroup E] [Module ℝ E]
[TopologicalSpace E] [IsModuleTopology ℝ E] {S : Set E} (hS : S.Nonempty) {x0 : E}
(hx0 : x0 ∉ affineSpan ℝ S) : ∃ f : StrongDual ℝ E, ∀ x ∈ S, f x < f x0
If C is relatively open in its affine span, then translating by a point c0 ∈ C yields an open subset of the direction submodule (affineSpan ℝ C).direction.
open scoped Pointwise
lemma isOpen_translate_in_direction_of_relOpen {E : Type*} [AddCommGroup E] [Module ℝ E]
[TopologicalSpace E] [IsModuleTopology ℝ E] {C : Set E}
(hCrelopen : ∃ U : Set E, IsOpen U ∧ C = U ∩ (affineSpan ℝ C : Set E)) {c0 : E} (hc0 : c0 ∈ C) :
IsOpen {v : (affineSpan ℝ C).direction | ((v : E) + c0) ∈ C}
If O is open in the direction submodule V0, then its image in the mapped direction Submodule.map π V0 is open. This is the key topological step in the hard case of Theorem 11.2.
open scoped Pointwise
lemma isOpen_image_direction_mkQ_of_isOpen {n : Nat} (C : Set (Fin n → ℝ))
(D : Submodule ℝ (Fin n → ℝ)) :
let Q := (Fin n → ℝ) ⧸ D
let π : (Fin n → ℝ) →ₗ[ℝ] Q := D.mkQ
let V0 : Submodule ℝ (Fin n → ℝ) := (affineSpan ℝ C).direction
let f0 : V0 →ₗ[ℝ] Q := π.comp V0.subtype
let V : Submodule ℝ Q := Submodule.map π V0
∀ O : Set V0, IsOpen O →
IsOpen ((fun v : V0 => (⟨f0 v, by exact ⟨(v : Fin n → ℝ), v.2, rfl⟩⟩ : V)) '' O)§11 — Separation Theorems — Part 2
Lean authors: Zichen Wang, Wanli Ma, Suwan Wu, Yunxi Duan, Zaiwen Wen.
Hard-case separation: in the quotient by a submodule, a convex relatively open set C can be strictly separated from a point in the affine span of its image but outside the image.
open scoped Pointwise
lemma exists_strictSep_point_of_mem_affineSpan_openIn_convex
(n : Nat) (C : Set (Fin n → ℝ)) (D : Submodule ℝ (Fin n → ℝ))
(hCne : C.Nonempty) (hCconv : Convex ℝ C)
(hCrelopen :
∃ U : Set (Fin n → ℝ), IsOpen U ∧ C = U ∩ (affineSpan ℝ C : Set (Fin n → ℝ)))
{m0 : Fin n → ℝ}
(hq0S : D.mkQ m0 ∉ (D.mkQ '' C)) (hq0A : D.mkQ m0 ∈ affineSpan ℝ (D.mkQ '' C)) :
∃ f : StrongDual ℝ ((Fin n → ℝ) ⧸ D), ∀ c ∈ C, f (D.mkQ c) < f (D.mkQ m0)
**Key separation lemma for Theorem 11.2** (to be filled): in the quotient by M.direction, the convex relatively-open set C can be strictly separated from the point corresponding to M.
open scoped Pointwise
lemma exists_strictSep_point_image_mkQ_of_disjoint_convex_relativelyOpen
(n : Nat) (C : Set (Fin n → ℝ)) (M : AffineSubspace ℝ (Fin n → ℝ))
(hCne : C.Nonempty) (hCconv : Convex ℝ C)
(hCrelopen :
∃ U : Set (Fin n → ℝ), IsOpen U ∧ C = U ∩ (affineSpan ℝ C : Set (Fin n → ℝ)))
{m0 : Fin n → ℝ} (hm0 : m0 ∈ M) (hCM : Disjoint C (M : Set (Fin n → ℝ))) :
∃ f : StrongDual ℝ ((Fin n → ℝ) ⧸ M.direction),
(∀ c ∈ C, f (M.direction.mkQ c) < f (M.direction.mkQ m0)) ∨
(∀ c ∈ C, f (M.direction.mkQ m0) < f (M.direction.mkQ c))
A continuous linear functional on ℝ^n is determined by its values on the coordinate vectors.
open scoped Pointwise
lemma strongDual_apply_eq_sum_mul_single {n : Nat} (g : StrongDual ℝ (Fin n → ℝ))
(x : Fin n → ℝ) :
g x =
∑ i : Fin n,
x i * g (Pi.single (M := fun _ : Fin n => ℝ) i (1 : ℝ))
Convert a continuous linear functional on ℝ^n into a dotProduct against a vector.
open scoped Pointwise
lemma strongDual_apply_eq_dotProduct {n : Nat} (g : StrongDual ℝ (Fin n → ℝ)) :
∀ x : Fin n → ℝ, g x = x ⬝ᵥ (fun i => g (Pi.single (M := fun _ : Fin n => ℝ) i (1 : ℝ)))
Theorem 11.2: Let C be a non-empty relatively open convex set in R^n, and let M be a non-empty affine set in R^n not meeting C. Then there exists a hyperplane H containing M, such that one of the open half-spaces associated with H contains C.
open scoped Pointwise
theorem exists_hyperplane_contains_affine_of_convex_relativelyOpen
(n : Nat) (C : Set (Fin n → Real)) (M : AffineSubspace Real (Fin n → Real))
(hCne : C.Nonempty) (hCconv : Convex Real C)
(hCrelopen :
∃ U : Set (Fin n → Real), IsOpen U ∧ C = U ∩ (affineSpan Real C : Set (Fin n → Real)))
(hMne : (M : Set (Fin n → Real)).Nonempty) (hCM : Disjoint C (M : Set (Fin n → Real))) :
∃ H : Set (Fin n → Real),
(M : Set (Fin n → Real)) ⊆ H ∧
∃ (b : Fin n → Real) (β : Real),
b ≠ 0 ∧
H = {x : Fin n → Real | x ⬝ᵥ b = β} ∧
(C ⊆ {x : Fin n → Real | x ⬝ᵥ b < β} ∨
C ⊆ {x : Fin n → Real | β < x ⬝ᵥ b})
Minkowski subtraction is Minkowski addition with pointwise negation.
open scoped Pointwise
lemma set_sub_eq_add_neg {E : Type*} [AddGroup E] (A B : Set E) : A - B = A + (-B)
Pointwise negation equals scalar multiplication by -1.
open scoped Pointwise
lemma neg_set_eq_smul {E : Type*} [AddCommGroup E] [Module Real E] (C : Set E) :
-C = ((-1 : Real) • C)
A continuous linear equivalence maps intrinsic interiors.
open scoped Pointwise
lemma ContinuousLinearEquiv.image_intrinsicInterior {𝕜 E F : Type*} [NontriviallyNormedField 𝕜]
[NormedAddCommGroup E] [NormedSpace 𝕜 E] [NormedAddCommGroup F] [NormedSpace 𝕜 F]
(e : E ≃L[𝕜] F) (s : Set E) :
intrinsicInterior 𝕜 (e '' s) = e '' intrinsicInterior 𝕜 s§11 — Separation Theorems — Part 3
Lean authors: Zichen Wang, Wanli Ma, Suwan Wu, Yunxi Duan, Zaiwen Wen.
The book's euclideanRelativeInterior is contained in mathlib's intrinsicInterior in Euclidean space.
open scoped Pointwise
lemma euclideanRelativeInterior_subset_intrinsicInterior (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) :
euclideanRelativeInterior n C ⊆ intrinsicInterior Real C§11 — Separation Theorems — Part 4
Lean authors: Zichen Wang, Wanli Ma, Suwan Wu, Yunxi Duan, Zaiwen Wen.
intrinsicInterior equals euclideanRelativeInterior in Euclidean space.
open scoped Pointwise
lemma intrinsicInterior_eq_euclideanRelativeInterior (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) :
intrinsicInterior Real C = euclideanRelativeInterior n C
Relative interior commutes with Minkowski subtraction for convex sets in Fin n → ℝ.
open scoped Pointwise
lemma intrinsicInterior_sub_eq (n : Nat) (C₁ C₂ : Set (Fin n → Real))
(hC₁ : Convex Real C₁) (hC₂ : Convex Real C₂) :
intrinsicInterior Real (C₁ - C₂) =
intrinsicInterior Real C₁ - intrinsicInterior Real C₂
Disjointness of intrinsic interiors is equivalent to 0 not lying in the intrinsic interior of the Minkowski difference.
open scoped Pointwise
lemma disjoint_intrinsicInterior_iff_zero_not_mem_intrinsicInterior_sub (n : Nat)
(C₁ C₂ : Set (Fin n → Real)) (hC₁ : Convex Real C₁) (hC₂ : Convex Real C₂) :
Disjoint (intrinsicInterior Real C₁) (intrinsicInterior Real C₂) ↔
(0 : Fin n → Real) ∉ intrinsicInterior Real (C₁ - C₂)
intrinsicInterior is relatively open in its affine span.
open scoped Pointwise
lemma exists_isOpen_inter_affineSpan_eq_intrinsicInterior (n : Nat) (s : Set (Fin n → Real)) :
∃ U : Set (Fin n → Real), IsOpen U ∧
intrinsicInterior Real s = U ∩ (affineSpan Real s : Set (Fin n → Real))
The intrinsic interior of a convex set in Fin n → ℝ is convex.
open scoped Pointwise
lemma convex_intrinsicInterior_of_convex (n : Nat) (C : Set (Fin n → Real))
(hC : Convex Real C) : Convex Real (intrinsicInterior Real C)
For a nonempty convex set in Fin n → ℝ, the affine span of the intrinsic interior equals the affine span of the set.
open scoped Pointwise
lemma affineSpan_intrinsicInterior_eq_of_convex_nonempty (n : Nat) (C : Set (Fin n → Real))
(hC : Convex Real C) (hCne : C.Nonempty) :
affineSpan Real (intrinsicInterior Real C) = affineSpan Real C
For a nonempty convex set in Fin n → ℝ, closure C ⊆ closure (intrinsicInterior C).
open scoped Pointwise
lemma closure_subset_closure_intrinsicInterior_of_convex_nonempty (n : Nat)
(C : Set (Fin n → Real)) (hC : Convex Real C) (hCne : C.Nonempty) :
closure C ⊆ closure (intrinsicInterior Real C)
Proper separation implies 0 is not in the intrinsic interior of the Minkowski difference.
open scoped Pointwise
lemma zero_not_mem_intrinsicInterior_sub_of_exists_hyperplaneSeparatesProperly (n : Nat)
(C₁ C₂ : Set (Fin n → Real)) (hC₁ne : C₁.Nonempty) (hC₂ne : C₂.Nonempty) :
(∃ H, HyperplaneSeparatesProperly n H C₁ C₂) →
(0 : Fin n → Real) ∉ intrinsicInterior Real (C₁ - C₂)
If 0 is not in the intrinsic interior of the Minkowski difference, then the sets admit a properly separating hyperplane.
open scoped Pointwise
lemma exists_hyperplaneSeparatesProperly_of_zero_not_mem_intrinsicInterior_sub (n : Nat)
(C₁ C₂ : Set (Fin n → Real)) (hC₁ne : C₁.Nonempty) (hC₂ne : C₂.Nonempty)
(hC₁conv : Convex Real C₁) (hC₂conv : Convex Real C₂) :
(0 : Fin n → Real) ∉ intrinsicInterior Real (C₁ - C₂) →
∃ H, HyperplaneSeparatesProperly n H C₁ C₂
Theorem 11.3: Let C₁ and C₂ be non-empty convex sets in ℝ^n. In order that there exist a hyperplane separating C₁ and C₂ properly, it is necessary and sufficient that ri C₁ and ri C₂ have no point in common, where ri denotes the relative (intrinsic) interior.
open scoped Pointwise
theorem exists_hyperplaneSeparatesProperly_iff_disjoint_intrinsicInterior (n : Nat)
(C₁ C₂ : Set (Fin n → Real)) (hC₁ne : C₁.Nonempty) (hC₂ne : C₂.Nonempty)
(hC₁conv : Convex Real C₁) (hC₂conv : Convex Real C₂) :
(∃ H, HyperplaneSeparatesProperly n H C₁ C₂) ↔
Disjoint (intrinsicInterior Real C₁) (intrinsicInterior Real C₂)§11 — Separation Theorems — Part 5
Lean authors: Zichen Wang, Wanli Ma, Suwan Wu, Yunxi Duan, Zaiwen Wen.
The closed unit ball {x | ‖x‖ ≤ 1} in Fin n → ℝ agrees with Metric.closedBall 0 1.
open scoped Pointwise
lemma unitBall_eq_closedBall (n : Nat) :
({x : Fin n → Real | ‖x‖ ≤ (1 : Real)} : Set (Fin n → Real)) =
Metric.closedBall (0 : Fin n → Real) 1
Scalar multiples of the unit ball are the corresponding centered closed balls (for r ≥ 0).
open scoped Pointwise
lemma smul_unitBall_eq_closedBall (n : Nat) {r : Real} (hr : 0 ≤ r) :
r • ({x : Fin n → Real | ‖x‖ ≤ (1 : Real)} : Set (Fin n → Real)) =
Metric.closedBall (0 : Fin n → Real) r
Given b ≠ 0, there is a unit-ball vector whose dot product with b is positive.
open scoped Pointwise
lemma exists_mem_unitBall_and_dotProduct_pos {n : Nat} {b : Fin n → Real} :
b ≠ 0 →
∃ v : Fin n → Real,
v ∈ ({x : Fin n → Real | ‖x‖ ≤ (1 : Real)} : Set (Fin n → Real)) ∧ 0 < v ⬝ᵥ b
If a δ-thickening of S is contained in a *closed* half-space {x | x ⬝ᵥ b ≤ β}, then S is contained in the corresponding *open* half-space {x | x ⬝ᵥ b < β} (for δ > 0).
open scoped Pointwise
lemma strictify_halfspace_le_of_thickening {n : Nat} {S : Set (Fin n → Real)}
{b : Fin n → Real} {β δ : Real} (hb0 : b ≠ 0) (hδ : 0 < δ)
(hS : S + (δ • ({x : Fin n → Real | ‖x‖ ≤ (1 : Real)} : Set (Fin n → Real))) ⊆
{x : Fin n → Real | x ⬝ᵥ b ≤ β}) :
S ⊆ {x : Fin n → Real | x ⬝ᵥ b < β}
If a δ-thickening of S is contained in a *closed* half-space {x | β ≤ x ⬝ᵥ b}, then S is contained in the corresponding *open* half-space {x | β < x ⬝ᵥ b} (for δ > 0).
open scoped Pointwise
lemma strictify_halfspace_ge_of_thickening {n : Nat} {S : Set (Fin n → Real)}
{b : Fin n → Real} {β δ : Real} (hb0 : b ≠ 0) (hδ : 0 < δ)
(hS : S + (δ • ({x : Fin n → Real | ‖x‖ ≤ (1 : Real)} : Set (Fin n → Real))) ⊆
{x : Fin n → Real | β ≤ x ⬝ᵥ b}) :
S ⊆ {x : Fin n → Real | β < x ⬝ᵥ b}
If ‖x₁ - x₂‖ ≤ 2ε, then the midpoint lies in both ε-thickenings of {x₁} and {x₂}.
open scoped Pointwise
lemma midpoint_mem_add_smul_unitBall_of_norm_sub_le {n : Nat} {x₁ x₂ : Fin n → Real} {ε : Real}
(hε : 0 ≤ ε) (h : ‖x₁ - x₂‖ ≤ 2 * ε) :
∃ z : Fin n → Real,
z ∈ ({x₁} + (ε • ({x : Fin n → Real | ‖x‖ ≤ (1 : Real)} : Set (Fin n → Real)))) ∧
z ∈ ({x₂} + (ε • ({x : Fin n → Real | ‖x‖ ≤ (1 : Real)} : Set (Fin n → Real))))
Disjointness of ε-thickenings is equivalent to positivity of the infimum of pairwise distances.
open scoped Pointwise
lemma exists_eps_disjoint_thickenings_iff_sInf_norm_sub_pos (n : Nat)
(C₁ C₂ : Set (Fin n → Real)) (hC₁ne : C₁.Nonempty) (hC₂ne : C₂.Nonempty) :
(∃ ε : Real, 0 < ε ∧
let B : Set (Fin n → Real) := {x : Fin n → Real | ‖x‖ ≤ (1 : Real)}
Disjoint (C₁ + (ε • B)) (C₂ + (ε • B))) ↔
0 < sInf (Set.image2 (fun x₁ x₂ : Fin n → Real => ‖x₁ - x₂‖) C₁ C₂)
Strong separation is equivalent to disjointness of some ε-thickenings.
open scoped Pointwise
lemma exists_hyperplaneSeparatesStrongly_iff_exists_eps_disjoint_thickenings (n : Nat)
(C₁ C₂ : Set (Fin n → Real)) (hC₁ne : C₁.Nonempty) (hC₂ne : C₂.Nonempty)
(hC₁conv : Convex Real C₁) (hC₂conv : Convex Real C₂) :
(∃ H, HyperplaneSeparatesStrongly n H C₁ C₂) ↔
∃ ε : Real, 0 < ε ∧
let B : Set (Fin n → Real) := {x : Fin n → Real | ‖x‖ ≤ (1 : Real)}
Disjoint (C₁ + (ε • B)) (C₂ + (ε • B))
The distance-set between C₁ and C₂ is the image of dist 0 on C₁ - C₂.
open scoped Pointwise
lemma image2_norm_sub_eq_image_dist_sub (n : Nat) (C₁ C₂ : Set (Fin n → Real)) :
Set.image2 (fun x₁ x₂ : Fin n → Real => ‖x₁ - x₂‖) C₁ C₂ =
(fun z : Fin n → Real => dist (0 : Fin n → Real) z) '' (C₁ - C₂)
Positivity of the infimum distance-set is equivalent to 0 ∉ closure (C₁ - C₂).
open scoped Pointwise
lemma sInf_norm_sub_pos_iff_zero_not_mem_closure_sub (n : Nat)
(C₁ C₂ : Set (Fin n → Real)) (hC₁ne : C₁.Nonempty) (hC₂ne : C₂.Nonempty) :
0 < sInf (Set.image2 (fun x₁ x₂ : Fin n → Real => ‖x₁ - x₂‖) C₁ C₂) ↔
(0 : Fin n → Real) ∉ closure (C₁ - C₂)
Theorem 11.4: Let C₁ and C₂ be non-empty convex sets in ℝ^n. In order that there exist a hyperplane separating C₁ and C₂ strongly, it is necessary and sufficient that inf { ‖x₁ - x₂‖ | x₁ ∈ C₁, x₂ ∈ C₂ } > 0, in other words that 0 ∉ closure (C₁ - C₂).
open scoped Pointwise
theorem exists_hyperplaneSeparatesStrongly_iff_inf_sub_norm_pos (n : Nat)
(C₁ C₂ : Set (Fin n → Real)) (hC₁ne : C₁.Nonempty) (hC₂ne : C₂.Nonempty)
(hC₁conv : Convex Real C₁) (hC₂conv : Convex Real C₂) :
(∃ H, HyperplaneSeparatesStrongly n H C₁ C₂) ↔
0 < sInf (Set.image2 (fun x₁ x₂ : Fin n → Real => ‖x₁ - x₂‖) C₁ C₂)
Theorem 11.4 (closure form): Let C₁ and C₂ be non-empty convex sets in ℝ^n. There exists a hyperplane separating C₁ and C₂ strongly if and only if 0 ∉ closure (C₁ - C₂).
open scoped Pointwise
theorem exists_hyperplaneSeparatesStrongly_iff_zero_not_mem_closure_sub (n : Nat)
(C₁ C₂ : Set (Fin n → Real)) (hC₁ne : C₁.Nonempty) (hC₂ne : C₂.Nonempty)
(hC₁conv : Convex Real C₁) (hC₂conv : Convex Real C₂) :
(∃ H, HyperplaneSeparatesStrongly n H C₁ C₂) ↔
(0 : Fin n → Real) ∉ closure (C₁ - C₂)
Disjointness implies 0 ∉ (C₁ - C₂).
open scoped Pointwise
lemma zero_not_mem_sub_of_disjoint {E : Type*} [AddGroup E] {C₁ C₂ : Set E}
(hdisj : Disjoint C₁ C₂) :
(0 : E) ∉ (C₁ - C₂)
Recession directions of a negated set are the negations of recession directions.
open scoped Pointwise
lemma recessionCone_neg_set_iff {E : Type*} [AddCommGroup E] [Module Real E] (C : Set E) (y : E) :
y ∈ Set.recessionCone (-C) ↔ (-y) ∈ Set.recessionCone C
Membership in a recession cone is preserved under preimages by linear equivalences.
open scoped Pointwise
lemma mem_recessionCone_preimage_linearEquiv_iff
{E F : Type*} [AddCommGroup E] [Module Real E] [AddCommGroup F] [Module Real F]
(φ : E ≃ₗ[Real] F) (C : Set F) (y : E) :
y ∈ Set.recessionCone (φ ⁻¹' C) ↔ φ y ∈ Set.recessionCone C
Under the recession-condition hypothesis, the Minkowski difference C₁ - C₂ is closed.
open scoped Pointwise
lemma isClosed_sub_of_noCommonRecessionDirections (n : Nat) (C₁ C₂ : Set (Fin n → Real))
(hC₁ne : C₁.Nonempty) (hC₂ne : C₂.Nonempty) (hC₁closed : IsClosed C₁) (hC₂closed : IsClosed C₂)
(hC₁conv : Convex Real C₁) (hC₂conv : Convex Real C₂)
(hrec : ∀ y, y ≠ (0 : Fin n → Real) → y ∈ Set.recessionCone C₁ → y ∈ Set.recessionCone C₂ →
False) :
IsClosed (C₁ - C₂)
Corollary 11.4.1. Let C₁ and C₂ be non-empty disjoint closed convex sets in ℝ^n having no common directions of recession. Then there exists a hyperplane separating C₁ and C₂ strongly.
open scoped Pointwise
theorem exists_hyperplaneSeparatesStrongly_of_disjoint_closed_convex_noCommonRecessionDirections
(n : Nat) (C₁ C₂ : Set (Fin n → Real)) (hC₁ne : C₁.Nonempty) (hC₂ne : C₂.Nonempty)
(hC₁closed : IsClosed C₁) (hC₂closed : IsClosed C₂) (hC₁conv : Convex Real C₁)
(hC₂conv : Convex Real C₂) (hdisj : Disjoint C₁ C₂)
(hrec : ∀ y, y ≠ (0 : Fin n → Real) → y ∈ Set.recessionCone C₁ → y ∈ Set.recessionCone C₂ →
False) :
∃ H, HyperplaneSeparatesStrongly n H C₁ C₂
Bounded sets in Fin n → ℝ have only the zero recession direction.
open scoped Pointwise
lemma recessionCone_subset_singleton_of_bounded_fin {n : Nat} {C : Set (Fin n → Real)}
(hCne : C.Nonempty) (hCbdd : Bornology.IsBounded C) :
Set.recessionCone C ⊆ {0}
If one of two sets is bounded, then their closures have no common nonzero recession direction.
open scoped Pointwise
lemma noCommonRecessionDirections_closure_of_bounded_side {n : Nat} {C₁ C₂ : Set (Fin n → Real)}
(hC₁ne : C₁.Nonempty) (hC₂ne : C₂.Nonempty)
(hbounded : Bornology.IsBounded C₁ ∨ Bornology.IsBounded C₂) :
∀ y, y ≠ (0 : Fin n → Real) → y ∈ Set.recessionCone (closure C₁) →
y ∈ Set.recessionCone (closure C₂) → False
Strong hyperplane separation is monotone under shrinking the separated sets.
open scoped Pointwise
lemma hyperplaneSeparatesStrongly_mono_sets {n : Nat} {H A₁ A₂ B₁ B₂ : Set (Fin n → Real)}
(hH : HyperplaneSeparatesStrongly n H A₁ A₂) (hB₁ : B₁ ⊆ A₁) (hB₂ : B₂ ⊆ A₂)
(hB₁ne : B₁.Nonempty) (hB₂ne : B₂.Nonempty) :
HyperplaneSeparatesStrongly n H B₁ B₂
Corollary 11.4.2. Let C₁ and C₂ be non-empty convex sets in ℝ^n whose closures are disjoint. If either set is bounded, there exists a hyperplane separating C₁ and C₂ strongly.
open scoped Pointwise
theorem exists_hyperplaneSeparatesStrongly_of_disjoint_closure_convex_of_bounded (n : Nat)
(C₁ C₂ : Set (Fin n → Real)) (hC₁ne : C₁.Nonempty) (hC₂ne : C₂.Nonempty)
(hC₁conv : Convex Real C₁) (hC₂conv : Convex Real C₂)
(hdisj : Disjoint (closure C₁) (closure C₂))
(hbounded : Bornology.IsBounded C₁ ∨ Bornology.IsBounded C₂) :
∃ H, HyperplaneSeparatesStrongly n H C₁ C₂
A single dot-product inequality defines a closed half-space.
open scoped Pointwise
lemma isClosed_setOf_dotProduct_le {n : Nat} (b : Fin n → Real) (β : Real) :
IsClosed {x : Fin n → Real | x ⬝ᵥ b ≤ β}
A single dot-product inequality defines a convex half-space.
open scoped Pointwise
lemma convex_setOf_dotProduct_le {n : Nat} (b : Fin n → Real) (β : Real) :
Convex Real {x : Fin n → Real | x ⬝ᵥ b ≤ β}
A system of dot-product inequalities is the intersection of the corresponding half-spaces.
open scoped Pointwise
lemma setOf_forall_dotProduct_le_eq_iInter (n : Nat) (I : Type*) (b : I → Fin n → Real)
(β : I → Real) :
{x : Fin n → Real | ∀ i : I, x ⬝ᵥ b i ≤ β i} =
⋂ i : I, {x : Fin n → Real | x ⬝ᵥ b i ≤ β i}
Text 11.2.1: The solution set S of a system of weak linear inequalities ⟪x, b i⟫ ≤ β i (i.e. x ⬝ᵥ b i ≤ β i, for all i ∈ I) is a closed convex set.
open scoped Pointwise
theorem isClosed_and_convex_setOf_dotProduct_le (n : Nat) (I : Type*) (b : I → Fin n → Real)
(β : I → Real) :
IsClosed {x : Fin n → Real | ∀ i : I, x ⬝ᵥ b i ≤ β i} ∧
Convex Real {x : Fin n → Real | ∀ i : I, x ⬝ᵥ b i ≤ β i}
Text 11.3.1: Let C be a convex set in ℝ^n. A supporting half-space to C is a closed half-space which contains C and has a point of C in its boundary.
open scoped Pointwise
def IsClosedHalfspace (n : Nat) (H : Set (Fin n → Real)) : Prop :=
∃ (b : Fin n → Real) (β : Real),
b ≠ 0 ∧ (H = {x : Fin n → Real | x ⬝ᵥ b ≤ β} ∨ H = {x : Fin n → Real | β ≤ x ⬝ᵥ b})
Text 11.3.1: Let C be a convex set in ℝ^n. A supporting half-space to C is a closed half-space which contains C and has a point of C in its boundary.
open scoped Pointwise
def IsSupportingHalfspace (n : Nat) (C H : Set (Fin n → Real)) : Prop :=
Convex Real C ∧ IsClosedHalfspace n H ∧ C ⊆ H ∧ ∃ x, x ∈ C ∧ x ∈ frontier H
Text 11.3.2: A supporting hyperplane to C is a hyperplane which is the boundary of a supporting half-space to C. Equivalently, a supporting hyperplane has the form H = {x | x ⬝ᵥ b = β} with b ≠ 0, such that x ⬝ᵥ b ≤ β for every x ∈ C and x ⬝ᵥ b = β for at least one x ∈ C.
open scoped Pointwise
def IsSupportingHyperplane (n : Nat) (C H : Set (Fin n → Real)) : Prop :=
∃ (b : Fin n → Real) (β : Real),
b ≠ 0 ∧
H = {x : Fin n → Real | x ⬝ᵥ b = β} ∧
(∀ x, x ∈ C → x ⬝ᵥ b ≤ β) ∧ ∃ x, x ∈ C ∧ x ⬝ᵥ b = β
If a nonempty affine subspace is not ⊤, then its direction is a proper submodule.
open scoped Pointwise
lemma direction_ne_top_of_ne_top_of_nonempty {n : Nat} {A : AffineSubspace ℝ (Fin n → Real)}
(hAne : (A : Set (Fin n → Real)).Nonempty) (hA : A ≠ ⊤) :
A.direction ≠ ⊤
If a linear functional vanishes on the direction of an affine subspace, then it is constant on that affine subspace.
open scoped Pointwise
lemma affineSubspace_subset_setOf_apply_eq_of_le_ker_direction
{V : Type*} [AddCommGroup V] [Module ℝ V] (A : AffineSubspace ℝ V) {x0 : V} (hx0 : x0 ∈ A)
(f : Module.Dual ℝ V) (hdir : A.direction ≤ LinearMap.ker f) :
(A : Set V) ⊆ {x : V | f x = f x0}
In Fin n → ℝ with 0 < n, one can choose an explicit nonzero vector.
open scoped Pointwise
lemma empty_case_choose_nonzero_b {n : Nat} (hn : 0 < n) :
∃ b : Fin n → Real, b ≠ 0
Text 11.3.3: If C is not n-dimensional (so that affineSpan ℝ C ≠ ⊤), then there exists a hyperplane containing (and hence extending) affineSpan ℝ C, and therefore containing all of C.
open scoped Pointwise
theorem exists_hyperplane_superset_affineSpan_of_affineSpan_ne_top (n : Nat) (hn : 0 < n)
(C : Set (Fin n → Real))
(hC : affineSpan ℝ C ≠ (⊤ : AffineSubspace ℝ (Fin n → Real))) :
∃ (b : Fin n → Real) (β : Real),
b ≠ 0 ∧
C ⊆ {x : Fin n → Real | x ⬝ᵥ b = β} ∧
(affineSpan ℝ C : Set (Fin n → Real)) ⊆ {x : Fin n → Real | x ⬝ᵥ b = β}
Text 11.3.4: We usually speak only of non-trivial supporting hyperplanes to C, i.e. those supporting hyperplanes H which do not contain C itself.
open scoped Pointwise
def IsNontrivialSupportingHyperplane (n : Nat) (C H : Set (Fin n → Real)) : Prop :=
IsSupportingHyperplane n C H ∧ ¬ C ⊆ H
The dot product with b as a continuous linear functional, viewed in the StrongDual.
open scoped Pointwise
noncomputable def dotProduct_strongDual {n : Nat} (b : Fin n → Real) :
StrongDual ℝ (Fin n → Real) :=
LinearMap.toContinuousLinearMap
(IsLinearMap.mk' (fun x : Fin n → Real => x ⬝ᵥ b) (isLinearMap_dotProduct_left (n := n) b))
open scoped Pointwise
@[simp] lemma dotProduct_strongDual_apply {n : Nat} (b x : Fin n → Real) :
dotProduct_strongDual (n := n) b x = x ⬝ᵥ b
If a ∉ C and C is a nonempty closed convex set in ℝ^n, then there exists a continuous linear functional that strictly separates a from C.
open scoped Pointwise
lemma exists_strongDual_strict_sep_point_of_not_mem_isClosed_convex {n : Nat}
{a : Fin n → Real} {C : Set (Fin n → Real)} (hCconv : Convex Real C) (hCclosed : IsClosed C)
(hCne : C.Nonempty) (ha : a ∉ C) :
∃ l : StrongDual ℝ (Fin n → Real), ∀ y ∈ C, l y < l a§11 — Separation Theorems — Part 6
Lean authors: Zichen Wang, Wanli Ma, Suwan Wu, Yunxi Duan, Zaiwen Wen.
Theorem 11.5: A closed convex set C is the intersection of the closed half-spaces which contain it. In mathlib, a convenient way to express the closed half-spaces containing C is to quantify over continuous linear functionals l : StrongDual ℝ (Fin n → ℝ) and consider the half-space {x | ∃ y ∈ C, l x ≤ l y}.
open scoped Pointwise
theorem isClosed_convex_eq_iInter_halfSpaces (n : Nat) (C : Set (Fin n → Real))
(hCconv : Convex Real C) (hCclosed : IsClosed C) :
(⋂ l : StrongDual ℝ (Fin n → Real), {x : Fin n → Real | ∃ y ∈ C, l x ≤ l y}) = C
closure (convexHull ℝ S) is a closed set.
open scoped Pointwise
lemma closure_convexHull_isClosed (n : Nat) (S : Set (Fin n → Real)) :
IsClosed (closure (convexHull ℝ S))
closure (convexHull ℝ S) is a convex set.
open scoped Pointwise
lemma closure_convexHull_convex (n : Nat) (S : Set (Fin n → Real)) :
Convex Real (closure (convexHull ℝ S))
Corollary 11.5.1. Let S be any subset of ℝ^n. Then cl (conv S) (i.e. closure (convexHull ℝ S)) is the intersection of all the closed half-spaces containing S.
open scoped Pointwise
theorem closure_convexHull_eq_iInter_halfSpaces (n : Nat) (S : Set (Fin n → Real)) :
closure (convexHull ℝ S) =
⋂ l : StrongDual ℝ (Fin n → Real),
{x : Fin n → Real | ∃ y ∈ closure (convexHull ℝ S), l x ≤ l y}§11 — Separation Theorems — Part 7
Lean authors: Zichen Wang, Wanli Ma, Suwan Wu, Yunxi Duan, Zaiwen Wen.
If a ∉ C, then {a} and C have disjoint intrinsic interiors.
open scoped Pointwise
lemma cor11_5_2_disjoint_intrinsicInterior_singleton {n : Nat} {a : Fin n → Real}
{C : Set (Fin n → Real)} (ha : a ∉ C) :
Disjoint (intrinsicInterior Real ({a} : Set (Fin n → Real))) (intrinsicInterior Real C)
A point outside a nonempty convex set can be properly separated from it by a hyperplane.
open scoped Pointwise
lemma cor11_5_2_exists_hyperplaneSeparatesProperly_point {n : Nat} {C : Set (Fin n → Real)}
{a : Fin n → Real} (hCne : C.Nonempty) (hCconv : Convex Real C) (ha : a ∉ C) :
∃ H, HyperplaneSeparatesProperly n H ({a} : Set (Fin n → Real)) C
From a proper separating hyperplane between {a} and C, extract a nontrivial closed half-space of the form {x | x ⬝ᵥ b ≤ β} containing C.
open scoped Pointwise
lemma cor11_5_2_extract_closedHalfspace_from_separation {n : Nat} {C : Set (Fin n → Real)}
{a : Fin n → Real} {H : Set (Fin n → Real)}
(hsep : HyperplaneSeparatesProperly n H ({a} : Set (Fin n → Real)) C) :
∃ (b : Fin n → Real) (β : Real), b ≠ 0 ∧ C ⊆ {x : Fin n → Real | x ⬝ᵥ b ≤ β}
Corollary 11.5.2. Let C be a convex subset of ℝ^n other than ℝ^n itself. Then there exists a closed half-space containing C. In other words, there exists some b ∈ ℝ^n such that the linear function x ↦ ⟪x, b⟫ (i.e. x ⬝ᵥ b) is bounded above on C.
open scoped Pointwise
theorem exists_closedHalfspace_superset_of_convex_ne_univ (n : Nat) (C : Set (Fin n → Real))
(hn : 0 < n) (hCconv : Convex Real C) (hCne : C ≠ (Set.univ : Set (Fin n → Real))) :
∃ (b : Fin n → Real) (β : Real), b ≠ 0 ∧ C ⊆ {x : Fin n → Real | x ⬝ᵥ b ≤ β}
If D ⊆ C, a nontrivial supporting hyperplane to C containing D is the same as a properly separating hyperplane between D and C.
open scoped Pointwise
lemma exists_nontrivialSupportingHyperplane_containing_iff_exists_hyperplaneSeparatesProperly
{n : Nat} {C D : Set (Fin n → Real)} (hDne : D.Nonempty) (hDC : D ⊆ C) :
(∃ H, IsNontrivialSupportingHyperplane n C H ∧ D ⊆ H) ↔
(∃ H, HyperplaneSeparatesProperly n H D C)
If D ⊆ C, then ri D is disjoint from ri C iff D is disjoint from ri C.
open scoped Pointwise
lemma disjoint_intrinsicInterior_left_iff_disjoint_of_subset {n : Nat} {C D : Set (Fin n → Real)}
(hCconv : Convex Real C) (hDne : D.Nonempty) (hDconv : Convex Real D) (hDC : D ⊆ C) :
Disjoint (intrinsicInterior ℝ D) (intrinsicInterior ℝ C) ↔ Disjoint D (intrinsicInterior ℝ C)
Theorem 11.6: Let C be a convex set, and let D be a non-empty convex subset of C (for instance, a subset consisting of a single point). In order that there exist a non-trivial supporting hyperplane to C containing D, it is necessary and sufficient that D be disjoint from ri C (the relative/intrinsic interior of C).
open scoped Pointwise
theorem exists_nontrivialSupportingHyperplane_containing_iff_disjoint_intrinsicInterior (n : Nat)
(C D : Set (Fin n → Real)) (hCconv : Convex Real C) (hDne : D.Nonempty) (hDconv : Convex Real D)
(hDC : D ⊆ C) :
(∃ H, IsNontrivialSupportingHyperplane n C H ∧ D ⊆ H) ↔
Disjoint D (intrinsicInterior ℝ C)§11 — Separation Theorems — Part 8
Lean authors: Zichen Wang, Wanli Ma, Suwan Wu, Yunxi Duan, Zaiwen Wen.
If affineSpan ℝ C = ⊤, then intrinsicInterior ℝ C agrees with the usual interior C.
open scoped Pointwise
lemma cor11_6_1_intrinsicInterior_eq_interior_of_affineSpan_eq_top {n : Nat}
(C : Set (Fin n → Real)) (hspan : affineSpan ℝ C = ⊤) :
intrinsicInterior ℝ C = interior C
If affineSpan ℝ C ≠ ⊤ and 0 < n, then every point of C admits a nonzero normal vector coming from a hyperplane containing C.
open scoped Pointwise
lemma cor11_6_1_exists_normal_of_affineSpan_ne_top {n : Nat} (hn : 0 < n)
{C : Set (Fin n → Real)} (hspan : affineSpan ℝ C ≠ ⊤) :
∀ x, x ∈ C → ∃ b : Fin n → Real, b ≠ 0 ∧ ∀ y, y ∈ C → y ⬝ᵥ b ≤ x ⬝ᵥ b
If affineSpan ℝ C = ⊤, then a point in the (ambient) boundary of C is disjoint from the intrinsic interior of C (as a singleton set).
open scoped Pointwise
lemma cor11_6_1_disjoint_singleton_intrinsicInterior_of_mem_frontier_of_affineSpan_eq_top
{n : Nat} {C : Set (Fin n → Real)} (hspan : affineSpan ℝ C = ⊤) {x : Fin n → Real}
(hxfr : x ∈ frontier C) :
Disjoint ({x} : Set (Fin n → Real)) (intrinsicInterior ℝ C)
From Theorem 11.6 (with D = {x}), extract a nonzero normal vector to C at x.
open scoped Pointwise
lemma cor11_6_1_exists_normal_of_disjoint_singleton_intrinsicInterior {n : Nat}
{C : Set (Fin n → Real)} (hC : Convex Real C) {x : Fin n → Real} (hxC : x ∈ C)
(hdisj : Disjoint ({x} : Set (Fin n → Real)) (intrinsicInterior ℝ C)) :
∃ b : Fin n → Real, b ≠ 0 ∧ ∀ y, y ∈ C → y ⬝ᵥ b ≤ x ⬝ᵥ b
Corollary 11.6.1. A convex set has a non-zero normal at each of its boundary points.
open scoped Pointwise
theorem exists_nonzero_normal_of_mem_frontier_of_convex (n : Nat) (C : Set (Fin n → Real))
(hC : Convex Real C) :
∀ x, x ∈ C → x ∈ frontier C → ∃ b : Fin n → Real, b ≠ 0 ∧ ∀ y, y ∈ C → y ⬝ᵥ b ≤ x ⬝ᵥ b
If x ∈ intrinsicFrontier ℝ C, then the singleton {x} is disjoint from intrinsicInterior ℝ C.
open scoped Pointwise
lemma cor11_6_2_disjoint_singleton_intrinsicInterior_of_mem_intrinsicFrontier {n : Nat}
{C : Set (Fin n → Real)} {x : Fin n → Real} (hx : x ∈ intrinsicFrontier ℝ C) :
Disjoint ({x} : Set (Fin n → Real)) (intrinsicInterior ℝ C)
A nontrivial supporting hyperplane gives a point of C where the defining linear functional is strictly smaller than its boundary value.
open scoped Pointwise
lemma cor11_6_2_exists_lt_of_isSupportingHyperplane_of_not_subset {n : Nat}
{C H : Set (Fin n → Real)} (hH : IsSupportingHyperplane n C H) (hCnot : ¬ C ⊆ H) :
∃ (b : Fin n → Real) (β : Real),
b ≠ 0 ∧ H = {x : Fin n → Real | x ⬝ᵥ b = β} ∧ (∀ x, x ∈ C → x ⬝ᵥ b ≤ β) ∧
∃ y0, y0 ∈ C ∧ y0 ⬝ᵥ b < β
If a linear functional attains a strict maximum over C at x, then x is not in the intrinsic interior of C.
open scoped Pointwise
lemma cor11_6_2_not_mem_intrinsicInterior_of_isMaxOn_of_exists_lt {n : Nat}
{C : Set (Fin n → Real)} {x : Fin n → Real} {l : StrongDual ℝ (Fin n → Real)}
(hexists : ∃ y, y ∈ C ∧ l y < l x) (hxmax : IsMaxOn (fun y => l y) C x) :
x ∉ intrinsicInterior ℝ C
Corollary 11.6.2. Let C be a convex set. An x ∈ C is a relative boundary point of C if and only if there exists a linear function h not constant on C such that h achieves its maximum over C at x. (Here "relative boundary" is interpreted as membership in intrinsicFrontier ℝ C.)
open scoped Pointwise
theorem mem_intrinsicFrontier_iff_exists_isMaxOn_linearFunctional (n : Nat)
(C : Set (Fin n → Real)) (hC : Convex Real C) (x : Fin n → Real) :
x ∈ C ∧ x ∈ intrinsicFrontier ℝ C ↔
∃ l : StrongDual ℝ (Fin n → Real),
(∃ y, y ∈ C ∧ l y < l x) ∧ IsMaxOn (fun y => l y) C x ∧ x ∈ C
HyperplaneSeparatesProperly is symmetric in the two sets.
open scoped Pointwise
lemma hyperplaneSeparatesProperly_comm {n : Nat} {H C₁ C₂ : Set (Fin n → Real)} :
HyperplaneSeparatesProperly n H C₁ C₂ → HyperplaneSeparatesProperly n H C₂ C₁
If x ⬝ᵥ b ≤ β for all x ∈ C, then ((fun x ↦ x ⬝ᵥ b) '' C) is bounded above.
open scoped Pointwise
lemma bddAbove_image_dotProduct_of_forall_le {n : Nat} {C : Set (Fin n → Real)}
(b : Fin n → Real) (β : Real) (h : ∀ x, x ∈ C → x ⬝ᵥ b ≤ β) :
BddAbove ((fun x : Fin n → Real => x ⬝ᵥ b) '' C)
On a cone, a dot-product bounded above must be nonpositive everywhere.
open scoped Pointwise
lemma thm11_7_dotProduct_le_zero_of_isConeSet_of_bddAbove {n : Nat} {C : Set (Fin n → Real)}
(hcone : IsConeSet n C) {b : Fin n → Real}
(hbdd : BddAbove ((fun x : Fin n → Real => x ⬝ᵥ b) '' C)) :
∀ x, x ∈ C → x ⬝ᵥ b ≤ 0
For a nonempty cone, a bounded-above dot-product has supremum 0.
open scoped Pointwise
lemma thm11_7_sSup_image_dotProduct_eq_zero_of_isConeSet {n : Nat} {C : Set (Fin n → Real)}
(hCne : C.Nonempty) (hcone : IsConeSet n C) (b : Fin n → Real)
(hbdd : BddAbove ((fun x : Fin n → Real => x ⬝ᵥ b) '' C)) :
sSup ((fun x : Fin n → Real => x ⬝ᵥ b) '' C) = (0 : Real)
Theorem 11.7. Let C₁ and C₂ be non-empty subsets of ℝ^n, at least one of which is a cone. If there exists a hyperplane which separates C₁ and C₂ properly, then there exists a hyperplane which separates C₁ and C₂ properly and passes through the origin.
open scoped Pointwise
theorem exists_hyperplaneSeparatesProperly_through_origin_of_isConeSet (n : Nat)
(C₁ C₂ : Set (Fin n → Real)) (hC₁ne : C₁.Nonempty) (hC₂ne : C₂.Nonempty)
(hcone : IsConeSet n C₁ ∨ IsConeSet n C₂) :
(∃ H, HyperplaneSeparatesProperly n H C₁ C₂) →
∃ H, HyperplaneSeparatesProperly n H C₁ C₂ ∧ (0 : Fin n → Real) ∈ H
If a ∉ K for a nonempty closed convex set K, then there is a dot-product functional strictly separating a from K in the sense x ⬝ᵥ b ≤ β < a ⬝ᵥ b.
open scoped Pointwise
lemma cor11_7_1_exists_strict_dotProduct_separator_of_not_mem {n : Nat}
{K : Set (Fin n → Real)} (hKne : K.Nonempty) (hKclosed : IsClosed K) (hKconv : Convex Real K)
{a : Fin n → Real} (ha : a ∉ K) :
∃ (b : Fin n → Real) (β : Real),
b ≠ 0 ∧ (∀ x ∈ K, x ⬝ᵥ b ≤ β) ∧ β < a ⬝ᵥ b
If a ∉ K for a nonempty closed convex cone K, then some homogeneous closed half-space contains K but excludes a.
open scoped Pointwise
lemma cor11_7_1_exists_homogeneous_halfspace_superset_excluding_of_not_mem {n : Nat}
{K : Set (Fin n → Real)} (hKne : K.Nonempty) (hKclosed : IsClosed K) (hK : IsConvexCone n K)
{a : Fin n → Real} (ha : a ∉ K) :
∃ b : Fin n → Real, b ≠ 0 ∧ K ⊆ {x : Fin n → Real | x ⬝ᵥ b ≤ 0} ∧ 0 < a ⬝ᵥ b
Corollary 11.7.1: A non-empty closed convex cone K ⊆ ℝ^n is the intersection of the homogeneous closed half-spaces which contain it (a homogeneous half-space being one with the origin on its boundary).
open scoped Pointwise
theorem isClosed_convexCone_eq_iInter_homogeneous_closedHalfspaces (n : Nat)
(K : Set (Fin n → Real)) (hKne : K.Nonempty) (hKclosed : IsClosed K) (hK : IsConvexCone n K) :
(⋂ (b : Fin n → Real) (_hb : b ≠ 0)
(_hKb : K ⊆ {x : Fin n → Real | x ⬝ᵥ b ≤ 0}), {x : Fin n → Real | x ⬝ᵥ b ≤ 0}) = K
The closure of a cone set (closed under positive scalar multiplication) is again a cone set.
open scoped Pointwise
lemma cor11_7_2_isConeSet_closure {n : Nat} {K : Set (Fin n → Real)} (hK : IsConeSet n K) :
IsConeSet n (closure K)
A homogeneous closed half-space {x | x ⬝ᵥ b ≤ 0} is closed.
open scoped Pointwise
lemma cor11_7_2_isClosed_homogeneousClosedHalfspace {n : Nat} (b : Fin n → Real) :
IsClosed {x : Fin n → Real | x ⬝ᵥ b ≤ 0}
The homogeneous closed half-space {x | x ⬝ᵥ b ≤ 0} as a bundled ConvexCone.
open scoped Pointwise
def cor11_7_2_homogeneousClosedHalfspaceCone {n : Nat} (b : Fin n → Real) :
ConvexCone Real (Fin n → Real)
If a homogeneous closed half-space {x | x ⬝ᵥ b ≤ 0} contains S, then it contains the closure of the convex cone hull of S.
open scoped Pointwise
lemma cor11_7_2_closure_hull_subset_halfspace_of_subset {n : Nat} (S : Set (Fin n → Real))
(b : Fin n → Real) (hSb : S ⊆ {x : Fin n → Real | x ⬝ᵥ b ≤ 0}) :
closure (ConvexCone.hull Real S : Set (Fin n → Real)) ⊆ {x : Fin n → Real | x ⬝ᵥ b ≤ 0}
For K = closure (ConvexCone.hull ℝ S), intersecting all homogeneous closed half-spaces containing S is the same as intersecting those containing K.
open scoped Pointwise
lemma cor11_7_2_intersections_over_S_vs_over_closure_hull (n : Nat) (S : Set (Fin n → Real)) :
(⋂ (b : Fin n → Real) (_hb : b ≠ 0)
(_hSb : S ⊆ {x : Fin n → Real | x ⬝ᵥ b ≤ 0}), {x : Fin n → Real | x ⬝ᵥ b ≤ 0}) =
⋂ (b : Fin n → Real) (_hb : b ≠ 0)
(_hKb : closure (ConvexCone.hull Real S : Set (Fin n → Real)) ⊆
{x : Fin n → Real | x ⬝ᵥ b ≤ 0}),
{x : Fin n → Real | x ⬝ᵥ b ≤ 0}
Corollary 11.7.2: Let S be any subset of ℝ^n, and let K be the closure of the convex cone generated by S. Then K is the intersection of all homogeneous closed half-spaces containing S (a homogeneous closed half-space is one of the form {x | x ⬝ᵥ b ≤ 0} with b ≠ 0).
open scoped Pointwise
theorem closure_convexConeHull_eq_iInter_homogeneous_closedHalfspaces (n : Nat)
(S : Set (Fin n → Real)) (hSne : S.Nonempty) :
closure (ConvexCone.hull Real S : Set (Fin n → Real)) =
⋂ (b : Fin n → Real) (_hb : b ≠ 0)
(_hSb : S ⊆ {x : Fin n → Real | x ⬝ᵥ b ≤ 0}), {x : Fin n → Real | x ⬝ᵥ b ≤ 0}
In positive dimension, there exists a nonzero vector in ℝ^n (as functions Fin n → ℝ).
open scoped Pointwise
lemma cor11_7_3_exists_nonzero_vector_of_pos {n : Nat} (hn : 0 < n) :
∃ b : Fin n → Real, b ≠ 0
If K = ∅ in positive dimension, then K lies in a homogeneous closed half-space {x | x ⬝ᵥ b ≤ 0} with b ≠ 0.
open scoped Pointwise
lemma cor11_7_3_exists_subset_halfspace_of_empty {n : Nat} (hn : 0 < n) (K : Set (Fin n → Real))
(hKempty : K = ∅) :
∃ b : Fin n → Real, b ≠ 0 ∧ K ⊆ {x : Fin n → Real | x ⬝ᵥ b ≤ 0}
From a proper separating hyperplane through the origin between {a} and K, extract a nonzero normal b for a homogeneous closed half-space containing K.
open scoped Pointwise
lemma cor11_7_3_extract_homogeneous_halfspace_of_sep_through_origin {n : Nat}
{K : Set (Fin n → Real)} {a : Fin n → Real} {H : Set (Fin n → Real)}
(hsep : HyperplaneSeparatesProperly n H ({a} : Set (Fin n → Real)) K)
(h0 : (0 : Fin n → Real) ∈ H) :
∃ b : Fin n → Real, b ≠ 0 ∧ K ⊆ {x : Fin n → Real | x ⬝ᵥ b ≤ 0}
Corollary 11.7.3: Let K be a convex cone in ℝ^n other than ℝ^n itself. Then K is contained in some homogeneous closed half-space of ℝ^n. In other words, there exists some vector b ≠ 0 such that x ⬝ᵥ b ≤ 0 for every x ∈ K.
open scoped Pointwise
theorem exists_subset_homogeneous_closedHalfspace_of_isConvexCone_ne_univ (n : Nat) (hn : 0 < n)
(K : Set (Fin n → Real)) (hK : IsConvexCone n K) (hKne : K ≠ Set.univ) :
∃ b : Fin n → Real, b ≠ 0 ∧ K ⊆ {x : Fin n → Real | x ⬝ᵥ b ≤ 0}§12 — Conjugates of Convex Functions — Part 1
Lean authors: Zichen Wang, Wanli Ma, Zhiyan Wang, Zaiwen Wen.
Defn 12.1: A closed half-space in ℝ^(n+1) (written with coordinates (x, μ) where x ∈ ℝ^n and μ ∈ ℝ) is classified into three types by its defining inequality: 1. (Vertical) a set {(x, μ) | ⟪x, b⟫ ≤ β} with b ≠ 0; 2. (Upper) a set {(x, μ) | μ ≥ ⟪x, b⟫ - β} (the epigraph of an affine function); 3. (Lower) a set {(x, μ) | μ ≤ ⟪x, b⟫ - β}.
def IsVerticalHalfSpace (n : Nat) (H : Set ((Fin n → Real) × Real)) : Prop :=
∃ (b : Fin n → Real) (β : Real),
b ≠ 0 ∧ H = {p : (Fin n → Real) × Real | p.1 ⬝ᵥ b ≤ β}
Defn 12.1 (Upper): H ⊆ ℝ^n × ℝ has the form {(x, μ) | μ ≥ ⟪x, b⟫ - β}.
def IsUpperHalfSpace (n : Nat) (H : Set ((Fin n → Real) × Real)) : Prop :=
∃ (b : Fin n → Real) (β : Real),
H = {p : (Fin n → Real) × Real | p.2 ≥ p.1 ⬝ᵥ b - β}
Defn 12.1 (Lower): H ⊆ ℝ^n × ℝ has the form {(x, μ) | μ ≤ ⟪x, b⟫ - β}.
def IsLowerHalfSpace (n : Nat) (H : Set ((Fin n → Real) × Real)) : Prop :=
∃ (b : Fin n → Real) (β : Real),
H = {p : (Fin n → Real) × Real | p.2 ≤ p.1 ⬝ᵥ b - β}
If b ≠ 0, then b ⬝ᵥ b ≠ 0.
lemma dotProduct_self_ne_zero {n : Nat} (b : Fin n → Real) (hb : b ≠ 0) : b ⬝ᵥ b ≠ 0
For a nonzero vector b, there exists an x with x ⬝ᵥ b ≤ β.
lemma exists_dotProduct_le {n : Nat} (b : Fin n → Real) (hb : b ≠ 0) (β : Real) :
∃ x : Fin n → Real, x ⬝ᵥ b ≤ β
If H has a representation with t = 0, then it is a vertical half-space.
lemma closedHalfSpace_isVertical_of_repr_t_eq_zero {n : Nat} {H : Set ((Fin n → Real) × Real)}
{b : Fin n → Real} {t β : Real} (hnonzero : b ≠ 0 ∨ t ≠ 0) (ht : t = 0)
(hrepr : H = {p : (Fin n → Real) × Real | p.1 ⬝ᵥ b + p.2 * t ≤ β}) :
IsVerticalHalfSpace n H
If H = {p | p.1 ⬝ᵥ b + p.2 * t ≤ β} with t < 0, then H is an upper half-space.
lemma closedHalfSpace_isUpper_of_repr_t_neg {n : Nat} {H : Set ((Fin n → Real) × Real)}
{b : Fin n → Real} {t β : Real} (ht : t < 0)
(hrepr : H = {p : (Fin n → Real) × Real | p.1 ⬝ᵥ b + p.2 * t ≤ β}) :
IsUpperHalfSpace n H
If H = {p | p.1 ⬝ᵥ b + p.2 * t ≤ β} with 0 < t, then H is a lower half-space.
lemma closedHalfSpace_isLower_of_repr_t_pos {n : Nat} {H : Set ((Fin n → Real) × Real)}
{b : Fin n → Real} {t β : Real} (ht : 0 < t)
(hrepr : H = {p : (Fin n → Real) × Real | p.1 ⬝ᵥ b + p.2 * t ≤ β}) :
IsLowerHalfSpace n H
The three half-space types (vertical, upper, lower) are pairwise incompatible.
lemma halfSpaceTypes_pairwise_disjoint {n : Nat} {H : Set ((Fin n → Real) × Real)} :
(¬ (IsVerticalHalfSpace n H ∧ IsUpperHalfSpace n H)) ∧
(¬ (IsVerticalHalfSpace n H ∧ IsLowerHalfSpace n H)) ∧
(¬ (IsUpperHalfSpace n H ∧ IsLowerHalfSpace n H))
Text 12.0.1: Every closed half-space in ℝ^(n+1) belongs to exactly one of the following categories: vertical, upper, or lower.
theorem closedHalfSpace_trichotomy (n : Nat) (H : Set ((Fin n → Real) × Real))
(hH :
∃ (b : Fin n → Real) (t β : Real),
(b ≠ 0 ∨ t ≠ 0) ∧
H = {p : (Fin n → Real) × Real | p.1 ⬝ᵥ b + p.2 * t ≤ β}) :
(IsVerticalHalfSpace n H ∨ IsUpperHalfSpace n H ∨ IsLowerHalfSpace n H) ∧
¬ (IsVerticalHalfSpace n H ∧ IsUpperHalfSpace n H) ∧
¬ (IsVerticalHalfSpace n H ∧ IsLowerHalfSpace n H) ∧
¬ (IsUpperHalfSpace n H ∧ IsLowerHalfSpace n H)
A continuous linear functional on E × ℝ evaluates as l (x, μ) = l (x, 0) + μ * l (0, 1).
lemma strongDual_apply_prod {E : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E]
(l : StrongDual ℝ (E × ℝ)) (x : E) (μ : ℝ) :
l (x, μ) = l (x, (0 : ℝ)) + μ * l ((0 : E), (1 : ℝ))
The epigraph {(x, μ) | f x ≤ μ} of a convex lower-semicontinuous function is closed and convex.
lemma epigraph_closed_convex_univ {E : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E]
(f : E → ℝ) (hf_convex : ConvexOn ℝ Set.univ f) (hf_closed : LowerSemicontinuous f) :
let C : Set (E × ℝ) := {p : E × ℝ | f p.1 ≤ p.2}
Convex ℝ C ∧ IsClosed C
Any point strictly below the epigraph can be strictly separated by an affine minorant.
lemma exists_affineMap_strictly_above_below_point {E : Type*} [NormedAddCommGroup E]
[NormedSpace ℝ E] (f : E → ℝ) (hf_convex : ConvexOn ℝ Set.univ f)
(hf_closed : LowerSemicontinuous f) :
∀ (x0 : E) (μ0 : ℝ),
μ0 < f x0 →
∃ h : AffineMap ℝ E ℝ, (∀ y : E, h y ≤ f y) ∧ μ0 < h x0
Theorem 12.1. A closed convex function f is the pointwise supremum of the collection of all affine maps h such that h ≤ f.
theorem closedConvex_eq_sSup_affineMap_le {E : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E]
(f : E → ℝ) (hf_convex : ConvexOn ℝ Set.univ f) (hf_closed : LowerSemicontinuous f) :
∀ x : E,
f x =
sSup
((fun h : AffineMap ℝ E ℝ => h x) '' {h : AffineMap ℝ E ℝ | ∀ y : E, h y ≤ f y})
clConv n f is a Lean stand-in for the book's cl (conv f) for an extended-real-valued function f : ℝ^n → [-∞, ∞]. It is defined pointwise as the supremum of all affine functions majorized by f.
noncomputable def clConv (n : Nat) (f : (Fin n → Real) → EReal) : (Fin n → Real) → EReal :=
fun x =>
sSup
((fun h : AffineMap ℝ (Fin n → Real) Real => (h x : EReal)) ''
{h : AffineMap ℝ (Fin n → Real) Real | ∀ y : Fin n → Real, (h y : EReal) ≤ f y})
Corollary 12.1.1. If f : ℝ^n → [-∞, ∞] is any function, then cl (conv f) (here represented by clConv n f) is the pointwise supremum of the collection of all affine functions on ℝ^n majorized by f.
theorem clConv_eq_sSup_affineMap_le (n : Nat) (f : (Fin n → Real) → EReal) :
∀ x : Fin n → Real,
clConv n f x =
sSup
((fun h : AffineMap ℝ (Fin n → Real) Real => (h x : EReal)) ''
{h : AffineMap ℝ (Fin n → Real) Real | ∀ y : Fin n → Real, (h y : EReal) ≤ f y})
A proper convex function on ℝ^n has a point where it takes a finite real value.
lemma properConvexFunctionOn_exists_finite_point {n : Nat} {f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
∃ x0 : Fin n → Real, ∃ r0 : Real, f x0 = (r0 : EReal)
The Real-valued function x ↦ (f x).toReal is convex on the effective domain of a proper convex function f.
lemma convexOn_toReal_effectiveDomain {n : Nat} {f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
ConvexOn ℝ (effectiveDomain (Set.univ : Set (Fin n → Real)) f) (fun x => (f x).toReal)
If x0 lies in the intrinsic interior of the effective domain of a proper convex function, then the point (x0, f x0 - 1) lies outside the closure of the epigraph.
lemma point_below_not_mem_closure_epigraph_of_mem_intrinsicInterior {n : Nat}
{f : (Fin n → Real) → EReal} (hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
{x0 : Fin n → Real} {r0 : Real}
(hx0 : x0 ∈ intrinsicInterior ℝ (effectiveDomain (Set.univ : Set (Fin n → Real)) f))
(hr0 : f x0 = (r0 : EReal)) :
((x0, r0 - 1) : (Fin n → Real) × Real) ∉
closure (epigraph (Set.univ : Set (Fin n → Real)) f)
A linear functional on ℝ^n can be represented as a dot product with a coefficient vector.
lemma linearMap_exists_dotProduct_representation {n : Nat} (φ : (Fin n → Real) →ₗ[ℝ] Real) :
∃ b : Fin n → Real, ∀ x : Fin n → Real, φ x = x ⬝ᵥ b
Corollary 12.1.2. Given any proper convex function f on ℝ^n, there exist b : ℝ^n and β : ℝ such that f x ≥ ⟪x, b⟫ - β for every x.
theorem properConvexFunctionOn_exists_linear_lowerBound {n : Nat} {f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
∃ (b : Fin n → Real) (β : Real),
∀ x : Fin n → Real, f x ≥ ((x ⬝ᵥ b - β : Real) : EReal)
A real-valued function g (coerced to EReal) is pointwise below indicatorFunction C iff g x ≤ 0 for every x ∈ C.
lemma le_indicatorFunction_iff_forall_mem_le_zero {n : Nat} {C : Set (Fin n → Real)}
{g : (Fin n → Real) → Real} :
(fun x : Fin n → Real => ((g x : Real) : EReal)) ≤ indicatorFunction C ↔
∀ x, x ∈ C → g x ≤ 0
The inequality x ⬝ᵥ b - β ≤ 0 on C is equivalent to the half-space containment C ⊆ {x | x ⬝ᵥ b ≤ β}.
lemma forall_mem_sub_le_zero_iff_subset_halfspace {n : Nat} {C : Set (Fin n → Real)}
(b : Fin n → Real) (β : Real) :
(∀ x, x ∈ C → x ⬝ᵥ b - β ≤ 0) ↔ C ⊆ {x : Fin n → Real | x ⬝ᵥ b ≤ β}
Text 12.1.1: If f is the indicator function of a convex set C and h x = ⟪x, b⟫ - β, then h ≤ f if and only if h x ≤ 0 for every x ∈ C, i.e. if and only if C ⊆ {x | ⟪x, b⟫ ≤ β}.
theorem affine_le_indicatorFunction_iff_subset_halfspace {n : Nat} {C : Set (Fin n → Real)}
(hC : Convex ℝ C) (b : Fin n → Real) (β : Real) :
(fun x : Fin n → Real => ((x ⬝ᵥ b - β : Real) : EReal)) ≤ indicatorFunction C ↔
C ⊆ {x : Fin n → Real | x ⬝ᵥ b ≤ β}
Rearranging x ⬝ᵥ xStar - muStar ≤ f x yields x ⬝ᵥ xStar - f x ≤ muStar.
lemma affine_majorized_iff_forall_dotSub_le_mu {n : Nat} (f : (Fin n → Real) → Real)
(xStar : Fin n → Real) (muStar : Real) :
(∀ x : Fin n → Real, x ⬝ᵥ xStar - muStar ≤ f x) ↔
(∀ x : Fin n → Real, x ⬝ᵥ xStar - f x ≤ muStar)
A pointwise upper bound yields BddAbove for the range.
lemma bddAbove_range_of_forall_le {α β : Type*} [Preorder β] (g : α → β) (a : β)
(h : ∀ x, g x ≤ a) : BddAbove (Set.range g)
Text 12.1.2: Let f be a function on ℝ^n (contextually: closed convex), and let h x = ⟪x, x*⟫ - μ* be an affine function. Then h is majorized by f (i.e. h x ≤ f x for every x) if and only if μ* ≥ sup {⟪x, x*⟫ - f x | x ∈ ℝ^n}.
theorem affine_majorized_iff_mu_ge_sSup (n : Nat) (f : (Fin n → Real) → Real)
(xStar : Fin n → Real) (muStar : Real) :
(∀ x : Fin n → Real, x ⬝ᵥ xStar - muStar ≤ f x) ↔
BddAbove (Set.range (fun x : Fin n → Real => x ⬝ᵥ xStar - f x)) ∧
muStar ≥ sSup (Set.range (fun x : Fin n → Real => x ⬝ᵥ xStar - f x))§12 — Conjugates of Convex Functions — Part 2
Lean authors: Zichen Wang, Wanli Ma, Zhiyan Wang, Zaiwen Wen.
Defn 12.2: The conjugate f* of a function f on ℝ^n is the function on ℝ^n defined by f*(x*) = sup_x (⟪x, x*⟫ - f x) = - inf_x (f x - ⟪x, x*⟫).
noncomputable def fenchelConjugate (n : Nat) (f : (Fin n → Real) → EReal) :
(Fin n → Real) → EReal :=
fun xStar =>
sSup (Set.range (fun x : Fin n → Real => ((x ⬝ᵥ xStar : Real) : EReal) - f x))
Text 12.2.1: If f₁ and f₂ are functions on ℝ^n satisfying f₁ x ≤ f₂ x for every x, then their conjugates satisfy the reverse inequality: f₁*(x*) ≥ f₂*(x*) for every x*.
theorem fenchelConjugate_antitone (n : Nat) :
Antitone (fun f : (Fin n → Real) → EReal => fenchelConjugate n f)
fenchelConjugate rewritten as a pointwise iSup (over Set.range).
lemma fenchelConjugate_eq_iSup (n : Nat) (f : (Fin n → Real) → EReal) (xStar : Fin n → Real) :
fenchelConjugate n f xStar =
iSup (fun x : Fin n → Real => ((x ⬝ᵥ xStar : Real) : EReal) - f x)
The dot product with a fixed vector is continuous.
lemma continuous_dotProduct_const {n : Nat} (x : Fin n → Real) :
Continuous (fun xStar : Fin n → Real => x ⬝ᵥ xStar)
Lower semicontinuity of the affine pieces xStar ↦ ⟪x, xStar⟫ - c used in fenchelConjugate.
lemma lowerSemicontinuous_dotProduct_sub_const {n : Nat} (x : Fin n → Real) (c : EReal) :
LowerSemicontinuous (fun xStar : Fin n → Real => ((x ⬝ᵥ xStar : Real) : EReal) - c)
f*(b) ≤ μ iff the affine function x ↦ ⟪x,b⟫ - μ is pointwise below f.
lemma fenchelConjugate_le_coe_iff_affine_le (n : Nat) (f : (Fin n → Real) → EReal)
(b : Fin n → Real) (μ : Real) :
fenchelConjugate n f b ≤ (μ : EReal) ↔
∀ x : Fin n → Real, ((x ⬝ᵥ b - μ : Real) : EReal) ≤ f x
Any affine map h : ℝ^n → ℝ can be written as x ↦ ⟪x,b⟫ - β.
lemma affineMap_exists_dotProduct_sub {n : Nat} (h : AffineMap ℝ (Fin n → Real) Real) :
∃ (b : Fin n → Real) (β : Real), ∀ x : Fin n → Real, h x = x ⬝ᵥ b - β
clConv n f is pointwise bounded above by f.
lemma clConv_le (n : Nat) (f : (Fin n → Real) → EReal) : clConv n f ≤ f
Any affine minorant of f is pointwise below clConv n f.
lemma affine_le_clConv (n : Nat) (f : (Fin n → Real) → EReal) (h : AffineMap ℝ (Fin n → Real) Real)
(hh : ∀ y : Fin n → Real, (h y : EReal) ≤ f y) :
∀ x : Fin n → Real, (h x : EReal) ≤ clConv n f x
Two EReal values are equal if they have the same real upper bounds.
lemma EReal.eq_of_forall_le_coe_iff {a b : EReal}
(h : ∀ μ : Real, a ≤ (μ : EReal) ↔ b ≤ (μ : EReal)) : a = b
The Fenchel conjugate is always closed and convex.
lemma fenchelConjugate_closedConvex (n : Nat) (f : (Fin n → Real) → EReal) :
LowerSemicontinuous (fenchelConjugate n f) ∧ ConvexFunction (fenchelConjugate n f)
For a real x and a : EReal, x - a is the supremum of x - β over real β with a ≤ β.
lemma sub_eq_sSup_image_sub_of_le (x : Real) (a : EReal) :
(x : EReal) - a =
sSup ((fun β : Real => ((x - β : Real) : EReal)) '' {β : Real | a ≤ (β : EReal)})
The Fenchel biconjugate coincides with clConv.
lemma fenchelConjugate_biconjugate_eq_clConv (n : Nat) (f : (Fin n → Real) → EReal) :
fenchelConjugate n (fenchelConjugate n f) = clConv n f
The conjugate of clConv agrees with the conjugate.
lemma fenchelConjugate_clConv_eq (n : Nat) (f : (Fin n → Real) → EReal) :
fenchelConjugate n (clConv n f) = fenchelConjugate n f
If f is identically ⊤, then clConv n f is identically ⊤.
lemma clConv_eq_const_top_of_forall_eq_top {n : Nat} (f : (Fin n → Real) → EReal)
(hf : ∀ x : Fin n → Real, f x = ⊤) :
clConv n f = fun _ : Fin n → Real => (⊤ : EReal)
A proper convex function has a proper Fenchel conjugate.
lemma proper_fenchelConjugate_of_proper (n : Nat) {f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f)
Properness is preserved by Fenchel conjugation (for convex functions).
lemma fenchelConjugate_proper_iff (n : Nat) (f : (Fin n → Real) → EReal) (hf : ConvexFunction f) :
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f) ↔
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f
Theorem 12.2. Let f be a convex function. Then its conjugate f* (here represented by fenchelConjugate n f) is a closed convex function, and it is proper if and only if f is proper. Moreover, the conjugate of the closure equals the conjugate, and the biconjugate equals the closure: (cl f)^* = f^* and f^{**} = cl f (here cl f is represented by clConv n f).
theorem fenchelConjugate_closedConvex_proper_iff_and_biconjugate (n : Nat)
(f : (Fin n → Real) → EReal) (hf : ConvexFunction f) :
(LowerSemicontinuous (fenchelConjugate n f) ∧ ConvexFunction (fenchelConjugate n f)) ∧
(ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f) ↔
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) ∧
fenchelConjugate n (clConv n f) = fenchelConjugate n f ∧
fenchelConjugate n (fenchelConjugate n f) = clConv n f
Lower semicontinuity implies closedness of the book's epigraph epigraph univ f.
lemma isClosed_epigraph_univ_of_lowerSemicontinuous {n : Nat} {f : (Fin n → Real) → EReal}
(hf : LowerSemicontinuous f) :
IsClosed (epigraph (S := (Set.univ : Set (Fin n → Real))) f)
Supporting affine minorant for a closed proper convex EReal-valued function: if (μ0 : EReal) < f x0, then there is an affine h with h ≤ f and μ0 < h x0.
lemma exists_affineMap_strictly_above_below_point_ereal {n : Nat} {f : (Fin n → Real) → EReal}
(hf_closed : LowerSemicontinuous f)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
∀ (x0 : Fin n → Real) (μ0 : Real), ((μ0 : EReal) < f x0) →
∃ h : AffineMap ℝ (Fin n → Real) Real,
(∀ y : Fin n → Real, (h y : EReal) ≤ f y) ∧ μ0 < h x0
A closed proper convex function agrees with its clConv envelope.
lemma clConv_eq_of_closedProperConvex {n : Nat} {f : (Fin n → Real) → EReal}
(hf_closed : LowerSemicontinuous f)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
clConv n f = f§12 — Conjugates of Convex Functions — Part 3
Lean authors: Zichen Wang, Wanli Ma, Zhiyan Wang, Zaiwen Wen.
Corollary 12.2.1. The conjugacy operation f ↦ f* (here f* = fenchelConjugate n f) induces a symmetric one-to-one correspondence in the class of all closed proper convex functions on ℝ^n.
theorem fenchelConjugate_symm_bijOn_closedProperConvex (n : Nat) :
Set.BijOn (fun f : (Fin n → Real) → EReal => fenchelConjugate n f)
{f :
(Fin n → Real) → EReal |
LowerSemicontinuous f ∧ ConvexFunction f ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f}
{f :
(Fin n → Real) → EReal |
LowerSemicontinuous f ∧ ConvexFunction f ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f} ∧
∀ f : (Fin n → Real) → EReal,
(LowerSemicontinuous f ∧ ConvexFunction f ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) →
fenchelConjugate n (fenchelConjugate n f) = f
Taking the convex-function closure does not change the Fenchel conjugate.
lemma fenchelConjugate_eq_of_convexFunctionClosure (n : Nat) (f : (Fin n → Real) → EReal) :
fenchelConjugate n (convexFunctionClosure f) = fenchelConjugate n f
The effective domain of x ↦ if x ∈ ri(dom f) then f x else ⊤ is exactly ri(dom f).
lemma effectiveDomain_if_eq_intrinsicInterior (n : Nat) (f : (Fin n → Real) → EReal) :
let C := effectiveDomain (Set.univ : Set (Fin n → Real)) f
let A := intrinsicInterior ℝ C
let g : (Fin n → Real) → EReal := fun x => by
classical
exact if x ∈ A then f x else ⊤
effectiveDomain (Set.univ : Set (Fin n → Real)) g = A
Extending a convex function by ⊤ outside a convex set preserves convexity.
lemma convexFunction_if_eq_top_on_compl (n : Nat) (f : (Fin n → Real) → EReal) (hf : ConvexFunction f)
(A : Set (Fin n → Real)) (hA : Convex ℝ A) :
ConvexFunction (fun x : Fin n → Real => by
classical
exact if x ∈ A then f x else ⊤)
For g(x)=f(x) on A and g(x)=⊤ outside, g* is the supremum over A.
lemma fenchelConjugate_if_eq_sSup_image (n : Nat) (f : (Fin n → Real) → EReal) (A : Set (Fin n → Real)) :
∀ xStar : Fin n → Real,
fenchelConjugate n (fun x : Fin n → Real => by
classical
exact if x ∈ A then f x else ⊤) xStar =
sSup ((fun x : Fin n → Real => ((x ⬝ᵥ xStar : Real) : EReal) - f x) '' A)
Relative interior is idempotent for convex sets in Euclidean space.
lemma euclideanRelativeInterior_idem_of_convex (n : Nat) (C : Set (EuclideanSpace Real (Fin n)))
(hC : Convex ℝ C) :
euclideanRelativeInterior n (euclideanRelativeInterior n C) = euclideanRelativeInterior n C
Corollary 12.2.2. For a convex function f on ℝ^n, one can compute the Fenchel conjugate f*(x*) = sup_x (⟪x, x*⟫ - f x) by restricting to points x in ri (dom f). Here f* = fenchelConjugate n f, dom f = effectiveDomain Set.univ f, and ri = intrinsicInterior ℝ.
theorem fenchelConjugate_eq_sSup_intrinsicInterior_effectiveDomain (n : Nat)
(f : (Fin n → Real) → EReal) (hf : ConvexFunction f) :
∀ xStar : Fin n → Real,
fenchelConjugate n f xStar =
sSup
((fun x : Fin n → Real => ((x ⬝ᵥ xStar : Real) : EReal) - f x) ''
intrinsicInterior ℝ (effectiveDomain (Set.univ : Set (Fin n → Real)) f))
Text 12.2.2: A pair of functions (f, g) : (ℝ^n → [-∞, ∞]) × (ℝ^n → [-∞, ∞]) satisfies the generalized Fenchel inequality if ⟪x, y⟫ ≤ f x + g y for all x, y ∈ ℝ^n.
def SatisfiesGeneralizedFenchelInequality (n : Nat)
(p : ((Fin n → Real) → EReal) × ((Fin n → Real) → EReal)) : Prop :=
∀ x y : Fin n → Real, ((x ⬝ᵥ y : Real) : EReal) ≤ p.1 x + p.2 y
Text 12.2.2: The set 𝓦 of all pairs (f, g) satisfying the generalized Fenchel inequality.
abbrev GeneralizedFenchelPairs (n : Nat) :=
{p : ((Fin n → Real) → EReal) × ((Fin n → Real) → EReal) //
SatisfiesGeneralizedFenchelInequality n p}
From the generalized Fenchel inequality for (f, g), we get the pointwise bound f* ≤ g.
lemma satisfiesGeneralizedFenchelInequality_conjugate_le_snd (n : Nat)
{f g : (Fin n → Real) → EReal} (h : SatisfiesGeneralizedFenchelInequality n (f, g)) :
∀ y : Fin n → Real, fenchelConjugate n f y ≤ g y
A generalized Fenchel pair cannot take the value ⊥ in either component.
lemma satisfiesGeneralizedFenchelInequality_no_bot (n : Nat)
{f g : (Fin n → Real) → EReal} (h : SatisfiesGeneralizedFenchelInequality n (f, g)) :
(∀ x : Fin n → Real, f x ≠ (⊥ : EReal)) ∧ (∀ y : Fin n → Real, g y ≠ (⊥ : EReal))
A minimal generalized Fenchel pair has some point where the first component is not ⊤.
lemma generalizedFenchelPairs_exists_fst_ne_top_of_isMin (n : Nat) (p : GeneralizedFenchelPairs n)
(hpmin : IsMin p) : ∃ x : Fin n → Real, p.1.1 x ≠ (⊤ : EReal)
A minimal generalized Fenchel pair has some point where the second component is not ⊤.
lemma generalizedFenchelPairs_exists_snd_ne_top_of_isMin (n : Nat) (p : GeneralizedFenchelPairs n)
(hpmin : IsMin p) : ∃ y : Fin n → Real, p.1.2 y ≠ (⊤ : EReal)
Under the generalized Fenchel inequality, swapping the pair preserves the property.
lemma satisfiesGeneralizedFenchelInequality_swap (n : Nat) {f g : (Fin n → Real) → EReal}
(h : SatisfiesGeneralizedFenchelInequality n (f, g)) :
SatisfiesGeneralizedFenchelInequality n (g, f)
If f has no ⊥ values and is not identically ⊤, then f* is never ⊥.
lemma fenchelConjugate_ne_bot_of_exists_ne_top (n : Nat) (f : (Fin n → Real) → EReal)
(h : ∃ x0 : Fin n → Real, f x0 ≠ (⊤ : EReal)) :
∀ y : Fin n → Real, fenchelConjugate n f y ≠ (⊥ : EReal)
Fenchel's inequality: the pair (f, f*) satisfies the generalized Fenchel inequality.
lemma satisfiesGeneralizedFenchelInequality_self_conjugate (n : Nat) {f : (Fin n → Real) → EReal}
(hf_bot : ∀ x : Fin n → Real, f x ≠ (⊥ : EReal))
(hf_top : ∃ x0 : Fin n → Real, f x0 ≠ (⊤ : EReal)) :
SatisfiesGeneralizedFenchelInequality n (f, fenchelConjugate n f)
Text 12.2.2: For 𝓦 the set of pairs satisfying the generalized Fenchel inequality, ordered by (f', g') ≼ (f, g) meaning f' ≤ f and g' ≤ g pointwise, a pair is a minimal element of 𝓦 if and only if f and g are mutually conjugate, i.e. g = f^* and f = g^* (here f^* is fenchelConjugate).
theorem generalizedFenchelPairs_isMin_iff_mutuallyConjugate (n : Nat)
(p : GeneralizedFenchelPairs n) :
IsMin p ↔
p.1.2 = fenchelConjugate n p.1.1 ∧ p.1.1 = fenchelConjugate n p.1.2
Text 12.2.3 (Fenchel's inequality): For any proper convex function f on ℝ^n and its conjugate f*, the inequality ⟪x, x*⟫ ≤ f x + f*(x*) holds for every x ∈ ℝ^n and every x* ∈ ℝ^n. Here the conjugate f* is represented by fenchelConjugate n f.
theorem fenchel_inequality (n : Nat) (f : (Fin n → Real) → EReal)
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
∀ x xStar : Fin n → Real, ((x ⬝ᵥ xStar : Real) : EReal) ≤ f x + fenchelConjugate n f xStar
The Fenchel conjugate of the indicator of a set is the supremum of the dot product over that set.
lemma fenchelConjugate_indicatorFunction_eq_sSup_dotProduct_image (n : Nat)
(L : Submodule ℝ (Fin n → Real)) :
∀ xStar : Fin n → Real,
fenchelConjugate n (indicatorFunction (L : Set (Fin n → Real))) xStar =
sSup ((fun x : Fin n → Real => ((x ⬝ᵥ xStar : Real) : EReal)) '' (L : Set (Fin n → Real)))
Scaling the left argument scales the dot product.
lemma dotProduct_smul_left_real {n : Nat} (a : Real) (y xStar : Fin n → Real) :
(a • y) ⬝ᵥ xStar = a * (y ⬝ᵥ xStar)
If xStar is orthogonal to every x ∈ L, then the supremum of ⟪x,xStar⟫ over L is 0.
lemma sSup_dotProduct_image_eq_zero_of_forall_mem_eq_zero {n : Nat}
(L : Submodule ℝ (Fin n → Real)) (xStar : Fin n → Real)
(h : ∀ x : Fin n → Real, x ∈ L → x ⬝ᵥ xStar = 0) :
sSup ((fun x : Fin n → Real => ((x ⬝ᵥ xStar : Real) : EReal)) '' (L : Set (Fin n → Real))) = 0
If some y ∈ L has nonzero dot product with xStar, then the supremum over L is +∞.
lemma sSup_dotProduct_image_eq_top_of_exists_mem_ne_zero {n : Nat}
(L : Submodule ℝ (Fin n → Real)) (xStar : Fin n → Real)
(h : ∃ y : Fin n → Real, y ∈ L ∧ y ⬝ᵥ xStar ≠ 0) :
sSup ((fun x : Fin n → Real => ((x ⬝ᵥ xStar : Real) : EReal)) '' (L : Set (Fin n → Real))) = ⊤
Text 12.2.3 (indicator function of a subspace): Let f be the indicator function of a subspace L of ℝ^n, i.e. f(x) = δ(x | L). Then the conjugate function f* is the indicator function of the orthogonal complement L^{⊥} (the set of vectors x* such that ⟪x, x*⟫ = 0 for all x ∈ L), denoted δ(· | L^{⊥}).
theorem fenchelConjugate_indicatorFunction_submodule (n : Nat) (L : Submodule ℝ (Fin n → Real)) :
fenchelConjugate n (indicatorFunction (L : Set (Fin n → Real))) =
indicatorFunction {xStar : Fin n → Real | ∀ x : Fin n → Real, x ∈ L → x ⬝ᵥ xStar = 0}§12 — Conjugates of Convex Functions — Part 4
Lean authors: Zichen Wang, Wanli Ma, Zhiyan Wang, Zaiwen Wen.
A standard inequality for the exponential function: for a > 0, the function x ↦ x * a - exp x is bounded above by a * log a - a.
lemma exp_linear_sub_exp_le_conjugateValue {a x : ℝ} (ha : 0 < a) :
x * a - Real.exp x ≤ a * Real.log a - a
If a < 0, the function x ↦ x * a - exp x is unbounded above.
lemma exists_gt_linear_sub_exp_of_neg {a μ : ℝ} (ha : a < 0) :
∃ x : ℝ, μ < x * a - Real.exp x
For any negative μ, one can make -exp x exceed μ by choosing x very negative.
lemma exists_gt_neg_exp_of_neg_mu {μ : ℝ} (hμ : μ < 0) :
∃ x : ℝ, μ < -Real.exp x
The xStar 0 > 0 case for the conjugate of x ↦ exp (x 0) in dimension 1.
lemma fenchelConjugate_exp_pos_case (xStar : Fin 1 → Real) (ha : 0 < xStar 0) :
fenchelConjugate 1 (fun x : Fin 1 → Real => ((Real.exp (x 0) : Real) : EReal)) xStar =
((xStar 0 * Real.log (xStar 0) - xStar 0 : Real) : EReal)
The xStar 0 = 0 case for the conjugate of x ↦ exp (x 0) in dimension 1.
lemma fenchelConjugate_exp_zero_case (xStar : Fin 1 → Real) (ha0 : xStar 0 = 0) :
fenchelConjugate 1 (fun x : Fin 1 → Real => ((Real.exp (x 0) : Real) : EReal)) xStar = 0
The xStar 0 < 0 case for the conjugate of x ↦ exp (x 0) in dimension 1.
lemma fenchelConjugate_exp_neg_case (xStar : Fin 1 → Real) (ha : xStar 0 < 0) :
fenchelConjugate 1 (fun x : Fin 1 → Real => ((Real.exp (x 0) : Real) : EReal)) xStar = ⊤
Text 12.2.4: The conjugate function of the exponential function f(x) = e^x on ℝ is given by the piecewise formula f*(x^*) = x^* log x^* - x^* if x^* > 0, f*(0) = 0, and f*(x^*) = +∞ if x^* < 0.
theorem fenchelConjugate_exp (xStar : Fin 1 → Real) :
fenchelConjugate 1 (fun x : Fin 1 → Real => ((Real.exp (x 0) : Real) : EReal)) xStar =
if xStar 0 > 0 then ((xStar 0 * Real.log (xStar 0) - xStar 0 : Real) : EReal)
else if xStar 0 = 0 then (0 : EReal) else ⊤
Turn the hypotheses 1 < p and 1/p + 1/q = 1 into p.HolderConjugate q.
lemma holderConjugate_of_oneDiv_add_oneDiv_eq_one (p q : ℝ) (hp : 1 < p)
(hpq : 1 / p + 1 / q = (1 : ℝ)) : p.HolderConjugate q
Young's inequality rearranged as an upper bound for the Fenchel-conjugate range terms.
lemma range_term_le_conj_bound_oneDiv_mul_abs_rpow {p q a x : ℝ} (hpq : p.HolderConjugate q) :
x * a - (1 / p) * Real.rpow (|x|) p ≤ (1 / q) * Real.rpow (|a|) q
Value of x*a - (1/p)*|x|^p at the explicit maximizer x = sign(a) * |a|^(q-1).
lemma opt_value_oneDiv_mul_abs_rpow {p q : ℝ} (hpq : p.HolderConjugate q) (a : ℝ) :
let x0 : ℝ := if 0 ≤ a then Real.rpow (|a|) (q - 1) else -Real.rpow (|a|) (q - 1)
x0 * a - (1 / p) * Real.rpow (|x0|) p = (1 / q) * Real.rpow (|a|) q
Text 12.2.5: Let p ∈ (1, +∞). The conjugate of the function f(x) = (1/p) * |x|^p on ℝ is given by f*(x*) = (1/q) * |x*|^q, where 1/p + 1/q = 1.
theorem fenchelConjugate_oneDiv_mul_abs_rpow (p q : Real) (hp : 1 < p)
(hpq : 1 / p + 1 / q = (1 : Real)) :
∀ xStar : Fin 1 → Real,
fenchelConjugate 1
(fun x : Fin 1 → Real =>
(((1 / p) * Real.rpow (|x 0|) p : Real) : EReal))
xStar =
(((1 / q) * Real.rpow (|xStar 0|) q : Real) : EReal)
Basic identities for q when 0 < p < 1 and 1/p + 1/q = 1.
lemma q_neg_and_coeff_identities_of_oneDiv_add_oneDiv_eq_one {p q : Real} (hp0 : 0 < p)
(hp1 : p < 1) (hpq : 1 / p + 1 / q = (1 : Real)) :
q ≠ 0 ∧ q < 0 ∧ (1 / p - 1 = -(1 / q)) ∧ (q - 1) * p = q
For 0 < p < 1 and t ≥ 0, we have the tangent-line upper bound t^p ≤ (1 - p) + p * t.
lemma rpow_le_affine_at_one_of_lt_one {p t : Real} (hp0 : 0 < p) (hp1 : p < 1) (ht : 0 ≤ t) :
Real.rpow t p ≤ (1 - p) + p * t
Value of x*a + (1/p)*x^p at the candidate maximizer x0 = |a|^(q-1) in the a < 0 case.
lemma opt_value_neg_oneDiv_mul_rpow {p q a : Real} (hp0 : 0 < p) (hp1 : p < 1)
(hpq : 1 / p + 1 / q = (1 : Real)) (ha : a < 0) :
let x0 : Real := Real.rpow (|a|) (q - 1)
x0 * a + (1 / p) * Real.rpow x0 p = (-(1 / q)) * Real.rpow (|a|) q
For a < 0, each finite range term in the Fenchel conjugate is bounded above by the optimal value -(1/q) * |a|^q.
lemma range_term_le_conjValue_neg_case {p q a x : Real} (hp0 : 0 < p) (hp1 : p < 1)
(hpq : 1 / p + 1 / q = (1 : Real)) (ha : a < 0) (hx : 0 ≤ x) :
x * a + (1 / p) * Real.rpow x p ≤ (-(1 / q)) * Real.rpow (|a|) q
The xStar 0 ≥ 0 case: the conjugate is unbounded above, hence ⊤.
lemma fenchelConjugate_neg_oneDiv_mul_rpow_top_case (p : Real) (hp0 : 0 < p) (xStar : Fin 1 → ℝ)
(ha : 0 ≤ xStar 0) :
fenchelConjugate 1
(fun x : Fin 1 → ℝ =>
if 0 ≤ x 0 then ((-(1 / p) * Real.rpow (x 0) p : ℝ) : EReal) else ⊤)
xStar = ⊤
Text 12.2.6: Let 0 < p < 1. Consider the extended-real function on ℝ given by f(x) = -(1/p) * x^p for x ≥ 0 and f(x) = +∞ otherwise. Its conjugate has the piecewise formula f*(x*) = -(1/q) * |x*|^q for x* < 0 and f*(x*) = +∞ for x* ≥ 0, where 1/p + 1/q = 1 (in particular q < 0).
theorem fenchelConjugate_neg_oneDiv_mul_rpow (p q : Real) (hp0 : 0 < p) (hp1 : p < 1)
(hpq : 1 / p + 1 / q = (1 : Real)) :
∀ xStar : Fin 1 → Real,
fenchelConjugate 1
(fun x : Fin 1 → Real =>
if 0 ≤ x 0 then ((-(1 / p) * Real.rpow (x 0) p : Real) : EReal) else ⊤)
xStar =
if xStar 0 < 0 then ((-(1 / q) * Real.rpow (|xStar 0|) q : Real) : EReal) else ⊤§12 — Conjugates of Convex Functions — Part 5
Lean authors: Zichen Wang, Wanli Ma, Zhiyan Wang, Zaiwen Wen.
Simplify the Fenchel-conjugate range term for the function x ↦ if |x| ≤ a then -√(a^2-x^2) else +∞ in dimension 1.
lemma range_term_neg_sqrt_sq_sub_sq_simp (a : Real) (x xStar : Fin 1 → Real) :
(((x ⬝ᵥ xStar : Real) : EReal) -
(if |x 0| ≤ a then ((-Real.sqrt (a ^ 2 - (x 0) ^ 2) : Real) : EReal) else ⊤)) =
if |x 0| ≤ a then
((x 0 * xStar 0 + Real.sqrt (a ^ 2 - (x 0) ^ 2) : Real) : EReal)
else ⊥
A Cauchy–Schwarz-type bound for the range terms x*s + √(a^2-x^2) under the constraint |x| ≤ a and a ≥ 0.
lemma mul_add_sqrt_sq_sub_sq_le {a x s : Real} (ha : 0 ≤ a) (hx : |x| ≤ a) :
x * s + Real.sqrt (a ^ 2 - x ^ 2) ≤ a * Real.sqrt (1 + s ^ 2)
A concrete point x0 achieving the upper bound in mul_add_sqrt_sq_sub_sq_le.
lemma exists_argmax_mul_add_sqrt_sq_sub_sq (a s : Real) (ha : 0 ≤ a) :
∃ x0 : Real, |x0| ≤ a ∧ x0 * s + Real.sqrt (a ^ 2 - x0 ^ 2) = a * Real.sqrt (1 + s ^ 2)
Text 12.2.7: Let f : ℝ → [-∞, +∞] be the extended-real function f(x) = -√(a^2 - x^2) for |x| ≤ a (with a ≥ 0) and f(x) = +∞ otherwise. Then its conjugate is given by f*(x*) = a * √(1 + x*^2). Here the conjugate is represented by fenchelConjugate.
theorem fenchelConjugate_neg_sqrt_sq_sub_sq (a : Real) (ha : 0 ≤ a) :
∀ xStar : Fin 1 → Real,
fenchelConjugate 1
(fun x : Fin 1 → Real =>
if |x 0| ≤ a then ((-Real.sqrt (a ^ 2 - (x 0) ^ 2) : Real) : EReal) else ⊤)
xStar =
((a * Real.sqrt (1 + (xStar 0) ^ 2) : Real) : EReal)
Text 12.2.8: Consider f(x) = -log x - 1/2 for x > 0 (and +∞ otherwise). The conjugate is given by f*(x*) = -log(-x*) - 1/2 if x* < 0 and f*(x*) = +∞ if x* ≥ 0. Here the conjugate is represented by fenchelConjugate.
theorem fenchelConjugate_neg_log_sub_oneHalf :
∀ xStar : Fin 1 → Real,
fenchelConjugate 1
(fun x : Fin 1 → Real =>
if 0 < x 0 then ((-Real.log (x 0) - (1 / 2 : Real) : Real) : EReal) else ⊤)
xStar =
if xStar 0 < 0 then ((-Real.log (-xStar 0) - (1 / 2 : Real) : Real) : EReal) else ⊤
The quadratic function w(x) = (1/2) * ⟪x, x⟫ on ℝ^n, written on coordinates x : Fin n → ℝ and valued in EReal.
noncomputable def quadraticHalfInner (n : Nat) : (Fin n → Real) → EReal :=
fun x => (((1 / 2 : Real) * (x ⬝ᵥ x) : Real) : EReal)§12 — Conjugates of Convex Functions — Part 6
Lean authors: Zichen Wang, Wanli Ma, Zhiyan Wang, Zaiwen Wen.
Nonnegativity of the dot product of a vector with itself.
lemma dotProduct_self_nonneg {n : Nat} (v : Fin n → Real) : 0 ≤ v ⬝ᵥ v
The key inequality for the conjugate of quadraticHalfInner: ⟪x, x*⟫ - (1/2)⟪x, x⟫ ≤ (1/2)⟪x*, x*⟫.
lemma range_term_quadraticHalfInner_le (n : Nat) (x xStar : Fin n → Real) :
((x ⬝ᵥ xStar : Real) : EReal) - quadraticHalfInner n x ≤ quadraticHalfInner n xStar
The Fenchel conjugate of w(x) = (1/2)⟪x,x⟫ is itself.
lemma fenchelConjugate_quadraticHalfInner (n : Nat) :
fenchelConjugate n (quadraticHalfInner n) = quadraticHalfInner n
The quadratic function w(x) = (1/2)⟪x,x⟫ is convex on ℝ^n.
lemma convexFunction_quadraticHalfInner (n : Nat) : ConvexFunction (quadraticHalfInner n)
If f* = f, then f is pointwise bounded below by w(x) = (1/2)⟪x,x⟫.
lemma quadraticHalfInner_le_of_fenchelConjugate_eq_self (n : Nat)
(f : (Fin n → Real) → EReal) (hf : fenchelConjugate n f = f) :
∀ x : Fin n → Real, quadraticHalfInner n x ≤ f x
Text 12.2.9: The fixed-point identity f* = f for Fenchel conjugation has a unique solution in the class of convex functions on ℝ^n, namely f = w where w(x) = (1/2) * ⟪x, x⟫. Here f* is fenchelConjugate n f.
theorem fenchelConjugate_eq_self_unique_solution (n : Nat) :
(ConvexFunction (quadraticHalfInner n) ∧
fenchelConjugate n (quadraticHalfInner n) = quadraticHalfInner n) ∧
∀ f : (Fin n → Real) → EReal,
ConvexFunction f → fenchelConjugate n f = f → f = quadraticHalfInner n
Text 12.1.3: Let f_{-∞} and f_{+∞} denote the constant functions on ℝ^n taking values -∞ and +∞ respectively. Then they are conjugate to each other (with conjugation given here by fenchelConjugate), i.e. (f_{-∞})^* = f_{+∞} and (f_{+∞})^* = f_{-∞}.
def constNegInf (n : Nat) : (Fin n → Real) → EReal :=
fun _ => (⊥ : EReal)
The constant function on ℝ^n with value +∞.
def constPosInf (n : Nat) : (Fin n → Real) → EReal :=
fun _ => (⊤ : EReal)
The Fenchel conjugate of the constant -∞ function is pointwise +∞.
lemma fenchelConjugate_constNegInf_apply (n : Nat) (xStar : Fin n → Real) :
fenchelConjugate n (constNegInf n) xStar = (⊤ : EReal)
The Fenchel conjugate of the constant +∞ function is pointwise -∞.
lemma fenchelConjugate_constPosInf_apply (n : Nat) (xStar : Fin n → Real) :
fenchelConjugate n (constPosInf n) xStar = (⊥ : EReal)
For f_{-∞} the constant function x ↦ -∞ on ℝ^n, its Fenchel conjugate is f_{+∞}.
theorem fenchelConjugate_constNegInf (n : Nat) :
fenchelConjugate n (constNegInf n) = constPosInf n
For f_{+∞} the constant function x ↦ +∞ on ℝ^n, its Fenchel conjugate is f_{-∞}.
theorem fenchelConjugate_constPosInf (n : Nat) :
fenchelConjugate n (constPosInf n) = constNegInf n
Rewriting a dot product after applying the inverse linear equivalence, using the assumed adjoint relation.
lemma dotProduct_linearEquiv_symm_eq (n : Nat) (A AStar : (Fin n → Real) ≃ₗ[ℝ] (Fin n → Real))
(hAStar : ∀ x y : Fin n → Real, (A x) ⬝ᵥ y = x ⬝ᵥ AStar y) (y v : Fin n → Real) :
(A.symm y) ⬝ᵥ v = y ⬝ᵥ AStar.symm v
Cancel a finite real constant on the right in an EReal inequality.
lemma ereal_add_coe_le_coe_iff_cancel (t : EReal) (c μ : Real) :
t + (c : EReal) ≤ (μ : EReal) ↔ t ≤ ((μ - c : Real) : EReal)
Change-of-variables lemma for the affine-minorant characterization used in Theorem 12.3.
lemma affineTransform_pointwise_inequality_iff (n : Nat) (h : (Fin n → Real) → EReal)
(A AStar : (Fin n → Real) ≃ₗ[ℝ] (Fin n → Real))
(hAStar : ∀ x y : Fin n → Real, (A x) ⬝ᵥ y = x ⬝ᵥ AStar y) (a aStar : Fin n → Real)
(α : Real) (xStar : Fin n → Real) (μ : Real) :
let b' : Fin n → Real := AStar.symm (xStar - aStar)
let μ' : Real := μ + α - (a ⬝ᵥ (xStar - aStar))
(∀ x : Fin n → Real,
((x ⬝ᵥ xStar - μ : Real) : EReal) ≤
h (A (x - a)) + ((x ⬝ᵥ aStar : Real) : EReal) + (α : EReal)) ↔
(∀ y : Fin n → Real, ((y ⬝ᵥ b' - μ' : Real) : EReal) ≤ h y)
Upper-bound equivalence for the Fenchel conjugate under the affine transformation in Theorem 12.3.
lemma fenchelConjugate_affineTransform_le_coe_iff (n : Nat) (h : (Fin n → Real) → EReal)
(A AStar : (Fin n → Real) ≃ₗ[ℝ] (Fin n → Real))
(hAStar : ∀ x y : Fin n → Real, (A x) ⬝ᵥ y = x ⬝ᵥ AStar y) (a aStar : Fin n → Real)
(α : Real) (xStar : Fin n → Real) (μ : Real) :
let f : (Fin n → Real) → EReal :=
fun x => h (A (x - a)) + ((x ⬝ᵥ aStar : Real) : EReal) + (α : EReal)
let b' : Fin n → Real := AStar.symm (xStar - aStar)
let μ' : Real := μ + α - (a ⬝ᵥ (xStar - aStar))
fenchelConjugate n f xStar ≤ (μ : EReal) ↔ fenchelConjugate n h b' ≤ (μ' : EReal)
Theorem 12.3. Let h be a convex function on ℝ^n, and define f(x) = h(A(x - a)) + ⟪x, a^*⟫ + α, where A is a one-to-one linear transformation of ℝ^n, a and a^* are vectors in ℝ^n, and α ∈ ℝ. Then f^*(x^*) = h^*(A^{*-1}(x^* - a^*)) + ⟪x^*, a⟫ + α^*, where A^* is the adjoint of A and α^* = -α - ⟪a, a^*⟫. Here f^* and h^* are represented by fenchelConjugate.
theorem fenchelConjugate_affineTransform (n : Nat) (h : (Fin n → Real) → EReal)
(A : (Fin n → Real) ≃ₗ[ℝ] (Fin n → Real))
(AStar : (Fin n → Real) ≃ₗ[ℝ] (Fin n → Real))
(hAStar : ∀ x y : Fin n → Real, (A x) ⬝ᵥ y = x ⬝ᵥ AStar y) (a aStar : Fin n → Real)
(α : Real) :
let f : (Fin n → Real) → EReal :=
fun x => h (A (x - a)) + ((x ⬝ᵥ aStar : Real) : EReal) + (α : EReal)
let αStar : Real := -α - (a ⬝ᵥ aStar)
fenchelConjugate n f =
fun xStar =>
fenchelConjugate n h (AStar.symm (xStar - aStar)) + ((xStar ⬝ᵥ a : Real) : EReal) +
(αStar : EReal)
Simplify the Fenchel-conjugate range term for a Tucker-type partial affine function. Outside the primal constraint set, the function value is ⊤ and the range term is ⊥.
lemma range_term_partialAffine_tucker_simp (n m : Nat) (α00 : Real) (α0 : Fin n → Real)
(αi0 : Fin m → Real) (α : Fin m → Fin n → Real) (x xStar : Fin (n + m) → Real) :
let f : (Fin (n + m) → Real) → EReal :=
fun y =>
if _ :
(∀ i : Fin m,
y (Fin.natAdd n i) =
(∑ j : Fin n, α i j * y (Fin.castAdd m j)) - αi0 i) then
((∑ j : Fin n, α0 j * y (Fin.castAdd m j) - α00 : Real) : EReal)
else
⊤
(((x ⬝ᵥ xStar : Real) : EReal) - f x) =
if _ :
(∀ i : Fin m,
x (Fin.natAdd n i) =
(∑ j : Fin n, α i j * x (Fin.castAdd m j)) - αi0 i) then
((x ⬝ᵥ xStar - (∑ j : Fin n, α0 j * x (Fin.castAdd m j) - α00) : Real) : EReal)
else
⊥
Expand the real expression ⟪x,x*⟫ - f(x) for a Tucker-type partial affine function under the primal feasibility constraint, grouping the free coordinates x (Fin.castAdd m ·).
lemma range_term_partialAffine_tucker_real_formula (n m : Nat) (α00 : Real) (α0 : Fin n → Real)
(αi0 : Fin m → Real) (α : Fin m → Fin n → Real) (x xStar : Fin (n + m) → Real)
(hx :
∀ i : Fin m,
x (Fin.natAdd n i) =
(∑ j : Fin n, α i j * x (Fin.castAdd m j)) - αi0 i) :
let xHead : Fin n → Real := fun j => x (Fin.castAdd m j)
let coeff : Fin n → Real :=
fun j =>
xStar (Fin.castAdd m j) + (∑ i : Fin m, α i j * xStar (Fin.natAdd n i)) - α0 j
let const : Real := α00 - (∑ i : Fin m, αi0 i * xStar (Fin.natAdd n i))
x ⬝ᵥ xStar - (∑ j : Fin n, α0 j * xHead j - α00) =
(∑ j : Fin n, xHead j * coeff j) + const
The dual feasibility constraint in Tucker form is equivalent to vanishing of the coefficients of the free primal variables in the conjugate range term.
lemma tucker_dualConstraint_iff_coeff_eq_zero (n m : Nat) (α0 : Fin n → Real)
(α : Fin m → Fin n → Real) (xStar : Fin (n + m) → Real) :
let coeff : Fin n → Real :=
fun j =>
xStar (Fin.castAdd m j) + (∑ i : Fin m, α i j * xStar (Fin.natAdd n i)) - α0 j
(∀ j : Fin n,
xStar (Fin.castAdd m j) =
(∑ i : Fin m, (-α i j) * xStar (Fin.natAdd n i)) + α0 j) ↔
(∀ j : Fin n, coeff j = 0)
If some coefficient in a linear form is nonzero, we can choose the variables to make the form arbitrarily large (even achieving any prescribed value up to a constant shift).
lemma exists_head_makes_linear_combo_arbitrarily_large {n : Nat} (coeff : Fin n → Real)
(const : Real) (hcoeff : ∃ j : Fin n, coeff j ≠ 0) :
∀ μ : Real, ∃ xHead : Fin n → Real, μ ≤ (∑ j : Fin n, xHead j * coeff j) + const§12 — Conjugates of Convex Functions — Part 7
Lean authors: Zichen Wang, Wanli Ma, Zhiyan Wang, Zaiwen Wen.
Text 12.3.1: Let f be an n-dimensional partial affine function on ℝ^N with 0 < n < N and m = N - n, given (after a partition of coordinates x = (ξ₁, …, ξ_N)) by a Tucker-type representation: f(x) = (∑_{j=1}^n α₀ⱼ ξⱼ - α₀₀) if ξ_{n+i} = ∑_{j=1}^n αᵢⱼ ξⱼ - αᵢ₀ for i = 1, …, m, and f(x) = +∞ otherwise. Then the conjugate f* is: f*(x*) = (∑_{i=1}^m β₀ᵢ ξ*_{n+i} - β₀₀) if ξ*ⱼ = ∑_{i=1}^m βⱼᵢ ξ*_{n+i} - βⱼ₀ for j = 1, …, n, and f*(x*) = +∞ otherwise, where βⱼᵢ = -αᵢⱼ for all i ∈ {0, …, m}, j ∈ {0, …, n}.
theorem fenchelConjugate_partialAffine_tucker (n m : Nat) (hn : 0 < n) (hm : 0 < m) (α00 : Real)
(α0 : Fin n → Real) (αi0 : Fin m → Real) (α : Fin m → Fin n → Real) :
let f : (Fin (n + m) → Real) → EReal :=
fun x =>
if _ :
(∀ i : Fin m,
x (Fin.natAdd n i) =
(∑ j : Fin n, α i j * x (Fin.castAdd m j)) - αi0 i) then
((∑ j : Fin n, α0 j * x (Fin.castAdd m j) - α00 : Real) : EReal)
else
⊤
let fStar : (Fin (n + m) → Real) → EReal :=
fun xStar =>
if _ :
(∀ j : Fin n,
xStar (Fin.castAdd m j) =
(∑ i : Fin m, (-α i j) * xStar (Fin.natAdd n i)) + α0 j) then
((∑ i : Fin m, (-αi0 i) * xStar (Fin.natAdd n i) + α00 : Real) : EReal)
else
⊤
fenchelConjugate (n + m) f = fStar
The quadratic extended-real function associated to a linear map Q on ℝ^n, x ↦ (1/2) * ⟪x, Q x⟫. Here the inner product is written as the dot product ⬝ᵥ on coordinates x : Fin n → ℝ.
noncomputable def quadraticHalfLinear (n : Nat)
(Q : (Fin n → Real) →ₗ[ℝ] (Fin n → Real)) : (Fin n → Real) → EReal :=
fun x => (((1 / 2 : Real) * (x ⬝ᵥ Q x) : Real) : EReal)
A function f on ℝ^n is *symmetric with respect to a set G of orthogonal linear transformations* if it is invariant under every element of G, i.e. f (g x) = f x for all g ∈ G and all x.
def IsSymmetricWrtOrthogonalSet {α : Type*} (n : Nat)
(G : Set ((Fin n → Real) ≃ₗᵢ[ℝ] (Fin n → Real)))
(f : (Fin n → Real) → α) : Prop :=
∀ g ∈ G, ∀ x : Fin n → Real, f (g x) = f x
If g preserves dot products, then g.symm is the adjoint of g for the dot product.
lemma dotProduct_orthogonal_symm_apply {n : Nat}
(g : (Fin n → Real) ≃ₗᵢ[ℝ] (Fin n → Real))
(hg : ∀ x y : Fin n → Real, (g x) ⬝ᵥ (g y) = x ⬝ᵥ y) :
∀ x y : Fin n → Real, (g x) ⬝ᵥ y = x ⬝ᵥ (g.symm y)
The Fenchel conjugate commutes with precomposition by an "orthogonal" map (for the dot product), with the inverse appearing on the dual side.
lemma fenchelConjugate_precomp_orthogonal (n : Nat) (f : (Fin n → Real) → EReal)
(g : (Fin n → Real) ≃ₗᵢ[ℝ] (Fin n → Real))
(hAStar : ∀ x y : Fin n → Real, (g x) ⬝ᵥ y = x ⬝ᵥ (g.symm y)) :
fenchelConjugate n (fun x => f (g x)) = fun xStar => fenchelConjugate n f (g xStar)
A lower semicontinuous convex function that never takes the value -∞ equals its Fenchel biconjugate.
lemma fenchelConjugate_biconjugate_eq_of_closedConvex (n : Nat) (f : (Fin n → Real) → EReal)
(hf_closed : LowerSemicontinuous f) (hf_convex : ConvexFunction f)
(hf_ne_bot : ∀ x : Fin n → Real, f x ≠ (⊥ : EReal)) :
fenchelConjugate n (fenchelConjugate n f) = f
If f is invariant under a set G of dot-product-orthogonal maps, then its Fenchel conjugate is invariant under G.
lemma isSymmetricWrtOrthogonalSet_fenchelConjugate_of_isSymmetricWrtOrthogonalSet (n : Nat)
(G : Set ((Fin n → Real) ≃ₗᵢ[ℝ] (Fin n → Real))) (f : (Fin n → Real) → EReal)
(hG :
∀ g ∈ G, ∀ x y : Fin n → Real, (g x) ⬝ᵥ (g y) = x ⬝ᵥ y) :
IsSymmetricWrtOrthogonalSet n G f →
IsSymmetricWrtOrthogonalSet n G (fenchelConjugate n f)
Corollary 12.3.1. A closed convex function f is symmetric with respect to a given set G of orthogonal linear transformations if and only if f* is symmetric with respect to G.
theorem isSymmetricWrtOrthogonalSet_iff_isSymmetricWrtOrthogonalSet_fenchelConjugate (n : Nat)
(G : Set ((Fin n → Real) ≃ₗᵢ[ℝ] (Fin n → Real)))
(f : (Fin n → Real) → EReal)
(hf_closed : LowerSemicontinuous f)
(hf_convex : ConvexFunction f)
(hf_ne_bot : ∀ x : Fin n → Real, f x ≠ (⊥ : EReal))
(hG :
∀ g ∈ G, ∀ x y : Fin n → Real, (g x) ⬝ᵥ (g y) = x ⬝ᵥ y) :
IsSymmetricWrtOrthogonalSet n G f ↔
IsSymmetricWrtOrthogonalSet n G (fenchelConjugate n f)
If P_L acts as the identity on L and Q ∘ Q' = P_L, then Q (Q' xStar) = xStar on L.
lemma Q_Q'_apply_eq_of_mem_L {n : Nat}
(Q Q' P_L : (Fin n → Real) →ₗ[ℝ] (Fin n → Real)) (L : Submodule ℝ (Fin n → Real))
(hQQ' : Q.comp Q' = P_L) (hP_L_id : ∀ xStar : Fin n → Real, xStar ∈ L → P_L xStar = xStar)
{xStar : Fin n → Real} (hxStar : xStar ∈ L) : Q (Q' xStar) = xStar
Completed-square inequality for the range term of the quadratic function x ↦ (1/2) * (x ⬝ᵥ Q x) when xStar ∈ L = range(Q).
lemma range_term_quadraticHalfLinear_le_on_L {n : Nat}
(Q Q' P_L : (Fin n → Real) →ₗ[ℝ] (Fin n → Real)) (L : Submodule ℝ (Fin n → Real))
(hQQ' : Q.comp Q' = P_L) (hP_L_id : ∀ xStar : Fin n → Real, xStar ∈ L → P_L xStar = xStar)
(hQsymm : ∀ x y : Fin n → Real, (Q x) ⬝ᵥ y = x ⬝ᵥ Q y)
(hQpsd : ∀ x : Fin n → Real, 0 ≤ x ⬝ᵥ Q x) {xStar : Fin n → Real} (hxStar : xStar ∈ L) :
∀ x : Fin n → Real,
((x ⬝ᵥ xStar : Real) : EReal) - quadraticHalfLinear n Q x ≤ quadraticHalfLinear n Q' xStar
Text 12.3.2: Let Q be a symmetric positive semi-definite n × n matrix and consider the quadratic convex function h(x) = (1/2) * ⟪x, Qx⟫. Let L = range(Q) and let Q' be the Moore–Penrose pseudoinverse, characterized by Q Q' = Q' Q = P_L where P_L is the orthogonal projection onto L. Then the conjugate function h* has the piecewise form h*(x*) = (1/2) * ⟪x*, Q' x*⟫ for x* ∈ L and h*(x*) = +∞ for x* ∉ L. (If Q is nonsingular, then L = ℝ^n and Q' = Q⁻¹.)
theorem fenchelConjugate_quadraticHalfLinear_pseudoinverse (n : Nat)
(Q Q' P_L : (Fin n → Real) →ₗ[ℝ] (Fin n → Real)) (L : Submodule ℝ (Fin n → Real))
(hL : L = LinearMap.range Q) (hQQ' : Q.comp Q' = P_L) (_hQ'Q : Q'.comp Q = P_L)
(hP_L_id : ∀ xStar : Fin n → Real, xStar ∈ L → P_L xStar = xStar)
(hQsymm : ∀ x y : Fin n → Real, (Q x) ⬝ᵥ y = x ⬝ᵥ Q y)
(hQpsd : ∀ x : Fin n → Real, 0 ≤ x ⬝ᵥ Q x) :
fenchelConjugate n (quadraticHalfLinear n Q) =
(by
classical
exact fun xStar =>
if xStar ∈ L then quadraticHalfLinear n Q' xStar else (⊤ : EReal))
Text 12.3.3: A proper convex function f on ℝ^n is a *partial quadratic convex function* if it can be written as f = q + δ(· | M), where q is a finite quadratic convex function and M is an affine set. Moreover, f is partial quadratic iff, after an affine change of coordinates, it has the form x ↦ h (A (x - a)) + ⟪x, a*⟫ + α, where h is an elementary partial quadratic convex function, A is a bijective linear map, and a, a* ∈ ℝ^n, α ∈ ℝ.
def IsPartialQuadraticConvexFunction (n : Nat) (f : (Fin n → Real) → EReal) : Prop :=
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f ∧
∃ (Q : (Fin n → Real) →ₗ[ℝ] (Fin n → Real)) (b : Fin n → Real) (c : Real)
(M : AffineSubspace ℝ (Fin n → Real)),
(∀ x : Fin n → Real, 0 ≤ x ⬝ᵥ Q x) ∧
f =
fun x =>
quadraticHalfLinear n Q x + ((x ⬝ᵥ b : Real) : EReal) + (c : EReal) +
indicatorFunction (M : Set (Fin n → Real)) x
Text 12.3.3 (elementary model): An *elementary partial quadratic convex function* is a prototype partial quadratic convex function given by a diagonal quadratic form plus the indicator of a linear subspace.
def IsElementaryPartialQuadraticConvexFunction (n : Nat) (h : (Fin n → Real) → EReal) : Prop :=
∃ (d : Fin n → Real) (L : Submodule ℝ (Fin n → Real)),
(∀ i : Fin n, 0 ≤ d i) ∧
h =
fun x =>
(((1 / 2 : Real) * (∑ i : Fin n, d i * (x i) ^ 2) : Real) : EReal) +
indicatorFunction (L : Set (Fin n → Real)) x
If a proper convex function can be written as r + δ(· | M), then the affine subspace M must be nonempty. This is extracted by taking a point in the (nonempty) epigraph of f and observing that off M the indicator term forces f = ⊤.
lemma exists_mem_affineSubspace_of_proper_of_eq_indicator {n : Nat}
{f r : (Fin n → Real) → EReal} {M : AffineSubspace ℝ (Fin n → Real)}
(hfprop : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hr_ne_bot : ∀ x : Fin n → Real, r x ≠ (⊥ : EReal))
(hfEq : f = fun x => r x + indicatorFunction (M : Set (Fin n → Real)) x) :
∃ a : Fin n → Real, a ∈ M
Re-center the indicator of an affine subspace at a base point: translating by a ∈ M turns δ(· | M) into the indicator of the direction submodule M.direction.
lemma indicatorFunction_affineSubspace_eq_indicator_direction_sub {n : Nat}
(M : AffineSubspace ℝ (Fin n → Real)) {a : Fin n → Real} (ha : a ∈ M) :
∀ x : Fin n → Real,
indicatorFunction (M : Set (Fin n → Real)) x =
indicatorFunction (M.direction : Set (Fin n → Real)) (x - a)
Convert dot products on Fin n → ℝ into inner products on EuclideanSpace ℝ (Fin n) via EuclideanSpace.equiv. This is the bridge that allows the use of spectral-theorem APIs.
lemma dotProduct_eq_inner_euclideanSpace (n : Nat) (x y : Fin n → Real) :
(x ⬝ᵥ y : Real) =
inner ℝ ((EuclideanSpace.equiv (Fin n) ℝ).symm x)
((EuclideanSpace.equiv (Fin n) ℝ).symm y)
Replace a linear map by its dot-product symmetric part without changing the associated quadratic form x ↦ x ⬝ᵥ Q x. This is implemented by transporting to EuclideanSpace and symmetrizing using the adjoint.
lemma linearMap_symmPart_dotProduct_preserves_quadratic {n : Nat}
(Q : (Fin n → Real) →ₗ[ℝ] (Fin n → Real)) :
∃ Qs : (Fin n → Real) →ₗ[ℝ] (Fin n → Real),
(∀ x y : Fin n → Real, (Qs x) ⬝ᵥ y = x ⬝ᵥ Qs y) ∧
∀ x : Fin n → Real, (x ⬝ᵥ Qs x : Real) = (x ⬝ᵥ Q x : Real)
Transporting an indicator function along a linear equivalence using Submodule.comap.
lemma indicatorFunction_submodule_comap_linearEquiv {n : Nat} (L : Submodule ℝ (Fin n → Real))
(A : (Fin n → Real) ≃ₗ[ℝ] (Fin n → Real)) :
∀ v : Fin n → Real,
indicatorFunction ((L.comap A.toLinearMap : Submodule ℝ (Fin n → Real)) : Set (Fin n → Real))
v =
indicatorFunction (L : Set (Fin n → Real)) (A v)
Translation identity for the quadratic term when Q is symmetric with respect to the dot product. This is the algebraic step used to rewrite a quadratic form in x in terms of x-a, plus a linear term and a constant.
lemma quadraticHalfLinear_translate_of_dotProduct_symmetric {n : Nat}
(Q : (Fin n → Real) →ₗ[ℝ] (Fin n → Real))
(hQsymm : ∀ x y : Fin n → Real, (Q x) ⬝ᵥ y = x ⬝ᵥ Q y) :
∀ a x : Fin n → Real,
quadraticHalfLinear n Q x =
quadraticHalfLinear n Q (x - a) + ((x ⬝ᵥ Q a : Real) : EReal) +
(((-(1 / 2 : Real)) * (a ⬝ᵥ Q a) : Real) : EReal)
Diagonalize a dot-product symmetric positive semidefinite linear operator on ℝ^n by transporting to EuclideanSpace and applying the spectral theorem.
lemma exists_linearEquiv_diagonalize_psd_dotProduct {n : Nat}
(Q : (Fin n → Real) →ₗ[ℝ] (Fin n → Real))
(hQsymm : ∀ x y : Fin n → Real, (Q x) ⬝ᵥ y = x ⬝ᵥ Q y)
(hpsd : ∀ x : Fin n → Real, 0 ≤ x ⬝ᵥ Q x) :
∃ (A : (Fin n → Real) ≃ₗ[ℝ] (Fin n → Real)) (d : Fin n → Real),
(∀ i : Fin n, 0 ≤ d i) ∧ ∀ x : Fin n → Real, (x ⬝ᵥ Q x : Real) = ∑ i, d i * (A x i) ^ 2§12 — Conjugates of Convex Functions — Part 8
Lean authors: Zichen Wang, Wanli Ma, Zhiyan Wang, Zaiwen Wen.
Given a linear equivalence A and diagonal weights d, build a dot-product symmetric linear map Q whose quadratic form is x ↦ ∑ i, d i * (A x i)^2.
lemma exists_Q_of_linearEquiv_diag {n : Nat} (A : (Fin n → Real) ≃ₗ[ℝ] (Fin n → Real))
(d : Fin n → Real) :
∃ Q : (Fin n → Real) →ₗ[ℝ] (Fin n → Real),
(∀ x y : Fin n → Real, (Q x) ⬝ᵥ y = x ⬝ᵥ Q y) ∧
∀ x : Fin n → Real, (x ⬝ᵥ Q x : Real) = ∑ i : Fin n, d i * (A x i) ^ 2§12 — Conjugates of Convex Functions — Part 9
Lean authors: Zichen Wang, Wanli Ma, Zhiyan Wang, Zaiwen Wen.
Translation identity for the shifted argument: rewrite quadraticHalfLinear n Q (x - a) in terms of quadraticHalfLinear n Q x, plus a linear term and a constant, assuming dot-product symmetry of Q.
lemma quadraticHalfLinear_translate_sub_of_dotProduct_symmetric {n : Nat}
(Q : (Fin n → Real) →ₗ[ℝ] (Fin n → Real))
(hQsymm : ∀ x y : Fin n → Real, (Q x) ⬝ᵥ y = x ⬝ᵥ Q y) :
∀ a x : Fin n → Real,
quadraticHalfLinear n Q (x - a) =
quadraticHalfLinear n Q x + ((x ⬝ᵥ (-Q a) : Real) : EReal) +
(((1 / 2 : Real) * (a ⬝ᵥ Q a) : Real) : EReal)
Convexity of the square function: t ↦ ((1 - t)a + t b)^2 lies below the chord between a^2 and b^2 for t ∈ [0,1].
lemma sq_segment_le (a b t : Real) (ht0 : 0 ≤ t) (ht1 : t ≤ 1) :
((1 - t) * a + t * b) ^ 2 ≤ (1 - t) * (a ^ 2) + t * (b ^ 2)
A weighted diagonal quadratic function with nonnegative weights is a proper convex function on ℝ^n.
lemma properConvexFunctionOn_diagonalQuadratic {n : Nat} (d : Fin n → Real)
(hd : ∀ i : Fin n, 0 ≤ d i) :
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real))
(fun x =>
(((1 / 2 : Real) * (∑ i : Fin n, d i * (x i) ^ 2) : Real) : EReal))
Elementary partial quadratic convex functions are proper convex on Set.univ.
lemma properConvexFunctionOn_elementaryPartialQuadratic {n : Nat}
{h : (Fin n → Real) → EReal} (hElem : IsElementaryPartialQuadraticConvexFunction n h) :
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) h
Precomposition by a linear equivalence preserves proper convexity on Set.univ.
lemma properConvexFunctionOn_precomp_linearEquiv {n : Nat}
(A : (Fin n → Real) ≃ₗ[ℝ] (Fin n → Real)) {f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (fun x => f (A x))
Translating the input preserves proper convexity on Set.univ.
lemma properConvexFunctionOn_translate {n : Nat} (a : Fin n → Real)
{f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (fun x => f (x - a))
The affine function x ↦ ⟪x, v⟫ + α is a proper convex function on Set.univ.
lemma properConvexFunctionOn_affine_dotProduct {n : Nat} (v : Fin n → Real) (α : Real) :
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real))
(fun x => ((x ⬝ᵥ v + α : Real) : EReal))
Text 12.3.3: Characterization of partial quadratic convex functions by affine change of coordinates to an elementary partial quadratic convex function plus a linear term.
theorem isPartialQuadraticConvexFunction_iff_exists_elementary_affine (n : Nat)
(f : (Fin n → Real) → EReal) :
IsPartialQuadraticConvexFunction n f ↔
∃ (h : (Fin n → Real) → EReal) (A : (Fin n → Real) ≃ₗ[ℝ] (Fin n → Real))
(a aStar : Fin n → Real) (α : Real),
IsElementaryPartialQuadraticConvexFunction n h ∧
f =
fun x =>
h (A (x - a)) + ((x ⬝ᵥ aStar : Real) : EReal) + (α : EReal)
Negation turns iInf into iSup in EReal.
lemma ereal_iSup_neg_eq_neg_iInf {ι : Sort*} (g : ι → EReal) :
(iSup fun i => -g i) = - (iInf fun i => g i)
The value of the Fenchel conjugate at 0 is the negative of the infimum of f.
lemma fenchelConjugate_zero_eq_neg_iInf (n : Nat) (f : (Fin n → Real) → EReal) :
fenchelConjugate n f 0 = - (iInf fun x : Fin n → Real => f x)
For closed proper convex f, the infimum of the Fenchel conjugate is -f 0.
lemma iInf_fenchelConjugate_eq_neg_eval_zero (n : Nat) (f : (Fin n → Real) → EReal)
(hf_closed : LowerSemicontinuous f) (hf_convex : ConvexFunction f)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
(iInf fun xStar : Fin n → Real => fenchelConjugate n f xStar) = -f 0
Text 12.3.4: Let f be a closed proper convex function on ℝ^n. Then inf_x f x = f 0 = 0 if and only if inf_{x*} f* x* = f* 0 = 0, where f* is the Fenchel conjugate (here f* = fenchelConjugate n f).
theorem inf_eq_at_zero_iff_conjugate_inf_eq_at_zero (n : Nat) (f : (Fin n → Real) → EReal)
(hf_closed : LowerSemicontinuous f) (hf_convex : ConvexFunction f)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
((iInf fun x : Fin n → Real => f x) = f 0 ∧ f 0 = 0) ↔
((iInf fun xStar : Fin n → Real => fenchelConjugate n f xStar) =
fenchelConjugate n f 0 ∧
fenchelConjugate n f 0 = 0)
Defn 12.3: A function f on ℝ^n is *symmetric* (or *even*) if it satisfies f (-x) = f x for all x ∈ ℝ^n.
abbrev IsSymmetricFunction {α : Type*} (n : Nat) (f : (Fin n → Real) → α) : Prop :=
Function.Even f
Reindex an iSup by the involution x ↦ -x.
lemma iSup_comp_neg {α β : Type*} [AddGroup α] [CompleteLattice β] (g : α → β) :
(iSup fun x => g (-x)) = iSup g
If f is even, then its Fenchel conjugate f* is even.
lemma fenchelConjugate_even_of_even (n : Nat) {f : (Fin n → Real) → EReal}
(hf : Function.Even f) : Function.Even (fenchelConjugate n f)
If the Fenchel conjugate of f is even, then the closed convex envelope clConv n f is even.
lemma even_clConv_of_even_fenchelConjugate (n : Nat) {f : (Fin n → Real) → EReal}
(hf_convex : ConvexFunction f) (hEvenConj : Function.Even (fenchelConjugate n f)) :
Function.Even (clConv n f)
For closed proper convex f, evenness of its Fenchel conjugate implies evenness of f.
lemma even_of_even_fenchelConjugate_of_closedProperConvex (n : Nat) {f : (Fin n → Real) → EReal}
(hf_closed : LowerSemicontinuous f) (hf_convex : ConvexFunction f)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hEvenConj : Function.Even (fenchelConjugate n f)) : Function.Even f
Text 12.3.5: Let f be a closed proper convex function on ℝ^n. Then f is symmetric (i.e. f (-x) = f x for all x) if and only if its conjugate f* is symmetric (here f* = fenchelConjugate n f).
theorem isSymmetricFunction_iff_isSymmetricFunction_fenchelConjugate (n : Nat)
(f : (Fin n → Real) → EReal)
(hf_closed : LowerSemicontinuous f)
(hf_convex : ConvexFunction f)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
IsSymmetricFunction n f ↔ IsSymmetricFunction n (fenchelConjugate n f)
Text 12.3.5: A function f on ℝ^n is symmetric with respect to all orthogonal transformations (rotationally symmetric) if and only if it has the form f x = g (‖x‖), for some function g on [0, +∞), where ‖·‖ is the Euclidean norm. Furthermore, for such an f, it is a closed proper convex function on ℝ^n if and only if g satisfies: 1. g is convex on [0, +∞); 2. g is non-decreasing; 3. g is lower semicontinuous; 4. g 0 is finite.
def IsRotationallySymmetric {α : Type*} (n : Nat) (f : (Fin n → Real) → α) : Prop :=
let eL : EuclideanSpace ℝ (Fin n) ≃L[ℝ] (Fin n → Real) := EuclideanSpace.equiv (Fin n) ℝ
∀ O : EuclideanSpace ℝ (Fin n) ≃ₗᵢ[ℝ] EuclideanSpace ℝ (Fin n), ∀ x : Fin n → Real,
f (eL (O (eL.symm x))) = f x
The Euclidean norm on Fin n → ℝ, defined via the equivalence with EuclideanSpace.
noncomputable def euclidNorm (n : Nat) (x : Fin n → Real) : Real :=
‖(EuclideanSpace.equiv (Fin n) ℝ).symm x‖
The nonnegative ray in Fin 1 → ℝ used to model [0, +∞).
def nonnegRay : Set (Fin 1 → Real) := {r : Fin 1 → Real | 0 ≤ r 0}
The nonnegative ray nonnegRay is convex.
lemma convex_nonnegRay : Convex ℝ (nonnegRay)
In Euclidean space, vectors of equal norm are related by some linear isometry.
lemma exists_linearIsometryEquiv_euclideanSpace_apply_eq_of_norm_eq (n : Nat)
(x y : EuclideanSpace ℝ (Fin n)) (hxy : ‖x‖ = ‖y‖) :
∃ O : EuclideanSpace ℝ (Fin n) ≃ₗᵢ[ℝ] EuclideanSpace ℝ (Fin n), O x = y
Text 12.3.5: Rotational symmetry is equivalent to depending only on the Euclidean norm.
theorem isRotationallySymmetric_iff_exists_norm_comp {α : Type*} (n : Nat)
(f : (Fin n → Real) → α) :
IsRotationallySymmetric n f ↔
∃ g : (Fin 1 → Real) → α, ∀ x : Fin n → Real, f x = g (fun _ : Fin 1 => euclidNorm n x)§12 — Conjugates of Convex Functions — Part 10
Lean authors: Zichen Wang, Wanli Ma, Zhiyan Wang, Zaiwen Wen.
If g is monotone on nonnegRay and g 0 is finite, then x ↦ g (euclidNorm n x) never takes the value ⊥.
lemma radial_ne_bot_of_mono_g0_finite {n : Nat} {g : (Fin 1 → Real) → EReal}
(hg_mono : MonotoneOn g (nonnegRay))
(hg0 : ∃ c : Real, g 0 = (c : EReal)) :
∀ x : Fin n → Real, g (fun _ : Fin 1 => euclidNorm n x) ≠ (⊥ : EReal)
Convexity of the radial composition x ↦ g (‖x‖) from convexity and monotonicity of g on the nonnegative ray, assuming g 0 is finite.
lemma radial_convex_of_convexOn_monoOn (n : Nat) {g : (Fin 1 → Real) → EReal}
(hg_conv : ConvexFunctionOn (nonnegRay) g)
(hg_mono : MonotoneOn g (nonnegRay))
(hg0 : ∃ c : Real, g 0 = (c : EReal)) :
let f : (Fin n → Real) → EReal := fun x => g (fun _ : Fin 1 => euclidNorm n x)
ConvexFunction f
Lower semicontinuity of g on nonnegRay implies lower semicontinuity of the radial composition x ↦ g (‖x‖).
lemma radial_lsc_of_lscOn (n : Nat) (g : (Fin 1 → Real) → EReal)
(hg : LowerSemicontinuousOn g (nonnegRay)) :
let f : (Fin n → Real) → EReal := fun x => g (fun _ : Fin 1 => euclidNorm n x)
LowerSemicontinuous f
Properness of the radial function follows from finiteness at 0 and monotonicity of g on the ray, once convexity is known.
lemma radial_proper_of_mono_g0_finite (n : Nat) {g : (Fin 1 → Real) → EReal}
(hg_mono : MonotoneOn g (nonnegRay)) (hg0 : ∃ c : Real, g 0 = (c : EReal))
(hf_conv :
let f : (Fin n → Real) → EReal := fun x => g (fun _ : Fin 1 => euclidNorm n x)
ConvexFunction f) :
let f : (Fin n → Real) → EReal := fun x => g (fun _ : Fin 1 => euclidNorm n x)
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f
From closed proper convexity of the radial function x ↦ g (‖x‖), recover the one-dimensional ray conditions on g.
lemma radial_ray_props_of_closedProperConvex (n : Nat) (hn : 0 < n) (g : (Fin 1 → Real) → EReal) :
let f : (Fin n → Real) → EReal := fun x => g (fun _ : Fin 1 => euclidNorm n x)
(LowerSemicontinuous f ∧ ConvexFunction f ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) →
(ConvexFunctionOn (nonnegRay) g ∧
MonotoneOn g (nonnegRay) ∧
LowerSemicontinuousOn g (nonnegRay) ∧
∃ c : Real, g 0 = (c : EReal))
theorem radial_closedProperConvex_iff (n : Nat) (hn : 0 < n) (g : (Fin 1 → Real) → EReal) :
let f : (Fin n → Real) → EReal := fun x => g (fun _ : Fin 1 => euclidNorm n x)
(LowerSemicontinuous f ∧ ConvexFunction f ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) ↔
(ConvexFunctionOn {r : Fin 1 → Real | 0 ≤ r 0} g ∧
MonotoneOn g {r : Fin 1 → Real | 0 ≤ r 0} ∧
LowerSemicontinuousOn g {r : Fin 1 → Real | 0 ≤ r 0} ∧
∃ c : Real, g 0 = (c : EReal))
Defn 12.4: Let g be a function on the nonnegative orthant ℝ^n_+. Its *monotone conjugate* g⁺ is defined for y* ∈ ℝ^n_+ by g⁺(y*) = sup_{y ≥ 0} (⟪y, y*⟫ - g y).
noncomputable def monotoneConjugate (n : Nat) (g : (Fin n → NNReal) → Real) :
(Fin n → NNReal) → Real :=
fun yStar =>
sSup <|
Set.range fun y : Fin n → NNReal =>
((fun i => (y i : Real)) ⬝ᵥ (fun i => (yStar i : Real))) - g y
EReal-valued monotone conjugate: same supremum, but living in [-∞, ∞] so it can take ⊤.
noncomputable def monotoneConjugateEReal (n : Nat) (g : (Fin n → NNReal) → Real) :
(Fin n → NNReal) → EReal :=
fun yStar =>
sSup <|
Set.range fun y : Fin n → NNReal =>
(((fun i => (y i : Real)) ⬝ᵥ (fun i => (yStar i : Real)) : Real) : EReal) - (g y : EReal)
Text 12.3.6: For x : ℝ^n, define abs x = (|x₁|, …, |xₙ|) as an element of the nonnegative orthant ℝ^n_+ (modeled as Fin n → NNReal).
def absVecNN (n : Nat) (x : Fin n → Real) : Fin n → NNReal :=
fun i => ⟨|x i|, abs_nonneg (x i)⟩
Coercing a nonnegative vector to Real and taking componentwise absolute values recovers it.
lemma absVecNN_coe (n : Nat) (y : Fin n → NNReal) :
absVecNN n (fun i => (y i : Real)) = y
Componentwise absolute values ignore global sign changes.
lemma absVecNN_neg (n : Nat) (x : Fin n → Real) : absVecNN n (-x) = absVecNN n x
Lower semicontinuity is preserved by coercion Real → EReal.
lemma lowerSemicontinuous_coe_real_toEReal {α : Type*} [TopologicalSpace α] {h : α → Real} :
LowerSemicontinuous h → LowerSemicontinuous (fun x => (h x : EReal))
Lower semicontinuity of Real → EReal coercions implies lower semicontinuity of the real map.
lemma lowerSemicontinuous_of_coe_real_toEReal {α : Type*} [TopologicalSpace α] {h : α → Real} :
LowerSemicontinuous (fun x => (h x : EReal)) → LowerSemicontinuous h
The coercion Fin n → NNReal → Fin n → Real is continuous.
lemma continuous_coeVecNN (n : Nat) :
Continuous (fun y : Fin n → NNReal => fun i => (y i : Real))
The map x ↦ absVecNN n x is continuous.
lemma continuous_absVecNN (n : Nat) : Continuous (absVecNN n)
Pointwise triangle inequality for affine combinations: | (1-t)x + t y | is bounded by the corresponding convex combination of |x| and |y| when t ∈ [0,1].
lemma abs_combo_le (a b t : Real) (ht0 : 0 ≤ t) (ht1 : 0 ≤ 1 - t) :
Choose signs so that x ⬝ᵥ xStar becomes |x| ⬝ᵥ |xStar|.
lemma exists_sign_choice_dotProduct_eq (n : Nat) (xStar : Fin n → Real) (y : Fin n → NNReal) :
∃ x : Fin n → Real,
absVecNN n x = y ∧
(x ⬝ᵥ xStar : Real) =
((fun i => (y i : Real)) ⬝ᵥ (fun i => (absVecNN n xStar i : Real)) : Real)
Text 12.3.6 (1): Let f : ℝ^n → ℝ be of the form f(x) = g(abs x) with g : ℝ^n_+ → ℝ and abs x = (|x₁|, …, |xₙ|). Then f is closed proper convex on ℝ^n iff g is lower semicontinuous, convex, finite at 0, and non-decreasing (i.e. 0 ≤ y ≤ y' → g y ≤ g y').
theorem closedProperConvex_abs_comp_iff (n : Nat) (g : (Fin n → NNReal) → Real) :
let f : (Fin n → Real) → EReal := fun x => (g (absVecNN n x) : EReal)
(LowerSemicontinuous f ∧ ConvexFunction f ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) ↔
(LowerSemicontinuous g ∧ ConvexOn NNReal (Set.univ : Set (Fin n → NNReal)) g ∧
(∃ c : Real, g 0 = c) ∧ Monotone g)
Text 12.3.6 (2): The conjugate of f(x) = g(abs x) is f*(x*) = g⁺(abs x*), where g⁺ is the EReal-valued monotone conjugate (here g⁺ = monotoneConjugateEReal n g).
theorem fenchelConjugate_abs_comp_eq_monotoneConjugate (n : Nat) (g : (Fin n → NNReal) → Real) :
let f : (Fin n → Real) → EReal := fun x => (g (absVecNN n x) : EReal)
∀ xStar : Fin n → Real,
fenchelConjugate n f xStar = monotoneConjugateEReal n g (absVecNN n xStar)§12 — Conjugates of Convex Functions — Part 11
Lean authors: Zichen Wang, Wanli Ma, Zhiyan Wang, Zaiwen Wen.
EReal-valued monotone conjugate for an EReal-valued function on ℝ^n_+ (modeled as Fin n → NNReal). This version allows the input function to take the value ⊤, so it can be iterated.
noncomputable def monotoneConjugateERealEReal (n : Nat) (g : (Fin n → NNReal) → EReal) :
(Fin n → NNReal) → EReal :=
fun yStar =>
sSup <|
Set.range fun y : Fin n → NNReal =>
(((fun i => (y i : Real)) ⬝ᵥ (fun i => (yStar i : Real)) : Real) : EReal) - g y
Theorem 12.4. Let g be a non-decreasing lower semicontinuous convex function on the nonnegative orthant ℝ^n_+ (modeled as Fin n → NNReal) such that g 0 is finite. Then the monotone conjugate g⁺ (here g⁺ = monotoneConjugateEReal n g) has the same properties, and the monotone conjugate of g⁺ is g (as an EReal-valued function).
theorem monotoneConjugate_properties_and_involutive (n : Nat) (g : (Fin n → NNReal) → Real)
(hg_mono : Monotone g) (hg_lsc : LowerSemicontinuous g)
(hg_convex : ConvexOn NNReal (Set.univ : Set (Fin n → NNReal)) g) (hg0 : ∃ c : Real, g 0 = c) :
let gPlus : (Fin n → NNReal) → EReal := monotoneConjugateEReal n g
(Monotone gPlus ∧ LowerSemicontinuous gPlus ∧
(∀ x y : Fin n → NNReal, ∀ a b : NNReal, a + b = 1 →
gPlus (a • x + b • y) ≤ ((a : Real) : EReal) * gPlus x + ((b : Real) : EReal) * gPlus y) ∧
∃ c : Real, gPlus 0 = (c : EReal)) ∧
monotoneConjugateERealEReal n gPlus = fun y => (g y : EReal)§13 — Support Functions — Part 1
Lean authors: Zichen Wang, Wanli Ma, Siyuan Shao, Yunxi Duan, Zaiwen Wen.
Text 13.0.1: Let C be a convex set. The support function δ^*(· | C) of C is defined by δ^*(x^* | C) = sup { ⟪x, x^*⟫ | x ∈ C }.
open scoped Pointwise
noncomputable abbrev deltaStar {n : Nat} (C : Set (Fin n → Real)) : (Fin n → Real) → ℝ :=
supportFunction C
Rewrite supportFunction as an sSup over an image.
open scoped Pointwise
lemma supportFunction_eq_sSup_image_dotProduct {n : Nat} (C : Set (Fin n → Real))
(x : Fin n → Real) :
supportFunction C x = sSup (Set.image (fun y : Fin n → Real => dotProduct x y) C)
The dot-product image-set at -xStar is the (-1)-scaled dot-product image-set at xStar.
open scoped Pointwise
lemma image_dotProduct_neg_left_eq_neg_smul_image_dotProduct {n : Nat} (C : Set (Fin n → Real))
(xStar : Fin n → Real) :
Set.image (fun y : Fin n → Real => dotProduct (-xStar) y) C =
(-1 : ℝ) • Set.image (fun y : Fin n → Real => dotProduct y xStar) C
Scaling a real set by -1 swaps sSup and sInf.
open scoped Pointwise
lemma sSup_neg_one_smul_eq_neg_sInf (S : Set ℝ) : sSup ((-1 : ℝ) • S) = -sInf S
Text 13.0.2: Let C be a convex set. Then inf { ⟪x, x^*⟫ | x ∈ C } = -δ^*(-x^* | C), where δ^*(· | C) is the support function of C.
open scoped Pointwise
theorem sInf_dotProduct_eq_neg_deltaStar {n : Nat} (C : Set (Fin n → Real)) (xStar : Fin n → Real)
(_hC : Convex ℝ C) :
sInf (Set.image (fun x => dotProduct x xStar) C) = -(deltaStar C (-xStar))
Rewrite deltaStar as an sSup of dotProduct x xStar over x ∈ C.
open scoped Pointwise
lemma deltaStar_eq_sSup_image_dotProduct_right {n : Nat} (C : Set (Fin n → Real))
(xStar : Fin n → Real) :
deltaStar C xStar = sSup (Set.image (fun x : Fin n → Real => dotProduct x xStar) C)
Convert the pointwise bound on the dot-product image-set to a halfspace inclusion.
open scoped Pointwise
lemma forall_image_dotProduct_le_iff_subset_halfspace {n : Nat} (C : Set (Fin n → Real))
(xStar : Fin n → Real) (β : ℝ) :
(∀ r ∈ Set.image (fun x : Fin n → Real => dotProduct x xStar) C, r ≤ β) ↔
C ⊆ { x | dotProduct x xStar ≤ β }
A csSup inequality over the dot-product image-set is equivalent to a pointwise bound.
open scoped Pointwise
lemma csSup_image_dotProduct_le_iff {n : Nat} (C : Set (Fin n → Real)) (xStar : Fin n → Real)
(β : ℝ) (hCne : C.Nonempty)
(hbdd : BddAbove (Set.image (fun x : Fin n → Real => dotProduct x xStar) C)) :
(sSup (Set.image (fun x : Fin n → Real => dotProduct x xStar) C) ≤ β) ↔
(∀ r ∈ Set.image (fun x : Fin n → Real => dotProduct x xStar) C, r ≤ β)
Text 13.0.3: Let C be a convex set. Then C ⊆ { x | ⟪x, x^*⟫ ≤ β } if and only if β ≥ δ^*(x^* | C).
open scoped Pointwise
theorem subset_halfspace_iff_ge_deltaStar {n : Nat} (C : Set (Fin n → Real)) (xStar : Fin n → Real)
(β : ℝ) (_hC : Convex ℝ C) (hCne : C.Nonempty)
(hbdd : BddAbove (Set.image (fun x : Fin n → Real => dotProduct x xStar) C)) :
C ⊆ { x | dotProduct x xStar ≤ β } ↔ β ≥ deltaStar C xStar
Text 13.0.4: Let C ⊆ ℝ^n be a nonempty convex set. The barrier cone of C, equivalently the effective domain of the support function δ^*(· | C), is bar C = dom (δ^*(· | C)) = {x^* | sup {⟪x, x^*⟫ | x ∈ C} < +∞ }.
open scoped Pointwise
noncomputable def barrierCone13 {n : Nat} (C : Set (Fin n → Real)) : Set (Fin n → Real) :=
effectiveDomain (Set.univ : Set (Fin n → Real)) (fun xStar =>
sSup {z : EReal | ∃ x ∈ C, z = ((dotProduct x xStar : Real) : EReal)})
The support function is unchanged by taking the (topological) closure of the set.
open scoped Pointwise
lemma deltaStar_eq_deltaStar_closure {n : Nat} (C : Set (Fin n → Real)) (xStar : Fin n → Real) :
deltaStar C xStar = deltaStar (closure C) xStar
For a convex set, the closure of its intrinsic (relative) interior equals its closure.
open scoped Pointwise
lemma closure_intrinsicInterior_eq_closure_of_convex {n : Nat} (C : Set (Fin n → Real))
(hC : Convex ℝ C) :
closure (intrinsicInterior ℝ C) = closure C
Text 13.0.5: For any convex set C, one has δ^*(x^* | C) = δ^*(x^* | cl C) = δ^*(x^* | ri C) for all x^*, where cl denotes closure and ri denotes relative (intrinsic) interior.
open scoped Pointwise
theorem deltaStar_eq_closure_eq_intrinsicInterior {n : Nat} (C : Set (Fin n → Real))
(hC : Convex ℝ C) :
∀ xStar : Fin n → Real,
deltaStar C xStar = deltaStar (closure C) xStar ∧
deltaStar (closure C) xStar = deltaStar (intrinsicInterior ℝ C) xStar
If the dot-product image is bounded above, any member of C satisfies the supporting inequality.
open scoped Pointwise
lemma section13_dotProduct_le_deltaStar_of_mem {n : Nat} {C : Set (Fin n → Real)}
{xStar y : Fin n → Real}
(hbdd : BddAbove (Set.image (fun x : Fin n → Real => dotProduct x xStar) C)) (hy : y ∈ C) :
dotProduct y xStar ≤ deltaStar C xStar
A bound on the dot-product image at -xStar gives a lower bound on the dot-product image at xStar.
open scoped Pointwise
lemma section13_bddBelow_image_dotProduct_of_bddAbove_image_dotProduct_neg {n : Nat}
(C : Set (Fin n → Real)) (xStar : Fin n → Real)
(hbdd : BddAbove (Set.image (fun x : Fin n → Real => dotProduct x (-xStar)) C)) :
BddBelow (Set.image (fun x : Fin n → Real => dotProduct x xStar) C)
For a point x ∈ C, being a maximizer of dotProduct · xStar is equivalent to equality with deltaStar.
open scoped Pointwise
lemma section13_isMaxOn_dotProduct_iff_eq_deltaStar_of_mem {n : Nat} {C : Set (Fin n → Real)}
(hCne : C.Nonempty) (xStar x : Fin n → Real) (hxC : x ∈ C)
(hbdd : BddAbove (Set.image (fun y : Fin n → Real => dotProduct y xStar) C)) :
IsMaxOn (fun y : Fin n → Real => dotProduct y xStar) C x ↔
dotProduct x xStar = deltaStar C xStar
If -deltaStar (-xStar) ≠ deltaStar xStar, then a maximizer of dotProduct · xStar is not a minimizer.
open scoped Pointwise
lemma section13_exists_lt_dotProduct_of_ne_neg_deltaStar {n : Nat} {C : Set (Fin n → Real)}
(hC : Convex ℝ C) (x xStar : Fin n → Real) (hxC : x ∈ C)
(hxEq : dotProduct x xStar = deltaStar C xStar)
(hne : -(deltaStar C (-xStar)) ≠ deltaStar C xStar)
(hbdd : BddAbove (Set.image (fun y : Fin n → Real => dotProduct y xStar) C)) :
∃ y, y ∈ C ∧ dotProduct y xStar < dotProduct x xStar
Theorem 13.1: Let C be a convex set. Then x ∈ cl C if and only if ⟪x, x^*⟫ ≤ δ^*(x^* | C) for every x^*. Moreover, x ∈ ri C if and only if the same condition holds, but with strict inequality for each x^* such that -δ^*(-x^* | C) ≠ δ^*(x^* | C). One has x ∈ int C if and only if ⟪x, x^*⟫ < δ^*(x^* | C) for every x^*.
open scoped Pointwise
theorem mem_closure_iff_forall_dotProduct_le_deltaStar {n : Nat} (C : Set (Fin n → Real))
(x : Fin n → Real) (hC : Convex ℝ C) (hCne : C.Nonempty)
(hCbd : ∀ xStar : Fin n → Real,
BddAbove (Set.image (fun y : Fin n → Real => dotProduct y xStar) C)) :
x ∈ closure C ↔ ∀ xStar : Fin n → Real, dotProduct x xStar ≤ deltaStar C xStar
Theorem 13.1 (relative interior characterization): for convex C, membership in the relative (intrinsic) interior is equivalent to the supporting-inequality condition, with strict inequality for each x^* such that -δ^*(-x^* | C) ≠ δ^*(x^* | C).
open scoped Pointwise
theorem mem_intrinsicInterior_iff_forall_dotProduct_le_deltaStar_and_forall_of_ne {n : Nat}
(C : Set (Fin n → Real)) (x : Fin n → Real) (hC : Convex ℝ C) (hCne : C.Nonempty)
(hCbd : ∀ xStar : Fin n → Real,
BddAbove (Set.image (fun y : Fin n → Real => dotProduct y xStar) C)) :
x ∈ intrinsicInterior ℝ C ↔
x ∈ C ∧
(∀ xStar : Fin n → Real, dotProduct x xStar ≤ deltaStar C xStar) ∧
∀ xStar : Fin n → Real,
(-(deltaStar C (-xStar)) ≠ deltaStar C xStar) →
dotProduct x xStar < deltaStar C xStar
Theorem 13.1 (interior characterization): for convex C, membership in the (topological) interior is equivalent to strict supporting inequalities ⟪x, x^*⟫ < δ^*(x^* | C) for all x^*.
open scoped Pointwise
theorem mem_interior_iff_forall_dotProduct_lt_deltaStar {n : Nat} (C : Set (Fin n → Real))
(x : Fin n → Real) (hC : Convex ℝ C)
(hCbd : ∀ xStar : Fin n → Real,
BddAbove (Set.image (fun y : Fin n → Real => dotProduct y xStar) C)) :
x ∈ interior C ↔
x ∈ C ∧ ∀ xStar : Fin n → Real, xStar ≠ 0 → dotProduct x xStar < deltaStar C xStar
If the dot-product image of C is bounded above, then so is the dot-product image of closure C.
open scoped Pointwise
lemma section13_bddAbove_image_dotProduct_closure {n : Nat} (C : Set (Fin n → Real))
(xStar : Fin n → Real)
(hbdd : BddAbove (Set.image (fun x : Fin n → Real => dotProduct x xStar) C)) :
BddAbove (Set.image (fun x : Fin n → Real => dotProduct x xStar) (closure C))
Monotonicity of deltaStar under set inclusion (assuming nonempty domain and bounded range).
open scoped Pointwise
lemma section13_deltaStar_le_of_subset {n : Nat} {C₁ C₂ : Set (Fin n → Real)} (xStar : Fin n → Real)
(hsub : C₁ ⊆ C₂) (hC₁ne : C₁.Nonempty)
(hC₂bd : BddAbove (Set.image (fun x : Fin n → Real => dotProduct x xStar) C₂)) :
deltaStar C₁ xStar ≤ deltaStar C₂ xStar
Corollary 13.1.1. For convex sets C₁ and C₂ in ℝ^n, one has cl C₁ ⊆ cl C₂ if and only if δ^*(· | C₁) ≤ δ^*(· | C₂).
open scoped Pointwise
theorem closure_subset_iff_deltaStar_le {n : Nat} (C₁ C₂ : Set (Fin n → Real))
(hC₁ : Convex ℝ C₁) (hC₂ : Convex ℝ C₂) (hC₁ne : C₁.Nonempty) (hC₂ne : C₂.Nonempty)
(hC₁bd : ∀ xStar : Fin n → Real,
BddAbove (Set.image (fun y : Fin n → Real => dotProduct y xStar) C₁))
(hC₂bd : ∀ xStar : Fin n → Real,
BddAbove (Set.image (fun y : Fin n → Real => dotProduct y xStar) C₂)) :
closure C₁ ⊆ closure C₂ ↔ deltaStar C₁ ≤ deltaStar C₂
For a convex set C with well-defined support function, the set of points satisfying all supporting inequalities is the closure of C.
open scoped Pointwise
lemma section13_setOf_forall_dotProduct_le_deltaStar_eq_closure {n : Nat} (C : Set (Fin n → Real))
(hC : Convex ℝ C) (hCne : C.Nonempty)
(hCbd : ∀ xStar : Fin n → Real,
BddAbove (Set.image (fun y : Fin n → Real => dotProduct y xStar) C)) :
{x : Fin n → Real | ∀ xStar : Fin n → Real, dotProduct x xStar ≤ deltaStar C xStar} =
closure C
Text 13.1.2: Let C ⊆ ℝ^n be a closed convex set. Define D := { x | ∀ x^*, ⟪x, x^*⟫ ≤ δ^*(x^* | C) }. Then D = C. In particular, C is completely determined by its support function.
open scoped Pointwise
theorem setOf_forall_dotProduct_le_deltaStar_eq_of_isClosed {n : Nat} (C : Set (Fin n → Real))
(hC : Convex ℝ C) (hCclosed : IsClosed C) (hCne : C.Nonempty)
(hCbd : ∀ xStar : Fin n → Real,
BddAbove (Set.image (fun y : Fin n → Real => dotProduct y xStar) C)) :
{x : Fin n → Real | ∀ xStar : Fin n → Real, dotProduct x xStar ≤ deltaStar C xStar} = C
Text 13.1.2 (support function determines a closed convex set): if two closed convex sets have the same (finite) support function, then they are equal.
open scoped Pointwise
theorem eq_of_isClosed_of_convex_of_deltaStar_eq {n : Nat} {C₁ C₂ : Set (Fin n → Real)}
(hC₁ : Convex ℝ C₁) (hC₂ : Convex ℝ C₂) (hC₁closed : IsClosed C₁) (hC₂closed : IsClosed C₂)
(hC₁ne : C₁.Nonempty) (hC₂ne : C₂.Nonempty)
(hC₁bd : ∀ xStar : Fin n → Real,
BddAbove (Set.image (fun y : Fin n → Real => dotProduct y xStar) C₁))
(hC₂bd : ∀ xStar : Fin n → Real,
BddAbove (Set.image (fun y : Fin n → Real => dotProduct y xStar) C₂))
(hδ : deltaStar C₁ = deltaStar C₂) : C₁ = C₂
The dot-product image of a Minkowski sum is the pointwise sum of the dot-product images.
open scoped Pointwise
lemma section13_image_dotProduct_add_eq {n : Nat} (C₁ C₂ : Set (Fin n → Real))
(xStar : Fin n → Real) :
Set.image (fun x : Fin n → Real => dotProduct x xStar) (C₁ + C₂) =
Set.image (fun x : Fin n → Real => dotProduct x xStar) C₁ +
Set.image (fun x : Fin n → Real => dotProduct x xStar) C₂
If both dot-product image-sets are bounded above, then the supremum over their sum splits.
open scoped Pointwise
lemma section13_sSup_image_dotProduct_add {n : Nat} (C₁ C₂ : Set (Fin n → Real))
(xStar : Fin n → Real)
(hC₁ne : Set.Nonempty C₁) (hC₂ne : Set.Nonempty C₂)
(hbdd₁ : BddAbove (Set.image (fun x : Fin n → Real => dotProduct x xStar) C₁))
(hbdd₂ : BddAbove (Set.image (fun x : Fin n → Real => dotProduct x xStar) C₂)) :
sSup
(Set.image (fun x : Fin n → Real => dotProduct x xStar) C₁ +
Set.image (fun x : Fin n → Real => dotProduct x xStar) C₂) =
sSup (Set.image (fun x : Fin n → Real => dotProduct x xStar) C₁) +
sSup (Set.image (fun x : Fin n → Real => dotProduct x xStar) C₂)
Boundedness in direction xStar makes the Real-valued support function additive under sums.
open scoped Pointwise
lemma section13_deltaStar_add_of_bddAbove {n : Nat} (C₁ C₂ : Set (Fin n → Real))
(xStar : Fin n → Real)
(hC₁ne : Set.Nonempty C₁) (hC₂ne : Set.Nonempty C₂)
(hbdd₁ : BddAbove (Set.image (fun x : Fin n → Real => dotProduct x xStar) C₁))
(hbdd₂ : BddAbove (Set.image (fun x : Fin n → Real => dotProduct x xStar) C₂)) :
deltaStar (C₁ + C₂) xStar = deltaStar C₁ xStar + deltaStar C₂ xStar
Text 13.1.3: The support function of the sum of two non-empty convex sets C₁ and C₂ is given by δ^*(x^* | C₁ + C₂) = δ^*(x^* | C₁) + δ^*(x^* | C₂).
open scoped Pointwise
theorem deltaStar_add {n : Nat} (C₁ C₂ : Set (Fin n → Real)) (xStar : Fin n → Real)
(_hC₁ : Convex ℝ C₁) (_hC₂ : Convex ℝ C₂) (hC₁ne : Set.Nonempty C₁) (hC₂ne : Set.Nonempty C₂)
(hbdd₁ : BddAbove (Set.image (fun x : Fin n → Real => dotProduct x xStar) C₁))
(hbdd₂ : BddAbove (Set.image (fun x : Fin n → Real => dotProduct x xStar) C₂)) :
deltaStar (C₁ + C₂) xStar = deltaStar C₁ xStar + deltaStar C₂ xStar
For indicator functions, points outside the set contribute ⊥ to the Fenchel conjugate.
open scoped Pointwise
lemma section13_fenchelConjugate_indicatorFunction_eq_sSup_image_dotProduct {n : Nat}
(C : Set (Fin n → Real)) (xStar : Fin n → Real) :
fenchelConjugate n (indicatorFunction C) xStar =
sSup ((fun x : Fin n → Real => ((x ⬝ᵥ xStar : Real) : EReal)) '' C)
Coercion commutes with sSup for a nonempty, bounded-above set of reals.
open scoped Pointwise
lemma section13_sSup_image_coe_real_eq_coe_sSup (S : Set ℝ) (hne : S.Nonempty) (hbdd : BddAbove S) :
sSup ((fun r : ℝ => (r : EReal)) '' S) = ((sSup S : ℝ) : EReal)
Text 13.1.4: Let C ⊆ ℝ^n be a convex set, and let δ(· | C) be its indicator function, δ(x | C) = 0 for x ∈ C and δ(x | C) = +∞ for x ∉ C. Then the convex conjugate of δ(· | C) is the support function of C. More precisely, for every xStar, δ*(xStar | C) = sup_x (⟪x, xStar⟫ - δ(x | C)) = sup_{x ∈ C} ⟪x, xStar⟫.
open scoped Pointwise
theorem fenchelConjugate_indicatorFunction_eq_deltaStar {n : Nat} (C : Set (Fin n → Real))
(xStar : Fin n → Real) (hCne : C.Nonempty)
(hbdd : BddAbove (Set.image (fun x : Fin n → Real => dotProduct x xStar) C)) :
fenchelConjugate n (indicatorFunction C) xStar = ((deltaStar C xStar : ℝ) : EReal)
Under directional boundedness, the Fenchel conjugate of the indicator function equals the Real-valued support function (coerced to EReal).
open scoped Pointwise
lemma section13_fenchelConjugate_indicatorFunction_eq_deltaStar_fun {n : Nat} (C : Set (Fin n → Real))
(hCne : C.Nonempty)
(hCbd : ∀ xStar : Fin n → Real,
BddAbove (Set.image (fun x : Fin n → Real => dotProduct x xStar) C)) :
fenchelConjugate n (indicatorFunction C) =
fun xStar : Fin n → Real => ((deltaStar C xStar : ℝ) : EReal)
If x ∈ closure C, then any affine map majorized by indicatorFunction C is nonpositive at x.
open scoped Pointwise
lemma section13_affine_le_indicatorFunction_le_zero_on_closure {n : Nat} {C : Set (Fin n → Real)}
(h : AffineMap ℝ (Fin n → Real) Real)
(hh : ∀ y : Fin n → Real, (h y : EReal) ≤ indicatorFunction C y) :
∀ x : Fin n → Real, x ∈ closure C → h x ≤ 0§13 — Support Functions — Part 2
Lean authors: Zichen Wang, Wanli Ma, Siyuan Shao, Yunxi Duan, Zaiwen Wen.
If x ∉ closure C and C is a nonempty convex set, then for any real μ there exists an affine map bounded above by indicatorFunction C whose value at x exceeds μ.
open scoped Pointwise
lemma section13_exists_affineMap_le_indicatorFunction_gt_of_not_mem_closure {n : Nat}
(C : Set (Fin n → Real)) (hC : Convex ℝ C) (hCne : C.Nonempty) {x : Fin n → Real}
(hx : x ∉ closure C) (μ : Real) :
∃ h : AffineMap ℝ (Fin n → Real) Real,
(∀ y : Fin n → Real, (h y : EReal) ≤ indicatorFunction C y) ∧
(μ : EReal) < (h x : EReal)
Under directional boundedness, the conjugate of the (Real-valued) support function is the closed convex envelope of the indicator function.
open scoped Pointwise
lemma section13_fenchelConjugate_deltaStar_eq_clConv_indicatorFunction {n : Nat}
(C : Set (Fin n → Real)) (hCne : C.Nonempty)
(hCbd : ∀ xStar : Fin n → Real,
BddAbove (Set.image (fun x : Fin n → Real => dotProduct x xStar) C)) :
fenchelConjugate n (fun xStar : Fin n → Real => ((deltaStar C xStar : ℝ) : EReal)) =
clConv n (indicatorFunction C)
For a nonempty convex set C, the closed convex envelope of indicatorFunction C is the indicator function of closure C.
open scoped Pointwise
lemma section13_clConv_indicatorFunction_eq_indicatorFunction_closure {n : Nat}
(C : Set (Fin n → Real)) (hC : Convex ℝ C) (hCne : C.Nonempty) :
clConv n (indicatorFunction C) = indicatorFunction (closure C)
Text 13.1.5: Let C ⊆ ℝ^n be a convex set and let δ(· | C) be its indicator function. Then (δ^*(· | C))^* = cl δ(· | C) = δ(· | cl C). In this development, cl δ(·|C) is represented by clConv n (indicatorFunction C).
open scoped Pointwise
theorem fenchelConjugate_deltaStar_eq_clConv_indicatorFunction_eq_indicatorFunction_closure
{n : Nat} (C : Set (Fin n → Real)) (hC : Convex ℝ C) (hCne : C.Nonempty)
(hCbd : ∀ xStar : Fin n → Real,
BddAbove (Set.image (fun x : Fin n → Real => dotProduct x xStar) C)) :
fenchelConjugate n (fun xStar : Fin n → Real => ((deltaStar C xStar : ℝ) : EReal)) =
clConv n (indicatorFunction C) ∧
clConv n (indicatorFunction C) = indicatorFunction (closure C)
Text 13.1.4 (EReal version): the Fenchel conjugate of an indicator function is the EReal-valued support function.
open scoped Pointwise
lemma section13_fenchelConjugate_indicatorFunction_eq_supportFunctionEReal {n : Nat}
(C : Set (Fin n → Real)) :
fenchelConjugate n (indicatorFunction C) = supportFunctionEReal C
For a > 0, multiplication by (a : EReal) cancels with multiplication by (a⁻¹ : EReal).
open scoped Pointwise
lemma section13_mul_mul_inv_cancel_pos_real {a : Real} (ha : 0 < a) (z : EReal) :
((a : Real) : EReal) * (((a⁻¹ : Real) : EReal) * z) = z
For a > 0, multiplication by (a⁻¹ : EReal) cancels with multiplication by (a : EReal).
open scoped Pointwise
lemma section13_mul_inv_mul_cancel_pos_real {a : Real} (ha : 0 < a) (z : EReal) :
((a⁻¹ : Real) : EReal) * (((a : Real) : EReal) * z) = z
Multiplying by (a : EReal) for a > 0 commutes with iSup.
open scoped Pointwise
lemma section13_ereal_mul_iSup_of_pos {ι : Sort _} (a : Real) (ha : 0 < a) (f : ι → EReal) :
((a : Real) : EReal) * iSup f = iSup fun i => ((a : Real) : EReal) * f i
The Fenchel conjugate of a positive scalar multiple is the corresponding right scalar multiple of the Fenchel conjugate.
open scoped Pointwise
lemma section13_fenchelConjugate_smul_eq_rightScalarMultiple {n : Nat}
(f : (Fin n → Real) → EReal) (lam : Real) (hlam : 0 < lam) :
fenchelConjugate n (fun x => ((lam : Real) : EReal) * f x) =
rightScalarMultiple (fenchelConjugate n f) lam
A closed proper convex function is 0/⊤-valued iff its Fenchel conjugate is positively homogeneous.
open scoped Pointwise
lemma section13_only_zero_top_iff_fenchelConjugate_posHom {n : Nat}
(f : (Fin n → Real) → EReal) (hf_closed : ClosedConvexFunction f)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
(∀ x : Fin n → Real, f x = 0 ∨ f x = ⊤) ↔ PositivelyHomogeneous (fenchelConjugate n f)
A 0/⊤-valued function with no ⊥ values is an indicator function of its 0-sublevel.
open scoped Pointwise
lemma section13_eq_indicatorFunction_setOf_le_zero_of_only_zero_top {n : Nat}
(g : (Fin n → Real) → EReal) (hzeroTop : ∀ x : Fin n → Real, g x = 0 ∨ g x = ⊤) :
g = indicatorFunction {x : Fin n → Real | g x ≤ (0 : EReal)}
Theorem 13.2: The indicator function and the support function of a closed convex set are conjugate to each other. The functions which are the support functions of non-empty convex sets are the closed proper convex functions which are positively homogeneous.
open scoped Pointwise
theorem indicatorFunction_conjugate_supportFunctionEReal_of_isClosed {n : Nat}
(C : Set (Fin n → Real)) (hC : Convex ℝ C) (hCclosed : IsClosed C) :
fenchelConjugate n (indicatorFunction C) = supportFunctionEReal C ∧
fenchelConjugate n (supportFunctionEReal C) = indicatorFunction C
Theorem 13.2 (characterization of support functions): an EReal-valued function on ℝ^n is a support function of a nonempty convex set iff it is closed, proper, convex, and positively homogeneous.
open scoped Pointwise
theorem exists_supportFunctionEReal_iff_closedProperConvex_posHom {n : Nat}
(f : (Fin n → Real) → EReal) :
(∃ C : Set (Fin n → Real), Set.Nonempty C ∧ Convex ℝ C ∧ f = supportFunctionEReal C) ↔
(ClosedConvexFunction f ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f ∧ PositivelyHomogeneous f)
The set {xStar | ∀ x, ⟪x, xStar⟫ ≤ f x} is the 0-sublevel set of the Fenchel conjugate f*.
open scoped Pointwise
lemma section13_setOf_forall_dotProduct_le_eq_setOf_fenchelConjugate_le_zero {n : Nat}
(f : (Fin n → Real) → EReal) :
{xStar : Fin n → Real |
∀ x : Fin n → Real, ((dotProduct x xStar : Real) : EReal) ≤ f x} =
{xStar : Fin n → Real | fenchelConjugate n f xStar ≤ (0 : EReal)}
If there exists a strict supporting violation f x < ⟪x, xStar⟫, then f*(xStar) = ⊤.
open scoped Pointwise
lemma section13_fenchelConjugate_eq_top_of_exists_dotProduct_gt {n : Nat}
(f : (Fin n → Real) → EReal) (hpos : PositivelyHomogeneous f) (xStar : Fin n → Real)
(hx : ∃ x : Fin n → Real, f x < ((dotProduct x xStar : Real) : EReal)) :
fenchelConjugate n f xStar = (⊤ : EReal)
If all supporting inequalities ⟪x, xStar⟫ ≤ f x hold (and f is not identically ⊤), then f*(xStar) = 0.
open scoped Pointwise
lemma section13_fenchelConjugate_eq_zero_of_forall_dotProduct_le {n : Nat}
(f : (Fin n → Real) → EReal) (hpos : PositivelyHomogeneous f)
(hnotTop : ¬ (∀ x : Fin n → Real, f x = ⊤)) (xStar : Fin n → Real)
(hx : ∀ x : Fin n → Real, ((dotProduct x xStar : Real) : EReal) ≤ f x) :
fenchelConjugate n f xStar = (0 : EReal)
Corollary 13.2.1. Let f be a positively homogeneous convex function which is not identically ⊤ (+∞). Then cl f (here represented by clConv n f) is the support function of the closed convex set C = { xStar | ∀ x, ⟪x, xStar⟫ ≤ f x }.
open scoped Pointwise
theorem clConv_eq_supportFunctionEReal_setOf_forall_dotProduct_le {n : Nat}
(f : (Fin n → Real) → EReal) (hpos : PositivelyHomogeneous f) (_hconv : ConvexFunction f)
(hnotTop : ¬ (∀ x : Fin n → Real, f x = ⊤)) :
∃ C : Set (Fin n → Real),
IsClosed C ∧ Convex ℝ C ∧
clConv n f = supportFunctionEReal C ∧
C =
{xStar : Fin n → Real |
∀ x : Fin n → Real, ((dotProduct x xStar : Real) : EReal) ≤ f x}
Auxiliary definition: the EReal-valued support function xStar ↦ sup_{x ∈ C} ⟪x, xStar⟫.
open scoped Pointwise
noncomputable def supportFunctionEReal {n : Nat} (C : Set (Fin n → Real)) :
(Fin n → Real) → EReal :=
fun xStar => sSup {z : EReal | ∃ x ∈ C, z = ((dotProduct x xStar : Real) : EReal)}§13 — Support Functions — Part 3
Lean authors: Zichen Wang, Wanli Ma, Siyuan Shao, Yunxi Duan, Zaiwen Wen.
A crude upper bound for the dot product in terms of the sup norm.
open scoped Pointwise
lemma section13_dotProduct_le_norm_mul_sum_abs {n : Nat} (x y : Fin n → Real) :
dotProduct x y ≤ ‖x‖ * (Finset.univ.sum fun i : Fin n => |y i|)
A nonempty set has a support function that never takes the value ⊥.
open scoped Pointwise
lemma section13_supportFunctionEReal_ne_bot_of_nonempty {n : Nat} {C : Set (Fin n → Real)}
(hCne : Set.Nonempty C) (xStar : Fin n → Real) :
supportFunctionEReal C xStar ≠ (⊥ : EReal)
A bounded set has a support function that never takes the value ⊤.
open scoped Pointwise
lemma section13_supportFunctionEReal_ne_top_of_isBounded {n : Nat} {C : Set (Fin n → Real)}
(hCbd : Bornology.IsBounded C) (xStar : Fin n → Real) :
supportFunctionEReal C xStar ≠ (⊤ : EReal)
If the support function of a nonempty set is finite everywhere, then the set is bounded.
open scoped Pointwise
lemma section13_isBounded_of_supportFunctionEReal_finite {n : Nat} {C : Set (Fin n → Real)}
(hfinite :
∀ xStar,
supportFunctionEReal C xStar ≠ (⊤ : EReal) ∧ supportFunctionEReal C xStar ≠ (⊥ : EReal)) :
Bornology.IsBounded C
A convex function finite everywhere is automatically closed and proper.
open scoped Pointwise
lemma section13_closedProper_of_convex_finite {n : Nat} {f : (Fin n → Real) → EReal}
(hf : ConvexFunction f) (hfinite : ∀ x, f x ≠ ⊤ ∧ f x ≠ ⊥) :
ClosedConvexFunction f ∧ ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f
Corollary 13.2.2. The support functions of the non-empty bounded convex sets are the finite positively homogeneous convex functions.
open scoped Pointwise
theorem exists_supportFunctionEReal_iff_boundedConvex_posHom_finite {n : Nat}
(f : (Fin n → Real) → EReal) :
(∃ C : Set (Fin n → Real),
Set.Nonempty C ∧ Convex ℝ C ∧ Bornology.IsBounded C ∧ f = supportFunctionEReal C) ↔
(ConvexFunction f ∧ PositivelyHomogeneous f ∧ ∀ x : Fin n → Real, f x ≠ ⊤ ∧ f x ≠ ⊥)§13 — Support Functions — Part 4
Lean authors: Zichen Wang, Wanli Ma, Siyuan Shao, Yunxi Duan, Zaiwen Wen.
For a nonempty convex set C, its EReal-valued support function is closed, proper, convex, and positively homogeneous.
open scoped Pointwise
lemma section13_supportFunctionEReal_closedProperConvex_posHom {n : Nat} (C : Set (Fin n → Real))
(hCne : Set.Nonempty C) (hC : Convex ℝ C) :
ClosedConvexFunction (supportFunctionEReal C) ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (supportFunctionEReal C) ∧
PositivelyHomogeneous (supportFunctionEReal C)
The EReal-valued support function of a nonempty convex set never takes the value ⊥.
open scoped Pointwise
lemma section13_supportFunctionEReal_ne_bot {n : Nat} (C : Set (Fin n → Real))
(hCne : Set.Nonempty C) (hC : Convex ℝ C) :
∀ xStar : Fin n → Real, supportFunctionEReal C xStar ≠ (⊥ : EReal)
The EReal-valued support function of a nonempty convex set is subadditive.
open scoped Pointwise
lemma section13_supportFunctionEReal_subadditive {n : Nat} (C : Set (Fin n → Real))
(hCne : Set.Nonempty C) (hC : Convex ℝ C) :
∀ xStar₁ xStar₂ : Fin n → Real,
supportFunctionEReal C (xStar₁ + xStar₂) ≤
supportFunctionEReal C xStar₁ + supportFunctionEReal C xStar₂
Text 13.2.3: Let C ⊆ ℝ^n be a nonempty convex set. Then the (extended-real) support function is lower semicontinuous on ℝ^n, and it satisfies the subadditivity inequality δ^*(xStar₁ + xStar₂ | C) ≤ δ^*(xStar₁ | C) + δ^*(xStar₂ | C) for all xStar₁, xStar₂.
open scoped Pointwise
theorem deltaStar_lowerSemicontinuous_and_add_le {n : Nat} (C : Set (Fin n → Real))
(hCne : Set.Nonempty C) (hC : Convex ℝ C) :
LowerSemicontinuous (supportFunctionEReal C) ∧
∀ xStar₁ xStar₂ : Fin n → Real,
supportFunctionEReal C (xStar₁ + xStar₂) ≤
supportFunctionEReal C xStar₁ + supportFunctionEReal C xStar₂
Bridge Real.sqrt (dotProduct x x) to the Euclidean norm on EuclideanSpace.
open scoped Pointwise
lemma section13_sqrt_dotProduct_self_eq_norm_euclideanSpace {n : Nat} (x : Fin n → Real) :
Real.sqrt (dotProduct x x) =
‖((EuclideanSpace.equiv (Fin n) ℝ).symm x : EuclideanSpace ℝ (Fin n))‖
Cauchy--Schwarz in the dotProduct notation used throughout this section.
open scoped Pointwise
lemma section13_dotProduct_le_sqrt_mul_sqrt {n : Nat} (x y : Fin n → Real) :
dotProduct x y ≤ Real.sqrt (dotProduct x x) * Real.sqrt (dotProduct y y)
If y lies in the unit Euclidean ball, then dotProduct x y is bounded by the Euclidean norm of x.
open scoped Pointwise
lemma section13_dotProduct_le_sqrt_of_sqrt_dotProduct_self_le_one {n : Nat} (x y : Fin n → Real)
(hy : Real.sqrt (dotProduct y y) ≤ (1 : Real)) :
dotProduct x y ≤ Real.sqrt (dotProduct x x)
For any x, the supremum of y ↦ dotProduct x y over the unit Euclidean ball is attained by a normalized vector.
open scoped Pointwise
lemma section13_exists_dotProduct_eq_sqrt {n : Nat} (x : Fin n → Real) :
∃ y : Fin n → Real,
Real.sqrt (dotProduct y y) ≤ (1 : Real) ∧
dotProduct x y = Real.sqrt (dotProduct x x)
Text 13.2.4: Let B := { y ∈ ℝ^n | |y| ≤ 1 } be the unit Euclidean ball. Then for every x ∈ ℝ^n, |x| = sup { ⟪x, y⟫ | y ∈ B } = δ^*(x | B).
open scoped Pointwise
theorem euclideanNorm_eq_sSup_dotProduct_unitEuclideanBall_eq_deltaStar {n : Nat}
(x : Fin n → Real) :
let B : Set (Fin n → Real) := {y | Real.sqrt (dotProduct y y) ≤ (1 : Real)}
Real.sqrt (dotProduct x x) =
sSup (Set.image (fun y : Fin n → Real => dotProduct x y) B) ∧
sSup (Set.image (fun y : Fin n → Real => dotProduct x y) B) = deltaStar B x
Membership in {a} + λ • B can be rewritten using the explicit parametrization a + λ • y, with y ∈ B.
open scoped Pointwise
lemma section13_mem_singleton_add_smul_unitEuclideanBall_iff {n : Nat} (a z : Fin n → Real)
(lam : ℝ) :
let B : Set (Fin n → Real) := {y | Real.sqrt (dotProduct y y) ≤ (1 : Real)}
z ∈ ({a} : Set (Fin n → Real)) + (lam • B) ↔ ∃ y, y ∈ B ∧ z = a + lam • y
Every dot product over the translated/scaled unit Euclidean ball is bounded by ⟪x, a⟫ + λ‖x‖ when λ ≥ 0.
open scoped Pointwise
lemma section13_dotProduct_le_dotProduct_add_mul_sqrt_of_mem_singleton_add_smul_unitEuclideanBall
{n : Nat} (a x z : Fin n → Real) (lam : ℝ) (hlam : 0 ≤ lam) :
let B : Set (Fin n → Real) := {y | Real.sqrt (dotProduct y y) ≤ (1 : Real)}
z ∈ ({a} : Set (Fin n → Real)) + (lam • B) →
dotProduct x z ≤ dotProduct x a + lam * Real.sqrt (dotProduct x x)
The supremum of z ↦ ⟪x, z⟫ over {a} + λ • B is attained at a + λ • y, where y maximizes ⟪x, ·⟫ over B.
open scoped Pointwise
lemma section13_exists_mem_singleton_add_smul_unitEuclideanBall_dotProduct_eq {n : Nat}
(a x : Fin n → Real) (lam : ℝ) :
let B : Set (Fin n → Real) := {y | Real.sqrt (dotProduct y y) ≤ (1 : Real)}
∃ z ∈ ({a} : Set (Fin n → Real)) + (lam • B),
dotProduct x z = dotProduct x a + lam * Real.sqrt (dotProduct x x)
Text 13.2.5: Let B := { y ∈ ℝ^n | |y| ≤ 1 } be the unit Euclidean ball. More generally, for any a ∈ ℝ^n and λ ≥ 0, the support function of the ball a + λ B is given by δ^*(x | a + λ B) = ⟪x, a⟫ + λ |x|.
open scoped Pointwise
theorem deltaStar_singleton_add_smul_unitEuclideanBall {n : Nat} (a x : Fin n → Real) (lam : ℝ)
(hlam : 0 ≤ lam) :
let B : Set (Fin n → Real) := {y | Real.sqrt (dotProduct y y) ≤ (1 : Real)}
deltaStar (({a} : Set (Fin n → Real)) + (lam • B)) x =
dotProduct x a + lam * Real.sqrt (dotProduct x x)§13 — Support Functions — Part 5
Lean authors: Zichen Wang, Wanli Ma, Siyuan Shao, Yunxi Duan, Zaiwen Wen.
For a probability vector x (nonnegative with total mass 1), dotProduct x xStar is bounded above by the supremum of the coordinates of xStar.
open scoped Pointwise
lemma section13_dotProduct_le_sSup_coord_of_simplex {n : Nat} (x xStar : Fin n → Real)
(hx0 : ∀ j : Fin n, 0 ≤ x j)
(hsum : Finset.univ.sum (fun j : Fin n => x j) = (1 : Real)) :
dotProduct x xStar ≤ sSup (Set.range fun j : Fin n => xStar j)
On the finite type Fin n with 0 < n, the supremum of the coordinate range is achieved by some coordinate.
open scoped Pointwise
lemma section13_exists_coord_eq_sSup {n : Nat} (xStar : Fin n → Real) (hn : 0 < n) :
∃ j0 : Fin n, xStar j0 = sSup (Set.range fun j : Fin n => xStar j)
A simplex vertex Pi.single j0 1 has nonnegative coordinates summing to 1, and its dot product with xStar is the j0-th coordinate.
open scoped Pointwise
lemma section13_simplexVertex_mem_and_dotProduct {n : Nat} (j0 : Fin n) (xStar : Fin n → Real) :
(∀ j : Fin n, 0 ≤ Pi.single (M := fun _ : Fin n => Real) j0 (1 : Real) j) ∧
(Finset.univ.sum
(fun j : Fin n => Pi.single (M := fun _ : Fin n => Real) j0 (1 : Real) j) =
(1 : Real)) ∧
dotProduct (Pi.single (M := fun _ : Fin n => Real) j0 (1 : Real)) xStar = xStar j0
Text 13.2.6: As further examples, the support functions of the sets C₁ = {x = (ξ₁, …, ξₙ) | (∀ j, 0 ≤ ξⱼ) ∧ (ξ₁ + ⋯ + ξₙ = 1)}, C₂ = {x = (ξ₁, …, ξₙ) | |ξ₁| + ⋯ + |ξₙ| ≤ 1}, C₃ = {x = (ξ₁, ξ₂) | ξ₁ < 0 ∧ ξ₂ ≤ ξ₁⁻¹}, C₄ = {x = (ξ₁, ξ₂) | 2 ξ₁ + ξ₂² ≤ 0}, are given by the following formulas. (Here we use supportFunctionEReal so that ⊤ represents +∞.)
open scoped Pointwise
theorem supportFunctionEReal_C1_eq_sSup_coord {n : Nat} (xStar : Fin n → Real) :
0 < n →
let C₁ : Set (Fin n → Real) :=
{x | (∀ j : Fin n, 0 ≤ x j) ∧ (Finset.univ.sum (fun j : Fin n => x j) = (1 : Real))}
supportFunctionEReal C₁ xStar =
((sSup (Set.range fun j : Fin n => xStar j) : Real) : EReal)
For x in the ℓ¹ unit ball, dotProduct x xStar is bounded above by the ℓ∞ norm of xStar, expressed as sSup (range fun j ↦ |xStar j|).
open scoped Pointwise
lemma section13_dotProduct_le_sSup_abs_coord_of_l1Ball {n : Nat} (x xStar : Fin n → Real)
(hn : 0 < n) (hx : Finset.univ.sum (fun j : Fin n => |x j|) ≤ (1 : Real)) :
dotProduct x xStar ≤ sSup (Set.range fun j : Fin n => |xStar j|)
The signed coordinate vector Pi.single j0 (±1) has ℓ¹ norm 1, hence lies in the ℓ¹ unit ball.
open scoped Pointwise
lemma section13_l1Ball_single_sign_mem {n : Nat} (xStar : Fin n → Real) (j0 : Fin n) :
Finset.univ.sum
(fun j : Fin n =>
|Pi.single (M := fun _ : Fin n => Real) j0
(if 0 ≤ xStar j0 then (1 : Real) else -1) j|)
≤ (1 : Real)
The dot product of the signed coordinate vector Pi.single j0 (±1) with xStar equals |xStar j0|.
open scoped Pointwise
lemma section13_dotProduct_single_sign_eq_abs {n : Nat} (xStar : Fin n → Real) (j0 : Fin n) :
dotProduct
(Pi.single (M := fun _ : Fin n => Real) j0 (if 0 ≤ xStar j0 then (1 : Real) else -1))
xStar =
Text 13.2.6 (support function of C₂): for the ℓ¹ unit ball C₂ = {x | ∑ j, |xⱼ| ≤ 1}, the support function is max_j |xStarⱼ|.
open scoped Pointwise
theorem supportFunctionEReal_C2_eq_sSup_abs_coord {n : Nat} (xStar : Fin n → Real) :
0 < n →
let C₂ : Set (Fin n → Real) :=
{x | Finset.univ.sum (fun j : Fin n => |x j|) ≤ (1 : Real)}
supportFunctionEReal C₂ xStar =
((sSup (Set.range fun j : Fin n => |xStar j|) : Real) : EReal)
If one coordinate of xStar is negative, then the support function of C₃ is ⊤.
open scoped Pointwise
lemma section13_supportFunctionEReal_C3_eq_top_of_neg_coord (xStar : Fin 2 → Real)
(hneg : xStar 0 < 0 ∨ xStar 1 < 0) :
let C₃ : Set (Fin 2 → Real) := {x | x 0 < 0 ∧ x 1 ≤ (x 0)⁻¹}
supportFunctionEReal C₃ xStar = ⊤
For x ∈ C₃ and xStar ≥ 0, the dot product is bounded above by -2 * sqrt (xStar 0 * xStar 1).
open scoped Pointwise
lemma section13_dotProduct_le_supportBound_C3 (x xStar : Fin 2 → Real)
(hx0 : x 0 < 0) (hx1 : x 1 ≤ (x 0)⁻¹) (hxStar0 : 0 ≤ xStar 0) (hxStar1 : 0 ≤ xStar 1) :
dotProduct x xStar ≤ -2 * Real.sqrt (xStar 0 * xStar 1)
If xStar ≥ 0 and xStar 0 * xStar 1 = 0, then the support function of C₃ is 0.
open scoped Pointwise
lemma section13_supportFunctionEReal_C3_eq_zero_of_nonneg_and_mul_eq_zero (xStar : Fin 2 → Real)
(hxStar0 : 0 ≤ xStar 0) (hxStar1 : 0 ≤ xStar 1) (hprod : xStar 0 * xStar 1 = 0) :
let C₃ : Set (Fin 2 → Real) := {x | x 0 < 0 ∧ x 1 ≤ (x 0)⁻¹}
supportFunctionEReal C₃ xStar = (0 : EReal)
When xStar 0, xStar 1 > 0, the bound -2 * sqrt (xStar 0 * xStar 1) is attained on C₃.
open scoped Pointwise
lemma section13_exists_C3_point_dotProduct_eq_supportBound (xStar : Fin 2 → Real)
(hxStar0 : 0 < xStar 0) (hxStar1 : 0 < xStar 1) :
∃ x : Fin 2 → Real,
x 0 < 0 ∧ x 1 ≤ (x 0)⁻¹ ∧ dotProduct x xStar = -2 * Real.sqrt (xStar 0 * xStar 1)
Text 13.2.6 (support function of C₃): for C₃ = {x = (ξ₁, ξ₂) | ξ₁ < 0 ∧ ξ₂ ≤ ξ₁⁻¹}, one has δ^*(xStar | C₃) = -2 * sqrt (ξ₁^* ξ₂^*) if xStar ≥ 0, and +∞ otherwise.
open scoped Pointwise
theorem supportFunctionEReal_C3_eq_piecewise (xStar : Fin 2 → Real) :
let C₃ : Set (Fin 2 → Real) := {x | x 0 < 0 ∧ x 1 ≤ (x 0)⁻¹}
supportFunctionEReal C₃ xStar =
if (0 ≤ xStar 0 ∧ 0 ≤ xStar 1) then
((-2 * Real.sqrt (xStar 0 * xStar 1) : Real) : EReal)
else
⊤
For 0 < a, the concave quadratic t ↦ (-(t^2)/2) * a + t * b is bounded above by b^2 / (2 * a).
open scoped Pointwise
lemma section13_neg_mul_sq_div_two_add_mul_le (a b t : Real) (ha : 0 < a) :
(-(t ^ 2) / 2) * a + t * b ≤ b ^ 2 / (2 * a)
For x ∈ C₄ and 0 < xStar 0, the dot product is bounded above by (xStar 1)^2 / (2 * xStar 0).
open scoped Pointwise
lemma section13_dotProduct_le_supportBound_C4 (x xStar : Fin 2 → Real)
(hxC4 : (2 : Real) * x 0 + (x 1) ^ 2 ≤ 0) (hxStar0 : 0 < xStar 0) :
dotProduct x xStar ≤ (xStar 1) ^ 2 / (2 * xStar 0)
For 0 < xStar 0, there exists a point of C₄ attaining the bound (xStar 1)^2 / (2 * xStar 0) in the dot product.
open scoped Pointwise
lemma section13_exists_C4_point_dotProduct_eq_supportBound (xStar : Fin 2 → Real)
(hxStar0 : 0 < xStar 0) :
∃ x : Fin 2 → Real, ((2 : Real) * x 0 + (x 1) ^ 2 ≤ 0) ∧
dotProduct x xStar = (xStar 1) ^ 2 / (2 * xStar 0)
If xStar 0 < 0, then the support function of C₄ is ⊤.
open scoped Pointwise
lemma section13_supportFunctionEReal_C4_eq_top_of_xStar0_lt_zero (xStar : Fin 2 → Real)
(h0 : xStar 0 < 0) :
let C₄ : Set (Fin 2 → Real) := {x | (2 : Real) * x 0 + (x 1) ^ 2 ≤ 0}
supportFunctionEReal C₄ xStar = ⊤
If xStar 0 = 0 and xStar 1 ≠ 0, then the support function of C₄ is ⊤.
open scoped Pointwise
lemma section13_supportFunctionEReal_C4_eq_top_of_xStar0_eq_zero_and_xStar1_ne_zero
(xStar : Fin 2 → Real) (h0 : xStar 0 = 0) (h1 : xStar 1 ≠ 0) :
let C₄ : Set (Fin 2 → Real) := {x | (2 : Real) * x 0 + (x 1) ^ 2 ≤ 0}
supportFunctionEReal C₄ xStar = ⊤
If xStar = 0, then the support function of C₄ is 0.
open scoped Pointwise
lemma section13_supportFunctionEReal_C4_eq_zero_of_xStar_eq_zero (xStar : Fin 2 → Real)
(h0 : xStar 0 = 0) (h1 : xStar 1 = 0) :
let C₄ : Set (Fin 2 → Real) := {x | (2 : Real) * x 0 + (x 1) ^ 2 ≤ 0}
supportFunctionEReal C₄ xStar = (0 : EReal)
Text 13.2.6 (support function of C₄): for C₄ = {x = (ξ₁, ξ₂) | 2 ξ₁ + ξ₂² ≤ 0}, one has δ^*(xStar | C₄) = (ξ₂^*)² / (2 * ξ₁^*) if ξ₁^* > 0, 0 if xStar = 0, and +∞ otherwise.
open scoped Pointwise
theorem supportFunctionEReal_C4_eq_piecewise (xStar : Fin 2 → Real) :
let C₄ : Set (Fin 2 → Real) := {x | (2 : Real) * x 0 + (x 1) ^ 2 ≤ 0}
supportFunctionEReal C₄ xStar =
if xStar 0 > 0 then
(((xStar 1) ^ 2 / (2 * xStar 0) : Real) : EReal)
else if xStar 0 = 0 ∧ xStar 1 = 0 then
(0 : EReal)
else
⊤
For the EReal-valued support function, an upper bound by a real μ is equivalent to a pointwise upper bound on all dot products defining the supremum.
open scoped Pointwise
lemma section13_supportFunctionEReal_effectiveDomain_le_coe_iff {n : Nat}
(f : (Fin n → Real) → EReal) (y : Fin n → Real) (μ : Real) :
supportFunctionEReal (effectiveDomain (Set.univ : Set (Fin n → Real)) f) y ≤ (μ : EReal) ↔
∀ x ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) f, (dotProduct x y : Real) ≤ μ
For the sSup-of-differences recession formula, bounding by a real μ is equivalent to a uniform pointwise bound on each difference term.
open scoped Pointwise
lemma section13_recessionSup_le_coe_iff {n : Nat}
(g : (Fin n → Real) → EReal) (g0_plus : (Fin n → Real) → EReal)
(hg0_plus :
∀ y : Fin n → Real,
g0_plus y =
sSup
{r : EReal |
∃ x ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) g,
r = g (x + y) - g x})
(y : Fin n → Real) (μ : Real) :
g0_plus y ≤ (μ : EReal) ↔
∀ x ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) g, g (x + y) - g x ≤ (μ : EReal)
The dot product is additive in the second argument, hence it commutes with nsmul there.
open scoped Pointwise
lemma section13_dotProduct_nsmul_right {n : Nat} (x y : Fin n → Real) (m : ℕ) :
dotProduct x (m • y) = (m : Real) * dotProduct x y
If all dot products ⟪x,y⟫ over x ∈ dom f are bounded by μ, then every difference f*(xStar+y) - f*(xStar) with xStar ∈ dom f* is bounded by μ.
open scoped Pointwise
lemma section13_fenchelConjugate_add_le_add_supportFunctionEReal {n : Nat}
(f : (Fin n → Real) → EReal)
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
{y : Fin n → Real} {μ : Real}
(hxy :
∀ x ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) f, (dotProduct x y : Real) ≤ μ) :
∀ xStar ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f),
fenchelConjugate n f (xStar + y) - fenchelConjugate n f xStar ≤ (μ : EReal)
Theorem 13.3: Let f be a proper convex function. The support function of dom f is the recession function (f^*)0+ of the conjugate f^*. If f is closed, the support function of dom f^* is the recession function f0+ of f.
open scoped Pointwise
theorem supportFunctionEReal_effectiveDomain_eq_recession_fenchelConjugate {n : Nat}
(f : (Fin n → Real) → EReal)
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(fStar0_plus : (Fin n → Real) → EReal)
(hfStar0_plus :
∀ y : Fin n → Real,
fStar0_plus y =
sSup
{r : EReal |
∃ xStar ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f),
r = fenchelConjugate n f (xStar + y) - fenchelConjugate n f xStar}) :
supportFunctionEReal (effectiveDomain (Set.univ : Set (Fin n → Real)) f) = fStar0_plus
Theorem 13.3 (closed case): if f is closed, then the support function of dom f^* is the recession function f0+ of f.
open scoped Pointwise
theorem supportFunctionEReal_effectiveDomain_fenchelConjugate_eq_recession {n : Nat}
(f : (Fin n → Real) → EReal)
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hclosed : ClosedConvexFunction f)
(f0_plus : (Fin n → Real) → EReal)
(hf0_plus :
∀ y : Fin n → Real,
f0_plus y =
sSup
{r : EReal |
∃ x ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) f,
r = f (x + y) - f x}) :
supportFunctionEReal (effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f)) =
f0_plus§13 — Support Functions — Part 6
Lean authors: Zichen Wang, Wanli Ma, Siyuan Shao, Yunxi Duan, Zaiwen Wen.
Text 13.3.1: A convex function f is called co-finite if f is closed and proper and epi f contains no non-vertical half-lines; equivalently, its recession function f₀⁺ satisfies f₀⁺(y) = +∞ for all y ≠ 0.
open scoped Pointwise
def CoFiniteConvexFunction {n : Nat} (f : (Fin n → Real) → EReal) : Prop :=
ClosedConvexFunction f ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f ∧
∀ y : Fin n → Real,
y ≠ (0 : Fin n → Real) →
sSup
{r : EReal |
∃ x ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) f,
r = f (x + y) - f x} =
(⊤ : EReal)
If f is proper on univ, then its effective domain is nonempty.
open scoped Pointwise
lemma section13_effectiveDomain_nonempty_of_proper {n : Nat} {f : (Fin n → Real) → EReal} :
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f →
Set.Nonempty (effectiveDomain (Set.univ : Set (Fin n → Real)) f)
A bounded set cannot contain all iterates x + n • y for a nonzero vector y.
open scoped Pointwise
lemma section13_exists_mem_add_not_mem_effectiveDomain_of_isBounded {n : Nat}
{D : Set (Fin n → Real)} (hbounded : Bornology.IsBounded D) (hne : Set.Nonempty D)
{y : Fin n → Real} (hy : y ≠ (0 : Fin n → Real)) :
∃ x ∈ D, x + y ∉ D
If x ∈ dom f but x+y ∉ dom f, then the recession sSup in direction y is ⊤.
open scoped Pointwise
lemma section13_sSup_recession_eq_top_of_exists_not_mem {n : Nat} {f : (Fin n → Real) → EReal}
{y : Fin n → Real} :
(∃ x ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) f,
x + y ∉ effectiveDomain (Set.univ : Set (Fin n → Real)) f) →
sSup
{r : EReal |
∃ x ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) f, r = f (x + y) - f x} =
(⊤ : EReal)
Text 13.3.2: Let f : ℝ^n → (-∞, +∞] be closed, proper, and convex. If dom f is bounded, then f is co-finite.
open scoped Pointwise
theorem coFiniteConvexFunction_of_isBounded_effectiveDomain {n : Nat}
(f : (Fin n → Real) → EReal) (hclosed : ClosedConvexFunction f)
(hproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hbounded : Bornology.IsBounded (effectiveDomain (Set.univ : Set (Fin n → Real)) f)) :
CoFiniteConvexFunction f
The support function of Set.univ is ⊤ in any nonzero direction.
open scoped Pointwise
lemma section13_supportFunctionEReal_univ_eq_top_of_ne_zero {n : Nat} {y : Fin n → Real}
(hy : y ≠ (0 : Fin n → Real)) :
supportFunctionEReal (Set.univ : Set (Fin n → Real)) y = (⊤ : EReal)
If f* is finite everywhere (no ⊥ and dom f* = univ), then f is proper.
open scoped Pointwise
lemma section13_proper_of_conjugate_finite_everywhere {n : Nat} {f : (Fin n → Real) → EReal}
(hclosed : ClosedConvexFunction f)
(hfiniteStar :
effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f) = Set.univ ∧
∀ xStar : Fin n → Real, fenchelConjugate n f xStar ≠ (⊥ : EReal)) :
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f
Corollary 13.3.1. Let f be a closed convex function on ℝ^n. Then f^* is finite everywhere (equivalently dom f^* = ℝ^n) if and only if f is co-finite.
open scoped Pointwise
theorem effectiveDomain_fenchelConjugate_eq_univ_iff_coFinite {n : Nat}
(f : (Fin n → Real) → EReal) (hclosed : ClosedConvexFunction f) :
(effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f) = Set.univ ∧
∀ xStar : Fin n → Real, fenchelConjugate n f xStar ≠ (⊥ : EReal)) ↔
CoFiniteConvexFunction f
For the EReal-valued support function, an upper bound by a real μ is equivalent to a pointwise upper bound on all dot products defining the supremum.
open scoped Pointwise
lemma section13_supportFunctionEReal_le_coe_iff {n : Nat} (C : Set (Fin n → Real))
(y : Fin n → Real) (μ : Real) :
supportFunctionEReal C y ≤ (μ : EReal) ↔ ∀ x ∈ C, (dotProduct x y : Real) ≤ μ
If C is nonempty and supportFunctionEReal C y ≠ ⊤, then supportFunctionEReal C y is the coercion of its toReal.
open scoped Pointwise
lemma section13_supportFunctionEReal_coe_toReal {n : Nat} {C : Set (Fin n → Real)}
(hCne : C.Nonempty) {y : Fin n → Real} (hy : supportFunctionEReal C y ≠ ⊤) :
((supportFunctionEReal C y).toReal : EReal) = supportFunctionEReal C y
On an affine subspace, if the support function is finite at y, then it is symmetric.
open scoped Pointwise
lemma section13_supportFunctionEReal_affineSubspace_symmetry {n : Nat}
(A : AffineSubspace ℝ (Fin n → Real)) (hAne : (A : Set (Fin n → Real)).Nonempty)
{y : Fin n → Real} (hy : supportFunctionEReal (A : Set (Fin n → Real)) y ≠ ⊤) :
-supportFunctionEReal (A : Set (Fin n → Real)) (-y) =
supportFunctionEReal (A : Set (Fin n → Real)) y
A convex set is affine iff its EReal support function is symmetric on the directions where it is finite.
open scoped Pointwise
lemma section13_isAffineSet_iff_supportFunction_symmetry {n : Nat} {C : Set (Fin n → Real)}
(hCconv : Convex ℝ C) (hCne : C.Nonempty) :
(∃ A : AffineSubspace ℝ (Fin n → Real), (A : Set (Fin n → Real)) = C) ↔
∀ y : Fin n → Real,
supportFunctionEReal C y ≠ ⊤ →
-supportFunctionEReal C (-y) = supportFunctionEReal C y
The support function of dom f* is the recession function recessionFunction f.
open scoped Pointwise
lemma section13_supportFunctionEReal_dom_fenchelConjugate_eq_recessionFunction {n : Nat}
(f : (Fin n → Real) → EReal) (hclosed : ClosedConvexFunction f)
(hproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
supportFunctionEReal
(effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f)) =
recessionFunction f
Corollary 13.3.2. Let f be a closed proper convex function. In order that dom f* be an affine set, it is necessary and sufficient that (f0+)(y) = +∞ for every y which is not actually in the linearity space of f.
open scoped Pointwise
theorem effectiveDomain_fenchelConjugate_isAffine_iff_forall_recessionFunction_eq_top_of_not_mem
{n : Nat} (f : (Fin n → Real) → EReal) (hclosed : ClosedConvexFunction f)
(hproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
(∃ A : AffineSubspace ℝ (Fin n → Real),
(A : Set (Fin n → Real)) =
effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f)) ↔
∀ y : Fin n → Real, y ∉ linearitySpace f → recessionFunction f y = (⊤ : EReal)
Auxiliary definition: the recession function f₀⁺ of an EReal-valued function f on ℝ^n, given by f₀⁺(y) = sup { f(x+y) - f(x) | x ∈ dom f }.
open scoped Pointwise
noncomputable def recessionFunction {n : Nat} (f : (Fin n → Real) → EReal) :
(Fin n → Real) → EReal :=
fun y =>
sSup
{r : EReal |
∃ x ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) f, r = f (x + y) - f x}
Auxiliary definition: the linearity space of a function f, defined as the set of directions y for which the recession function is finite and symmetric: -(f₀⁺)(-y) = (f₀⁺)(y) whenever (f₀⁺)(y) < +∞.
open scoped Pointwise
def linearitySpace {n : Nat} (f : (Fin n → Real) → EReal) : Set (Fin n → Real) :=
{y | recessionFunction f y ≠ (⊤ : EReal) ∧ -(recessionFunction f (-y)) = recessionFunction f y}§13 — Support Functions — Part 7
Lean authors: Zichen Wang, Wanli Ma, Siyuan Shao, Yunxi Duan, Zaiwen Wen.
If f is finite everywhere and satisfies LipschitzCondition f α, then its recession function is bounded above by α‖y‖.
open scoped Pointwise
lemma section13_recessionFunction_le_mul_norm_of_LipschitzCondition {n : Nat}
{f : (Fin n → Real) → EReal} {α : ℝ}
(hfinite : ∀ z : Fin n → Real, f z ≠ ⊤ ∧ f z ≠ ⊥) (hLip : LipschitzCondition f α) :
∀ y : Fin n → Real, recessionFunction f y ≤ ((α * ‖y‖ : Real) : EReal)
If a nonempty convex set has support function ⊤ in every nonzero direction, then the set is all of ℝ^n.
open scoped Pointwise
lemma section13_set_eq_univ_of_supportFunctionEReal_eq_top {n : Nat} (C : Set (Fin n → Real))
(hconv : Convex ℝ C) (hne : Set.Nonempty C)
(htop : ∀ y : Fin n → Real, y ≠ (0 : Fin n → Real) → supportFunctionEReal C y = (⊤ : EReal)) :
C = Set.univ
If f is proper convex and dom f* is bounded, then f is finite everywhere.
open scoped Pointwise
lemma section13_finite_everywhere_of_isBounded_effectiveDomain_fenchelConjugate {n : Nat}
{f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hbounded :
Bornology.IsBounded
(effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f))) :
∀ z : Fin n → Real, f z ≠ ⊤ ∧ f z ≠ ⊥
If C is bounded, then its support function is dominated by R‖y‖ for some R ≥ 0 with respect to the sup norm on ℝ^n.
open scoped Pointwise
lemma section13_supportFunctionEReal_le_mul_norm_of_isBounded {n : Nat} (C : Set (Fin n → Real))
(hbounded : Bornology.IsBounded C) :
∃ R ≥ 0,
∀ y : Fin n → Real, supportFunctionEReal C y ≤ ((R * ‖y‖ : Real) : EReal)
If f is finite everywhere and f₀⁺(y) ≤ α‖y‖ for all y, then f satisfies the global Lipschitz condition with constant α.
open scoped Pointwise
lemma section13_lipschitzCondition_of_recessionFunction_le_mul_norm {n : Nat}
{f : (Fin n → Real) → EReal} {α : ℝ}
(hfinite : ∀ z : Fin n → Real, f z ≠ ⊤ ∧ f z ≠ ⊥)
(hrec : ∀ y : Fin n → Real, recessionFunction f y ≤ ((α * ‖y‖ : Real) : EReal)) :
LipschitzCondition f α
If f is proper convex and dom f* is bounded, then f satisfies a global Lipschitz condition for some α ≥ 0.
open scoped Pointwise
lemma section13_lipschitzCondition_of_isBounded_effectiveDomain_fenchelConjugate {n : Nat}
{f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hbounded :
Bornology.IsBounded
(effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f)))
(hfinite : ∀ z : Fin n → Real, f z ≠ ⊤ ∧ f z ≠ ⊥) :
∃ α : ℝ, 0 ≤ α ∧ LipschitzCondition f α
Corollary 13.3.3. Let f be a proper convex function. The effective domain dom f^* is bounded if and only if f is finite everywhere and there exists α ≥ 0 such that |f(z) - f(x)| ≤ α * ‖z - x‖ for all z and x (a global Lipschitz condition).
open scoped Pointwise
theorem isBounded_effectiveDomain_fenchelConjugate_iff_finite_and_exists_lipschitz {n : Nat}
(f : (Fin n → Real) → EReal)
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
Bornology.IsBounded
(effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f)) ↔
(∀ z : Fin n → Real, f z ≠ ⊤ ∧ f z ≠ ⊥) ∧ ∃ α : ℝ, 0 ≤ α ∧ LipschitzCondition f α
Hölder-style estimate for the sup-norm: ⟪x,y⟫ ≤ ‖x‖₁‖y‖.
open scoped Pointwise
lemma section13_dotProduct_le_l1Norm_mul_norm {n : Nat} (x y : Fin n → Real) :
dotProduct x y ≤ l1Norm x * ‖y‖
l1Norm is bounded above on a bounded set.
open scoped Pointwise
lemma section13_bddAbove_image_l1Norm_of_isBounded {n : Nat} (C : Set (Fin n → Real))
(hbounded : Bornology.IsBounded C) :
BddAbove (Set.image (fun x : Fin n → Real => l1Norm x) C)
If xStar ∈ dom f*, then ‖xStar‖₁ ≤ conjugateDomainRadius f.
open scoped Pointwise
lemma section13_l1Norm_le_conjugateDomainRadius_of_mem_dom {n : Nat} (f : (Fin n → Real) → EReal)
(hbounded :
Bornology.IsBounded
(effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f))) :
∀ xStar : Fin n → Real,
xStar ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f) →
l1Norm xStar ≤ conjugateDomainRadius f
The support function of dom f* is bounded by conjugateDomainRadius f * ‖y‖.
open scoped Pointwise
lemma section13_supportFunctionEReal_le_mul_norm_conjugateDomainRadius {n : Nat}
(f : (Fin n → Real) → EReal)
(hbounded :
Bornology.IsBounded
(effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f))) :
∀ y : Fin n → Real,
supportFunctionEReal
(effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f)) y ≤
((conjugateDomainRadius f * ‖y‖ : Real) : EReal)
If f is finite everywhere and satisfies a global Lipschitz condition with constant α, then any xStar with α < ‖xStar‖₁ lies outside dom f* (equivalently, f* xStar = ⊤).
open scoped Pointwise
lemma section13_fenchelConjugate_eq_top_of_LipschitzCondition_lt_l1Norm {n : Nat}
{f : (Fin n → Real) → EReal} {α : ℝ}
(hfinite : ∀ z : Fin n → Real, f z ≠ ⊤ ∧ f z ≠ ⊥) (hα0 : 0 ≤ α)
(hLip : LipschitzCondition f α) :
∀ xStar : Fin n → Real, α < l1Norm xStar → fenchelConjugate n f xStar = ⊤
Corollary 13.3.3 (optimal Lipschitz constant): assuming f is finite everywhere and dom f^* is bounded, the smallest α for which the global Lipschitz condition holds is α = sup { ‖xStar‖ | xStar ∈ dom f^* } (here conjugateDomainRadius f).
open scoped Pointwise
theorem lipschitzCondition_conjugateDomainRadius_and_minimal {n : Nat}
(f : (Fin n → Real) → EReal)
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hbounded :
Bornology.IsBounded
(effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f)))
(hfinite : ∀ z : Fin n → Real, f z ≠ ⊤ ∧ f z ≠ ⊥) :
LipschitzCondition f (conjugateDomainRadius f) ∧
∀ α : ℝ, 0 ≤ α → LipschitzCondition f α → conjugateDomainRadius f ≤ α
The EReal-valued affine functional x ↦ ⟪x, a⟫ is proper convex on univ.
open scoped Pointwise
lemma section13_properConvexFunctionOn_dotProduct {n : Nat} (a : Fin n → Real) :
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real))
(fun x : Fin n → Real => ((dotProduct x a : Real) : EReal))
Theorem 12.3 specialization: conjugating x ↦ f x - ⟪x, xStar⟫ translates the conjugate.
open scoped Pointwise
lemma section13_fenchelConjugate_sub_dotProduct {n : Nat}
(f : (Fin n → Real) → EReal) (xStar : Fin n → Real) :
let g : (Fin n → Real) → EReal := fun x => f x - ((dotProduct x xStar : Real) : EReal)
fenchelConjugate n g = fun yStar => fenchelConjugate n f (yStar + xStar)
The effective domain of (f - ⟪·, xStar⟫)^* is the translate dom f^* - xStar.
open scoped Pointwise
lemma section13_effectiveDomain_fenchelConjugate_sub_dotProduct {n : Nat}
(f : (Fin n → Real) → EReal) (xStar : Fin n → Real) :
let domFstar : Set (Fin n → Real) :=
effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f)
let g : (Fin n → Real) → EReal := fun x => f x - ((dotProduct x xStar : Real) : EReal)
effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n g) =
domFstar - ({xStar} : Set (Fin n → Real))
Translating a set preserves membership in closure (specialized to translation by -a).
open scoped Pointwise
lemma section13_translate_mem_closure_iff {n : Nat} (S : Set (Fin n → Real)) (a : Fin n → Real) :
(a ∈ closure S ↔ (0 : Fin n → Real) ∈ closure (S - ({a} : Set (Fin n → Real))))
Translating a set preserves membership in interior (specialized to translation by -a).
open scoped Pointwise
lemma section13_translate_mem_interior_iff {n : Nat} (S : Set (Fin n → Real)) (a : Fin n → Real) :
(a ∈ interior S ↔ (0 : Fin n → Real) ∈ interior (S - ({a} : Set (Fin n → Real))))
Translating a set preserves membership in intrinsicInterior.
open scoped Pointwise
lemma section13_translate_mem_intrinsicInterior_iff {n : Nat} (S : Set (Fin n → Real))
(a : Fin n → Real) :
(a ∈ intrinsicInterior ℝ S ↔
(0 : Fin n → Real) ∈ intrinsicInterior ℝ (S - ({a} : Set (Fin n → Real))))
Translating a set preserves membership in its affineSpan.
open scoped Pointwise
lemma section13_translate_mem_affineSpan_iff {n : Nat} (S : Set (Fin n → Real)) (a : Fin n → Real) :
(a ∈ (affineSpan ℝ S : Set (Fin n → Real)) ↔
(0 : Fin n → Real) ∈
(affineSpan ℝ (S - ({a} : Set (Fin n → Real))) : Set (Fin n → Real)))
Translation invariance of closure, intrinsic interior, interior and affine span at the origin.
open scoped Pointwise
lemma section13_translate_mem_closure_ri_interior_affineSpan_iff_zero {n : Nat}
(S : Set (Fin n → Real)) (a : Fin n → Real) :
(a ∈ closure S ↔ (0 : Fin n → Real) ∈ closure (S - ({a} : Set (Fin n → Real)))) ∧
(a ∈ intrinsicInterior ℝ S ↔
(0 : Fin n → Real) ∈ intrinsicInterior ℝ (S - ({a} : Set (Fin n → Real)))) ∧
(a ∈ interior S ↔ (0 : Fin n → Real) ∈ interior (S - ({a} : Set (Fin n → Real)))) ∧
(a ∈ (affineSpan ℝ S : Set (Fin n → Real)) ↔
(0 : Fin n → Real) ∈
(affineSpan ℝ (S - ({a} : Set (Fin n → Real))) : Set (Fin n → Real)))
Support-function characterization of (0 : ℝ^n) ∈ cl C for convex nonempty C.
open scoped Pointwise
lemma section13_zero_mem_closure_iff_forall_zero_le_supportFunctionEReal {n : Nat}
(C : Set (Fin n → Real)) (hCconv : Convex ℝ C) (hCne : C.Nonempty) :
(0 : Fin n → Real) ∈ closure C ↔ ∀ y : Fin n → Real, (0 : EReal) ≤ supportFunctionEReal C y
Support-function characterization of (0 : ℝ^n) ∈ int C for convex nonempty C.
open scoped Pointwise
lemma section13_zero_mem_interior_iff_forall_supportFunctionEReal_pos {n : Nat}
(C : Set (Fin n → Real)) (hCconv : Convex ℝ C) (hCne : C.Nonempty) :
(0 : Fin n → Real) ∈ interior C ↔
∀ y : Fin n → Real, y ≠ 0 → supportFunctionEReal C y > (0 : EReal)
Auxiliary definition: the global Lipschitz condition |f(z) - f(x)| ≤ α * ‖z - x‖, expressed via toReal for an EReal-valued function.
open scoped Pointwise
def LipschitzCondition {n : Nat} (f : (Fin n → Real) → EReal) (α : ℝ) : Prop :=
∀ z x : Fin n → Real, |(f z).toReal - (f x).toReal| ≤ α * ‖z - x‖
Auxiliary definition: the ℓ¹-norm on ℝ^n (with ℝ^n represented as Fin n → ℝ).
open scoped Pointwise
noncomputable def l1Norm {n : Nat} (x : Fin n → Real) : ℝ :=
Finset.univ.sum fun i : Fin n => ‖x i‖
Auxiliary definition: the quantity sup { ‖xStar‖₁ | xStar ∈ dom f^* }, where dom f^* is the effective domain of the Fenchel conjugate f^*.
open scoped Pointwise
noncomputable def conjugateDomainRadius {n : Nat} (f : (Fin n → Real) → EReal) : ℝ :=
sSup
(Set.image (fun xStar : Fin n → Real => l1Norm xStar)
(effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f)))§13 — Support Functions — Part 8
Lean authors: Zichen Wang, Wanli Ma, Siyuan Shao, Yunxi Duan, Zaiwen Wen.
Support-function characterization of (0 : ℝ^n) ∈ ri C for convex nonempty C.
open scoped Pointwise
lemma section13_zero_mem_intrinsicInterior_iff_supportFunctionEReal {n : Nat}
(C : Set (Fin n → Real)) (hCconv : Convex ℝ C) (hCne : C.Nonempty) :
(0 : Fin n → Real) ∈ intrinsicInterior ℝ C ↔
∀ y : Fin n → Real,
(¬ (-(supportFunctionEReal C (-y)) = supportFunctionEReal C y ∧ supportFunctionEReal C y = 0)) →
supportFunctionEReal C y > (0 : EReal)
Support-function characterization of 0 ∈ affineSpan C.
open scoped Pointwise
lemma section13_zero_mem_affineSpan_iff_supportFunctionEReal {n : Nat}
(C : Set (Fin n → Real)) (hCne : C.Nonempty) :
(0 : Fin n → Real) ∈ (affineSpan ℝ C : Set (Fin n → Real)) ↔
∀ y : Fin n → Real,
(-(supportFunctionEReal C (-y)) = supportFunctionEReal C y) →
supportFunctionEReal C y = 0
Corollary 13.3.4. Let f be a closed proper convex function. Let xStar be a fixed vector and let g x = f x - ⟪x, xStar⟫. Then: (a) xStar ∈ cl (dom f^*) if and only if (g₀⁺) y ≥ 0 for every y; (b) xStar ∈ ri (dom f^*) if and only if (g₀⁺) y > 0 for all y except those satisfying -(g₀⁺ (-y)) = (g₀⁺) y = 0; (c) xStar ∈ int (dom f^*) if and only if (g₀⁺) y > 0 for every y ≠ 0; (d) xStar ∈ aff (dom f^*) if and only if (g₀⁺) y = 0 for every y such that -(g₀⁺ (-y)) = (g₀⁺) y. Here f^* is fenchelConjugate n f, its domain is effectiveDomain univ (fenchelConjugate n f), and g₀⁺ is recessionFunction g.
open scoped Pointwise
theorem mem_closure_ri_interior_affineSpan_effectiveDomain_fenchelConjugate_iff_recessionFunction
{n : Nat} (f : (Fin n → Real) → EReal) (hclosed : ClosedConvexFunction f)
(hproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) (xStar : Fin n → Real) :
(let domFstar :
Set (Fin n → Real) :=
effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f)
let g : (Fin n → Real) → EReal := fun x => f x - ((dotProduct x xStar : Real) : EReal)
(xStar ∈ closure domFstar ↔ ∀ y : Fin n → Real, 0 ≤ recessionFunction g y) ∧
(xStar ∈ intrinsicInterior ℝ domFstar ↔
∀ y : Fin n → Real,
(¬
(-(recessionFunction g (-y)) = recessionFunction g y ∧
recessionFunction g y = 0)) →
recessionFunction g y > 0) ∧
(xStar ∈ interior domFstar ↔
∀ y : Fin n → Real, y ≠ 0 → recessionFunction g y > 0) ∧
(xStar ∈ (affineSpan ℝ domFstar : Set (Fin n → Real)) ↔
∀ y : Fin n → Real,
(-(recessionFunction g (-y)) = recessionFunction g y) →
recessionFunction g y = 0))
Rewrite linearitySpace (f*) using Theorem 13.3: it is a finiteness+symmetry condition on the support function of dom f.
open scoped Pointwise
lemma section13_linearitySpace_fenchelConjugate_iff_supportFunctionEReal_effectiveDomain {n : Nat}
(f : (Fin n → Real) → EReal) (hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(y : Fin n → Real) :
y ∈ linearitySpace (fenchelConjugate n f) ↔
supportFunctionEReal (effectiveDomain (Set.univ : Set (Fin n → Real)) f) y ≠ ⊤ ∧
(-(supportFunctionEReal (effectiveDomain (Set.univ : Set (Fin n → Real)) f) (-y)) =
supportFunctionEReal (effectiveDomain (Set.univ : Set (Fin n → Real)) f) y)
If a support function is finite and symmetric at y, then the dot-product functional x ↦ ⟪x,y⟫ is constant on the underlying set.
open scoped Pointwise
lemma section13_supportFunctionEReal_symm_finite_imp_dotProduct_const {n : Nat}
{C : Set (Fin n → Real)} (hCne : C.Nonempty) {y : Fin n → Real}
(hy : supportFunctionEReal C y ≠ ⊤)
(hsymm : -(supportFunctionEReal C (-y)) = supportFunctionEReal C y) :
∃ μ : Real, ∀ x ∈ C, dotProduct x y = μ
The dot-product functional x ↦ ⟪x,y⟫ is constant on a nonempty set C iff y is orthogonal to the direction of affineSpan C.
open scoped Pointwise
lemma section13_dotProduct_const_iff_mem_orthogonalComplement_direction_affineSpan {n : Nat}
{C : Set (Fin n → Real)} (hCne : C.Nonempty) (y : Fin n → Real) :
(∃ μ : Real, ∀ x ∈ C, dotProduct x y = μ) ↔
y ∈ orthogonalComplement n ((affineSpan ℝ C).direction)
Finite-dimensional formula: dim(Lᗮ) = n - dim(L) for the book’s orthogonalComplement.
open scoped Pointwise
lemma section13_finrank_orthogonalComplement {n : Nat} (L : Submodule ℝ (Fin n → Real)) :
Module.finrank ℝ (orthogonalComplement n L) = n - Module.finrank ℝ L§13 — Support Functions — Part 9
Lean authors: Zichen Wang, Wanli Ma, Siyuan Shao, Yunxi Duan, Zaiwen Wen.
Theorem 13.4 (primal part): for proper convex f, the linearity space of f* is the orthogonal complement of the direction of affineSpan (dom f).
open scoped Pointwise
lemma section13_span_linearitySpace_fenchelConjugate_eq_orthogonal_direction_affineSpan_effectiveDomain
{n : Nat} (f : (Fin n → Real) → EReal)
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
Submodule.span ℝ (linearitySpace (fenchelConjugate n f)) =
orthogonalComplement n
((affineSpan ℝ (effectiveDomain (Set.univ : Set (Fin n → Real)) f)).direction)
Theorem 13.4: Let f be a proper convex function on ℝ^n. The linearity space of f^* is the orthogonal complement of the subspace parallel to aff (dom f). Dually, if f is closed, the subspace parallel to aff (dom f^*) is the orthogonal complement of the linearity space of f, and one has linearity f^* = n - dimension f, and dimension f^* = n - linearity f.
open scoped Pointwise
theorem linearitySpace_fenchelConjugate_eq_orthogonal_direction_affineSpan_effectiveDomain_and_dual
{n : Nat} (f : (Fin n → Real) → EReal)
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
(let domF : Set (Fin n → Real) := effectiveDomain (Set.univ : Set (Fin n → Real)) f
let domFstar : Set (Fin n → Real) :=
effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f)
(Submodule.span ℝ (linearitySpace (fenchelConjugate n f)) =
orthogonalComplement n ((affineSpan ℝ domF).direction)) ∧
(ClosedConvexFunction f →
(affineSpan ℝ domFstar).direction =
orthogonalComplement n (Submodule.span ℝ (linearitySpace f)) ∧
(Module.finrank ℝ (Submodule.span ℝ (linearitySpace (fenchelConjugate n f))) =
n - Module.finrank ℝ ((affineSpan ℝ domF).direction)) ∧
(Module.finrank ℝ ((affineSpan ℝ domFstar).direction) =
n - Module.finrank ℝ (Submodule.span ℝ (linearitySpace f)))))§13 — Support Functions — Part 10
Lean authors: Zichen Wang, Wanli Ma, Siyuan Shao, Yunxi Duan, Zaiwen Wen.
For any f : ℝ^n → (-∞, +∞], the direction of affineSpan (dom f) has finrank at most n.
open scoped Pointwise
lemma section13_finrank_direction_affineSpan_effectiveDomain_le {n : Nat} (f : (Fin n → Real) → EReal) :
Module.finrank ℝ
((affineSpan ℝ (effectiveDomain (Set.univ : Set (Fin n → Real)) f)).direction) ≤ n
The book-defined rank dim(aff(dom f)) - lin(f) is invariant under Fenchel conjugation.
open scoped Pointwise
lemma section13_rank_eq_rank_fenchelConjugate {n : Nat} {f : (Fin n → Real) → EReal}
(hf_closed : ClosedConvexFunction f)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
(let rank (h : (Fin n → Real) → EReal) : Nat :=
Module.finrank ℝ
(affineSpan ℝ (effectiveDomain (Set.univ : Set (Fin n → Real)) h)).direction -
Module.finrank ℝ (Submodule.span ℝ (linearitySpace h));
rank (fenchelConjugate n f) = rank f)
Corollary 13.4.1. Closed proper convex functions conjugate to each other have the same rank.
open scoped Pointwise
theorem rank_eq_of_conjugate {n : Nat} {f g : (Fin n → Real) → EReal}
(hf_closed : ClosedConvexFunction f)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hg_closed : ClosedConvexFunction g)
(hg_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) g)
(hfg : fenchelConjugate n f = g) (hgf : fenchelConjugate n g = f) :
(let rank (h : (Fin n → Real) → EReal) : Nat :=
Module.finrank ℝ
(affineSpan ℝ (effectiveDomain (Set.univ : Set (Fin n → Real)) h)).direction -
Module.finrank ℝ (Submodule.span ℝ (linearitySpace h));
rank f = rank g)
The affine hyperplane {x | ⟪y, x⟫ = a} as an AffineSubspace.
open scoped Pointwise
def section13_dotProductHyperplane {n : Nat} (y : Fin n → Real) (a : Real) :
AffineSubspace ℝ (Fin n → Real)
If f is affine along a line, then its Fenchel conjugate is ⊤ off the corresponding hyperplane in the dual variable.
open scoped Pointwise
lemma section13_fenchelConjugate_eq_top_of_affineLine {n : Nat} (f : (Fin n → Real) → EReal)
{x y xStar : Fin n → Real} {a b : Real}
(hline : ∀ t : ℝ, f (x + t • y) = ((a * t + b : ℝ) : EReal))
(hcoeff : dotProduct y xStar ≠ a) :
fenchelConjugate n f xStar = ⊤
If f is affine along a line with direction y ≠ 0, then dom (f*) has empty interior.
open scoped Pointwise
lemma section13_not_nonempty_interior_dom_fenchelConjugate_of_exists_affineLine {n : Nat}
(f : (Fin n → Real) → EReal)
(hline :
∃ (x y : Fin n → Real),
y ≠ 0 ∧ ∃ a b : ℝ, ∀ t : ℝ, f (x + t • y) = ((a * t + b : ℝ) : EReal)) :
¬ Set.Nonempty (interior (effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f)))
Addition by a real constant as an order isomorphism on EReal.
open scoped Pointwise
def section13_addRightOrderIso (c : ℝ) : EReal ≃o EReal
sSup commutes with addition by a real constant on EReal.
open scoped Pointwise
lemma section13_sSup_image_add_right (c : ℝ) (s : Set EReal) :
sSup ((fun z : EReal => z + (c : EReal)) '' s) = sSup s + (c : EReal)
If the dot product xStar ↦ ⟪xStar, y⟫ is constant on dom g, then fenchelConjugate n g is affine along the line x + t • y.
open scoped Pointwise
lemma section13_fenchelConjugate_translate_of_dotProduct_const {n : Nat}
(g : (Fin n → Real) → EReal)
(hg_bot : ∀ xStar : Fin n → Real, g xStar ≠ (⊥ : EReal))
{y : Fin n → Real} {μ : Real}
(hμ :
∀ xStar ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) g, dotProduct xStar y = μ)
(x : Fin n → Real) (t : ℝ) :
fenchelConjugate n g (x + t • y) = fenchelConjugate n g x + ((t * μ : ℝ) : EReal)
For the book’s orthogonalComplement, one has Lᗮ = ⊤ iff L = ⊥.
open scoped Pointwise
lemma section13_orthogonalComplement_eq_top_iff {n : Nat} (L : Submodule ℝ (Fin n → Real)) :
orthogonalComplement n L = ⊤ ↔ L = ⊥
In finite dimension, the book orthogonal complement is involutive: (Lᗮ)ᗮ = L.
open scoped Pointwise
lemma section13_orthogonalComplement_orthogonalComplement_eq {n : Nat}
(L : Submodule ℝ (Fin n → Real)) :
orthogonalComplement n (orthogonalComplement n L) = L
If dom (f*) has empty interior, then linearitySpace f contains a nonzero direction.
open scoped Pointwise
lemma section13_exists_nonzero_mem_linearitySpace_of_not_nonempty_interior_dom_fenchelConjugate
{n : Nat} (f : (Fin n → Real) → EReal) (hclosed : ClosedConvexFunction f)
(hproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hno : ¬ Set.Nonempty
(interior (effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f)))) :
∃ y : Fin n → Real, y ≠ 0 ∧ y ∈ linearitySpace f
A nonzero direction in linearitySpace f yields an affine line along which f is finite and affine.
open scoped Pointwise
lemma section13_exists_affineLine_of_mem_linearitySpace {n : Nat} (f : (Fin n → Real) → EReal)
(hclosed : ClosedConvexFunction f)
(hproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
{y : Fin n → Real} (hy0 : y ≠ 0) (hy : y ∈ linearitySpace f) :
∃ (x y : Fin n → Real),
y ≠ 0 ∧ ∃ a b : ℝ, ∀ t : ℝ, f (x + t • y) = ((a * t + b : ℝ) : EReal)
Corollary 13.4.2. Let f be a closed proper convex function. Then dom f^* has a non-empty interior if and only if there are no lines along which f is (finite and) affine. Here f^* is the Fenchel conjugate fenchelConjugate n f, and dom f^* is its effective domain effectiveDomain univ (fenchelConjugate n f). The absence of a line along which f is finite and affine is expressed by the nonexistence of x and a nonzero direction y such that t ↦ f (x + t • y) agrees with an affine function of t.
open scoped Pointwise
theorem effectiveDomain_fenchelConjugate_interior_nonempty_iff_not_exists_affineLine {n : Nat}
(f : (Fin n → Real) → EReal) (hclosed : ClosedConvexFunction f)
(hproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
Set.Nonempty
(interior (effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f))) ↔
¬ ∃ (x y : Fin n → Real),
y ≠ 0 ∧ ∃ a b : ℝ, ∀ t : ℝ, f (x + t • y) = ((a * t + b : ℝ) : EReal)
Rewriting a dot-product level set as a ≤ 0 sublevel set by moving affine terms across the inequality in EReal.
open scoped Pointwise
lemma section13_levelSet_eq_setOf_sub_dotProduct_sub_le_zero {n : Nat}
(h : (Fin n → Real) → EReal) (bStar : Fin n → Real) (β : ℝ) :
{x | h x ≤ ((β : ℝ) : EReal) + ((dotProduct x bStar : Real) : EReal)} =
{x |
h x - ((dotProduct x bStar : Real) : EReal) - ((β : ℝ) : EReal) ≤ (0 : EReal)}
Subtracting a real constant from the primal function adds that constant to its Fenchel conjugate.
open scoped Pointwise
lemma section13_fenchelConjugate_sub_const {n : Nat} (g : (Fin n → Real) → EReal) (β : ℝ) :
fenchelConjugate n (fun x => g x - ((β : ℝ) : EReal)) =
fun xStar => fenchelConjugate n g xStar + ((β : ℝ) : EReal)
Text 13.4.3: Let h : ℝ^n → (-∞, +∞] be a proper convex function, let bStar ∈ ℝ^n, and let β ∈ ℝ. Define the level set C := {x | h x ≤ β + ⟪x, bStar⟫} and the shifted function f x := h x - ⟪x, bStar⟫ - β. Then C = {x | f x ≤ 0}. Moreover, the Fenchel conjugate of f satisfies f^*(xStar) = h^*(xStar + bStar) + β for all xStar ∈ ℝ^n.
open scoped Pointwise
theorem levelSet_le_add_dotProduct_eq_setOf_sub_dotProduct_sub_le_zero_and_fenchelConjugate
{n : Nat} (h : (Fin n → Real) → EReal)
(bStar : Fin n → Real) (β : ℝ) :
(let C : Set (Fin n → Real) :=
{x | h x ≤ ((β : ℝ) : EReal) + ((dotProduct x bStar : Real) : EReal)}
let f : (Fin n → Real) → EReal :=
fun x => h x - ((dotProduct x bStar : Real) : EReal) - ((β : ℝ) : EReal)
C = {x | f x ≤ (0 : EReal)} ∧
∀ xStar : Fin n → Real,
fenchelConjugate n f xStar =
fenchelConjugate n h (xStar + bStar) + ((β : ℝ) : EReal))
The linear functional x ↦ ⟪x, xStar⟫ is positively homogeneous when viewed as an EReal-valued function.
open scoped Pointwise
lemma section13_posHom_dotProduct {n : Nat} (xStar : Fin n → Real) :
PositivelyHomogeneous (fun x : Fin n → Real => ((dotProduct x xStar : Real) : EReal))
The linear functional x ↦ ⟪x, xStar⟫ is convex on ℝ^n (its epigraph is a half-space).
open scoped Pointwise
lemma section13_convexFunctionOn_dotProduct {n : Nat} (xStar : Fin n → Real) :
ConvexFunctionOn (Set.univ : Set (Fin n → Real))
(fun x : Fin n → Real => ((dotProduct x xStar : Real) : EReal))
Majorants of the positively homogeneous hull coincide with majorants of the original convex function.
open scoped Pointwise
lemma section13_setOf_forall_dotProduct_le_posHomGenerated_eq {n : Nat}
{f : (Fin n → Real) → EReal} (hf : ConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
let k := positivelyHomogeneousConvexFunctionGenerated f
{xStar : Fin n → Real |
∀ x : Fin n → Real, ((dotProduct x xStar : Real) : EReal) ≤ k x} =
{xStar : Fin n → Real |
∀ x : Fin n → Real, ((dotProduct x xStar : Real) : EReal) ≤ f x}
The closed positively homogeneous hull of a closed proper convex function f is the support function of the polar set {xStar | f*(xStar) ≤ 0}.
open scoped Pointwise
lemma section13_clConv_posHomGenerated_eq_supportFunctionEReal_setOf_fenchelConjugate_le_zero
{n : Nat} (f : (Fin n → Real) → EReal)
(hf_closed : ClosedConvexFunction f)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
clConv n (positivelyHomogeneousConvexFunctionGenerated f) =
supportFunctionEReal {xStar : Fin n → Real | fenchelConjugate n f xStar ≤ (0 : EReal)}
The support function of the 0-sublevel set of f is the closed positively homogeneous hull of f*.
open scoped Pointwise
lemma section13_supportFunctionEReal_setOf_le_zero_eq_clConv_posHomGenerated_fenchelConjugate
{n : Nat} (f : (Fin n → Real) → EReal)
(hf_closed : ClosedConvexFunction f)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
supportFunctionEReal {x : Fin n → Real | f x ≤ (0 : EReal)} =
clConv n (positivelyHomogeneousConvexFunctionGenerated (fenchelConjugate n f))
Theorem 13.5. Let f be a closed proper convex function. The support function of { x | f x ≤ 0 } is cl g, where g is the positively homogeneous convex function generated by f*. Dually, the closure of the positively homogeneous convex function k generated by f is the support function of { xStar | f*(xStar) ≤ 0 }. Here f* is the Fenchel conjugate fenchelConjugate n f; the support function is represented by supportFunctionEReal; the closure cl is represented by clConv n; and the positively homogeneous convex function generated by a function h is positivelyHomogeneousConvexFunctionGenerated h.
open scoped Pointwise
theorem supportFunctionEReal_setOf_le_zero_eq_clConv_posHomGenerated_fenchelConjugate_and_dual
{n : Nat} (f : (Fin n → Real) → EReal)
(hf_closed : ClosedConvexFunction f)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
supportFunctionEReal {x : Fin n → Real | f x ≤ (0 : EReal)} =
clConv n (positivelyHomogeneousConvexFunctionGenerated (fenchelConjugate n f)) ∧
clConv n (positivelyHomogeneousConvexFunctionGenerated f) =
supportFunctionEReal {xStar : Fin n → Real | fenchelConjugate n f xStar ≤ (0 : EReal)}§13 — Support Functions — Part 11
Lean authors: Zichen Wang, Wanli Ma, Siyuan Shao, Yunxi Duan, Zaiwen Wen.
Splitting a dot product in ℝ^{n+1} into its head and tail parts.
open scoped Pointwise
lemma section13_dotProduct_head_succ {n : Nat} (v w : Fin (n + 1) → Real) :
dotProduct v w =
v 0 * w 0 +
dotProduct (fun i : Fin n => v i.succ) (fun i : Fin n => w i.succ)
The set C = { (λ^*, x^*) | λ^* ≤ -f^*(x^*) } ⊆ ℝ^{n+1} from Corollary 13.5.1.
open scoped Pointwise
def section13_supportSet {n : Nat} (f : (Fin n → Real) → EReal) : Set (Fin (n + 1) → Real) :=
{vStar |
((vStar 0 : Real) : EReal) ≤ -(fenchelConjugate n f (fun i : Fin n => vStar i.succ))}
For λ > 0, the support function of section13_supportSet f at (λ,x) is the Fenchel conjugate of the scaled conjugate λ • f^*.
open scoped Pointwise
lemma section13_supportFunctionEReal_supportSet_pos {n : Nat} (f : (Fin n → Real) → EReal)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(v : Fin (n + 1) → Real) (hpos : 0 < v 0) :
supportFunctionEReal (section13_supportSet (n := n) f) v =
fenchelConjugate n
(fun xStar : Fin n → Real =>
((v 0 : Real) : EReal) * fenchelConjugate n f xStar)
(fun i : Fin n => v i.succ)
For λ > 0, supportFunctionEReal (section13_supportSet f) matches the rightScalarMultiple branch of Corollary 13.5.1.
open scoped Pointwise
lemma section13_supportFunctionEReal_supportSet_pos_eq_rightScalarMultiple {n : Nat}
(f : (Fin n → Real) → EReal) (hf_closed : ClosedConvexFunction f)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(v : Fin (n + 1) → Real) (hpos : 0 < v 0) :
supportFunctionEReal (section13_supportSet (n := n) f) v =
rightScalarMultiple f (v 0) (fun i : Fin n => v i.succ)
For λ = 0, the support function of section13_supportSet f reduces to the support function of dom f^*, evaluated at the tail vector.
open scoped Pointwise
lemma section13_supportFunctionEReal_supportSet_zero {n : Nat} (f : (Fin n → Real) → EReal)
(v : Fin (n + 1) → Real) (hzero : v 0 = 0) :
supportFunctionEReal (section13_supportSet (n := n) f) v =
supportFunctionEReal
(effectiveDomain (Set.univ : Set (Fin n → Real)) (fenchelConjugate n f))
(fun i : Fin n => v i.succ)
For λ < 0, the support function of section13_supportSet f is ⊤.
open scoped Pointwise
lemma section13_supportFunctionEReal_supportSet_neg {n : Nat} (f : (Fin n → Real) → EReal)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(v : Fin (n + 1) → Real) (hneg : v 0 < 0) :
supportFunctionEReal (section13_supportSet (n := n) f) v = ⊤
Corollary 13.5.1. Let f be a closed proper convex function on ℝ^n. Define a function k : ℝ^{n+1} → (-∞, +∞] by k(λ, x) = (f λ)(x) if λ > 0, k(0, x) = (f0^+)(x) if λ = 0, and k(λ, x) = +∞ if λ < 0. Then k is the support function of the set C = { (λ^*, x^*) | λ^* ≤ -f^*(x^*) } ⊆ ℝ^{n+1}.
open scoped Pointwise
theorem supportFunctionEReal_setOf_le_neg_fenchelConjugate_eq_piecewise_rightScalarMultiple
{n : Nat} (f : (Fin n → Real) → EReal) (hf_closed : ClosedConvexFunction f)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
(let k : (Fin (n + 1) → Real) → EReal :=
fun v =>
let lam : Real := v 0
let x : Fin n → Real := fun i => v i.succ
if _hpos : 0 < lam then
rightScalarMultiple f lam x
else if _hzero : lam = 0 then
recessionFunction f x
else
⊤
let C : Set (Fin (n + 1) → Real) :=
{vStar |
((vStar 0 : Real) : EReal) ≤
-(fenchelConjugate n f (fun i : Fin n => vStar i.succ))}
k = supportFunctionEReal C)
Move mulVec from the right argument of dotProduct to the left.
open scoped Pointwise
lemma section13_dotProduct_mulVec_right {n : Nat} (A : Matrix (Fin n) (Fin n) ℝ)
(x y : Fin n → ℝ) :
dotProduct x (A.mulVec y) = dotProduct (A.transpose.mulVec x) y
Support function under a linear image y ↦ A y: δ^*(x | A '' C) = δ^*(Aᵀ x | C).
open scoped Pointwise
lemma section13_deltaStar_image_mulVec {n : Nat} (A : Matrix (Fin n) (Fin n) ℝ)
(C : Set (Fin n → ℝ)) (xStar : Fin n → ℝ) :
deltaStar ((fun y => A.mulVec y) '' C) xStar = deltaStar C (A.transpose.mulVec xStar)
The support function of a singleton is just the dot product.
open scoped Pointwise
lemma section13_deltaStar_singleton {n : Nat} (a xStar : Fin n → ℝ) :
deltaStar ({a} : Set (Fin n → ℝ)) xStar = dotProduct xStar a
Quadratic forms from positive definite matrices are nonnegative.
open scoped Pointwise
lemma section13_posDef_dotProduct_mulVec_nonneg {n : Nat} {Q : Matrix (Fin n) (Fin n) ℝ}
(hQ : Q.PosDef) (x : Fin n → ℝ) :
0 ≤ dotProduct x (Q.mulVec x)
A real positive definite matrix factors as Bᵀ * B with B invertible.
open scoped Pointwise
open scoped MatrixOrder
lemma section13_posDef_exists_isUnit_transpose_mul_self {n : Nat} {Q : Matrix (Fin n) (Fin n) ℝ}
(hQ : Q.PosDef) :
∃ B : Matrix (Fin n) (Fin n) ℝ, IsUnit B ∧ Q = B.transpose * B
Completing the square for the elliptic quadratic inequality.
open scoped Pointwise
open scoped MatrixOrder
lemma section13_ellipticSet_completeSquare_mem {n : Nat} (Q : Matrix (Fin n) (Fin n) ℝ)
(a : Fin n → ℝ) (α : ℝ) (hQ : Q.PosDef) :
let C : Set (Fin n → ℝ) :=
{x | (1 / 2 : ℝ) * dotProduct x (Q.mulVec x) + dotProduct a x + α ≤ 0}
let b : Fin n → ℝ := -((Q⁻¹).mulVec a)
let β : ℝ := (1 / 2 : ℝ) * dotProduct a ((Q⁻¹).mulVec a) - α
∀ x, x ∈ C ↔ (1 / 2 : ℝ) * dotProduct (x - b) (Q.mulVec (x - b)) ≤ β
Nonemptiness of the elliptic set forces the radius parameter β to be nonnegative.
open scoped Pointwise
open scoped MatrixOrder
lemma section13_beta_nonneg_of_nonempty {n : Nat} (Q : Matrix (Fin n) (Fin n) ℝ) (a : Fin n → ℝ)
(α : ℝ) (hQ : Q.PosDef) :
let C : Set (Fin n → ℝ) :=
{x | (1 / 2 : ℝ) * dotProduct x (Q.mulVec x) + dotProduct a x + α ≤ 0}
let β : ℝ := (1 / 2 : ℝ) * dotProduct a ((Q⁻¹).mulVec a) - α
Set.Nonempty C → 0 ≤ β
The set {z | (1/2) * ⟪z, z⟫ ≤ β} is a scaling of the unit Euclidean ball.
open scoped Pointwise
open scoped MatrixOrder
lemma section13_setOf_half_dotProduct_self_le_eq_smul_unitEuclideanBall {n : Nat} (β : ℝ)
(hβ : 0 ≤ β) :
{z : Fin n → ℝ | (1 / 2 : ℝ) * dotProduct z z ≤ β} =
(Real.sqrt (2 * β)) • {y : Fin n → ℝ | Real.sqrt (dotProduct y y) ≤ (1 : ℝ)}
Boundedness in direction xStar for the centered ellipsoid {x | (1/2) ⟪x, Qx⟫ ≤ β}.
open scoped Pointwise
open scoped MatrixOrder
lemma section13_bddAbove_image_dotProduct_centeredSublevel_posDef {n : Nat}
(Q : Matrix (Fin n) (Fin n) ℝ) (β : ℝ) (xStar : Fin n → ℝ) (hQ : Q.PosDef) (hβ : 0 ≤ β) :
BddAbove (Set.image (fun x : Fin n → ℝ => dotProduct x xStar)
{x : Fin n → ℝ | (1 / 2 : ℝ) * dotProduct x (Q.mulVec x) ≤ β})
Support function of the centered ellipsoid {x | (1/2) ⟪x, Qx⟫ ≤ β} for positive definite Q.
open scoped Pointwise
open scoped MatrixOrder
lemma section13_deltaStar_centeredSublevel_posDef {n : Nat} (Q : Matrix (Fin n) (Fin n) ℝ) (β : ℝ)
(xStar : Fin n → ℝ) (hQ : Q.PosDef) (hβ : 0 ≤ β) :
deltaStar {x : Fin n → ℝ | (1 / 2 : ℝ) * dotProduct x (Q.mulVec x) ≤ β} xStar =
Real.sqrt (2 * β * dotProduct xStar ((Q⁻¹).mulVec xStar))
Text 13.5.2: For the “elliptic” convex set C = {x | (1/2) ⟪x, Qx⟫ + ⟪a, x⟫ + α ≤ 0}, where Q is a positive definite symmetric matrix, one has (assuming C ≠ ∅) δ^*(xStar | C) = ⟪b, xStar⟫ + (2β ⟪xStar, Q⁻¹ xStar⟫)^{1/2}, where b = -Q⁻¹ a and β = (1/2) ⟪a, Q⁻¹ a⟫ - α.
open scoped Pointwise
open scoped MatrixOrder
theorem deltaStar_ellipticSet_eq {n : Nat} (Q : Matrix (Fin n) (Fin n) ℝ) (a : Fin n → ℝ)
(α : ℝ) (xStar : Fin n → ℝ) (hQ : Q.PosDef) :
let C : Set (Fin n → ℝ) :=
{x |
(1 / 2 : ℝ) * dotProduct x (Q.mulVec x) + dotProduct a x + α ≤ 0}
let b : Fin n → ℝ := -((Q⁻¹).mulVec a)
let β : ℝ := (1 / 2 : ℝ) * dotProduct a ((Q⁻¹).mulVec a) - α
Set.Nonempty C →
deltaStar C xStar =
dotProduct xStar b + Real.sqrt (2 * β * dotProduct xStar ((Q⁻¹).mulVec xStar))§14 — Polars of Convex Sets — Part 1
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yuhao Jiang, Zaiwen Wen.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
noncomputable local instance section14_instTopologicalSpace_dualWeak_part1 :
TopologicalSpace (Module.Dual ℝ E)
Text 14.0.1: The polar of a non-empty convex cone K is the set Kᵒ = {x★ | ∀ x ∈ K, ⟪x, x★⟫ ≤ 0}. In this formalization, we interpret x★ as a linear functional Module.Dual ℝ E and the pairing ⟪x, x★⟫ as evaluation x★ x.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
def polarCone (K : Set E) : Set (Module.Dual ℝ E) :=
(PointedCone.dual (R := ℝ) (-LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) K : Set _)
The barrier cone of a set C is the set of all linear functionals φ that are bounded above on C, i.e. there exists β with φ x ≤ β for every x ∈ C.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
def Set.barrierCone (C : Set E) : Set (Module.Dual ℝ E) :=
{φ : Module.Dual ℝ E | ∃ β : ℝ, ∀ x ∈ C, φ x ≤ β}
An EReal-valued indicator function: 0 on a set and ⊤ off it.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
noncomputable def erealIndicator (K : Set E) : E → EReal
The (Fenchel–Legendre) conjugate of an EReal-valued function with respect to a bilinear pairing p, defined as sup_x (p x y - f x).
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
noncomputable def fenchelConjugateBilin {F : Type*} [AddCommGroup F] [Module ℝ F]
(p : E →ₗ[ℝ] F →ₗ[ℝ] ℝ) (f : E → EReal) : F → EReal :=
fun y => sSup (Set.range fun x => (p x y : EReal) - f x)
Membership in the dual cone ProperCone.dual (-p) K is exactly the inequality p x y ≤ 0 for all x ∈ K.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma mem_dual_neg_iff {F : Type*}
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E]
[TopologicalSpace F] [AddCommGroup F] [IsTopologicalAddGroup F] [Module ℝ F]
[ContinuousSMul ℝ F]
(p : E →ₗ[ℝ] F →ₗ[ℝ] ℝ) [(-p).IsContPerfPair] (K : ProperCone ℝ E) {y : F} :
y ∈ ProperCone.dual (-p) (K : Set E) ↔ ∀ x ∈ (K : Set E), p x y ≤ 0
The Fenchel conjugate of the EReal-indicator of a cone is 0 at points in the dual cone.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma fenchelConjugate_erealIndicator_eq_zero_of_mem_dual {F : Type*}
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E]
[TopologicalSpace F] [AddCommGroup F] [IsTopologicalAddGroup F] [Module ℝ F]
[ContinuousSMul ℝ F]
(p : E →ₗ[ℝ] F →ₗ[ℝ] ℝ) [(-p).IsContPerfPair] (K : ProperCone ℝ E) {y : F}
(hy : y ∈ ProperCone.dual (-p) (K : Set E)) :
fenchelConjugateBilin p (erealIndicator (K : Set E)) y = 0
The Fenchel conjugate of the EReal-indicator of a cone is ⊤ at points outside the dual cone.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma fenchelConjugate_erealIndicator_eq_top_of_not_mem_dual {F : Type*}
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E]
[TopologicalSpace F] [AddCommGroup F] [IsTopologicalAddGroup F] [Module ℝ F]
[ContinuousSMul ℝ F]
(p : E →ₗ[ℝ] F →ₗ[ℝ] ℝ) [(-p).IsContPerfPair] (K : ProperCone ℝ E) {y : F}
(hy : y ∉ ProperCone.dual (-p) (K : Set E)) :
fenchelConjugateBilin p (erealIndicator (K : Set E)) y = ⊤
The Fenchel conjugate of the EReal-indicator of a proper cone is the indicator of the dual cone.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma fenchelConjugate_erealIndicator_properCone {F : Type*}
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E]
[TopologicalSpace F] [AddCommGroup F] [IsTopologicalAddGroup F] [Module ℝ F]
[ContinuousSMul ℝ F]
(p : E →ₗ[ℝ] F →ₗ[ℝ] ℝ) [(-p).IsContPerfPair] (K : ProperCone ℝ E) :
fenchelConjugateBilin p (erealIndicator (K : Set E)) =
erealIndicator (ProperCone.dual (-p) (K : Set E) : Set F)
Theorem 14.1: If K is a non-empty closed convex cone, then its polar Kᵒ is also a non-empty closed convex cone and the bipolar identity Kᵒᵒ = K holds. Moreover, the indicator functions of K and Kᵒ are Fenchel–Legendre conjugates of each other.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
theorem polarCone_polar_polar_eq_and_indicator_conjugate {F : Type*}
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [LocallyConvexSpace ℝ E]
[TopologicalSpace F] [AddCommGroup F] [IsTopologicalAddGroup F]
[Module ℝ F] [ContinuousSMul ℝ F]
(p : E →ₗ[ℝ] F →ₗ[ℝ] ℝ) [(-p).IsContPerfPair] (K : ProperCone ℝ E) :
((ProperCone.dual (-p) (K : Set E) : Set F).Nonempty ∧
IsClosed (ProperCone.dual (-p) (K : Set E) : Set F) ∧
Convex ℝ (ProperCone.dual (-p) (K : Set E) : Set F)) ∧
ProperCone.dual (-p).flip (ProperCone.dual (-p) (K : Set E) : Set F) = K ∧
fenchelConjugateBilin p (erealIndicator (K : Set E)) =
erealIndicator (ProperCone.dual (-p) (K : Set E) : Set F) ∧
fenchelConjugateBilin p.flip
(erealIndicator (ProperCone.dual (-p) (K : Set E) : Set F)) =
erealIndicator (K : Set E)
Membership in polarCone is exactly the pointwise inequality φ x ≤ 0 on the set.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma mem_polarCone_iff (K : Set E) (φ : Module.Dual ℝ E) :
φ ∈ polarCone (E := E) K ↔ ∀ x, x ∈ K → φ x ≤ 0
For a submodule K, the polar cone coincides with the dual annihilator.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma mem_polarCone_submodule_iff_mem_dualAnnihilator (K : Submodule ℝ E) (φ : Module.Dual ℝ E) :
φ ∈ polarCone (E := E) (K : Set E) ↔ φ ∈ K.dualAnnihilator
Text 14.0.2: If K is a subspace of ℝ^n, then its polar equals its orthogonal complement: Kᵒ = Kᗮ = {x★ ∈ ℝ^n | ∀ x ∈ K, ⟪x, x★⟫ = 0}. In this file we interpret Kᵒ as polarCone (K : Set E) (a subset of the dual space Module.Dual ℝ E), and the orthogonal complement as the dual annihilator K.dualAnnihilator = {x★ | ∀ x ∈ K, x★ x = 0}.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
theorem polarCone_submodule_eq_dualAnnihilator (K : Submodule ℝ E) :
polarCone (K : Set E) = (K.dualAnnihilator : Set (Module.Dual ℝ E))
The convex cone of points where a linear functional is nonpositive.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
def nonposCone (φ : Module.Dual ℝ E) : ConvexCone ℝ E
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
@[simp] lemma mem_nonposCone_iff (φ : Module.Dual ℝ E) (x : E) :
x ∈ nonposCone (E := E) φ ↔ φ x ≤ 0
If a linear functional is nonpositive on a generating set, it is nonpositive on the cone hull.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma hull_le_nonposCone_of_forall_range {I : Type*} (a : I → E) (φ : Module.Dual ℝ E)
(h : ∀ i, φ (a i) ≤ 0) :
ConvexCone.hull ℝ (Set.range a) ≤ nonposCone (E := E) φ
If a linear functional is nonpositive on each generator a i, it is nonpositive on the hull.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma le_zero_on_hull_of_le_zero_on_generators {I : Type*} (a : I → E) (φ : Module.Dual ℝ E)
(h : ∀ i, φ (a i) ≤ 0) :
∀ x, x ∈ (ConvexCone.hull ℝ (Set.range a) : Set E) → φ x ≤ 0
Text 14.0.3: If K is the convex cone generated by a non-empty vector collection {a_i | i ∈ I}, then K consists of all non-negative linear combinations of the a_i. Consequently, the polar cone satisfies Kᵒ = {x★ | ∀ x ∈ K, x★ x ≤ 0} = {x★ | ∀ i ∈ I, x★ (a i) ≤ 0}.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
theorem polarCone_convexConeGenerated_range_eq {I : Type*} [Nonempty I] (a : I → E) :
polarCone (Set.insert (0 : E) ((ConvexCone.hull ℝ (Set.range a) : ConvexCone ℝ E) : Set E)) =
{φ : Module.Dual ℝ E | ∀ i : I, φ (a i) ≤ 0}
Membership in the polar of polarCone K, expressed as a pointwise inequality.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma mem_polarCone_polar_iff (K : Set E) (x : E) :
x ∈
PointedCone.dual (R := ℝ)
((-LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ))).flip (polarCone (E := E) K) ↔
∀ φ : Module.Dual ℝ E, (∀ y, y ∈ K → φ y ≤ 0) → φ x ≤ 0
If a functional is nonpositive on K, then it is nonpositive on closure K.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma closure_subset_polarCone_polar [TopologicalSpace E] [IsTopologicalAddGroup E]
[ContinuousSMul ℝ E] [T2Space E] [FiniteDimensional ℝ E] (K : Set E) :
closure K ⊆
PointedCone.dual (R := ℝ)
((-LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ))).flip (polarCone (E := E) K)
The polar of polarCone K is contained in closure K for a nonempty convex cone K.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma polarCone_polar_subset_closure_of_convexCone [TopologicalSpace E] [IsTopologicalAddGroup E]
[ContinuousSMul ℝ E] [LocallyConvexSpace ℝ E] (K : ConvexCone ℝ E)
(hKne : (K : Set E).Nonempty) :
(PointedCone.dual (R := ℝ)
((-LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ))).flip
(polarCone (E := E) (K : Set E)) : Set E) ⊆
closure (K : Set E)
Text 14.0.4: The polar of Kᵒ is the closure of K, i.e. (Kᵒ)ᵒ = cl K.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
theorem polarCone_polar_eq_closure [TopologicalSpace E] [IsTopologicalAddGroup E]
[ContinuousSMul ℝ E] [T2Space E] [FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
(K : ConvexCone ℝ E) (hKne : (K : Set E).Nonempty) :
PointedCone.dual (R := ℝ)
((-LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ))).flip
(polarCone (E := E) (K : Set E)) =
closure (K : Set E)
Text 14.0.5: The polar of a non-empty convex set C is C^∘ = {x★ | δ*(x★ | C) - 1 ≤ 0} = {x★ | ∀ x ∈ C, ⟪x, x★⟫ ≤ 1}. Here we model δ*(x★ | C) as the Fenchel–Legendre conjugate of the EReal-indicator of C with respect to the evaluation pairing.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
noncomputable def polar (C : Set E) : Set (Module.Dual ℝ E) :=
{φ : Module.Dual ℝ E |
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) (erealIndicator C) φ ≤
(1 : EReal)}
If fenchelConjugateBilin (eval) (erealIndicator C) φ ≤ 1, then φ is bounded above by 1 on C.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_le_one_on_C_of_fenchelConjugate_indicator_le_one {C : Set E} {φ : Module.Dual ℝ E}
(hφ :
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) (erealIndicator C) φ ≤
(1 : EReal)) :
∀ x ∈ C, φ x ≤ 1
If φ is bounded above by 1 on C, then fenchelConjugateBilin (eval) (erealIndicator C) φ ≤ 1.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_fenchelConjugate_indicator_le_one_of_le_one_on_C {C : Set E} {φ : Module.Dual ℝ E}
(hφ : ∀ x ∈ C, φ x ≤ 1) :
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) (erealIndicator C) φ ≤
(1 : EReal)
Text 14.0.5 (membership form): x★ ∈ C^∘ iff ⟪x, x★⟫ ≤ 1 for all x ∈ C.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
theorem mem_polar_iff {C : Set E} {φ : Module.Dual ℝ E} :
φ ∈ polar (E := E) C ↔ ∀ x ∈ C, φ x ≤ 1
If φ ∈ polar C, then φ is bounded above by 1 on C. This is a convenient lemma for Text 14.0.6 that avoids unfolding mem_polar_iff.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_le_one_of_mem_polar {C : Set E} {φ : Module.Dual ℝ E}
(hφ : φ ∈ polar (E := E) C) :
∀ x ∈ C, φ x ≤ 1
If φ ∈ polar C and C is absorbent, then fenchelConjugateBilin (eval) (gauge C) φ = 0.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_fenchelConjugate_gauge_eq_zero_of_mem_polar {C : Set E} (hCabs : Absorbent ℝ C)
{φ : Module.Dual ℝ E} (hφ : φ ∈ polar (E := E) C) :
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ))
(fun x => (gauge C x : EReal)) φ = 0
If φ ∉ polar C, then fenchelConjugateBilin (eval) (gauge C) φ = ⊤.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_fenchelConjugate_gauge_eq_top_of_not_mem_polar {C : Set E} {φ : Module.Dual ℝ E}
(hφ : φ ∉ polar (E := E) C) :
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ))
(fun x => (gauge C x : EReal)) φ = ⊤
The Fenchel conjugate of the (real-valued) gauge is the EReal indicator of polar C, assuming C is absorbent.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_fenchelConjugate_gauge_eq_erealIndicator_polar {C : Set E} (hCabs : Absorbent ℝ C) :
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ))
(fun x => (gauge C x : EReal)) =
erealIndicator (E := Module.Dual ℝ E) (polar (E := E) C)
Text 14.0.6: Let C be a non-empty convex set. Then the closure cl γ(·|C) of the gauge equals the support function δ*(·|C^∘) of the polar set. In this file we model γ(·|C) as mathlib's gauge C, and we model the closure operation cl as the Fenchel–Legendre biconjugate with respect to the evaluation pairing. We model δ*(·|C^∘) as the Fenchel–Legendre conjugate of the EReal-indicator of polar C with respect to the flipped evaluation pairing. Note: mathlib's gauge C is real-valued, so to match the intended convex-analytic gauge (which may take the value ⊤ when C does not absorb directions), we explicitly assume C is absorbent.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
theorem fenchelBiconjugate_gauge_eq_fenchelConjugate_indicator_polar {C : Set E}
(hCabs : Absorbent ℝ C) :
let p := LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)
fenchelConjugateBilin p.flip (fenchelConjugateBilin p (fun x => (gauge C x : EReal))) =
fenchelConjugateBilin p.flip (erealIndicator (E := Module.Dual ℝ E) (polar (E := E) C))
polarCone K is an intersection of pointwise half-spaces.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_polarCone_eq_iInter (K : Set E) :
polarCone (E := E) K =
⋂ x : E, ⋂ (_ : x ∈ K), {φ : Module.Dual ℝ E | φ x ≤ 0}
polarCone is closed in the weak topology on Module.Dual.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_isClosed_polarCone (K : Set E) :
IsClosed (polarCone (E := E) K)
polarCone is convex (as a subset of the dual space).
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_convex_polarCone (K : Set E) :
Convex ℝ (polarCone (E := E) K)
The zero functional always belongs to the polar cone.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_zero_mem_polarCone (K : Set E) :
(0 : Module.Dual ℝ E) ∈ polarCone (E := E) K
Text 14.0.7. Let K be a non-empty convex cone in ℝ^n. Then the polar cone Kᵒ defined in Text 14.0.1 is a closed convex cone containing the origin. In this file, Kᵒ is modeled as polarCone (E := E) (K : Set E), a subset of the dual space Module.Dual ℝ E.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
theorem isClosed_convex_polarCone_and_zero_mem (K : ConvexCone ℝ E) :
IsClosed (polarCone (E := E) (K : Set E)) ∧
Convex ℝ (polarCone (E := E) (K : Set E)) ∧
(0 : Module.Dual ℝ E) ∈ polarCone (E := E) (K : Set E)
polarCone is order-reversing: enlarging the set shrinks its polar cone.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_polarCone_antitone {A B : Set E} (hAB : A ⊆ B) :
polarCone (E := E) B ⊆ polarCone (E := E) A
If every linear functional is continuous (e.g. in finite dimension), then polar membership extends from a set to its topological closure.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_polarCone_subset_polarCone_closure
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E]
[T2Space E] [FiniteDimensional ℝ E] (K : Set E) :
polarCone (E := E) K ⊆ polarCone (E := E) (closure K)
Text 14.0.8. Let K be a non-empty convex cone in ℝ^n. Then the polar of cl K coincides with the polar of K, i.e. (cl K)^∘ = K^∘. In this file, we express this using the polar cone polarCone (Text 14.0.1): polarCone (closure K) = polarCone K.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
theorem polarCone_closure_eq [TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E]
[T2Space E] [FiniteDimensional ℝ E]
(K : ConvexCone ℝ E) :
polarCone (E := E) (closure (K : Set E)) = polarCone (E := E) (K : Set E)
The normal cone to a set C at the origin, modeled in the dual space: N(0 | C).
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
def normalConeAtOrigin (C : Set E) : Set (Module.Dual ℝ E) :=
{φ : Module.Dual ℝ E | ∀ x ∈ C, φ x ≤ 0}
The normal cone to a set C in the dual space at the origin, modeled in the primal space: N(0 | C).
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
def normalConeAtOriginDual (C : Set (Module.Dual ℝ E)) : Set E :=
{x : E | ∀ φ ∈ C, φ x ≤ 0}
The polar cone agrees with the normal cone to the set at the origin (in the dual space).
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_polarCone_eq_normalConeAtOrigin (K : Set E) :
polarCone (E := E) K = normalConeAtOrigin (E := E) K
The normal cone at the origin in the primal agrees with the dual cone defined via evaluation.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_normalConeAtOriginDual_eq_pointedCone_dual_flip (C : Set (Module.Dual ℝ E)) :
normalConeAtOriginDual (E := E) C =
PointedCone.dual (R := ℝ)
((-LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ))).flip C
Text 14.0.9. Let K be a non-empty closed convex cone in ℝ^n. Then the polar cone K^∘ consists of all vectors normal to K at the origin, and conversely K consists of all vectors normal to K^∘ at the origin. Equivalently, writing N(0 | C) for the normal cone of C at 0, one has K^∘ = N(0 | K) and K = N(0 | K^∘). In this file, K^∘ is modeled by polarCone (E := E) (K : Set E). The normal cone at 0 is modeled by normalConeAtOrigin (in the dual) and normalConeAtOriginDual (in the primal).
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
theorem polarCone_eq_normalConeAtOrigin_and_normalConeAtOriginDual_polarCone_eq
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E]
[T2Space E] [FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
(K : ConvexCone ℝ E) (hKne : (K : Set E).Nonempty) (hKclosed : IsClosed (K : Set E)) :
polarCone (E := E) (K : Set E) = normalConeAtOrigin (E := E) (K : Set E) ∧
(K : Set E) = normalConeAtOriginDual (E := E) (polarCone (E := E) (K : Set E))
Unpack membership in the dual cone w.r.t. the pairing -(innerₗ _).
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_mem_pointedConeDual_negInner_iff (n : ℕ)
(K : Set (EuclideanSpace ℝ (Fin n))) (y : EuclideanSpace ℝ (Fin n)) :
y ∈ (PointedCone.dual (R := ℝ) (-innerₗ (EuclideanSpace ℝ (Fin n))) K :
Set (EuclideanSpace ℝ (Fin n))) ↔
∀ x, x ∈ K → ⟪x, y⟫ ≤ 0
A vector in the dual cone of the nonnegative orthant has all coordinates nonpositive.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_coord_nonpos_of_mem_dual_nonnegOrthant (n : ℕ) (y : EuclideanSpace ℝ (Fin n))
(hy :
y ∈ (PointedCone.dual (R := ℝ) (-innerₗ (EuclideanSpace ℝ (Fin n)))
({x : EuclideanSpace ℝ (Fin n) | ∀ i, 0 ≤ x i} : Set (EuclideanSpace ℝ (Fin n))) :
Set (EuclideanSpace ℝ (Fin n)))) :
∀ i, y i ≤ 0
Membership in the negation of the nonnegative orthant is coordinatewise nonpositivity.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_mem_neg_nonnegOrthant_iff_coord_nonpos (n : ℕ) (y : EuclideanSpace ℝ (Fin n)) :
y ∈ -({x : EuclideanSpace ℝ (Fin n) | ∀ i, 0 ≤ x i} : Set (EuclideanSpace ℝ (Fin n))) ↔
∀ i, y i ≤ 0
If x has nonnegative coordinates and y has nonpositive coordinates, then ⟪x, y⟫ ≤ 0.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_inner_nonpos_of_coords_nonneg_nonpos (n : ℕ) (x y : EuclideanSpace ℝ (Fin n))
(hx : ∀ i, 0 ≤ x i) (hy : ∀ i, y i ≤ 0) :
⟪x, y⟫ ≤ 0
Text 14.0.10. Let K ⊆ ℝ^n be the non-negative orthant K = {x = (ξ₁, …, ξₙ) | ξⱼ ≥ 0 for j = 1, …, n}. Then the polar cone of K is the non-positive orthant, i.e. K^∘ = -K. In Lean, we model ℝ^n as EuclideanSpace ℝ (Fin n) and the polar cone as the dual cone with respect to the pairing -(innerₗ _), so that y is in the polar cone of K iff ⟪x, y⟫ ≤ 0 for all x ∈ K.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
theorem polarCone_nonnegOrthant_eq_neg (n : ℕ) :
(PointedCone.dual (R := ℝ) (-innerₗ (EuclideanSpace ℝ (Fin n)))
({x : EuclideanSpace ℝ (Fin n) | ∀ i, 0 ≤ x i} : Set (EuclideanSpace ℝ (Fin n))) :
Set (EuclideanSpace ℝ (Fin n))) =
-({x : EuclideanSpace ℝ (Fin n) | ∀ i, 0 ≤ x i} : Set (EuclideanSpace ℝ (Fin n)))
Text 14.0.11. Polarity is order-inverting: if C₁ ⊆ C₂ are closed convex sets containing the origin, then their polars satisfy C₁^∘ ⊇ C₂^∘. In this file, C^∘ is modeled as polar (E := E) C.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
theorem polar_antitone_of_subset [TopologicalSpace E] {C₁ C₂ : Set E}
(hC : C₁ ⊆ C₂) :
polar (E := E) C₂ ⊆ polar (E := E) C₁
A linear functional on ℝ^n is determined by its values on the coordinate vectors.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_dual_apply_eq_sum_mul_single {n : ℕ}
(φ : Module.Dual ℝ (EuclideanSpace ℝ (Fin n))) (x : EuclideanSpace ℝ (Fin n)) :
φ x =
Finset.univ.sum (fun i : Fin n => x i * φ (EuclideanSpace.single i (1 : ℝ)))
Each coordinate vector e_j belongs to the subprobability simplex {x | x ≥ 0, ∑ x ≤ 1}.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_single_one_mem_subprobabilitySimplex (n : ℕ) (j : Fin n) :
(EuclideanSpace.single j (1 : ℝ) : EuclideanSpace ℝ (Fin n)) ∈
{x : EuclideanSpace ℝ (Fin n) |
(∀ j, 0 ≤ x j) ∧ (Finset.univ.sum (fun j => x j)) ≤ (1 : ℝ)}
If φ is bounded by 1 on all coordinate vectors, then it is bounded by 1 on C₁.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_le_one_of_forall_single_le_one_and_mem_C₁ {n : ℕ}
(φ : Module.Dual ℝ (EuclideanSpace ℝ (Fin n))) (x : EuclideanSpace ℝ (Fin n))
(hφ : ∀ j : Fin n, φ (EuclideanSpace.single j (1 : ℝ)) ≤ 1)
(hxnonneg : ∀ j : Fin n, 0 ≤ x j)
(hxsum : (Finset.univ.sum fun j : Fin n => x j) ≤ (1 : ℝ)) :
φ x ≤ 1
Text 14.0.12 (Example). Define C₁ = {x = (ξ₁, …, ξₙ) | ξⱼ ≥ 0, ξ₁ + ⋯ + ξₙ ≤ 1}. Then its polar is C₁^∘ = {x★ = (ξ₁★, …, ξₙ★) | ξⱼ★ ≤ 1 for j = 1, …, n}. In this file, C^∘ is modeled as polar (E := E) C : Set (Module.Dual ℝ E). For ℝ^n we use E = EuclideanSpace ℝ (Fin n), and we express the coordinate inequalities as φ (Pi.single j 1) ≤ 1.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
theorem polar_subprobabilitySimplex_eq (n : ℕ) :
let C₁ : Set (EuclideanSpace ℝ (Fin n)) :=
{x | (∀ j, 0 ≤ x j) ∧ (Finset.univ.sum (fun j => x j)) ≤ (1 : ℝ)}
polar (E := EuclideanSpace ℝ (Fin n)) C₁ =
{φ : Module.Dual ℝ (EuclideanSpace ℝ (Fin n)) |
∀ j, φ (EuclideanSpace.single j (1 : ℝ)) ≤ 1}§14 — Polars of Convex Sets — Part 2
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yuhao Jiang, Zaiwen Wen.
The effective domain dom f of an EReal-valued function, as the set where f is finite (i.e. strictly below ⊤).
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
def erealDom {F : Type*} (f : F → EReal) : Set F :=
{x | f x < ⊤}
A basic notion of properness for an EReal-valued function: it never takes the value ⊥ and is finite at some point.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
def ProperERealFunction {F : Type*} (f : F → EReal) : Prop :=
(∀ x, f x ≠ ⊥) ∧ ∃ x, f x ≠ ⊤
A Jensen-style convexity condition for an EReal-valued function on a real vector space.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
def ConvexERealFunction {F : Type*} [AddCommGroup F] [Module ℝ F] (f : F → EReal) : Prop :=
∀ ⦃x y : F⦄ ⦃a b : ℝ⦄,
0 ≤ a → 0 ≤ b → a + b = 1 →
f (a • x + b • y) ≤ (a : EReal) * f x + (b : EReal) * f y
A "proper convex function" (proper + Jensen convex) for EReal-valued functions.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
def ProperConvexERealFunction {F : Type*} [AddCommGroup F] [Module ℝ F] (f : F → EReal) : Prop :=
ProperERealFunction f ∧ ConvexERealFunction f
The recession function f₀⁺(y) = sup { f(x+y) - f(x) | x ∈ dom f } for an EReal-valued function.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
noncomputable def recessionFunctionEReal {F : Type*} [AddCommGroup F] (f : F → EReal) :
F → EReal :=
fun y => sSup {r : EReal | ∃ x ∈ erealDom f, r = f (x + y) - f x}
The recession cone {y | f₀⁺(y) ≤ 0} associated to recessionFunctionEReal.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
def recessionConeEReal {F : Type*} [AddCommGroup F] (f : F → EReal) : Set F :=
{y | recessionFunctionEReal f y ≤ (0 : EReal)}
Membership in recessionConeEReal is equivalent to a pointwise nonpositivity condition on increment differences over the effective domain.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_mem_recessionConeEReal_iff {F : Type*} [AddCommGroup F] (g : F → EReal) (y : F) :
y ∈ recessionConeEReal (F := F) g ↔
∀ x, x ∈ erealDom g → g (x + y) - g x ≤ (0 : EReal)
For f : E → EReal, membership in the polar cone of the cone hull of erealDom f is equivalent to being nonpositive on erealDom f.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_mem_polarCone_hull_erealDom_iff {f : E → EReal} (φ : Module.Dual ℝ E) :
φ ∈ polarCone (E := E) ((ConvexCone.hull ℝ (erealDom f) : ConvexCone ℝ E) : Set E) ↔
∀ x, x ∈ erealDom f → φ x ≤ 0
If y★ is nonpositive on erealDom f, then the Fenchel conjugate of f is nonincreasing under translation by y★.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_fenchelConjugate_add_le_of_le_zero_on_dom {f : E → EReal}
(yStar : Module.Dual ℝ E) (hy : ∀ x, x ∈ erealDom f → yStar x ≤ 0) :
∀ xStar : Module.Dual ℝ E,
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f (xStar + yStar) ≤
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f xStar
A finite, non-⊥ EReal value is a real coercion.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_eq_coe_of_lt_top {z : EReal} (hzTop : z < ⊤) (hzBot : z ≠ ⊥) :
∃ r : ℝ, z = (r : EReal)
The Fenchel conjugate of a proper function never takes the value ⊥.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_fenchelConjugate_ne_bot {f : E → EReal} (hf : ProperERealFunction f) :
∀ xStar : Module.Dual ℝ E,
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f xStar ≠ ⊥
If y ∈ recessionConeEReal g and x ∈ erealDom g, then translating by y does not increase g, and the translate remains in the effective domain.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_step_le_of_mem_recessionCone {F : Type*} [AddCommGroup F] (g : F → EReal) {x y : F}
(hy : y ∈ recessionConeEReal (F := F) g) (hx : x ∈ erealDom g) :
g (x + y) ≤ g x ∧ x + y ∈ erealDom g
If n * r is uniformly bounded above for all natural n, then r ≤ 0.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_real_nonpos_of_nat_mul_le (r C : ℝ) (h : ∀ n : ℕ, (n : ℝ) * r ≤ C) : r ≤ 0
Text 14.0.13 (Example). Define C₂ = {x = (ξ₁, …, ξₙ) | ξⱼ ≥ 0, ξ₁ + ⋯ + ξₙ ≤ 1}. Then its polar is C₂^∘ = {x★ = (ξ₁★, …, ξₙ★) | ξⱼ★ ≤ 1 for j = 1, …, n}. In this file, C^∘ is modeled as `polar (E := EuclideanSpace ℝ (Fin n)) C : Set (Module.Dual ℝ _)`, and we express the coordinate inequalities as φ (EuclideanSpace.single j 1) ≤ 1.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
theorem polar_sum_le_one_eq (n : ℕ) :
let C₂ : Set (EuclideanSpace ℝ (Fin n)) :=
{x | (∀ j, 0 ≤ x j) ∧ (Finset.univ.sum (fun j => x j)) ≤ (1 : ℝ)}
polar (E := EuclideanSpace ℝ (Fin n)) C₂ =
{φ : Module.Dual ℝ (EuclideanSpace ℝ (Fin n)) |
∀ j, φ (EuclideanSpace.single j (1 : ℝ)) ≤ 1}
The coefficient vector of a linear functional on ℝ^2 with respect to the standard basis.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
noncomputable def section14_coeffVec (φ : Module.Dual ℝ (EuclideanSpace ℝ (Fin 2))) :
EuclideanSpace ℝ (Fin 2) :=
Finset.univ.sum fun i : Fin 2 =>
(φ (EuclideanSpace.single i (1 : ℝ))) • EuclideanSpace.single i (1 : ℝ)
The center (1,0) of the shifted unit disk C₃.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
noncomputable def section14_C3_center : EuclideanSpace ℝ (Fin 2) :=
EuclideanSpace.single 0 (1 : ℝ)
Expanding the squared distance from the center (1,0) gives the defining quadratic inequality for C₃.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_norm_sq_sub_C3_center (x : EuclideanSpace ℝ (Fin 2)) :
‖x - section14_C3_center‖ ^ 2 = (x 0 - 1) ^ 2 + (x 1) ^ 2
Membership in C₃ is equivalent to being within distance 1 from the center (1,0).
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_mem_C3_iff_norm_sub_le (x : EuclideanSpace ℝ (Fin 2)) :
((x 0 - 1) ^ 2 + (x 1) ^ 2 ≤ (1 : ℝ)) ↔ ‖x - section14_C3_center‖ ≤ 1
A linear functional on ℝ^2 acts as an inner product with its coefficient vector.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_dual_apply_eq_inner_coeffVec (φ : Module.Dual ℝ (EuclideanSpace ℝ (Fin 2)))
(x : EuclideanSpace ℝ (Fin 2)) :
φ x = inner ℝ x (section14_coeffVec φ)
If φ satisfies the support-function bound for the shifted unit disk, then φ belongs to its polar.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_mem_polar_C3_of_le_support_bound
(φ : Module.Dual ℝ (EuclideanSpace ℝ (Fin 2)))
(hφ : φ (EuclideanSpace.single 0 (1 : ℝ)) + ‖section14_coeffVec φ‖ ≤ 1) :
φ ∈ polar (E := EuclideanSpace ℝ (Fin 2))
({x : EuclideanSpace ℝ (Fin 2) | (x 0 - 1) ^ 2 + (x 1) ^ 2 ≤ (1 : ℝ)} : Set _)
If φ belongs to the polar of the shifted unit disk, then it satisfies the support-function bound φ(1,0) + ‖vφ‖ ≤ 1.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_support_bound_of_mem_polar_C3
(φ : Module.Dual ℝ (EuclideanSpace ℝ (Fin 2)))
(hφ : φ ∈ polar (E := EuclideanSpace ℝ (Fin 2))
({x : EuclideanSpace ℝ (Fin 2) | (x 0 - 1) ^ 2 + (x 1) ^ 2 ≤ (1 : ℝ)} : Set _)) :
φ (EuclideanSpace.single 0 (1 : ℝ)) + ‖section14_coeffVec φ‖ ≤ 1
Text 14.0.14 (Example). Define C₃ = {x = (ξ₁, ξ₂) | (ξ₁ - 1)^2 + ξ₂^2 ≤ 1}. Then its polar can be described as C₃^∘ = {x★ = (ξ₁★, ξ₂★) | ξ₁★ + ‖(ξ₁★, ξ₂★)‖ ≤ 1}. In this file, C^∘ is modeled as `polar (E := EuclideanSpace ℝ (Fin 2)) C : Set (Module.Dual ℝ _), and we interpret the coordinates ξ₁★, ξ₂★of a functionalφ` as the values φ (EuclideanSpace.single 0 1) and φ (EuclideanSpace.single 1 1).
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
theorem polar_shiftedUnitDisk_eq :
let C₃ : Set (EuclideanSpace ℝ (Fin 2)) :=
{x | (x 0 - 1) ^ 2 + (x 1) ^ 2 ≤ (1 : ℝ)}
polar (E := EuclideanSpace ℝ (Fin 2)) C₃ =
{φ : Module.Dual ℝ (EuclideanSpace ℝ (Fin 2)) |
φ (EuclideanSpace.single 0 (1 : ℝ)) + ‖section14_coeffVec φ‖ ≤ 1}
A concrete counterexample showing that, for the current definition of C₄ and P, the set P is not a subset of the polar C₄^∘. This indicates that polar_C4_eq_convexHull cannot be proven as stated.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_counterexample_P_not_subset_polar_C4 :
let C₄ : Set (EuclideanSpace ℝ (Fin 2)) :=
{x | x 0 ≤ 1 - Real.sqrt (1 + (x 1) ^ 2)}
let P : Set (Module.Dual ℝ (EuclideanSpace ℝ (Fin 2))) :=
{φ |
Real.sqrt (1 + (φ (EuclideanSpace.single 1 (1 : ℝ))) ^ 2) ≤
φ (EuclideanSpace.single 0 (1 : ℝ))}
¬ P ⊆ polar (E := EuclideanSpace ℝ (Fin 2)) C₄
The standard basis vector e₀ in ℝ².
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
noncomputable abbrev section14_e0 : EuclideanSpace ℝ (Fin 2) :=
EuclideanSpace.single 0 (1 : ℝ)
The standard basis vector e₁ in ℝ².
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
noncomputable abbrev section14_e1 : EuclideanSpace ℝ (Fin 2) :=
EuclideanSpace.single 1 (1 : ℝ)
The set C₄ = {(ξ₁, ξ₂) | ξ₁ ≤ 1 - √(1+ξ₂^2)} from Text 14.0.15.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
noncomputable def section14_C4 : Set (EuclideanSpace ℝ (Fin 2)) :=
{x | x 0 ≤ 1 - Real.sqrt (1 + (x 1) ^ 2)}
The parabola region P₄ = {(a,b) | b^2 + 1 ≤ 2a} in the dual space.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
noncomputable def section14_P4 : Set (Module.Dual ℝ (EuclideanSpace ℝ (Fin 2))) :=
{φ | (φ section14_e1) ^ 2 + 1 ≤ 2 * φ section14_e0}
Coordinate expansion of a linear functional on ℝ².
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_dual_apply_eq_two_coords
(φ : Module.Dual ℝ (EuclideanSpace ℝ (Fin 2))) (x : EuclideanSpace ℝ (Fin 2)) :
φ x = x 0 * φ section14_e0 + x 1 * φ section14_e1
The basic inequality |t| ≤ √(1+t^2).
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_abs_le_sqrt_one_add_sq (t : ℝ) : |t| ≤ Real.sqrt (1 + t ^ 2)
A squaring identity gives the key inequality used in the C₄ polar computation.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_sqrt_mul_sub_ge_sqrt_sub {a b t : ℝ} (hb : |b| ≤ a) :
Real.sqrt (a ^ 2 - b ^ 2) ≤ a * Real.sqrt (1 + t ^ 2) - b * t
The polar of any set is convex (as a subset of the dual).
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_convex_polar (C : Set E) : Convex ℝ (polar (E := E) C)
The parabola region P₄ is contained in the polar of C₄.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_P4_subset_polar_C4 :
section14_P4 ⊆ polar (E := EuclideanSpace ℝ (Fin 2)) section14_C4
The easy inclusion conv(P₄ ∪ {0}) ⊆ C₄^∘.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_convexHull_P4_union_zero_subset_polar_C4 :
convexHull ℝ (section14_P4 ∪ {0}) ⊆
polar (E := EuclideanSpace ℝ (Fin 2)) section14_C4
A linear functional on ℝ² is determined by its values on the standard basis.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_dual_eq_zero_of_apply_e0_e1_eq_zero
(φ : Module.Dual ℝ (EuclideanSpace ℝ (Fin 2)))
(h0 : φ section14_e0 = 0) (h1 : φ section14_e1 = 0) :
φ = 0
A convenient estimate: for n ≥ 1, √(1+n^2) - n ≤ 1/n.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_sqrt_one_add_sq_sub_nat_le_inv_nat (n : ℕ) (hn : 1 ≤ n) :
Real.sqrt (1 + (n : ℝ) ^ 2) - (n : ℝ) ≤ 1 / (n : ℝ)
For 0 ≤ a and b^2 < a^2, the choice t = b / √(a^2-b^2) attains the minimum in the expression a * √(1+t^2) - b * t.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_eval_opt_t {a b : ℝ} (ha : 0 ≤ a) (hba : b ^ 2 < a ^ 2) :
let t : ℝ := b / Real.sqrt (a ^ 2 - b ^ 2)
a * Real.sqrt (1 + t ^ 2) - b * t = Real.sqrt (a ^ 2 - b ^ 2)
Membership in the polar of C₄ forces the coordinate constraints needed for the description via P₄.
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_mem_polar_C4_coord_consequences
{φ : Module.Dual ℝ (EuclideanSpace ℝ (Fin 2))}
(hφ : φ ∈ polar (E := EuclideanSpace ℝ (Fin 2)) section14_C4) :
0 ≤ φ section14_e0 ∧
The hard inclusion C₄^∘ ⊆ conv(P₄ ∪ {0}) (to be proven).
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_polar_C4_subset_convexHull_P4_union_zero :
polar (E := EuclideanSpace ℝ (Fin 2)) section14_C4 ⊆ convexHull ℝ (section14_P4 ∪ {0})
Text 14.0.15 (Example). Define C₄ = {x = (ξ₁, ξ₂) | ξ₁ ≤ 1 - (1 + ξ₂^2)^{1/2}}. Then its polar can be written as C₄^∘ = conv(P ∪ {0}), where P = {x★ = (ξ₁★, ξ₂★) | ξ₁★ ≥ (1 + (ξ₂★)^2)^{1/2}}. Note: the set P above (a "sqrt cone") is not contained in the polar for the current definition of polar (via Fenchel conjugates), as shown by section14_counterexample_P_not_subset_polar_C4. The correct description of C₄^∘ in this formalization uses instead the parabola region P₄ = {x★ = (a,b) | b^2 + 1 ≤ 2a}, which is the closed convex hull of all supporting halfspaces through the origin. In this file, C^∘ is modeled as `polar (E := EuclideanSpace ℝ (Fin 2)) C : Set (Module.Dual ℝ _), and we interpret the coordinates ξ₁★, ξ₂★of a functionalφ` as the values φ (EuclideanSpace.single 0 1) and φ (EuclideanSpace.single 1 1).
open scoped Pointwise
open scoped Topology
open scoped RealInnerProductSpace
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
theorem polar_C4_eq_convexHull :
let C₄ : Set (EuclideanSpace ℝ (Fin 2)) :=
{x | x 0 ≤ 1 - Real.sqrt (1 + (x 1) ^ 2)}
let P₄ : Set (Module.Dual ℝ (EuclideanSpace ℝ (Fin 2))) :=
{φ |
(φ (EuclideanSpace.single 1 (1 : ℝ))) ^ 2 + 1 ≤
2 * φ (EuclideanSpace.single 0 (1 : ℝ))}
polar (E := EuclideanSpace ℝ (Fin 2)) C₄ = convexHull ℝ (P₄ ∪ {0})§14 — Polars of Convex Sets — Part 3
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yuhao Jiang, Zaiwen Wen.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see also `section14_part1`).
noncomputable local instance section14_instTopologicalSpace_dualWeak_part3 :
TopologicalSpace (Module.Dual ℝ E)
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see also `section14_part1`).
noncomputable local instance section14_instIsTopologicalAddGroup_dualWeak_part3 :
IsTopologicalAddGroup (Module.Dual ℝ E)
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see also `section14_part1`).
noncomputable local instance section14_instContinuousSMul_dualWeak_part3 :
ContinuousSMul ℝ (Module.Dual ℝ E)
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see also `section14_part1`).
noncomputable local instance section14_instLocallyConvexSpace_dualWeak_part3 :
LocallyConvexSpace ℝ (Module.Dual ℝ E)
If y★ is in the recession cone of f* and f* is finite somewhere, then y★ is nonpositive on erealDom f.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see also `section14_part1`).
lemma section14_le_zero_on_dom_of_mem_recessionCone_fenchelConjugate {f : E → EReal}
(hf : ProperERealFunction f) {yStar : Module.Dual ℝ E}
(hy :
yStar ∈
recessionConeEReal
(F := Module.Dual ℝ E)
(fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f))
(hxStar0 :
∃ xStar0 : Module.Dual ℝ E,
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f xStar0 < ⊤) :
∀ x, x ∈ erealDom f → yStar x ≤ 0
Theorem 14.2. Let f be a proper convex function. The polar of the convex cone generated by dom f is the recession cone of the Fenchel conjugate f*. Dually, if f is closed, the polar of the recession cone of f is the closure of the convex cone generated by dom f*.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see also `section14_part1`).
theorem polar_coneGenerated_dom_eq_recessionCone_fenchelConjugate {f : E → EReal}
(hf : ProperConvexERealFunction (F := E) f)
(hdom :
∃ xStar0 : Module.Dual ℝ E,
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f xStar0 < ⊤) :
polarCone (E := E) ((ConvexCone.hull ℝ (erealDom f) : ConvexCone ℝ E) : Set E) =
recessionConeEReal
(F := Module.Dual ℝ E)
(fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f)
The polar cone of a set, packaged as a ConvexCone, so that we can use ConvexCone.hull minimality arguments.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see also `section14_part1`).
def section14_polarConeConvexCone (K : Set E) : ConvexCone ℝ (Module.Dual ℝ E)
If x★ is in the effective domain of f*, then x★ lies in the polar cone of the recession cone of f.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see also `section14_part1`).
lemma section14_mem_polar_recessionCone_of_mem_dom_fenchelConjugate {f : E → EReal}
(hf : ProperERealFunction f) {xStar : Module.Dual ℝ E}
(hxStar :
xStar ∈
erealDom (fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f)) :
xStar ∈ polarCone (E := E) (recessionConeEReal (F := E) f)
The closed conic hull of dom f* is contained in the polar of the recession cone of f.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see also `section14_part1`).
lemma section14_closure_coneHull_dom_fenchelConjugate_subset_polar_recessionCone [TopologicalSpace E]
{f : E → EReal} (hf : ProperConvexERealFunction (F := E) f) :
closure
((ConvexCone.hull ℝ
(erealDom
(fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f)) :
ConvexCone ℝ (Module.Dual ℝ E)) : Set (Module.Dual ℝ E)) ⊆
polarCone (E := E) (recessionConeEReal (F := E) f)
Any continuous linear functional on Module.Dual ℝ E equipped with the weak topology induced by evaluation is itself an evaluation at some y : E.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see also `section14_part1`).
lemma section14_exists_eval_of_continuousLinearMap_dualWeak
(ℓ : (Module.Dual ℝ E) →L[ℝ] ℝ) :
∃ y : E, ∀ xStar : Module.Dual ℝ E, ℓ xStar = xStar y
If x★ is not in the (weak) closure of the conic hull of dom f*, then one can separate it from that cone by evaluation at some y : E.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see also `section14_part1`).
lemma section14_exists_eval_sep_of_not_mem_closure_coneHull_dom_fenchelConjugate {f : E → EReal}
{xStar : Module.Dual ℝ E}
(hDom :
(erealDom
(fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f)).Nonempty)
(hxStar :
xStar ∉
closure
((ConvexCone.hull ℝ
(erealDom
(fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f)) :
ConvexCone ℝ (Module.Dual ℝ E)) : Set (Module.Dual ℝ E))) :
∃ y : E,
(∀ z : Module.Dual ℝ E,
z ∈
closure
((ConvexCone.hull ℝ
(erealDom
(fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f)) :
ConvexCone ℝ (Module.Dual ℝ E)) : Set (Module.Dual ℝ E)) →
z y ≤ 0) ∧
0 < xStar y
A Hahn–Banach separation gadget for a proper convex lower semicontinuous function: given a point (x0, μ0) strictly below the epigraph of f, produce a continuous affine minorant x ↦ x★ x - β that lies below f everywhere and is strictly above μ0 at x0. In particular, the Fenchel conjugate is finite at x★.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see also `section14_part1`).
lemma section14_exists_affine_minorant_strict_of_lt [TopologicalSpace E] [IsTopologicalAddGroup E]
[ContinuousSMul ℝ E] [LocallyConvexSpace ℝ E] {f : E → EReal}
(hf : ProperConvexERealFunction (F := E) f) (hf_closed : LowerSemicontinuous f) {x0 : E}
{μ0 : ℝ} (hμ0 : (μ0 : EReal) < f x0) :
∃ (xStar : Module.Dual ℝ E) (β : ℝ),
(∀ x : E, ((xStar x - β : ℝ) : EReal) ≤ f x) ∧ (μ0 : ℝ) < xStar x0 - β ∧
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f xStar < ⊤
For a proper convex lower semicontinuous function, the Fenchel conjugate is finite somewhere.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see also `section14_part1`).
lemma section14_exists_mem_dom_fenchelConjugate_of_properConvex_lsc [TopologicalSpace E]
[IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [LocallyConvexSpace ℝ E]
{f : E → EReal} (hf : ProperConvexERealFunction (F := E) f) (hf_closed : LowerSemicontinuous f) :
(erealDom (fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f)).Nonempty
If y is nonpositive on dom f* and f is proper convex and closed, then y is a recession direction of f.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see also `section14_part1`).
lemma section14_mem_recessionConeEReal_of_forall_mem_dom_fenchelConjugate_le_zero
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [LocallyConvexSpace ℝ E]
{f : E → EReal} (hf : ProperConvexERealFunction (F := E) f)
(hf_closed : LowerSemicontinuous f) {y : E}
(hy : ∀ xStar : Module.Dual ℝ E,
xStar ∈ erealDom (fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f) →
xStar y ≤ 0) :
y ∈ recessionConeEReal (F := E) f
If x★ is in polarCone (recessionConeEReal f), then x★ lies in the weak closure of the conic hull of dom f* (dual part of Theorem 14.2).
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see also `section14_part1`).
lemma section14_polar_recessionCone_subset_closure_coneHull_dom_fenchelConjugate
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [LocallyConvexSpace ℝ E]
{f : E → EReal} (hf : ProperConvexERealFunction (F := E) f) (hf_closed : LowerSemicontinuous f) :
polarCone (E := E) (recessionConeEReal (F := E) f) ⊆
closure
((ConvexCone.hull ℝ
(erealDom
(fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f)) :
ConvexCone ℝ (Module.Dual ℝ E)) : Set (Module.Dual ℝ E))
Theorem 14.2 (dual statement). If f is closed, then the polar of the recession cone of f is the closure of the convex cone generated by dom f*.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see also `section14_part1`).
theorem polar_recessionCone_eq_closure_coneGenerated_dom_fenchelConjugate [TopologicalSpace E]
[IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [LocallyConvexSpace ℝ E]
{f : E → EReal} (hf : ProperConvexERealFunction (F := E) f) (hf_closed : LowerSemicontinuous f) :
polarCone (E := E) (recessionConeEReal (F := E) f) =
closure
((ConvexCone.hull ℝ
(erealDom
(fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f)) :
ConvexCone ℝ (Module.Dual ℝ E)) : Set (Module.Dual ℝ E))
Membership in a dual cone with respect to the flipped evaluation pairing is exactly a pointwise nonpositivity condition.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see also `section14_part1`).
lemma section14_mem_dual_flip_eval_iff_forall_le_zero (S : Set (Module.Dual ℝ E)) (x : E) :
x ∈
(PointedCone.dual (R := ℝ)
((-LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ))).flip S :
Set E) ↔
∀ φ : Module.Dual ℝ E, φ ∈ S → φ x ≤ 0
Recession directions lie in the polar (with respect to flipped evaluation) of the barrier cone.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see also `section14_part1`).
lemma section14_recessionCone_subset_polar_barrierCone {C : Set E} (hCne : C.Nonempty) :
Set.recessionCone C ⊆
(PointedCone.dual (R := ℝ)
((-LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ))).flip (Set.barrierCone (E := E) C) :
Set E)§14 — Polars of Convex Sets — Part 4
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yuhao Jiang, Zaiwen Wen.
Points in the polar (w.r.t. flipped evaluation) of the barrier cone of a closed convex set are recession directions.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_polar_barrierCone_subset_recessionCone [TopologicalSpace E] [IsTopologicalAddGroup E]
[ContinuousSMul ℝ E] [LocallyConvexSpace ℝ E] {C : Set E} (hCclosed : IsClosed C)
(hCconv : Convex ℝ C) :
(PointedCone.dual (R := ℝ)
((-LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ))).flip (Set.barrierCone (E := E) C) :
Set E) ⊆
Set.recessionCone C
Corollary 14.2.1. The polar of the barrier cone of a non-empty closed convex set C is the recession cone of C.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
theorem polar_barrierCone_eq_recessionCone_part4 [TopologicalSpace E] [IsTopologicalAddGroup E]
[ContinuousSMul ℝ E] [LocallyConvexSpace ℝ E] {C : Set E} (hCne : C.Nonempty)
(hCclosed : IsClosed C) (hCconv : Convex ℝ C) :
(PointedCone.dual (R := ℝ)
((-LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ))).flip (Set.barrierCone (E := E) C) :
Set E) =
Set.recessionCone C
The zero vector always lies in the recession cone of an EReal function.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_recessionConeEReal_zero_mem {F : Type*} [AddCommGroup F] (f : F → EReal) :
(0 : F) ∈ recessionConeEReal (F := F) f
If a set is contained in {0}, then its polar cone is all of the dual space.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_polarCone_eq_univ_of_subset_singleton_zero (S : Set E) (hS : S ⊆ ({0} : Set E)) :
polarCone (E := E) S = (Set.univ : Set (Module.Dual ℝ E))
A polar cone is all of the dual space iff the original set is contained in {0}.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_polarCone_eq_univ_iff_subset_singleton_zero (S : Set E) :
polarCone (E := E) S = (Set.univ : Set (Module.Dual ℝ E)) ↔ S ⊆ ({0} : Set E)
A set equals {0} iff its polar cone is univ, provided it contains 0.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_eq_singleton_zero_iff_polarCone_eq_univ {S : Set E} (h0 : (0 : E) ∈ S) :
S = ({0} : Set E) ↔ polarCone (E := E) S = (Set.univ : Set (Module.Dual ℝ E))
If C is closed under translation by y, then all natural translates stay in C.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_add_mem_of_add_mem_nat_smul {C : Set E} {y : E} (hadd : ∀ x ∈ C, x + y ∈ C) :
∀ {x} (_hx : x ∈ C) (m : ℕ), x + (m : ℝ) • y ∈ C
If C is convex and C + y ⊆ C, then y lies in the recession cone.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_mem_recessionCone_of_add_mem {C : Set E} (hconv : Convex ℝ C) {y : E}
(hadd : ∀ x ∈ C, x + y ∈ C) : y ∈ Set.recessionCone C
Sublevel sets of a Jensen-convex EReal function are convex.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_convex_sublevel {f : E → EReal} (hfconv : ConvexERealFunction (F := E) f)
(α : ℝ) : Convex ℝ {x : E | f x ≤ (α : EReal)}
A recession direction of f is a recession direction of every real sublevel set of f.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_recessionConeEReal_subset_recessionCone_sublevel {f : E → EReal}
(hfconv : ConvexERealFunction (F := E) f) {α : ℝ} :
recessionConeEReal (F := E) f ⊆ Set.recessionCone {x : E | f x ≤ (α : EReal)}
If f is proper, convex, and lower semicontinuous, then any recession direction of a nonempty real sublevel set is a recession direction of f.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_recessionCone_sublevel_subset_recessionConeEReal {E : Type*} [NormedAddCommGroup E]
[NormedSpace ℝ E] {f : E → EReal} (hf : ProperConvexERealFunction (F := E) f)
(hf_closed : LowerSemicontinuous f) {α : ℝ}
(hne : ({x : E | f x ≤ (α : EReal)}).Nonempty) :
Set.recessionCone {x : E | f x ≤ (α : EReal)} ⊆ recessionConeEReal (F := E) f
Corollary 14.2.2. Let f be a closed proper convex function. In order that the sublevel set {x | f x ≤ α} be bounded for every α ∈ ℝ, it is necessary and sufficient that 0 ∈ int (dom f*). Here f* is the Fenchel–Legendre conjugate and dom denotes the effective domain (the set where the value is finite). In this formalization we record a topology-free equivalent criterion: the polar cone of the recession cone of f is all of the dual space.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
theorem sublevelSets_bounded_iff_zero_mem_interior_dom_fenchelConjugate {n : ℕ}
{f : EuclideanSpace ℝ (Fin n) → EReal}
(hf : ProperConvexERealFunction (F := EuclideanSpace ℝ (Fin n)) f)
(hf_closed : LowerSemicontinuous f) :
(∀ α : ℝ, Bornology.IsBounded {x : EuclideanSpace ℝ (Fin n) | f x ≤ (α : EReal)}) ↔
polarCone (E := EuclideanSpace ℝ (Fin n))
(recessionConeEReal (F := EuclideanSpace ℝ (Fin n)) f) =
(Set.univ : Set (Module.Dual ℝ (EuclideanSpace ℝ (Fin n))))§14 — Polars of Convex Sets — Part 5
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yuhao Jiang, Zaiwen Wen.
If inf f < 0, then the 0-sublevel set of f is nonempty.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_sublevelZero_nonempty {F : Type*} {f : F → EReal}
(hInf : sInf (Set.range f) < (0 : EReal)) : ({x : F | f x ≤ (0 : EReal)}).Nonempty
If inf f < 0, then there exists a point with f x < 0.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_exists_lt_zero_of_sInf_lt_zero {F : Type*} {f : F → EReal}
(hInf : sInf (Set.range f) < (0 : EReal)) : ∃ x : F, f x < (0 : EReal)
The inner dual cone is unchanged by replacing a set with the closure of its conic hull.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
open scoped InnerProductSpace
variable {H : Type*} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
lemma section14_innerDualCone_closure_coneHull_eq (s : Set H) :
ProperCone.innerDual (E := H)
(closure ((ConvexCone.hull ℝ (Set.insert (0 : H) s) : ConvexCone ℝ H) : Set H)) =
ProperCone.innerDual (E := H) s
Bipolar theorem (inner-product version): the double inner dual cone of a set is the closure of its conic hull.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
open scoped InnerProductSpace
variable {H : Type*} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
lemma section14_innerDualCone_innerDualCone_eq_closure_coneHull (s : Set H) :
(ProperCone.innerDual (E := H) (ProperCone.innerDual (E := H) s : Set H) : Set H) =
closure ((ConvexCone.hull ℝ (Set.insert (0 : H) s) : ConvexCone ℝ H) : Set H)
In a real inner product space, the polar cone condition for toDual y is the inequality ⟪y, x⟫ ≤ 0 on the set.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
open scoped InnerProductSpace
variable {H : Type*} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
lemma section14_toDual_mem_polarCone_iff (s : Set H) (y : H) :
((InnerProductSpace.toDual ℝ H y : StrongDual ℝ H) : Module.Dual ℝ H) ∈
polarCone (E := H) s ↔
∀ x, x ∈ s → ⟪y, x⟫_ℝ ≤ 0
In a real inner product space, the polar cone condition for toDual y is equivalent to membership of -y in the inner dual cone of the set.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
open scoped InnerProductSpace
variable {H : Type*} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
lemma section14_toDual_mem_polarCone_iff_neg_mem_innerDualCone (s : Set H) (y : H) :
((InnerProductSpace.toDual ℝ H y : StrongDual ℝ H) : Module.Dual ℝ H) ∈
polarCone (E := H) s ↔ (-y) ∈ (ProperCone.innerDual (E := H) s : Set H)
If f* φ ≤ 0, then φ is nonpositive on the 0-sublevel set of f.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_le_zero_on_sublevelZero_of_fenchelConjugate_le_zero {f : E → EReal}
(φ : Module.Dual ℝ E)
(hφ : fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f φ ≤ (0 : EReal)) :
∀ x, f x ≤ (0 : EReal) → φ x ≤ 0
The Fenchel conjugate is nonpositive iff the functional is dominated by the primal function: f* φ ≤ 0 iff φ x ≤ f x for all x.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_fenchelConjugate_le_zero_iff {f : E → EReal} (φ : Module.Dual ℝ E) :
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f φ ≤ (0 : EReal) ↔
∀ x : E, (φ x : EReal) ≤ f x
The 0-sublevel set of f* lies in the polar cone of the convex cone generated by the 0-sublevel set of f.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_sublevel_fenchelConjugate_le_zero_subset_polarCone_hull_sublevelZero
{f : E → EReal} :
{φ : Module.Dual ℝ E |
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f φ ≤ (0 : EReal)} ⊆
polarCone (E := E)
((ConvexCone.hull ℝ {x : E | f x ≤ (0 : EReal)} : ConvexCone ℝ E) : Set E)
If a set of dual elements is contained in the polar cone of K, then the closed convex cone generated by that set is also contained in the polar cone of K.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_closure_coneHull_subset_polarCone_of_subset
[TopologicalSpace (Module.Dual ℝ E)] {K : Set E} {T : Set (Module.Dual ℝ E)}
(hT : T ⊆ polarCone (E := E) K) (hclosed : IsClosed (polarCone (E := E) K)) :
closure
((ConvexCone.hull ℝ T : ConvexCone ℝ (Module.Dual ℝ E)) :
Set (Module.Dual ℝ E)) ⊆
polarCone (E := E) K
The polar cone is unchanged by replacing a set with the convex cone it generates.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_polarCone_hull_eq (S : Set E) :
polarCone (E := E) ((ConvexCone.hull ℝ S : ConvexCone ℝ E) : Set E) =
polarCone (E := E) S
In a finite-dimensional real topological vector space, membership in a polar cone propagates from a set to its closure (and conversely).
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_polarCone_closure_eq
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E] (K : Set E) :
polarCone (E := E) (closure K) = polarCone (E := E) K
Restrict a polar-cone condition from the closed conic hull of a set to the set itself.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_mem_polarCone_of_mem_polarCone_closure_coneHull
(S : Set E) [TopologicalSpace E] (φ : Module.Dual ℝ E)
(hφ :
φ ∈
polarCone (E := E)
(closure ((ConvexCone.hull ℝ S : ConvexCone ℝ E) : Set E))) :
φ ∈ polarCone (E := E) S
The 0-sublevel set of the Fenchel conjugate lies in the polar cone of the 0-sublevel set of the primal function.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_sublevel_fenchelConjugate_le_zero_subset_polarCone_sublevelZero
{f : E → EReal} :
{φ : Module.Dual ℝ E |
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f φ ≤ (0 : EReal)} ⊆
polarCone (E := E) {x : E | f x ≤ (0 : EReal)}
The weak topology on the algebraic dual induced by the evaluation pairing. This is the canonical topology in which closure statements about polar sets and conjugate sublevel sets are meaningful; it ensures that all evaluation maps φ ↦ φ x are continuous.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
noncomputable local instance section14_instTopologicalSpace_dualWeak :
TopologicalSpace (Module.Dual ℝ E)
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
noncomputable local instance section14_instContinuousAdd_dualWeak :
ContinuousAdd (Module.Dual ℝ E)
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
noncomputable local instance section14_instIsTopologicalAddGroup_dualWeak :
IsTopologicalAddGroup (Module.Dual ℝ E)
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
noncomputable local instance section14_instContinuousSMul_dualWeak :
ContinuousSMul ℝ (Module.Dual ℝ E)
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
noncomputable local instance section14_instLocallyConvexSpace_dualWeak :
LocallyConvexSpace ℝ (Module.Dual ℝ E)
If T is nonempty, then the closure of the convex cone generated by T is nonempty.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_closure_coneHull_nonempty_of_nonempty [TopologicalSpace (Module.Dual ℝ E)]
{T : Set (Module.Dual ℝ E)} (hT : T.Nonempty) :
(closure
((ConvexCone.hull ℝ T : ConvexCone ℝ (Module.Dual ℝ E)) :
Set (Module.Dual ℝ E))).Nonempty
If φ is not in the weak-closure of the convex cone generated by T, then there is a point x : E such that all elements of that closed cone are nonpositive on x, but φ x > 0. This is a separation argument in the weak topology; the separating functional is represented as an evaluation map using reflexivity of finite-dimensional spaces.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_exists_eval_pos_of_not_mem_closure_coneHull
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
(T : Set (Module.Dual ℝ E))
(hKne :
(closure
((ConvexCone.hull ℝ T : ConvexCone ℝ (Module.Dual ℝ E)) :
Set (Module.Dual ℝ E))).Nonempty)
(φ : Module.Dual ℝ E)
(hφ :
φ ∉
closure
((ConvexCone.hull ℝ T : ConvexCone ℝ (Module.Dual ℝ E)) :
Set (Module.Dual ℝ E))) :
∃ x : E,
(∀ ψ : Module.Dual ℝ E,
ψ ∈
closure
((ConvexCone.hull ℝ T : ConvexCone ℝ (Module.Dual ℝ E)) :
Set (Module.Dual ℝ E)) →
ψ x ≤ 0) ∧
(0 : ℝ) < φ x
(Theorem 14.3, auxiliary) Nonemptiness of the 0-sublevel set of the Fenchel conjugate. In the textbook proof this is deduced from the Fenchel–Moreau identity at 0: f 0 = - inf (f*), so f 0 > 0 forces inf (f*) < 0 and hence {φ | f* φ ≤ 0} is nonempty. In this development, the corresponding Section 13/Fenchel–Moreau bridge is not yet imported due to a global name collision on _root_.fenchelConjugateBilin.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_sublevelZero_fenchelConjugate_nonempty
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
{f : E → EReal}
(hf : ProperConvexERealFunction (F := E) f) (hf_closed : LowerSemicontinuous f)
(hf0 : (0 : EReal) < f 0) (hf0_ltTop : f 0 < ⊤) :
({φ : Module.Dual ℝ E |
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f φ ≤ (0 : EReal)}
: Set (Module.Dual ℝ E)).Nonempty
Positively homogeneous hull generated by f: k(x) = inf_{t>0} t * f(t⁻¹ • x). This is the standard conic hull construction used in the proof of Theorem 14.3.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
noncomputable def section14_posHomHull (f : E → EReal) : E → EReal :=
fun x =>
sInf {r : EReal | ∃ t : ℝ, 0 < t ∧ r = (t : EReal) * f (t⁻¹ • x)}
Multiplication by a positive real number is an order isomorphism on EReal.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
noncomputable def section14_mulPosOrderIso (t : ℝ) (ht : 0 < t) : EReal ≃o EReal
Rewrite section14_posHomHull as an indexed infimum over positive scalars.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_posHomHull_eq_iInf (f : E → EReal) (x : E) :
section14_posHomHull (E := E) f x =
⨅ t : {t : ℝ // 0 < t}, (t.1 : EReal) * f (t.1⁻¹ • x)
Positive homogeneity of section14_posHomHull for strictly positive scalars.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_posHomHull_smul (f : E → EReal) {t : ℝ} (ht : 0 < t) (x : E) :
section14_posHomHull (E := E) f (t • x) =
(t : EReal) * section14_posHomHull (E := E) f x
The positively-homogeneous hull never exceeds the original function (take t = 1).
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_posHomHull_le (f : E → EReal) (x : E) : section14_posHomHull (E := E) f x ≤ f x
A linear functional dominated by f is also dominated by the positively-homogeneous hull of f.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_le_posHomHull_of_le (f : E → EReal) (φ : Module.Dual ℝ E)
(hφ : ∀ x : E, (φ x : EReal) ≤ f x) :
∀ x : E, (φ x : EReal) ≤ section14_posHomHull (E := E) f x
Pointwise domination by the positively-homogeneous hull of f is equivalent to domination by f.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_le_posHomHull_iff_le (f : E → EReal) (φ : Module.Dual ℝ E) :
(∀ x : E, (φ x : EReal) ≤ section14_posHomHull (E := E) f x) ↔ ∀ x : E, (φ x : EReal) ≤ f x
The 0-sublevel set of f* is unchanged if f is replaced by its positively-homogeneous hull.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_sublevelZero_fenchelConjugate_posHomHull_eq (f : E → EReal) :
{φ : Module.Dual ℝ E |
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) (section14_posHomHull (E := E) f)
φ ≤ (0 : EReal)} =
{φ : Module.Dual ℝ E |
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f φ ≤ (0 : EReal)}
A basic scaling estimate for the positively-homogeneous hull: choosing the same scaling parameter in the infimum gives k (t • x) ≤ t * f x for t > 0.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_posHomHull_smul_le (f : E → EReal) {t : ℝ} (ht : 0 < t) (x : E) :
section14_posHomHull (E := E) f (t • x) ≤ (t : EReal) * f x§14 — Polars of Convex Sets — Part 6
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yuhao Jiang, Zaiwen Wen.
(Theorem 14.3, auxiliary) If f* φ is strictly positive, then (posHomHull f)* φ = ⊤. This is the scaling argument used to identify the effective domain of the conjugate of the positively-homogeneous hull.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_fenchelConjugate_posHomHull_eq_top_of_fenchelConjugate_pos {f : E → EReal}
(φ : Module.Dual ℝ E)
(hpos :
(0 : EReal) <
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f φ) :
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ))
(section14_posHomHull (E := E) f) φ = ⊤
(Theorem 14.3, auxiliary) The effective domain of the conjugate of the positively homogeneous hull is exactly the 0-sublevel set of the original conjugate.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_erealDom_fenchelConjugate_posHomHull_eq_sublevelZero {f : E → EReal} :
erealDom
(fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ))
(section14_posHomHull (E := E) f)) =
{φ : Module.Dual ℝ E |
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f φ ≤ (0 : EReal)}
The positively-homogeneous hull takes value 0 at the origin, provided f 0 is finite and strictly positive.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_posHomHull_zero {f : E → EReal} (hf0 : (0 : EReal) < f 0) (hf0_ltTop : f 0 < ⊤)
(hf0_ne_bot : f 0 ≠ ⊥) :
section14_posHomHull (E := E) f 0 = 0
Strict negativity for the positively-homogeneous hull forces membership in the conic hull of the 0-sublevel set of f.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_strictSublevel_posHomHull_subset_coneHull_sublevelZero
{f : E → EReal} (hf : ProperConvexERealFunction (F := E) f) :
{x : E | section14_posHomHull (E := E) f x < (0 : EReal)} ⊆
((ConvexCone.hull ℝ {x : E | f x ≤ (0 : EReal)} : ConvexCone ℝ E) : Set E)
(Theorem 14.3, auxiliary) Under the hypotheses ensuring nonemptiness of the dual 0-sublevel set, the positively homogeneous hull never takes the value ⊥.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_posHomHull_ne_bot
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
{f : E → EReal}
(hf : ProperConvexERealFunction (F := E) f) (hf_closed : LowerSemicontinuous f)
(hf0 : (0 : EReal) < f 0) (hf0_ltTop : f 0 < ⊤) :
∀ x : E, section14_posHomHull (E := E) f x ≠ ⊥
(Theorem 14.3, auxiliary) Subadditivity of the positively homogeneous hull.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_posHomHull_subadd
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
{f : E → EReal}
(hf : ProperConvexERealFunction (F := E) f) (hf_closed : LowerSemicontinuous f)
(hf0 : (0 : EReal) < f 0) (hf0_ltTop : f 0 < ⊤) :
∀ x y : E,
section14_posHomHull (E := E) f (x + y) ≤
section14_posHomHull (E := E) f x + section14_posHomHull (E := E) f y
(Theorem 14.3, auxiliary) The positively homogeneous hull is a proper convex function.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_properConvex_posHomHull
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
{f : E → EReal}
(hf : ProperConvexERealFunction (F := E) f) (hf_closed : LowerSemicontinuous f)
(hf0 : (0 : EReal) < f 0) (hf0_ltTop : f 0 < ⊤) :
ProperConvexERealFunction (F := E) (section14_posHomHull (E := E) f)
In the weak topology on the algebraic dual induced by evaluation, every evaluation map φ ↦ φ x is continuous.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
noncomputable local instance section14_instTopologicalSpace_dualWeak_part6 :
TopologicalSpace (Module.Dual ℝ E)
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_continuous_dual_apply (x : E) :
Continuous fun φ : Module.Dual ℝ E => φ x
The Fenchel conjugate (with respect to the evaluation pairing) is lower semicontinuous in the weak topology on Module.Dual ℝ E.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_lowerSemicontinuous_fenchelConjugate_dual (f : E → EReal) :
LowerSemicontinuous
(fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f)
The Fenchel conjugate with respect to the flipped evaluation pairing is lower semicontinuous as a function of the primal variable (in finite dimensions, where all linear functionals are continuous).
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_lowerSemicontinuous_fenchelConjugate_flip
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E]
(g : Module.Dual ℝ E → EReal) :
LowerSemicontinuous
(fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)).flip g)
(Theorem 14.3, auxiliary) The *closed hull* cl k of the positively homogeneous hull is modeled as the Fenchel biconjugate (k*)*, and is lower semicontinuous.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_lowerSemicontinuous_posHomHull
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E]
{f : E → EReal} (_hf_closed : LowerSemicontinuous f) :
let p := LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)
LowerSemicontinuous
(fenchelConjugateBilin p.flip (fenchelConjugateBilin p (section14_posHomHull (E := E) f)))
(Theorem 14.3, auxiliary) The Fenchel conjugate of a proper EReal function (with respect to any real bilinear pairing) never takes the value ⊥.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_fenchelConjugate_ne_bot_of_proper {E₁ F₁ : Type*} [AddCommGroup E₁] [Module ℝ E₁]
[AddCommGroup F₁] [Module ℝ F₁] (p : E₁ →ₗ[ℝ] F₁ →ₗ[ℝ] ℝ) {f : E₁ → EReal}
(hf : ProperERealFunction f) :
∀ y : F₁, fenchelConjugateBilin (E := E₁) p f y ≠ ⊥
(Theorem 14.3, auxiliary) Antitonicity of the Fenchel conjugate in the primal function.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_fenchelConjugate_anti {E₁ F₁ : Type*} [AddCommGroup E₁] [Module ℝ E₁]
[AddCommGroup F₁] [Module ℝ F₁] (p : E₁ →ₗ[ℝ] F₁ →ₗ[ℝ] ℝ) {f g : E₁ → EReal}
(hfg : ∀ x, f x ≤ g x) :
∀ y : F₁, fenchelConjugateBilin (E := E₁) p g y ≤ fenchelConjugateBilin (E := E₁) p f y
(Theorem 14.3, auxiliary) Fenchel biconjugate inequality: if f is proper, then (f*)* ≤ f.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_fenchelBiconjugate_le_of_proper {E₁ F₁ : Type*} [AddCommGroup E₁] [Module ℝ E₁]
[AddCommGroup F₁] [Module ℝ F₁] (p : E₁ →ₗ[ℝ] F₁ →ₗ[ℝ] ℝ) {f : E₁ → EReal}
(hf : ProperERealFunction f) :
∀ x : E₁, fenchelConjugateBilin (E := F₁) p.flip (fenchelConjugateBilin (E := E₁) p f) x ≤ f x
(Theorem 14.3, auxiliary) The Fenchel triconjugate with respect to evaluation satisfies ((f*)*)* = f* for a proper function f.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_fenchelConjugate_triconjugate_eq_of_proper {E₁ F₁ : Type*} [AddCommGroup E₁]
[Module ℝ E₁] [AddCommGroup F₁] [Module ℝ F₁] (p : E₁ →ₗ[ℝ] F₁ →ₗ[ℝ] ℝ) {f : E₁ → EReal}
(hf : ProperERealFunction f)
(hdom : ∃ y0 : F₁, fenchelConjugateBilin (E := E₁) p f y0 < ⊤) :
fenchelConjugateBilin (E := E₁) p (fenchelConjugateBilin (E := F₁) p.flip (fenchelConjugateBilin (E := E₁) p f)) =
fenchelConjugateBilin (E := E₁) p f
(Theorem 14.3, auxiliary) The Fenchel biconjugate of the positively homogeneous hull is a proper convex function.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_properConvex_fenchelBiconjugate_posHomHull
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
{f : E → EReal}
(hf : ProperConvexERealFunction (F := E) f) (hf_closed : LowerSemicontinuous f)
(hf0 : (0 : EReal) < f 0) (hf0_ltTop : f 0 < ⊤) :
let p := LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)
let k : E → EReal := section14_posHomHull (E := E) f
ProperConvexERealFunction (F := E) (fenchelConjugateBilin p.flip (fenchelConjugateBilin p k))§14 — Polars of Convex Sets — Part 7
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yuhao Jiang, Zaiwen Wen.
(Theorem 14.3, auxiliary) If inf f < 0, then the 0-sublevel set of the positively homogeneous hull is contained in the closure of its strict 0-sublevel set.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_sublevel_posHomHull_le_zero_subset_closure_strictSublevel
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
{f : E → EReal} (hf : ProperConvexERealFunction (F := E) f)
(hf_closed : LowerSemicontinuous f) (hf0 : (0 : EReal) < f 0) (hf0_ltTop : f 0 < ⊤)
(hInf : sInf (Set.range f) < (0 : EReal)) :
{x : E | section14_posHomHull (E := E) f x ≤ (0 : EReal)} ⊆
closure {x : E | section14_posHomHull (E := E) f x < (0 : EReal)}
(Theorem 14.3, auxiliary) Recession directions of posHomHull f lie in the closed conic hull of the 0-sublevel set {x | f x ≤ 0}.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_recessionCone_posHomHull_subset_closure_coneHull_sublevelZero
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
{f : E → EReal}
(hf : ProperConvexERealFunction (F := E) f) (hf_closed : LowerSemicontinuous f)
(hf0 : (0 : EReal) < f 0) (hf0_ltTop : f 0 < ⊤) (hInf : sInf (Set.range f) < (0 : EReal)) :
recessionConeEReal (F := E) (section14_posHomHull (E := E) f) ⊆
closure ((ConvexCone.hull ℝ {x : E | f x ≤ (0 : EReal)} : ConvexCone ℝ E) : Set E)
(Theorem 14.3, auxiliary) If A ⊆ closure B and φ is nonpositive on B, then φ is nonpositive on A. This is the form needed to transfer a polar condition along an inclusion into a closure.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_mem_polarCone_of_mem_polarCone_closure_of_subset
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E]
{A B : Set E} {φ : Module.Dual ℝ E} (hA : A ⊆ closure B)
(hφ : φ ∈ polarCone (E := E) B) :
φ ∈ polarCone (E := E) A
Transport membership in recessionConeEReal along a linear equivalence.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_mem_recessionConeEReal_iff_map_mem_recessionConeEReal_of_linearEquiv
{E₁ F₁ : Type*} [AddCommGroup E₁] [Module ℝ E₁] [AddCommGroup F₁] [Module ℝ F₁]
(e : E₁ ≃ₗ[ℝ] F₁) (g : E₁ → EReal) (y : E₁) :
y ∈ recessionConeEReal (F := E₁) g ↔
e y ∈ recessionConeEReal (F := F₁) (fun z : F₁ => g (e.symm z))
recessionConeEReal commutes with linear equivalences, as a set.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_recessionConeEReal_image_linearEquiv
{E₁ F₁ : Type*} [AddCommGroup E₁] [Module ℝ E₁] [AddCommGroup F₁] [Module ℝ F₁]
(e : E₁ ≃ₗ[ℝ] F₁) (g : E₁ → EReal) :
e '' recessionConeEReal (F := E₁) g =
recessionConeEReal (F := F₁) (fun z : F₁ => g (e.symm z))
Recession directions of a subadditive function with g 0 = 0 are exactly its 0-sublevel set.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_recessionConeEReal_eq_sublevel_of_subadd_zero {F : Type*} [AddCommGroup F]
(g : F → EReal) (hg0 : g 0 = 0) (hg_subadd : ∀ x y, g (x + y) ≤ g x + g y) :
recessionConeEReal (F := F) g = {y : F | g y ≤ (0 : EReal)}
Recession directions give a pointwise inequality at the origin: if y ∈ recessionConeEReal g and 0 ∈ dom g, then g y ≤ g 0.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_le_at_zero_of_mem_recessionConeEReal {F : Type*} [AddCommGroup F] (g : F → EReal)
{y : F} (hy : y ∈ recessionConeEReal (F := F) g) (h0dom : (0 : F) ∈ erealDom g) :
g y ≤ g 0
For kcl = (k*)* coming from the positively homogeneous hull k of f, we have 0 ∈ dom kcl and kcl 0 ≤ 0.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_kcl_zero_dom_and_le_zero
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
{f : E → EReal}
(hf : ProperConvexERealFunction (F := E) f) (hf_closed : LowerSemicontinuous f)
(hf0 : (0 : EReal) < f 0) (hf0_ltTop : f 0 < ⊤) :
let p := LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)
let k : E → EReal := section14_posHomHull (E := E) f
let kcl : E → EReal := fenchelConjugateBilin p.flip (fenchelConjugateBilin p k)
(0 : E) ∈ erealDom kcl ∧ kcl 0 ≤ (0 : EReal)
(Theorem 14.3, auxiliary) If y is not in the closure of the strict ε-sublevel set of k, then (y, ε/2) is not in the closure of the (real) epigraph of k.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_not_mem_closure_epigraph_of_not_mem_closure_strictSublevel
{F : Type*} [TopologicalSpace F] {k : F → EReal} {y : F} {ε : ℝ} (hε : 0 < ε)
(hy : y ∉ closure {z : F | k z < (ε : EReal)}) :
let r0 : ℝ := ε / 2
let epi : Set (F × ℝ) := {p : F × ℝ | k p.1 ≤ (p.2 : EReal)}
(y, r0) ∉ closure epi
(Theorem 14.3, auxiliary) From an affine bound ψ + c ≤ k, one gets k* ψ ≤ -c.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_fenchelConjugate_le_neg_const_of_add_const_le
{k : E → EReal} {ψ : Module.Dual ℝ E} {c : ℝ}
(hk_ne_bot : ∀ x : E, k x ≠ ⊥) (hψ : ∀ x : E, (ψ x : EReal) + (c : EReal) ≤ k x) :
let p := LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)
fenchelConjugateBilin p k ψ ≤ (-(c : EReal))
(Theorem 14.3, auxiliary) A dual certificate forces strict positivity of the biconjugate.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_fenchelBiconjugate_pos_of_dual_certificate
{k : E → EReal} {y : E} {ψ : Module.Dual ℝ E} {c r0 : ℝ}
(hr0 : (r0 : EReal) < (ψ y : EReal) + (c : EReal))
(hconj :
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) k ψ ≤ (-(c : EReal))) :
let p := LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)
let kcl : E → EReal := fenchelConjugateBilin p.flip (fenchelConjugateBilin p k)
(r0 : EReal) < kcl y
(Theorem 14.3, auxiliary) The closure of a real epigraph is upward closed in the ℝ-coordinate. This is the basic monotonicity property inherited from the definition k x ≤ r.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_closure_epigraph_upwardClosed
{F : Type*} [TopologicalSpace F] {k : F → EReal} {x : F} {r r' : ℝ}
(hr : (x, r) ∈ closure {p : F × ℝ | k p.1 ≤ (p.2 : EReal)}) (hrr' : r ≤ r') :
(x, r') ∈ closure {p : F × ℝ | k p.1 ≤ (p.2 : EReal)}
(Theorem 14.3, auxiliary) If ψ₁ + c₁ ≤ k, then g(x,r) = ψ₁ x - r is bounded above on closure (epi k).
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_linearPerturb_bound_on_closure_epigraph
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E]
{k : E → EReal} {ψ₁ : Module.Dual ℝ E} {c₁ : ℝ}
(hminor : ∀ x, (ψ₁ x : EReal) + (c₁ : EReal) ≤ k x) :
let epi : Set (E × ℝ) := {p : E × ℝ | k p.1 ≤ (p.2 : EReal)}
let g : StrongDual ℝ (E × ℝ) :=
(ψ₁.toContinuousLinearMap.comp (ContinuousLinearMap.fst ℝ E ℝ)) -
(ContinuousLinearMap.snd ℝ E ℝ)
∀ p ∈ closure epi, g p ≤ -c₁
(Theorem 14.3, auxiliary) Separation of a point from closure (epi k) yields a dual certificate ψ + c ≤ k that is strictly above the point in the epigraph direction.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_sep_dual_certificate_of_not_mem_closure_epigraph
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
{k : E → EReal} (hk : ProperConvexERealFunction (F := E) k) (hk0 : k 0 = 0)
(hminor : ∃ (ψ₁ : Module.Dual ℝ E) (c₁ : ℝ), ∀ x, (ψ₁ x : EReal) + (c₁ : EReal) ≤ k x)
{y : E} {r0 : ℝ} :
let epi : Set (E × ℝ) := {p : E × ℝ | k p.1 ≤ (p.2 : EReal)}
(y, r0) ∉ closure epi →
∃ (ψ : Module.Dual ℝ E) (c : ℝ),
(r0 : EReal) < (ψ y : EReal) + (c : EReal) ∧
∀ x, (ψ x : EReal) + (c : EReal) ≤ k x
(Theorem 14.3, auxiliary) Approximation of kcl-sublevel points by strict sublevels of k. This is the analytic heart of Theorem 14.3: from kcl y ≤ 0 one should be able to find points arbitrarily close to y where the (possibly non-closed) positively homogeneous hull k takes an arbitrarily small (strict) value.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_fenchelBiconjugate_sublevelZero_approx_by_posHomHull
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
{f : E → EReal}
(hf : ProperConvexERealFunction (F := E) f) (hf_closed : LowerSemicontinuous f)
(hf0 : (0 : EReal) < f 0) (hf0_ltTop : f 0 < ⊤) :
let p := LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)
let k : E → EReal := section14_posHomHull (E := E) f
let kcl : E → EReal := fenchelConjugateBilin p.flip (fenchelConjugateBilin p k)
∀ {y : E}, kcl y ≤ (0 : EReal) →
∀ {ε : ℝ}, 0 < ε → y ∈ closure {z : E | k z < (ε : EReal)}
(Theorem 14.3, auxiliary) If k z is small and moving in a fixed direction makes k very negative, then the translated point lies in the strict 0-sublevel set.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_strictSublevel_mem_of_smallValue_add_smul_negDir
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
{f : E → EReal}
(hf : ProperConvexERealFunction (F := E) f) (hf_closed : LowerSemicontinuous f)
(hf0 : (0 : EReal) < f 0) (hf0_ltTop : f 0 < ⊤)
{x0 z : E} {ε t : ℝ} (ht : 0 < t)
(hz : section14_posHomHull (E := E) f z < (ε : EReal))
(hx0 : (t : EReal) * f x0 < -(ε : EReal)) :
section14_posHomHull (E := E) f (z + t • x0) < (0 : EReal)
(Theorem 14.3, auxiliary) From kcl y ≤ 0 we can reach the strict 0-sublevel of k by arbitrarily small perturbations, hence y lies in the closure of {k < 0}.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_sublevelZero_fenchelBiconjugate_posHomHull_subset_closure_strictSublevel_posHomHull
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
{f : E → EReal}
(hf : ProperConvexERealFunction (F := E) f) (hf_closed : LowerSemicontinuous f)
(hf0 : (0 : EReal) < f 0) (hf0_ltTop : f 0 < ⊤) (hInf : sInf (Set.range f) < (0 : EReal)) :
let p := LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)
let k : E → EReal := section14_posHomHull (E := E) f
let kcl : E → EReal := fenchelConjugateBilin p.flip (fenchelConjugateBilin p k)
{y : E | kcl y ≤ (0 : EReal)} ⊆ closure {z : E | k z < (0 : EReal)}§14 — Polars of Convex Sets — Part 8
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yuhao Jiang, Zaiwen Wen.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
noncomputable local instance section14_instTopologicalSpace_dualWeak_part8 :
TopologicalSpace (Module.Dual ℝ E)
(Theorem 14.3, auxiliary) Analytic core: a point in the 0-sublevel set of the Fenchel biconjugate of k = posHomHull f lies in the closed conic hull of {x | f x ≤ 0}.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_sublevelZero_fenchelBiconjugate_posHomHull_subset_closure_coneHull_sublevelZero
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
{f : E → EReal}
(hf : ProperConvexERealFunction (F := E) f) (hf_closed : LowerSemicontinuous f)
(hf0 : (0 : EReal) < f 0) (hf0_ltTop : f 0 < ⊤) (hInf : sInf (Set.range f) < (0 : EReal)) :
let p := LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)
let k : E → EReal := section14_posHomHull (E := E) f
let kcl : E → EReal := fenchelConjugateBilin p.flip (fenchelConjugateBilin p k)
{y : E | kcl y ≤ (0 : EReal)} ⊆
closure ((ConvexCone.hull ℝ {x : E | f x ≤ (0 : EReal)} : ConvexCone ℝ E) : Set E)
(Theorem 14.3, auxiliary) Recession directions of the Fenchel biconjugate (k*)* lie in the closure of the recession directions of k. This is the missing structural bridge in the conic-separation proof of Theorem 14.3.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_recessionConeEReal_fenchelBiconjugate_subset_closure
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
{f : E → EReal}
(hf : ProperConvexERealFunction (F := E) f) (hf_closed : LowerSemicontinuous f)
(hf0 : (0 : EReal) < f 0) (hf0_ltTop : f 0 < ⊤) (hInf : sInf (Set.range f) < (0 : EReal)) :
let p := LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)
let k : E → EReal := section14_posHomHull (E := E) f
let kcl : E → EReal := fenchelConjugateBilin p.flip (fenchelConjugateBilin p k)
recessionConeEReal (F := E) kcl ⊆ closure (recessionConeEReal (F := E) k)
(Theorem 14.3, auxiliary) Geometric conversion step for the separation proof. This is the missing implication in the contradiction argument: if a point x : E is nonpositive under every functional in the closed cone generated by {φ | f* φ ≤ 0}, then x lies in the closed cone generated by {x | f x ≤ 0}. The textbook proof routes this through the positively-homogeneous hull k of f, the conjugacy (cl k)* = ι_{ {φ | f* φ ≤ 0} }, and the polar/recession correspondence (Theorem 14.2).
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_dual_flip_closure_coneHull_sublevelZero_fenchelConjugate_subset_closure_coneHull_sublevelZero
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
{f : E → EReal}
(hf : ProperConvexERealFunction (F := E) f) (hf_closed : LowerSemicontinuous f)
(hf0 : (0 : EReal) < f 0) (hf0_ltTop : f 0 < ⊤) (hInf : sInf (Set.range f) < (0 : EReal)) :
(PointedCone.dual (R := ℝ)
((-LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ))).flip
(closure
((ConvexCone.hull ℝ
{φ : Module.Dual ℝ E |
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f φ ≤
(0 : EReal)} :
ConvexCone ℝ (Module.Dual ℝ E)) :
Set (Module.Dual ℝ E))) :
Set E) ⊆
closure ((ConvexCone.hull ℝ {x : E | f x ≤ (0 : EReal)} : ConvexCone ℝ E) : Set E)
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_polarCone_sublevelZero_subset_closure_coneHull_sublevelZero_fenchelConjugate
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
{f : E → EReal}
(hf : ProperConvexERealFunction (F := E) f) (hf_closed : LowerSemicontinuous f)
(hf0 : (0 : EReal) < f 0) (hf0_ltTop : f 0 < ⊤) (hInf : sInf (Set.range f) < (0 : EReal)) :
polarCone (E := E) {x : E | f x ≤ (0 : EReal)} ⊆
closure
((ConvexCone.hull ℝ
{φ : Module.Dual ℝ E |
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f φ ≤
(0 : EReal)} :
ConvexCone ℝ (Module.Dual ℝ E)) :
Set (Module.Dual ℝ E))§14 — Polars of Convex Sets — Part 9
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yuhao Jiang, Zaiwen Wen.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
noncomputable local instance section14_instTopologicalSpace_dualWeak_part9 :
TopologicalSpace (Module.Dual ℝ E)
A canonical bilinear pairing on triples (λ, x, μ) induced by a bilinear pairing p on E: ⟪(λ, x, μ), (λ★, x★, μ★)⟫ = λ * λ★ + p x x★ + μ * μ★.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
noncomputable def section14_triplePairing (p : E →ₗ[ℝ] E →ₗ[ℝ] ℝ) :
(ℝ × E × ℝ) →ₗ[ℝ] (ℝ × E × ℝ) →ₗ[ℝ] ℝ
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
@[simp] lemma section14_triplePairing_apply (p : E →ₗ[ℝ] E →ₗ[ℝ] ℝ) (z w : ℝ × E × ℝ) :
section14_triplePairing (E := E) p z w = z.1 * w.1 + p z.2.1 w.2.1 + z.2.2 * w.2.2
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
@[simp] lemma section14_triplePairing_one_x_mu_negMuStar_xStar_negOne (p : E →ₗ[ℝ] E →ₗ[ℝ] ℝ)
(x xStar : E) (μ μStar : ℝ) :
section14_triplePairing (E := E) p (1, x, μ) (-μStar, xStar, -1) =
(-μStar + p x xStar - μ)
A pointwise reformulation of fenchelConjugateBilin p f x★ ≤ μ★.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_fenchelConjugate_le_iff_forall {F : Type*} [AddCommGroup F] [Module ℝ F]
(p : E →ₗ[ℝ] F →ₗ[ℝ] ℝ) (f : E → EReal) (xStar : F) (μStar : EReal) :
fenchelConjugateBilin p f xStar ≤ μStar ↔ ∀ x, (p x xStar : EReal) - f x ≤ μStar
In finite dimensions, PointedCone.dual p s is a closed set.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_isClosed_pointedCone_dual {M N : Type*} [AddCommGroup M] [AddCommGroup N]
[Module ℝ M] [Module ℝ N] [TopologicalSpace N] [IsTopologicalAddGroup N] [ContinuousSMul ℝ N]
[T2Space N] [FiniteDimensional ℝ N]
(p : M →ₗ[ℝ] N →ₗ[ℝ] ℝ) (s : Set M) :
IsClosed (PointedCone.dual (R := ℝ) p s : Set N)
Slice characterization for the lifted-epigraph cone (core of Theorem 14.4).
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_mem_dual_liftedEpigraphCone_slice_iff
[TopologicalSpace E]
{f : E → EReal} (hf : ProperConvexERealFunction (F := E) f)
(p : E →ₗ[ℝ] E →ₗ[ℝ] ℝ) (xStar : E) (μStar : ℝ) :
let q := section14_triplePairing (E := E) p
let K : Set (ℝ × E × ℝ) :=
((ConvexCone.hull ℝ {z : ℝ × E × ℝ | ∃ x μ, z = (1, x, μ) ∧ f x ≤ (μ : EReal)} :
ConvexCone ℝ (ℝ × E × ℝ)) :
Set (ℝ × E × ℝ))
(-μStar, xStar, -1) ∈ (PointedCone.dual (R := ℝ) (-q) K : Set (ℝ × E × ℝ)) ↔
fenchelConjugateBilin p f xStar ≤ (μStar : EReal)
Key inclusion for Theorem 14.3: any functional nonpositive on the closed cone generated by the 0-sublevel set {x | f x ≤ 0} lies in the closed cone generated by the 0-sublevel set of the Fenchel conjugate {φ | f* φ ≤ 0}.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_polarCone_closure_coneHull_sublevelZero_subset_closure_coneHull_sublevelZero_fenchelConjugate
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
{f : E → EReal}
(hf : ProperConvexERealFunction (F := E) f) (hf_closed : LowerSemicontinuous f)
(hf0 : (0 : EReal) < f 0) (hf0_ltTop : f 0 < ⊤) (hInf : sInf (Set.range f) < (0 : EReal)) :
polarCone (E := E)
(closure
((ConvexCone.hull ℝ {x : E | f x ≤ (0 : EReal)} : ConvexCone ℝ E) : Set E)) ⊆
closure
((ConvexCone.hull ℝ
{φ : Module.Dual ℝ E |
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f φ ≤
(0 : EReal)} :
ConvexCone ℝ (Module.Dual ℝ E)) :
Set (Module.Dual ℝ E))
Theorem 14.3. Let f be a closed proper convex function such that f 0 > 0 > inf f. Then the closed convex cones generated by the sublevel sets {x | f x ≤ 0} and {x★ | f* x★ ≤ 0} are polar to each other, where f* denotes the Fenchel–Legendre conjugate of f with respect to the evaluation pairing.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
theorem polar_closure_coneHull_sublevelZero_eq_closure_coneHull_sublevelZero_fenchelConjugate
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
{f : E → EReal}
(hf : ProperConvexERealFunction (F := E) f) (hf_closed : LowerSemicontinuous f)
(hf0 : (0 : EReal) < f 0) (hf0_ltTop : f 0 < ⊤) (hInf : sInf (Set.range f) < (0 : EReal)) :
polarCone (E := E)
(closure
((ConvexCone.hull ℝ {x : E | f x ≤ (0 : EReal)} : ConvexCone ℝ E) : Set E)) =
closure
((ConvexCone.hull ℝ
{φ : Module.Dual ℝ E |
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f φ ≤
(0 : EReal)} :
ConvexCone ℝ (Module.Dual ℝ E)) :
Set (Module.Dual ℝ E)) ∧
PointedCone.dual (R := ℝ)
((-LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ))).flip
(closure
((ConvexCone.hull ℝ
{φ : Module.Dual ℝ E |
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f φ ≤
(0 : EReal)} :
ConvexCone ℝ (Module.Dual ℝ E)) :
Set (Module.Dual ℝ E))) =
closure
((ConvexCone.hull ℝ {x : E | f x ≤ (0 : EReal)} : ConvexCone ℝ E) : Set E)
Coordinate involution used in Theorem 14.4: swapNeg (λ, x, μ) = (-μ, x, -λ).
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
def section14_swapNegₗ (E : Type*) [AddCommGroup E] [Module ℝ E] :
(ℝ × E × ℝ) →ₗ[ℝ] (ℝ × E × ℝ)
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
@[simp] lemma section14_swapNegₗ_apply (z : ℝ × E × ℝ) :
section14_swapNegₗ E z = (-z.2.2, z.2.1, -z.1)
The linear equivalence associated to section14_swapNegₗ.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
def section14_swapNeg (E : Type*) [AddCommGroup E] [Module ℝ E] :
(ℝ × E × ℝ) ≃ₗ[ℝ] (ℝ × E × ℝ)
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
@[simp] lemma section14_swapNeg_apply (z : ℝ × E × ℝ) :
section14_swapNeg E z = (-z.2.2, z.2.1, -z.1)
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
@[simp] lemma section14_swapNeg_swapNeg (z : ℝ × E × ℝ) :
section14_swapNeg E (section14_swapNeg E z) = z
Elements of the dual of the lifted-epigraph cone have nonpositive last coordinate.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_dual_liftedEpigraphCone_last_nonpos
[TopologicalSpace E]
{f : E → EReal} (hf : ProperConvexERealFunction (F := E) f)
(p : E →ₗ[ℝ] E →ₗ[ℝ] ℝ) :
let q := section14_triplePairing (E := E) p
let K : Set (ℝ × E × ℝ) :=
((ConvexCone.hull ℝ {z : ℝ × E × ℝ | ∃ x μ, z = (1, x, μ) ∧ f x ≤ (μ : EReal)} :
ConvexCone ℝ (ℝ × E × ℝ)) :
Set (ℝ × E × ℝ))
∀ w, w ∈ (PointedCone.dual (R := ℝ) (-q) K : Set (ℝ × E × ℝ)) → w.2.2 ≤ 0
Cone-slicing reconstruction for the lifted-epigraph polar (reverse inclusion in Theorem 14.4). If swapNeg z lies in the polar of the lifted-epigraph cone, then z lies in the closure of the cone generated by the epigraph of the Fenchel conjugate.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_swapNeg_preimage_dual_subset_closure_Kstar
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E]
{f : E → EReal} (hf : ProperConvexERealFunction (F := E) f)
(p : E →ₗ[ℝ] E →ₗ[ℝ] ℝ)
(hdom : ∃ x0 : E, ∃ μ0 : ℝ, fenchelConjugateBilin p f x0 ≤ (μ0 : EReal)) :
let q := section14_triplePairing (E := E) p
let K : Set (ℝ × E × ℝ) :=
((ConvexCone.hull ℝ {z : ℝ × E × ℝ | ∃ x μ, z = (1, x, μ) ∧ f x ≤ (μ : EReal)} :
ConvexCone ℝ (ℝ × E × ℝ)) :
Set (ℝ × E × ℝ))
let Kstar : Set (ℝ × E × ℝ) :=
((ConvexCone.hull ℝ
{z : ℝ × E × ℝ |
∃ x μ, z = (1, x, μ) ∧ fenchelConjugateBilin p f x ≤ (μ : EReal)} :
ConvexCone ℝ (ℝ × E × ℝ)) :
Set (ℝ × E × ℝ))
{z : ℝ × E × ℝ |
section14_swapNeg E z ∈ (PointedCone.dual (R := ℝ) (-q) K : Set (ℝ × E × ℝ))} ⊆
closure Kstar
Theorem 14.4. Let f be a closed proper convex function on ℝⁿ. Let K be the convex cone generated by points (1, x, μ) ∈ ℝ × ℝⁿ × ℝ such that μ ≥ f x, and let K★ be the convex cone generated by points (1, x★, μ★) such that μ★ ≥ f★ x★, where f★ is the Fenchel–Legendre conjugate of f. Then cl K★ = { (λ★, x★, μ★) | (-μ★, x★, -λ★) ∈ Kᵒ }. In this formalization, f★ is fenchelConjugateBilin p f for a chosen bilinear pairing p on E, and Kᵒ is modeled as PointedCone.dual (-q) K for a chosen bilinear pairing q on ℝ × E × ℝ.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
theorem closure_coneHull_liftedEpigraph_fenchelConjugate_eq_setOf_swapNeg_mem_dual
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E]
{f : E → EReal} (hf : ProperConvexERealFunction (F := E) f)
(_hf_closed : LowerSemicontinuous f) (p : E →ₗ[ℝ] E →ₗ[ℝ] ℝ)
(q : (ℝ × E × ℝ) →ₗ[ℝ] (ℝ × E × ℝ) →ₗ[ℝ] ℝ)
(hq : q = section14_triplePairing (E := E) p) :
(∃ x0 : E, ∃ μ0 : ℝ, fenchelConjugateBilin p f x0 ≤ (μ0 : EReal)) →
let K : Set (ℝ × E × ℝ) :=
((ConvexCone.hull ℝ {z : ℝ × E × ℝ | ∃ x μ, z = (1, x, μ) ∧ f x ≤ (μ : EReal)} :
ConvexCone ℝ (ℝ × E × ℝ)) :
Set (ℝ × E × ℝ))
let Kstar : Set (ℝ × E × ℝ) :=
((ConvexCone.hull ℝ
{z : ℝ × E × ℝ |
∃ x μ, z = (1, x, μ) ∧ fenchelConjugateBilin p f x ≤ (μ : EReal)} :
ConvexCone ℝ (ℝ × E × ℝ)) :
Set (ℝ × E × ℝ))
closure Kstar =
{z : ℝ × E × ℝ |
(-z.2.2, z.2.1, -z.1) ∈ (PointedCone.dual (R := ℝ) (-q) K : Set (ℝ × E × ℝ))}
Sanity check for Theorem 14.5: mathlib's gauge is real-valued, so it cannot match a support function that can take the value ⊤ (e.g. for C = {0}).
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
-- The weak topology on the algebraic dual induced by evaluation (see `section14_part3`).
lemma section14_sanity_gauge_counterexample :
(fun x : ℝ => (gauge ({0} : Set ℝ) x : EReal)) ≠
(let p := LinearMap.applyₗ (R := ℝ) (M := ℝ) (M₂ := ℝ)
fenchelConjugateBilin p.flip
(erealIndicator (E := Module.Dual ℝ ℝ) (polar (E := ℝ) ({0} : Set ℝ))))§14 — Polars of Convex Sets — Part 10
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yuhao Jiang, Zaiwen Wen.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
noncomputable local instance section14_instTopologicalSpace_dualWeak_part10 :
TopologicalSpace (Module.Dual ℝ E)
Polar sets are intersections of half-spaces {φ | φ x ≤ 1}.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_polar_eq_iInter_halfspaces (C : Set E) :
polar (E := E) C = ⋂ x : E, ⋂ (_ : x ∈ C), {φ : Module.Dual ℝ E | φ x ≤ 1}
Under the weak topology on the algebraic dual induced by evaluation, polar sets are closed.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_isClosed_polar (C : Set E) : IsClosed (polar (E := E) C)
Convexity and 0-membership of a polar set (topology-free).
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_convex_and_zeroMem_polar (C : Set E) :
Convex ℝ (polar (E := E) C) ∧ (0 : Module.Dual ℝ E) ∈ polar (E := E) C
The bipolar identity for closed convex sets containing the origin.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_bipolar_eq_of_closed_convex
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [LocallyConvexSpace ℝ E]
{C : Set E} (hCclosed : IsClosed C) (hCconv : Convex ℝ C) (hC0 : (0 : E) ∈ C) :
{x : E | ∀ φ ∈ polar (E := E) C, φ x ≤ 1} = C
An EReal-valued gauge: sInf {r | 0 < r ∧ x ∈ r • C} (with ⊤ for points not in any positive scaling of C). This is the textbook gauge, unlike mathlib's real-valued gauge.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
noncomputable def section14_egauge (C : Set E) (x : E) : EReal :=
sInf ((fun r : ℝ => (r : EReal)) '' {r : ℝ | 0 < r ∧ x ∈ r • C})
For r ≠ 0, membership in r • C is equivalent to membership of (1 / r) • x in C.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_mem_smul_set_iff (C : Set E) {r : ℝ} (hr : r ≠ 0) {x : E} :
x ∈ r • C ↔ (1 / r) • x ∈ C
If C satisfies the bipolar identity, then x ∈ r • C is equivalent to φ x ≤ r for all φ ∈ polar C (for r > 0).
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_mem_smul_set_iff_forall_polar_le
{C : Set E} (hCpolar : {x : E | ∀ φ ∈ polar (E := E) C, φ x ≤ 1} = C) {r : ℝ} (hr : 0 < r)
{x : E} :
x ∈ r • C ↔ ∀ φ ∈ polar (E := E) C, φ x ≤ r
fenchelConjugateBilin of an erealIndicator is a support function: for the flipped evaluation pairing, it is sSup {φ x | φ ∈ polar C}.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_fenchelConjugate_flip_erealIndicator_polar_eq_sSup (C : Set E) (x : E) :
let p := LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)
fenchelConjugateBilin p.flip
(erealIndicator (E := Module.Dual ℝ E) (polar (E := E) C)) x =
sSup ((fun φ : Module.Dual ℝ E => (φ x : EReal)) '' polar (E := E) C)
fenchelConjugateBilin of an erealIndicator is a support function: for the evaluation pairing, it is sSup {φ x | x ∈ C}.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_fenchelConjugate_erealIndicator_eq_sSup (C : Set E) (φ : Module.Dual ℝ E) :
let p := LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)
fenchelConjugateBilin p (erealIndicator (E := E) C) φ =
sSup ((fun x : E => (φ x : EReal)) '' C)
The EReal-valued gauge of a closed convex set containing 0 agrees with the support function of its polar set.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_egauge_eq_fenchelConjugate_indicator_polar
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [LocallyConvexSpace ℝ E]
{C : Set E} (hCclosed : IsClosed C) (hCconv : Convex ℝ C) (hC0 : (0 : E) ∈ C) :
let p := LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)
(fun x : E => section14_egauge (E := E) C x) =
fenchelConjugateBilin p.flip
(erealIndicator (E := Module.Dual ℝ E) (polar (E := E) C))
Membership in r • polar C is equivalent to a uniform bound φ x ≤ r on C (for r > 0).
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_mem_smul_polar_iff_forall_le (C : Set E) {r : ℝ} (hr : 0 < r)
{φ : Module.Dual ℝ E} :
φ ∈ r • polar (E := E) C ↔ ∀ x ∈ C, φ x ≤ r
The EReal-valued gauge of the polar set agrees with the support function of C.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_egauge_polar_eq_fenchelConjugate_indicator
{C : Set E} (hC0 : (0 : E) ∈ C) :
let p := LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)
(fun φ : Module.Dual ℝ E => section14_egauge (E := Module.Dual ℝ E) (polar (E := E) C) φ) =
fenchelConjugateBilin p (erealIndicator (E := E) C)
Theorem 14.5: Let C be a closed convex set containing the origin. Then the polar C^∘ is another closed convex set containing the origin, and the bipolar identity C^{∘∘} = C holds. The gauge function of C is the support function of C^∘. Dually, the gauge function of C^∘ is the support function of C.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
theorem polar_closed_convex_bipolar_eq_and_gauge_eq_support
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [LocallyConvexSpace ℝ E]
{C : Set E} :
IsClosed C → Convex ℝ C → (0 : E) ∈ C →
(IsClosed (polar (E := E) C) ∧ Convex ℝ (polar (E := E) C) ∧
(0 : Module.Dual ℝ E) ∈ polar (E := E) C) ∧
{x : E | ∀ φ ∈ polar (E := E) C, φ x ≤ 1} = C ∧
(let p := LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)
(fun x : E => section14_egauge (E := E) C x) =
fenchelConjugateBilin p.flip
(erealIndicator (E := Module.Dual ℝ E) (polar (E := E) C)) ∧
(fun φ : Module.Dual ℝ E =>
section14_egauge (E := Module.Dual ℝ E) (polar (E := E) C) φ) =
fenchelConjugateBilin p (erealIndicator (E := E) C))
In a finite-dimensional Hausdorff real topological vector space, a convex set contains the origin in its interior iff every vector can be scaled into it.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_zero_mem_interior_iff_forall_exists_pos_smul_mem
{E : Type*} [AddCommGroup E] [Module ℝ E]
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E]
[FiniteDimensional ℝ E] [T2Space E]
{C : Set E} (hCconv : Convex ℝ C) :
(0 : E) ∈ interior C ↔ ∀ y : E, ∃ ε : ℝ, 0 < ε ∧ ε • y ∈ C
Corollary 14.5.1. Let C be a closed convex set containing the origin. Then the polar C^∘ is bounded if and only if 0 ∈ int C. Dually, C is bounded if and only if 0 ∈ int C^∘.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
theorem polar_bounded_iff_zero_mem_interior_and_dual
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [LocallyConvexSpace ℝ E]
[FiniteDimensional ℝ E] [T2Space E] {C : Set E}
(hCclosed : IsClosed C) (hCconv : Convex ℝ C) (hC0 : (0 : E) ∈ C) :
(@Bornology.IsBounded (Module.Dual ℝ E)
(Bornology.induced (fun φ : Module.Dual ℝ E => fun x : E => φ x))
(polar (E := E) C) ↔ (0 : E) ∈ interior C) ∧
(@Bornology.IsBounded E
(Bornology.induced (fun x : E => fun φ : Module.Dual ℝ E => φ x)) C ↔
(0 : Module.Dual ℝ E) ∈ interior (polar (E := E) C))
If C is given as {x | ∀ φ ∈ Cpolar, φ x ≤ 1}, then its recession cone is exactly the set of directions on which every φ ∈ Cpolar is nonpositive.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_recessionCone_eq_forall_polar_le_zero_of_bipolar
{C : Set E} {Cpolar : Set (Module.Dual ℝ E)}
(hbipolar : {x : E | ∀ φ ∈ Cpolar, φ x ≤ 1} = C) (hC0 : (0 : E) ∈ C) :
Set.recessionCone C = {y : E | ∀ φ ∈ Cpolar, φ y ≤ 0}
Under the same bipolar hypothesis, the set of "two-sided" recession directions is exactly the set of vectors annihilated by all φ ∈ Cpolar.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_neg_rec_inter_rec_eq_forall_polar_eq_zero_of_bipolar
{C : Set E} {Cpolar : Set (Module.Dual ℝ E)}
(hbipolar : {x : E | ∀ φ ∈ Cpolar, φ x ≤ 1} = C) (hC0 : (0 : E) ∈ C) :
(-Set.recessionCone C) ∩ Set.recessionCone C = {y : E | ∀ φ ∈ Cpolar, φ y = 0}
Under the bipolar hypothesis, the lineality subspace of C (spanned by (-rec C) ∩ rec C) is the dual coannihilator of the span of Cpolar.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_dualCoannihilator_span_polar_eq_span_negRec_inter_rec_of_bipolar
{C : Set E} {Cpolar : Set (Module.Dual ℝ E)}
(hbipolar : {x : E | ∀ φ ∈ Cpolar, φ x ≤ 1} = C) (hC0 : (0 : E) ∈ C) :
(Submodule.span ℝ Cpolar).dualCoannihilator =
Submodule.span ℝ ((-Set.recessionCone C) ∩ Set.recessionCone C)
If Cpolar = polar C and 0 ∈ Cpolar, then recession directions in Cpolar are exactly the functionals nonpositive on C.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_recessionCone_Cpolar_eq_forall_primal_le_zero_of_hpolar
{C : Set E} {Cpolar : Set (Module.Dual ℝ E)}
(hpolar : Cpolar = polar (E := E) C) (hCpolar0 : (0 : Module.Dual ℝ E) ∈ Cpolar) :
Set.recessionCone Cpolar = {ψ : Module.Dual ℝ E | ∀ x ∈ C, ψ x ≤ 0}
If Cpolar = polar C and 0 ∈ Cpolar, then the two-sided recession directions in Cpolar are exactly the functionals vanishing on C.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_lineality_Cpolar_eq_forall_primal_eq_zero_of_hpolar
{C : Set E} {Cpolar : Set (Module.Dual ℝ E)}
(hpolar : Cpolar = polar (E := E) C) (hCpolar0 : (0 : Module.Dual ℝ E) ∈ Cpolar) :
(-Set.recessionCone Cpolar) ∩ Set.recessionCone Cpolar =
{ψ : Module.Dual ℝ E | ∀ x ∈ C, ψ x = 0}§14 — Polars of Convex Sets — Part 11
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yuhao Jiang, Zaiwen Wen.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
noncomputable local instance : TopologicalSpace (Module.Dual ℝ E)
If Cpolar = polar C and 0 ∈ Cpolar, then the span of C is the orthogonal complement of the lineality subspace of Cpolar (expressed via dual annihilators).
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_dualAnnihilator_span_primal_eq_span_lineality_Cpolar_of_hpolar
{C : Set E} {Cpolar : Set (Module.Dual ℝ E)}
(hpolar : Cpolar = polar (E := E) C) (hCpolar0 : (0 : Module.Dual ℝ E) ∈ Cpolar) :
(Submodule.span ℝ C).dualAnnihilator =
Submodule.span ℝ ((-Set.recessionCone Cpolar) ∩ Set.recessionCone Cpolar)
Theorem 14.6: Let C and C^∘ be a polar pair of closed convex sets containing the origin. Then the recession cone of C and the closure of the convex cone generated by C^∘ are polar to each other. The lineality space of C and the subspace generated by C^∘ are orthogonally complementary. Dually, the same statements hold with C and C^∘ interchanged.
open scoped Pointwise
open scoped Topology
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
theorem recessionCone_polar_eq_closure_coneHull_and_lineality_orthogonal_of_polarPair
[TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E] [T2Space E]
[FiniteDimensional ℝ E] [LocallyConvexSpace ℝ E]
{C : Set E} {Cpolar : Set (Module.Dual ℝ E)}
(_hCclosed : IsClosed C) (_hCconv : Convex ℝ C) (hC0 : (0 : E) ∈ C)
(_hCpolar_closed : IsClosed Cpolar) (_hCpolar_conv : Convex ℝ Cpolar)
(hCpolar0 : (0 : Module.Dual ℝ E) ∈ Cpolar)
(hpolar : Cpolar = polar (E := E) C) (hbipolar : {x : E | ∀ φ ∈ Cpolar, φ x ≤ 1} = C) :
polarCone (E := E) (Set.recessionCone C) =
closure ((ConvexCone.hull ℝ Cpolar : ConvexCone ℝ (Module.Dual ℝ E)) : Set _) ∧
(PointedCone.dual (R := ℝ)
((-LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ))).flip
(closure ((ConvexCone.hull ℝ Cpolar : ConvexCone ℝ (Module.Dual ℝ E)) : Set _)) :
Set E) =
Set.recessionCone C ∧
(Submodule.span ℝ ((-Set.recessionCone C) ∩ Set.recessionCone C)).dualAnnihilator =
Submodule.span ℝ Cpolar ∧
(Submodule.span ℝ Cpolar).dualCoannihilator =
Submodule.span ℝ ((-Set.recessionCone C) ∩ Set.recessionCone C) ∧
(PointedCone.dual (R := ℝ)
((-LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ))).flip
(Set.recessionCone Cpolar) :
Set E) =
closure ((ConvexCone.hull ℝ C : ConvexCone ℝ E) : Set _) ∧
polarCone (E := E) (closure ((ConvexCone.hull ℝ C : ConvexCone ℝ E) : Set _)) =
Set.recessionCone Cpolar ∧
(Submodule.span ℝ C).dualAnnihilator =
Submodule.span ℝ ((-Set.recessionCone Cpolar) ∩ Set.recessionCone Cpolar) ∧
(Submodule.span ℝ ((-Set.recessionCone Cpolar) ∩ Set.recessionCone Cpolar)).dualCoannihilator
=
Submodule.span ℝ C§14 — Polars of Convex Sets — Part 12
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yuhao Jiang, Zaiwen Wen.
Finite-dimensional formula: dim(Wᵃⁿⁿ) = dim(E) - dim(W) for dual annihilators.
open scoped Pointwise
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_finrank_dualAnnihilator_eq_sub [FiniteDimensional ℝ E] (W : Submodule ℝ E) :
Module.finrank ℝ (W.dualAnnihilator) = Module.finrank ℝ E - Module.finrank ℝ W
Orthogonality relation: the span of polar C is the dual annihilator of the lineality subspace of C (spanned by (-rec C) ∩ rec C).
open scoped Pointwise
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_span_polar_eq_dualAnnihilator_span_lineality
[FiniteDimensional ℝ E] [TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E]
[LocallyConvexSpace ℝ E] {C : Set E} (hCclosed : IsClosed C) (hCconv : Convex ℝ C)
(hC0 : (0 : E) ∈ C) :
Submodule.span ℝ (polar (E := E) C) =
(Submodule.span ℝ ((-Set.recessionCone C) ∩ Set.recessionCone C)).dualAnnihilator
Orthogonality relation (dual direction): the span of the lineality subspace of polar C is the dual annihilator of span C.
open scoped Pointwise
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_span_lineality_polar_eq_dualAnnihilator_span {C : Set E} :
Submodule.span ℝ
((-Set.recessionCone (polar (E := E) C)) ∩ Set.recessionCone (polar (E := E) C)) =
(Submodule.span ℝ C).dualAnnihilator
Corollary 14.6.1. Let C be a closed convex set in ℝ^n containing the origin. Then dimension C^∘ = n - lineality C, lineality C^∘ = n - dimension C, and rank C^∘ = rank C. In this file, we model C^∘ as polar C : Set (Module.Dual ℝ E). We interpret the book's dimension S as Module.finrank ℝ (Submodule.span ℝ S) and the book's lineality S as Module.finrank ℝ (Submodule.span ℝ ((-Set.recessionCone S) ∩ Set.recessionCone S)), so that rank S = dimension S - lineality S.
open scoped Pointwise
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
theorem finrank_span_polar_eq_finrank_sub_finrank_span_lineality_and_dual_and_rank
[FiniteDimensional ℝ E] [TopologicalSpace E] [IsTopologicalAddGroup E] [ContinuousSMul ℝ E]
[LocallyConvexSpace ℝ E] {C : Set E} (hCclosed : IsClosed C) (hCconv : Convex ℝ C)
(hC0 : (0 : E) ∈ C) :
(Module.finrank ℝ (Submodule.span ℝ (polar (E := E) C)) =
Module.finrank ℝ E -
Module.finrank ℝ (Submodule.span ℝ ((-Set.recessionCone C) ∩ Set.recessionCone C))) ∧
(Module.finrank ℝ
(Submodule.span ℝ
((-Set.recessionCone (polar (E := E) C)) ∩ Set.recessionCone (polar (E := E) C))) =
Module.finrank ℝ E - Module.finrank ℝ (Submodule.span ℝ C)) ∧
(Module.finrank ℝ (Submodule.span ℝ (polar (E := E) C)) -
Module.finrank ℝ
(Submodule.span ℝ
((-Set.recessionCone (polar (E := E) C)) ∩
Set.recessionCone (polar (E := E) C))) =
Module.finrank ℝ (Submodule.span ℝ C) -
Module.finrank ℝ
(Submodule.span ℝ ((-Set.recessionCone C) ∩ Set.recessionCone C)))
If f is nonnegative and vanishes at 0, then its Fenchel conjugate (for evaluation) is also nonnegative and vanishes at 0.
open scoped Pointwise
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_fenchelConjugate_nonneg_of_nonneg_and_zero
{f : E → EReal} (hf_nonneg : ∀ x, (0 : EReal) ≤ f x) (hf0 : f 0 = 0) :
(∀ xStar : Module.Dual ℝ E,
(0 : EReal) ≤
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f xStar) ∧
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f 0 = 0
Fenchel–Young inequality for the evaluation pairing: ⟪x★, x⟫ ≤ f x + f★ x★.
open scoped Pointwise
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_eval_le_add_fenchelConjugate
{f : E → EReal} (x : E) (xStar : Module.Dual ℝ E) (hfbot : f x ≠ ⊥) (hftop : f x ≠ ⊤) :
((xStar x : ℝ) : EReal) ≤
f x +
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f xStar
The α⁻¹-scaled α-sublevel set of the conjugate lies in 2 • polar of the primal α-sublevel set.
open scoped Pointwise
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_inv_smul_sublevel_fenchelConjugate_subset_two_smul_polar_sublevel
{f : E → EReal} (hf_nonneg : ∀ x, (0 : EReal) ≤ f x) {α : ℝ} (hα : 0 < α) :
(α⁻¹ : ℝ) •
{xStar : Module.Dual ℝ E |
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f xStar ≤
(α : EReal)} ⊆
(2 : ℝ) • polar (E := E) {x : E | f x ≤ (α : EReal)}
The polar of the primal α-sublevel set is contained in the α⁻¹-scaled α-sublevel set of the conjugate, for a convex function vanishing at 0.
open scoped Pointwise
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
lemma section14_polar_sublevel_subset_inv_smul_sublevel_fenchelConjugate
{f : E → EReal} (hf_nonneg : ∀ x, (0 : EReal) ≤ f x) (hf_conv : ConvexERealFunction (F := E) f)
(hf0 : f 0 = 0) {α : ℝ} (hα : 0 < α) :
polar (E := E) {x : E | f x ≤ (α : EReal)} ⊆
(α⁻¹ : ℝ) •
{xStar : Module.Dual ℝ E |
fenchelConjugateBilin (LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)) f xStar ≤
(α : EReal)}
Theorem 14.7. Let f be a non-negative closed convex function which vanishes at the origin. Then the Fenchel–Legendre conjugate f* (with respect to the evaluation pairing) is also non-negative and vanishes at the origin. Moreover, for 0 < α < ∞ one has {x | f x ≤ α}^∘ = α⁻¹ • {x★ | f* x★ ≤ α} ⊆ 2 • {x | f x ≤ α}^∘, where ^∘ denotes the polar of a set (modeled here as polar).
open scoped Pointwise
variable {E : Type*} [AddCommGroup E] [Module ℝ E]
theorem polar_sublevel_eq_inv_smul_sublevel_fenchelConjugate_and_subset_two_smul
[TopologicalSpace E] {f : E → EReal} (hf_nonneg : ∀ x, (0 : EReal) ≤ f x)
(hf_conv : ConvexERealFunction (F := E) f)
(hf0 : f 0 = 0) {α : ℝ} (hα : 0 < α) :
let p := LinearMap.applyₗ (R := ℝ) (M := E) (M₂ := ℝ)
let fstar := fenchelConjugateBilin p f
(∀ xStar : Module.Dual ℝ E, (0 : EReal) ≤ fstar xStar) ∧
fstar 0 = 0 ∧
polar (E := E) {x : E | f x ≤ (α : EReal)} ⊆
(α⁻¹ : ℝ) • {xStar : Module.Dual ℝ E | fstar xStar ≤ (α : EReal)} ∧
(α⁻¹ : ℝ) • {xStar : Module.Dual ℝ E | fstar xStar ≤ (α : EReal)} ⊆
(2 : ℝ) • polar (E := E) {x : E | f x ≤ (α : EReal)}§15 — Polars of Convex Functions — Part 1
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
Text 15.0.1: A function k : ℝⁿ → (-∞, +∞] is a gauge if it is convex, nonnegative, positively homogeneous of degree 1 (i.e. k (λ x) = λ k x for all λ ≥ 0), and satisfies k 0 = 0. Equivalently, its epigraph epi k = {(x, μ) | k x ≤ μ} is a convex cone in ℝⁿ⁺¹ containing (0, 0) and containing no vector (x, μ) with μ < 0. In this formalization, we use Fin n → ℝ for ℝⁿ, EReal for (-∞, +∞], and the epigraph construction epigraph (S := Set.univ) k from Books.ConvexAnalysis_Rockafellar_1970.Chap01.section04_part1.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
def IsGauge {n : ℕ} (k : (Fin n → ℝ) → EReal) : Prop :=
ConvexFunctionOn (S := (Set.univ : Set (Fin n → ℝ))) k ∧
(∀ x, 0 ≤ k x) ∧
(∀ x (r : ℝ), 0 ≤ r → k (r • x) = (r : EReal) * k x) ∧
k 0 = 0
A gauge never attains ⊥ because it is nonnegative.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma IsGauge.ne_bot {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsGauge k) (x : Fin n → ℝ) :
k x ≠ (⊥ : EReal)
A gauge is subadditive.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma IsGauge.add_le {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsGauge k)
(x y : Fin n → ℝ) : k (x + y) ≤ k x + k y
The unit sublevel set of a gauge is nonempty (it contains 0).
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma IsGauge.sublevel_one_nonempty {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsGauge k) :
({x : Fin n → ℝ | k x ≤ (1 : EReal)}).Nonempty
The unit sublevel set of a gauge is convex.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma IsGauge.convex_sublevel_one {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsGauge k) :
Convex ℝ {x : Fin n → ℝ | k x ≤ (1 : EReal)}
The gauge of the unit sublevel set of a finite-valued gauge agrees with toReal.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma IsGauge.gaugeFunction_sublevel_one_eq_toReal {n : ℕ} {k : (Fin n → ℝ) → EReal}
(hk : IsGauge k) (hk_finite : ∀ x, k x ≠ ⊤) :
∀ x, gaugeFunction {x : Fin n → ℝ | k x ≤ (1 : EReal)} x = (k x).toReal
The gauge of the unit sublevel set of a finite-valued gauge agrees with k.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma IsGauge.gaugeFunction_sublevel_one_eq {n : ℕ} {k : (Fin n → ℝ) → EReal}
(hk : IsGauge k) (hk_finite : ∀ x, k x ≠ ⊤) :
(fun x => (gaugeFunction {x : Fin n → ℝ | k x ≤ (1 : EReal)} x : EReal)) = k
Text 15.0.2: Every gauge k can be represented as k = γ(· | C) for some nonempty convex set C. In this development, γ(· | C) is represented by fun x => (gaugeFunction C x : EReal).
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma IsGauge.exists_set_eq_gaugeFunction {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsGauge k)
(hk_finite : ∀ x, k x ≠ ⊤) :
∃ C : Set (Fin n → ℝ), C.Nonempty ∧ Convex ℝ C ∧ k = fun x => (gaugeFunction C x : EReal)
Text 15.0.3: For a nonempty convex set C = {x ∈ ℝⁿ | k x ≤ 1}, one has γ(· | C) = k. In this development, γ(· | C) is represented by fun x => (gaugeFunction C x : EReal).
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma gaugeFunction_eq_of_convex_nonempty_eq_sublevel {n : ℕ} {k : (Fin n → ℝ) → EReal}
{C : Set (Fin n → ℝ)} (hC : C = {x | k x ≤ (1 : EReal)}) (hk : IsGauge k)
(hk_finite : ∀ x, k x ≠ ⊤) :
(fun x => (gaugeFunction C x : EReal)) = k
A closed convex absorbing set is the unit sublevel set of its gauge function.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma set_eq_gaugeFunction_sublevel_one {n : ℕ} {D : Set (Fin n → ℝ)}
(hDclosed : IsClosed D) (hDconv : Convex ℝ D) (hD0 : (0 : Fin n → ℝ) ∈ D)
(hDabs :
∀ x : Fin n → ℝ, ∃ lam : ℝ, 0 ≤ lam ∧ x ∈ (fun y => lam • y) '' D) :
D = {x | (gaugeFunction D x : EReal) ≤ (1 : EReal)}
If a set's gauge function agrees with k, then the set is the k ≤ 1 sublevel.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma eq_sublevel_one_of_gaugeFunction_eq {n : ℕ} {k : (Fin n → ℝ) → EReal} {D : Set (Fin n → ℝ)}
(hDclosed : IsClosed D) (hDconv : Convex ℝ D) (hD0 : (0 : Fin n → ℝ) ∈ D)
(hDabs :
∀ x : Fin n → ℝ, ∃ lam : ℝ, 0 ≤ lam ∧ x ∈ (fun y => lam • y) '' D)
(hg : (fun x => (gaugeFunction D x : EReal)) = k) :
D = {x | k x ≤ (1 : EReal)}
Text 15.0.4: If k is closed (and finite, positive away from 0), then the set C = {x ∈ ℝⁿ | k x ≤ 1} is the unique closed convex set containing 0 such that γ(· | C) = k. In this development, we express "closedness of k" as closedness of the book epigraph epigraph univ k, and γ(· | C) is represented by fun x => (gaugeFunction C x : EReal).
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma sublevel_one_unique_closed_convex_of_isClosed_epigraph {n : ℕ} {k : (Fin n → ℝ) → EReal}
(hk : IsGauge k)
(hk_finite : ∀ x, k x ≠ ⊤)
(hk_pos : ∀ x, x ≠ 0 → (0 : EReal) < k x)
(hk_closed : IsClosed (epigraph (S := (Set.univ : Set (Fin n → ℝ))) k)) :
let C : Set (Fin n → ℝ) := {x | k x ≤ (1 : EReal)}
IsClosed C ∧
Convex ℝ C ∧
(0 : Fin n → ℝ) ∈ C ∧
(fun x => (gaugeFunction C x : EReal)) = k ∧
∀ D : Set (Fin n → ℝ),
IsClosed D →
Convex ℝ D →
(0 : Fin n → ℝ) ∈ D →
(fun x => (gaugeFunction D x : EReal)) = k → D = C
Text 15.0.5: Let k be a gauge on ℝⁿ. Its polar kᵒ : ℝⁿ → [0, +∞] is defined by kᵒ x⋆ = inf { μ⋆ ≥ 0 | ⟪x, x⋆⟫ ≤ μ⋆ * k x for all x }. In this development, ℝⁿ is Fin n → ℝ and we represent the inf as sInf in EReal.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
noncomputable def polarGauge {n : ℕ} (k : (Fin n → ℝ) → EReal) (xStar : Fin n → ℝ) : EReal :=
sInf {μStar : EReal | 0 ≤ μStar ∧ ∀ x : Fin n → ℝ, ((x ⬝ᵥ xStar : ℝ) : EReal) ≤ μStar * k x}
The polar gauge is nonnegative.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_nonneg {n : ℕ} {k : (Fin n → ℝ) → EReal} (xStar : Fin n → ℝ) :
(0 : EReal) ≤ polarGauge k xStar
The polar gauge of any function vanishes at 0.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_zero {n : ℕ} (k : (Fin n → ℝ) → EReal) : polarGauge k (0 : Fin n → ℝ) = 0
The polar gauge never attains ⊥.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_ne_bot {n : ℕ} {k : (Fin n → ℝ) → EReal} (xStar : Fin n → ℝ) :
polarGauge k xStar ≠ (⊥ : EReal)
Scaling a feasible multiplier remains feasible for the scaled dual vector.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_feasible_smul {n : ℕ} {k : (Fin n → ℝ) → EReal} {xStar : Fin n → ℝ}
{μ : EReal} (hμ : 0 ≤ μ ∧ ∀ x : Fin n → ℝ, ((x ⬝ᵥ xStar : ℝ) : EReal) ≤ μ * k x)
{t : ℝ} (ht : 0 ≤ t) :
0 ≤ ((t : ℝ) : EReal) * μ ∧
∀ x : Fin n → ℝ,
((x ⬝ᵥ (t • xStar) : ℝ) : EReal) ≤ (((t : ℝ) : EReal) * μ) * k x
Adding feasible multipliers is feasible for the sum of dual vectors.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_feasible_add {n : ℕ} {k : (Fin n → ℝ) → EReal} {xStar₁ xStar₂ : Fin n → ℝ}
{μ₁ μ₂ : EReal}
(hμ₁ : 0 ≤ μ₁ ∧ ∀ x : Fin n → ℝ, ((x ⬝ᵥ xStar₁ : ℝ) : EReal) ≤ μ₁ * k x)
(hμ₂ : 0 ≤ μ₂ ∧ ∀ x : Fin n → ℝ, ((x ⬝ᵥ xStar₂ : ℝ) : EReal) ≤ μ₂ * k x) :
0 ≤ μ₁ + μ₂ ∧
∀ x : Fin n → ℝ,
((x ⬝ᵥ (xStar₁ + xStar₂) : ℝ) : EReal) ≤ (μ₁ + μ₂) * k x
A feasible sum bounds the polar gauge of the summed dual vector.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_le_of_feasible_add {n : ℕ} {k : (Fin n → ℝ) → EReal}
{xStar₁ xStar₂ : Fin n → ℝ} {μ₁ μ₂ : EReal}
(hμ₁ : 0 ≤ μ₁ ∧ ∀ x : Fin n → ℝ, ((x ⬝ᵥ xStar₁ : ℝ) : EReal) ≤ μ₁ * k x)
(hμ₂ : 0 ≤ μ₂ ∧ ∀ x : Fin n → ℝ, ((x ⬝ᵥ xStar₂ : ℝ) : EReal) ≤ μ₂ * k x) :
polarGauge k (xStar₁ + xStar₂) ≤ μ₁ + μ₂
A feasible scaling bounds the polar gauge of the scaled dual vector.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_le_of_feasible_smul {n : ℕ} {k : (Fin n → ℝ) → EReal} {xStar : Fin n → ℝ}
{μ : EReal} (hμ : 0 ≤ μ ∧ ∀ x : Fin n → ℝ, ((x ⬝ᵥ xStar : ℝ) : EReal) ≤ μ * k x)
{t : ℝ} (ht : 0 ≤ t) :
polarGauge k (t • xStar) ≤ ((t : ℝ) : EReal) * μ
The polar gauge is positively homogeneous.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_posHom {n : ℕ} {k : (Fin n → ℝ) → EReal} :
PositivelyHomogeneous (polarGauge k)
The polar gauge is subadditive.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_add_le {n : ℕ} {k : (Fin n → ℝ) → EReal}
(xStar₁ xStar₂ : Fin n → ℝ) :
polarGauge k (xStar₁ + xStar₂) ≤ polarGauge k xStar₁ + polarGauge k xStar₂
The polar gauge is a gauge.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_isGauge {n : ℕ} (k : (Fin n → ℝ) → EReal) : IsGauge (polarGauge k)
A feasible μ for the polar gauge bounds each ratio ⟪x, x⋆⟫ / k x when x ≠ 0.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_ratio_le_of_feasible {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsGauge k)
(hk_finite : ∀ x, k x ≠ ⊤) (hk_pos : ∀ x, x ≠ 0 → (0 : EReal) < k x)
{xStar : Fin n → ℝ} {μ : EReal} (hμ0 : 0 ≤ μ)
(hμ : ∀ x, ((x ⬝ᵥ xStar : ℝ) : EReal) ≤ μ * k x) {x : Fin n → ℝ} (hx : x ≠ 0) :
(((x ⬝ᵥ xStar : ℝ) / (k x).toReal : ℝ) : EReal) ≤ μ
A bound on the ratio yields the product bound when the coefficient is finite.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_mul_le_of_ratio_le {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsGauge k)
(hk_finite : ∀ x, k x ≠ ⊤) (hk_pos : ∀ x, x ≠ 0 → (0 : EReal) < k x)
{xStar : Fin n → ℝ} {μ : EReal} (hμ_top : μ ≠ ⊤) (hμ_ne_bot : μ ≠ ⊥)
{x : Fin n → ℝ} (hx : x ≠ 0)
(hμ : (((x ⬝ᵥ xStar : ℝ) / (k x).toReal : ℝ) : EReal) ≤ μ) :
((x ⬝ᵥ xStar : ℝ) : EReal) ≤ μ * k x
Text 15.0.6: If a gauge k is finite everywhere and positive except at the origin, then kᵒ admits the equivalent formula kᵒ x⋆ = sup_{x ≠ 0} ⟪x, x⋆⟫ / k x. In this development, kᵒ is polarGauge k, and we express sup as a sSup over the image of x ↦ ⟪x, x⋆⟫ / (k x) (using EReal.toReal to form a real quotient).
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_eq_sSup_div {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsGauge k)
(hk_finite : ∀ x, k x ≠ ⊤) (hk_pos : ∀ x, x ≠ 0 → (0 : EReal) < k x)
(h_nonempty : ∃ x : Fin n → ℝ, x ≠ 0) (xStar : Fin n → ℝ) :
polarGauge k xStar =
sSup
((fun x : Fin n → ℝ =>
(((x ⬝ᵥ xStar : ℝ) / (k x).toReal : ℝ) : EReal)) '' {x | x ≠ 0})§15 — Polars of Convex Functions — Part 2
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
Outside the polar cone, the polar gauge of an indicator is ⊤.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_erealIndicator_eq_top_of_not_mem_Kpolar {n : ℕ}
(K : ConvexCone ℝ (Fin n → ℝ)) {xStar : Fin n → ℝ}
(hxStar : ¬ ∀ x ∈ (K : Set (Fin n → ℝ)), (x ⬝ᵥ xStar : ℝ) ≤ 0) :
polarGauge (erealIndicator (K : Set (Fin n → ℝ))) xStar = ⊤
On the polar cone, the polar gauge of an indicator is 0.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_erealIndicator_eq_zero_of_mem_Kpolar {n : ℕ}
(K : ConvexCone ℝ (Fin n → ℝ)) {xStar : Fin n → ℝ}
(hxStar : ∀ x ∈ (K : Set (Fin n → ℝ)), (x ⬝ᵥ xStar : ℝ) ≤ 0) :
polarGauge (erealIndicator (K : Set (Fin n → ℝ))) xStar = 0
On the polar cone, the conjugate of the indicator is 0.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma kStar_eq_zero_of_mem_Kpolar {n : ℕ} (K : ConvexCone ℝ (Fin n → ℝ)) {xStar : Fin n → ℝ}
(h0K : (0 : Fin n → ℝ) ∈ (K : Set (Fin n → ℝ)))
(hxStar : ∀ x ∈ (K : Set (Fin n → ℝ)), (x ⬝ᵥ xStar : ℝ) ≤ 0) :
sSup
(Set.range fun x : Fin n → ℝ =>
((x ⬝ᵥ xStar : ℝ) : EReal) - erealIndicator (K : Set (Fin n → ℝ)) x) = 0
Outside the polar cone, the conjugate of the indicator is ⊤.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma kStar_eq_top_of_not_mem_Kpolar {n : ℕ} (K : ConvexCone ℝ (Fin n → ℝ)) {xStar : Fin n → ℝ}
(hxStar : ¬ ∀ x ∈ (K : Set (Fin n → ℝ)), (x ⬝ᵥ xStar : ℝ) ≤ 0) :
sSup
(Set.range fun x : Fin n → ℝ =>
((x ⬝ᵥ xStar : ℝ) : EReal) - erealIndicator (K : Set (Fin n → ℝ)) x) = ⊤
Text 15.0.7: Let K ⊆ ℝⁿ be a convex cone and let k = δ(· | K) be its indicator function. Then the polar kᵒ agrees with the conjugate k*, and it is the indicator of the polar cone Kᵒ := {x⋆ ∈ ℝⁿ | ⟪x, x⋆⟫ ≤ 0 for all x ∈ K}. In this development, ℝⁿ is Fin n → ℝ, the indicator δ(· | K) is erealIndicator K, the polar kᵒ is polarGauge k, the conjugate k* is given by x⋆ ↦ sup_x (⟪x, x⋆⟫ - k x), and ⟪x, x⋆⟫ is x ⬝ᵥ x⋆.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
theorem polarGauge_erealIndicator_eq_conjugate_eq_erealIndicator_polarCone {n : ℕ}
(K : ConvexCone ℝ (Fin n → ℝ)) (h0K : (0 : Fin n → ℝ) ∈ (K : Set (Fin n → ℝ))) :
let k : (Fin n → ℝ) → EReal := erealIndicator (K : Set (Fin n → ℝ))
let Kpolar : Set (Fin n → ℝ) :=
{xStar | ∀ x ∈ (K : Set (Fin n → ℝ)), (x ⬝ᵥ xStar : ℝ) ≤ 0}
let kStar : (Fin n → ℝ) → EReal :=
fun xStar => sSup (Set.range fun x : Fin n → ℝ => ((x ⬝ᵥ xStar : ℝ) : EReal) - k x)
polarGauge k = kStar ∧ polarGauge k = erealIndicator Kpolar
Text 15.0.8: Define k : ℝ² → [0, +∞) by k(ξ₁, ξ₂) = sqrt (ξ₁^2 + ξ₂^2) + ξ₁. Then k is a closed gauge function. Its polar gauge k^∘ is given by the piecewise formula k^∘(ξ₁⋆, ξ₂⋆) = (1/2) * ((ξ₂⋆)^2 / ξ₁⋆ + ξ₁⋆) if 0 < ξ₁⋆, k^∘(0, 0) = 0, and k^∘(ξ₁⋆, ξ₂⋆) = +∞ otherwise. Neither k nor k^∘ is a norm (e.g. neither is an even function).
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
noncomputable def gaugeSqrtSumSq_add_fst (x : Fin 2 → ℝ) : EReal :=
((Real.sqrt ((x 0) ^ 2 + (x 1) ^ 2) + x 0) : ℝ)
Rewrite gaugeSqrtSumSq_add_fst using the Euclidean norm on ℝ².
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma gaugeSqrtSumSq_add_fst_eq_euclideanNorm_add_fst (x : Fin 2 → ℝ) :
gaugeSqrtSumSq_add_fst x =
((‖(EuclideanSpace.equiv (Fin 2) ℝ).symm x‖ + x 0 : ℝ) : EReal)
Convexity of gaugeSqrtSumSq_add_fst on Set.univ.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma gaugeSqrtSumSq_add_fst_convexFunctionOn :
ConvexFunctionOn (S := (Set.univ : Set (Fin 2 → ℝ))) gaugeSqrtSumSq_add_fst
Nonnegativity of gaugeSqrtSumSq_add_fst.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma gaugeSqrtSumSq_add_fst_nonneg (x : Fin 2 → ℝ) :
(0 : EReal) ≤ gaugeSqrtSumSq_add_fst x
Positive homogeneity of gaugeSqrtSumSq_add_fst.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma gaugeSqrtSumSq_add_fst_smul (x : Fin 2 → ℝ) (r : ℝ) (hr : 0 ≤ r) :
gaugeSqrtSumSq_add_fst (r • x) = (r : EReal) * gaugeSqrtSumSq_add_fst x
The epigraph of gaugeSqrtSumSq_add_fst is closed.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma isClosed_epigraph_gaugeSqrtSumSq_add_fst :
IsClosed (epigraph (S := (Set.univ : Set (Fin 2 → ℝ))) gaugeSqrtSumSq_add_fst)
Text 15.0.8 (closed gauge): the function gaugeSqrtSumSq_add_fst is a gauge and its epigraph is closed.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
theorem gaugeSqrtSumSq_add_fst_isGauge_isClosed :
IsGauge gaugeSqrtSumSq_add_fst ∧
IsClosed (epigraph (S := (Set.univ : Set (Fin 2 → ℝ))) gaugeSqrtSumSq_add_fst)
Text 15.0.8 (polar formula): explicit piecewise formula for the polar gauge of gaugeSqrtSumSq_add_fst.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
noncomputable def polarGaugeSqrtSumSq_add_fst (xStar : Fin 2 → ℝ) : EReal :=
if 0 < xStar 0 then
((1 / 2 : ℝ) * (((xStar 1) ^ 2) / (xStar 0) + xStar 0) : ℝ)
else if xStar 0 = 0 ∧ xStar 1 = 0 then
(0 : EReal)
else
⊤
Bound sqrt(t^2 + b^2) - t by b^2 / (2t) for t > 0.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma sqrt_sq_add_sub_le (t b : ℝ) (ht : 0 < t) :
Real.sqrt (t ^ 2 + b ^ 2) - t ≤ b ^ 2 / (2 * t)
Cauchy–Schwarz inequality for Fin 2 → ℝ in dot-product form.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma dotProduct_le_mul_sqrt_dotProduct_self_fin2 (x y : Fin 2 → ℝ) :
(x ⬝ᵥ y : ℝ) ≤
Real.sqrt ((x 0) ^ 2 + (x 1) ^ 2) * Real.sqrt ((y 0) ^ 2 + (y 1) ^ 2)
If xStar 0 < 0, the polar gauge of gaugeSqrtSumSq_add_fst is ⊤.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_gaugeSqrtSumSq_add_fst_eq_top_of_fst_lt_zero (xStar : Fin 2 → ℝ)
(hx : xStar 0 < 0) : polarGauge gaugeSqrtSumSq_add_fst xStar = ⊤
If xStar 0 = 0 and xStar 1 ≠ 0, the polar gauge is ⊤.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_gaugeSqrtSumSq_add_fst_eq_top_of_fst_eq_zero_snd_ne_zero
(xStar : Fin 2 → ℝ) (hx0 : xStar 0 = 0) (hx1 : xStar 1 ≠ 0) :
polarGauge gaugeSqrtSumSq_add_fst xStar = ⊤
Feasible multiplier for the polar gauge when xStar 0 > 0.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_gaugeSqrtSumSq_add_fst_mu0_feasible_of_fst_pos (xStar : Fin 2 → ℝ)
(hpos : 0 < xStar 0) :
let μ0 : ℝ := (1 / 2) * ((xStar 1) ^ 2 / (xStar 0) + xStar 0)
(0 : EReal) ≤ (μ0 : EReal) ∧
∀ x : Fin 2 → ℝ,
((x ⬝ᵥ xStar : ℝ) : EReal) ≤ (μ0 : EReal) * gaugeSqrtSumSq_add_fst x
A point that attains the polar inequality when xStar 0 > 0.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_gaugeSqrtSumSq_add_fst_witness_of_fst_pos (xStar : Fin 2 → ℝ)
(hpos : 0 < xStar 0) :
let t : ℝ := xStar 1 / xStar 0
let x : Fin 2 → ℝ := ![(1 - t ^ 2) / 2, t]
gaugeSqrtSumSq_add_fst x = (1 : EReal) ∧
(x ⬝ᵥ xStar : ℝ) = (1 / 2) * ((xStar 1) ^ 2 / (xStar 0) + xStar 0)
Compute the polar gauge value for xStar 0 > 0.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_gaugeSqrtSumSq_add_fst_eq_mu0_of_fst_pos (xStar : Fin 2 → ℝ)
(hpos : 0 < xStar 0) :
polarGauge gaugeSqrtSumSq_add_fst xStar =
((1 / 2) * ((xStar 1) ^ 2 / (xStar 0) + xStar 0) : ℝ)
Text 15.0.8 (polar gauge): the polar gauge of gaugeSqrtSumSq_add_fst agrees with the explicit formula polarGaugeSqrtSumSq_add_fst.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
theorem polarGauge_gaugeSqrtSumSq_add_fst :
polarGauge gaugeSqrtSumSq_add_fst = polarGaugeSqrtSumSq_add_fst
Evaluate gaugeSqrtSumSq_add_fst on ![1,0] and its negation.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma gaugeSqrtSumSq_add_fst_eval_vec10 :
gaugeSqrtSumSq_add_fst (![1, 0]) = (2 : EReal) ∧
gaugeSqrtSumSq_add_fst (-![1, 0]) = (0 : EReal)
The witness ![1,0] shows gaugeSqrtSumSq_add_fst is not even.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma gaugeSqrtSumSq_add_fst_not_even_via_vec10 :
¬ (∀ x : Fin 2 → ℝ, gaugeSqrtSumSq_add_fst (-x) = gaugeSqrtSumSq_add_fst x)
Text 15.0.8 (not a norm): gaugeSqrtSumSq_add_fst is not even.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
theorem gaugeSqrtSumSq_add_fst_not_even :
¬ (∀ x : Fin 2 → ℝ, gaugeSqrtSumSq_add_fst (-x) = gaugeSqrtSumSq_add_fst x)
Evaluate polarGaugeSqrtSumSq_add_fst on ![1,0] and its negation.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGaugeSqrtSumSq_add_fst_eval_vec10 :
polarGaugeSqrtSumSq_add_fst (![1, 0]) = ((1 / 2 : ℝ) : EReal) ∧
polarGaugeSqrtSumSq_add_fst (-![1, 0]) = (⊤ : EReal)
The witness ![1,0] shows polarGaugeSqrtSumSq_add_fst is not even.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGaugeSqrtSumSq_add_fst_not_even_via_vec10 :
¬ (∀ xStar : Fin 2 → ℝ,
polarGaugeSqrtSumSq_add_fst (-xStar) = polarGaugeSqrtSumSq_add_fst xStar)
Text 15.0.8 (not a norm): the polar gauge polarGaugeSqrtSumSq_add_fst is not even.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
theorem polarGaugeSqrtSumSq_add_fst_not_even :
¬ (∀ xStar : Fin 2 → ℝ,
polarGaugeSqrtSumSq_add_fst (-xStar) = polarGaugeSqrtSumSq_add_fst xStar)
Approximate the polar gauge inequality in ℝ with an ε-margin.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma inner_le_mul_polarGauge_real_eps {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsGauge k)
(x xStar : Fin n → ℝ) (hx : k x ≠ ⊤) (hxStar : polarGauge k xStar ≠ ⊤) :
∀ ε > 0, (x ⬝ᵥ xStar : ℝ) ≤ ((polarGauge k xStar).toReal + ε) * (k x).toReal
Letting ε→0 in the real inequality gives the sharp product bound.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma inner_le_mul_polarGauge_real {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsGauge k)
(x xStar : Fin n → ℝ) (hx : k x ≠ ⊤) (hxStar : polarGauge k xStar ≠ ⊤) :
(x ⬝ᵥ xStar : ℝ) ≤ (k x).toReal * (polarGauge k xStar).toReal
Cast the real inequality back to EReal.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma inner_le_mul_polarGauge_via_toReal {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsGauge k)
(x xStar : Fin n → ℝ) (hx : k x ≠ ⊤) (hxStar : polarGauge k xStar ≠ ⊤) :
((x ⬝ᵥ xStar : ℝ) : EReal) ≤ k x * polarGauge k xStar
Text 15.0.9: Let k be a gauge and let kᵒ be its polar gauge. Then for every x in the effective domain of k and every x⋆ in the effective domain of kᵒ, ⟪x, x⋆⟫ ≤ k x * kᵒ x⋆. In this development, we represent kᵒ by polarGauge k, the pairing ⟪x, x⋆⟫ by x ⬝ᵥ x⋆, and we model domain assumptions as k x ≠ ⊤ and polarGauge k x⋆ ≠ ⊤. If k is a norm (i.e. k (-x) = k x and k is finite and positive away from 0), then |⟪x, x⋆⟫| ≤ k x * kᵒ x⋆ for all x, x⋆.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
theorem inner_le_mul_polarGauge {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsGauge k)
(x xStar : Fin n → ℝ) (hx : k x ≠ ⊤) (hxStar : polarGauge k xStar ≠ ⊤) :
((x ⬝ᵥ xStar : ℝ) : EReal) ≤ k x * polarGauge k xStar
Absolute-value form of the polar gauge inequality under norm-like hypotheses.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma abs_inner_le_mul_polarGauge_of_norm_like_casework {n : ℕ} {k : (Fin n → ℝ) → EReal}
(hk : IsGauge k) (hk_even : ∀ x, k (-x) = k x) (hk_finite : ∀ x, k x ≠ ⊤)
(hk_pos : ∀ x, x ≠ 0 → (0 : EReal) < k x) (x xStar : Fin n → ℝ) :
((|x ⬝ᵥ xStar| : ℝ) : EReal) ≤ k x * polarGauge k xStar
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
theorem abs_inner_le_mul_polarGauge_of_norm_like {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsGauge k)
(hk_even : ∀ x, k (-x) = k x) (hk_finite : ∀ x, k x ≠ ⊤)
(hk_pos : ∀ x, x ≠ 0 → (0 : EReal) < k x) (x xStar : Fin n → ℝ) :
((|x ⬝ᵥ xStar| : ℝ) : EReal) ≤ k x * polarGauge k xStar
Feasible multipliers for the inf-definition in Text 15.0.10.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
def innerMulFeasible {n : ℕ} {J : Set (Fin n → ℝ)} (j : J → EReal) (x : Fin n → ℝ) :
Set EReal :=
{μ : EReal | 0 ≤ μ ∧ ∀ y : J, ((x ⬝ᵥ (y : Fin n → ℝ) : ℝ) : EReal) ≤ μ * j y}
Candidate gauge defined by infimum of feasible multipliers.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
noncomputable def innerMulGauge {n : ℕ} {J : Set (Fin n → ℝ)} (j : J → EReal)
(x : Fin n → ℝ) : EReal :=
sInf (innerMulFeasible (J := J) j x)
The infimum definition yields a nonnegative value.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma innerMulGauge_nonneg {n : ℕ} {J : Set (Fin n → ℝ)} (j : J → EReal) (x : Fin n → ℝ) :
(0 : EReal) ≤ innerMulGauge (J := J) j x
The infimum definition yields zero at the origin.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma innerMulGauge_zero {n : ℕ} {J : Set (Fin n → ℝ)} (j : J → EReal) :
innerMulGauge (J := J) j (0 : Fin n → ℝ) = 0
If r ≥ 0 and a ≤ (μ + ε) * r for all ε > 0, then a ≤ μ * r.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma le_mul_of_forall_pos_add_mul {a μ r : ℝ} (hr : 0 ≤ r)
(h : ∀ ε > 0, a ≤ (μ + ε) * r) : a ≤ μ * r
If h is nonnegative and provides a feasible multiplier on H, then innerMulGauge ≤ h on H and hence innerMulGauge is finite on H.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma innerMulGauge_le_h_on_H {n : ℕ} {H J : Set (Fin n → ℝ)} (h : H → EReal) (j : J → EReal)
(h_nonneg : ∀ x : H, (0 : EReal) ≤ h x)
(h_ne_top : ∀ x : H, h x ≠ ⊤)
(hineq : ∀ x : H, ∀ y : J,
((x.1 ⬝ᵥ (y : Fin n → ℝ) : ℝ) : EReal) ≤ h x * j y) :
∀ x : H,
innerMulGauge (J := J) j x.1 ≤ h x ∧ innerMulGauge (J := J) j x.1 ≠ ⊤
On the effective domain, innerMulGauge satisfies the defining inequality.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma inner_le_innerMulGauge_mul_j_of_ne_top {n : ℕ} {J : Set (Fin n → ℝ)} (j : J → EReal)
(j_nonneg : ∀ y : J, (0 : EReal) ≤ j y) (j_ne_top : ∀ y : J, j y ≠ ⊤) :
∀ x, innerMulGauge (J := J) j x ≠ ⊤ → ∀ y : J,
((x ⬝ᵥ (y : Fin n → ℝ) : ℝ) : EReal) ≤ innerMulGauge (J := J) j x * j y
The inner-multiplier gauge never attains ⊥.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma innerMulGauge_ne_bot {n : ℕ} {J : Set (Fin n → ℝ)} (j : J → EReal) (x : Fin n → ℝ) :
innerMulGauge (J := J) j x ≠ (⊥ : EReal)
Scaling a feasible multiplier scales feasibility for the scaled vector.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma innerMulFeasible_smul {n : ℕ} {J : Set (Fin n → ℝ)} (j : J → EReal)
{x : Fin n → ℝ} {μ : EReal} (hμ : μ ∈ innerMulFeasible (J := J) j x) {t : ℝ}
(ht : 0 ≤ t) :
((t : ℝ) : EReal) * μ ∈ innerMulFeasible (J := J) j (t • x)
The inner-multiplier gauge is subadditive.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma innerMulGauge_add_le {n : ℕ} {J : Set (Fin n → ℝ)} (j : J → EReal)
(x z : Fin n → ℝ) :
innerMulGauge (J := J) j (x + z) ≤
innerMulGauge (J := J) j x + innerMulGauge (J := J) j z
The inner-multiplier gauge is positively homogeneous.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma innerMulGauge_posHom {n : ℕ} {J : Set (Fin n → ℝ)} (j : J → EReal) :
PositivelyHomogeneous (innerMulGauge (J := J) j)
The inner-multiplier gauge is a gauge.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma innerMulGauge_isGauge {n : ℕ} {J : Set (Fin n → ℝ)} (j : J → EReal) :
IsGauge (innerMulGauge (J := J) j)
Epigraph as an intersection of halfspaces for innerMulGauge.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma epigraph_innerMulGauge_eq_inter_iInter_halfspaces {n : ℕ} {J : Set (Fin n → ℝ)}
(j : J → EReal) (j_nonneg : ∀ y : J, (0 : EReal) ≤ j y)
(j_ne_top : ∀ y : J, j y ≠ ⊤) :
epigraph (S := (Set.univ : Set (Fin n → ℝ))) (innerMulGauge (J := J) j) =
{p : (Fin n → ℝ) × ℝ | 0 ≤ p.2} ∩
⋂ y : J, {p : (Fin n → ℝ) × ℝ | dotProduct p.1 (y : Fin n → ℝ) ≤
p.2 * (j y).toReal}
The epigraph of innerMulGauge is closed.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma isClosed_epigraph_innerMulGauge {n : ℕ} {J : Set (Fin n → ℝ)} (j : J → EReal)
(j_nonneg : ∀ y : J, (0 : EReal) ≤ j y) (j_ne_top : ∀ y : J, j y ≠ ⊤) :
IsClosed (epigraph (S := (Set.univ : Set (Fin n → ℝ))) (innerMulGauge (J := J) j))§15 — Polars of Convex Functions — Part 3
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
The polar gauge of innerMulGauge is bounded by j on J.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_innerMulGauge_le_j_on_J {n : ℕ} {J : Set (Fin n → ℝ)} (j : J → EReal)
(j_nonneg : ∀ y : J, (0 : EReal) ≤ j y) (j_ne_top : ∀ y : J, j y ≠ ⊤) :
∀ y : J, polarGauge (innerMulGauge (J := J) j) (y : Fin n → ℝ) ≤ j y ∧
polarGauge (innerMulGauge (J := J) j) (y : Fin n → ℝ) ≠ ⊤
Text 15.0.10: Given subsets H, J ⊆ ℝⁿ and functions h : H → [0, +∞], j : J → [0, +∞] such that ⟪x, y⟫ ≤ h x * j y for all x ∈ H, y ∈ J, define k(x) := inf { μ ≥ 0 | ⟪x, y⟫ ≤ μ * j y for all y ∈ J }. Then k is a closed gauge and satisfies ⟪x, y⟫ ≤ k x * j y for all x ∈ dom k, y ∈ J, with dom k ⊇ H and k x ≤ h x for all x ∈ H. Moreover, its polar gauge k^∘ satisfies dom k^∘ ⊇ J and k^∘ y ≤ j y for all y ∈ J, hence ⟪x, y⟫ ≤ k x * k^∘ y for all x ∈ dom k, y ∈ dom k^∘. In this development, ℝⁿ is Fin n → ℝ, [0, +∞] is modeled by EReal with nonnegativity, and effective-domain assumptions are modeled by k x ≠ ⊤.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
theorem exists_closedGauge_of_inner_le_mul {n : ℕ} {H J : Set (Fin n → ℝ)} (h : H → EReal)
(j : J → EReal)
(h_nonneg : ∀ x : H, (0 : EReal) ≤ h x)
(h_ne_top : ∀ x : H, h x ≠ ⊤)
(j_nonneg : ∀ y : J, (0 : EReal) ≤ j y)
(j_ne_top : ∀ y : J, j y ≠ ⊤)
(hineq : ∀ x : H, ∀ y : J, ((x.1 ⬝ᵥ (y : Fin n → ℝ) : ℝ) : EReal) ≤ h x * j y) :
let k : (Fin n → ℝ) → EReal := innerMulGauge (J := J) j
(IsGauge k ∧ IsClosed (epigraph (S := (Set.univ : Set (Fin n → ℝ))) k)) ∧
(∀ x : Fin n → ℝ,
k x ≠ ⊤ → ∀ y : J, ((x ⬝ᵥ (y : Fin n → ℝ) : ℝ) : EReal) ≤ k x * j y) ∧
(∀ x : H, k x.1 ≠ ⊤) ∧
(∀ x : H, k x.1 ≤ h x) ∧
(∀ y : J, polarGauge k (y : Fin n → ℝ) ≠ ⊤) ∧
(∀ y : J, polarGauge k (y : Fin n → ℝ) ≤ j y) ∧
(∀ x y : Fin n → ℝ,
k x ≠ ⊤ →
polarGauge k y ≠ ⊤ → ((x ⬝ᵥ y : ℝ) : EReal) ≤ k x * polarGauge k y)
Text 15.0.11: Let k(x) = ‖x‖₂ be the Euclidean norm on ℝⁿ. Then k is both the gauge function and the support function of the Euclidean unit ball B = {x | ‖x‖₂ ≤ 1}. In particular, the polar gauge satisfies k^∘ = k. Consequently, the inequality associated with the polar pair (k, k^∘) is the Schwarz inequality ⟪x, y⟫ ≤ ‖x‖₂ ‖y‖₂ for all x, y ∈ ℝⁿ. In this development, we work with Fin n → ℝ for ℝⁿ, define ‖x‖₂ as Real.sqrt (dotProduct x x), take the unit ball to be {x | Real.sqrt (dotProduct x x) ≤ 1}, and use gaugeFunction and supportFunctionEReal for the gauge and support functions.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
noncomputable def euclideanNormEReal {n : ℕ} (x : Fin n → ℝ) : EReal :=
((Real.sqrt (dotProduct x x) : ℝ) : EReal)
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
def piEuclideanUnitBall (n : ℕ) : Set (Fin n → ℝ) :=
{x | Real.sqrt (dotProduct x x) ≤ (1 : ℝ)}
Every vector can be written as its Euclidean norm times a unit-ball vector.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma exists_unitBall_smul_eq {n : ℕ} (x : Fin n → ℝ) :
∃ y ∈ piEuclideanUnitBall n, x = Real.sqrt (dotProduct x x) • y
The Euclidean unit ball is absorbing.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma piEuclideanUnitBall_absorbing {n : ℕ} :
∀ x : Fin n → ℝ, ∃ r : ℝ, 0 ≤ r ∧ ∃ y ∈ piEuclideanUnitBall n, x = r • y
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
theorem euclideanNormEReal_eq_gaugeFunction_euclideanUnitBall {n : ℕ} :
euclideanNormEReal (n := n) = fun x => (gaugeFunction (piEuclideanUnitBall n) x : EReal)
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
theorem euclideanNormEReal_eq_supportFunctionEReal_euclideanUnitBall {n : ℕ} :
euclideanNormEReal (n := n) = supportFunctionEReal (piEuclideanUnitBall n)
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
theorem dotProduct_le_mul_sqrt_dotProduct_self {n : ℕ} (x y : Fin n → ℝ) :
dotProduct x y ≤ Real.sqrt (dotProduct x x) * Real.sqrt (dotProduct y y)
Text 15.0.12: A gauge k : ℝⁿ → [0, +∞) is called a norm if it is finite everywhere, symmetric (k (-x) = k x for all x), and satisfies k x > 0 for all x ≠ 0. Equivalently, k is real-valued and satisfies: (a) k x > 0 for all x ≠ 0; (b) k (x₁ + x₂) ≤ k x₁ + k x₂ for all x₁, x₂; (c) k (λ • x) = λ * k x for all x and all λ > 0; (d) k (-x) = k x for all x. In particular, (c) and (d) imply k (λ • x) = |λ| * k x for all x and all λ ∈ ℝ. In this development, ℝⁿ is Fin n → ℝ, codomain is EReal, and "finite everywhere" is ∀ x, k x ≠ ⊤.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
def IsNormGauge {n : ℕ} (k : (Fin n → ℝ) → EReal) : Prop :=
IsGauge k ∧ (∀ x, k x ≠ ⊤) ∧ (∀ x, k (-x) = k x) ∧ (∀ x, x ≠ 0 → (0 : EReal) < k x)
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma IsNormGauge.smul_eq_abs {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsNormGauge k)
(x : Fin n → ℝ) (r : ℝ) :
k (r • x) = ((|r| : ℝ) : EReal) * k x
The epigraph of a gauge is a convex cone in ℝⁿ × ℝ.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
def epigraph_isConvexCone_of_isGauge {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsGauge k) :
ConvexCone ℝ ((Fin n → ℝ) × ℝ)
The closure-slice infimum used in Theorem 15.1 matches epigraphClosureInf.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma kCl_eq_epigraphClosureInf {n : ℕ} (k : (Fin n → ℝ) → EReal) :
(fun x =>
sInf
{μ : EReal |
∃ r : ℝ,
μ = (r : EReal) ∧
(x, r) ∈ closure (epigraph (S := (Set.univ : Set (Fin n → ℝ))) k)}) =
epigraphClosureInf k
The support function is bounded above by the polar gauge of a gauge-function.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma supportFunctionEReal_le_polarGauge_of_gaugeFunction {n : ℕ} (C : Set (Fin n → ℝ))
(xStar : Fin n → ℝ) :
supportFunctionEReal C xStar ≤
polarGauge (fun x => (gaugeFunction C x : EReal)) xStar
The support function of an absorbing set is nonnegative.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma supportFunctionEReal_nonneg_of_absorbing {n : ℕ} {C : Set (Fin n → ℝ)} (hCne : C.Nonempty)
(hCabs : ∀ x : Fin n → ℝ, ∃ r : ℝ, 0 ≤ r ∧ ∃ y ∈ C, x = r • y) :
∀ xStar : Fin n → ℝ, 0 ≤ supportFunctionEReal C xStar
The polar gauge of a gauge-function is bounded above by the support function under absorbing.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_of_gaugeFunction_le_supportFunctionEReal {n : ℕ} {C : Set (Fin n → ℝ)}
(hCne : C.Nonempty)
(hCabs : ∀ x : Fin n → ℝ, ∃ r : ℝ, 0 ≤ r ∧ ∃ y ∈ C, x = r • y) :
∀ xStar : Fin n → ℝ,
polarGauge (fun x => (gaugeFunction C x : EReal)) xStar ≤ supportFunctionEReal C xStar
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
theorem polarGauge_euclideanNormEReal {n : ℕ} :
polarGauge (euclideanNormEReal (n := n)) = euclideanNormEReal (n := n)
The sign-flip on the first coordinate of EuclideanSpace × ℝ is an involution.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma signflip_euclidean_involutive {n : ℕ} :
let E₂ := EuclideanSpace ℝ (Fin n)
let H₂ := E₂ × ℝ
let σ₂ : H₂ → H₂ := fun p => (-p.1, p.2)
∀ p : H₂, σ₂ (σ₂ p) = p
The inner product commutes with the sign-flip on the first coordinate in H₂.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma inner_signflip_H2 {n : ℕ} :
let E₂ := EuclideanSpace ℝ (Fin n)
let H₂ := WithLp 2 (E₂ × ℝ)
let σ₂ : H₂ → H₂ := fun p => WithLp.toLp 2 (- (WithLp.ofLp p).1, (WithLp.ofLp p).2)
∀ p q : H₂,
inner (𝕜 := ℝ) (σ₂ p) q = inner (𝕜 := ℝ) p (σ₂ q)
The inner dual cone commutes with sign-flip on the first coordinate in WithLp 2 (E × ℝ).
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma innerDual_preimage_signflip {n : ℕ}
(A : Set (WithLp 2 (EuclideanSpace ℝ (Fin n) × ℝ))) :
(ProperCone.innerDual
(E := WithLp 2 (EuclideanSpace ℝ (Fin n) × ℝ))
((fun p => WithLp.toLp 2 (- (WithLp.ofLp p).1, (WithLp.ofLp p).2)) ⁻¹' A) :
Set (WithLp 2 (EuclideanSpace ℝ (Fin n) × ℝ))) =
(fun p => WithLp.toLp 2 (- (WithLp.ofLp p).1, (WithLp.ofLp p).2)) ⁻¹'
(ProperCone.innerDual
(E := WithLp 2 (EuclideanSpace ℝ (Fin n) × ℝ)) A :
Set (WithLp 2 (EuclideanSpace ℝ (Fin n) × ℝ)))
Continuous linear equivalence between the inner-product model H₂ and ((Fin n → ℝ) × ℝ).
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
noncomputable def eH_transport_to_H2 (n : ℕ) :
WithLp 2 (EuclideanSpace ℝ (Fin n) × ℝ) ≃L[ℝ] ((Fin n → ℝ) × ℝ)
Epigraph points of a gauge have nonnegative second coordinate.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma mem_epigraph_snd_nonneg_of_isGauge {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsGauge k)
{x : Fin n → ℝ} {r : ℝ}
(hx : (x, r) ∈ epigraph (S := (Set.univ : Set (Fin n → ℝ))) k) : 0 ≤ r
The sign-flip on the first coordinate of H₂ is an involution.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma signflip_H2_involutive {n : ℕ} :
let E₂ := EuclideanSpace ℝ (Fin n)
let H₂ := WithLp 2 (E₂ × ℝ)
let σ₂ : H₂ → H₂ := fun p => WithLp.toLp 2 (- (WithLp.ofLp p).1, (WithLp.ofLp p).2)
∀ p : H₂, σ₂ (σ₂ p) = p
The inner dual cone in H₂ is closed.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma isClosed_innerDual_H2 {n : ℕ}
(S : Set (WithLp 2 (EuclideanSpace ℝ (Fin n) × ℝ))) :
IsClosed ((ProperCone.innerDual
(E := WithLp 2 (EuclideanSpace ℝ (Fin n) × ℝ)) S :
Set (WithLp 2 (EuclideanSpace ℝ (Fin n) × ℝ))))
Explicit formula for the inverse of eH_transport_to_H2.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma eH_transport_to_H2_symm_apply {n : ℕ} (x : Fin n → ℝ) (r : ℝ) :
let eH := eH_transport_to_H2 n
eH.symm (x, r) = WithLp.toLp 2 (WithLp.toLp 2 x, r)
Inner product formula for the transported sign-flip in H₂.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma inner_eH_symm_signflip {n : ℕ} (x xStar : Fin n → ℝ) (r μ : ℝ) :
inner (𝕜 := ℝ) ((eH_transport_to_H2 n).symm (x, r))
((fun p : WithLp 2 (EuclideanSpace ℝ (Fin n) × ℝ) =>
WithLp.toLp 2 (- (WithLp.ofLp p).1, (WithLp.ofLp p).2))
((eH_transport_to_H2 n).symm (xStar, μ))) =
- dotProduct x xStar + r * μ§15 — Polars of Convex Functions — Part 4
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
Epigraph of the polar gauge as a sign-flipped inner-dual cone in H₂.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma epigraph_polarGauge_eq_preimage_innerDualCone_H2 {n : ℕ} {k : (Fin n → ℝ) → EReal}
(hk : IsGauge k) (xStar : Fin n → ℝ) (μ : ℝ) :
((xStar, μ) ∈ epigraph (S := (Set.univ : Set (Fin n → ℝ))) (polarGauge k)) ↔
(let E₂ := EuclideanSpace ℝ (Fin n)
let H₂ := WithLp 2 (E₂ × ℝ)
let eH := eH_transport_to_H2 n
let σ₂ : H₂ → H₂ := fun p => WithLp.toLp 2 (- (WithLp.ofLp p).1, (WithLp.ofLp p).2)
let S₂ : Set H₂ := eH.symm '' epigraph (S := (Set.univ : Set (Fin n → ℝ))) k
σ₂ (eH.symm (xStar, μ)) ∈
(ProperCone.innerDual (E := H₂) S₂ : Set H₂))
The epigraph of the polar gauge is closed.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma isClosed_epigraph_polarGauge {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsGauge k) :
IsClosed (epigraph (S := (Set.univ : Set (Fin n → ℝ))) (polarGauge k))§15 — Polars of Convex Functions — Part 5
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
The bipolar of a gauge equals the epigraph-closure infimum.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_bipolar_eq_kCl {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsGauge k) :
let kCl : (Fin n → ℝ) → EReal :=
fun x =>
sInf
{μ : EReal |
∃ r : ℝ,
μ = (r : EReal) ∧
(x, r) ∈ closure (epigraph (S := (Set.univ : Set (Fin n → ℝ))) k)}
polarGauge (polarGauge k) = kCl
Theorem 15.1: If k is a gauge function, then its polar k^∘ is a closed gauge function and k^{∘∘} = cl k. Moreover, if k = γ(· | C) for some nonempty convex set C ⊆ ℝ^n that is absorbing, then k^∘ agrees with the support function of C. In this development, k^∘ is polarGauge k; "closed" is expressed as IsClosed (epigraph univ ·); and cl k is modeled as the extended-real function whose epigraph is the topological closure of epigraph univ k.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
theorem polarGauge_isGauge_isClosed_and_bipolar_eq_closure_epigraph {n : ℕ}
{k : (Fin n → ℝ) → EReal} (hk : IsGauge k) :
let kPolar : (Fin n → ℝ) → EReal := polarGauge k
let kCl : (Fin n → ℝ) → EReal :=
fun x =>
sInf
{μ : EReal |
∃ r : ℝ,
μ = (r : EReal) ∧
(x, r) ∈ closure (epigraph (S := (Set.univ : Set (Fin n → ℝ))) k)}
IsGauge kPolar ∧
IsClosed (epigraph (S := (Set.univ : Set (Fin n → ℝ))) kPolar) ∧
polarGauge kPolar = kCl ∧
∀ C : Set (Fin n → ℝ),
C.Nonempty →
Convex ℝ C →
(∀ x : Fin n → ℝ, ∃ r : ℝ, 0 ≤ r ∧ ∃ y ∈ C, x = r • y) →
k = (fun x => (gaugeFunction C x : EReal)) →
kPolar = supportFunctionEReal C
A gauge with closed epigraph equals its epigraph-closure infimum.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma epigraphClosureInf_eq_of_isGauge_isClosed_epigraph {n : ℕ} {k : (Fin n → ℝ) → EReal}
(hk : IsGauge k)
(hclosed : IsClosed (epigraph (S := (Set.univ : Set (Fin n → ℝ))) k)) :
epigraphClosureInf k = k
Closed gauges are fixed by double polarity.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_involutive_of_isGauge_isClosed_epigraph {n : ℕ} {k : (Fin n → ℝ) → EReal}
(hk : IsGauge k)
(hclosed : IsClosed (epigraph (S := (Set.univ : Set (Fin n → ℝ))) k)) :
polarGauge (polarGauge k) = k
The polar gauge of a closed gauge is again a closed gauge.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_mapsTo_closedGauges {n : ℕ} {k : (Fin n → ℝ) → EReal} :
(IsGauge k ∧ IsClosed (epigraph (S := (Set.univ : Set (Fin n → ℝ))) k)) →
(IsGauge (polarGauge k) ∧
IsClosed (epigraph (S := (Set.univ : Set (Fin n → ℝ))) (polarGauge k)))
Corollary 15.1.1: The polarity mapping k ↦ k^∘ induces a one-to-one symmetric correspondence on the class of all closed gauges on ℝⁿ. Two closed convex sets containing the origin are polar to each other if and only if their gauge functions are polar to each other.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
theorem polarGauge_bijOn_closedGauges {n : ℕ} :
Set.BijOn (fun k : (Fin n → ℝ) → EReal => polarGauge k)
{k | IsGauge k ∧ IsClosed (epigraph (S := (Set.univ : Set (Fin n → ℝ))) k)}
{k | IsGauge k ∧ IsClosed (epigraph (S := (Set.univ : Set (Fin n → ℝ))) k)}
The polar set is the ≤ 1 sublevel set of the support function.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma supportFunctionEReal_sublevel_one_eq_polarSet {n : ℕ} {C : Set (Fin n → ℝ)} :
{xStar | supportFunctionEReal C xStar ≤ (1 : EReal)} =
{xStar | ∀ x ∈ C, (x ⬝ᵥ xStar : ℝ) ≤ 1}
The polar gauge of a gauge-function equals the support function under absorbing.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma polarGauge_gaugeFunction_eq_supportFunctionEReal {n : ℕ} {C : Set (Fin n → ℝ)}
(hCne : C.Nonempty)
(hCabs : ∀ x : Fin n → ℝ, ∃ lam : ℝ, 0 ≤ lam ∧ x ∈ (fun y => lam • y) '' C) :
polarGauge (fun x => (gaugeFunction C x : EReal)) = supportFunctionEReal C
The support function is bounded above by the gauge of a polar set.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma supportFunctionEReal_le_gaugeFunction_of_polarSet {n : ℕ} {C D : Set (Fin n → ℝ)}
(hDabs : ∀ x : Fin n → ℝ, ∃ lam : ℝ, 0 ≤ lam ∧ x ∈ (fun y => lam • y) '' D)
(hD : D = {xStar | ∀ x ∈ C, (x ⬝ᵥ xStar : ℝ) ≤ 1}) :
∀ xStar : Fin n → ℝ,
supportFunctionEReal C xStar ≤ (gaugeFunction D xStar : EReal)
The support function of a polar set agrees with the corresponding gauge function.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
lemma supportFunctionEReal_eq_gaugeFunction_of_polarSet {n : ℕ} {C D : Set (Fin n → ℝ)}
(hC0 : (0 : Fin n → ℝ) ∈ C)
(hCabs : ∀ x : Fin n → ℝ, ∃ lam : ℝ, 0 ≤ lam ∧ x ∈ (fun y => lam • y) '' C)
(hDabs : ∀ x : Fin n → ℝ, ∃ lam : ℝ, 0 ≤ lam ∧ x ∈ (fun y => lam • y) '' D)
(hD : D = {xStar | ∀ x ∈ C, (x ⬝ᵥ xStar : ℝ) ≤ 1}) :
supportFunctionEReal C = fun xStar => (gaugeFunction D xStar : EReal)
Corollary 15.1.1 (sets version): Two closed convex absorbing sets containing the origin are polar to each other if and only if their gauge functions are polar to each other.
open scoped BigOperators
open scoped Pointwise
open scoped RealInnerProductSpace
theorem closed_convex_sets_polar_iff_gaugeFunctions_polar {n : ℕ} {C D : Set (Fin n → ℝ)}
(hC_closed : IsClosed C) (hC_conv : Convex ℝ C) (hC0 : (0 : Fin n → ℝ) ∈ C)
(hCabs : ∀ x : Fin n → ℝ, ∃ lam : ℝ, 0 ≤ lam ∧ x ∈ (fun y => lam • y) '' C)
(hD_closed : IsClosed D) (hD_conv : Convex ℝ D) (hD0 : (0 : Fin n → ℝ) ∈ D)
(hDabs : ∀ x : Fin n → ℝ, ∃ lam : ℝ, 0 ≤ lam ∧ x ∈ (fun y => lam • y) '' D) :
(D = {xStar | ∀ x ∈ C, (x ⬝ᵥ xStar : ℝ) ≤ 1} ∧
C = {x | ∀ xStar ∈ D, (x ⬝ᵥ xStar : ℝ) ≤ 1}) ↔
(polarGauge (fun x => (gaugeFunction C x : EReal)) =
(fun xStar => (gaugeFunction D xStar : EReal)) ∧
polarGauge (fun xStar => (gaugeFunction D xStar : EReal)) =
(fun x => (gaugeFunction C x : EReal)))§15 — Polars of Convex Functions — Part 6
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
The gauge of a closed convex absorbing set containing 0 is a gauge.
open scoped BigOperators
open scoped Pointwise
lemma gaugeFunctionEReal_isGauge_of_closed_convex_zeroMem_absorbing {n : ℕ} {C : Set (Fin n → ℝ)}
(hC_closed : IsClosed C) (hC_conv : Convex ℝ C) (hC0 : (0 : Fin n → ℝ) ∈ C)
(hCabs : ∀ x : Fin n → ℝ, ∃ lam : ℝ, 0 ≤ lam ∧ x ∈ (fun y => lam • y) '' C) :
IsGauge (fun x => (gaugeFunction C x : EReal))
The gauge function of a closed convex absorbing set has closed epigraph.
open scoped BigOperators
open scoped Pointwise
lemma isClosed_epigraph_gaugeFunctionEReal_of_closed_convex_zeroMem_absorbing
{n : ℕ} {C : Set (Fin n → ℝ)} (hC_closed : IsClosed C) (hC_conv : Convex ℝ C)
(hC0 : (0 : Fin n → ℝ) ∈ C)
(hCabs : ∀ x : Fin n → ℝ, ∃ lam : ℝ, 0 ≤ lam ∧ x ∈ (fun y => lam • y) '' C) :
IsClosed (epigraph (S := (Set.univ : Set (Fin n → ℝ)))
(fun x => (gaugeFunction C x : EReal)))
The polar gauge of the support function agrees with the gauge of the set.
open scoped BigOperators
open scoped Pointwise
lemma polarGauge_supportFunctionEReal_eq_gaugeFunctionEReal {n : ℕ} {C : Set (Fin n → ℝ)}
(hCne : C.Nonempty)
(hCabs : ∀ x : Fin n → ℝ, ∃ lam : ℝ, 0 ≤ lam ∧ x ∈ (fun y => lam • y) '' C)
(hγ : IsGauge (fun x => (gaugeFunction C x : EReal)))
(hγclosed :
IsClosed (epigraph (S := (Set.univ : Set (Fin n → ℝ)))
(fun x => (gaugeFunction C x : EReal)))) :
polarGauge (supportFunctionEReal C) = (fun x => (gaugeFunction C x : EReal))
Corollary 15.1.2: If C ⊆ ℝ^n is a closed convex set containing 0, then the gauge γ(· | C) and the support function δ^*(· | C) are closed gauges polar to each other. In this development, γ(· | C) is fun x => (gaugeFunction C x : EReal), the support function is supportFunctionEReal C, "closed" is expressed as IsClosed (epigraph univ ·), and polarity of gauges is given by polarGauge. We additionally assume C is absorbing.
open scoped BigOperators
open scoped Pointwise
theorem gaugeFunction_and_supportFunctionEReal_closedGauges_polar {n : ℕ} {C : Set (Fin n → ℝ)}
(hC_closed : IsClosed C) (hC_conv : Convex ℝ C) (hC0 : (0 : Fin n → ℝ) ∈ C)
(hCabs : ∀ x : Fin n → ℝ, ∃ lam : ℝ, 0 ≤ lam ∧ x ∈ (fun y => lam • y) '' C) :
let γ : (Fin n → ℝ) → EReal := fun x => (gaugeFunction C x : EReal)
let δ : (Fin n → ℝ) → EReal := supportFunctionEReal C
(IsGauge γ ∧ IsClosed (epigraph (S := (Set.univ : Set (Fin n → ℝ))) γ)) ∧
(IsGauge δ ∧ IsClosed (epigraph (S := (Set.univ : Set (Fin n → ℝ))) δ)) ∧
polarGauge γ = δ ∧ polarGauge δ = γ
If 0 lies in the interior of a convex set, then the set is absorbing.
open scoped BigOperators
open scoped Pointwise
lemma absorbing_of_zero_mem_interior {n : ℕ} {C : Set (Fin n → ℝ)} (hCconv : Convex ℝ C)
(hC0 : (0 : Fin n → ℝ) ∈ interior C) :
∀ x : Fin n → ℝ, ∃ lam : ℝ, 0 ≤ lam ∧ x ∈ (fun y => lam • y) '' C
A norm gauge is continuous on ℝ^n.
open scoped BigOperators
open scoped Pointwise
lemma normGauge_continuous {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsNormGauge k) :
Continuous k
The unit sublevel set of a norm gauge is closed.
open scoped BigOperators
open scoped Pointwise
lemma normGauge_sublevel_one_closed {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsNormGauge k) :
IsClosed {x : Fin n → ℝ | k x ≤ (1 : EReal)}
The unit sublevel set of a norm gauge contains the origin in its interior.
open scoped BigOperators
open scoped Pointwise
lemma normGauge_sublevel_one_zero_mem_interior {n : ℕ} {k : (Fin n → ℝ) → EReal}
(hk : IsNormGauge k) :
(0 : Fin n → ℝ) ∈ interior {x : Fin n → ℝ | k x ≤ (1 : EReal)}
The recession cone of the unit sublevel set of a norm gauge is {0}.
open scoped BigOperators
open scoped Pointwise
lemma normGauge_sublevel_one_recessionCone_eq_singleton_zero {n : ℕ}
{k : (Fin n → ℝ) → EReal} (hk : IsNormGauge k) :
Set.recessionCone {x : Fin n → ℝ | k x ≤ (1 : EReal)} = ({0} : Set (Fin n → ℝ))
The unit sublevel set of a norm gauge is bounded.
open scoped BigOperators
open scoped Pointwise
lemma normGauge_sublevel_one_bounded {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsNormGauge k) :
Bornology.IsBounded {x : Fin n → ℝ | k x ≤ (1 : EReal)}
The unit sublevel set of a norm gauge is symmetric, closed, bounded, convex, and contains 0 in its interior.
open scoped BigOperators
open scoped Pointwise
lemma normGauge_sublevelOne_symmClosedBoundedConvex {n : ℕ} {k : (Fin n → ℝ) → EReal}
(hk : IsNormGauge k) :
let C := {x : Fin n → ℝ | k x ≤ (1 : EReal)};
IsClosed C ∧
Convex ℝ C ∧
Bornology.IsBounded C ∧
(0 : Fin n → ℝ) ∈ interior C ∧ (∀ x, x ∈ C → -x ∈ C)
The gauge of a symmetric closed bounded convex set with 0 in its interior is a norm gauge.
open scoped BigOperators
open scoped Pointwise
lemma gaugeFunctionEReal_isNormGauge_of_symmClosedBoundedConvex_zeroInterior {n : ℕ}
{C : Set (Fin n → ℝ)} (hC_closed : IsClosed C) (hC_conv : Convex ℝ C)
(hC_bdd : Bornology.IsBounded C) (hC0 : (0 : Fin n → ℝ) ∈ interior C)
(hC_symm : ∀ x, x ∈ C → -x ∈ C) :
IsNormGauge (fun x => (gaugeFunction C x : EReal))
The support function of a symmetric closed bounded convex set with 0 in its interior is a norm gauge.
open scoped BigOperators
open scoped Pointwise
lemma supportFunctionEReal_isNormGauge_of_symmClosedBoundedConvex_zeroInterior {n : ℕ}
{C : Set (Fin n → ℝ)} (hCne : C.Nonempty) (hC_conv : Convex ℝ C)
(hC_bdd : Bornology.IsBounded C) (hC0 : (0 : Fin n → ℝ) ∈ interior C)
(hC_symm : ∀ x, x ∈ C → -x ∈ C) :
IsNormGauge (supportFunctionEReal C)
The polar gauge of a norm gauge is the support function of its unit sublevel set.
open scoped BigOperators
open scoped Pointwise
lemma polarGauge_eq_supportFunctionEReal_unitSublevel {n : ℕ} {k : (Fin n → ℝ) → EReal}
(hk : IsNormGauge k) :
let C := {x : Fin n → ℝ | k x ≤ (1 : EReal)};
polarGauge k = supportFunctionEReal C
Theorem 15.2: The relations k(x) = γ(x | C) and C = {x ∈ ℝ^n | k x ≤ 1} define a one-to-one correspondence between norms k on ℝ^n and symmetric closed bounded convex sets C ⊆ ℝ^n such that 0 ∈ interior C. In this development, ℝ^n is Fin n → ℝ; the unit ball of k is {x | k x ≤ 1}; and γ(x | C) is represented by fun x => (gaugeFunction C x : EReal).
open scoped BigOperators
open scoped Pointwise
noncomputable def normGaugeEquiv_symmClosedBoundedConvex {n : ℕ} :
{k : (Fin n → ℝ) → EReal // IsNormGauge k} ≃
{C : Set (Fin n → ℝ) //
IsClosed C ∧
Convex ℝ C ∧
Bornology.IsBounded C ∧
(0 : Fin n → ℝ) ∈ interior C ∧ (∀ x, x ∈ C → -x ∈ C)}
Theorem 15.2 (polar): Under the norm/set correspondence of Theorem 15.2, the polar of a norm is a norm. In this development, the polar of k is polarGauge k.
open scoped BigOperators
open scoped Pointwise
theorem polarGauge_isNormGauge {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsNormGauge k) :
IsNormGauge (polarGauge k)
Text 15.0.13: Define k : ℝⁿ → [0, +∞) by k(x) = max {|ξ₁|, …, |ξₙ|} for x = (ξ₁, …, ξₙ). Then k is a norm, and its polar gauge k^∘ is the ℓ₁-norm: k^∘(x⋆) = |ξ₁⋆| + ⋯ + |ξₙ⋆| for x⋆ = (ξ₁⋆, …, ξₙ⋆). Consequently, k and k^∘ form a polar pair of norms. In this development, ℝⁿ is Fin n → ℝ, and we model real-valued norms as EReal-valued functions via coercion ℝ → EReal.
open scoped BigOperators
open scoped Pointwise
noncomputable def linftyNormEReal {n : ℕ} (x : Fin n → ℝ) : EReal :=
((‖x‖ : ℝ) : EReal)
open scoped BigOperators
open scoped Pointwise
noncomputable def l1NormEReal {n : ℕ} (x : Fin n → ℝ) : EReal :=
((∑ i, |x i|) : ℝ)
The sign function satisfies sign r * r = |r|.
open scoped BigOperators
open scoped Pointwise
lemma real_sign_mul_self_eq_abs (r : ℝ) : Real.sign r * r = |r|
The absolute value of Real.sign is bounded by 1.
open scoped BigOperators
open scoped Pointwise
lemma abs_sign_le_one (r : ℝ) : |Real.sign r| ≤ (1 : ℝ)
The support function of the ℓ∞ unit ball is the ℓ1 norm.
open scoped BigOperators
open scoped Pointwise
lemma supportFunctionEReal_linftyUnitBall_eq_l1NormEReal {n : ℕ} :
let C : Set (Fin n → ℝ) := {x | (‖x‖ : ℝ) ≤ 1};
supportFunctionEReal C = l1NormEReal (n := n)
The support function of the ℓ1 unit ball is the ℓ∞ norm.
open scoped BigOperators
open scoped Pointwise
lemma supportFunctionEReal_l1UnitBall_eq_linftyNormEReal {n : ℕ} :
let C : Set (Fin n → ℝ) := {x | (∑ i, |x i|) ≤ 1};
supportFunctionEReal C = linftyNormEReal (n := n)
open scoped BigOperators
open scoped Pointwise
theorem linftyNormEReal_isNormGauge {n : ℕ} : IsNormGauge (linftyNormEReal (n := n))
open scoped BigOperators
open scoped Pointwise
theorem polarGauge_linftyNormEReal {n : ℕ} :
polarGauge (linftyNormEReal (n := n)) = l1NormEReal (n := n)
open scoped BigOperators
open scoped Pointwise
theorem polarGauge_l1NormEReal {n : ℕ} : polarGauge (l1NormEReal (n := n)) = linftyNormEReal§15 — Polars of Convex Functions — Part 7
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
Text 15.0.14: A metric on ℝⁿ is a function ρ : ℝⁿ × ℝⁿ → ℝ such that: (1) ρ(x, y) > 0 if x ≠ y, and ρ(x, y) = 0 if x = y; (2) ρ(x, y) = ρ(y, x) for all x, y; (3) ρ(x, z) ≤ ρ(x, y) + ρ(y, z) for all x, y, z. The quantity ρ(x, y) is interpreted as the distance between x and y. In this development, we use Fin n → ℝ for ℝⁿ. The usual mathlib notion of a metric space is MetricSpace, and IsMetricFun ρ records the book axioms for the distance function ρ.
open scoped BigOperators
open scoped Pointwise
def IsMetricFun {n : ℕ} (ρ : (Fin n → ℝ) → (Fin n → ℝ) → ℝ) : Prop :=
(∀ x y, x ≠ y → 0 < ρ x y) ∧
(∀ x, ρ x x = 0) ∧
(∀ x y, ρ x y = ρ y x) ∧ ∀ x y z, ρ x z ≤ ρ x y + ρ y z
A metric function vanishes only on the diagonal.
open scoped BigOperators
open scoped Pointwise
lemma eq_of_IsMetricFun_eq_zero {n : ℕ} {ρ : (Fin n → ℝ) → (Fin n → ℝ) → ℝ}
(hρ : IsMetricFun ρ) {x y : Fin n → ℝ} (hxy : ρ x y = 0) : x = y
Build a metric space structure from a metric function.
open scoped BigOperators
open scoped Pointwise
lemma exists_metricSpace_of_IsMetricFun {n : ℕ}
(ρ : (Fin n → ℝ) → (Fin n → ℝ) → ℝ) (hρ : IsMetricFun ρ) :
∃ ms : MetricSpace (Fin n → ℝ), ms.dist = ρ
A metric space with distance ρ satisfies the metric-function axioms.
open scoped BigOperators
open scoped Pointwise
lemma isMetricFun_of_exists_metricSpace {n : ℕ}
(ρ : (Fin n → ℝ) → (Fin n → ℝ) → ℝ)
(hρ : ∃ ms : MetricSpace (Fin n → ℝ), ms.dist = ρ) : IsMetricFun ρ
open scoped BigOperators
open scoped Pointwise
theorem isMetricFun_iff_exists_metricSpace {n : ℕ} (ρ : (Fin n → ℝ) → (Fin n → ℝ) → ℝ) :
IsMetricFun ρ ↔ ∃ ms : MetricSpace (Fin n → ℝ), ms.dist = ρ
Text 15.0.15: An extreme example of a metric on ℝⁿ is the function ρ(x, y) = 0 if x = y and ρ(x, y) = 1 if x ≠ y (the discrete metric). This metric has no relation to the algebraic structure of ℝⁿ.
open scoped BigOperators
open scoped Pointwise
noncomputable def discreteMetricFun {n : ℕ} : (Fin n → ℝ) → (Fin n → ℝ) → ℝ
open scoped BigOperators
open scoped Pointwise
theorem discreteMetricFun_isMetricFun {n : ℕ} : IsMetricFun (discreteMetricFun (n := n))
Text 15.0.16: A metric ρ on ℝⁿ is called a Minkowski metric if it satisfies the metric axioms and: - Translation invariance: ρ (x + z) (y + z) = ρ x y for all x, y, z. - Segment homogeneity: ρ x ((1 - λ) • x + λ • y) = λ * ρ x y for all x, y and λ ∈ [0, 1].
open scoped BigOperators
open scoped Pointwise
def IsMinkowskiMetricFun {n : ℕ} (ρ : (Fin n → ℝ) → (Fin n → ℝ) → ℝ) : Prop :=
IsMetricFun ρ ∧
(∀ x y z, ρ (x + z) (y + z) = ρ x y) ∧
∀ x y (r : ℝ),
r ∈ Set.Icc (0 : ℝ) 1 → ρ x ((1 - r) • x + r • y) = r * ρ x y
A real-valued function k : ℝⁿ → ℝ is a norm if it is nonnegative, vanishes only at 0, satisfies the triangle inequality, and is absolutely homogeneous: k (r • x) = |r| * k x.
open scoped BigOperators
open scoped Pointwise
def IsNormFun {n : ℕ} (k : (Fin n → ℝ) → ℝ) : Prop :=
(∀ x, 0 ≤ k x) ∧
(∀ x, k x = 0 ↔ x = 0) ∧ (∀ x y, k (x + y) ≤ k x + k y) ∧
∀ x (r : ℝ), k (r • x) = |r| * k x
Given a norm function k, define the associated distance function ρ x y := k (x - y).
open scoped BigOperators
open scoped Pointwise
def normFunToMinkowskiMetricFun {n : ℕ} (k : (Fin n → ℝ) → ℝ) :
(Fin n → ℝ) → (Fin n → ℝ) → ℝ :=
fun x y => k (x - y)
Given a Minkowski metric ρ, define the associated norm function k x := ρ x 0.
open scoped BigOperators
open scoped Pointwise
def minkowskiMetricFunToNormFun {n : ℕ} (ρ : (Fin n → ℝ) → (Fin n → ℝ) → ℝ) :
(Fin n → ℝ) → ℝ :=
fun x => ρ x 0
Displacement to a convex combination is a scalar multiple of the difference.
open scoped BigOperators
open scoped Pointwise
lemma sub_segment_eq_smul_sub {n : ℕ} (x y : Fin n → ℝ) (r : ℝ) :
x - ((1 - r) • x + r • y) = r • (x - y)
A norm induces a Minkowski metric via ρ x y = k (x - y).
open scoped BigOperators
open scoped Pointwise
lemma isMinkowskiMetricFun_normFunToMinkowskiMetricFun {n : ℕ} {k : (Fin n → ℝ) → ℝ}
(hk : IsNormFun k) :
IsMinkowskiMetricFun (normFunToMinkowskiMetricFun (n := n) k)
For a Minkowski metric, the induced norm is homogeneous on [0,1].
open scoped BigOperators
open scoped Pointwise
lemma minkowskiMetricFunToNormFun_hom_Icc {n : ℕ}
{ρ : (Fin n → ℝ) → (Fin n → ℝ) → ℝ} (hρ : IsMinkowskiMetricFun ρ) :
∀ x r, r ∈ Set.Icc (0 : ℝ) 1 →
minkowskiMetricFunToNormFun (n := n) ρ (r • x) =
r * minkowskiMetricFunToNormFun (n := n) ρ x
The norm induced by a Minkowski metric is even.
open scoped BigOperators
open scoped Pointwise
lemma minkowskiMetricFunToNormFun_neg {n : ℕ}
{ρ : (Fin n → ℝ) → (Fin n → ℝ) → ℝ} (hρ : IsMinkowskiMetricFun ρ) :
∀ x,
minkowskiMetricFunToNormFun (n := n) ρ (-x) =
minkowskiMetricFunToNormFun (n := n) ρ x
The norm induced by a Minkowski metric is homogeneous for nonnegative scalars.
open scoped BigOperators
open scoped Pointwise
lemma minkowskiMetricFunToNormFun_hom_nonneg {n : ℕ}
{ρ : (Fin n → ℝ) → (Fin n → ℝ) → ℝ} (hρ : IsMinkowskiMetricFun ρ) :
∀ x r, 0 ≤ r →
minkowskiMetricFunToNormFun (n := n) ρ (r • x) =
r * minkowskiMetricFunToNormFun (n := n) ρ x
A Minkowski metric induces a norm via k x = ρ x 0.
open scoped BigOperators
open scoped Pointwise
lemma isNormFun_minkowskiMetricFunToNormFun {n : ℕ}
{ρ : (Fin n → ℝ) → (Fin n → ℝ) → ℝ} (hρ : IsMinkowskiMetricFun ρ) :
IsNormFun (minkowskiMetricFunToNormFun (n := n) ρ)
Translation invariance pins down a Minkowski metric by its values at the origin.
open scoped BigOperators
open scoped Pointwise
lemma minkowskiMetricFun_eq_normFunToMinkowskiMetricFun_minkowskiMetricFunToNormFun {n : ℕ}
{ρ : (Fin n → ℝ) → (Fin n → ℝ) → ℝ}
(htrans : ∀ x y z, ρ (x + z) (y + z) = ρ x y) :
∀ x y,
ρ x y =
normFunToMinkowskiMetricFun (n := n) (minkowskiMetricFunToNormFun (n := n) ρ) x y
Text 15.0.17: There is a one-to-one correspondence between Minkowski metrics and norms: if k is a norm then ρ x y := k (x - y) is a Minkowski metric, and conversely every Minkowski metric ρ arises uniquely in this way from the norm k x := ρ x 0.
open scoped BigOperators
open scoped Pointwise
noncomputable def normFunEquivMinkowskiMetricFun {n : ℕ} :
{k : (Fin n → ℝ) → ℝ // IsNormFun k} ≃
{ρ : (Fin n → ℝ) → (Fin n → ℝ) → ℝ // IsMinkowskiMetricFun ρ}
The gauge of a symmetric closed bounded convex set with 0 in its interior is a norm.
open scoped BigOperators
open scoped Pointwise
lemma isNormFun_gaugeFunction_of_symmClosedBoundedConvex_zeroInterior {n : ℕ}
{C : Set (Fin n → ℝ)} (hC_symm : ∀ x, x ∈ C ↔ -x ∈ C) (hC_closed : IsClosed C)
(hC_bdd : Bornology.IsBounded C) (hC_conv : Convex ℝ C)
(hC0 : (0 : Fin n → ℝ) ∈ interior C) : IsNormFun (gaugeFunction C)
The gauge of a closed convex set with 0 in its interior has unit sublevel set C.
open scoped BigOperators
open scoped Pointwise
lemma gaugeFunction_unitBall_eq {n : ℕ} {C : Set (Fin n → ℝ)} (hC_closed : IsClosed C)
(hC_conv : Convex ℝ C) (hC0 : (0 : Fin n → ℝ) ∈ interior C) :
{x : Fin n → ℝ | gaugeFunction C x ≤ (1 : ℝ)} = C
Metric balls from a norm are translates of scaled unit balls.
open scoped BigOperators
open scoped Pointwise
lemma ball_eq_image_add_smul_of_normFun_unitBall {n : ℕ} {k : (Fin n → ℝ) → ℝ}
{C : Set (Fin n → ℝ)} (hk : IsNormFun k) (hC : {x : Fin n → ℝ | k x ≤ 1} = C) :
∀ x : Fin n → ℝ, ∀ ε : ℝ, 0 < ε →
{y | normFunToMinkowskiMetricFun (n := n) k x y ≤ ε} = (fun c => x + ε • c) '' C
The induced norm of a Minkowski metric with prescribed balls is the gauge of C.
open scoped BigOperators
open scoped Pointwise
lemma minkowskiMetricFunToNormFun_eq_gaugeFunction_of_ball_property {n : ℕ}
{C : Set (Fin n → ℝ)} {ρ : (Fin n → ℝ) → (Fin n → ℝ) → ℝ}
(hρ : IsMinkowskiMetricFun ρ)
(hball : ∀ ε : ℝ, 0 < ε → {y | ρ 0 y ≤ ε} = (fun c => ε • c) '' C) :
∀ x, minkowskiMetricFunToNormFun (n := n) ρ x = gaugeFunction C x
Text 15.0.18: If C is a symmetric closed bounded convex set such that 0 ∈ interior C, then there is a unique Minkowski metric ρ such that the ρ-ball of radius ε > 0 around x is x + ε C, i.e. {y | ρ x y ≤ ε} = {x + ε • c | c ∈ C}.
open scoped BigOperators
open scoped Pointwise
theorem existsUnique_minkowskiMetricFun_ball_eq_image_add_smul {n : ℕ} {C : Set (Fin n → ℝ)}
(hC_symm : ∀ x, x ∈ C ↔ -x ∈ C) (hC_closed : IsClosed C) (hC_bdd : Bornology.IsBounded C)
(hC_conv : Convex ℝ C) (hC0 : (0 : Fin n → ℝ) ∈ interior C) :
∃! ρ : (Fin n → ℝ) → (Fin n → ℝ) → ℝ,
IsMinkowskiMetricFun ρ ∧
∀ x : Fin n → ℝ,
∀ ε : ℝ, 0 < ε → {y | ρ x y ≤ ε} = (fun c => x + ε • c) '' C
A distance function d : α → α → ℝ defines a notion of "open set" by the usual metric-ball criterion: s is open if for every x ∈ s there exists ε > 0 such that d x y < ε implies y ∈ s.
open scoped BigOperators
open scoped Pointwise
def IsOpenOfDistFun {α : Type*} (d : α → α → ℝ) (s : Set α) : Prop :=
∀ x ∈ s, ∃ ε : ℝ, 0 < ε ∧ ∀ y, d x y < ε → y ∈ s
open scoped BigOperators
open scoped Pointwise
def IsClosedOfDistFun {α : Type*} (d : α → α → ℝ) (s : Set α) : Prop :=
IsOpenOfDistFun d sᶜ
A sequence u : ℕ → α is Cauchy with respect to d if for every ε > 0 there exists N such that d (u m) (u n) < ε for all m,n ≥ N.
open scoped BigOperators
open scoped Pointwise
def CauchySeqOfDistFun {α : Type*} (d : α → α → ℝ) (u : ℕ → α) : Prop :=
∀ ε : ℝ, 0 < ε → ∃ N : ℕ, ∀ m n : ℕ, N ≤ m → N ≤ n → d (u m) (u n) < ε
Euclidean distance on ℝⁿ (modeled as Fin n → ℝ), defined by d(x,y) = ‖x - y‖₂ = sqrt (dotProduct (x - y) (x - y)). We use a pi-prefixed name to avoid clashing with mathlib's euclideanDist.
open scoped BigOperators
open scoped Pointwise
noncomputable def piEuclideanDist {n : ℕ} (x y : Fin n → ℝ) : ℝ :=
Real.sqrt (dotProduct (x - y) (x - y))§15 — Polars of Convex Functions — Part 8
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
The Euclidean norm scales linearly with scalar multiplication.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma piEuclideanNorm_smul {n : ℕ} (r : ℝ) (x : Fin n → ℝ) :
Real.sqrt (dotProduct (r • x) (r • x)) = |r| * Real.sqrt (dotProduct x x)
Comparison of the sup norm with the Euclidean norm on Fin n → ℝ.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma supNorm_le_piEuclideanNorm_and_piEuclideanNorm_le_sqrt_n_mul_supNorm {n : ℕ}
(x : Fin n → ℝ) :
‖x‖ ≤ Real.sqrt (dotProduct x x) ∧
Real.sqrt (dotProduct x x) ≤ Real.sqrt (n : ℝ) * ‖x‖
A neighborhood of 0 contains a scaled Euclidean unit ball.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma exists_pos_smul_piEuclideanUnitBall_subset_of_zero_mem_interior {n : ℕ}
{C : Set (Fin n → ℝ)} (hC0 : (0 : Fin n → ℝ) ∈ interior C) :
∃ α : ℝ, 0 < α ∧ (fun x => α • x) '' piEuclideanUnitBall n ⊆ C
A bounded set is contained in a scaled Euclidean unit ball.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma exists_pos_smul_piEuclideanUnitBall_superset_of_isBounded {n : ℕ}
{C : Set (Fin n → ℝ)} (hC_bdd : Bornology.IsBounded C) :
∃ β : ℝ, 0 < β ∧ (C ⊆ (fun x => β • x) '' piEuclideanUnitBall n)
Bounds on ρ from the unit-ball sandwich and the ball formula.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma rho_bounds_of_ball_eq_and_unitBall_sandwich {n : ℕ}
{C : Set (Fin n → ℝ)} {ρ : (Fin n → ℝ) → (Fin n → ℝ) → ℝ}
{α β : ℝ}
(hαpos : 0 < α) (hβpos : 0 < β)
(hCα : (fun x => α • x) '' piEuclideanUnitBall n ⊆ C)
(hCβ : C ⊆ (fun x => β • x) '' piEuclideanUnitBall n)
(hρdiag : ∀ x, ρ x x = 0)
(hρpos : ∀ x y, x ≠ y → 0 < ρ x y)
(hball : ∀ x ε, 0 < ε → {y | ρ x y ≤ ε} = (fun c => x + ε • c) '' C) :
∀ x y : Fin n → ℝ,
β⁻¹ * piEuclideanDist x y ≤ ρ x y ∧ ρ x y ≤ α⁻¹ * piEuclideanDist x y
Two-sided bounds between distance functions give equivalent open/closed/Cauchy notions.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma isOpen_isClosed_cauchy_equiv_of_two_sided_bounds {α : Type*}
{d₁ d₂ : α → α → ℝ} {c₁ c₂ : ℝ}
(hc₁ : 0 < c₁) (hc₂ : 0 < c₂)
(hbound : ∀ x y, c₁ * d₁ x y ≤ d₂ x y ∧ d₂ x y ≤ c₂ * d₁ x y) :
(∀ s : Set α, IsOpenOfDistFun d₂ s ↔ IsOpenOfDistFun d₁ s) ∧
(∀ s : Set α, IsClosedOfDistFun d₂ s ↔ IsClosedOfDistFun d₁ s) ∧
(∀ u : ℕ → α, CauchySeqOfDistFun d₂ u ↔ CauchySeqOfDistFun d₁ u)
Text 15.0.19: Since C is bounded and 0 ∈ int C, there exist α, β > 0 such that α B ⊆ C ⊆ β B, where B is the Euclidean unit ball. Writing d(x,y) = ‖x - y‖₂, the associated Minkowski metric ρ satisfies the two-sided comparison α^{-1} d(x,y) ≥ ρ(x,y) ≥ β^{-1} d(x,y) for all x,y. Consequently, Minkowski metrics on ℝⁿ are topologically equivalent to the Euclidean metric: they induce the same open/closed sets and the same Cauchy sequences.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
theorem exists_constants_unitBall_sandwich_and_topologicalEquiv_of_minkowskiMetricFun_ball_eq
{n : ℕ} {C : Set (Fin n → ℝ)} {ρ : (Fin n → ℝ) → (Fin n → ℝ) → ℝ}
(hρ : IsMinkowskiMetricFun ρ) (hC_bdd : Bornology.IsBounded C)
(hC0 : (0 : Fin n → ℝ) ∈ interior C)
(hball : ∀ x : Fin n → ℝ, ∀ ε : ℝ, 0 < ε → {y | ρ x y ≤ ε} = (fun c => x + ε • c) '' C) :
∃ α β : ℝ,
0 < α ∧
0 < β ∧
((fun x => α • x) '' piEuclideanUnitBall n ⊆ C) ∧
(C ⊆ (fun x => β • x) '' piEuclideanUnitBall n) ∧
(∀ x y : Fin n → ℝ,
β⁻¹ * piEuclideanDist x y ≤ ρ x y ∧ ρ x y ≤ α⁻¹ * piEuclideanDist x y) ∧
(∀ s : Set (Fin n → ℝ),
IsOpenOfDistFun ρ s ↔ IsOpenOfDistFun (piEuclideanDist (n := n)) s) ∧
(∀ s : Set (Fin n → ℝ),
IsClosedOfDistFun ρ s ↔ IsClosedOfDistFun (piEuclideanDist (n := n)) s) ∧
∀ u : ℕ → Fin n → ℝ,
CauchySeqOfDistFun ρ u ↔
CauchySeqOfDistFun (piEuclideanDist (n := n)) u
Text 15.0.20: An extended-real-valued convex function f : ℝⁿ → (-∞, +∞] is *gauge-like* if f(0) = inf f and all its sublevel sets {x | f x ≤ α} for f(0) < α < +∞ are proportional, i.e. each is a positive scalar multiple of any other such sublevel set.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
def IsGaugeLike {n : ℕ} (f : (Fin n → ℝ) → EReal) : Prop :=
ConvexFunctionOn (S := (Set.univ : Set (Fin n → ℝ))) f ∧
f 0 = sInf (Set.range f) ∧
∀ ⦃α β : EReal⦄,
f 0 < α → α < ⊤ → f 0 < β → β < ⊤ →
∃ t : ℝ, 0 < t ∧ {x | f x ≤ α} = t • {x | f x ≤ β}
Sublevel sets of a gauge-like function are homothetic.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma sublevel_sets_homothetic_of_IsGaugeLike {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf : IsGaugeLike f) {α β : EReal} (hα0 : f 0 < α) (hαtop : α < ⊤)
(hβ0 : f 0 < β) (hβtop : β < ⊤) :
∃ t : ℝ, 0 < t ∧ {x | f x ≤ α} = t • {x | f x ≤ β}
Sublevel sets of a gauge scale by positive factors.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma sublevel_eq_smul_sublevel_one_of_isGauge {n : ℕ} {k : (Fin n → ℝ) → EReal}
(hk : IsGauge k) {t : ℝ} (ht : 0 < t) :
{x | k x ≤ (t : EReal)} = t • {x | k x ≤ (1 : EReal)}
A gauge taking a positive finite value attains 1 on some point.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma exists_unit_level_of_pos_finite {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsGauge k) :
(∃ y : Fin n → ℝ, (0 : EReal) < k y ∧ k y < ⊤) → ∃ y1 : Fin n → ℝ, k y1 = (1 : EReal)
The support function of a nonempty convex set containing 0 is a gauge.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma supportFunctionEReal_isGauge_of_convex_zeroMem {n : ℕ} {C : Set (Fin n → ℝ)}
(hCne : C.Nonempty) (hCconv : Convex ℝ C) (hC0 : (0 : Fin n → ℝ) ∈ C) :
IsGauge (supportFunctionEReal C)
A gauge is *closed* if its epigraph over univ is closed.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
def IsClosedGauge {n : ℕ} (k : (Fin n → ℝ) → EReal) : Prop :=
IsGauge k ∧ IsClosed (epigraph (S := (Set.univ : Set (Fin n → ℝ))) k)
The support function of a nonempty convex set containing 0 is a closed gauge.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma supportFunctionEReal_isClosedGauge_of_convex_zeroMem {n : ℕ} {C : Set (Fin n → ℝ)}
(hCne : C.Nonempty) (hCconv : Convex ℝ C) (hC0 : (0 : Fin n → ℝ) ∈ C) :
IsClosedGauge (supportFunctionEReal C)
The support function of a set containing 0 is nonnegative.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma supportFunctionEReal_nonneg_of_zero_mem {n : ℕ} {C : Set (Fin n → ℝ)}
(hC0 : (0 : Fin n → ℝ) ∈ C) (xStar : Fin n → ℝ) :
(0 : EReal) ≤ supportFunctionEReal C xStar
A function is *gauge-like closed proper convex* if it is gauge-like and is a closed proper convex EReal-valued function on ℝⁿ (modeled as Fin n → ℝ).
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
def IsGaugeLikeClosedProperConvex {n : ℕ} (f : (Fin n → ℝ) → EReal) : Prop :=
IsGaugeLike f ∧ ClosedConvexFunction f ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) f
A gauge-like closed proper convex function has a finite level strictly above f 0.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma exists_real_beta_between_f0_and_top {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf : IsGaugeLikeClosedProperConvex f) :
∃ β : ℝ, f 0 < (β : EReal) ∧ (β : EReal) < ⊤
Sublevel sets of a closed convex function are closed and convex.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma sublevel_closed_convex_of_closedConvex {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf : ClosedConvexFunction f) (α : ℝ) :
IsClosed {x | f x ≤ (α : EReal)} ∧ Convex ℝ {x | f x ≤ (α : EReal)}
The bipolar set always contains the original set.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma subset_bipolar_of_any {n : ℕ} {C : Set (Fin n → ℝ)} :
C ⊆
{x | ∀ xStar ∈ {xStar | ∀ x ∈ C, (x ⬝ᵥ xStar : ℝ) ≤ 1},
(x ⬝ᵥ xStar : ℝ) ≤ 1}
Any linear functional on Fin n → ℝ is a dot product with its coordinates on the standard basis.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma linearMap_eq_dotProduct_piSingle {n : ℕ} (f : (Fin n → ℝ) →ₗ[ℝ] ℝ) :
∀ x, f x = (x ⬝ᵥ fun i => f (Pi.single i 1) : ℝ)
A closed convex set containing 0 coincides with its bipolar.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma bipolar_eq_of_closed_convex_zeroMem {n : ℕ} {C : Set (Fin n → ℝ)}
(hCclosed : IsClosed C) (hCconv : Convex ℝ C) (hC0 : (0 : Fin n → ℝ) ∈ C) :
C =
{x | ∀ xStar ∈ {xStar | ∀ x ∈ C, (x ⬝ᵥ xStar : ℝ) ≤ 1},
(x ⬝ᵥ xStar : ℝ) ≤ 1}
Build a closed gauge from a base sublevel of a gauge-like closed proper convex function.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma closedGauge_from_baseSublevel_supportFunctionEReal_polar_and_unitSublevel {n : ℕ}
{f : (Fin n → ℝ) → EReal} {β : ℝ}
(hf : IsGaugeLikeClosedProperConvex f)
(hβ0 : f 0 < (β : EReal)) :
let C : Set (Fin n → ℝ) := {x | f x ≤ (β : EReal)}
let Cpolar : Set (Fin n → ℝ) := {xStar | ∀ x ∈ C, (x ⬝ᵥ xStar : ℝ) ≤ 1}
let k : (Fin n → ℝ) → EReal := supportFunctionEReal Cpolar
IsClosedGauge k ∧ {x | k x ≤ (1 : EReal)} = C
Pick a base sublevel and a closed gauge whose unit sublevel matches it.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma exists_closedGauge_unitSublevel_eq_baseSublevel {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf : IsGaugeLikeClosedProperConvex f) :
∃ β : ℝ, f 0 < (β : EReal) ∧ (β : EReal) < ⊤ ∧
∃ k : (Fin n → ℝ) → EReal, IsClosedGauge k ∧
{x | k x ≤ (1 : EReal)} = {x | f x ≤ (β : EReal)}
The support function of the unit sublevel of a gauge is bounded by its polar gauge.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma supportFunction_unitSublevel_le_polarGauge_of_isGauge {n : ℕ} {k : (Fin n → ℝ) → EReal}
(xStar : Fin n → ℝ) :
supportFunctionEReal {x | k x ≤ (1 : EReal)} xStar ≤ polarGauge k xStar
If the unit-sublevel support function is finite, then any x with k x = 0 pairs nonpositively with the dual vector.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma dotProduct_le_zero_of_k_eq_zero_of_supportFunction_ne_top {n : ℕ}
{k : (Fin n → ℝ) → EReal} (hk : IsGauge k) {xStar : Fin n → ℝ}
(hμtop : supportFunctionEReal {x | k x ≤ (1 : EReal)} xStar ≠ ⊤)
{x : Fin n → ℝ} (hx0 : k x = 0) :
((x ⬝ᵥ xStar : ℝ) : EReal) ≤ 0§15 — Polars of Convex Functions — Part 9
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
The polar gauge is bounded above by the support function of the unit sublevel.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma polarGauge_le_supportFunction_unitSublevel_of_isGauge {n : ℕ} {k : (Fin n → ℝ) → EReal}
(hk : IsGauge k) (xStar : Fin n → ℝ) :
polarGauge k xStar ≤ supportFunctionEReal {x | k x ≤ (1 : EReal)} xStar
The support function of a gauge's unit sublevel equals the polar gauge.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma supportFunction_unitSublevel_eq_polarGauge_of_isGauge {n : ℕ} {k : (Fin n → ℝ) → EReal}
(hk : IsGauge k) (xStar : Fin n → ℝ) :
supportFunctionEReal {x | k x ≤ (1 : EReal)} xStar = polarGauge k xStar
The monotone conjugate g⁺ of a function g : [0, +∞] → (-∞, +∞], defined by g⁺(s) = sup_{t ≥ 0} (t * s - g t).
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
noncomputable def monotoneConjugateERealNonneg (g : EReal → EReal) (s : EReal) : EReal :=
sSup ((fun t : EReal => t * s - g t) '' {t : EReal | 0 ≤ t})
The monotone conjugate is monotone in its argument.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma monotoneConjugateERealNonneg_mono {g : EReal → EReal} :
Monotone (monotoneConjugateERealNonneg g)
Each affine term lies below the monotone conjugate.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma term_le_monotoneConjugateERealNonneg {g : EReal → EReal} {s t : EReal} (ht : 0 ≤ t) :
t * s - g t ≤ monotoneConjugateERealNonneg g s
The monotone conjugate takes value ⊤ at ⊤ once g is finite at some positive point.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma monotoneConjugateERealNonneg_top_of_exists_finite {g : EReal → EReal}
(hζ : ∃ ζ : ℝ, 0 < ζ ∧ g (ζ : EReal) ≠ ⊤ ∧ g (ζ : EReal) ≠ ⊥) :
monotoneConjugateERealNonneg g ⊤ = ⊤
Real cutoffs for the monotone conjugate are bounded above once g is finite at some positive point.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma bddAbove_cutoff_real_of_monotoneConjugateERealNonneg_of_exists_finite {g : EReal → EReal}
(hζ : ∃ ζ : ℝ, 0 < ζ ∧ g (ζ : EReal) ≠ ⊤ ∧ g (ζ : EReal) ≠ ⊥) {αr : ℝ} :
BddAbove
{s : ℝ | 0 ≤ s ∧ monotoneConjugateERealNonneg g (s : EReal) ≤ (αr : EReal)}
The cutoff for a monotone conjugate lies in its real sublevel.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma csSup_cutoff_mem_sublevel_monotoneConjugateERealNonneg {g : EReal → EReal}
(hg_top : g ⊤ = ⊤) {αr : ℝ}
(hA_bdd : BddAbove
{s : ℝ | 0 ≤ s ∧ monotoneConjugateERealNonneg g (s : EReal) ≤ (αr : EReal)})
(hA_nonempty :
({s : ℝ | 0 ≤ s ∧ monotoneConjugateERealNonneg g (s : EReal) ≤ (αr : EReal)}).Nonempty) :
0 ≤ (sSup
{s : ℝ | 0 ≤ s ∧ monotoneConjugateERealNonneg g (s : EReal) ≤ (αr : EReal)} : ℝ) ∧
monotoneConjugateERealNonneg g
((sSup {s : ℝ | 0 ≤ s ∧
monotoneConjugateERealNonneg g (s : EReal) ≤ (αr : EReal)} : ℝ) : EReal)
≤ (αr : EReal)
Sublevels of g⁺ ∘ kᵒ are kᵒ-sublevels at the real cutoff.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma sublevel_monotoneConjugate_comp_polarGauge_eq_sublevel_polarGauge_csSup {n : ℕ}
{k : (Fin n → ℝ) → EReal} {g : EReal → EReal} (hg_top : g ⊤ = ⊤)
(hζ : ∃ ζ : ℝ, 0 < ζ ∧ g (ζ : EReal) ≠ ⊤ ∧ g (ζ : EReal) ≠ ⊥) {αr : ℝ}
(hA_bdd : BddAbove
{s : ℝ | 0 ≤ s ∧ monotoneConjugateERealNonneg g (s : EReal) ≤ (αr : EReal)})
(hA_nonempty :
({s : ℝ | 0 ≤ s ∧ monotoneConjugateERealNonneg g (s : EReal) ≤ (αr : EReal)}).Nonempty) :
{xStar | monotoneConjugateERealNonneg g (polarGauge k xStar) ≤ (αr : EReal)} =
{xStar |
polarGauge k xStar ≤
((sSup
{s : ℝ | 0 ≤ s ∧
monotoneConjugateERealNonneg g (s : EReal) ≤ (αr : EReal)} : ℝ) : EReal)}
A convex interpolation bound for the monotone conjugate on the nonnegative ray.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma monotoneConjugateERealNonneg_le_affine {g : EReal → EReal} (hg_top : g ⊤ = ⊤)
(h0_ne_top : monotoneConjugateERealNonneg g (0 : EReal) ≠ ⊤) {s0 lam : ℝ} (hlam0 : 0 ≤ lam)
(hlam1 : lam ≤ 1) :
monotoneConjugateERealNonneg g ((lam * s0 : ℝ) : EReal) ≤
((1 - lam : ℝ) : EReal) * monotoneConjugateERealNonneg g (0 : EReal) +
((lam : ℝ) : EReal) * monotoneConjugateERealNonneg g (s0 : EReal)
Cutoffs for the monotone conjugate are either both zero or both positive.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma cutoff_pos_or_all_zero_of_monotoneConjugateERealNonneg {g : EReal → EReal} (hg_top : g ⊤ = ⊤)
(h0_ne_bot : monotoneConjugateERealNonneg g (0 : EReal) ≠ ⊥) {αr βr : ℝ}
(hα0 : monotoneConjugateERealNonneg g (0 : EReal) < (αr : EReal))
(hβ0 : monotoneConjugateERealNonneg g (0 : EReal) < (βr : EReal))
(hAα_bdd : BddAbove
{s : ℝ | 0 ≤ s ∧ monotoneConjugateERealNonneg g (s : EReal) ≤ (αr : EReal)})
(hAβ_bdd : BddAbove
{s : ℝ | 0 ≤ s ∧ monotoneConjugateERealNonneg g (s : EReal) ≤ (βr : EReal)}) :
(sSup {s : ℝ | 0 ≤ s ∧ monotoneConjugateERealNonneg g (s : EReal) ≤ (αr : EReal)} = 0 ∧
sSup {s : ℝ | 0 ≤ s ∧ monotoneConjugateERealNonneg g (s : EReal) ≤ (βr : EReal)} = 0) ∨
(0 <
sSup {s : ℝ | 0 ≤ s ∧ monotoneConjugateERealNonneg g (s : EReal) ≤ (αr : EReal)} ∧
0 <
sSup {s : ℝ | 0 ≤ s ∧ monotoneConjugateERealNonneg g (s : EReal) ≤ (βr : EReal)})
Admissible levels covering a k-sublevel by a sublevel of f.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
def profileSet {n : ℕ} (f k : (Fin n → ℝ) → EReal) (z : EReal) : Set EReal :=
{α : EReal | f 0 < α ∧ α < ⊤ ∧ {x | k x ≤ z} ⊆ {x | f x ≤ α}}
The admissible sets are antitone in the threshold.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma profileSet_mono {n : ℕ} {f k : (Fin n → ℝ) → EReal} {z₁ z₂ : EReal}
(hz : z₁ ≤ z₂) :
profileSet f k z₂ ⊆ profileSet f k z₁
The profile infimum is monotone in the threshold.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma profileFun_mono {n : ℕ} {f k : (Fin n → ℝ) → EReal} :
Monotone (fun z : EReal => sInf (profileSet f k z))
The profile infimum with a ⊤ guard is monotone.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma profileFun_with_top_mono {n : ℕ} {f k : (Fin n → ℝ) → EReal} :
Monotone (fun z : EReal => if z = ⊤ then ⊤ else sInf (profileSet f k z))
The profile infimum with a ⊤ guard sends ⊤ to ⊤.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma profileFun_with_top_top {n : ℕ} {f k : (Fin n → ℝ) → EReal} :
(fun z : EReal => if z = ⊤ then ⊤ else sInf (profileSet f k z)) ⊤ = ⊤
A convex function on t ≥ 0 with closed epigraph cannot take value ⊥ at 0 if it is finite somewhere on the positive ray.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma g0_ne_bot_of_convex_closed_and_exists_finite_nonneg {g : EReal → EReal}
(hg_conv :
ConvexFunctionOn (S := {t : Fin 1 → ℝ | 0 ≤ t 0}) (fun t => g (t 0 : EReal)))
(hg_closed :
IsClosed (epigraph (S := {t : Fin 1 → ℝ | 0 ≤ t 0}) (fun t => g (t 0 : EReal))))
(hζ : ∃ ζ : ℝ, 0 < ζ ∧ g (ζ : EReal) ≠ ⊤ ∧ g (ζ : EReal) ≠ ⊥) :
g (0 : EReal) ≠ ⊥
A monotone function is never ⊥ on nonnegative inputs once g 0 ≠ ⊥.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma monotone_ne_bot_of_nonneg {g : EReal → EReal} (hg_mono : Monotone g)
(h0 : g (0 : EReal) ≠ ⊥) :
∀ t : EReal, (0 : EReal) ≤ t → g t ≠ ⊥
Sublevels of g ∘ k are contained in the k-sublevel at the sSup cutoff.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma sublevel_comp_subset_sublevel_gauge_sSup {n : ℕ} {k : (Fin n → ℝ) → EReal}
{g : EReal → EReal} (hk : IsGauge k) {α : EReal} :
{x | g (k x) ≤ α} ⊆ {x | k x ≤ sSup {t : EReal | 0 ≤ t ∧ g t ≤ α}}
Closed epigraph on the nonnegative ray gives lower semicontinuity on that ray.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma lscOn_g_nonneg_of_isClosed_epigraph {g : EReal → EReal}
(hg_closed :
IsClosed
(epigraph (S := {t : Fin 1 → ℝ | 0 ≤ t 0}) (fun t => g (t 0 : EReal)))) :
LowerSemicontinuousOn (fun t : Fin 1 → ℝ => g (t 0 : EReal)) {t : Fin 1 → ℝ | 0 ≤ t 0}
Lower semicontinuity on t ≥ 0 makes the sSup cutoff attainable.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma g_le_csSup_of_lscOn_nonneg {g : EReal → EReal}
(hlsc :
LowerSemicontinuousOn (fun t : Fin 1 → ℝ => g (t 0 : EReal)) {t : Fin 1 → ℝ | 0 ≤ t 0})
{α : ℝ} (h0α : g (0 : EReal) ≤ (α : EReal))
(hA_bdd : BddAbove {s : ℝ | 0 ≤ s ∧ g (s : EReal) ≤ (α : EReal)}) :
g ((sSup {s : ℝ | 0 ≤ s ∧ g (s : EReal) ≤ (α : EReal)} : ℝ) : EReal) ≤ (α : EReal)§15 — Polars of Convex Functions — Part 10
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma bddAbove_sublevel_real_of_mono_convex_strict_nonneg {g : EReal → EReal}
(hg_mono : Monotone g)
(hg_conv : ConvexFunctionOn (S := {t : Fin 1 → ℝ | 0 ≤ t 0}) (fun t => g (t 0 : EReal)))
(hstrict : ∃ s : ℝ, 0 < s ∧ g (0 : EReal) < g (s : EReal)) {αr : ℝ}
(hα0 : g (0 : EReal) < (αr : EReal)) :
BddAbove {s : ℝ | 0 ≤ s ∧ g (s : EReal) ≤ (αr : EReal)}
There is a positive real in the cutoff sublevel once g is finite at some positive point.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma exists_pos_mem_sublevel_real_of_convex_finite_at_pos {g : EReal → EReal}
(hg_conv : ConvexFunctionOn (S := {t : Fin 1 → ℝ | 0 ≤ t 0}) (fun t => g (t 0 : EReal)))
(hζ : ∃ ζ : ℝ, 0 < ζ ∧ g (ζ : EReal) ≠ ⊤ ∧ g (ζ : EReal) ≠ ⊥) {αr : ℝ}
(hα0 : g (0 : EReal) < (αr : EReal)) :
∃ s : ℝ, 0 < s ∧ g (s : EReal) ≤ (αr : EReal)
A composition with a closed gauge is gauge-like closed proper convex under the hypotheses.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma comp_closedGauge_gives_gaugeLikeClosedProperConvex {n : ℕ} {k : (Fin n → ℝ) → EReal}
{g : EReal → EReal} (hk : IsClosedGauge k) (hg_mono : Monotone g)
(hg_conv : ConvexFunctionOn (S := {t : Fin 1 → ℝ | 0 ≤ t 0}) (fun t => g (t 0 : EReal)))
(hg_closed :
IsClosed
(epigraph (S := {t : Fin 1 → ℝ | 0 ≤ t 0}) (fun t => g (t 0 : EReal))))
(hg_top : g ⊤ = ⊤)
(hζ : ∃ ζ : ℝ, 0 < ζ ∧ g (ζ : EReal) ≠ ⊤ ∧ g (ζ : EReal) ≠ ⊥)
(hstrict : ∃ s : ℝ, 0 < s ∧ g (0 : EReal) < g (s : EReal)) :
IsGaugeLikeClosedProperConvex (fun x : Fin n → ℝ => g (k x))
Theorem 15.3: A function f is a gauge-like closed proper convex function if and only if it can be expressed as f(x) = g (k x), where k is a closed gauge and g : [0, +∞] → (-∞, +∞] is a non-decreasing, lower semicontinuous convex function which strictly increases at some positive point and satisfies g(ζ) finite for some ζ > 0 (with the convention g(+∞) = +∞).
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
theorem gaugeLikeClosedProperConvex_iff_exists_closedGauge_comp {n : ℕ}
(f : (Fin n → ℝ) → EReal) :
IsGaugeLikeClosedProperConvex f ↔
∃ (k : (Fin n → ℝ) → EReal) (g : EReal → EReal),
IsClosedGauge k ∧
(∃ ζ : ℝ, 0 < ζ ∧ g (ζ : EReal) ≠ ⊤ ∧ g (ζ : EReal) ≠ ⊥) ∧
(∃ s : ℝ, 0 < s ∧ g (0 : EReal) < g (s : EReal)) ∧
Monotone g ∧
ConvexFunctionOn (S := {t : Fin 1 → ℝ | 0 ≤ t 0}) (fun t => g (t 0 : EReal)) ∧
IsClosed
(epigraph (S := {t : Fin 1 → ℝ | 0 ≤ t 0}) (fun t => g (t 0 : EReal))) ∧
g ⊤ = ⊤ ∧
f = fun x => g (k x)
Formula part of Theorem 15.3 for the conjugate of g ∘ k.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma fenchelConjugate_eq_monotoneConjugate_polarGauge_of_comp_formula {n : ℕ}
{f : (Fin n → ℝ) → EReal} {k : (Fin n → ℝ) → EReal} {g : EReal → EReal}
(hk : IsClosedGauge k) (hfg : f = fun x => g (k x)) (hg_mono : Monotone g)
(hg_top : g ⊤ = ⊤)
(hζ : ∃ ζ : ℝ, 0 < ζ ∧ g (ζ : EReal) ≠ ⊤ ∧ g (ζ : EReal) ≠ ⊥) :
∀ xStar : Fin n → ℝ,
fenchelConjugate n f xStar = monotoneConjugateERealNonneg g (polarGauge k xStar)§15 — Polars of Convex Functions — Part 11
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
Theorem 15.3 (conjugate formula): If f(x) = g (k x) with k a closed gauge and g as above, then f* is gauge-like and satisfies f*(x⋆) = g⁺ (kᵒ x⋆), where g⁺ is monotoneConjugateERealNonneg g and kᵒ is the polar gauge polarGauge k.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
theorem fenchelConjugate_eq_monotoneConjugate_polarGauge_of_comp {n : ℕ}
{f : (Fin n → ℝ) → EReal} {k : (Fin n → ℝ) → EReal} {g : EReal → EReal}
(hk : IsClosedGauge k) (hfg : f = fun x => g (k x)) (hg_mono : Monotone g)
(hg_top : g ⊤ = ⊤)
(hζ : ∃ ζ : ℝ, 0 < ζ ∧ g (ζ : EReal) ≠ ⊤ ∧ g (ζ : EReal) ≠ ⊥) :
IsGaugeLike (fenchelConjugate n f) ∧
∀ xStar : Fin n → ℝ,
fenchelConjugate n f xStar = monotoneConjugateERealNonneg g (polarGauge k xStar)
Text 15.0.21: Let p > 0. A function f : ℝⁿ → (-∞, +∞] is positively homogeneous of degree p if f (λ x) = λ^p f x for all x and all λ > 0. In this development, ℝⁿ is Fin n → ℝ, (-∞, +∞] is EReal, and λ^p is Real.rpow λ p.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
def PositivelyHomogeneousOfDegree {n : ℕ} (p : ℝ) (f : (Fin n → ℝ) → EReal) : Prop :=
∀ x : Fin n → ℝ, ∀ t : ℝ, 0 < t → f (t • x) = ((Real.rpow t p : ℝ) : EReal) * f x
Sublevel sets scale under positive homogeneity.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma sublevel_eq_smul_sublevel_of_posHomogeneous {n : ℕ} {p : ℝ} {f : (Fin n → ℝ) → EReal}
(hp : 0 < p) (hf_hom : PositivelyHomogeneousOfDegree (n := n) p f) {αr βr : ℝ}
(hα : 0 < αr) (hβ : 0 < βr) :
let t : ℝ := Real.rpow (αr / βr) (1 / p)
{x | f x ≤ (αr : EReal)} = t • {x | f x ≤ (βr : EReal)}
Power profile used in the rpow representation.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
noncomputable def phiPow (r : ℝ) : EReal → EReal := fun z =>
if z = ⊥ then (0 : EReal) else if z = ⊤ then (⊤ : EReal) else (Real.rpow (max z.toReal 0) r : EReal)
Multiplying ⊤ by a positive real yields ⊤ in EReal.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma ereal_coe_mul_top_of_pos {t : ℝ} (ht : 0 < t) :
((t : ℝ) : EReal) * (⊤ : EReal) = (⊤ : EReal)
The power profile scales on nonnegative inputs.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma phiPow_mul_of_nonneg {r : ℝ} {z : EReal} {t : ℝ} (hz : 0 ≤ z) (ht : 0 < t) :
phiPow r (((t : ℝ) : EReal) * z) =
((Real.rpow t r : ℝ) : EReal) * phiPow r z
The power profile is monotone for nonnegative exponents.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma phiPow_mono {r : ℝ} (hr : 0 ≤ r) : Monotone (phiPow r)
Basic properties of gPow z = (1/p) * phiPow p z.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma gPow_basic {p : ℝ} (hp : 1 < p) :
Monotone (fun z : EReal => ((1 / p : ℝ) : EReal) * phiPow p z) ∧
(fun z : EReal => ((1 / p : ℝ) : EReal) * phiPow p z) ⊤ = ⊤ ∧
∃ ζ : ℝ, 0 < ζ ∧
(fun z : EReal => ((1 / p : ℝ) : EReal) * phiPow p z) (ζ : EReal) ≠ ⊤ ∧
(fun z : EReal => ((1 / p : ℝ) : EReal) * phiPow p z) (ζ : EReal) ≠ ⊥
The monotone conjugate of the power profile (1/p) * phiPow p on nonnegative inputs.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma monotoneConjugateERealNonneg_gPow_eq_one_div_q_phiPow {p q : ℝ} (hp : 1 < p)
(hpq : (1 / p) + (1 / q) = 1) {s : EReal} (hs : 0 ≤ s) :
monotoneConjugateERealNonneg (fun z : EReal => ((1 / p : ℝ) : EReal) * phiPow p z) s =
((1 / q : ℝ) : EReal) * phiPow q s
Sublevel sets of a closed gauge match sublevels of a positive-homogeneous function.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma sublevel_closedGauge_eq_sublevel_f_pow {n : ℕ} {p : ℝ} {f : (Fin n → ℝ) → EReal}
{k : (Fin n → ℝ) → EReal} (hp : 0 < p)
(hf_hom : PositivelyHomogeneousOfDegree (n := n) p f) (hk : IsGauge k)
(hunit : {x | k x ≤ (1 : EReal)} = {x | f x ≤ ((1 / p : ℝ) : EReal)}) :
∀ {r : ℝ}, 0 < r →
{x | k x ≤ (r : EReal)} =
{x | f x ≤ (((1 / p : ℝ) * Real.rpow r p) : ℝ)}
A closed proper convex positively homogeneous function is finite at the origin.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma f0_ne_top_of_closedProperConvex_posHomogeneous {n : ℕ} {p : ℝ}
{f : (Fin n → ℝ) → EReal} (hp : 1 < p)
(hf_closed : ClosedConvexFunction f)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) f)
(hf_hom : PositivelyHomogeneousOfDegree (n := n) p f) :
f 0 ≠ ⊤
A closed proper convex positively homogeneous function is gauge-like.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma isGaugeLike_of_closedProperConvex_posHomogeneous {n : ℕ} {p : ℝ}
{f : (Fin n → ℝ) → EReal} (hp : 1 < p)
(hf_closed : ClosedConvexFunction f)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) f)
(hf_hom : PositivelyHomogeneousOfDegree (n := n) p f) :
IsGaugeLike f§15 — Polars of Convex Functions — Part 12
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
A closed proper convex positively homogeneous function vanishes at the origin.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma f0_eq_zero_of_closedProperConvex_posHomogeneous {n : ℕ} {p : ℝ}
{f : (Fin n → ℝ) → EReal} (hp : 1 < p)
(hf_closed : ClosedConvexFunction f)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) f)
(hf_hom : PositivelyHomogeneousOfDegree (n := n) p f) :
f 0 = 0
Gauge-like functions with f 0 = 0 are nonnegative.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma nonneg_of_isGaugeLike_f0_eq_zero {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf : IsGaugeLike f) (h0 : f 0 = 0) : ∀ x : Fin n → ℝ, (0 : EReal) ≤ f x
A positively homogeneous closed proper convex function has a closed-gauge power representation.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma exists_closedGauge_rpow_representation_of_posHomogeneous {n : ℕ} {p : ℝ}
{f : (Fin n → ℝ) → EReal} (hp : 1 < p)
(hf_closed : ClosedConvexFunction f)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) f)
(hf_hom : PositivelyHomogeneousOfDegree (n := n) p f) :
∃ k : (Fin n → ℝ) → EReal, IsClosedGauge k ∧
f = fun x => ((1 / p : ℝ) : EReal) * phiPow p (k x)
If a profile φ scales on nonnegative inputs, then (1/p) * φ ∘ k is positively homogeneous of degree p.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma posHomogeneous_of_exists_closedGauge_scaled {n : ℕ} {p : ℝ}
{f : (Fin n → ℝ) → EReal} {φ : EReal → EReal}
(hφ : ∀ {z : EReal} {t : ℝ}, 0 ≤ z → 0 < t →
φ (((t : ℝ) : EReal) * z) = ((Real.rpow t p : ℝ) : EReal) * φ z) :
(∃ k : (Fin n → ℝ) → EReal, IsClosedGauge k ∧
f = fun x => ((1 / p : ℝ) : EReal) * φ (k x)) →
PositivelyHomogeneousOfDegree (n := n) p f
The conjugate of a closed-gauge power profile is a power profile with conjugate exponent.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma fenchelConjugate_rpow_closedGauge_formula_and_posHomogeneous {n : ℕ} {p q : ℝ}
(hp : 1 < p) (hpq : (1 / p) + (1 / q) = 1) {f : (Fin n → ℝ) → EReal}
{k : (Fin n → ℝ) → EReal} (hk : IsClosedGauge k)
(hfk : f = fun x => ((1 / p : ℝ) : EReal) * phiPow p (k x)) :
PositivelyHomogeneousOfDegree (n := n) q (fenchelConjugate n f) ∧
∀ xStar : Fin n → ℝ,
fenchelConjugate n f xStar = ((1 / q : ℝ) : EReal) * phiPow q (polarGauge k xStar)
Triangle inequality for affine combinations.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma section15_abs_combo_le (a b t : ℝ) (ht0 : 0 ≤ t) (ht1 : 0 ≤ 1 - t) :
Convexity inequality for |·|^p with p ≥ 1.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma section15_abs_rpow_combo_le {p : ℝ} (hp : 1 ≤ p) (a b t : ℝ)
(ht0 : 0 ≤ t) (ht1 : 0 ≤ 1 - t) :
Real.rpow (|((1 - t) * a + t * b)|) p ≤
(1 - t) * Real.rpow (|a|) p + t * Real.rpow (|b|) p
Convexity of the (1/p) * ∑ |xᵢ|^p function.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma section15_convexFunction_sum_abs_rpow_div {n : ℕ} {p : ℝ} (hp : 1 < p) :
ConvexFunction (fun x : Fin n → ℝ =>
(((1 / p : ℝ) * ∑ i : Fin n, Real.rpow (|x i|) p) : EReal))
Lower semicontinuity of the (1/p) * ∑ |xᵢ|^p function.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma section15_lowerSemicontinuous_sum_abs_rpow_div {n : ℕ} {p : ℝ} (hp : 1 < p) :
LowerSemicontinuous (fun x : Fin n → ℝ =>
(((1 / p : ℝ) * ∑ i : Fin n, Real.rpow (|x i|) p) : EReal))
Positive homogeneity of the (1/p) * ∑ |xᵢ|^p function.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma section15_posHomogeneous_sum_abs_rpow_div {n : ℕ} {p : ℝ} :
PositivelyHomogeneousOfDegree (n := n) p (fun x : Fin n → ℝ =>
(((1 / p : ℝ) * ∑ i : Fin n, Real.rpow (|x i|) p) : EReal))
Text 15.0.22: Let 1 < p < ∞ and define f : ℝⁿ → ℝ by f x = (1/p) * (|ξ₁|^p + ... + |ξₙ|^p) for x = (ξ₁, ..., ξₙ). Then f is a closed proper convex function and is positively homogeneous of degree p. In this development, ℝⁿ is Fin n → ℝ, f is treated as EReal-valued by coercion, closedness is ClosedConvexFunction, properness is ProperConvexFunctionOn univ, and |ξᵢ|^p is Real.rpow |ξᵢ| p.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
theorem sum_abs_rpow_div_closedProperConvex_posHomogeneous {n : ℕ} {p : ℝ} (hp : 1 < p) :
let f : (Fin n → ℝ) → EReal :=
fun x => (((1 / p : ℝ) * ∑ i : Fin n, Real.rpow (|x i|) p) : EReal)
ClosedConvexFunction f ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) f ∧
PositivelyHomogeneousOfDegree (n := n) p f
Conjugate-exponent data from 1 / p + 1 / q = 1.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma section15_holderConjugate_of_hpq {p q : ℝ} (hp : 1 < p) (hpq : (1 / p) + (1 / q) = 1) :
p.HolderConjugate q
Young's inequality in a rearranged single-term form.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma section15_young_term_bound {p q a t : ℝ} (hpq : p.HolderConjugate q) :
a * t - (1 / p) * Real.rpow (|t|) p ≤ (1 / q) * Real.rpow (|a|) q
Coordinatewise Young bounds summed over Fin n.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma section15_dot_sub_sum_abs_rpow_le {n : ℕ} {p q : ℝ} (hpq : p.HolderConjugate q)
(xStar x : Fin n → ℝ) :
(∑ i : Fin n, x i * xStar i) - (1 / p) * ∑ i : Fin n, Real.rpow (|x i|) p ≤
(1 / q) * ∑ i : Fin n, Real.rpow (|xStar i|) q
Multiplying by |Real.sign r| does not change |r|^a when a > 0.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma section15_abs_sign_mul_rpow_eq {r a : ℝ} (ha : 0 < a) :
Evaluation of the Fenchel expression at the explicit maximizer.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma section15_eval_at_explicit_maximizer {n : ℕ} {p q : ℝ} (hpq : p.HolderConjugate q)
(xStar : Fin n → ℝ) :
let x0 : Fin n → ℝ := fun i => Real.sign (xStar i) * Real.rpow (|xStar i|) (q - 1)
((x0 ⬝ᵥ xStar : ℝ) : EReal) -
(((1 / p : ℝ) * ∑ i : Fin n, Real.rpow (|x0 i|) p) : EReal) =
(((1 / q : ℝ) * ∑ i : Fin n, Real.rpow (|xStar i|) q) : EReal)
Text 15.0.23: Let 1 < p < ∞ and define f : ℝⁿ → ℝ by f x = (1/p) * (|ξ₁|^p + ... + |ξₙ|^p) for x = (ξ₁, ..., ξₙ). Its conjugate is f^*(x^*) = (1/q) * (|ξ₁^*|^q + ... + |ξₙ^*|^q), where 1/p + 1/q = 1. In this development, ℝⁿ is Fin n → ℝ, |ξᵢ|^p is Real.rpow |ξᵢ| p, and the conjugate is modeled as fenchelConjugate n f.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
theorem fenchelConjugate_sum_abs_rpow_div_eq_sum_abs_rpow_div {n : ℕ} {p q : ℝ} (hp : 1 < p)
(hpq : (1 / p) + (1 / q) = 1) :
let f : (Fin n → ℝ) → EReal :=
fun x => (((1 / p : ℝ) * ∑ i : Fin n, Real.rpow (|x i|) p) : EReal)
fenchelConjugate n f =
fun xStar => (((1 / q : ℝ) * ∑ i : Fin n, Real.rpow (|xStar i|) q) : EReal)
Corollary 15.3.1: Let 1 < p < ∞ and let q be its conjugate exponent, 1/p + 1/q = 1. For a closed proper convex function f, positive homogeneity of degree p is equivalent to the existence of a closed gauge k such that f(x) = (1/p) * k(x)^p. In this case, the Fenchel conjugate f^* (modeled as fenchelConjugate n f) is positively homogeneous of degree q and satisfies f^*(x^*) = (1/q) * (k^∘(x^*))^q, where k^∘ is the polar gauge (modeled as polarGauge k).
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
theorem closedProperConvex_posHomogeneous_degree_iff_exists_closedGauge_rpow {n : ℕ}
{p q : ℝ} (hp : 1 < p) (hpq : (1 / p) + (1 / q) = 1) {f : (Fin n → ℝ) → EReal}
(hf_closed : ClosedConvexFunction f)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) f) :
let φp : EReal → EReal := fun z =>
if z = ⊥ then (0 : EReal) else if z = ⊤ then (⊤ : EReal) else (Real.rpow (max z.toReal 0) p : EReal)
let φq : EReal → EReal := fun z =>
if z = ⊥ then (0 : EReal) else if z = ⊤ then (⊤ : EReal) else (Real.rpow (max z.toReal 0) q : EReal)
(PositivelyHomogeneousOfDegree (n := n) p f ↔
∃ k : (Fin n → ℝ) → EReal, IsClosedGauge k ∧
f = fun x => ((1 / p : ℝ) : EReal) * φp (k x)) ∧
(∀ k : (Fin n → ℝ) → EReal,
IsClosedGauge k →
f = (fun x => ((1 / p : ℝ) : EReal) * φp (k x)) →
PositivelyHomogeneousOfDegree (n := n) q (fenchelConjugate n f) ∧
∀ xStar : Fin n → ℝ,
fenchelConjugate n f xStar = ((1 / q : ℝ) : EReal) * φq (polarGauge k xStar))§15 — Polars of Convex Functions — Part 13
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
phiPow is inverted by the reciprocal exponent on nonnegative inputs.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma phiPow_one_div_comp_phiPow_of_nonneg {p : ℝ} (hp : 0 < p) {z : EReal} (hz : 0 ≤ z) :
phiPow (1 / p) (phiPow p z) = z
phiPow with exponent p inverts phiPow (1/p) on nonnegative inputs.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma phiPow_comp_phiPow_one_div_of_nonneg {p : ℝ} (hp : 0 < p) {z : EReal} (hz : 0 ≤ z) :
phiPow p (phiPow (1 / p) z) = z
The base sublevel of f equals the unit sublevel of k := (p f)^{1/p} under nonnegativity.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma sublevel_f_one_div_eq_unitSublevel_k {n : ℕ} {p : ℝ} {f : (Fin n → ℝ) → EReal}
(hp : 0 < p) (hnonneg : ∀ x, (0 : EReal) ≤ f x) :
{x | f x ≤ ((1 / p : ℝ) : EReal)} =
{x | phiPow (1 / p) (((p : ℝ) : EReal) * f x) ≤ (1 : EReal)}
The unit sublevel of a gauge corresponds to its polar set.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma unitSublevel_polarGauge_eq_polarSet {n : ℕ} {k : (Fin n → ℝ) → EReal} (hk : IsGauge k) :
let C : Set (Fin n → ℝ) := {x | k x ≤ (1 : EReal)}
{xStar | polarGauge k xStar ≤ (1 : EReal)} =
{xStar | ∀ x ∈ C, (x ⬝ᵥ xStar : ℝ) ≤ 1}§15 — Polars of Convex Functions — Part 14
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
The unit sublevel of a closed gauge is polar to the unit sublevel of its polar gauge.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma unitSublevel_polarGauge_eq_polarSet_and_bipolar {n : ℕ} {k : (Fin n → ℝ) → EReal}
(hk : IsClosedGauge k) :
let C : Set (Fin n → ℝ) := {x | k x ≤ (1 : EReal)}
let CStar : Set (Fin n → ℝ) := {xStar | polarGauge k xStar ≤ (1 : EReal)}
CStar = {xStar | ∀ x ∈ C, (x ⬝ᵥ xStar : ℝ) ≤ 1} ∧
C = {x | ∀ xStar ∈ CStar, (x ⬝ᵥ xStar : ℝ) ≤ 1}
Recover (p f)^{1/p} and (q f^*)^{1/q} from the power representation.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
lemma k_eq_closedGauge_from_cor1531 {n : ℕ} {p q : ℝ} (hp : 1 < p)
(hpq : (1 / p) + (1 / q) = 1) {f : (Fin n → ℝ) → EReal}
{k0 : (Fin n → ℝ) → EReal} (hk0 : IsClosedGauge k0)
(hfk0 : f = fun x => ((1 / p : ℝ) : EReal) * phiPow p (k0 x))
(hfc : ∀ xStar : Fin n → ℝ,
fenchelConjugate n f xStar =
((1 / q : ℝ) : EReal) * phiPow q (polarGauge k0 xStar)) :
(fun x => phiPow (1 / p) (((p : ℝ) : EReal) * f x)) = k0 ∧
(fun xStar =>
phiPow (1 / q) (((q : ℝ) : EReal) * fenchelConjugate n f xStar)) = polarGauge k0§15 — Polars of Convex Functions — Part 15
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
The ℓᵖ gauge as an EReal-valued function.
open scoped BigOperators
open scoped Pointwise
noncomputable def lpNormEReal {n : ℕ} (p : ℝ) (x : Fin n → ℝ) : EReal :=
((Real.rpow (∑ i : Fin n, Real.rpow (|x i|) p) (1 / p) : ℝ) : EReal)
lpNormEReal is the phiPow of the ℓ^p sum.
open scoped BigOperators
open scoped Pointwise
lemma lpNormEReal_eq_phiPow_sum_abs_rpow {n : ℕ} (p : ℝ) (x : Fin n → ℝ) :
lpNormEReal (n := n) p x =
phiPow (1 / p) ((∑ i : Fin n, Real.rpow (|x i|) p : ℝ) : EReal)
lpNormEReal is even.
open scoped BigOperators
open scoped Pointwise
lemma lpNormEReal_symm {n : ℕ} (p : ℝ) (x : Fin n → ℝ) :
lpNormEReal (n := n) p (-x) = lpNormEReal (n := n) p x
lpNormEReal is strictly positive away from the origin.
open scoped BigOperators
open scoped Pointwise
lemma lpNormEReal_pos_of_ne_zero {n : ℕ} {p : ℝ} (x : Fin n → ℝ) (hx : x ≠ 0) :
(0 : EReal) < lpNormEReal (n := n) p x
Closed gauge and polar gauge for power profiles of closed proper convex functions.
open scoped BigOperators
open scoped Pointwise
lemma closedGauge_rpow_mul_and_polar_eq_of_closedProperConvex_posHomogeneous {n : ℕ} {p q : ℝ}
(hp : 1 < p) (hpq : (1 / p) + (1 / q) = 1) {f : (Fin n → ℝ) → EReal}
(hf_closed : ClosedConvexFunction f)
(hf_proper : ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) f)
(hf_hom : PositivelyHomogeneousOfDegree (n := n) p f) :
let k : (Fin n → ℝ) → EReal := fun x => phiPow (1 / p) (((p : ℝ) : EReal) * f x)
let kStar : (Fin n → ℝ) → EReal :=
fun xStar => phiPow (1 / q) (((q : ℝ) : EReal) * fenchelConjugate n f xStar)
IsClosedGauge k ∧ polarGauge k = kStar
The ℓᵖ gauge is closed and its polar gauge is the ℓᵠ gauge.
open scoped BigOperators
open scoped Pointwise
lemma lpNormEReal_isClosedGauge_and_polarGauge_eq {n : ℕ} {p q : ℝ} (hp : 1 < p)
(hpq : (1 / p) + (1 / q) = 1) :
IsClosedGauge (lpNormEReal (n := n) p) ∧
polarGauge (lpNormEReal (n := n) p) = lpNormEReal (n := n) q
A gauge instance for lpNormEReal upgrades to a norm gauge.
open scoped BigOperators
open scoped Pointwise
lemma lpNormEReal_isNormGauge_of_isGauge {n : ℕ} {p : ℝ}
(hg : IsGauge (lpNormEReal (n := n) p)) : IsNormGauge (lpNormEReal (n := n) p)
A closed gauge is fixed by the polar gauge involution.
open scoped BigOperators
open scoped Pointwise
lemma lpNormEReal_polarGauge_eq_self_of_closedGauge {n : ℕ} {k kPolar : (Fin n → ℝ) → EReal}
(hk : IsClosedGauge k) (hpol : polarGauge k = kPolar) : polarGauge kPolar = k§15 — Polars of Convex Functions — Part 16
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
open scoped BigOperators
open scoped Pointwise
theorem lpNormEReal_isClosedGauge_and_polar {n : ℕ} {p q : ℝ} (hp : 1 < p)
(hpq : (1 / p) + (1 / q) = 1) :
let k : (Fin n → ℝ) → EReal := lpNormEReal (n := n) p
let kPolar : (Fin n → ℝ) → EReal := lpNormEReal (n := n) q
IsClosedGauge k ∧ polarGauge k = kPolar ∧ polarGauge kPolar = k ∧ IsNormGauge k ∧ IsNormGauge kPolar
Rewrite the quadratic form using quadraticHalfLinear and Matrix.toLin'.
open scoped BigOperators
open scoped Pointwise
lemma matrix_quadraticHalf_eq_quadraticHalfLinear {n : ℕ} (Q : Matrix (Fin n) (Fin n) ℝ) :
(fun x => (((1 / 2 : ℝ) * (x ⬝ᵥ (Q.mulVec x))) : EReal)) =
quadraticHalfLinear n (Matrix.toLin' Q)
The quadratic form is positively homogeneous of degree 2.
open scoped BigOperators
open scoped Pointwise
lemma matrix_quadratic_posHomogeneous {n : ℕ} (Q : Matrix (Fin n) (Fin n) ℝ) :
PositivelyHomogeneousOfDegree (n := n) 2
(fun x => (((1 / 2 : ℝ) * (x ⬝ᵥ (Q.mulVec x))) : EReal))
The quadratic form from a positive definite matrix is closed, convex, and proper.
open scoped BigOperators
open scoped Pointwise
lemma matrix_quadratic_closedConvex_and_proper {n : ℕ} (Q : Matrix (Fin n) (Fin n) ℝ)
(hQpos : Matrix.PosDef Q) :
ClosedConvexFunction (fun x => (((1 / 2 : ℝ) * (x ⬝ᵥ (Q.mulVec x))) : EReal)) ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ))
(fun x => (((1 / 2 : ℝ) * (x ⬝ᵥ (Q.mulVec x))) : EReal))
The Fenchel conjugate of the quadratic form under positive definiteness.
open scoped BigOperators
open scoped Pointwise
lemma fenchelConjugate_matrix_quadraticHalf_of_posDef {n : ℕ}
(Q : Matrix (Fin n) (Fin n) ℝ) (hQsymm : Q.IsSymm) (hQpos : Matrix.PosDef Q) :
fenchelConjugate n (fun x => (((1 / 2 : ℝ) * (x ⬝ᵥ (Q.mulVec x))) : EReal)) =
fun xStar => (((1 / 2 : ℝ) * (xStar ⬝ᵥ ((Q⁻¹).mulVec xStar))) : EReal)
Text 15.0.25: Let Q be a symmetric positive definite n × n matrix and define f(x) = (1/2) * ⟪x, Q x⟫. Then f is a closed proper convex function on ℝⁿ and its Fenchel conjugate satisfies f^*(x⋆) = (1/2) * ⟪x⋆, Q⁻¹ x⋆⟫. Since f is positively homogeneous of degree 2, Corollary 15.3.2 implies that k(x) = ⟪x, Q x⟫^{1/2} is a closed gauge (in fact, a norm) whose polar is k^∘(x⋆) = ⟪x⋆, Q⁻¹ x⋆⟫^{1/2}.
open scoped BigOperators
open scoped Pointwise
theorem matrix_quadratic_closedProperConvex_and_fenchelConjugate {n : ℕ}
(Q : Matrix (Fin n) (Fin n) ℝ) (hQsymm : Q.IsSymm) (hQpos : Matrix.PosDef Q) :
let f : (Fin n → ℝ) → EReal := fun x => (((1 / 2 : ℝ) * (x ⬝ᵥ (Q.mulVec x))) : EReal)
ClosedConvexFunction f ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) f ∧
fenchelConjugate n f =
fun xStar => (((1 / 2 : ℝ) * (xStar ⬝ᵥ ((Q⁻¹).mulVec xStar))) : EReal)
Text 15.0.25 (gauge/polar): with Q symmetric positive definite, the map k(x) = ⟪x, Q x⟫^{1/2} is a closed norm gauge and kᵒ(x⋆) = ⟪x⋆, Q⁻¹ x⋆⟫^{1/2}.
open scoped BigOperators
open scoped Pointwise
theorem matrix_quadraticGauge_isClosedGauge_isNormGauge_and_polar {n : ℕ}
(Q : Matrix (Fin n) (Fin n) ℝ) (hQsymm : Q.IsSymm) (hQpos : Matrix.PosDef Q) :
let k : (Fin n → ℝ) → EReal := fun x => ((Real.sqrt (x ⬝ᵥ (Q.mulVec x)) : ℝ) : EReal)
let kStar : (Fin n → ℝ) → EReal :=
fun xStar => ((Real.sqrt (xStar ⬝ᵥ ((Q⁻¹).mulVec xStar)) : ℝ) : EReal)
IsClosedGauge k ∧ IsNormGauge k ∧ polarGauge k = kStar
The quadratic sublevel set agrees with the unit sublevel of its square-root gauge.
open scoped BigOperators
open scoped Pointwise
lemma matrix_quadraticSublevel_eq_unitSublevel_sqrt {n : ℕ} (Q : Matrix (Fin n) (Fin n) ℝ) :
{x | x ⬝ᵥ (Q.mulVec x) ≤ (1 : ℝ)} =
{x | ((Real.sqrt (x ⬝ᵥ (Q.mulVec x)) : ℝ) : EReal) ≤ (1 : EReal)}
The inverse quadratic sublevel set agrees with the unit sublevel of its square-root gauge.
open scoped BigOperators
open scoped Pointwise
lemma matrix_quadraticInvSublevel_eq_unitSublevel_sqrt {n : ℕ} (Q : Matrix (Fin n) (Fin n) ℝ) :
{xStar | ((Real.sqrt (xStar ⬝ᵥ ((Q⁻¹).mulVec xStar)) : ℝ) : EReal) ≤ (1 : EReal)} =
{xStar | xStar ⬝ᵥ ((Q⁻¹).mulVec xStar) ≤ (1 : ℝ)}
Text 15.0.26: Let Q be a symmetric positive definite n × n matrix. Define the convex set C = {x | ⟪x, Q x⟫ ≤ 1}. Then its polar (with respect to the pairing ⟪x, x⋆⟫) is C^∘ = {x⋆ | ⟪x⋆, Q⁻¹ x⋆⟫ ≤ 1}.
open scoped BigOperators
open scoped Pointwise
theorem matrix_quadraticSublevel_polar {n : ℕ} (Q : Matrix (Fin n) (Fin n) ℝ) (hQsymm : Q.IsSymm)
(hQpos : Matrix.PosDef Q) :
let C : Set (Fin n → ℝ) := {x | x ⬝ᵥ (Q.mulVec x) ≤ (1 : ℝ)}
{xStar | ∀ x ∈ C, (x ⬝ᵥ xStar : ℝ) ≤ 1} = {xStar | xStar ⬝ᵥ ((Q⁻¹).mulVec xStar) ≤ (1 : ℝ)}
Ring.inverse on a unit-valued function agrees with pointwise inversion.
open scoped BigOperators
open scoped Pointwise
lemma ring_inverse_apply_of_isUnit {ι : Type*} (v : ι → ℝ) (hv : IsUnit v) (i : ι) :
Ring.inverse v i = (v i)⁻¹
Quadratic form of a diagonal matrix on Fin 2.
open scoped BigOperators
open scoped Pointwise
lemma polar_ellipticDisk_quadraticForm_diagonal_fun_fin2 (v : Fin 2 → ℝ) (x : Fin 2 → ℝ) :
x ⬝ᵥ ((Matrix.diagonal v).mulVec x) = v 0 * (x 0)^2 + v 1 * (x 1)^2
The diagonal matrix for the elliptic disk is positive definite.
open scoped BigOperators
open scoped Pointwise
lemma polar_ellipticDisk_posDef_diagonal {α₁ α₂ : ℝ} (hα₁ : α₁ ≠ 0) (hα₂ : α₂ ≠ 0) :
Matrix.PosDef (Matrix.diagonal ![(α₁ ^ 2)⁻¹, (α₂ ^ 2)⁻¹])
Quadratic form for the inverse diagonal matrix in the elliptic disk.
open scoped BigOperators
open scoped Pointwise
lemma polar_ellipticDisk_quadraticForm_inv_diagonal_fin2 {α₁ α₂ : ℝ} (hα₁ : α₁ ≠ 0)
(hα₂ : α₂ ≠ 0) (x : Fin 2 → ℝ) :
let Q : Matrix (Fin 2) (Fin 2) ℝ := Matrix.diagonal ![(α₁ ^ 2)⁻¹, (α₂ ^ 2)⁻¹]
x ⬝ᵥ ((Q⁻¹).mulVec x) = α₁ ^ 2 * (x 0)^2 + α₂ ^ 2 * (x 1)^2
Text 15.0.27: Let C ⊆ ℝ² be the elliptic disk C = {(ξ₁, ξ₂) | (ξ₁^2 / α₁^2) + (ξ₂^2 / α₂^2) ≤ 1}. Then its polar set is the elliptic disk C^∘ = {(ξ₁⋆, ξ₂⋆) | α₁^2 * (ξ₁⋆)^2 + α₂^2 * (ξ₂⋆)^2 ≤ 1}.
open scoped BigOperators
open scoped Pointwise
theorem polar_ellipticDisk {α₁ α₂ : ℝ} (hα₁ : α₁ ≠ 0) (hα₂ : α₂ ≠ 0) :
let C : Set (Fin 2 → ℝ) :=
{x | (x 0) ^ 2 / α₁ ^ 2 + (x 1) ^ 2 / α₂ ^ 2 ≤ (1 : ℝ)}
{xStar | ∀ x ∈ C, (x ⬝ᵥ xStar : ℝ) ≤ 1} =
{xStar | α₁ ^ 2 * (xStar 0) ^ 2 + α₂ ^ 2 * (xStar 1) ^ 2 ≤ (1 : ℝ)}
Convexity of the composition g ∘ k for the quadratic norm gauge.
open scoped BigOperators
open scoped Pointwise
lemma matrix_quadraticGauge_comp_convex {n : ℕ}
(Q : Matrix (Fin n) (Fin n) ℝ) (hQsymm : Q.IsSymm) (hQpos : Matrix.PosDef Q)
(g : EReal → EReal)
(hgζ : ∃ ζ : ℝ, 0 < ζ ∧ g (ζ : EReal) ≠ ⊤ ∧ g (ζ : EReal) ≠ ⊥)
(hg_mono : Monotone g)
(hg_convex :
ConvexFunctionOn (S := {t : Fin 1 → ℝ | 0 ≤ t 0}) (fun t => g (t 0 : EReal)))
(hg_closed :
IsClosed (epigraph (S := {t : Fin 1 → ℝ | 0 ≤ t 0}) (fun t => g (t 0 : EReal)))) :
let k : (Fin n → ℝ) → EReal := fun x => ((Real.sqrt (x ⬝ᵥ (Q.mulVec x)) : ℝ) : EReal)
let f : (Fin n → ℝ) → EReal := fun x => g (k x)
ConvexFunction f
Lower semicontinuity of g ∘ k from the closed epigraph of g.
open scoped BigOperators
open scoped Pointwise
lemma matrix_quadraticGauge_comp_lowerSemicontinuous {n : ℕ}
(Q : Matrix (Fin n) (Fin n) ℝ) (hQsymm : Q.IsSymm) (hQpos : Matrix.PosDef Q)
(g : EReal → EReal)
(hg_closed :
IsClosed (epigraph (S := {t : Fin 1 → ℝ | 0 ≤ t 0}) (fun t => g (t 0 : EReal)))) :
let k : (Fin n → ℝ) → EReal := fun x => ((Real.sqrt (x ⬝ᵥ (Q.mulVec x)) : ℝ) : EReal)
let f : (Fin n → ℝ) → EReal := fun x => g (k x)
LowerSemicontinuous f
Properness of g ∘ k once g is finite at the origin.
open scoped BigOperators
open scoped Pointwise
lemma matrix_quadraticGauge_comp_proper {n : ℕ}
(Q : Matrix (Fin n) (Fin n) ℝ) (hQsymm : Q.IsSymm) (hQpos : Matrix.PosDef Q)
(g : EReal → EReal)
(hgζ : ∃ ζ : ℝ, 0 < ζ ∧ g (ζ : EReal) ≠ ⊤ ∧ g (ζ : EReal) ≠ ⊥)
(hg_mono : Monotone g)
(hg_convex :
ConvexFunctionOn (S := {t : Fin 1 → ℝ | 0 ≤ t 0}) (fun t => g (t 0 : EReal)))
(hg_closed :
IsClosed (epigraph (S := {t : Fin 1 → ℝ | 0 ≤ t 0}) (fun t => g (t 0 : EReal)))) :
let k : (Fin n → ℝ) → EReal := fun x => ((Real.sqrt (x ⬝ᵥ (Q.mulVec x)) : ℝ) : EReal)
let f : (Fin n → ℝ) → EReal := fun x => g (k x)
ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) f
Conjugate of g ∘ k for the quadratic norm gauge.
open scoped BigOperators
open scoped Pointwise
lemma matrix_quadraticGauge_comp_fenchelConjugate {n : ℕ}
(Q : Matrix (Fin n) (Fin n) ℝ) (hQsymm : Q.IsSymm) (hQpos : Matrix.PosDef Q)
(g : EReal → EReal)
(hgζ : ∃ ζ : ℝ, 0 < ζ ∧ g (ζ : EReal) ≠ ⊤ ∧ g (ζ : EReal) ≠ ⊥)
(hg_mono : Monotone g) (hg_top : g ⊤ = ⊤) :
let k : (Fin n → ℝ) → EReal := fun x => ((Real.sqrt (x ⬝ᵥ (Q.mulVec x)) : ℝ) : EReal)
let f : (Fin n → ℝ) → EReal := fun x => g (k x)
let kStar : (Fin n → ℝ) → EReal :=
fun xStar => ((Real.sqrt (xStar ⬝ᵥ ((Q⁻¹).mulVec xStar)) : ℝ) : EReal)
fenchelConjugate n f = fun xStar => monotoneConjugateERealNonneg g (kStar xStar)
Text 15.0.28: Let Q be a symmetric positive definite matrix and define k(x) = ⟪x, Q x⟫^{1/2}. Let g : [0, +∞) → (-∞, +∞] satisfy the hypotheses of Theorem 15.3, and define f(x) = g(k(x)). Then f is a closed proper convex function and its conjugate satisfies f^*(x⋆) = g⁺(k^∘(x⋆)) = g⁺(⟪x⋆, Q⁻¹ x⋆⟫^{1/2}), where g⁺ is the monotone conjugate of g and k^∘ is the polar gauge of k.
open scoped BigOperators
open scoped Pointwise
theorem matrix_quadraticGauge_comp_closedProperConvex_and_fenchelConjugate {n : ℕ}
(Q : Matrix (Fin n) (Fin n) ℝ) (hQsymm : Q.IsSymm) (hQpos : Matrix.PosDef Q)
(g : EReal → EReal)
(hgζ : ∃ ζ : ℝ, 0 < ζ ∧ g (ζ : EReal) ≠ ⊤ ∧ g (ζ : EReal) ≠ ⊥)
(hg_mono : Monotone g)
(hg_convex :
ConvexFunctionOn (S := {t : Fin 1 → ℝ | 0 ≤ t 0}) (fun t => g (t 0 : EReal)))
(hg_closed :
IsClosed (epigraph (S := {t : Fin 1 → ℝ | 0 ≤ t 0}) (fun t => g (t 0 : EReal))))
(hg_top : g ⊤ = ⊤) :
let k : (Fin n → ℝ) → EReal := fun x => ((Real.sqrt (x ⬝ᵥ (Q.mulVec x)) : ℝ) : EReal)
let f : (Fin n → ℝ) → EReal := fun x => g (k x)
let kStar : (Fin n → ℝ) → EReal :=
fun xStar => ((Real.sqrt (xStar ⬝ᵥ ((Q⁻¹).mulVec xStar)) : ℝ) : EReal)
ClosedConvexFunction f ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) f ∧
fenchelConjugate n f =
fun xStar => monotoneConjugateERealNonneg g (kStar xStar)
Text 15.0.29: Let f : ℝⁿ → [0, +∞] be a convex function such that f 0 = 0. Its polar fᵒ : ℝⁿ → [0, +∞] is defined by fᵒ x⋆ = inf { μ⋆ ≥ 0 | ⟪x, x⋆⟫ ≤ 1 + μ⋆ * f x for all x }. If f is a gauge, this reduces to the polar gauge (Text 15.0.5). If f = δ(· | C) for a closed convex set C containing 0, then fᵒ = δ(· | Cᵒ). Furthermore, whenever x ∈ dom f and x⋆ ∈ dom fᵒ, one has ⟪x, x⋆⟫ ≤ 1 + f x * fᵒ x⋆. In this development, we represent [0, +∞] by EReal together with explicit nonnegativity assumptions, and effective-domain assumptions by f x ≠ ⊤.
open scoped BigOperators
open scoped Pointwise
noncomputable def polarConvex {n : ℕ} (f : (Fin n → ℝ) → EReal) (xStar : Fin n → ℝ) : EReal :=
sInf
{μStar : EReal |
0 ≤ μStar ∧
∀ x : Fin n → ℝ, ((x ⬝ᵥ xStar : ℝ) : EReal) ≤ (1 : EReal) + μStar * f x}
A polarConvex feasible multiplier is feasible for the polar gauge of a gauge.
open scoped BigOperators
open scoped Pointwise
lemma polarConvex_feasible_imp_polarGauge_feasible_of_isGauge {n : ℕ}
{f : (Fin n → ℝ) → EReal} (hf : IsGauge f) {xStar : Fin n → ℝ} {μ : EReal}
(hμ :
0 ≤ μ ∧
∀ x : Fin n → ℝ, ((x ⬝ᵥ xStar : ℝ) : EReal) ≤ (1 : EReal) + μ * f x) :
0 ≤ μ ∧ ∀ x : Fin n → ℝ, ((x ⬝ᵥ xStar : ℝ) : EReal) ≤ μ * f x
Text 15.0.29 (gauge case): if f is a gauge, then its polar (as in polarConvex) agrees with the polar gauge polarGauge.
open scoped BigOperators
open scoped Pointwise
theorem polarConvex_eq_polarGauge_of_isGauge {n : ℕ} {f : (Fin n → ℝ) → EReal} (hf : IsGauge f) :
polarConvex f = polarGauge f
Text 15.0.29 (indicator case): if f = δ(· | C) for a closed convex set C containing 0, then fᵒ = δ(· | Cᵒ) where Cᵒ = {x⋆ | ∀ x ∈ C, ⟪x, x⋆⟫ ≤ 1}.
open scoped BigOperators
open scoped Pointwise
theorem polarConvex_erealIndicator_eq_erealIndicator_polar {n : ℕ} (C : Set (Fin n → ℝ)) :
let Cpolar : Set (Fin n → ℝ) := {xStar | ∀ x ∈ C, ((x ⬝ᵥ xStar : ℝ) : EReal) ≤ (1 : EReal)}
polarConvex (erealIndicator C) = erealIndicator Cpolar
Text 15.0.29 (polar inequality): whenever x ∈ dom f and x⋆ ∈ dom fᵒ, ⟪x, x⋆⟫ ≤ 1 + f x * fᵒ x⋆.
open scoped BigOperators
open scoped Pointwise
theorem inner_le_one_add_mul_polarConvex {n : ℕ} (f : (Fin n → ℝ) → EReal) (x xStar : Fin n → ℝ)
(hx : f x ≠ ⊤) (hxStar : polarConvex f xStar ≠ ⊤) :
((x ⬝ᵥ xStar : ℝ) : EReal) ≤ (1 : EReal) + f x * polarConvex f xStar§15 — Polars of Convex Functions — Part 17
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
Text 15.0.30: Let f : ℝⁿ → [0, +∞] be a nonnegative closed convex function with f 0 = 0, and let fᵒ be its polar in the extended sense, defined by fᵒ x⋆ = inf { μ⋆ ≥ 0 | ⟪x, x⋆⟫ ≤ 1 + μ⋆ * f x for all x }. Then for all x ∈ dom f and x⋆ ∈ dom fᵒ, one has ⟪x, x⋆⟫ ≤ 1 + f x * fᵒ x⋆. In this development, ℝⁿ is Fin n → ℝ, fᵒ is polarConvex f, and effective-domain assumptions are modeled by f x ≠ ⊤ and polarConvex f x⋆ ≠ ⊤.
open scoped BigOperators
open scoped Pointwise
theorem inner_le_one_add_mul_polarConvex_of_nonneg_closedConvex_zero {n : ℕ}
(f : (Fin n → ℝ) → EReal) (x xStar : Fin n → ℝ) (hx : f x ≠ ⊤)
(hxStar : polarConvex f xStar ≠ ⊤) :
((x ⬝ᵥ xStar : ℝ) : EReal) ≤ (1 : EReal) + f x * polarConvex f xStar§15 — Polars of Convex Functions — Part 18
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
Fenchel–Young inequality from the conjugate definition, using f 0 = 0.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma dotProduct_le_add_fenchelConjugate {n : ℕ} (f : (Fin n → ℝ) → EReal)
(hf_nonneg : ∀ x, 0 ≤ f x) (hf0 : f 0 = 0) :
∀ x y : Fin n → ℝ,
((x ⬝ᵥ y : ℝ) : EReal) ≤ f x + fenchelConjugate n f y
Nonnegativity and f 0 = 0 give properness on univ for a closed convex function.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma properConvexFunctionOn_of_nonneg_closedConvex_zero {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0) :
ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) f
Biconjugation reduces to the original function under closedness and nonnegativity.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma fenchelConjugate_biconjugate_eq_of_nonneg_closedConvex_zero {n : ℕ}
(f : (Fin n → ℝ) → EReal) (hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f)
(hf0 : f 0 = 0) :
fenchelConjugate n (fenchelConjugate n f) = f
Nonnegativity and f 0 = 0 make the Fenchel conjugate nonnegative.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma fenchelConjugate_nonneg_of_nonneg_zero {n : ℕ} (f : (Fin n → ℝ) → EReal) (hf0 : f 0 = 0) :
∀ y : Fin n → ℝ, (0 : EReal) ≤ fenchelConjugate n f y
Dilation feasibility yields the polar inequality for a positive scale.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma dilation_le_one_pos_imp_polar_feasible {n : ℕ} (f : (Fin n → ℝ) → EReal)
(hf_nonneg : ∀ x, 0 ≤ f x) (hf0 : f 0 = 0) {x : Fin n → ℝ} {t : ℝ} (ht : 0 < t)
(hle : (t : EReal) * f ((t⁻¹) • x) ≤ (1 : EReal)) :
∀ y, ((y ⬝ᵥ x : ℝ) : EReal) ≤ (1 : EReal) + (t : EReal) * fenchelConjugate n f y
Polar feasibility implies the dilation inequality for a positive scale.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma polar_feasible_imp_dilation_le_one_pos {n : ℕ} (f : (Fin n → ℝ) → EReal)
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0)
{x : Fin n → ℝ} {t : ℝ} (ht : 0 < t)
(hfeas :
∀ y, ((y ⬝ᵥ x : ℝ) : EReal) ≤ (1 : EReal) + (t : EReal) * fenchelConjugate n f y) :
(t : EReal) * f ((t⁻¹) • x) ≤ (1 : EReal)
Nonzero vectors admit a dual witness with dot product exceeding 1.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma exists_dotProduct_gt_one_of_ne_zero {n : ℕ} {x : Fin n → ℝ} (hx : x ≠ 0) :
∃ y : Fin n → ℝ, (1 : ℝ) < (y ⬝ᵥ x : ℝ)
The infimum of positive real coefficients in EReal is zero.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma sInf_pos_real_eq_zero :
let S : Set EReal := {μ | ∃ t : ℝ, 0 < t ∧ μ = (t : EReal)}
sInf S = (0 : EReal)
Text 15.0.31: Let f : ℝⁿ → [0, +∞] be a nonnegative closed convex function with f 0 = 0, and define g := f^{*∘}. For λ > 0, define the dilation f_λ x := λ * f (λ⁻¹ • x). Then for every x ∈ ℝⁿ one has the representation g x = inf { λ > 0 | f_λ x ≤ 1 }. In this development, we model [0, +∞] by EReal together with nonnegativity assumptions, define g as polarConvex (fenchelConjugate n f), and represent the infimum as sInf in EReal.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
theorem polarFenchelConjugate_eq_sInf_dilation_le_one {n : ℕ} (f : (Fin n → ℝ) → EReal)
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0) :
let g : (Fin n → ℝ) → EReal := polarConvex (fenchelConjugate n f)
let fDil : ℝ → (Fin n → ℝ) → EReal := fun t x => (t : EReal) * f ((t⁻¹) • x)
∀ x : Fin n → ℝ,
g x =
sInf {μ : EReal | ∃ t : ℝ, 0 < t ∧ μ = (t : EReal) ∧ fDil t x ≤ (1 : EReal)}
Text 15.0.32: Let f : ℝⁿ → [0, +∞] be a nonnegative closed convex function with f(0) = 0. For λ > 0, define the scaled (perspective) function f_λ x := λ * f (x / λ). The obverse of f is the function g : ℝⁿ → [0, +∞] given by g x := inf {λ > 0 | f_λ x ≤ 1}. If f = δ(· | C) for a closed convex set C containing 0, then g = γ(· | C) is the gauge of C. If f = γ(· | C) is the gauge of such a set C, then g = δ(· | C). Thus the gauge and indicator functions of C are obverses of each other. In this development, ℝⁿ is Fin n → ℝ, [0, +∞] is modeled by EReal with nonnegativity assumptions, δ(· | C) is erealIndicator C, and γ(· | C) is represented by fun x => (gaugeFunction C x : EReal).
open scoped BigOperators
open scoped Pointwise
open scoped Topology
noncomputable def obverseConvex {n : ℕ} (f : (Fin n → ℝ) → EReal) : (Fin n → ℝ) → EReal :=
fun x =>
sInf
{μ : EReal |
∃ t : ℝ,
0 < t ∧
μ = (t : EReal) ∧
(t : EReal) * f ((t⁻¹) • x) ≤ (1 : EReal)}
The obverse constraint for an indicator reduces to scaled membership.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma obverseConvex_erealIndicator_apply_simp {n : ℕ} {C : Set (Fin n → ℝ)} {t : ℝ}
(ht : 0 < t) (x : Fin n → ℝ) :
((t : EReal) * erealIndicator C ((t⁻¹) • x) ≤ (1 : EReal)) ↔ x ∈ t • C
Text 15.0.32 (indicator case): if f = δ(· | C) for a closed convex absorbing set C containing 0, then its obverse is γ(· | C).
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma obverseConvex_erealIndicator_eq_gaugeFunction {n : ℕ} (C : Set (Fin n → ℝ))
(h0C : (0 : Fin n → ℝ) ∈ C)
(hCabs : ∀ x : Fin n → ℝ, ∃ lam : ℝ, 0 ≤ lam ∧ x ∈ (fun y => lam • y) '' C) :
obverseConvex (erealIndicator C) = fun x => (gaugeFunction C x : EReal)
Text 15.0.32 (gauge case): if f = γ(· | C) for a closed convex absorbing set C containing 0, then its obverse is δ(· | C).
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma obverseConvex_gaugeFunction_eq_erealIndicator {n : ℕ} (C : Set (Fin n → ℝ))
(hC_closed : IsClosed C) (hC_conv : Convex ℝ C) (h0C : (0 : Fin n → ℝ) ∈ C)
(hCabs : ∀ x : Fin n → ℝ, ∃ lam : ℝ, 0 ≤ lam ∧ x ∈ (fun y => lam • y) '' C) :
obverseConvex (fun x => (gaugeFunction C x : EReal)) = erealIndicator C
The obverse sInf set is the image of its positive real parameters.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma obverseConvex_set_eq_image_pos {n : ℕ} (f : (Fin n → ℝ) → EReal) (x : Fin n → ℝ) :
{μ : EReal |
∃ t : ℝ,
0 < t ∧ μ = (t : EReal) ∧
(t : EReal) * f ((t⁻¹) • x) ≤ (1 : EReal)} =
(fun t : ℝ => (t : EReal)) ''
{t : ℝ | 0 < t ∧ (t : EReal) * f ((t⁻¹) • x) ≤ (1 : EReal)}
Feasibility at a positive scalar gives an upper bound for the obverse.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma obverseConvex_le_coe_pos_of_perspective_le_one {n : ℕ} (f : (Fin n → ℝ) → EReal)
{t : ℝ} (ht : 0 < t) (x : Fin n → ℝ)
(hle : (t : EReal) * f ((t⁻¹) • x) ≤ (1 : EReal)) :
obverseConvex f x ≤ (t : EReal)
The obverse is always nonnegative.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma obverseConvex_nonneg {n : ℕ} (f : (Fin n → ℝ) → EReal) (x : Fin n → ℝ) :
(0 : EReal) ≤ obverseConvex f x
An element below a strict upper bound for sInf exists in a nonempty real set.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma exists_lt_of_sInf_lt {S : Set ℝ} (hne : S.Nonempty) {a : ℝ} (h : sInf S < a) :
∃ s ∈ S, s < a
For positive t, scaling inequality is equivalent to a reciprocal bound.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma mul_le_one_iff_le_inv_pos {t : ℝ} {a : EReal} (ht : 0 < t) (ha : 0 ≤ a) :
(t : EReal) * a ≤ (1 : EReal) ↔ a ≤ (t⁻¹ : EReal)
Feasibility at a smaller scale implies feasibility at a larger scale.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma perspective_isUpperSet {n : ℕ} (f : (Fin n → ℝ) → EReal)
(hf_nonneg : ∀ x, 0 ≤ f x)
(hf_conv : ConvexFunctionOn (S := (Set.univ : Set (Fin n → ℝ))) f) (hf0 : f 0 = 0)
{x : Fin n → ℝ} {s t : ℝ} (hs : 0 < s) (hst : s ≤ t)
(hle : (s : EReal) * f ((s⁻¹) • x) ≤ (1 : EReal)) :
(t : EReal) * f ((t⁻¹) • x) ≤ (1 : EReal)
Obverse at a positive scalar equals the perspective sublevel inequality.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma obverseConvex_le_coe_pos_iff_perspective_le_one {n : ℕ} (f : (Fin n → ℝ) → EReal)
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0)
{x : Fin n → ℝ} {lam : ℝ} (hlam : 0 < lam) :
obverseConvex f x ≤ (lam : EReal) ↔ (lam : EReal) * f ((lam⁻¹) • x) ≤ (1 : EReal)§15 — Polars of Convex Functions — Part 19
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
The obverse at 0 equals the recession function sublevel.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma obverseConvex_le_zero_iff_recessionFunction_le_one {n : ℕ} (f : (Fin n → ℝ) → EReal)
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0)
(x : Fin n → ℝ) :
obverseConvex f x ≤ (0 : EReal) ↔ recessionFunction f x ≤ (1 : EReal)
Text 15.0.33: Let f : ℝⁿ → [0, +∞] and let g be the function defined by the obverse formula g x = inf {λ > 0 | f_λ x ≤ 1}, where f_λ x := λ * f (x / λ) for λ > 0. Then the epigraph of g admits the geometric representation epi g = {(x, λ) | h (λ, x) ≤ 1}, where h (λ, x) = f_λ x if λ > 0, h (0, x) = (f₀⁺) x (the recession function), and h (λ, x) = +∞ if λ < 0.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma epigraph_obverseConvex_eq_sublevel_one {n : ℕ} (f : (Fin n → ℝ) → EReal)
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0) :
let g : (Fin n → ℝ) → EReal := obverseConvex f
let h : ℝ → (Fin n → ℝ) → EReal :=
fun t x =>
if 0 < t then
(t : EReal) * f ((t⁻¹) • x)
else if t = 0 then
recessionFunction f x
else
⊤
epigraph (S := (Set.univ : Set (Fin n → ℝ))) g = {p : (Fin n → ℝ) × ℝ | h p.2 p.1 ≤ 1}
At λ = 1, the auxiliary function h agrees with f.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma h_at_one_simp {n : ℕ} (f : (Fin n → ℝ) → EReal) :
let h : ℝ → (Fin n → ℝ) → EReal :=
fun t x =>
if 0 < t then
(t : EReal) * f ((t⁻¹) • x)
else if t = 0 then
recessionFunction f x
else
⊤
∀ x, h 1 x = f x
The λ = 1 slice of P = epi h projects to epi f.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma image_P_slice_lambda_one_eq_epigraph_f {n : ℕ} (f : (Fin n → ℝ) → EReal) :
let h : ℝ → (Fin n → ℝ) → EReal :=
fun t x =>
if 0 < t then
(t : EReal) * f ((t⁻¹) • x)
else if t = 0 then
recessionFunction f x
else
⊤
let P : Set ((ℝ × (Fin n → ℝ)) × ℝ) := {q | h q.1.1 q.1.2 ≤ (q.2 : EReal)}
((fun q : ((ℝ × (Fin n → ℝ)) × ℝ) => (q.1.2, q.2)) '' (P ∩ {q | q.1.1 = (1 : ℝ)}) =
epigraph (S := (Set.univ : Set (Fin n → ℝ))) f)
The μ = 1 slice of P = epi h projects to the h ≤ 1 sublevel set.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma image_P_slice_mu_one_eq_sublevel_h {n : ℕ} (f : (Fin n → ℝ) → EReal) :
let h : ℝ → (Fin n → ℝ) → EReal :=
fun t x =>
if 0 < t then
(t : EReal) * f ((t⁻¹) • x)
else if t = 0 then
recessionFunction f x
else
⊤
let P : Set ((ℝ × (Fin n → ℝ)) × ℝ) := {q | h q.1.1 q.1.2 ≤ (q.2 : EReal)}
((fun q : ((ℝ × (Fin n → ℝ)) × ℝ) => (q.1.2, q.1.1)) '' (P ∩ {q | q.2 = (1 : ℝ)}) =
{p : (Fin n → ℝ) × ℝ | h p.2 p.1 ≤ (1 : EReal)})
The sublevel set h ≤ 1 is the epigraph of obverseConvex f.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma sublevel_h_one_eq_epigraph_obverseConvex {n : ℕ} (f : (Fin n → ℝ) → EReal)
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0) :
let g : (Fin n → ℝ) → EReal := obverseConvex f
let h : ℝ → (Fin n → ℝ) → EReal :=
fun t x =>
if 0 < t then
(t : EReal) * f ((t⁻¹) • x)
else if t = 0 then
recessionFunction f x
else
⊤
{p : (Fin n → ℝ) × ℝ | h p.2 p.1 ≤ (1 : EReal)} =
epigraph (S := (Set.univ : Set (Fin n → ℝ))) g
Text 15.0.34: Let P = epi h ⊆ ℝ^{n+2} be the (closed convex) cone given as the epigraph of the function h (λ, x) from Text 15.0.33; it is the smallest closed convex cone containing the points (1, x, μ) with μ ≥ f x. Then the slice of P by the hyperplane λ = 1 corresponds to the epigraph epi f, and the slice by the hyperplane μ = 1 corresponds to the epigraph epi g of the obverse g.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma epigraph_h_slice_hyperplanes_correspond {n : ℕ} (f : (Fin n → ℝ) → EReal)
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0) :
let g : (Fin n → ℝ) → EReal := obverseConvex f
let h : ℝ → (Fin n → ℝ) → EReal :=
fun t x =>
if 0 < t then
(t : EReal) * f ((t⁻¹) • x)
else if t = 0 then
recessionFunction f x
else
⊤
let P : Set ((ℝ × (Fin n → ℝ)) × ℝ) := {q | h q.1.1 q.1.2 ≤ (q.2 : EReal)}
((fun q : ((ℝ × (Fin n → ℝ)) × ℝ) => (q.1.2, q.2)) '' (P ∩ {q | q.1.1 = (1 : ℝ)}) =
epigraph (S := (Set.univ : Set (Fin n → ℝ))) f) ∧
((fun q : ((ℝ × (Fin n → ℝ)) × ℝ) => (q.1.2, q.1.1)) '' (P ∩ {q | q.2 = (1 : ℝ)}) =
epigraph (S := (Set.univ : Set (Fin n → ℝ))) g)
The recession function is nonnegative when f ≥ 0 and f 0 = 0.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma recessionFunction_nonneg_of_nonneg_zero {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf0 : f 0 = 0) :
∀ y, (0 : EReal) ≤ recessionFunction f y
The recession function is positively homogeneous under closedness and nonnegativity.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma recessionFunction_posHom_of_nonneg_closedConvex_zero {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0) :
PositivelyHomogeneous (recessionFunction f)
Normalizing the h-inequality at positive μ.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma h_div_mu_le_one_of_h_le_mu {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0) :
let h : ℝ → (Fin n → ℝ) → EReal :=
fun t x =>
if 0 < t then
(t : EReal) * f ((t⁻¹) • x)
else if t = 0 then
recessionFunction f x
else
⊤
∀ {lam μ : ℝ} {x : Fin n → ℝ}, 0 < μ →
h lam x ≤ (μ : EReal) →
h (lam / μ) ((μ⁻¹) • x) ≤ (1 : EReal)
Points in P have nonnegative λ.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma mem_P_imp_lam_nonneg {n : ℕ} {f : (Fin n → ℝ) → EReal} :
let h : ℝ → (Fin n → ℝ) → EReal :=
fun t x =>
if 0 < t then
(t : EReal) * f ((t⁻¹) • x)
else if t = 0 then
recessionFunction f x
else
⊤
let P : Set ((ℝ × (Fin n → ℝ)) × ℝ) := {q | h q.1.1 q.1.2 ≤ (q.2 : EReal)}
∀ {q}, q ∈ P → 0 ≤ q.1.1
The auxiliary function h is nonnegative when λ ≥ 0.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma h_nonneg_of_lam_nonneg {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf0 : f 0 = 0) :
let h : ℝ → (Fin n → ℝ) → EReal :=
fun t x =>
if 0 < t then
(t : EReal) * f ((t⁻¹) • x)
else if t = 0 then
recessionFunction f x
else
⊤
∀ {lam x}, 0 ≤ lam → (0 : EReal) ≤ h lam x
Dot-product bound for points in the epigraph set P.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma dotProduct_le_mu_add_lam_mul_fStar_toReal_of_mem_P {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0) :
let fStar : (Fin n → ℝ) → EReal := fenchelConjugate n f
let C : Set (Fin n → ℝ) := effectiveDomain (Set.univ : Set (Fin n → ℝ)) fStar
let h : ℝ → (Fin n → ℝ) → EReal :=
fun t x =>
if 0 < t then
(t : EReal) * f ((t⁻¹) • x)
else if t = 0 then
recessionFunction f x
else
⊤
let P : Set ((ℝ × (Fin n → ℝ)) × ℝ) := {q | h q.1.1 q.1.2 ≤ (q.2 : EReal)}
∀ {q}, q ∈ P → ∀ y ∈ C,
dotProduct q.1.2 y ≤ q.2 + q.1.1 * (fStar y).toReal
The epigraph set P is closed.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma isClosed_P {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0) :
let h : ℝ → (Fin n → ℝ) → EReal :=
fun t x =>
if 0 < t then
(t : EReal) * f ((t⁻¹) • x)
else if t = 0 then
recessionFunction f x
else
⊤
let P : Set ((ℝ × (Fin n → ℝ)) × ℝ) := {q | h q.1.1 q.1.2 ≤ (q.2 : EReal)}
IsClosed P§15 — Polars of Convex Functions — Part 20
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
The cone P is the closure of its μ > 0 slice.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma P_eq_closure_P_inter_mu_pos {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0) :
let h : ℝ → (Fin n → ℝ) → EReal :=
fun t x =>
if 0 < t then
(t : EReal) * f ((t⁻¹) • x)
else if t = 0 then
recessionFunction f x
else
⊤
let P : Set ((ℝ × (Fin n → ℝ)) × ℝ) := {q | h q.1.1 q.1.2 ≤ (q.2 : EReal)}
P = closure (P ∩ {q | 0 < q.2})
Any closed convex cone containing the obverse points contains P.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma P_subset_any_closed_convexCone_containing_obverse_points {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0) :
let g : (Fin n → ℝ) → EReal := obverseConvex f
let h : ℝ → (Fin n → ℝ) → EReal :=
fun t x =>
if 0 < t then
(t : EReal) * f ((t⁻¹) • x)
else if t = 0 then
recessionFunction f x
else
⊤
let P : Set ((ℝ × (Fin n → ℝ)) × ℝ) := {q | h q.1.1 q.1.2 ≤ (q.2 : EReal)}
∀ K : ConvexCone ℝ ((ℝ × (Fin n → ℝ)) × ℝ),
IsClosed (K : Set ((ℝ × (Fin n → ℝ)) × ℝ)) →
(∀ (lam : ℝ) (x : Fin n → ℝ),
g x ≤ (lam : EReal) →
((lam, x), (1 : ℝ)) ∈ (K : Set ((ℝ × (Fin n → ℝ)) × ℝ))) →
P ⊆ (K : Set ((ℝ × (Fin n → ℝ)) × ℝ))
Text 15.0.35: Let P ⊆ ℝ^{n+2} be the closed convex cone P = epi h from Text 15.0.34, where h is built from f as in Text 15.0.33 and g is the obverse of f. Assuming f ≥ 0, the set P is the closure of its intersection with the open half-space μ > 0. Moreover, P is the smallest closed convex cone containing all points (λ, x, 1) such that λ ≥ g(x). Thus f and g determine the same closed convex cone in ℝ^{n+2}, with the roles of λ and μ reversed when passing between the two descriptions.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma epigraph_h_min_closedConvexCone_containing_obverse_points {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0) :
let g : (Fin n → ℝ) → EReal := obverseConvex f
let h : ℝ → (Fin n → ℝ) → EReal :=
fun t x =>
if 0 < t then
(t : EReal) * f ((t⁻¹) • x)
else if t = 0 then
recessionFunction f x
else
⊤
let P : Set ((ℝ × (Fin n → ℝ)) × ℝ) := {q | h q.1.1 q.1.2 ≤ (q.2 : EReal)}
(P = closure (P ∩ {q | 0 < q.2})) ∧
(∀ K : ConvexCone ℝ ((ℝ × (Fin n → ℝ)) × ℝ),
IsClosed (K : Set ((ℝ × (Fin n → ℝ)) × ℝ)) →
(∀ (lam : ℝ) (x : Fin n → ℝ),
g x ≤ (lam : EReal) →
((lam, x), (1 : ℝ)) ∈ (K : Set ((ℝ × (Fin n → ℝ)) × ℝ))) →
P ⊆ (K : Set ((ℝ × (Fin n → ℝ)) × ℝ)))
The polar convex function is nonnegative.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma polarConvex_nonneg {n : ℕ} (f : (Fin n → ℝ) → EReal) (xStar : Fin n → ℝ) :
(0 : EReal) ≤ polarConvex f xStar
The polar convex function vanishes at the origin.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma polarConvex_zero {n : ℕ} (f : (Fin n → ℝ) → EReal) :
polarConvex f (0 : Fin n → ℝ) = 0
Vertical reflection in the last coordinate.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
def vertReflect {n : ℕ} (p : (Fin n → ℝ) × ℝ) : (Fin n → ℝ) × ℝ :=
(p.1, -p.2)
The vertical reflection is an involution.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma vertReflect_involutive {n : ℕ} : Function.Involutive (vertReflect (n := n))
Product-space polar set using the dot-product-plus-scalar pairing.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
def polarSetProd {n : ℕ} (C : Set ((Fin n → ℝ) × ℝ)) : Set ((Fin n → ℝ) × ℝ) :=
{y | ∀ x ∈ C, (x.1 ⬝ᵥ y.1 : ℝ) + x.2 * y.2 ≤ 1}
polarSetProd is closed and convex as an intersection of closed halfspaces.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma polarSetProd_isClosed_and_convex {n : ℕ} (C : Set ((Fin n → ℝ) × ℝ)) :
IsClosed (polarSetProd (n := n) C) ∧ Convex ℝ (polarSetProd (n := n) C)
polarSetProd is unchanged by taking the closure of the set.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma polarSetProd_closure_eq {n : ℕ} (C : Set ((Fin n → ℝ) × ℝ)) :
polarSetProd (n := n) (closure C) = polarSetProd (n := n) C
polarSetProd commutes with vertical reflection preimages.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma polarSetProd_preimage_vertReflect {n : ℕ} (C : Set ((Fin n → ℝ) × ℝ)) :
polarSetProd (n := n) ((vertReflect (n := n)) ⁻¹' C) =
(vertReflect (n := n)) ⁻¹' (polarSetProd (n := n) C)
A linear functional on the product splits into a dot product and a scalar part.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma linearMap_eq_dotProduct_add_mul_prod {n : ℕ}
(φ : ((Fin n → ℝ) × ℝ) →ₗ[ℝ] ℝ) :
∀ p : (Fin n → ℝ) × ℝ,
φ p = (p.1 ⬝ᵥ fun i => φ (Pi.single i 1, (0 : ℝ)) : ℝ) + p.2 * φ (0, 1)
The dual bipolar set matches the polarSetProd bipolar.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
lemma bipolar_dualSet_eq_polarSetProd_polarSetProd {n : ℕ}
(C : Set ((Fin n → ℝ) × ℝ)) :
{x : (Fin n → ℝ) × ℝ |
∀ φ ∈ polar (E := (Fin n → ℝ) × ℝ) C, φ x ≤ (1 : ℝ)} =
polarSetProd (n := n) (polarSetProd (n := n) C)§15 — Polars of Convex Functions — Part 21
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
Bipolarity for polarSetProd, using the closed convex bipolar theorem.
open scoped BigOperators
open scoped Pointwise
lemma polarSetProd_bipolar_eq_closure_of_convex_zeroMem {n : ℕ}
(C : Set ((Fin n → ℝ) × ℝ)) (hCconv : Convex ℝ C)
(hC0 : (0 : (Fin n → ℝ) × ℝ) ∈ C) :
polarSetProd (n := n) (polarSetProd (n := n) C) = closure C
Nonnegativity rules out the value ⊥.
open scoped BigOperators
open scoped Pointwise
lemma hbot_of_nonneg {n : ℕ} {f : (Fin n → ℝ) → EReal} (hf_nonneg : ∀ x, 0 ≤ f x) :
∀ x, f x ≠ (⊥ : EReal)
If a reflected point lies in the polar of the epigraph and f 0 = 0, then the height is nonnegative.
open scoped BigOperators
open scoped Pointwise
lemma mu_nonneg_of_vertReflect_mem_polarSetProd_epigraph {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf0 : f 0 = 0) {xStar : Fin n → ℝ} {μ : ℝ}
(hmem :
vertReflect (n := n) (xStar, μ) ∈
polarSetProd (n := n) (epigraph (S := (Set.univ : Set (Fin n → ℝ))) f)) :
0 ≤ μ§15 — Polars of Convex Functions — Part 22
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
Epigraph points of the polar map to the product polar via vertical reflection.
open scoped BigOperators
open scoped Pointwise
lemma epigraph_polarConvex_subset_vertReflect_preimage_polarSetProd_epigraph {n : ℕ}
{f : (Fin n → ℝ) → EReal} (hf_nonneg : ∀ x, 0 ≤ f x) :
epigraph (S := (Set.univ : Set (Fin n → ℝ))) (polarConvex f) ⊆
(vertReflect (n := n)) ⁻¹'
polarSetProd (n := n) (epigraph (S := (Set.univ : Set (Fin n → ℝ))) f)
Reflected product-polar points are in the epigraph of the polar.
open scoped BigOperators
open scoped Pointwise
lemma vertReflect_preimage_polarSetProd_epigraph_subset_epigraph_polarConvex {n : ℕ}
{f : (Fin n → ℝ) → EReal} (hf_nonneg : ∀ x, 0 ≤ f x) (hf0 : f 0 = 0) :
(vertReflect (n := n)) ⁻¹'
polarSetProd (n := n) (epigraph (S := (Set.univ : Set (Fin n → ℝ))) f) ⊆
epigraph (S := (Set.univ : Set (Fin n → ℝ))) (polarConvex f)
Epigraph identity for the polar convex function.
open scoped BigOperators
open scoped Pointwise
lemma epigraph_polarConvex_mem_iff_vertReflect_mem_polarSetProd_epigraph {n : ℕ}
{f : (Fin n → ℝ) → EReal} (hf_nonneg : ∀ x, 0 ≤ f x) (hf0 : f 0 = 0)
(p : (Fin n → ℝ) × ℝ) :
p ∈ epigraph (S := (Set.univ : Set (Fin n → ℝ))) (polarConvex f) ↔
vertReflect (n := n) p ∈
polarSetProd (n := n) (epigraph (S := (Set.univ : Set (Fin n → ℝ))) f)
Bipolar epigraph of the polar convex function.
open scoped BigOperators
open scoped Pointwise
lemma epigraph_bipolar_polarConvex_eq_closure_epigraph {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf_conv : ConvexFunctionOn (S := (Set.univ : Set (Fin n → ℝ))) f)
(hf_nonneg : ∀ x, 0 ≤ f x) (hf0 : f 0 = 0) :
epigraph (S := (Set.univ : Set (Fin n → ℝ))) (polarConvex (polarConvex f)) =
closure (epigraph (S := (Set.univ : Set (Fin n → ℝ))) f)
open scoped BigOperators
open scoped Pointwise
theorem polarConvex_nonneg_closedConvex_and_bipolar_eq_convexFunctionClosure {n : ℕ}
{f : (Fin n → ℝ) → EReal}
(hf_conv : ConvexFunctionOn (S := (Set.univ : Set (Fin n → ℝ))) f)
(hf_nonneg : ∀ x, 0 ≤ f x) (hf0 : f 0 = 0) :
let fPolar : (Fin n → ℝ) → EReal := polarConvex f
(∀ xStar, 0 ≤ fPolar xStar) ∧
ClosedConvexFunction fPolar ∧
fPolar 0 = 0 ∧
polarConvex fPolar = convexFunctionClosure f
A closed convex function is convex on univ.
open scoped BigOperators
open scoped Pointwise
lemma convexFunctionOn_univ_of_closedConvexFunction {n : ℕ}
{f : (Fin n → ℝ) → EReal} (hf_closed : ClosedConvexFunction f) :
ConvexFunctionOn (S := (Set.univ : Set (Fin n → ℝ))) f
polarConvex preserves nonnegativity, closed convexity, and vanishing at 0 on the subtype.
open scoped BigOperators
open scoped Pointwise
lemma polarConvex_mem_subtype_nonneg_closedConvex_zero {n : ℕ}
(f : {g : (Fin n → ℝ) → EReal //
(∀ x, 0 ≤ g x) ∧ ClosedConvexFunction g ∧ g 0 = 0}) :
(∀ x, 0 ≤ polarConvex f.1 x) ∧
ClosedConvexFunction (polarConvex f.1) ∧
polarConvex f.1 0 = 0
polarConvex is involutive on the nonnegative closed convex subclass with f 0 = 0.
open scoped BigOperators
open scoped Pointwise
lemma polarConvex_involutive_on_subtype {n : ℕ}
(f : {g : (Fin n → ℝ) → EReal //
(∀ x, 0 ≤ g x) ∧ ClosedConvexFunction g ∧ g 0 = 0}) :
polarConvex (polarConvex f.1) = f.1§15 — Polars of Convex Functions — Part 23
Lean authors: Zichen Wang, Wanli Ma, Yijie Wang, Yunxi Duan, Zaiwen Wen.
Corollary 15.4.1: The polarity mapping f ↦ fᵒ induces a symmetric one-to-one correspondence on the class of all nonnegative closed convex functions on ℝⁿ that vanish at the origin. In this development, fᵒ is polarConvex f.
open scoped BigOperators
open scoped Pointwise
noncomputable def polarConvexEquiv_nonneg_closedConvex_zero {n : ℕ} :
{f : (Fin n → ℝ) → EReal //
(∀ x, 0 ≤ f x) ∧ ClosedConvexFunction f ∧ f 0 = 0} ≃
{f : (Fin n → ℝ) → EReal //
(∀ x, 0 ≤ f x) ∧ ClosedConvexFunction f ∧ f 0 = 0}
The obverse is the polar of the Fenchel conjugate under nonnegativity and closedness.
open scoped BigOperators
open scoped Pointwise
lemma obverseConvex_eq_polarConvex_fenchelConjugate {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0) :
obverseConvex f = polarConvex (fenchelConjugate n f)
Nonnegativity and f 0 = 0 force the Fenchel conjugate to vanish at 0.
open scoped BigOperators
open scoped Pointwise
lemma fenchelConjugate_zero_eq_zero_of_nonneg_zero {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf0 : f 0 = 0) : fenchelConjugate n f 0 = 0
The polar of the obverse recovers the Fenchel conjugate.
open scoped BigOperators
open scoped Pointwise
lemma polarConvex_obverseConvex_eq_fenchelConjugate {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0) :
polarConvex (obverseConvex f) = fenchelConjugate n f
The obverse vanishes at 0 under nonnegativity and f 0 = 0.
open scoped BigOperators
open scoped Pointwise
lemma obverseConvex_zero_of_nonneg_zero {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf0 : f 0 = 0) : obverseConvex f 0 = 0
A feasible dilation inequality bounds the obverse at a scaled point.
open scoped BigOperators
open scoped Pointwise
lemma obverseConvex_smul_le_coe_of_mul_le_one {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0)
{t : ℝ} (ht : 0 < t) {y : Fin n → ℝ} :
((t : EReal) * f y ≤ (1 : EReal)) → obverseConvex f (t • y) ≤ (t : EReal)
A polar-feasible μ⋆ bounds the Fenchel conjugate of the obverse.
open scoped BigOperators
open scoped Pointwise
lemma fenchelConjugate_obverseConvex_le_of_polar_feasible {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x)
{xStar : Fin n → ℝ} {μStar : EReal}
(hμ : 0 ≤ μStar ∧
∀ y, ((y ⬝ᵥ xStar : ℝ) : EReal) ≤ (1 : EReal) + μStar * f y) :
fenchelConjugate n (obverseConvex f) xStar ≤ μStar
Positive finite values of f yield a polar-feasible inner bound.
open scoped BigOperators
open scoped Pointwise
lemma inner_le_one_add_mul_of_fenchelConjugate_obverse_of_pos {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0)
{xStar : Fin n → ℝ} {μ ε : ℝ}
(hμ : fenchelConjugate n (obverseConvex f) xStar = (μ : EReal)) (hε : 0 < ε)
{y : Fin n → ℝ} {r : ℝ} (hfy : f y = (r : EReal)) (hr : 0 < r) :
((y ⬝ᵥ xStar : ℝ) : EReal) ≤ (1 : EReal) + ((μ + ε : ℝ) : EReal) * f y
A zero value of f forces the inner product to be at most one.
open scoped BigOperators
open scoped Pointwise
lemma inner_le_one_of_f_eq_zero_of_fenchelConjugate_obverse_finite {n : ℕ}
{f : (Fin n → ℝ) → EReal} (hf_nonneg : ∀ x, 0 ≤ f x)
(hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0)
{xStar : Fin n → ℝ} {μ : ℝ}
(hμ : fenchelConjugate n (obverseConvex f) xStar = (μ : EReal))
{y : Fin n → ℝ} (hfy : f y = (0 : EReal)) :
((y ⬝ᵥ xStar : ℝ) : EReal) ≤ (1 : EReal)
The Fenchel conjugate value yields a polar-feasible bound with μ + ε.
open scoped BigOperators
open scoped Pointwise
lemma polar_feasible_of_fenchelConjugate_obverse_add_pos {n : ℕ}
{f : (Fin n → ℝ) → EReal} (hf_nonneg : ∀ x, 0 ≤ f x)
(hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0)
{xStar : Fin n → ℝ} {μ ε : ℝ}
(hμ : fenchelConjugate n (obverseConvex f) xStar = (μ : EReal)) (hε : 0 < ε) :
0 ≤ ((μ + ε : ℝ) : EReal) ∧
∀ y, ((y ⬝ᵥ xStar : ℝ) : EReal) ≤ (1 : EReal) + ((μ + ε : ℝ) : EReal) * f y
The polar is bounded above by the Fenchel conjugate of the obverse.
open scoped BigOperators
open scoped Pointwise
lemma polarConvex_le_fenchelConjugate_obverseConvex {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0)
(xStar : Fin n → ℝ) :
polarConvex f xStar ≤ fenchelConjugate n (obverseConvex f) xStar
The Fenchel conjugate of the obverse is the polar (core symmetry).
open scoped BigOperators
open scoped Pointwise
lemma fenchelConjugate_obverseConvex_eq_polarConvex {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0) :
fenchelConjugate n (obverseConvex f) = polarConvex f
The obverse of the Fenchel conjugate is the polar.
open scoped BigOperators
open scoped Pointwise
lemma obverseConvex_fenchelConjugate_eq_polarConvex {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0) :
obverseConvex (fenchelConjugate n f) = polarConvex f
Theorem 15.5: Let f be a nonnegative closed convex function with f(0)=0, and let g be the obverse of f. Then g is also a nonnegative closed convex function with g(0)=0, and f is the obverse of g. Moreover, f^∘ = g^* and f^* = g^∘, and f^∘ and f^* are obverses of each other. In this development, the obverse is obverseConvex, the polar is polarConvex, and the Fenchel conjugate is fenchelConjugate n ·.
open scoped BigOperators
open scoped Pointwise
theorem obverseConvex_nonneg_closedConvex_zero_and_polar_eq_conjugate {n : ℕ}
{f g : (Fin n → ℝ) → EReal} (hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f)
(hf0 : f 0 = 0) (hg : g = obverseConvex f) :
(∀ x, 0 ≤ g x) ∧
ClosedConvexFunction g ∧
g 0 = 0 ∧
obverseConvex g = f ∧
polarConvex f = fenchelConjugate n g ∧
fenchelConjugate n f = polarConvex g ∧
obverseConvex (polarConvex f) = fenchelConjugate n f ∧
obverseConvex (fenchelConjugate n f) = polarConvex f
Corollary 15.5.1: If f is a nonnegative closed convex function on ℝⁿ with f(0)=0, then f^{∘*} = f^{*∘}; i.e. the Fenchel conjugate of the polar equals the polar of the Fenchel conjugate. In this development, fᵒ is polarConvex f and f* is fenchelConjugate n f.
open scoped BigOperators
open scoped Pointwise
theorem fenchelConjugate_polarConvex_eq_polarConvex_fenchelConjugate {n : ℕ}
{f : (Fin n → ℝ) → EReal} (hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f)
(hf0 : f 0 = 0) :
fenchelConjugate n (polarConvex f) = polarConvex (fenchelConjugate n f)
Obverse sublevel inequality is equivalent to a reciprocal bound on f.
open scoped BigOperators
open scoped Pointwise
lemma obverseConvex_le_coe_pos_iff_f_le_inv {n : ℕ} {f : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0)
{x : Fin n → ℝ} {α : ℝ} (hα : 0 < α) :
obverseConvex f x ≤ (α : EReal) ↔ f ((α⁻¹) • x) ≤ ((α⁻¹ : ℝ) : EReal)
Text 15.0.36: Let f be as in Theorem 15.5 and let g be its obverse. For α > 0, {x | g x ≤ α} = α • {x | f x ≤ α⁻¹}. In particular, the sublevel sets of g are homothetic to those of f with reciprocal levels.
open scoped BigOperators
open scoped Pointwise
theorem sublevelSet_obverseConvex_eq_smul_sublevelSet {n : ℕ} {f g : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0)
(hg : g = obverseConvex f) {α : ℝ} (hα : 0 < α) :
{x | g x ≤ (α : EReal)} = α • {x | f x ≤ ((α⁻¹ : ℝ) : EReal)}
Text 15.0.36: Since polarConvex f is the obverse of fenchelConjugate n f (Theorem 15.5), for every α > 0, {x⋆ | polarConvex f x⋆ ≤ α⁻¹} = α⁻¹ • {x⋆ | fenchelConjugate n f x⋆ ≤ α}. The set on the left is the “middle” set appearing in Theorem 14.7.
open scoped BigOperators
open scoped Pointwise
theorem sublevelSet_polarConvex_le_inv_eq_inv_smul_sublevelSet_fenchelConjugate {n : ℕ}
{f : (Fin n → ℝ) → EReal} (hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f)
(hf0 : f 0 = 0) {α : ℝ} (hα : 0 < α) :
{xStar | polarConvex f xStar ≤ ((α⁻¹ : ℝ) : EReal)} =
(α⁻¹) • {xStar | fenchelConjugate n f xStar ≤ (α : EReal)}
Normalize a dilation inequality in EReal to a ≤ 1 form.
open scoped BigOperators
open scoped Pointwise
lemma dilation_le_iff_scaled_le_one {lam mu : ℝ} (hmu : 0 < mu) (A : EReal) :
((lam : EReal) * A ≤ (mu : EReal)) ↔
(((lam / mu) : ℝ) : EReal) * A ≤ (1 : EReal)
Divide a dilated obverse inequality by a positive scalar.
open scoped BigOperators
open scoped Pointwise
lemma obverse_dilation_le_iff_obverse_le_div {n : ℕ} {f : (Fin n → ℝ) → EReal}
{x : Fin n → ℝ} {lam mu : ℝ} (hmu : 0 < mu) :
((mu : EReal) * obverseConvex f ((mu⁻¹) • x) ≤ (lam : EReal)) ↔
obverseConvex f ((mu⁻¹) • x) ≤ (((lam / mu) : ℝ) : EReal)
Combine reciprocal scalings in a nested smul.
open scoped BigOperators
open scoped Pointwise
lemma inv_div_inv_smul_smul_eq_inv_smul {n : ℕ} {x : Fin n → ℝ} {lam mu : ℝ}
(hmu : 0 < mu) :
(((lam / mu)⁻¹ : ℝ) • ((mu⁻¹) • x)) = (lam⁻¹) • x
Text 15.0.37: For the obverse g of f, one has (f_λ)(x) ≤ μ if and only if (g_μ)(x) ≤ λ, assuming λ > 0 and μ > 0. Here f_λ x := λ * f (λ⁻¹ • x) and g_μ x := μ * g (μ⁻¹ • x).
open scoped BigOperators
open scoped Pointwise
lemma dilation_le_iff_obverse_dilation_le {n : ℕ} {f g : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0)
(hg : g = obverseConvex f) (x : Fin n → ℝ) {lam mu : ℝ} (hlam : 0 < lam) (hmu : 0 < mu) :
((lam : EReal) * f ((lam⁻¹) • x) ≤ (mu : EReal)) ↔
((mu : EReal) * g ((mu⁻¹) • x) ≤ (lam : EReal))
Sublevel sets of the obverse match those of the perspective.
open scoped BigOperators
open scoped Pointwise
lemma sublevelSet_obverseConvex_eq_sublevelSet_perspective {n : ℕ} {f g : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0)
(hg : g = obverseConvex f) {α : ℝ} (hα : 0 < α) :
{x | g x ≤ (α : EReal)} =
{x | ((α : EReal) * f ((α⁻¹) • x)) ≤ (1 : EReal)}
Perspective sublevel sets are homothetic to those of f.
open scoped BigOperators
open scoped Pointwise
lemma sublevelSet_perspective_eq_smul_sublevelSet {n : ℕ} {f g : (Fin n → ℝ) → EReal}
(hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f) (hf0 : f 0 = 0)
(hg : g = obverseConvex f) {α : ℝ} (hα : 0 < α) :
{x | ((α : EReal) * f ((α⁻¹) • x)) ≤ (1 : EReal)} =
α • {x | f x ≤ ((α⁻¹ : ℝ) : EReal)}
Text 15.0.38: Let f : ℝⁿ → [0, +∞] be a nonnegative closed convex function with f 0 = 0, and let g be the obverse of f. For λ > 0 define (f_λ)(x) := λ * f (λ⁻¹ • x). Then for every α > 0, {x | g x ≤ α} = {x | (f_α) x ≤ 1} = α • {x | f x ≤ α⁻¹}.
open scoped BigOperators
open scoped Pointwise
theorem sublevelSet_obverseConvex_eq_sublevelSet_dilation_eq_smul_sublevelSet {n : ℕ}
{f g : (Fin n → ℝ) → EReal} (hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f)
(hf0 : f 0 = 0) (hg : g = obverseConvex f) {α : ℝ} (hα : 0 < α) :
let fα : (Fin n → ℝ) → EReal := fun x => (α : EReal) * f ((α⁻¹) • x)
{x | g x ≤ (α : EReal)} = {x | fα x ≤ (1 : EReal)} ∧
{x | fα x ≤ (1 : EReal)} = α • {x | f x ≤ ((α⁻¹ : ℝ) : EReal)}
Text 15.0.39: Let f : ℝⁿ → [0, +∞] be a nonnegative closed convex function with f 0 = 0. Then the polar f^∘ is the obverse of the Fenchel conjugate f*. Consequently, for every α > 0, {x⋆ | f^∘ x⋆ ≤ α⁻¹} = α⁻¹ • {x⋆ | f* x⋆ ≤ α}. In this development, f^∘ is polarConvex f and f* is fenchelConjugate n f.
open scoped BigOperators
open scoped Pointwise
theorem polarConvex_eq_obverseConvex_fenchelConjugate_and_sublevelSet {n : ℕ}
{f : (Fin n → ℝ) → EReal} (hf_nonneg : ∀ x, 0 ≤ f x) (hf_closed : ClosedConvexFunction f)
(hf0 : f 0 = 0) {α : ℝ} (hα : 0 < α) :
polarConvex f = obverseConvex (fenchelConjugate n f) ∧
{xStar | polarConvex f xStar ≤ ((α⁻¹ : ℝ) : EReal)} =
(α⁻¹) • {xStar | fenchelConjugate n f xStar ≤ (α : EReal)}§16 — Dual Operations — Part 1
Lean authors: Zichen Wang, Wanli Ma, Weiran Shi, Yunxi Duan, Zaiwen Wen.
Dot product is additive in the left argument.
open scoped BigOperators
open scoped Pointwise
lemma section16_dotProduct_add_left {n : ℕ} (u v w : Fin n → ℝ) :
dotProduct (u + v) w = dotProduct u w + dotProduct v w
Translate the Fenchel conjugate range under x ↦ x - a.
open scoped BigOperators
open scoped Pointwise
lemma section16_range_fenchelConjugate_translate {n : ℕ} (h : (Fin n → ℝ) → EReal)
(a xStar : Fin n → ℝ) :
Set.range (fun x => ((x ⬝ᵥ xStar : ℝ) : EReal) - h (x - a)) =
(fun z : EReal => z + ((dotProduct a xStar : ℝ) : EReal)) ''
Set.range (fun y => ((y ⬝ᵥ xStar : ℝ) : EReal) - h y)
Text 16.0.1: Basic operations on a convex function and their effect on the Fenchel conjugate. - If h is translated by a, i.e. f x = h (x - a), then f* x* = h* x* + ⟪a, x*⟫. - If a linear functional is added, i.e. f x = h x + ⟪x, a*⟫, then f* x* = h* (x* - a*). - For a real constant α, the conjugate of h + α is h* - α. As a special case, for a set C, the support function of the translate C + a is supportFunctionEReal C + ⟪a, ·⟫, since the support function is the conjugate of the indicator function.
open scoped BigOperators
open scoped Pointwise
lemma section16_fenchelConjugate_translate {n : ℕ} (h : (Fin n → ℝ) → EReal) (a : Fin n → ℝ) :
fenchelConjugate n (fun x => h (x - a)) =
fun xStar => fenchelConjugate n h xStar + ((dotProduct a xStar : ℝ) : EReal)
Rewriting the affine piece after adding a linear functional.
open scoped BigOperators
open scoped Pointwise
lemma section16_affine_piece_add_linear {n : ℕ} (h : (Fin n → ℝ) → EReal)
(aStar : Fin n → ℝ) (x xStar : Fin n → ℝ) :
((x ⬝ᵥ xStar : ℝ) : EReal) - (h x + ((dotProduct x aStar : ℝ) : EReal)) =
((x ⬝ᵥ (xStar - aStar) : ℝ) : EReal) - h x
Range rewrite for adding a linear functional inside fenchelConjugate.
open scoped BigOperators
open scoped Pointwise
lemma section16_range_fenchelConjugate_add_linear {n : ℕ} (h : (Fin n → ℝ) → EReal)
(aStar xStar : Fin n → ℝ) :
Set.range (fun x => ((x ⬝ᵥ xStar : ℝ) : EReal) - (h x + ((dotProduct x aStar : ℝ) : EReal))) =
Set.range (fun x => ((x ⬝ᵥ (xStar - aStar) : ℝ) : EReal) - h x)
open scoped BigOperators
open scoped Pointwise
lemma section16_fenchelConjugate_add_linear {n : ℕ} (h : (Fin n → ℝ) → EReal) (aStar : Fin n → ℝ) :
fenchelConjugate n (fun x => h x + ((dotProduct x aStar : ℝ) : EReal)) =
fun xStar => fenchelConjugate n h (xStar - aStar)
Rewriting the affine piece after adding a constant.
open scoped BigOperators
open scoped Pointwise
lemma section16_affine_piece_add_const {n : ℕ} (h : (Fin n → ℝ) → EReal) (α : ℝ)
(x xStar : Fin n → ℝ) :
((x ⬝ᵥ xStar : ℝ) : EReal) - (h x + (α : EReal)) =
(((x ⬝ᵥ xStar : ℝ) : EReal) - h x) + ((-α : ℝ) : EReal)
Range rewrite for adding a constant inside fenchelConjugate.
open scoped BigOperators
open scoped Pointwise
lemma section16_range_fenchelConjugate_add_const {n : ℕ} (h : (Fin n → ℝ) → EReal) (α : ℝ)
(xStar : Fin n → ℝ) :
Set.range (fun x => ((x ⬝ᵥ xStar : ℝ) : EReal) - (h x + (α : EReal))) =
(fun z : EReal => z + ((-α : ℝ) : EReal)) ''
Set.range (fun x => ((x ⬝ᵥ xStar : ℝ) : EReal) - h x)
open scoped BigOperators
open scoped Pointwise
lemma section16_fenchelConjugate_add_const {n : ℕ} (h : (Fin n → ℝ) → EReal) (α : ℝ) :
fenchelConjugate n (fun x => h x + (α : EReal)) =
fun xStar => fenchelConjugate n h xStar - (α : EReal)
open scoped BigOperators
open scoped Pointwise
lemma section16_supportFunctionEReal_translate {n : ℕ} (C : Set (Fin n → ℝ)) (a : Fin n → ℝ) :
supportFunctionEReal (Set.image (fun x : Fin n → ℝ => x + a) C) =
fun xStar => supportFunctionEReal C xStar + ((dotProduct a xStar : ℝ) : EReal)
Text 16.0.2: The polar of a convex set C is a ≤ 1 level set of the support function δ^*(· | C), namely C^∘ = {x^* | δ^*(x^* | C) ≤ 1}.
open scoped BigOperators
open scoped Pointwise
lemma section16_polar_eq_sublevel_deltaStar {n : ℕ} (C : Set (Fin n → ℝ)) :
{xStar | ∀ x ∈ C, (dotProduct x xStar : ℝ) ≤ 1} =
{xStar | supportFunctionEReal C xStar ≤ (1 : EReal)}
The conjugate of the constant zero function is the indicator of {0}.
open scoped BigOperators
open scoped Pointwise
lemma section16_fenchelConjugate_const_zero {n : ℕ} :
fenchelConjugate n (fun _ : (Fin n → ℝ) => (0 : EReal)) =
indicatorFunction ({0} : Set (Fin n → ℝ))
Scaling precomposition inside a Fenchel conjugate.
open scoped BigOperators
open scoped Pointwise
lemma section16_fenchelConjugate_precomp_smul {n : ℕ} (f : (Fin n → ℝ) → EReal) {lam : ℝ}
(hlam : lam ≠ 0) :
fenchelConjugate n (fun x => f (lam⁻¹ • x)) =
fun xStar => fenchelConjugate n f (lam • xStar)
The conjugate of the singleton indicator is the constant zero function.
open scoped BigOperators
open scoped Pointwise
lemma section16_fenchelConjugate_indicator_singleton_zero {n : ℕ} :
fenchelConjugate n (indicatorFunction ({0} : Set (Fin n → ℝ))) = (fun _ => (0 : EReal))
Theorem 16.1: For any proper convex function f, one has (λ f)^* = f^* λ and (f λ)^* = λ f^*, 0 ≤ λ < ∞. Here f^* is the Fenchel conjugate fenchelConjugate n f, λ f is pointwise multiplication x ↦ (λ : EReal) * f x, and f λ is the right scalar multiple rightScalarMultiple f λ.
open scoped BigOperators
open scoped Pointwise
theorem section16_fenchelConjugate_scaling {n : ℕ} (f : (Fin n → ℝ) → EReal)
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) f) {lam : ℝ} (hlam : 0 ≤ lam) :
fenchelConjugate n (fun x => ((lam : ℝ) : EReal) * f x) =
rightScalarMultiple (fenchelConjugate n f) lam ∧
fenchelConjugate n (rightScalarMultiple f lam) =
fun xStar => ((lam : ℝ) : EReal) * fenchelConjugate n f xStar
Scaling the dot-product image-set commutes with set scaling.
open scoped BigOperators
open scoped Pointwise
lemma section16_image_dotProduct_smul_set {n : ℕ} (C : Set (Fin n → ℝ)) (lam : ℝ)
(xStar : Fin n → ℝ) :
Set.image (fun x => dotProduct x xStar) (lam • C) =
lam • Set.image (fun x => dotProduct x xStar) C
Corollary 16.1.1. For any non-empty convex set C, one has δ^*(x^* | λ C) = λ δ^*(x^* | C), 0 ≤ λ < ∞.
open scoped BigOperators
open scoped Pointwise
theorem section16_deltaStar_scaling {n : ℕ} (C : Set (Fin n → ℝ)) {lam : ℝ} (hlam : 0 ≤ lam) :
deltaStar (lam • C) = fun xStar => lam * deltaStar C xStar
Membership in an inverse-scaled set can be rewritten using the original scaling.
open scoped BigOperators
open scoped Pointwise
lemma section16_mem_inv_smul_set_iff {n : ℕ} {lam : ℝ} (hlam0 : lam ≠ 0)
(S : Set (Fin n → ℝ)) (x : Fin n → ℝ) :
x ∈ lam⁻¹ • S ↔ lam • x ∈ S
The polar inequality for a scaled set is equivalent to a scaled dual variable.
open scoped BigOperators
open scoped Pointwise
lemma section16_polar_smul_iff {n : ℕ} (C : Set (Fin n → ℝ)) {lam : ℝ} (xStar : Fin n → ℝ) :
(∀ x ∈ (lam • C), (dotProduct x xStar : ℝ) ≤ 1) ↔
(∀ x ∈ C, (dotProduct x (lam • xStar) : ℝ) ≤ 1)
Corollary 16.1.2. For any non-empty convex set C one has (λ C)^∘ = λ^{-1} C^∘ for 0 < λ < ∞.
open scoped BigOperators
open scoped Pointwise
theorem section16_polar_scaling {n : ℕ} (C : Set (Fin n → ℝ)) {lam : ℝ} (hlam : 0 < lam) :
{xStar | ∀ x ∈ (lam • C), (dotProduct x xStar : ℝ) ≤ 1} =
lam⁻¹ • {xStar | ∀ x ∈ C, (dotProduct x xStar : ℝ) ≤ 1}
The intrinsic interior of a submodule (as a set) is the submodule itself.
open scoped BigOperators
open scoped Pointwise
lemma section16_intrinsicInterior_submodule_eq {n : Nat} (L : Submodule ℝ (Fin n → ℝ)) :
intrinsicInterior ℝ (L : Set (Fin n → ℝ)) = (L : Set (Fin n → ℝ))
Nonempty intersection in Euclidean space corresponds to non-disjoint intrinsic interiors.
open scoped BigOperators
open scoped Pointwise
lemma section16_nonempty_preimage_inter_ri_preimage_iff_not_disjoint_intrinsicInterior {n : Nat}
(L : Submodule ℝ (Fin n → ℝ)) (C : Set (Fin n → ℝ)) :
Set.Nonempty
(((fun x : EuclideanSpace ℝ (Fin n) => (x : Fin n → ℝ)) ⁻¹' (L : Set (Fin n → ℝ))) ∩
euclideanRelativeInterior n
((fun x : EuclideanSpace ℝ (Fin n) => (x : Fin n → ℝ)) ⁻¹' C)) ↔
¬ Disjoint (intrinsicInterior ℝ (L : Set (Fin n → ℝ))) (intrinsicInterior ℝ C)
The dot-product image over a submodule has sInf/sSup determined by orthogonality.
open scoped BigOperators
open scoped Pointwise
lemma section16_sInf_sSup_image_dotProduct_submodule {n : Nat} (L : Submodule ℝ (Fin n → ℝ))
(b : Fin n → ℝ) :
(sInf ((fun x : Fin n → ℝ => ((x ⬝ᵥ b : ℝ) : EReal)) '' (L : Set (Fin n → ℝ))) =
if b ∈ orthogonalComplement n L then (0 : EReal) else (⊥ : EReal)) ∧
(sSup ((fun x : Fin n → ℝ => ((x ⬝ᵥ b : ℝ) : EReal)) '' (L : Set (Fin n → ℝ))) =
if b ∈ orthogonalComplement n L then (0 : EReal) else (⊤ : EReal))
Negating the dual variable swaps the sign of the dot-product infimum condition.
open scoped BigOperators
open scoped Pointwise
lemma section16_sInf_lt_zero_iff_sSup_neg_gt_zero {n : Nat} (C : Set (Fin n → ℝ))
(b : Fin n → ℝ) :
(0 : EReal) > sInf ((fun x : Fin n → ℝ => ((x ⬝ᵥ b : ℝ) : EReal)) '' C) ↔
sSup ((fun x : Fin n → ℝ => ((x ⬝ᵥ (-b) : ℝ) : EReal)) '' C) > (0 : EReal)
Proper separation of a submodule and an effective domain in terms of recession inequalities.
open scoped BigOperators
open scoped Pointwise
lemma section16_exists_hyperplaneSeparatesProperly_submodule_effectiveDomain_iff_exists_orthogonal_recession_ineq
{n : Nat} (L : Submodule ℝ (Fin n → ℝ)) (f : (Fin n → ℝ) → EReal)
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) f) :
(∃ H, HyperplaneSeparatesProperly n H (L : Set (Fin n → ℝ))
(effectiveDomain (Set.univ : Set (Fin n → ℝ)) f)) ↔
∃ xStar : Fin n → ℝ,
xStar ∈ orthogonalComplement n L ∧
recessionFunction (fenchelConjugate n f) xStar ≤ (0 : EReal) ∧
recessionFunction (fenchelConjugate n f) (-xStar) > (0 : EReal)
Lemma 16.2. Let L be a subspace of ℝ^n and let f be a proper convex function. Then L meets ri (dom f) if and only if there exists no vector xStar ∈ L^⊥ such that (f^*0^+)(xStar) ≤ 0 and (f^*0^+)(-xStar) > 0. Here dom f is the effective domain effectiveDomain univ f, ri is euclideanRelativeInterior, and (f^*0^+) is represented as recessionFunction (fenchelConjugate n f).
open scoped BigOperators
open scoped Pointwise
lemma section16_subspace_meets_ri_effectiveDomain_iff_not_exists_orthogonal_recession_ineq
{n : Nat} (L : Submodule ℝ (Fin n → ℝ)) (f : (Fin n → ℝ) → EReal)
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) f) :
Set.Nonempty
((fun x : EuclideanSpace ℝ (Fin n) => (x : Fin n → ℝ)) ⁻¹' (L : Set (Fin n → ℝ)) ∩
euclideanRelativeInterior n
((fun x : EuclideanSpace ℝ (Fin n) => (x : Fin n → ℝ)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → ℝ)) f)) ↔
¬ ∃ xStar : Fin n → ℝ,
xStar ∈ orthogonalComplement n L ∧
recessionFunction (fenchelConjugate n f) xStar ≤ (0 : EReal) ∧
recessionFunction (fenchelConjugate n f) (-xStar) > (0 : EReal)
Coordinate form of a Euclidean linear map, viewed as a map into Fin m → ℝ.
open scoped BigOperators
open scoped Pointwise
noncomputable def section16_coordLinearMap {n m : Nat}
(A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) :
EuclideanSpace ℝ (Fin n) →ₗ[ℝ] (Fin m → ℝ) :=
(EuclideanSpace.equiv (ι := Fin m) (𝕜 := ℝ)).toLinearMap ∘ₗ A
Convert the range-preimage intersection into an explicit witness.
open scoped BigOperators
open scoped Pointwise
lemma section16_nonempty_preimage_range_inter_ri_effectiveDomain_iff {n m : Nat}
(A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) (g : (Fin m → ℝ) → EReal) :
Set.Nonempty
(((fun z : EuclideanSpace ℝ (Fin m) => (z : Fin m → ℝ)) ⁻¹'
(LinearMap.range (section16_coordLinearMap A) : Set (Fin m → ℝ))) ∩
euclideanRelativeInterior m
((fun z : EuclideanSpace ℝ (Fin m) => (z : Fin m → ℝ)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin m → ℝ)) g)) ↔
∃ x : EuclideanSpace ℝ (Fin n),
A x ∈
euclideanRelativeInterior m
((fun z : EuclideanSpace ℝ (Fin m) => (z : Fin m → ℝ)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin m → ℝ)) g)
Characterize the orthogonal complement of a range via the adjoint.
open scoped BigOperators
open scoped Pointwise
lemma section16_mem_orthogonalComplement_range_iff_adjoint_eq_zero {n m : Nat}
(A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m))
(yStar : EuclideanSpace ℝ (Fin m)) :
(yStar : Fin m → ℝ) ∈
orthogonalComplement m (LinearMap.range (section16_coordLinearMap A)) ↔
((LinearMap.adjoint :
(EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) ≃ₗ⋆[ℝ]
(EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n)))
A)
yStar = 0
Rewrite orthogonality over the range in an existential statement.
open scoped BigOperators
open scoped Pointwise
lemma section16_exists_orthogonalComplement_range_recession_iff_exists_adjoint_eq_zero_recession
{n m : Nat} (A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m))
(g : (Fin m → ℝ) → EReal) :
(∃ xStar : Fin m → ℝ,
xStar ∈ orthogonalComplement m (LinearMap.range (section16_coordLinearMap A)) ∧
recessionFunction (fenchelConjugate m g) xStar ≤ (0 : EReal) ∧
recessionFunction (fenchelConjugate m g) (-xStar) > (0 : EReal)) ↔
∃ yStar : EuclideanSpace ℝ (Fin m),
((LinearMap.adjoint :
(EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) ≃ₗ⋆[ℝ]
(EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n)))
A)
yStar = 0 ∧
recessionFunction (fenchelConjugate m g) (yStar : Fin m → ℝ) ≤ (0 : EReal) ∧
recessionFunction (fenchelConjugate m g) (-yStar : Fin m → ℝ) > (0 : EReal)§16 — Dual Operations — Part 2
Lean authors: Zichen Wang, Wanli Ma, Weiran Shi, Yunxi Duan, Zaiwen Wen.
The Euclidean-space equivalence reduces to the identity on Fin n → ℝ.
open scoped BigOperators
open scoped Pointwise
@[simp] lemma section16_euclideanSpace_equiv_toLp {n : Nat} (x : Fin n → ℝ) :
(EuclideanSpace.equiv (Fin n) ℝ) (WithLp.toLp 2 x) = x
The inverse Euclidean-space equivalence reduces to WithLp.toLp.
open scoped BigOperators
open scoped Pointwise
@[simp] lemma section16_euclideanSpace_equiv_symm_apply {n : Nat} (x : Fin n → ℝ) :
(EuclideanSpace.equiv (Fin n) ℝ).symm x = WithLp.toLp 2 x
Corollary 16.2.1. Let A be a linear transformation from ℝ^n to ℝ^m and let g be a proper convex function on ℝ^m. Then there exists no vector y* ∈ ℝ^m such that A^* y* = 0, (g^*0^+)(y*) ≤ 0, and (g^*0^+)(-y*) > 0 if and only if A x ∈ ri (dom g) for at least one x ∈ ℝ^n. Here dom g is the effective domain effectiveDomain univ g, ri is euclideanRelativeInterior, and (g^*0^+) is represented as recessionFunction (fenchelConjugate m g).
open scoped BigOperators
open scoped Pointwise
theorem section16_exists_image_mem_ri_effectiveDomain_iff_not_exists_adjoint_eq_zero_recession_ineq
{n m : Nat} (A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m))
(g : (Fin m → ℝ) → EReal)
(hg : ProperConvexFunctionOn (Set.univ : Set (Fin m → ℝ)) g) :
(¬ ∃ yStar : EuclideanSpace ℝ (Fin m),
((LinearMap.adjoint :
(EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) ≃ₗ⋆[ℝ]
(EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n)))
A)
yStar = 0 ∧
recessionFunction (fenchelConjugate m g) (yStar : Fin m → ℝ) ≤ (0 : EReal) ∧
recessionFunction (fenchelConjugate m g) (-yStar : Fin m → ℝ) > (0 : EReal)) ↔
∃ x : EuclideanSpace ℝ (Fin n),
A x ∈
euclideanRelativeInterior m
((fun z : EuclideanSpace ℝ (Fin m) => (z : Fin m → ℝ)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin m → ℝ)) g)
Block equivalence between flattened vectors and block families.
open scoped BigOperators
open scoped Pointwise
noncomputable def section16_blockEquivFun {n m : Nat} :
(Fin (m * n) → ℝ) ≃ₗ[ℝ] (Fin m → (Fin n → ℝ)) :=
blockEquivFun (n := n) (m := m)
Block projection linear map extracting the i-th block.
open scoped BigOperators
open scoped Pointwise
noncomputable def section16_blockLinearMap {n m : Nat} (i : Fin m) :
(Fin (m * n) → ℝ) →ₗ[ℝ] (Fin n → ℝ) :=
(LinearMap.proj i).comp (section16_blockEquivFun (n := n) (m := m)).toLinearMap
Flatten a block family into a single vector.
open scoped BigOperators
open scoped Pointwise
noncomputable def section16_unblock {n m : Nat} :
(Fin m → Fin n → ℝ) →ₗ[ℝ] (Fin (m * n) → ℝ) :=
(section16_blockEquivFun (n := n) (m := m)).symm.toLinearMap
Unblocking followed by block projection recovers the original block.
open scoped BigOperators
open scoped Pointwise
lemma section16_blockLinearMap_unblock {n m : Nat} (i : Fin m)
(x : Fin m → Fin n → ℝ) :
section16_blockLinearMap (n := n) (m := m) i (section16_unblock (n := n) (m := m) x) =
x i
Unblocking the block projections recovers the original vector.
open scoped BigOperators
open scoped Pointwise
lemma section16_unblock_blockLinearMap {n m : Nat} (x : Fin (m * n) → ℝ) :
section16_unblock (n := n) (m := m)
(fun i => section16_blockLinearMap (n := n) (m := m) i x) = x
Diagonal block embedding sending x to the constant block family.
open scoped BigOperators
open scoped Pointwise
noncomputable def section16_diagBlockLinearMap {n m : Nat} :
(Fin n → ℝ) →ₗ[ℝ] (Fin m → Fin n → ℝ)
Diagonal embedding into the flattened space.
open scoped BigOperators
open scoped Pointwise
noncomputable def section16_diagLinearMap {n m : Nat} :
(Fin n → ℝ) →ₗ[ℝ] (Fin (m * n) → ℝ) :=
(section16_unblock (n := n) (m := m)).comp (section16_diagBlockLinearMap (n := n) (m := m))
Block projection of the diagonal embedding is the identity.
open scoped BigOperators
open scoped Pointwise
lemma section16_blockLinearMap_diag {n m : Nat} (i : Fin m) (x : Fin n → ℝ) :
section16_blockLinearMap (n := n) (m := m) i (section16_diagLinearMap (n := n) (m := m) x) =
x
Each block projection is surjective.
open scoped BigOperators
open scoped Pointwise
lemma section16_blockLinearMap_surjective {n m : Nat} (i : Fin m) :
Function.Surjective (section16_blockLinearMap (n := n) (m := m) i)
Diagonal embedding as a linear map on Euclidean space.
open scoped BigOperators
open scoped Pointwise
noncomputable def section16_diagLinearMapE {n m : Nat} :
EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin (m * n)) :=
(EuclideanSpace.equiv (ι := Fin (m * n)) (𝕜 := ℝ)).symm.toLinearMap
∘ₗ (section16_diagLinearMap (n := n) (m := m))
∘ₗ (EuclideanSpace.equiv (ι := Fin n) (𝕜 := ℝ)).toLinearMap
Block projection as a linear map on Euclidean space.
open scoped BigOperators
open scoped Pointwise
noncomputable def section16_blockLinearMapE {n m : Nat} (i : Fin m) :
EuclideanSpace ℝ (Fin (m * n)) →ₗ[ℝ] EuclideanSpace ℝ (Fin n) :=
(EuclideanSpace.equiv (ι := Fin n) (𝕜 := ℝ)).symm.toLinearMap
∘ₗ (section16_blockLinearMap (n := n) (m := m) i)
∘ₗ (EuclideanSpace.equiv (ι := Fin (m * n)) (𝕜 := ℝ)).toLinearMap
Coercion form of the diagonal map.
open scoped BigOperators
open scoped Pointwise
lemma section16_diagLinearMapE_apply {n m : Nat} (x : EuclideanSpace ℝ (Fin n)) :
(section16_diagLinearMapE (n := n) (m := m) x : Fin (m * n) → ℝ) =
section16_diagLinearMap (n := n) (m := m) (x : Fin n → ℝ)
Coercion form of the block projection.
open scoped BigOperators
open scoped Pointwise
lemma section16_blockLinearMapE_apply {n m : Nat} (i : Fin m)
(y : EuclideanSpace ℝ (Fin (m * n))) :
(section16_blockLinearMapE (n := n) (m := m) i y : Fin n → ℝ) =
section16_blockLinearMap (n := n) (m := m) i (y : Fin (m * n) → ℝ)
Block projection of the diagonal embedding is the identity on Euclidean space.
open scoped BigOperators
open scoped Pointwise
lemma section16_blockLinearMapE_diag {n m : Nat} (i : Fin m) (x : EuclideanSpace ℝ (Fin n)) :
section16_blockLinearMapE (n := n) (m := m) i
(section16_diagLinearMapE (n := n) (m := m) x) =
x
Evaluate the block projection via the canonical index equivalence.
open scoped BigOperators
open scoped Pointwise
lemma section16_blockLinearMap_apply_equiv {n m : Nat} (i : Fin m) (x : Fin (m * n) → ℝ)
(j : Fin n) :
(section16_blockLinearMap (n := n) (m := m) i x) j =
x ((Fintype.equivFinOfCardEq (by simp) : Fin m × Fin n ≃ Fin (m * n)) (i, j))
Evaluate unblocking via the canonical index equivalence.
open scoped BigOperators
open scoped Pointwise
lemma section16_unblock_apply_equiv {n m : Nat} (xStar : Fin m → Fin n → ℝ) (i : Fin m)
(j : Fin n) :
(section16_unblock (n := n) (m := m) xStar)
((Fintype.equivFinOfCardEq (by simp) : Fin m × Fin n ≃ Fin (m * n)) (i, j)) =
xStar i j
Dot product against an unblocked vector splits by blocks.
open scoped BigOperators
open scoped Pointwise
lemma section16_dotProduct_unblock_eq_sum_blocks {n m : Nat}
(xStar : Fin m → Fin n → ℝ) (x : Fin (m * n) → ℝ) :
dotProduct x (section16_unblock (n := n) (m := m) xStar) =
∑ i : Fin m,
dotProduct (section16_blockLinearMap (n := n) (m := m) i x) (xStar i)
Dot product with an unblocked vector splits symmetrically by blocks.
open scoped BigOperators
open scoped Pointwise
lemma section16_dotProduct_unblock_eq_sum_blocks_symm {n m : Nat}
(xStar : Fin m → Fin n → ℝ) (x : Fin (m * n) → ℝ) :
dotProduct (section16_unblock (n := n) (m := m) xStar) x =
∑ i : Fin m,
dotProduct (xStar i) (section16_blockLinearMap (n := n) (m := m) i x)
The adjoint of the diagonal embedding is the sum of block projections.
open scoped BigOperators
open scoped Pointwise
lemma section16_diagLinearMapE_adjoint_eq_sum_blockLinearMapE {n m : Nat} :
(LinearMap.adjoint (section16_diagLinearMapE (n := n) (m := m))) =
∑ i : Fin m, section16_blockLinearMapE (n := n) (m := m) i
Effective domain of a block-sum is the intersection of block preimages.
open scoped BigOperators
open scoped Pointwise
lemma section16_effectiveDomain_blockSum_eq_iInter_preimage_blocks
{n m : Nat} (f : Fin m → (Fin n → ℝ) → EReal)
(hnotbot : ∀ i : Fin m, ∀ x, f i x ≠ (⊥ : EReal)) :
let g : (Fin (m * n) → ℝ) → EReal :=
fun x => ∑ i, f i (section16_blockLinearMap (n := n) (m := m) i x)
effectiveDomain (Set.univ : Set (Fin (m * n) → ℝ)) g =
⋂ i : Fin m,
(section16_blockLinearMap (n := n) (m := m) i) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → ℝ)) (f i)
Properness of the block-sum function assembled from proper convex summands.
open scoped BigOperators
open scoped Pointwise
lemma section16_properConvexFunctionOn_blockSum {n m : Nat}
(f : Fin m → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i)) :
ProperConvexFunctionOn (Set.univ : Set (Fin (m * n) → ℝ))
(fun x => ∑ i, f i (section16_blockLinearMap (n := n) (m := m) i x))
Image of the block projection on the block-sum effective domain.
open scoped BigOperators
open scoped Pointwise
lemma section16_blockLinearMapE_image_effectiveDomain_blockSum_eq {n m : Nat}
(f : Fin m → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i))
(i : Fin m) :
let g : (Fin (m * n) → ℝ) → EReal :=
fun x => ∑ j, f j (section16_blockLinearMap (n := n) (m := m) j x)
(section16_blockLinearMapE (n := n) (m := m) i) ''
((fun z : EuclideanSpace ℝ (Fin (m * n)) => (z : Fin (m * n) → ℝ)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin (m * n) → ℝ)) g) =
((fun x : EuclideanSpace ℝ (Fin n) => (x : Fin n → ℝ)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → ℝ)) (f i))
Coercion from ℝ to EReal commutes with finite sums.
open scoped BigOperators
open scoped Pointwise
lemma section16_sum_coe {ι : Type*} (s : Finset ι) (f : ι → ℝ) :
Finset.sum s (fun i => ((f i : ℝ) : EReal)) = ((Finset.sum s f : ℝ) : EReal)
Supremum of pairwise sums in EReal splits as the sum of suprema.
open scoped BigOperators
open scoped Pointwise
lemma section16_sSup_image2_add_eq_sSup_add {S T : Set EReal} :
sSup (Set.image2 (· + ·) S T) = sSup S + sSup T
Support function of a block-sum domain dominates the sum of blockwise support functions.
open scoped BigOperators
open scoped Pointwise
lemma section16_supportFunction_effectiveDomain_blockSum_unblock_ge_sum {n m : Nat}
(f : Fin m → (Fin n → ℝ) → EReal) (xStar : Fin m → Fin n → ℝ)
(hnotbot : ∀ i : Fin m, ∀ x, f i x ≠ (⊥ : EReal)) :
let g : (Fin (m * n) → ℝ) → EReal :=
fun x => ∑ i, f i (section16_blockLinearMap (n := n) (m := m) i x)
∑ i : Fin m,
supportFunctionEReal (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (f i)) (xStar i) ≤
supportFunctionEReal (effectiveDomain (Set.univ : Set (Fin (m * n) → ℝ)) g)
(section16_unblock (n := n) (m := m) xStar)
Support function of a block-sum effective domain splits by blocks.
open scoped BigOperators
open scoped Pointwise
lemma section16_supportFunction_effectiveDomain_blockSum_unblock_eq_sum {n m : Nat}
(f : Fin m → (Fin n → ℝ) → EReal) (xStar : Fin m → Fin n → ℝ)
(hnotbot : ∀ i : Fin m, ∀ x, f i x ≠ (⊥ : EReal)) :
let g : (Fin (m * n) → ℝ) → EReal :=
fun x => ∑ i, f i (section16_blockLinearMap (n := n) (m := m) i x)
supportFunctionEReal (effectiveDomain (Set.univ : Set (Fin (m * n) → ℝ)) g)
(section16_unblock (n := n) (m := m) xStar) =
∑ i : Fin m,
supportFunctionEReal (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (f i)) (xStar i)
Recession of the block-sum conjugate splits by blocks at an unblocked covector.
open scoped BigOperators
open scoped Pointwise
lemma section16_recession_blockSum_unblock_eq_sum {n m : Nat}
(f : Fin m → (Fin n → ℝ) → EReal) (xStar : Fin m → Fin n → ℝ)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i)) :
let g : (Fin (m * n) → ℝ) → EReal :=
fun x => ∑ i, f i (section16_blockLinearMap (n := n) (m := m) i x)
recessionFunction (fenchelConjugate (m * n) g) (section16_unblock (n := n) (m := m) xStar) =
∑ i : Fin m, recessionFunction (fenchelConjugate n (f i)) (xStar i)§16 — Dual Operations — Part 3
Lean authors: Zichen Wang, Wanli Ma, Weiran Shi, Yunxi Duan, Zaiwen Wen.
Diagonal points in the relative interior of a block-sum domain.
open scoped BigOperators
open scoped Pointwise
lemma section16_diag_mem_ri_effectiveDomain_blockSum_iff {n m : Nat}
(f : Fin m → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i)) :
let g : (Fin (m * n) → ℝ) → EReal :=
fun x => ∑ i, f i (section16_blockLinearMap (n := n) (m := m) i x)
(∃ x : EuclideanSpace ℝ (Fin n),
section16_diagLinearMapE (n := n) (m := m) x ∈
euclideanRelativeInterior (m * n)
((fun z : EuclideanSpace ℝ (Fin (m * n)) => (z : Fin (m * n) → ℝ)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin (m * n) → ℝ)) g)) ↔
Set.Nonempty
(⋂ i : Fin m,
euclideanRelativeInterior n
((fun x : EuclideanSpace ℝ (Fin n) => (x : Fin n → ℝ)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → ℝ)) (f i)))
Corollary 16.2.2. Let f₁, …, fₘ be proper convex functions on ℝⁿ. Then there do not exist vectors x₁⋆, …, xₘ⋆ such that - x₁⋆ + ⋯ + xₘ⋆ = 0, - (f₁⋆0⁺)(x₁⋆) + ⋯ + (fₘ⋆0⁺)(xₘ⋆) ≤ 0, - (f₁⋆0⁺)(-x₁⋆) + ⋯ + (fₘ⋆0⁺)(-xₘ⋆) > 0, if and only if ri (dom f₁) ∩ ⋯ ∩ ri (dom fₘ) ≠ ∅. Here dom fᵢ is the effective domain effectiveDomain univ (f i), ri is euclideanRelativeInterior, and (fᵢ⋆0⁺) is represented as recessionFunction (fenchelConjugate n (f i)).
open scoped BigOperators
open scoped Pointwise
theorem section16_nonempty_iInter_ri_effectiveDomain_iff_not_exists_sum_eq_zero_recession_ineq
{n m : Nat} (f : Fin m → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i)) :
(¬ ∃ xStar : Fin m → Fin n → ℝ,
(∑ i, xStar i) = 0 ∧
(∑ i, recessionFunction (fenchelConjugate n (f i)) (xStar i)) ≤ (0 : EReal) ∧
(∑ i, recessionFunction (fenchelConjugate n (f i)) (-xStar i)) > (0 : EReal)) ↔
Set.Nonempty
(⋂ i : Fin m,
euclideanRelativeInterior n
((fun x : EuclideanSpace ℝ (Fin n) => (x : Fin n → ℝ)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → ℝ)) (f i)))
If (a : EReal) - x = ⊥ with a real, then x = ⊤.
open scoped BigOperators
open scoped Pointwise
lemma section16_coe_sub_eq_bot_iff_eq_top {a : ℝ} {x : EReal} :
(a : EReal) - x = (⊥ : EReal) → x = ⊤
Subtracting an infimum equals the supremum of the translated values.
open scoped BigOperators
open scoped Pointwise
lemma section16_coeReal_sub_sInf_image_eq_sSup_image {α : Type*}
(S : Set α) (φ : α → EReal) (a : ℝ) :
((a : EReal) - sInf (φ '' S)) =
sSup ((fun x => (a : EReal) - φ x) '' S)
Collapsing the sup over fibers to a sup over the total space.
open scoped BigOperators
open scoped Pointwise
lemma section16_sSup_range_sSup_fiber_image_eq_sSup_range_total
{α β : Type*} (A : α → β) (g : α → EReal) :
sSup (Set.range fun y : β => sSup (g '' {x | A x = y})) = sSup (Set.range g)
Dot product with a linear map equals dot product with the adjoint.
open scoped BigOperators
open scoped Pointwise
lemma section16_dotProduct_map_eq_dotProduct_adjoint {n m : Nat}
(A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m))
(x : Fin n → ℝ) (yStar : Fin m → ℝ) :
dotProduct (A (WithLp.toLp 2 x) : Fin m → ℝ) yStar =
dotProduct x
((LinearMap.adjoint A) (WithLp.toLp 2 yStar) : Fin n → ℝ)
Dot-product identity with swapped arguments, to match support-function conventions.
open scoped BigOperators
open scoped Pointwise
lemma section16_dotProduct_adjoint_map_comm {n m : Nat}
(A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m))
(x : Fin n → ℝ) (yStar : Fin m → ℝ) :
dotProduct yStar (A (WithLp.toLp 2 x) : Fin m → ℝ) =
dotProduct ((LinearMap.adjoint A) (WithLp.toLp 2 yStar) : Fin n → ℝ) x
The dot-product image of a linear image is reindexed by the adjoint.
open scoped BigOperators
open scoped Pointwise
lemma section16_image_dotProduct_linearImage_eq_image_dotProduct_adjoint {n m : Nat}
(A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m))
(C : Set (Fin n → ℝ)) (yStar : Fin m → ℝ) :
Set.image (fun y : Fin m → ℝ => dotProduct yStar y)
(Set.image (fun x : Fin n → ℝ => (A (WithLp.toLp 2 x) : Fin m → ℝ)) C) =
Set.image (fun x : Fin n → ℝ =>
dotProduct ((LinearMap.adjoint A) (WithLp.toLp 2 yStar) : Fin n → ℝ) x) C
Reindexing a Set.range along WithLp.toLp does not change the range.
open scoped BigOperators
open scoped Pointwise
lemma section16_range_euclideanSpace_toLp_eq_range {n : Nat}
(g : EuclideanSpace ℝ (Fin n) → EReal) :
Set.range g = Set.range (fun x : Fin n → ℝ => g (WithLp.toLp 2 x))
Theorem 16.3.1: Let A : ℝ^n →ₗ[ℝ] ℝ^m be a linear transformation. For any convex function f on ℝ^n, the Fenchel conjugate of the direct image A f satisfies (A f)^* = f^* ∘ A^*, where (A f) y = inf {f x | A x = y} and A^* is the adjoint of A.
open scoped BigOperators
open scoped Pointwise
theorem section16_fenchelConjugate_linearImage {n m : Nat}
(A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m))
(f : (Fin n → ℝ) → EReal) :
fenchelConjugate m
(fun y =>
sInf
((fun x : EuclideanSpace ℝ (Fin n) => f (x : Fin n → ℝ)) '' {x | (A x : _ ) = y})) =
fun yStar : Fin m → ℝ =>
fenchelConjugate n f
(((LinearMap.adjoint :
(EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) ≃ₗ⋆[ℝ]
(EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n)))
A)
(WithLp.toLp 2 yStar) : Fin n → ℝ)
The first-coordinate projection ℝ² → ℝ as a linear map.
open scoped BigOperators
open scoped Pointwise
def section16_projFirstLinearMap :
EuclideanSpace ℝ (Fin 2) →ₗ[ℝ] EuclideanSpace ℝ (Fin 1)
The fiber image of the first-coordinate projection equals the range over the second coordinate.
open scoped BigOperators
open scoped Pointwise
lemma section16_image_fiber_projFirst_eq_range (f : (Fin 2 → ℝ) → EReal) (ξ1 : Fin 1 → ℝ) :
(fun x : EuclideanSpace ℝ (Fin 2) => f (x : Fin 2 → ℝ)) ''
{x | (section16_projFirstLinearMap x : Fin 1 → ℝ) = ξ1} =
Set.range (fun ξ2 : ℝ => f ![ξ1 0, ξ2])
The adjoint of the first-coordinate projection sends ξ1Star to (ξ1Star, 0).
open scoped BigOperators
open scoped Pointwise
lemma section16_adjoint_projFirst_to_pair (ξ1Star : Fin 1 → ℝ) :
(((LinearMap.adjoint :
(EuclideanSpace ℝ (Fin 2) →ₗ[ℝ] EuclideanSpace ℝ (Fin 1)) ≃ₗ⋆[ℝ]
(EuclideanSpace ℝ (Fin 1) →ₗ[ℝ] EuclideanSpace ℝ (Fin 2)))
section16_projFirstLinearMap)
(WithLp.toLp 2 ξ1Star) : Fin 2 → ℝ) = ![ξ1Star 0, (0 : ℝ)]
Text 16.0.3: As an illustration of Theorem 16.3.1, let f be a convex function on ℝ² and define h : ℝ → ℝ ∪ {±∞} by h(ξ₁) = inf_{ξ₂} f(ξ₁, ξ₂). Equivalently, h is the direct image A f for the projection A : (ξ₁, ξ₂) ↦ ξ₁. The adjoint map A^* sends ξ₁^* to (ξ₁^*, 0), hence h^*(ξ₁^*) = f^*(ξ₁^*, 0).
open scoped BigOperators
open scoped Pointwise
theorem section16_fenchelConjugate_inf_over_second_coordinate
(f : (Fin 2 → ℝ) → EReal) :
let h : (Fin 1 → ℝ) → EReal := fun ξ1 => sInf (Set.range fun ξ2 : ℝ => f ![ξ1 0, ξ2])
fenchelConjugate 1 h = fun ξ1Star : Fin 1 → ℝ => fenchelConjugate 2 f ![ξ1Star 0, (0 : ℝ)]
Corollary 16.3.1.1: Let A be a linear transformation from ℝ^n to ℝ^m. For any convex set C ⊆ ℝ^n, one has δ^*(y^* | A C) = δ^*(A^* y^* | C) for all y^* ∈ ℝ^m.
open scoped BigOperators
open scoped Pointwise
theorem section16_deltaStar_linearImage {n m : Nat}
(A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m))
(C : Set (Fin n → ℝ)) :
deltaStar (Set.image (fun x : Fin n → ℝ => (A (WithLp.toLp 2 x) : Fin m → ℝ)) C) =
fun yStar : Fin m → ℝ =>
deltaStar C
(((LinearMap.adjoint :
(EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) ≃ₗ⋆[ℝ]
(EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n)))
A)
(WithLp.toLp 2 yStar) : Fin n → ℝ)
The conjugate of the support function is the indicator of the closure.
open scoped BigOperators
open scoped Pointwise
lemma section16_fenchelConjugate_supportFunctionEReal_eq_indicatorFunction_closure {m : Nat}
(D : Set (Fin m → ℝ)) (hD : Convex ℝ D) :
fenchelConjugate m (supportFunctionEReal D) = indicatorFunction (closure D)
Rewriting the adjoint fiber using WithLp.toLp.
open scoped BigOperators
open scoped Pointwise
lemma section16_fiber_set_eq_toLp_fiber_set {n m : Nat}
(Aadj : EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n))
(xStar : Fin n → ℝ) :
{yStar : EuclideanSpace ℝ (Fin m) | Aadj yStar = WithLp.toLp 2 xStar} =
{yStar | (Aadj yStar : Fin n → ℝ) = xStar}
Linear images of convex functions via fiberwise sInf are convex.
open scoped BigOperators
open scoped Pointwise
lemma section16_convexFunction_linearImage_sInf_part3 {n m : Nat}
(Aadj : EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n))
(f : (Fin m → ℝ) → EReal) (hf : ConvexFunction f) :
ConvexFunction
(fun xStar : Fin n → ℝ =>
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) => f (yStar : Fin m → ℝ)) ''
{yStar | (Aadj yStar : Fin n → ℝ) = xStar}))
The Fenchel biconjugate of a convex function agrees with its convex-function closure.
open scoped BigOperators
open scoped Pointwise
lemma section16_fenchelConjugate_biconjugate_eq_convexFunctionClosure_part3 {n : Nat}
{f : (Fin n → ℝ) → EReal} (hf : ConvexFunction f) :
fenchelConjugate n (fenchelConjugate n f) = convexFunctionClosure f
Corollary 16.3.1.2: Let A be a linear transformation from ℝ^n to ℝ^m. For any convex set D ⊆ ℝ^m, one has δ^*(· | A⁻¹ (cl D)) = cl (A^* δ^*(· | D)). Here δ^*(·|D) is the support function, cl D is the topological closure of the set D, and A^* is the adjoint of A. In this development, the closure cl of a function is represented by convexFunctionClosure, and A^* δ^*(·|D) is modeled via an sInf over the affine fiber {yStar | A^* yStar = xStar}.
open scoped BigOperators
open scoped Pointwise
theorem section16_supportFunctionEReal_preimage_closure_eq_convexFunctionClosure_adjoint_image
{n m : Nat} (A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m))
(D : Set (Fin m → ℝ)) (hD : Convex ℝ D) :
supportFunctionEReal
(Set.preimage (fun x : Fin n → ℝ => (A (WithLp.toLp 2 x) : Fin m → ℝ)) (closure D)) =
convexFunctionClosure
(fun xStar : Fin n → ℝ =>
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
supportFunctionEReal D (yStar : Fin m → ℝ)) ''
{yStar |
((LinearMap.adjoint :
(EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) ≃ₗ⋆[ℝ]
(EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n)))
A)
yStar =
WithLp.toLp 2 xStar}))§16 — Dual Operations — Part 4
Lean authors: Zichen Wang, Wanli Ma, Weiran Shi, Yunxi Duan, Zaiwen Wen.
The Fenchel biconjugate of a convex function agrees with its convex-function closure.
open scoped BigOperators
open scoped Pointwise
lemma section16_fenchelConjugate_biconjugate_eq_convexFunctionClosure {n : Nat}
{f : (Fin n → ℝ) → EReal} (hf : ConvexFunction f) :
fenchelConjugate n (fenchelConjugate n f) = convexFunctionClosure f
Linear images of convex functions via fiberwise sInf are convex.
open scoped BigOperators
open scoped Pointwise
lemma section16_convexFunction_linearImage_sInf {n m : Nat}
(A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) (f : (Fin n → ℝ) → EReal)
(hf : ConvexFunction f) :
ConvexFunction
(fun y : Fin m → ℝ =>
sInf
((fun x : EuclideanSpace ℝ (Fin n) => f (x : Fin n → ℝ)) '' {x | (A x : _) = y}))
Applying Theorem 16.3.1 to the adjoint gives a biconjugate identity.
open scoped BigOperators
open scoped Pointwise
lemma section16_fenchelConjugate_adjoint_image_eq_precomp_biconjugate {n m : Nat}
(A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m))
(g : (Fin m → ℝ) → EReal) :
let gStar : (Fin m → ℝ) → EReal := fenchelConjugate m g
let h : (Fin n → ℝ) → EReal :=
fun xStar : Fin n → ℝ =>
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) => gStar (yStar : Fin m → ℝ)) ''
{yStar |
((LinearMap.adjoint :
(EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) ≃ₗ⋆[ℝ]
(EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n)))
A)
yStar =
WithLp.toLp 2 xStar})
fenchelConjugate n h =
fun x : Fin n → ℝ =>
fenchelConjugate m gStar (A (WithLp.toLp 2 x) : Fin m → ℝ)
The adjoint-image conjugate equals precomposition by A of the convex-function closure.
open scoped BigOperators
open scoped Pointwise
lemma section16_fenchelConjugate_adjoint_image_eq_precomp_convexFunctionClosure {n m : Nat}
(A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m))
(g : (Fin m → ℝ) → EReal) (hg : ConvexFunction g) :
let gStar : (Fin m → ℝ) → EReal := fenchelConjugate m g
let h : (Fin n → ℝ) → EReal :=
fun xStar : Fin n → ℝ =>
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) => gStar (yStar : Fin m → ℝ)) ''
{yStar |
((LinearMap.adjoint :
(EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) ≃ₗ⋆[ℝ]
(EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n)))
A)
yStar =
WithLp.toLp 2 xStar})
fenchelConjugate n h =
fun x : Fin n → ℝ => convexFunctionClosure g (A (WithLp.toLp 2 x) : Fin m → ℝ)
Theorem 16.3.2: Let A : ℝ^n →ₗ[ℝ] ℝ^m be a linear transformation. For any convex function g on ℝ^m, one has ((cl g)A)^* = cl (A^* g^*). Here cl denotes the convex-function closure (modeled as convexFunctionClosure), g^* is the Fenchel conjugate fenchelConjugate m g, and A^* g^* is the direct image of g^* under the adjoint of A (modeled by an sInf over the affine fiber {yStar | A^* yStar = xStar}).
open scoped BigOperators
open scoped Pointwise
theorem section16_fenchelConjugate_precomp_convexFunctionClosure_eq_convexFunctionClosure_adjoint_image
{n m : Nat} (A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m))
(g : (Fin m → ℝ) → EReal) (hg : ConvexFunction g) :
fenchelConjugate n (fun x => convexFunctionClosure g (A (WithLp.toLp 2 x) : Fin m → ℝ)) =
convexFunctionClosure
(fun xStar : Fin n → ℝ =>
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
fenchelConjugate m g (yStar : Fin m → ℝ)) ''
{yStar |
((LinearMap.adjoint :
(EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) ≃ₗ⋆[ℝ]
(EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n)))
A)
yStar =
WithLp.toLp 2 xStar}))
Corollary 16.3.2.1. Let A : ℝ^n →ₗ[ℝ] ℝ^m be a linear transformation. For any convex set C ⊆ ℝ^n, one has (A C)^* = A^{*-1} (C^*), where C^* denotes the polar set and A^{*-1} denotes the preimage under the adjoint A^*.
open scoped BigOperators
open scoped Pointwise
theorem section16_polar_linearImage {n m : Nat}
(A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m))
(C : Set (Fin n → ℝ)) :
{yStar : Fin m → ℝ |
∀ y ∈ Set.image (fun x : Fin n → ℝ => (A (WithLp.toLp 2 x) : Fin m → ℝ)) C,
(dotProduct y yStar : ℝ) ≤ 1} =
Set.preimage
(fun yStar : Fin m → ℝ =>
(((LinearMap.adjoint :
(EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) ≃ₗ⋆[ℝ]
(EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n)))
A)
(WithLp.toLp 2 yStar) : Fin n → ℝ))
{xStar : Fin n → ℝ | ∀ x ∈ C, (dotProduct x xStar : ℝ) ≤ 1}
Closure of a ≤ 1 sublevel set matches the ≤ 1 sublevel of the convex-function closure.
open scoped BigOperators
open scoped Pointwise
lemma section16_closure_sublevel_eq_sublevel_convexFunctionClosure_one {n : Nat}
{h : (Fin n → ℝ) → EReal}
(hh : ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) h)
(hInf : iInf (fun x => h x) < (1 : EReal)) :
closure {x : EuclideanSpace ℝ (Fin n) | h (x : Fin n → ℝ) ≤ (1 : EReal)} =
{x : EuclideanSpace ℝ (Fin n) |
convexFunctionClosure h (x : Fin n → ℝ) ≤ (1 : EReal)}
The adjoint image of the support-function sublevel lies in the fiberwise sInf sublevel.
open scoped BigOperators
open scoped Pointwise
lemma section16_image_adjoint_polar_subset_sublevel_sInf {n m : Nat}
(Aadj : EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n))
(D : Set (Fin m → ℝ)) :
Set.image
(fun yStar : Fin m → ℝ => (Aadj (WithLp.toLp 2 yStar) : Fin n → ℝ))
{yStar : Fin m → ℝ | supportFunctionEReal D yStar ≤ (1 : EReal)} ⊆
{xStar : Fin n → ℝ |
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
supportFunctionEReal D (yStar : Fin m → ℝ)) ''
{yStar | Aadj yStar = WithLp.toLp 2 xStar}) ≤ (1 : EReal)}
The fiberwise sInf at the origin is bounded by 0.
open scoped BigOperators
open scoped Pointwise
lemma section16_sInf_fiber_zero_le {n m : Nat}
(Aadj : EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n))
(D : Set (Fin m → ℝ)) :
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
supportFunctionEReal D (yStar : Fin m → ℝ)) ''
{yStar | Aadj yStar = WithLp.toLp 2 (0 : Fin n → ℝ)}) ≤
(0 : EReal)
Reverse inclusion for section16_closure_image_adjoint_polar_eq_closure_sublevel_sInf.
open scoped BigOperators
open scoped Pointwise
lemma section16_closure_image_adjoint_polar_eq_closure_sublevel_sInf_reverse {n m : Nat}
(Aadj : EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n))
(D : Set (Fin m → ℝ)) :
{xStar : Fin n → ℝ |
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
supportFunctionEReal D (yStar : Fin m → ℝ)) ''
{yStar | Aadj yStar = WithLp.toLp 2 xStar}) ≤ (1 : EReal)} ⊆
closure
(Set.image
(fun yStar : Fin m → ℝ => (Aadj (WithLp.toLp 2 yStar) : Fin n → ℝ))
{yStar : Fin m → ℝ | supportFunctionEReal D yStar ≤ (1 : EReal)})
The closure of the adjoint image of the polar sublevel equals the closure of the sInf sublevel.
open scoped BigOperators
open scoped Pointwise
lemma section16_closure_image_adjoint_polar_eq_closure_sublevel_sInf {n m : Nat}
(Aadj : EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n))
(D : Set (Fin m → ℝ)) :
closure
(Set.image
(fun yStar : Fin m → ℝ => (Aadj (WithLp.toLp 2 yStar) : Fin n → ℝ))
{yStar : Fin m → ℝ | supportFunctionEReal D yStar ≤ (1 : EReal)}) =
closure
{xStar : Fin n → ℝ |
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
supportFunctionEReal D (yStar : Fin m → ℝ)) ''
{yStar | Aadj yStar = WithLp.toLp 2 xStar}) ≤ (1 : EReal)}
Properness of the fiberwise sInf under 0 ∈ closure D.
open scoped BigOperators
open scoped Pointwise
lemma section16_properConvexFunctionOn_h_of_zero_mem_closure {n m : Nat}
(Aadj : EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n))
(D : Set (Fin m → ℝ)) (hD : Convex ℝ D)
(h0D : (0 : Fin m → ℝ) ∈ closure D) :
ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ))
(fun xStar : Fin n → ℝ =>
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
supportFunctionEReal D (yStar : Fin m → ℝ)) ''
{yStar | Aadj yStar = WithLp.toLp 2 xStar}))
The fiberwise infimum is strictly below 1 under 0 ∈ closure D.
open scoped BigOperators
open scoped Pointwise
lemma section16_iInf_h_lt_one_of_zero_mem_closure {n m : Nat}
(Aadj : EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n))
(D : Set (Fin m → ℝ)) :
iInf
(fun xStar : Fin n → ℝ =>
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
supportFunctionEReal D (yStar : Fin m → ℝ)) ''
{yStar | Aadj yStar = WithLp.toLp 2 xStar})) <
(1 : EReal)
Corollary 16.3.2.2. Let A be a linear transformation from ℝ^n to ℝ^m. For any convex set D ⊆ ℝ^m with 0 ∈ cl D, one has (A⁻¹ (cl D))^* = cl (A^* (D^*)), where cl is the topological closure of sets, D^* is the polar set, and A^* is the adjoint of A.
open scoped BigOperators
open scoped Pointwise
theorem section16_polar_preimage_closure_eq_closure_adjoint_image {n m : Nat}
(A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m))
(D : Set (Fin m → ℝ)) (hD : Convex ℝ D) (h0D : (0 : Fin m → ℝ) ∈ closure D) :
{xStar : Fin n → ℝ |
∀ x ∈ Set.preimage (fun x : Fin n → ℝ => (A (WithLp.toLp 2 x) : Fin m → ℝ)) (closure D),
(dotProduct x xStar : ℝ) ≤ 1} =
closure
(Set.image
(fun yStar : Fin m → ℝ =>
(((LinearMap.adjoint :
(EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) ≃ₗ⋆[ℝ]
(EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n)))
A)
(WithLp.toLp 2 yStar) : Fin n → ℝ))
{yStar : Fin m → ℝ | ∀ y ∈ D, (dotProduct y yStar : ℝ) ≤ 1})
Rewriting sInf of an image by an explicit existential description.
open scoped BigOperators
open scoped Pointwise
lemma section16_sInf_image_fiber_eq_sInf_exists {α : Type*} (φ : α → EReal) (S : Set α) :
sInf (φ '' S) = sInf {z : EReal | ∃ y ∈ S, z = φ y}
For a proper function, the recession function agrees with the unrestricted formula.
open scoped BigOperators
open scoped Pointwise
lemma section16_recessionFunction_eq_sSup_unrestricted {n : Nat}
{f : (Fin n → ℝ) → EReal} (y : Fin n → ℝ) :
recessionFunction f y =
sSup { r : EReal | ∃ x : Fin n → ℝ, r = f (x + y) - f x }§16 — Dual Operations — Part 5
Lean authors: Zichen Wang, Wanli Ma, Weiran Shi, Yunxi Duan, Zaiwen Wen.
In the improper case, the conjugates are identically ⊤.
open scoped BigOperators
open scoped Pointwise
lemma section16_fenchelConjugate_precomp_eq_top_of_improper_of_exists_mem_ri
{n m : Nat} (A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m))
(g : (Fin m → ℝ) → EReal) (hg : ConvexFunction g)
(hri :
∃ x : EuclideanSpace ℝ (Fin n),
A x ∈
euclideanRelativeInterior m
((fun z : EuclideanSpace ℝ (Fin m) => (z : Fin m → ℝ)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin m → ℝ)) g))
(hproper : ¬ ProperConvexFunctionOn (Set.univ : Set (Fin m → ℝ)) g) :
(fenchelConjugate n (fun x => g (A (WithLp.toLp 2 x) : Fin m → ℝ)) = constPosInf n) ∧
(fenchelConjugate m g = constPosInf m)
Closure commutes with linear precomposition under a relative-interior preimage point.
open scoped BigOperators
open scoped Pointwise
lemma section16_convexFunctionClosure_precomp_eq_precomp_convexFunctionClosure_of_exists_mem_ri
{n m : Nat} (A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m))
{g : (Fin m → ℝ) → EReal}
(hgproper : ProperConvexFunctionOn (Set.univ : Set (Fin m → ℝ)) g)
(hri :
∃ x : EuclideanSpace ℝ (Fin n),
A x ∈
euclideanRelativeInterior m
((fun z : EuclideanSpace ℝ (Fin m) => (z : Fin m → ℝ)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin m → ℝ)) g)) :
convexFunctionClosure (fun x : Fin n → ℝ => g (A (WithLp.toLp 2 x) : Fin m → ℝ)) =
fun x => convexFunctionClosure g (A (WithLp.toLp 2 x) : Fin m → ℝ)
Closedness and attainment for the adjoint-image infimum under the ri condition.
open scoped BigOperators
open scoped Pointwise
lemma section16_adjointImage_closed_and_attained_of_exists_mem_ri_effectiveDomain
{n m : Nat} (A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m))
(g : (Fin m → ℝ) → EReal)
(hgproper : ProperConvexFunctionOn (Set.univ : Set (Fin m → ℝ)) g)
(hri :
∃ x : EuclideanSpace ℝ (Fin n),
A x ∈
euclideanRelativeInterior m
((fun z : EuclideanSpace ℝ (Fin m) => (z : Fin m → ℝ)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin m → ℝ)) g)) :
let Aadj :
EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n) :=
((LinearMap.adjoint :
(EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) ≃ₗ⋆[ℝ]
(EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n)))
A)
let gStar : (Fin m → ℝ) → EReal := fenchelConjugate m g
let h : (Fin n → ℝ) → EReal :=
fun xStar : Fin n → ℝ =>
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) => gStar (yStar : Fin m → ℝ)) ''
{yStar | Aadj yStar = WithLp.toLp 2 xStar})
(convexFunctionClosure h = h) ∧
∀ xStar : Fin n → ℝ,
h xStar = ⊤ ∨
∃ yStar : EuclideanSpace ℝ (Fin m),
Aadj yStar = WithLp.toLp 2 xStar ∧
gStar (yStar : Fin m → ℝ) = h xStar
Properness and ri-compatibility for indicator functions under a feasible point.
open scoped BigOperators
open scoped Pointwise
lemma section16_properConvexFunctionOn_indicatorFunction_of_exists_mem_ri {n m : Nat}
(A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) (D : Set (Fin m → ℝ))
(hD : Convex ℝ D)
(hri :
∃ x : EuclideanSpace ℝ (Fin n),
A x ∈
euclideanRelativeInterior m
((fun z : EuclideanSpace ℝ (Fin m) => (z : Fin m → ℝ)) ⁻¹' D)) :
ProperConvexFunctionOn (Set.univ : Set (Fin m → ℝ)) (indicatorFunction D) ∧
∃ x : EuclideanSpace ℝ (Fin n),
A x ∈
euclideanRelativeInterior m
((fun z : EuclideanSpace ℝ (Fin m) => (z : Fin m → ℝ)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin m → ℝ)) (indicatorFunction D))
Precomposition of an indicator is the indicator of a preimage set.
open scoped BigOperators
open scoped Pointwise
lemma section16_indicator_precomp_eq_indicator_preimage {n m : Nat}
(A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) (D : Set (Fin m → ℝ)) :
(fun x : Fin n → ℝ => indicatorFunction D (A (WithLp.toLp 2 x) : Fin m → ℝ)) =
indicatorFunction
(Set.preimage (fun x : Fin n → ℝ => (A (WithLp.toLp 2 x) : Fin m → ℝ)) D)§16 — Dual Operations — Part 6
Lean authors: Zichen Wang, Wanli Ma, Weiran Shi, Yunxi Duan, Zaiwen Wen.
Corollary 16.3.1.3: Let A : ℝ^n →ₗ[ℝ] ℝ^m be a linear map and let D ⊆ ℝ^m be convex. If there exists x such that A x ∈ ri D, then the closure on D can be omitted in Corollary 16.3.1.2. Equivalently, for every xStar, δ^*(xStar | A⁻¹ D) = inf { δ^*(yStar | D) | A^* yStar = xStar }, and the infimum is attained (or the value is +∞ if the affine fiber is empty).
open scoped BigOperators
open scoped Pointwise
theorem section16_supportFunctionEReal_preimage_eq_adjoint_image_of_exists_mem_ri {n m : Nat}
(A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) (D : Set (Fin m → ℝ))
(hD : Convex ℝ D)
(hri :
∃ x : EuclideanSpace ℝ (Fin n),
A x ∈
euclideanRelativeInterior m
((fun z : EuclideanSpace ℝ (Fin m) => (z : Fin m → ℝ)) ⁻¹' D)) :
let Aadj : EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n) :=
((LinearMap.adjoint :
(EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) ≃ₗ⋆[ℝ]
(EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n)))
A)
((supportFunctionEReal
(Set.preimage (fun x : Fin n → ℝ => (A (WithLp.toLp 2 x) : Fin m → ℝ)) D) =
fun xStar : Fin n → ℝ =>
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
supportFunctionEReal D (yStar : Fin m → ℝ)) ''
{yStar | Aadj yStar = WithLp.toLp 2 xStar})) ∧
∀ xStar : Fin n → ℝ,
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
supportFunctionEReal D (yStar : Fin m → ℝ)) ''
{yStar | Aadj yStar = WithLp.toLp 2 xStar}) =
⊤ ∨
∃ yStar : EuclideanSpace ℝ (Fin m),
Aadj yStar = WithLp.toLp 2 xStar ∧
supportFunctionEReal D (yStar : Fin m → ℝ) =
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
supportFunctionEReal D (yStar : Fin m → ℝ)) ''
{yStar | Aadj yStar = WithLp.toLp 2 xStar}))§16 — Dual Operations — Part 7
Lean authors: Zichen Wang, Wanli Ma, Weiran Shi, Yunxi Duan, Zaiwen Wen.
Theorem 16.3.3: Let A : ℝ^n →ₗ[ℝ] ℝ^m be a linear transformation. For a convex function g on ℝ^m, if there exists x with A x ∈ ri (dom g), then the closure operation in Theorem 16.3.2 can be omitted, i.e. (g ∘ A)^* = A^* g^*. Equivalently, for each xStar, one has (g ∘ A)^*(xStar) = inf { g^*(yStar) | A^* yStar = xStar }, and the infimum is attained (or the value is +∞ if the affine fiber is empty).
open scoped BigOperators
open scoped Pointwise
theorem section16_fenchelConjugate_precomp_eq_adjoint_image_of_exists_mem_ri_effectiveDomain
{n m : Nat} (A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m))
(g : (Fin m → ℝ) → EReal) (hg : ConvexFunction g)
(hri :
∃ x : EuclideanSpace ℝ (Fin n),
A x ∈
euclideanRelativeInterior m
((fun z : EuclideanSpace ℝ (Fin m) => (z : Fin m → ℝ)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin m → ℝ)) g)) :
let Aadj :
EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n) :=
((LinearMap.adjoint :
(EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) ≃ₗ⋆[ℝ]
(EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n)))
A)
((fenchelConjugate n (fun x => g (A (WithLp.toLp 2 x) : Fin m → ℝ)) =
fun xStar : Fin n → ℝ =>
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
fenchelConjugate m g (yStar : Fin m → ℝ)) ''
{yStar | Aadj yStar = WithLp.toLp 2 xStar})) ∧
∀ xStar : Fin n → ℝ,
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
fenchelConjugate m g (yStar : Fin m → ℝ)) ''
{yStar | Aadj yStar = WithLp.toLp 2 xStar}) =
⊤ ∨
∃ yStar : EuclideanSpace ℝ (Fin m),
Aadj yStar = WithLp.toLp 2 xStar ∧
fenchelConjugate m g (yStar : Fin m → ℝ) =
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
fenchelConjugate m g (yStar : Fin m → ℝ)) ''
{yStar | Aadj yStar = WithLp.toLp 2 xStar}))
Text 16.0.4: Example of the conjugate of a function built from one-dimensional convex functions and linear functionals. Let h : ℝⁿ → ℝ ∪ {±∞} be given by h(x) = g₁(⟪a₁, x⟫) + ⋯ + gₘ(⟪aₘ, x⟫), where each gᵢ is closed proper convex on ℝ and a₁, …, aₘ ∈ ℝⁿ. Writing h = g ∘ A where A x = (⟪a₁, x⟫, …, ⟪aₘ, x⟫) and g(η₁, …, ηₘ) = g₁(η₁) + ⋯ + gₘ(ηₘ), one has g^*(η₁⋆, …, ηₘ⋆) = g₁^*(η₁⋆) + ⋯ + gₘ^*(ηₘ⋆). Therefore (A^* g^*)(x⋆) is the infimum of g₁^*(η₁⋆) + ⋯ + gₘ^*(ηₘ⋆) over all (η₁⋆, …, ηₘ⋆) such that η₁⋆ a₁ + ⋯ + ηₘ⋆ aₘ = x⋆. The conjugate h^* is the closure of A^* g^* (Theorem 16.3.2). If there exists x with ⟪aᵢ, x⟫ ∈ ri (dom gᵢ) for i = 1, …, m, then the infimum is attained and h^* = A^* g^* (Theorem 16.3.3).
open scoped BigOperators
open scoped Pointwise
theorem section16_fenchelConjugate_sum_inner_precomp_of_exists_mem_ri_effectiveDomain
{n m : Nat} (a : Fin m → EuclideanSpace ℝ (Fin n)) (g1 : Fin m → ℝ → EReal)
(hg : ConvexFunction (fun y : Fin m → ℝ => ∑ i, g1 i (y i)))
(hri :
∃ x : EuclideanSpace ℝ (Fin n),
(((WithLp.linearEquiv (2 : ENNReal) ℝ (Fin m → ℝ)).symm.toLinearMap).comp
(LinearMap.pi fun i => (innerSL ℝ (a i)).toLinearMap))
x ∈
euclideanRelativeInterior m
((fun z : EuclideanSpace ℝ (Fin m) => (z : Fin m → ℝ)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin m → ℝ))
(fun y : Fin m → ℝ => ∑ i, g1 i (y i)))) :
let g : (Fin m → ℝ) → EReal := fun y => ∑ i, g1 i (y i)
let A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m) :=
(((WithLp.linearEquiv (2 : ENNReal) ℝ (Fin m → ℝ)).symm.toLinearMap).comp
(LinearMap.pi fun i => (innerSL ℝ (a i)).toLinearMap))
let h : (Fin n → ℝ) → EReal := fun x => g (A (WithLp.toLp 2 x) : Fin m → ℝ)
let Aadj : EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n) :=
((LinearMap.adjoint :
(EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) ≃ₗ⋆[ℝ]
(EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n)))
A)
((fenchelConjugate n h =
fun xStar : Fin n → ℝ =>
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
fenchelConjugate m g (yStar : Fin m → ℝ)) ''
{yStar | Aadj yStar = WithLp.toLp 2 xStar})) ∧
∀ xStar : Fin n → ℝ,
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
fenchelConjugate m g (yStar : Fin m → ℝ)) ''
{yStar | Aadj yStar = WithLp.toLp 2 xStar}) =
⊤ ∨
∃ yStar : EuclideanSpace ℝ (Fin m),
Aadj yStar = WithLp.toLp 2 xStar ∧
fenchelConjugate m g (yStar : Fin m → ℝ) =
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
fenchelConjugate m g (yStar : Fin m → ℝ)) ''
{yStar | Aadj yStar = WithLp.toLp 2 xStar}))
Text 16.0.5: Interpreting the identity (g ∘ A)^* = A^* g^* (in the case where no closure is needed) as a sup/inf formula. For any xStar ∈ ℝⁿ, one has sup {⟪x, xStar⟫ - g (A x) | x ∈ ℝⁿ} = inf {g^* yStar | A^* yStar = xStar}.
open scoped BigOperators
open scoped Pointwise
lemma section16_sup_dotProduct_sub_precomp_eq_sInf_fenchelConjugate_fiber {n m : Nat}
(A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) (g : (Fin m → ℝ) → EReal)
(xStar : Fin n → ℝ)
(hEq :
let Aadj :
EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n) :=
((LinearMap.adjoint :
(EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) ≃ₗ⋆[ℝ]
(EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n)))
A)
fenchelConjugate n (fun x => g (A (WithLp.toLp 2 x) : Fin m → ℝ)) =
fun xStar : Fin n → ℝ =>
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
fenchelConjugate m g (yStar : Fin m → ℝ)) ''
{yStar | Aadj yStar = WithLp.toLp 2 xStar})) :
let Aadj :
EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n) :=
((LinearMap.adjoint :
(EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) ≃ₗ⋆[ℝ]
(EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n)))
A)
sSup
(Set.range fun x : Fin n → ℝ =>
((dotProduct x xStar : ℝ) : EReal) - g (A (WithLp.toLp 2 x) : Fin m → ℝ)) =
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
fenchelConjugate m g (yStar : Fin m → ℝ)) ''
{yStar | Aadj yStar = WithLp.toLp 2 xStar})
A finite sum of EReal values is not ⊥ if all terms are not ⊥.
open scoped BigOperators
open scoped Pointwise
lemma section16_sum_ne_bot_of_ne_bot {ι : Type*} (s : Finset ι) (b : ι → EReal)
(hb : ∀ i ∈ s, b i ≠ ⊥) : (Finset.sum s b) ≠ ⊥
Negation distributes over finite sums of non-⊥ EReal values.
open scoped BigOperators
open scoped Pointwise
lemma section16_neg_sum_eq_sum_neg {ι : Type*} (s : Finset ι) (b : ι → EReal)
(hb : ∀ i ∈ s, b i ≠ ⊥) :
-(Finset.sum s b) = Finset.sum s (fun i => -b i)
Supremum of a sum of independent variables equals the sum of suprema.
open scoped BigOperators
open scoped Pointwise
lemma section16_iSup_add_iSup_eq_iSup_prod {α β : Type*} (u : α → EReal) (v : β → EReal) :
(iSup u) + (iSup v) = iSup (fun p : α × β => u p.1 + v p.2)
Supremum over a Fin-indexed sum equals the sum of suprema.
open scoped BigOperators
open scoped Pointwise
lemma section16_iSup_fin_sum_eq_sum_iSup {n m : Nat}
(g : Fin m → (Fin n → ℝ) → EReal) :
iSup (fun x' : Fin m → Fin n → ℝ => ∑ i, g i (x' i)) =
∑ i, iSup (fun x : Fin n → ℝ => g i x)
Coercion of a real finite sum into EReal equals the finite sum of coercions.
open scoped BigOperators
open scoped Pointwise
lemma section16_coe_finset_sum {ι : Type*} (s : Finset ι) (b : ι → ℝ) :
((Finset.sum s b : ℝ) : EReal) = Finset.sum s (fun i => (b i : EReal))
Coercion of a dot-product sum into EReal equals the sum of coerced dot products.
open scoped BigOperators
open scoped Pointwise
lemma section16_coe_sum_dotProduct_eq_sum_coe_dotProduct {n m : Nat}
(x' : Fin m → Fin n → ℝ) (xStar : Fin n → ℝ) :
(((∑ i, x' i ⬝ᵥ xStar) : ℝ) : EReal) =
∑ i, (((x' i) ⬝ᵥ xStar : ℝ) : EReal)
Rewrite a dot-product minus a sum as a sum of dot-product differences.
open scoped BigOperators
open scoped Pointwise
lemma section16_sum_dotProduct_sub_sum_eq_sum_sub {n m : Nat}
(x' : Fin m → Fin n → ℝ) (xStar : Fin n → ℝ)
(f : Fin m → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i)) :
(((∑ i, x' i) ⬝ᵥ xStar : ℝ) : EReal) - ∑ i, f i (x' i) =
∑ i, ((((x' i) ⬝ᵥ xStar : ℝ) : EReal) - f i (x' i))
Theorem 16.4.1: Let f₁, …, fₘ be proper convex functions on ℝⁿ. Then the Fenchel conjugate of their infimal convolution is the sum of their Fenchel conjugates: (f₁ □ ⋯ □ fₘ)^* = f₁^* + ⋯ + fₘ^*. In this development, f₁ □ ⋯ □ fₘ is infimalConvolutionFamily f, and fᵢ^* is fenchelConjugate n (f i).
open scoped BigOperators
open scoped Pointwise
theorem section16_fenchelConjugate_infimalConvolutionFamily {n m : Nat}
(f : Fin m → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i)) :
fenchelConjugate n (infimalConvolutionFamily f) =
fun xStar => ∑ i, fenchelConjugate n (f i) xStar
The infimal convolution of indicator functions is the indicator of the Minkowski sum.
open scoped BigOperators
open scoped Pointwise
lemma section16_infimalConvolutionFamily_indicatorFunction_eq_indicatorFunction_sum {n m : Nat}
(C : Fin m → Set (Fin n → ℝ)) :
infimalConvolutionFamily (fun i => indicatorFunction (C i)) = indicatorFunction (∑ i, C i)
Corollary 16.4.1.1: Let C₁, …, Cₘ be non-empty convex sets in ℝⁿ. Then the support function δ^*(· | C) sends Minkowski sums to pointwise sums: δ^*(· | C₁ + ⋯ + Cₘ) = δ^*(· | C₁) + ⋯ + δ^*(· | Cₘ).
open scoped BigOperators
open scoped Pointwise
theorem section16_deltaStar_sum {n m : Nat} (C : Fin m → Set (Fin n → ℝ))
(hC : ∀ i, Convex ℝ (C i)) (hCne : ∀ i, (C i).Nonempty) :
supportFunctionEReal (∑ i, C i) = fun xStar => ∑ i, supportFunctionEReal (C i) xStar
The convex-function closure of an indicator function is the indicator of the set closure.
open scoped BigOperators
open scoped Pointwise
lemma section16_convexFunctionClosure_indicatorFunction_eq_indicatorFunction_closure {n : Nat}
(C : Set (Fin n → ℝ)) (hC : Convex ℝ C) (hCne : C.Nonempty) :
convexFunctionClosure (indicatorFunction C) = indicatorFunction (closure C)
The sum of indicator functions is the indicator of the intersection.
open scoped BigOperators
open scoped Pointwise
lemma section16_sum_indicatorFunction_eq_indicatorFunction_iInter {n m : Nat}
(C : Fin m → Set (Fin n → ℝ)) :
(fun x => ∑ i, indicatorFunction (C i) x) = indicatorFunction (⋂ i, C i)
Corollary 16.4.1.2. Let C₁, …, Cₘ be non-empty convex sets in ℝⁿ. Then the support function of the intersection of their closures is the convex-function closure of the infimal convolution of the support functions: δ^*(· | cl C₁ ∩ ⋯ ∩ cl Cₘ) = cl (δ^*(· | C₁) ⊕ ⋯ ⊕ δ^*(· | Cₘ)). In this development, δ^* is supportFunctionEReal, cl is convexFunctionClosure, and ⊕ is modeled by infimalConvolutionFamily.
open scoped BigOperators
open scoped Pointwise
theorem section16_supportFunctionEReal_iInter_closure_eq_convexFunctionClosure_infimalConvolutionFamily
{n m : Nat} (C : Fin m → Set (Fin n → ℝ)) (hC : ∀ i, Convex ℝ (C i))
(hCne : ∀ i, (C i).Nonempty) :
supportFunctionEReal (⋂ i, closure (C i)) =
convexFunctionClosure (infimalConvolutionFamily fun i => supportFunctionEReal (C i))
The conjugate of the infimal convolution of conjugates is the sum of closures.
open scoped BigOperators
open scoped Pointwise
lemma section16_fenchelConjugate_infimalConvolutionFamily_conjugates_eq_sum_closure {n m : Nat}
(f : Fin m → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i)) :
fenchelConjugate n (infimalConvolutionFamily fun i => fenchelConjugate n (f i)) =
fun xStar => ∑ i, convexFunctionClosure (f i) xStar
The infimal convolution of conjugates is a convex function.
open scoped BigOperators
open scoped Pointwise
lemma section16_convexFunction_infimalConvolutionFamily_conjugates {n m : Nat}
(f : Fin m → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i)) :
ConvexFunction (infimalConvolutionFamily fun i => fenchelConjugate n (f i))
Conjugating the sum-closure identity yields the closure statement.
open scoped BigOperators
open scoped Pointwise
lemma section16_fenchelConjugate_sum_convexFunctionClosure_eq_convexFunctionClosure_infimalConvolutionFamily_proof_step
{n m : Nat} (f : Fin m → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i)) :
fenchelConjugate n (fun x => ∑ i, convexFunctionClosure (f i) x) =
convexFunctionClosure (infimalConvolutionFamily fun i => fenchelConjugate n (f i))
Theorem 16.4.2: Let f₁, …, fₘ be proper convex functions on ℝⁿ. Then (cl f₁ + ⋯ + cl fₘ)^* = cl (f₁^* □ ⋯ □ fₘ^*). Here cl is the convex-function closure convexFunctionClosure, ^* is the Fenchel conjugate fenchelConjugate n, the sum is pointwise (written as ∑ i), and □ is the infimal convolution family infimalConvolutionFamily.
open scoped BigOperators
open scoped Pointwise
theorem section16_fenchelConjugate_sum_convexFunctionClosure_eq_convexFunctionClosure_infimalConvolutionFamily
{n m : Nat} (f : Fin m → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i)) :
fenchelConjugate n (fun x => ∑ i, convexFunctionClosure (f i) x) =
convexFunctionClosure (infimalConvolutionFamily fun i => fenchelConjugate n (f i))
Under a common relative-interior point, the sum of closures equals the closure of the sum.
open scoped BigOperators
open scoped Pointwise
lemma section16_sum_convexFunctionClosure_eq_convexFunctionClosure_sum_of_nonempty_iInter_ri_effectiveDomain
{n m : Nat} (f : Fin m → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i))
(hri :
Set.Nonempty
(⋂ i : Fin m,
euclideanRelativeInterior n
((fun x : EuclideanSpace ℝ (Fin n) => (x : Fin n → ℝ)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → ℝ)) (f i)))) :
(fun x => ∑ i, convexFunctionClosure (f i) x) =
convexFunctionClosure (fun x => ∑ i, f i x)§16 — Dual Operations — Part 8
Lean authors: Zichen Wang, Wanli Ma, Weiran Shi, Yunxi Duan, Zaiwen Wen.
Specializing Theorem 16.4.3 to indicator functions yields the support-function identity and the usual attainment/⊤ disjunction.
open scoped BigOperators
open scoped Pointwise
lemma section16_supportFunctionEReal_iInter_eq_infimalConvolutionFamily_of_nonempty_iInter_ri_proof_step
{n m : Nat} (C : Fin m → Set (Fin n → ℝ)) (hC : ∀ i, Convex ℝ (C i))
(hCne : ∀ i, (C i).Nonempty)
(hri :
Set.Nonempty
(⋂ i : Fin m,
euclideanRelativeInterior n
((fun x : EuclideanSpace ℝ (Fin n) => (x : Fin n → ℝ)) ⁻¹' (C i)))) :
supportFunctionEReal (⋂ i, C i) =
infimalConvolutionFamily (fun i => supportFunctionEReal (C i)) ∧
∀ xStar : Fin n → ℝ,
infimalConvolutionFamily (fun i => supportFunctionEReal (C i)) xStar = ⊤ ∨
∃ xStarFamily : Fin m → Fin n → ℝ,
(∑ i, xStarFamily i) = xStar ∧
(∑ i, supportFunctionEReal (C i) (xStarFamily i)) =
infimalConvolutionFamily (fun i => supportFunctionEReal (C i)) xStar
When m > 0, any xStar admits a family summing to xStar.
open scoped BigOperators
open scoped Pointwise
lemma section16_exists_xStarFamily_sum_eq_of_pos {n m : Nat} (hm : 0 < m)
(xStar : Fin n → ℝ) :
∃ xStarFamily : Fin m → Fin n → ℝ, (∑ i, xStarFamily i) = xStar
If the infimal convolution is ⊤ and m > 0, any fixed decomposition attains ⊤.
open scoped BigOperators
open scoped Pointwise
lemma section16_attainment_when_infimalConvolutionFamily_eq_top_of_pos {n m : Nat} (hm : 0 < m)
(g : Fin m → (Fin n → ℝ) → EReal) (xStar : Fin n → ℝ)
(htop : infimalConvolutionFamily g xStar = ⊤) :
∃ xStarFamily : Fin m → Fin n → ℝ,
(∑ i, xStarFamily i) = xStar ∧
(∑ i, g i (xStarFamily i)) = infimalConvolutionFamily g xStar
Corollary 16.4.1.3: Let C₁, …, Cₘ be non-empty convex sets in ℝⁿ. If the sets ri Cᵢ have a point in common, then the closure operation can be omitted from Corollary 16.4.1.2, and δ^*(· | C₁ ∩ ⋯ ∩ Cₘ) = inf { δ^*(· | C₁) + ⋯ + δ^*(· | Cₘ) | x₁⋆ + ⋯ + xₘ⋆ = x⋆ }, where for each x⋆ the infimum is attained. In this development, δ^* is supportFunctionEReal, ri is euclideanRelativeInterior, and the right-hand side is modeled by infimalConvolutionFamily.
open scoped BigOperators
open scoped Pointwise
theorem section16_supportFunctionEReal_iInter_eq_infimalConvolutionFamily_of_nonempty_iInter_ri
{n m : Nat} (C : Fin m → Set (Fin n → ℝ)) (hC : ∀ i, Convex ℝ (C i))
(hCne : ∀ i, (C i).Nonempty)
(hri :
Set.Nonempty
(⋂ i : Fin m,
euclideanRelativeInterior n
((fun x : EuclideanSpace ℝ (Fin n) => (x : Fin n → ℝ)) ⁻¹' (C i))))
(hm : 0 < m) :
supportFunctionEReal (⋂ i, C i) = infimalConvolutionFamily (fun i => supportFunctionEReal (C i)) ∧
∀ xStar : Fin n → ℝ,
∃ xStarFamily : Fin m → Fin n → ℝ,
(∑ i, xStarFamily i) = xStar ∧
(∑ i, supportFunctionEReal (C i) (xStarFamily i)) =
infimalConvolutionFamily (fun i => supportFunctionEReal (C i)) xStar§16 — Dual Operations — Part 9
Lean authors: Zichen Wang, Wanli Ma, Weiran Shi, Yunxi Duan, Zaiwen Wen.
Taking the convex-function closure does not change the Fenchel conjugate.
open scoped BigOperators
open scoped Pointwise
lemma section16_fenchelConjugate_convexFunctionClosure_eq {n : Nat}
(g : (Fin n → ℝ) → EReal) :
fenchelConjugate n (convexFunctionClosure g) = fenchelConjugate n g
The Fenchel conjugate is closed and convex.
open scoped BigOperators
open scoped Pointwise
lemma section16_closedConvexFunction_fenchelConjugate {n : Nat}
(g : (Fin n → ℝ) → EReal) :
ClosedConvexFunction (fenchelConjugate n g)
The recession function of a Fenchel conjugate is proper and positively homogeneous.
open scoped BigOperators
open scoped Pointwise
lemma section16_recessionFunction_fenchelConjugate_proper_posHom {n : Nat}
(f : (Fin n → ℝ) → EReal)
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) f) :
ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (recessionFunction (fenchelConjugate n f)) ∧
PositivelyHomogeneous (recessionFunction (fenchelConjugate n f))
The epigraph recession cone of a Fenchel conjugate is the epigraph of its recession function.
open scoped BigOperators
open scoped Pointwise
lemma section16_recessionCone_epigraph_eq_epigraph_recessionFunction_fenchelConjugate {n : Nat}
(f : (Fin n → ℝ) → EReal)
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) f) :
Set.recessionCone (epigraph (Set.univ : Set (Fin n → ℝ)) (fenchelConjugate n f)) =
epigraph (Set.univ : Set (Fin n → ℝ)) (recessionFunction (fenchelConjugate n f))
Closedness and attainment for the infimal convolution of conjugates under the ri condition.
open scoped BigOperators
open scoped Pointwise
lemma section16_infimalConvolutionFamily_fenchelConjugate_closed_and_attained_of_nonempty_iInter_ri
{n m : Nat} (f : Fin m → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i))
(hri :
Set.Nonempty
(⋂ i : Fin m,
euclideanRelativeInterior n
((fun x : EuclideanSpace ℝ (Fin n) => (x : Fin n → ℝ)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → ℝ)) (f i)))) :
(convexFunctionClosure (infimalConvolutionFamily fun i => fenchelConjugate n (f i)) =
infimalConvolutionFamily fun i => fenchelConjugate n (f i)) ∧
∀ xStar : Fin n → ℝ,
infimalConvolutionFamily (fun i => fenchelConjugate n (f i)) xStar = ⊤ ∨
∃ xStarFamily : Fin m → Fin n → ℝ,
(∑ i, xStarFamily i) = xStar ∧
(∑ i, fenchelConjugate n (f i) (xStarFamily i)) =
infimalConvolutionFamily (fun i => fenchelConjugate n (f i)) xStar§16 — Dual Operations — Part 10
Lean authors: Zichen Wang, Wanli Ma, Weiran Shi, Yunxi Duan, Zaiwen Wen.
Remove both closures under the ri hypothesis and record attainment.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_fenchelConjugate_sum_eq_infimalConvolutionFamily_of_nonempty_iInter_ri_effectiveDomain_proof_step
{n m : Nat} (f : Fin m → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i))
(hri :
Set.Nonempty
(⋂ i : Fin m,
euclideanRelativeInterior n
((fun x : EuclideanSpace ℝ (Fin n) => (x : Fin n → ℝ)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → ℝ)) (f i)))) :
(fenchelConjugate n (fun x => ∑ i, f i x) =
infimalConvolutionFamily (fun i => fenchelConjugate n (f i))) ∧
∀ xStar : Fin n → ℝ,
infimalConvolutionFamily (fun i => fenchelConjugate n (f i)) xStar = ⊤ ∨
∃ xStarFamily : Fin m → Fin n → ℝ,
(∑ i, xStarFamily i) = xStar ∧
(∑ i, fenchelConjugate n (f i) (xStarFamily i)) =
infimalConvolutionFamily (fun i => fenchelConjugate n (f i)) xStar
Theorem 16.4.3 (sum vs infimal convolution without closures).
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
theorem section16_fenchelConjugate_sum_eq_infimalConvolutionFamily_of_nonempty_iInter_ri_effectiveDomain
{n m : Nat} (f : Fin m → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i))
(hri :
Set.Nonempty
(⋂ i : Fin m,
euclideanRelativeInterior n
((fun x : EuclideanSpace ℝ (Fin n) => (x : Fin n → ℝ)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → ℝ)) (f i)))) :
(fenchelConjugate n (fun x => ∑ i, f i x) =
infimalConvolutionFamily (fun i => fenchelConjugate n (f i))) ∧
∀ xStar : Fin n → ℝ,
infimalConvolutionFamily (fun i => fenchelConjugate n (f i)) xStar = ⊤ ∨
∃ xStarFamily : Fin m → Fin n → ℝ,
(∑ i, xStarFamily i) = xStar ∧
(∑ i, fenchelConjugate n (f i) (xStarFamily i)) =
infimalConvolutionFamily (fun i => fenchelConjugate n (f i)) xStar
If a + t * b ≤ 0 for all positive t, then a ≤ 0.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_le_zero_of_forall_pos_add_mul_le {a b : ℝ}
(h : ∀ t : ℝ, 0 < t → a + t * b ≤ 0) : a ≤ 0
The intersection of the polars is contained in the polar of the sum.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_iInter_polar_subset_polar_sum {n m : Nat}
(K : Fin m → ConvexCone ℝ (Fin n → ℝ)) :
(⋂ i : Fin m,
{xStar : Fin n → ℝ | ∀ x ∈ (K i : Set (Fin n → ℝ)), (dotProduct x xStar : ℝ) ≤ 0}) ⊆
{xStar : Fin n → ℝ |
∀ x ∈ (∑ i, (K i : Set (Fin n → ℝ))),
(dotProduct x xStar : ℝ) ≤ 0}
The polar of the sum is contained in the intersection of the polars.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_polar_sum_subset_iInter_polar {n m : Nat}
(K : Fin m → ConvexCone ℝ (Fin n → ℝ))
(hKne : ∀ i, (K i : Set (Fin n → ℝ)).Nonempty) :
{xStar : Fin n → ℝ |
∀ x ∈ (∑ i, (K i : Set (Fin n → ℝ))),
(dotProduct x xStar : ℝ) ≤ 0} ⊆
⋂ i : Fin m,
{xStar : Fin n → ℝ | ∀ x ∈ (K i : Set (Fin n → ℝ)), (dotProduct x xStar : ℝ) ≤ 0}
Corollary 16.4.2.1: Let K₁, …, Kₘ be non-empty convex cones in ℝⁿ. Then the polar cone of their Minkowski sum is the intersection of their polar cones: (K₁ + ⋯ + Kₘ)^∘ = K₁^∘ ∩ ⋯ ∩ Kₘ^∘.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
theorem section16_polar_sum_convexCones {n m : Nat} (K : Fin m → ConvexCone ℝ (Fin n → ℝ))
(hKne : ∀ i, (K i : Set (Fin n → ℝ)).Nonempty) :
{xStar : Fin n → ℝ |
∀ x ∈ (∑ i, (K i : Set (Fin n → ℝ))),
(dotProduct x xStar : ℝ) ≤ 0} =
⋂ i : Fin m,
{xStar : Fin n → ℝ | ∀ x ∈ (K i : Set (Fin n → ℝ)), (dotProduct x xStar : ℝ) ≤ 0}
Negating a set flips the inner dual cone.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_innerDual_neg_eq_neg_innerDual {H : Type*} [NormedAddCommGroup H]
[InnerProductSpace ℝ H] [CompleteSpace H] (s : Set H) :
(ProperCone.innerDual (E := H) (-s) : Set H) =
-(ProperCone.innerDual (E := H) s : Set H)
The inner-product polar set is the negation of the inner dual cone.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_inner_polar_eq_neg_innerDual {H : Type*} [NormedAddCommGroup H]
[InnerProductSpace ℝ H] [CompleteSpace H] (s : Set H) :
{y : H | ∀ x ∈ s, ⟪x, y⟫_ℝ ≤ 0} =
-(ProperCone.innerDual (E := H) s : Set H)
The double inner-product polar of a convex cone is its closure.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_inner_polar_polar_eq_closure_convexCone {n : Nat}
(K : ConvexCone ℝ (EuclideanSpace ℝ (Fin n)))
(hKne : (K : Set (EuclideanSpace ℝ (Fin n))).Nonempty) :
{xStar : EuclideanSpace ℝ (Fin n) |
∀ x ∈
{z : EuclideanSpace ℝ (Fin n) |
∀ x ∈ (K : Set (EuclideanSpace ℝ (Fin n))), ⟪x, z⟫_ℝ ≤ 0},
⟪x, xStar⟫_ℝ ≤ 0} =
closure (K : Set (EuclideanSpace ℝ (Fin n)))
The dot-product double polar of a convex cone is its closure.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_polar_polar_eq_closure_convexCone {n : Nat}
(K : ConvexCone ℝ (Fin n → ℝ)) (hKne : (K : Set (Fin n → ℝ)).Nonempty) :
{xStar : Fin n → ℝ |
∀ x ∈ {z : Fin n → ℝ | ∀ x ∈ (K : Set (Fin n → ℝ)), dotProduct x z ≤ 0},
dotProduct x xStar ≤ 0} =
closure (K : Set (Fin n → ℝ))
The polar of the sum of polars is the intersection of the closures.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_polar_sum_polars_eq_iInter_closure {n m : Nat}
(K : Fin m → ConvexCone ℝ (Fin n → ℝ)) (hKne : ∀ i, (K i : Set (Fin n → ℝ)).Nonempty) :
{xStar : Fin n → ℝ |
∀ x ∈
(∑ i : Fin m,
{xStar : Fin n → ℝ |
∀ x ∈ (K i : Set (Fin n → ℝ)), dotProduct x xStar ≤ 0}),
dotProduct x xStar ≤ 0} =
⋂ i : Fin m, closure (K i : Set (Fin n → ℝ))
The double polar of the sum of polars is the closure of that sum.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_polar_polar_sum_polars_eq_closure_sum_polars {n m : Nat}
(K : Fin m → ConvexCone ℝ (Fin n → ℝ)) :
{xStar : Fin n → ℝ |
∀ x ∈
{y : Fin n → ℝ |
∀ z ∈
(∑ i : Fin m,
{xStar : Fin n → ℝ |
∀ x ∈ (K i : Set (Fin n → ℝ)), dotProduct x xStar ≤ 0}),
dotProduct z y ≤ 0},
dotProduct x xStar ≤ 0} =
closure
(∑ i : Fin m,
{xStar : Fin n → ℝ |
∀ x ∈ (K i : Set (Fin n → ℝ)), dotProduct x xStar ≤ 0})
Corollary 16.4.2.2. Let K₁, …, Kₘ be non-empty convex cones in ℝⁿ. Then (cl K₁ ∩ ⋯ ∩ cl Kₘ)^∘ = cl (K₁^∘ + ⋯ + Kₘ^∘). In this development, cl is closure, ∩ is ⋂, and K^∘ is represented as the set {xStar | ∀ x ∈ K, ⟪x, xStar⟫ ≤ 0}.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
theorem section16_polar_iInter_closure_eq_closure_sum_polars {n m : Nat}
(K : Fin m → ConvexCone ℝ (Fin n → ℝ)) (hKne : ∀ i, (K i : Set (Fin n → ℝ)).Nonempty) :
{xStar : Fin n → ℝ |
∀ x ∈ (⋂ i : Fin m, closure (K i : Set (Fin n → ℝ))),
(dotProduct x xStar : ℝ) ≤ 0} =
closure
(∑ i : Fin m,
{xStar : Fin n → ℝ | ∀ x ∈ (K i : Set (Fin n → ℝ)), (dotProduct x xStar : ℝ) ≤ 0})
The support function of a nonempty convex cone is the indicator of its polar.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_supportFunctionEReal_convexCone_eq_indicatorFunction_polar {n : Nat}
(K : ConvexCone ℝ (Fin n → ℝ)) (hKne : (K : Set (Fin n → ℝ)).Nonempty) :
supportFunctionEReal (K : Set (Fin n → ℝ)) =
indicatorFunction
{xStar : Fin n → ℝ | ∀ x ∈ (K : Set (Fin n → ℝ)), dotProduct x xStar ≤ 0}
Indicator functions determine their sets.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_set_eq_of_indicatorFunction_eq {n : Nat} {A B : Set (Fin n → ℝ)}
(h : indicatorFunction A = indicatorFunction B) : A = B
Corollary 16.4.2.3: Let K₁, …, Kₘ be non-empty convex cones in ℝⁿ. If the cones ri Kᵢ, i = 1, …, m, have a point in common, then (K₁ ∩ ⋯ ∩ Kₘ)^∘ = (K₁^∘ + ⋯ + Kₘ^∘). In this development, K^∘ is represented as the set {xStar | ∀ x ∈ K, ⟪x, xStar⟫ ≤ 0} and ri is euclideanRelativeInterior.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
theorem section16_polar_iInter_eq_sum_polars_of_nonempty_iInter_ri {n m : Nat}
(K : Fin m → ConvexCone ℝ (Fin n → ℝ)) (hKne : ∀ i, (K i : Set (Fin n → ℝ)).Nonempty)
(hri :
Set.Nonempty
(⋂ i : Fin m,
euclideanRelativeInterior n
((fun x : EuclideanSpace ℝ (Fin n) => (x : Fin n → ℝ)) ⁻¹' (K i)))) :
{xStar : Fin n → ℝ |
∀ x ∈ (⋂ i : Fin m, (K i : Set (Fin n → ℝ))),
(dotProduct x xStar : ℝ) ≤ 0} =
∑ i : Fin m, {xStar : Fin n → ℝ | ∀ x ∈ (K i : Set (Fin n → ℝ)), (dotProduct x xStar : ℝ) ≤ 0}§16 — Dual Operations — Part 11
Lean authors: Zichen Wang, Wanli Ma, Weiran Shi, Yunxi Duan, Zaiwen Wen.
Theorem 16.4.3: Let f₁, …, fₘ be proper convex functions on ℝⁿ. For any convex set D ⊆ ℝᵐ, if there exists x such that A x ∈ ri D, then the closure operation can be omitted in Theorem 16.4.2, and δ^*(xStar | A⁻¹ D) = inf { δ^*(yStar | D) | A^* yStar = xStar }, where the infimum is attained (or is +∞ vacuously). In this development, δ^* is represented by supportFunctionEReal, ri is euclideanRelativeInterior, and the right-hand side is modeled by an sInf over the affine fiber {yStar | A^* yStar = xStar}.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
theorem section16_supportFunctionEReal_preimage_eq_sInf_of_exists_mem_ri {n m : Nat}
(A : EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m))
(D : Set (Fin m → ℝ)) (hD : Convex ℝ D)
(hri :
∃ x : EuclideanSpace ℝ (Fin n),
A x ∈
euclideanRelativeInterior m
((fun z : EuclideanSpace ℝ (Fin m) => (z : Fin m → ℝ)) ⁻¹' D)) :
let Aadj : EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n) :=
((LinearMap.adjoint :
(EuclideanSpace ℝ (Fin n) →ₗ[ℝ] EuclideanSpace ℝ (Fin m)) ≃ₗ⋆[ℝ]
(EuclideanSpace ℝ (Fin m) →ₗ[ℝ] EuclideanSpace ℝ (Fin n)))
A)
((supportFunctionEReal
(Set.preimage (fun x : Fin n → ℝ => (A (WithLp.toLp 2 x) : Fin m → ℝ)) D) =
fun xStar : Fin n → ℝ =>
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
supportFunctionEReal D (yStar : Fin m → ℝ)) ''
{yStar | Aadj yStar = WithLp.toLp 2 xStar})) ∧
∀ xStar : Fin n → ℝ,
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
supportFunctionEReal D (yStar : Fin m → ℝ)) ''
{yStar | Aadj yStar = WithLp.toLp 2 xStar}) =
⊤ ∨
∃ yStar : EuclideanSpace ℝ (Fin m),
Aadj yStar = WithLp.toLp 2 xStar ∧
supportFunctionEReal D (yStar : Fin m → ℝ) =
sInf
((fun yStar : EuclideanSpace ℝ (Fin m) =>
supportFunctionEReal D (yStar : Fin m → ℝ)) ''
{yStar | Aadj yStar = WithLp.toLp 2 xStar}))
Rewriting the distance function as an infimal convolution of the norm and indicator.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_infDist_as_infimalConvolution_norm_indicator {n : Nat} {C : Set (Fin n → ℝ)}
(hCne : C.Nonempty) :
(fun x : Fin n → ℝ => ((Metric.infDist x C : ℝ) : EReal)) =
infimalConvolution (fun y : Fin n → ℝ => ((‖y‖ : ℝ) : EReal)) (indicatorFunction C)
The Fenchel conjugate of the norm is the indicator of the ℓ¹ unit ball.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_fenchelConjugate_norm_eq_indicator_unitBall {n : ℕ} :
fenchelConjugate n (fun x : Fin n → ℝ => ((‖x‖ : ℝ) : EReal)) =
fun xStar : Fin n → ℝ =>
if l1Norm xStar ≤ (1 : ℝ) then (0 : EReal) else (⊤ : EReal)
The Fenchel conjugate of the distance function splits into the norm conjugate and the support function.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_fenchelConjugate_infDist_eq_indicatorBall_add_supportFunctionEReal {n : ℕ}
(C : Set (Fin n → ℝ)) (hC : Convex ℝ C) (hCne : C.Nonempty) :
fenchelConjugate n (fun x : Fin n → ℝ => ((Metric.infDist x C : ℝ) : EReal)) =
fun xStar : Fin n → ℝ =>
(if l1Norm xStar ≤ (1 : ℝ) then (0 : EReal) else (⊤ : EReal)) +
supportFunctionEReal C xStar
Text 16.4.1: Let f(x) = d(x, C) = inf {‖x - y‖ | y ∈ C} be the distance function, where C is a given nonempty convex set. Then f^*(x^*) = δ^*(x^* | C) if ‖x^*‖₁ ≤ 1, and f^*(x^*) = +∞ otherwise. In this development, f^* is represented by fenchelConjugate n f, d(x, C) is Metric.infDist, and δ^*(· | C) is the support function supportFunctionEReal C.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
theorem section16_fenchelConjugate_infDist {n : ℕ} (C : Set (Fin n → ℝ)) (hC : Convex ℝ C)
(hCne : C.Nonempty) :
fenchelConjugate n (fun x : Fin n → ℝ => ((Metric.infDist x C : ℝ) : EReal)) =
fun xStar : Fin n → ℝ =>
if l1Norm xStar ≤ (1 : ℝ) then supportFunctionEReal C xStar else (⊤ : EReal)
The span of a finite family agrees with the range of its linear-combination map.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_span_range_eq_linearMap_range_sum_smul {n m : Nat} (a : Fin m → Fin n → ℝ) :
let A : (Fin m → ℝ) →ₗ[ℝ] (Fin n → ℝ) :=
{ toFun := fun ξ => ∑ t, (ξ t) • a t
map_add' := by
intro ξ η
ext j
simp [Finset.sum_add_distrib, add_smul]
map_smul' := by
intro c ξ
ext j
simpa [smul_eq_mul, mul_comm, mul_left_comm, mul_assoc] using
(Finset.mul_sum (s := (Finset.univ : Finset (Fin m)))
(f := fun x => a x j * ξ x) (a := c)).symm }
Submodule.span ℝ (Set.range a) = LinearMap.range A
The infimum over all coefficient vectors is the distance to the range.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_sInf_range_norm_sub_eq_infDist_range {n m : Nat}
(A : (Fin m → ℝ) →ₗ[ℝ] (Fin n → ℝ)) (x : Fin n → ℝ) :
sInf (Set.range fun ξ : Fin m → ℝ => ‖x - A ξ‖) =
Metric.infDist x (Set.range A)
The distance to a submodule is closed, proper, and convex.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_closedProperConvex_coe_infDist_submodule {n : Nat}
(L : Submodule ℝ (Fin n → ℝ)) :
ClosedConvexFunction
(fun x : Fin n → ℝ => ((Metric.infDist x (L : Set (Fin n → ℝ)) : ℝ) : EReal)) ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ))
(fun x : Fin n → ℝ => ((Metric.infDist x (L : Set (Fin n → ℝ)) : ℝ) : EReal))
The support function of a submodule is the indicator of its orthogonal complement.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_supportFunctionEReal_submodule {n : Nat} (L : Submodule ℝ (Fin n → ℝ)) :
supportFunctionEReal (L : Set (Fin n → ℝ)) =
fun b : Fin n → ℝ =>
if b ∈ orthogonalComplement n L then (0 : EReal) else (⊤ : EReal)
The conjugate of the distance to a submodule is an indicator on D ∩ Lᗮ.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_fenchelConjugate_coe_infDist_submodule_eq_indicator_D_inter_orth {n : Nat}
(L : Submodule ℝ (Fin n → ℝ)) :
fenchelConjugate n
(fun x : Fin n → ℝ =>
((Metric.infDist x (L : Set (Fin n → ℝ)) : ℝ) : EReal)) =
indicatorFunction
({b : Fin n → ℝ | l1Norm b ≤ (1 : ℝ)} ∩
(orthogonalComplement n L : Set (Fin n → ℝ)))
Dot products over D ∩ Lᗮ are bounded above.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_bddAbove_dotProduct_D_inter_orth {n : Nat} (L : Submodule ℝ (Fin n → ℝ))
(x : Fin n → ℝ) :
BddAbove
(Set.image (fun y : Fin n → ℝ => dotProduct y x)
({y : Fin n → ℝ | l1Norm y ≤ (1 : ℝ)} ∩
(orthogonalComplement n L : Set (Fin n → ℝ))))
Text 16.4.2: Consider the function f(x) = inf { ‖x - ξ₁ a₁ - ⋯ - ξₘ aₘ‖∞ | ξₜ ∈ ℝ }, where a₁, …, aₘ ∈ ℝⁿ are given and ‖x‖∞ = max_j |x j| for x ∈ ℝⁿ. Then f is the support function of the (polyhedral) convex set D ∩ L^⊥, where L is the subspace spanned by a₁, …, aₘ and D = {xStar | |xStar 1| + ⋯ + |xStar n| ≤ 1}.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
theorem section16_infNorm_sub_lincomb_eq_deltaStar_inter_orthogonalComplement {n m : Nat}
(a : Fin m → Fin n → ℝ) :
let L : Submodule ℝ (Fin n → ℝ) := Submodule.span ℝ (Set.range a)
let D : Set (Fin n → ℝ) := {xStar | (∑ j, |xStar j|) ≤ (1 : ℝ)}
let f : (Fin n → ℝ) → ℝ :=
fun x => sInf (Set.range fun ξ : Fin m → ℝ => ‖x - ∑ t, (ξ t) • a t‖)
f = deltaStar (D ∩ (orthogonalComplement n L : Set (Fin n → ℝ)))
Rewriting the piecewise nonnegative extension as a sum with the indicator.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_piecewise_nonneg_eq_add_indicator_nonnegOrthant {n : Nat}
(h : (Fin n → ℝ) → EReal) (hnotbot : ∀ x, h x ≠ (⊥ : EReal)) :
(fun x : Fin n → ℝ => if 0 ≤ x then h x else (⊤ : EReal)) =
fun x => h x + indicatorFunction {x : Fin n → ℝ | 0 ≤ x} x
The dot-product polar of the nonnegative orthant is the nonpositive orthant.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_polar_nonnegOrthant_eq_nonpos {n : Nat} :
{xStar : Fin n → ℝ |
∀ x ∈ {x : Fin n → ℝ | 0 ≤ x}, dotProduct x xStar ≤ 0} =
{xStar : Fin n → ℝ | xStar ≤ 0}
The Fenchel conjugate of the nonnegative-orthant indicator is the nonpositive-orthant indicator.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_fenchelConjugate_indicator_nonnegOrthant_eq_indicator_nonposOrthant {n : Nat} :
fenchelConjugate n (indicatorFunction {x : Fin n → ℝ | 0 ≤ x}) =
indicatorFunction {xStar : Fin n → ℝ | xStar ≤ 0}
Infimal convolution with the nonpositive orthant indicator is an infimum over upper bounds.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_infimalConvolution_with_indicator_nonpos_eq_sInf_image_ge {n : Nat}
(p : (Fin n → ℝ) → EReal) (hp : ∀ z, p z ≠ (⊥ : EReal)) :
infimalConvolution p (indicatorFunction {u : Fin n → ℝ | u ≤ 0}) =
fun xStar => sInf (p '' {zStar : Fin n → ℝ | xStar ≤ zStar})
Text 16.4.3: Let h be a closed proper convex function on ℝⁿ. Define f by f x = h x if x ≥ 0, and f x = +∞ otherwise, where x ≥ 0 is the coordinatewise order (the non-negative orthant). Then the Fenchel conjugate f^* is the closure of the convex function g xStar = inf { h^* zStar | zStar ≥ xStar }, again with respect to the coordinatewise order. Here h^* is fenchelConjugate n h and the closure is convexFunctionClosure.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
theorem section16_fenchelConjugate_piecewise_nonneg_eq_convexFunctionClosure_sInf_ge {n : ℕ}
(h : (Fin n → ℝ) → EReal) (hclosed : ClosedConvexFunction h)
(hproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) h) :
let f : (Fin n → ℝ) → EReal := by
classical
exact fun x => if 0 ≤ x then h x else (⊤ : EReal)
let g : (Fin n → ℝ) → EReal :=
fun xStar => sInf ((fun zStar => fenchelConjugate n h zStar) '' {zStar | xStar ≤ zStar})
fenchelConjugate n f = convexFunctionClosure g
Softmax lies in the simplex and has a positive normalizer.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_softmax_mem_simplex {n : Nat} (hn : 0 < n) (xStar : Fin n → ℝ) :
let Z := ∑ j : Fin n, Real.exp (xStar j)
let ξ0 : Fin n → ℝ := fun j => Real.exp (xStar j) / Z
(0 ≤ ξ0) ∧ (∑ j, ξ0 j = 1) ∧ 0 < Z
A weighted exponential term is bounded by the unweighted exponential.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_mul_exp_sub_log_le_exp {ξ a : ℝ} (hξ : 0 ≤ ξ) :
ξ * Real.exp (a - Real.log ξ) ≤ Real.exp a
Gibbs inequality for entropy on the simplex.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_entropy_rangeTerm_le_log_sum_exp {n : Nat} (hn : 0 < n) (ξ : Fin n → ℝ)
(hξ : 0 ≤ ξ) (hsum : ∑ j, ξ j = (1 : ℝ)) (xStar : Fin n → ℝ) :
dotProduct ξ xStar - ∑ j, ξ j * Real.log (ξ j) ≤
Real.log (∑ j, Real.exp (xStar j))
The entropy range term attains log (∑ exp) at the softmax point.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_entropy_rangeTerm_eq_log_sum_exp_at_softmax {n : Nat} (hn : 0 < n)
(xStar : Fin n → ℝ) :
let Z := ∑ j : Fin n, Real.exp (xStar j)
let ξ0 : Fin n → ℝ := fun j => Real.exp (xStar j) / Z
dotProduct ξ0 xStar - ∑ j, ξ0 j * Real.log (ξ0 j) = Real.log Z§16 — Dual Operations — Part 12
Lean authors: Zichen Wang, Wanli Ma, Weiran Shi, Yunxi Duan, Zaiwen Wen.
Text 16.4.4: Let f : ℝⁿ → ℝ ∪ {+∞} be defined by f(ξ₁, …, ξₙ) = ξ₁ log ξ₁ + ⋯ + ξₙ log ξₙ if ξⱼ ≥ 0 for all j and ∑ᵢ ξᵢ = 1, and f = +∞ otherwise (with the convention 0 log 0 = 0). Then the Fenchel conjugate satisfies f^*(x^*) = log (exp(λ₁^*) + ⋯ + exp(λₙ^*)).
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
theorem section16_fenchelConjugate_sum_mul_log_eq_log_sum_exp {n : ℕ} (hn : 0 < n) :
let f : (Fin n → ℝ) → EReal := by
classical
exact fun ξ =>
if 0 ≤ ξ ∧ (∑ j : Fin n, ξ j) = (1 : ℝ) then
((∑ j : Fin n, ξ j * Real.log (ξ j) : ℝ) : EReal)
else (⊤ : EReal)
fenchelConjugate n f =
fun xStar : Fin n → ℝ => (Real.log (∑ j : Fin n, Real.exp (xStar j)) : EReal)
The affine functional x ↦ ⟪x,b⟫ - μ is convex as an EReal-valued function.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_convexFunctionOn_dotProduct_sub_const {n : ℕ} (b : Fin n → ℝ) (μ : ℝ) :
ConvexFunctionOn (Set.univ : Set (Fin n → ℝ))
(fun x => ((x ⬝ᵥ b - μ : ℝ) : EReal))
Affine minorants of the convex hull coincide with affine minorants of each family member.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_affine_le_convexHullFunctionFamily_iff_forall_le {n : ℕ} {ι : Sort _}
(f : ι → (Fin n → ℝ) → EReal) (b : Fin n → ℝ) (μ : ℝ) :
(∀ x, ((x ⬝ᵥ b - μ : ℝ) : EReal) ≤ convexHullFunctionFamily f x) ↔
(∀ i x, ((x ⬝ᵥ b - μ : ℝ) : EReal) ≤ f i x)
The epigraph inequality for the convex hull reduces to the family componentwise.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_fenchelConjugate_convexHullFunctionFamily_le_coe_iff_forall {n : ℕ} {ι : Sort _}
(f : ι → (Fin n → ℝ) → EReal) (xStar : Fin n → ℝ) (μ : ℝ) :
fenchelConjugate n (convexHullFunctionFamily f) xStar ≤ (μ : EReal) ↔
∀ i, fenchelConjugate n (f i) xStar ≤ (μ : EReal)
Nonempty families yield a non-⊥ supremum of conjugates.
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
lemma section16_sSup_range_fenchelConjugate_ne_bot_of_nonempty {n : ℕ} {ι : Sort _} [Nonempty ι]
(f : ι → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i))
(xStar : Fin n → ℝ) :
sSup (Set.range fun i => fenchelConjugate n (f i) xStar) ≠ (⊥ : EReal)
Theorem 16.5.1: Let f i be a proper convex function on ℝ^n for each i in an (arbitrary) index set. Then the Fenchel conjugate of the convex hull of the family is the pointwise supremum of the conjugates: (conv {f_i | i ∈ I})^* = sup {f_i^* | i ∈ I}. Here conv {f_i | i ∈ I} is represented by convexHullFunctionFamily f, and f_i^* is fenchelConjugate n (f i).
open scoped BigOperators
open scoped Pointwise
open scoped InnerProductSpace
theorem section16_fenchelConjugate_convexHullFunctionFamily {n : ℕ} {ι : Sort _}
(f : ι → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i)) :
fenchelConjugate n (convexHullFunctionFamily f) =
fun xStar => sSup (Set.range fun i => fenchelConjugate n (f i) xStar)§16 — Dual Operations — Part 13
Lean authors: Zichen Wang, Wanli Ma, Weiran Shi, Yunxi Duan, Zaiwen Wen.
The indicator function of a nonempty convex set is proper convex on Set.univ.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_properConvexFunctionOn_indicatorFunction_univ {n : ℕ} {C : Set (Fin n → ℝ)}
(hC : Convex ℝ C) (hCne : C.Nonempty) :
ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (indicatorFunction C)
The convex hull of indicator functions is the indicator of the convex hull of the union.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_convexHullFunctionFamily_indicatorFunction_eq_indicatorFunction_convexHull_iUnion
{n : ℕ} {ι : Sort _} (C : ι → Set (Fin n → ℝ)) :
convexHullFunctionFamily (fun i => indicatorFunction (C i)) =
indicatorFunction (convexHull ℝ (⋃ i, C i))
Support function of the convex hull of a union as a supremum of support functions.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_supportFunctionEReal_convexHull_iUnion {n : ℕ} {ι : Sort _}
(C : ι → Set (Fin n → ℝ)) (hC : ∀ i, Convex ℝ (C i)) (hCne : ∀ i, (C i).Nonempty) :
supportFunctionEReal (convexHull ℝ (⋃ i, C i)) =
fun xStar => sSup (Set.range fun i => supportFunctionEReal (C i) xStar)
Corollary 16.5.1.1: Let C i be a non-empty convex set in ℝⁿ for each i in an index set. Let D be the convex hull of the union ⋃ i, C i. Then the support function satisfies δ^*(· | D) = sup { δ^*(· | C i) | i }, where δ^*(· | C) is represented by supportFunctionEReal C.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
theorem section16_deltaStar_convexHull_iUnion {n : ℕ} {ι : Sort _} (C : ι → Set (Fin n → ℝ))
(hC : ∀ i, Convex ℝ (C i)) (hCne : ∀ i, (C i).Nonempty) :
supportFunctionEReal (convexHull ℝ (⋃ i, C i)) =
fun xStar => sSup (Set.range fun i => supportFunctionEReal (C i) xStar)
Applying Theorem 16.5.1 to the family of conjugates yields the supremum of closures.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_fenchelConjugate_convexHullFunctionFamily_fenchelConjugate_eq_sSup_convexFunctionClosure
{n : ℕ} {ι : Sort _} (f : ι → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i)) :
fenchelConjugate n (convexHullFunctionFamily fun i => fenchelConjugate n (f i)) =
fun x => sSup (Set.range fun i => convexFunctionClosure (f i) x)
The convex hull of the conjugate family is a convex function.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_convexFunction_convexHullFunctionFamily_fenchelConjugate {n : ℕ} {ι : Sort _}
(f : ι → (Fin n → ℝ) → EReal) :
ConvexFunction (convexHullFunctionFamily fun i => fenchelConjugate n (f i))
The conjugate of the supremum of closures is the closure of the convex hull of conjugates.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_fenchelConjugate_sSup_convexFunctionClosure_eq_via_biconjugate
{n : ℕ} {ι : Sort _} (f : ι → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i)) :
fenchelConjugate n (fun x => sSup (Set.range fun i => convexFunctionClosure (f i) x)) =
convexFunctionClosure (convexHullFunctionFamily fun i => fenchelConjugate n (f i))
Theorem 16.5.2: Let f i be a proper convex function on ℝ^n for each i in an (arbitrary) index set. Then the Fenchel conjugate of the pointwise supremum of the closures cl f_i is the closure of the convex hull of the conjugates: (sup {cl f_i | i ∈ I})^* = cl (conv {f_i^* | i ∈ I}). Here cl is the convex-function closure convexFunctionClosure, sup is modeled pointwise by x ↦ sSup (range fun i => ·), conv is convexHullFunctionFamily, and f_i^* is fenchelConjugate n (f i).
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
theorem section16_fenchelConjugate_sSup_convexFunctionClosure_eq_convexFunctionClosure_convexHullFunctionFamily
{n : ℕ} {ι : Sort _} (f : ι → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i)) :
fenchelConjugate n (fun x => sSup (Set.range fun i => convexFunctionClosure (f i) x)) =
convexFunctionClosure (convexHullFunctionFamily fun i => fenchelConjugate n (f i))
Pointwise supremum of indicator functions of closures is the indicator of the intersection.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_sSup_range_indicatorFunction_closure_eq_indicatorFunction_iInter_closure
{n : ℕ} {ι : Sort _} [Nonempty ι] (C : ι → Set (Fin n → ℝ)) :
(fun x => sSup (Set.range fun i => indicatorFunction (closure (C i)) x)) =
indicatorFunction (⋂ i, closure (C i))
Support function of the intersection of closures as a convex hull closure.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_supportFunctionEReal_iInter_closure_eq_convexFunctionClosure_convexHullFunctionFamily
{n : ℕ} {ι : Sort _} [Nonempty ι] (C : ι → Set (Fin n → ℝ))
(hC : ∀ i, Convex ℝ (C i)) (hCne : ∀ i, (C i).Nonempty) :
supportFunctionEReal (⋂ i, closure (C i)) =
convexFunctionClosure (convexHullFunctionFamily fun i => supportFunctionEReal (C i))
Corollary 16.5.1.2: Let C i be a non-empty convex set in ℝ^n for each i in an index set. Let C be the intersection ⋂ i, closure (C i). Then the support function satisfies δ^*(· | C) = cl (conv { δ^*(· | C i) | i }). Moreover, δ^*(· | D) = sup { δ^*(· | C i) | i }. In this development, δ^* is supportFunctionEReal, cl is convexFunctionClosure, and conv is convexHullFunctionFamily.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
theorem section16_supportFunctionEReal_iInter_closure_eq_convexFunctionClosure_convexHullFunctionFamily_and_convexHull_iUnion
{n : ℕ} {ι : Sort _} [Nonempty ι] (C : ι → Set (Fin n → ℝ))
(hC : ∀ i, Convex ℝ (C i)) (hCne : ∀ i, (C i).Nonempty) :
let C' : Set (Fin n → ℝ) := ⋂ i, closure (C i)
let D : Set (Fin n → ℝ) := convexHull ℝ (⋃ i, C i)
(supportFunctionEReal C' =
convexFunctionClosure (convexHullFunctionFamily fun i => supportFunctionEReal (C i))) ∧
(supportFunctionEReal D = fun xStar => sSup (Set.range fun i => supportFunctionEReal (C i) xStar))
The sublevel set {x | dotProduct x xStar ≤ 1} is convex.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_convex_set_dotProduct_le_one {n : ℕ} (xStar : Fin n → ℝ) :
Convex ℝ {x : Fin n → ℝ | (dotProduct x xStar : ℝ) ≤ 1}
Polar inclusion from the convex hull of a union to the intersection of polars.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_polar_convexHull_iUnion_subset_iInter_polars {n : ℕ} {ι : Sort _}
(C : ι → Set (Fin n → ℝ)) (xStar : Fin n → ℝ)
(h : ∀ x ∈ convexHull ℝ (⋃ i, C i), (dotProduct x xStar : ℝ) ≤ 1) :
∀ i, ∀ x ∈ C i, (dotProduct x xStar : ℝ) ≤ 1
Polar inclusion from the intersection of polars to the polar of the convex hull of a union.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_iInter_polars_subset_polar_convexHull_iUnion {n : ℕ} {ι : Sort _}
(C : ι → Set (Fin n → ℝ)) (xStar : Fin n → ℝ)
(h : ∀ i, ∀ x ∈ C i, (dotProduct x xStar : ℝ) ≤ 1) :
∀ x ∈ convexHull ℝ (⋃ i, C i), (dotProduct x xStar : ℝ) ≤ 1
Corollary 16.5.2.1: Let C i be a convex set in ℝ^n for each i in an index set. Then the polar of the convex hull of the family equals the intersection of the polars: (conv { C_i | i ∈ I })^∘ = ⋂ { C_i^∘ | i ∈ I }. In this development, conv { C_i | i ∈ I } is represented by convexHull ℝ (⋃ i, C i), and the polar of a set S is represented by {xStar | ∀ x ∈ S, (dotProduct x xStar : ℝ) ≤ 1}.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
theorem section16_polar_convexHull_iUnion_eq_iInter_polars {n : ℕ} {ι : Sort _}
(C : ι → Set (Fin n → ℝ)) :
{xStar : Fin n → ℝ | ∀ x ∈ convexHull ℝ (⋃ i, C i), (dotProduct x xStar : ℝ) ≤ 1} =
⋂ i, {xStar : Fin n → ℝ | ∀ x ∈ C i, (dotProduct x xStar : ℝ) ≤ 1}
A common point in the intersection of closures yields a proper convex hull of support functions with a finite ≤ 1 sublevel.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_properConvexFunctionOn_convexHullFunctionFamily_supportFamily_of_exists_common_point
{n : ℕ} {ι : Sort _} [Nonempty ι] (C : ι → Set (Fin n → ℝ))
(hC : ∀ i, Convex ℝ (C i)) (hInter : (⋂ i, closure (C i)).Nonempty) :
let g : ι → (Fin n → ℝ) → EReal := fun i => supportFunctionEReal (C i)
let h := convexHullFunctionFamily g
ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) h ∧
iInf (fun xStar => h xStar) < (1 : EReal)
Raising the second coordinate preserves membership in the convex hull of a union of epigraphs.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_convexHull_union_epigraph_mono_snd {n : ℕ} {ι : Sort _}
{g : ι → (Fin n → ℝ) → EReal} {x : Fin n → ℝ} {μ μ' : ℝ}
(hmem :
(x, μ) ∈
convexHull ℝ (⋃ i, epigraph (S := (Set.univ : Set (Fin n → ℝ))) (g i)))
(hle : μ ≤ μ') :
(x, μ') ∈
convexHull ℝ (⋃ i, epigraph (S := (Set.univ : Set (Fin n → ℝ))) (g i))
A small epigraph height above 1 appears in the convex hull of the union of epigraphs.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_exists_mem_convexHull_union_epigraph_lt_one_add_eps_of_hle_one {n : ℕ} {ι : Sort _}
(g : ι → (Fin n → ℝ) → EReal) {xStar : Fin n → ℝ} {ε : ℝ} (hε : 0 < ε)
(hle : convexHullFunctionFamily g xStar ≤ (1 : EReal)) :
∃ μ : ℝ, μ < 1 + ε ∧
(xStar, μ) ∈
convexHull ℝ (⋃ i, epigraph (S := (Set.univ : Set (Fin n → ℝ))) (g i))
Positive homogeneity scales epigraph points.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_epigraph_univ_posHom_smul_mem {n : ℕ} {g : (Fin n → ℝ) → EReal}
(hpos : PositivelyHomogeneous g) {x : Fin n → ℝ} {μ t : ℝ} (ht : 0 < t)
(hmem : (x, μ) ∈ epigraph (S := (Set.univ : Set (Fin n → ℝ))) g) :
(t • x, t * μ) ∈ epigraph (S := (Set.univ : Set (Fin n → ℝ))) g
Scaling preserves membership in the convex hull of a union of epigraphs.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_convexHull_union_epigraph_smul_mem {n : ℕ} {ι : Sort _}
{g : ι → (Fin n → ℝ) → EReal} (hpos : ∀ i, PositivelyHomogeneous (g i))
{x : Fin n → ℝ} {μ t : ℝ} (ht : 0 < t)
(hmem :
(x, μ) ∈
convexHull ℝ (⋃ i, epigraph (S := (Set.univ : Set (Fin n → ℝ))) (g i))) :
(t • x, t * μ) ∈
convexHull ℝ (⋃ i, epigraph (S := (Set.univ : Set (Fin n → ℝ))) (g i))
Closedness and convexity of the closure of the convex hull of level-1 sublevels.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_closedConvex_closure_convexHull_iUnion_sublevels
{n : ℕ} {ι : Sort _} {g : ι → (Fin n → ℝ) → EReal} :
IsClosed
(closure
(convexHull ℝ (⋃ i, {xStar : Fin n → ℝ | g i xStar ≤ (1 : EReal)}))) ∧
Convex ℝ
(closure
(convexHull ℝ (⋃ i, {xStar : Fin n → ℝ | g i xStar ≤ (1 : EReal)})))
A separating continuous linear functional for a point outside the closed convex hull.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_exists_sep_clm_of_not_mem_closure_convexHull_iUnion_sublevels
{n : ℕ} {ι : Sort _} {g : ι → (Fin n → ℝ) → EReal} {xStar : Fin n → ℝ}
(hx :
xStar ∉
closure
(convexHull ℝ (⋃ i, {xStar : Fin n → ℝ | g i xStar ≤ (1 : EReal)}))) :
∃ φ : (Fin n → ℝ) →L[ℝ] ℝ, ∃ u : ℝ,
(∀ y ∈
closure
(convexHull ℝ (⋃ i, {xStar : Fin n → ℝ | g i xStar ≤ (1 : EReal)})),
φ y ≤ u) ∧
u < φ xStar
A separator bounded on the level-1 sublevel controls positive-height epigraph points.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_clm_le_mul_snd_of_mem_epigraph_of_pos {n : ℕ} {ι : Sort _}
{g : ι → (Fin n → ℝ) → EReal} {i : ι} (hpos : PositivelyHomogeneous (g i))
{φ : (Fin n → ℝ) →L[ℝ] ℝ} {u : ℝ}
(hbound : ∀ z, g i z ≤ (1 : EReal) → φ z ≤ u)
{x : Fin n → ℝ} {μ : ℝ}
(hmem : (x, μ) ∈ epigraph (S := (Set.univ : Set (Fin n → ℝ))) (g i)) (hμ : 0 < μ) :
φ x ≤ u * μ
The μ = 1 slice of a convex hull of epigraphs lies in the closure of the hull of slices.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_slice_one_convexHull_union_epigraph_subset_closure_convexHull_sublevels
{n : ℕ} {ι : Sort _} {g : ι → (Fin n → ℝ) → EReal}
(hpos : ∀ i, PositivelyHomogeneous (g i))
(hnonneg : ∀ i x, (0 : EReal) ≤ g i x) :
{xStar : Fin n → ℝ |
(xStar, (1 : ℝ)) ∈
convexHull ℝ (⋃ i, epigraph (S := (Set.univ : Set (Fin n → ℝ))) (g i))} ⊆
closure
(convexHull ℝ (⋃ i, {xStar : Fin n → ℝ | g i xStar ≤ (1 : EReal)}))
Normalize a convex-hull epigraph point at height 1 + ε to the level-1 sublevel hull, landing in the closure.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_mem_convexHull_iUnion_sublevels_one_of_mem_convexHull_union_epigraph_one_add_eps
{n : ℕ} {ι : Sort _} {g : ι → (Fin n → ℝ) → EReal}
(hpos : ∀ i, PositivelyHomogeneous (g i))
(hnonneg : ∀ i x, (0 : EReal) ≤ g i x) {xStar : Fin n → ℝ} {ε : ℝ} (hε : 0 < ε)
(hmem :
(xStar, 1 + ε) ∈
convexHull ℝ (⋃ i, epigraph (S := (Set.univ : Set (Fin n → ℝ))) (g i))) :
((1 + ε)⁻¹) • xStar ∈
closure
(convexHull ℝ (⋃ i, {xStar : Fin n → ℝ | g i xStar ≤ (1 : EReal)}))§16 — Dual Operations — Part 14
Lean authors: Zichen Wang, Wanli Ma, Weiran Shi, Yunxi Duan, Zaiwen Wen.
Closure of a ≤ 1 sublevel set matches the ≤ 1 sublevel of the convex-function closure when the domain is Fin n → ℝ.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_closure_sublevel_eq_sublevel_convexFunctionClosure_one_fin {n : Nat}
{h : (Fin n → ℝ) → EReal}
(hh : ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) h)
(hInf : iInf (fun x => h x) < (1 : EReal)) :
closure {xStar : Fin n → ℝ | h xStar ≤ (1 : EReal)} =
{xStar : Fin n → ℝ | convexFunctionClosure h xStar ≤ (1 : EReal)}
The support function of a nonempty convex set agrees with that of its closure.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_supportFunctionEReal_eq_supportFunctionEReal_closure {n : ℕ}
{C : Set (Fin n → ℝ)} (hC : Convex ℝ C) (hCne : C.Nonempty) :
supportFunctionEReal C = supportFunctionEReal (closure C)
The support function is nonnegative when 0 lies in the closure.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_supportFunctionEReal_nonneg_of_zero_mem_closure {n : ℕ}
{C : Set (Fin n → ℝ)} (hC : Convex ℝ C) (hCne : C.Nonempty)
(h0 : (0 : Fin n → ℝ) ∈ closure C) (xStar : Fin n → ℝ) :
(0 : EReal) ≤ supportFunctionEReal C xStar
Points in the ≤ 1 sublevel of the hull lie in the closure of the convex hull of the ≤ 1 sublevels.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_mem_closure_convexHull_iUnion_sublevels_of_hle_one {n : ℕ} {ι : Sort _}
[Nonempty ι] (C : ι → Set (Fin n → ℝ)) (hC : ∀ i, Convex ℝ (C i))
(hCne : ∀ i, (C i).Nonempty) (h0 : ∀ i, (0 : Fin n → ℝ) ∈ closure (C i))
{xStar : Fin n → ℝ}
(hle : convexHullFunctionFamily (fun i => supportFunctionEReal (C i)) xStar ≤ (1 : EReal)) :
xStar ∈
closure
(convexHull ℝ
(⋃ i, {xStar : Fin n → ℝ | supportFunctionEReal (C i) xStar ≤ (1 : EReal)}))
The ≤ 1 sublevel of the convex-hull support family has the same closure as the convex hull of the ≤ 1 sublevels of the family.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_closure_sublevel_convexHullFunctionFamily_supportFamily_one_eq_closure_convexHull_iUnion_sublevels
{n : ℕ} {ι : Sort _} [Nonempty ι] (C : ι → Set (Fin n → ℝ))
(hC : ∀ i, Convex ℝ (C i)) (hCne : ∀ i, (C i).Nonempty)
(h0 : ∀ i, (0 : Fin n → ℝ) ∈ closure (C i)) :
let g : ι → (Fin n → ℝ) → EReal := fun i => supportFunctionEReal (C i)
let h := convexHullFunctionFamily g
closure {xStar : Fin n → ℝ | h xStar ≤ (1 : EReal)} =
closure (convexHull ℝ (⋃ i, {xStar : Fin n → ℝ | g i xStar ≤ (1 : EReal)}))
Corollary 16.5.2.2: Let C i be a convex set in ℝ^n for each i in a nonempty index set, assume the intersection of the closures is nonempty, and that this intersection contains 0. Then (⋂ i, closure (C i))^∘ = closure (convexHull ℝ (⋃ i, (C i)^∘)). In this development, the polar of a set S is represented by {xStar | ∀ x ∈ S, (dotProduct x xStar : ℝ) ≤ 1}.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
theorem section16_polar_iInter_closure_eq_closure_convexHull_iUnion_polars {n : ℕ} {ι : Sort _}
[Nonempty ι] (C : ι → Set (Fin n → ℝ)) (hC : ∀ i, Convex ℝ (C i))
(h0 : (0 : Fin n → ℝ) ∈ ⋂ i, closure (C i)) (hInter : (⋂ i, closure (C i)).Nonempty) :
{xStar : Fin n → ℝ | ∀ x ∈ (⋂ i, closure (C i)), (dotProduct x xStar : ℝ) ≤ 1} =
closure
(convexHull ℝ (⋃ i, {xStar : Fin n → ℝ | ∀ x ∈ C i, (dotProduct x xStar : ℝ) ≤ 1}))
A common relative-interior point yields finite iSup for a finite family.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_exists_common_ri_point_finite_iSup_ne_top_of_common_closure_effectiveDomain
{n : ℕ} {ι : Type _} [Fintype ι] [Nonempty ι]
(f : ι → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i)) (C : Set (Fin n → ℝ))
(hC : ∀ i, closure (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (f i)) = C) :
∃ x0 : EuclideanSpace ℝ (Fin n),
(∀ i,
x0 ∈
euclideanRelativeInterior n
(Set.image (EuclideanSpace.equiv (Fin n) ℝ).symm
(effectiveDomain (Set.univ : Set (Fin n → ℝ)) (f i)))) ∧
(iSup (fun i => f i ((EuclideanSpace.equiv (Fin n) ℝ) x0))) ≠ (⊤ : EReal)
Closure commutes with a finite iSup under a common domain closure.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_convexFunctionClosure_iSup_eq_iSup_convexFunctionClosure_of_fintype_of_common_closure_effectiveDomain
{n : ℕ} {ι : Type _} [Fintype ι] [Nonempty ι]
(f : ι → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i)) (C : Set (Fin n → ℝ))
(hC : ∀ i, closure (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (f i)) = C) :
convexFunctionClosure (fun x => iSup (fun i => f i x)) =
fun x => iSup (fun i => convexFunctionClosure (f i) x)
Conjugates share a common recession function under a common domain closure.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_common_recession_fenchelConjugate_of_common_closure_effectiveDomain
{n : ℕ} {ι : Type _} [Fintype ι] [Nonempty ι]
(f : ι → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i)) (C : Set (Fin n → ℝ))
(hC : ∀ i, closure (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (f i)) = C) :
∀ i, recessionFunction (fenchelConjugate n (f i)) = supportFunctionEReal C
Closedness and attainment for the convex hull of conjugates under common domain closure.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_convexHullFunctionFamily_fenchelConjugate_closed_and_attained_of_common_closure_effectiveDomain
{n : ℕ} {ι : Type _} [Fintype ι] [Nonempty ι]
(f : ι → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i)) (C : Set (Fin n → ℝ))
(hC : ∀ i, closure (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (f i)) = C) :
let g : ι → (Fin n → ℝ) → EReal := fun i => fenchelConjugate n (f i)
(convexFunctionClosure (convexHullFunctionFamily g) = convexHullFunctionFamily g) ∧
(∀ xStar : Fin n → ℝ,
∃ (lam : ι → ℝ) (xStar_i : ι → Fin n → ℝ),
(∀ i, 0 ≤ lam i) ∧ (∑ i, lam i) = 1 ∧
(∑ i, lam i • xStar_i i) = xStar ∧
convexHullFunctionFamily g xStar =
∑ i, ((lam i : ℝ) : EReal) * g i (xStar_i i))
Reindexing a function family by an equivalence does not change its convex hull.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_convexHullFunctionFamily_congr_equiv {n : ℕ} {ι κ : Sort _}
(e : ι ≃ κ) (g : ι → (Fin n → ℝ) → EReal) :
convexHullFunctionFamily (fun k => g (e.symm k)) = convexHullFunctionFamily g
Convex-combination formula with full-index sums on a finite index type.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
lemma section16_convexHullFunctionFamily_eq_sInf_convexCombination_univ_fintype
{n : ℕ} {ι : Type _} [Fintype ι] [Nonempty ι]
(f : ι → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i)) :
∀ x : Fin n → ℝ,
convexHullFunctionFamily f x =
sInf { r : EReal |
∃ (lam : ι → ℝ) (x' : ι → Fin n → ℝ),
(∀ i, 0 ≤ lam i) ∧
(∑ i, lam i) = 1 ∧
(∑ i, lam i • x' i) = x ∧
r = ∑ i, ((lam i : ℝ) : EReal) * f i (x' i) }
Theorem 16.5.3: Let f i be a proper convex function on ℝ^n for each i in a finite index set. If the sets cl (dom f i) are all the same set C, then (supᵢ f i)^* = conv { (f i)^* | i }. Moreover, for each xStar, (supᵢ f i)^*(xStar) can be expressed as the infimum of ∑ i, λ i * (f i)^*(xStarᵢ) over all representations of xStar as a convex combination xStar = ∑ i, λ i • xStarᵢ, and this infimum is attained. Here dom f i is modeled by effectiveDomain univ (f i) and cl denotes topological closure of sets.
open scoped BigOperators
open scoped Pointwise
open scoped Topology
open Filter
open Set
theorem section16_fenchelConjugate_sSup_eq_convexHullFunctionFamily_of_finite_of_closure_effectiveDomain_eq
{n : ℕ} {ι : Type _} [Fintype ι] [Nonempty ι] (f : ι → (Fin n → ℝ) → EReal)
(hf : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ)) (f i)) (C : Set (Fin n → ℝ))
(hC : ∀ i, closure (effectiveDomain (Set.univ : Set (Fin n → ℝ)) (f i)) = C) :
(fenchelConjugate n (fun x => sSup (Set.range fun i => f i x)) =
convexHullFunctionFamily fun i => fenchelConjugate n (f i)) ∧
(∀ xStar : Fin n → ℝ,
let S : Set EReal :=
{r |
∃ (lam : ι → ℝ) (xStar_i : ι → Fin n → ℝ),
(∀ i, 0 ≤ lam i) ∧ (∑ i, lam i) = 1 ∧ (∑ i, (lam i) • (xStar_i i)) = xStar ∧
r = ∑ i, ((lam i : ℝ) : EReal) * fenchelConjugate n (f i) (xStar_i i)}
fenchelConjugate n (fun x => sSup (Set.range fun i => f i x)) xStar = sInf S ∧
∃ r, r ∈ S ∧ sInf S = r)§16 — Dual Operations — Part 15
Lean authors: Zichen Wang, Wanli Ma, Weiran Shi, Yunxi Duan, Zaiwen Wen.
The conjugate of x ↦ ‖x - a‖ is the unit-ball indicator plus a linear term.
open scoped BigOperators
open scoped Pointwise
lemma section16_fenchelConjugate_norm_sub_eq_indicator_add_dot {n : ℕ} (a : Fin n → ℝ) :
fenchelConjugate n (fun x => ((‖x - a‖ : ℝ) : EReal)) =
fun xStar : Fin n → ℝ =>
(if l1Norm xStar ≤ (1 : ℝ) then (0 : EReal) else (⊤ : EReal)) +
((dotProduct a xStar : ℝ) : EReal)
Translating the norm preserves proper convexity on Set.univ.
open scoped BigOperators
open scoped Pointwise
lemma section16_properConvexFunctionOn_norm_sub {n : ℕ} (a : Fin n → ℝ) :
ProperConvexFunctionOn (Set.univ : Set (Fin n → ℝ))
(fun x => ((‖x - a‖ : ℝ) : EReal))
The effective domain of x ↦ ‖x - a‖ is all of ℝ^n.
open scoped BigOperators
open scoped Pointwise
lemma section16_closure_effectiveDomain_norm_sub_eq_univ {n : ℕ} (a : Fin n → ℝ) :
closure (effectiveDomain (Set.univ : Set (Fin n → ℝ))
(fun x => ((‖x - a‖ : ℝ) : EReal))) = (Set.univ : Set (Fin n → ℝ))
Text 16.5.1: Let f(x) = max { ‖x - aᵢ‖ | i = 1, …, m }, where a₁, …, aₘ ∈ ℝⁿ. For each x⋆, the conjugate f⋆(x⋆) is the minimum of λ₁ ⟪a₁, x₁⋆⟫ + ⋯ + λₘ ⟪aₘ, xₘ⋆⟫ over all xᵢ⋆ and λᵢ satisfying - λ₁ x₁⋆ + ⋯ + λₘ xₘ⋆ = x⋆, - ‖xᵢ⋆‖ ≤ 1, - λᵢ ≥ 0 and λ₁ + ⋯ + λₘ = 1.
open scoped BigOperators
open scoped Pointwise
theorem section16_fenchelConjugate_sSup_norm_sub_points_eq_sInf {n m : ℕ} (a : Fin m → Fin n → ℝ)
(hm : 0 < m) :
let f : (Fin n → ℝ) → EReal :=
fun x => sSup (Set.range fun i : Fin m => ((‖x - a i‖ : ℝ) : EReal))
∀ xStar : Fin n → ℝ,
let S : Set EReal :=
{r |
∃ (lam : Fin m → ℝ) (xStar_i : Fin m → Fin n → ℝ),
(∀ i, 0 ≤ lam i) ∧
(∑ i, lam i) = 1 ∧
(∑ i, (lam i) • (xStar_i i)) = xStar ∧
(∀ i, l1Norm (xStar_i i) ≤ (1 : ℝ)) ∧
r = ((∑ i, lam i * (dotProduct (a i) (xStar_i i) : ℝ) : ℝ) : EReal)}
fenchelConjugate n f xStar = sInf S ∧ (S.Nonempty → ∃ r, r ∈ S ∧ sInf S = r)