Part II — Topological Properties
Proofs omitted. Hover over highlighted terms to see their Lean types.
Reading sections
§5 — Functional Operations
— source only (outside the published documentation root) (Lean source)
Lean authors: Zichen Wang, Wanli Ma, Yifan Bai, Zaiwen Wen.
Definiton 8.5.0. Let f be a proper convex function on R^n and let f0^+ be its recession function. The set of all vectors y such that (f0^+)(y) <= 0 is called the recession cone of f.
def Function.recessionCone {n : Nat} (f : EuclideanSpace Real (Fin n) → Real)
(f0_plus : EuclideanSpace Real (Fin n) → Real) :
Set (EuclideanSpace Real (Fin n)) :=
{ y | f0_plus y <= 0 }
§6 — Relative Interiors of Convex Sets — Part 1
Lean authors: Zichen Wang, Wanli Ma, Yunxi Duan, Zaiwen Wen.
Text 6.1: The Euclidean distance between two points x and y in R^n is defined by d(x, y) = |x - y| = sqrt (inner (x - y) (x - y)).
open scoped Pointwise Topology
def euclideanDist (n : Nat) (x y : EuclideanSpace Real (Fin n)) : Real :=
dist x y
The norm of the difference on a product space is convex on all of R^{2n}.
open scoped Pointwise Topology
lemma convexOn_univ_norm_comp_sub (n : Nat) :
ConvexOn Real Set.univ
(fun p : (EuclideanSpace Real (Fin n)) × (EuclideanSpace Real (Fin n)) => ‖p.1 - p.2‖)
Text 6.2: The function d, the Euclidean metric, is convex as a function on R^{2n}.
open scoped Pointwise Topology
theorem euclideanDist_convex (n : Nat) :
ConvexOn Real Set.univ
(fun p : (EuclideanSpace Real (Fin n)) × (EuclideanSpace Real (Fin n)) =>
dist p.1 p.2)
Text 6.3: Throughout this section, the Euclidean unit ball in R^n is B = {x | ‖x‖ ≤ 1} = {x | dist x 0 ≤ 1}.
open scoped Pointwise Topology
def euclideanUnitBall (n : Nat) : Set (EuclideanSpace Real (Fin n)) :=
{x | ‖x‖ ≤ (1 : Real)}
open scoped Pointwise Topology
theorem euclideanUnitBall_eq_dist (n : Nat) :
euclideanUnitBall n = {x | dist x 0 ≤ (1 : Real)}
Text 6.4: The Euclidean unit ball B is a closed convex set.
open scoped Pointwise Topology
theorem euclideanUnitBall_closed_convex (n : Nat) :
IsClosed (euclideanUnitBall n) ∧ Convex Real (euclideanUnitBall n)
Text 6.5: For any a ∈ R^n, the ball with radius ε > 0 and center a is {x | d(x, a) ≤ ε} = {a + y | |y| ≤ ε} = a + ε B. The closed ball centered at a is a translate of the radius-ε norm ball.
open scoped Pointwise Topology
lemma euclidean_closedBall_eq_translate_left (n : Nat) (a : EuclideanSpace Real (Fin n))
(ε : Real) :
{x | dist x a ≤ ε} = Set.image (fun y => a + y) {y | ‖y‖ ≤ ε}
Scaling the unit ball by ε ≥ 0 gives the radius-ε norm ball.
open scoped Pointwise Topology
lemma euclidean_normBall_eq_smul_unitBall (n : Nat) (ε : Real) (hε : 0 ≤ ε) :
{y | ‖y‖ ≤ ε} = ε • euclideanUnitBall n
Translating the ε-scaled unit ball equals scaling after translating.
open scoped Pointwise Topology
lemma image_translate_smul_unitBall (n : Nat) (a : EuclideanSpace Real (Fin n)) (ε : Real) :
Set.image (fun y => a + y) (ε • euclideanUnitBall n) =
Set.image (fun y => a + ε • y) (euclideanUnitBall n)
open scoped Pointwise Topology
theorem euclidean_closedBall_eq_translate (n : Nat) (a : EuclideanSpace Real (Fin n)) (ε : Real)
(hε : 0 < ε) :
{x | dist x a ≤ ε} = Set.image (fun y => a + y) {y | ‖y‖ ≤ ε} ∧
Set.image (fun y => a + y) {y | ‖y‖ ≤ ε} =
Set.image (fun y => a + ε • y) (euclideanUnitBall n)
Text 6.6: For any set C in R^n, the set of points x whose distance from C does not exceed ε is {x | ∃ y ∈ C, d(x, y) ≤ ε} = ⋃ {y + ε B | y ∈ C} = C + ε B.
open scoped Pointwise Topology
theorem euclidean_neighborhood_eq_iUnion_translate (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (ε : Real) (hε : 0 ≤ ε) :
{x | ∃ y ∈ C, dist x y ≤ ε} =
⋃ y ∈ C, (fun z => y + z) '' (ε • euclideanUnitBall n) ∧
(⋃ y ∈ C, (fun z => y + z) '' (ε • euclideanUnitBall n)) =
C + ε • euclideanUnitBall n
Text 6.7: The closure cl C of C can be expressed by cl C = ⋂ { C + ε B | ε > 0 }.
open scoped Pointwise Topology
theorem euclidean_closure_eq_iInter_thickening (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) :
closure C = ⋂ ε : {ε : Real // 0 < ε}, C + ε.1 • euclideanUnitBall n
Text 6.7: The interior int C of C can be expressed by int C = { x | ∃ ε > 0, x + ε B ⊆ C }.
open scoped Pointwise Topology
theorem euclidean_interior_eq_setOf_exists_thickening (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) :
interior C =
{x | ∃ ε : ℝ, 0 < ε ∧ (fun y => x + y) '' (ε • euclideanUnitBall n) ⊆ C}
Text 6.8: The relative interior ri C of a convex set C in R^n is the interior of C in its affine hull aff C. Equivalently, ri C = { x ∈ aff C | ∃ ε > 0, (x + ε B) ∩ aff C ⊆ C }.
open scoped Pointwise Topology
def euclideanRelativeInterior (n : Nat) (C : Set (EuclideanSpace Real (Fin n))) :
Set (EuclideanSpace Real (Fin n)) :=
{x |
x ∈ (affineSpan Real C : Set (EuclideanSpace Real (Fin n))) ∧
∃ ε : ℝ, 0 < ε ∧
((fun y => x + y) '' (ε • euclideanUnitBall n)) ∩
(affineSpan Real C : Set (EuclideanSpace Real (Fin n))) ⊆ C}
Text 6.9: Needless to say, ri C ⊂ C ⊂ cl C.
open scoped Pointwise Topology
theorem euclideanRelativeInterior_subset_closure (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) :
euclideanRelativeInterior n C ⊆ C ∧ C ⊆ closure C
Text 6.10: The relative boundary of C is the set difference (cl C) \ (ri C).
open scoped Pointwise Topology
def euclideanRelativeBoundary (n : Nat) (C : Set (EuclideanSpace Real (Fin n))) :
Set (EuclideanSpace Real (Fin n)) :=
closure C \ euclideanRelativeInterior n C
Text 6.11: Naturally, C is said to be relatively open if ri C = C.
open scoped Pointwise Topology
def euclideanRelativelyOpen (n : Nat) (C : Set (EuclideanSpace Real (Fin n))) : Prop :=
euclideanRelativeInterior n C = C
Text 6.12: For an n-dimensional convex set, aff C = R^n by definition, so ri C = int C.
open scoped Pointwise Topology
theorem euclideanRelativeInterior_eq_interior_of_affineSpan_eq_univ (n : Nat)
(C : Set (EuclideanSpace Real (Fin n)))
(hC : (affineSpan Real C : Set (EuclideanSpace Real (Fin n))) = Set.univ) :
euclideanRelativeInterior n C = interior C
The relative interior of an affine subspace is the subspace itself.
open scoped Pointwise Topology
lemma euclideanRelativeInterior_affineSubspace_eq (n : Nat)
(s : AffineSubspace Real (EuclideanSpace Real (Fin n))) :
euclideanRelativeInterior n (s : Set (EuclideanSpace Real (Fin n))) = s
Every affine subspace of Euclidean space is closed.
open scoped Pointwise Topology
lemma affineSubspace_isClosed (n : Nat)
(s : AffineSubspace Real (EuclideanSpace Real (Fin n))) :
IsClosed (s : Set (EuclideanSpace Real (Fin n)))
Text 6.13: An affine set is relatively open by definition. Every affine set is at the same time closed.
open scoped Pointwise Topology
theorem affineSubspace_relativelyOpen_closed (n : Nat)
(s : AffineSubspace Real (EuclideanSpace Real (Fin n))) :
euclideanRelativelyOpen n (s : Set (EuclideanSpace Real (Fin n))) ∧
IsClosed (s : Set (EuclideanSpace Real (Fin n)))
Text 6.14: Observe that cl C ⊆ cl (aff C) = aff C for any C.
open scoped Pointwise Topology
theorem euclidean_closure_subset_closure_affineSpan (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) :
closure C ⊆ closure (affineSpan Real C : Set (EuclideanSpace Real (Fin n))) ∧
closure (affineSpan Real C : Set (EuclideanSpace Real (Fin n))) =
(affineSpan Real C : Set (EuclideanSpace Real (Fin n)))
The translated scaled unit ball is the metric closed ball.
open scoped Pointwise Topology
lemma translate_smul_unitBall_eq_closedBall (n : Nat) (a : EuclideanSpace Real (Fin n)) (ε : Real)
(hε : 0 < ε) :
(fun y => a + y) '' (ε • euclideanUnitBall n) = Metric.closedBall a ε
Text 6.15: Closures and relative interiors are preserved under translations and, more generally, under any one-to-one affine transformation of R^n onto itself. Such maps preserve affine hulls and are continuous in both directions.
open scoped Pointwise Topology
theorem closure_image_affineEquiv (n : Nat) (C : Set (EuclideanSpace Real (Fin n)))
(e : EuclideanSpace Real (Fin n) ≃ᵃ[Real] EuclideanSpace Real (Fin n)) :
closure (e '' C) = e '' closure C
Affine equivalences send affine spans to affine spans of the images.
open scoped Pointwise Topology
lemma affineSpan_image_affineEquiv (n : Nat) (C : Set (EuclideanSpace Real (Fin n)))
(e : EuclideanSpace Real (Fin n) ≃ᵃ[Real] EuclideanSpace Real (Fin n)) :
e '' (affineSpan Real C : Set (EuclideanSpace Real (Fin n))) =
(affineSpan Real (e '' C) : Set (EuclideanSpace Real (Fin n)))
The inverse affine equivalence cancels the direct image of a set.
open scoped Pointwise Topology
lemma affineEquiv_symm_image_image (n : Nat) (C : Set (EuclideanSpace Real (Fin n)))
(e : EuclideanSpace Real (Fin n) ≃ᵃ[Real] EuclideanSpace Real (Fin n)) :
e.symm '' (e '' C) = C
An affine equivalence sends relative interior into relative interior.
open scoped Pointwise Topology
lemma ri_image_affineEquiv_subset (n : Nat) (C : Set (EuclideanSpace Real (Fin n)))
(e : EuclideanSpace Real (Fin n) ≃ᵃ[Real] EuclideanSpace Real (Fin n)) :
e '' euclideanRelativeInterior n C ⊆ euclideanRelativeInterior n (e '' C)
Text 6.15: Relative interiors are preserved under one-to-one affine transformations of R^n onto itself (hence in particular under translations).
open scoped Pointwise Topology
theorem euclideanRelativeInterior_image_affineEquiv (n : Nat)
(C : Set (EuclideanSpace Real (Fin n)))
(e : EuclideanSpace Real (Fin n) ≃ᵃ[Real] EuclideanSpace Real (Fin n)) :
euclideanRelativeInterior n (e '' C) = e '' euclideanRelativeInterior n C
The coordinate subspace where coordinates m, ..., n-1 vanish.
open scoped Pointwise Topology
def coordinateSubspace (n m : Nat) : Set (EuclideanSpace Real (Fin n)) :=
{x | ∀ i : Fin n, m ≤ (i : Nat) → x i = 0}
Evaluate LinearEquiv.piCongrLeft on a constant codomain.
open scoped Pointwise Topology
@[simp] lemma piCongrLeft_apply_const {ι ι' : Type*} (e : ι' ≃ ι) (f : ι' → ℝ) (i : ι) :
(LinearEquiv.piCongrLeft (R := ℝ) (φ := fun _ : ι => ℝ) e f) i = f (e.symm i)
Evaluate the inverse of LinearEquiv.piCongrLeft on a constant codomain.
open scoped Pointwise Topology
@[simp] lemma piCongrLeft_symm_apply_const {ι ι' : Type*} (e : ι' ≃ ι) (f : ι → ℝ) (i : ι') :
((LinearEquiv.piCongrLeft (R := ℝ) (φ := fun _ : ι => ℝ) e).symm f) i = f (e i)
Text 6.16: For example, if C is an m-dimensional convex set in R^n, there exists by Corollary 1.6.1 a one-to-one affine transformation T of R^n onto itself which carries aff C onto the subspace L = { x = (xi_1, ..., xi_n) | xi_{m+1} = 0, ..., xi_n = 0 }. This L can be regarded as a copy of R^m.
open scoped Pointwise Topology
theorem exists_affineEquiv_affineSpan_eq_coordinateSubspace (n m : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (hCnonempty : C.Nonempty)
(hCdim : Module.finrank Real (affineSpan Real C).direction = m) :
∃ T : EuclideanSpace Real (Fin n) ≃ᵃ[Real] EuclideanSpace Real (Fin n),
T '' (affineSpan Real C : Set (EuclideanSpace Real (Fin n))) =
coordinateSubspace n m
A half-radius translated ball around y stays inside the translated ball around x when y is within the half-radius ball around x.
open scoped Pointwise Topology
lemma translate_smul_unitBall_subset_translate_smul_unitBall (n : Nat)
{x y : EuclideanSpace Real (Fin n)} {ε : Real} (hε : 0 < ε)
(hy : y ∈ (fun v => x + v) '' ((ε / 2) • euclideanUnitBall n)) :
(fun v => y + v) '' ((ε / 2) • euclideanUnitBall n) ⊆
(fun v => x + v) '' (ε • euclideanUnitBall n)
A half-radius relative ball around a relative interior point is still in the relative interior.
open scoped Pointwise Topology
lemma small_ball_subset_relativeInterior (n : Nat)
{C : Set (EuclideanSpace Real (Fin n))} {x : EuclideanSpace Real (Fin n)}
(hx : x ∈ euclideanRelativeInterior n C) :
∃ ε : ℝ, 0 < ε ∧
((fun y => x + y) '' ((ε / 2) • euclideanUnitBall n)) ∩
(affineSpan Real C : Set (EuclideanSpace Real (Fin n))) ⊆
euclideanRelativeInterior n C
Text 6.17: For any set C in R^n, convex or not, the laws cl (cl C) = cl C and ri (ri C) = ri C are valid.
open scoped Pointwise Topology
theorem euclidean_closure_closure_eq_and_relativeInterior_relativeInterior_eq (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) :
closure (closure C) = closure C ∧
euclideanRelativeInterior n (euclideanRelativeInterior n C) =
euclideanRelativeInterior n C
Express the direct-sum set as the image of the product under the append map.
open scoped Pointwise Topology
lemma directSumSetEuclidean_eq_image_append (m p : Nat)
(C1 : Set (EuclideanSpace Real (Fin m))) (C2 : Set (EuclideanSpace Real (Fin p))) :
let append :
EuclideanSpace Real (Fin m) → EuclideanSpace Real (Fin p) →
EuclideanSpace Real (Fin (m + p)) :=
fun x y =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (m + p))).symm
((Fin.appendIsometry m p)
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin m)) x,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin p)) y))
directSumSetEuclidean m p C1 C2 = (fun xy => append xy.1 xy.2) '' Set.prod C1 C2
The append map on Euclidean spaces is a homeomorphism.
open scoped Pointwise Topology
noncomputable def appendHomeomorph (m p : Nat) :
(EuclideanSpace Real (Fin m) × EuclideanSpace Real (Fin p)) ≃ₜ
EuclideanSpace Real (Fin (m + p)) :=
(Homeomorph.prodCongr
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin m)).toHomeomorph
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin p)).toHomeomorph
).trans
((Fin.appendIsometry m p).toHomeomorph.trans
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (m + p))).symm.toHomeomorph)
The append map on Euclidean spaces is an affine equivalence.
open scoped Pointwise Topology
noncomputable def appendAffineEquiv (m p : Nat) :
(EuclideanSpace Real (Fin m) × EuclideanSpace Real (Fin p)) ≃ᵃ[Real]
EuclideanSpace Real (Fin (m + p))
The append affine equivalence acts as the explicit append map on coordinates.
open scoped Pointwise Topology
lemma appendAffineEquiv_apply (m p : Nat) (x : EuclideanSpace Real (Fin m))
(y : EuclideanSpace Real (Fin p)) :
appendAffineEquiv m p (x, y) =
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (m + p))).symm
((Fin.appendIsometry m p)
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin m)) x,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin p)) y))
Express the direct-sum set as the image of the product under appendAffineEquiv.
open scoped Pointwise Topology
lemma directSumSetEuclidean_eq_image_appendAffineEquiv (m p : Nat)
(C1 : Set (EuclideanSpace Real (Fin m))) (C2 : Set (EuclideanSpace Real (Fin p))) :
directSumSetEuclidean m p C1 C2 = (appendAffineEquiv m p) '' Set.prod C1 C2
The affine span of a product is the product of affine spans.
open scoped Pointwise Topology
lemma affineSpan_prod_eq (m p : Nat)
(C1 : Set (EuclideanSpace Real (Fin m))) (C2 : Set (EuclideanSpace Real (Fin p))) :
(affineSpan Real (Set.prod C1 C2) :
Set ((EuclideanSpace Real (Fin m)) × (EuclideanSpace Real (Fin p)))) =
Set.prod (affineSpan Real C1 : Set (EuclideanSpace Real (Fin m)))
(affineSpan Real C2 : Set (EuclideanSpace Real (Fin p)))
The affine span of the direct-sum set factors as the direct sum of affine spans.
open scoped Pointwise Topology
lemma affineSpan_directSumSetEuclidean_eq (m p : Nat)
(C1 : Set (EuclideanSpace Real (Fin m))) (C2 : Set (EuclideanSpace Real (Fin p))) :
(affineSpan Real (directSumSetEuclidean m p C1 C2) :
Set (EuclideanSpace Real (Fin (m + p)))) =
directSumSetEuclidean m p (affineSpan Real C1 : Set (EuclideanSpace Real (Fin m)))
(affineSpan Real C2 : Set (EuclideanSpace Real (Fin p)))
The closure of the direct-sum set factors as the direct sum of closures.
open scoped Pointwise Topology
lemma closure_directSumSetEuclidean_eq (m p : Nat)
(C1 : Set (EuclideanSpace Real (Fin m))) (C2 : Set (EuclideanSpace Real (Fin p))) :
closure (directSumSetEuclidean m p C1 C2) =
directSumSetEuclidean m p (closure C1) (closure C2)
Forward inclusion for the relative interior of a direct sum.
open scoped Pointwise Topology
lemma euclideanRelativeInterior_directSumSetEuclidean_subset (m p : Nat)
(C1 : Set (EuclideanSpace Real (Fin m))) (C2 : Set (EuclideanSpace Real (Fin p))) :
euclideanRelativeInterior (m + p) (directSumSetEuclidean m p C1 C2) ⊆
directSumSetEuclidean m p (euclideanRelativeInterior m C1)
(euclideanRelativeInterior p C2)
Auxiliary: the direct-sum set C1 ⊕ C2 in R^{m+p}, built via Fin.appendIsometry.
open scoped Pointwise Topology
def directSumSetEuclidean (m p : Nat)
(C1 : Set (EuclideanSpace Real (Fin m))) (C2 : Set (EuclideanSpace Real (Fin p))) :
Set (EuclideanSpace Real (Fin (m + p))) :=
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (m + p))).symm ''
((Fin.appendIsometry m p) '' Set.prod
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin m)) '' C1)
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin p)) '' C2))§6 — Relative Interiors of Convex Sets — Part 2
Lean authors: Zichen Wang, Wanli Ma, Yunxi Duan, Zaiwen Wen.
Reverse inclusion for the relative interior of a direct sum.
open scoped Pointwise
lemma euclideanRelativeInterior_directSumSetEuclidean_superset (m p : Nat)
(C1 : Set (EuclideanSpace Real (Fin m))) (C2 : Set (EuclideanSpace Real (Fin p))) :
directSumSetEuclidean m p (euclideanRelativeInterior m C1)
(euclideanRelativeInterior p C2) ⊆
euclideanRelativeInterior (m + p) (directSumSetEuclidean m p C1 C2)
Text 6.18: For the direct sum C1 ⊕ C2 in R^{m+p} of convex sets C1 ⊆ R^m and C2 ⊆ R^p (modeled as directSumSetEuclidean m p C1 C2), one has ri (C1 ⊕ C2) = ri C1 ⊕ ri C2 and cl (C1 ⊕ C2) = cl C1 ⊕ cl C2.
open scoped Pointwise
theorem euclideanRelativeInterior_directSumSetEuclidean_eq_and_closure_eq (m p : Nat)
(C1 : Set (EuclideanSpace Real (Fin m))) (C2 : Set (EuclideanSpace Real (Fin p)))
:
euclideanRelativeInterior (m + p) (directSumSetEuclidean m p C1 C2) =
directSumSetEuclidean m p (euclideanRelativeInterior m C1)
(euclideanRelativeInterior p C2) ∧
closure (directSumSetEuclidean m p C1 C2) =
directSumSetEuclidean m p (closure C1) (closure C2)
Points in the closure lie in the affine span.
open scoped Pointwise
lemma mem_affineSpan_of_mem_closure {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
{y : EuclideanSpace Real (Fin n)} (hy : y ∈ closure C) :
y ∈ (affineSpan Real C : Set (EuclideanSpace Real (Fin n)))
A point in a scaled unit ball has norm bounded by the scale.
open scoped Pointwise
lemma norm_le_of_mem_smul_unitBall {n : Nat} {ε : Real} (hε : 0 ≤ ε)
{v : EuclideanSpace Real (Fin n)} (hv : v ∈ ε • euclideanUnitBall n) : ‖v‖ ≤ ε
If a point has norm at most ε, it lies in the scaled unit ball ε • B.
open scoped Pointwise
lemma mem_smul_unitBall_of_norm_le {n : Nat} {ε : Real} (hε : 0 < ε)
{v : EuclideanSpace Real (Fin n)} (hv : ‖v‖ ≤ ε) :
v ∈ ε • euclideanUnitBall n
Theorem 6.1: Let C be a convex set in R^n. Let x ∈ ri C and y ∈ cl C. Then (1 - λ) x + λ y belongs to ri C (and hence in particular to C) for 0 ≤ λ < 1.
open scoped Pointwise
theorem euclideanRelativeInterior_convex_combination_mem (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (hC : Convex Real C)
(x y : EuclideanSpace Real (Fin n))
(hx : x ∈ euclideanRelativeInterior n C)
(hy : y ∈ closure C) (t : Real) (ht0 : 0 ≤ t) (ht1 : t < 1) :
(1 - t) • x + t • y ∈ euclideanRelativeInterior n C
The closure of a convex set is convex.
open scoped Pointwise
lemma convex_closure_of_convex (n : Nat) (C : Set (EuclideanSpace Real (Fin n)))
(hC : Convex Real C) : Convex Real (closure C)§6 — Relative Interiors of Convex Sets — Part 3
Lean authors: Zichen Wang, Wanli Ma, Yunxi Duan, Zaiwen Wen.
The relative interior of a convex set is convex.
open scoped Pointwise
lemma convex_euclideanRelativeInterior (n : Nat) (C : Set (EuclideanSpace Real (Fin n)))
(hC : Convex Real C) : Convex Real (euclideanRelativeInterior n C)
The affine span of the closure equals the affine span of the set.
open scoped Pointwise
lemma affineSpan_closure_eq (n : Nat) (C : Set (EuclideanSpace Real (Fin n))) :
affineSpan Real (closure C) = affineSpan Real C
The intrinsic interior is contained in the Euclidean relative interior.
open scoped Pointwise
lemma intrinsicInterior_subset_euclideanRelativeInterior (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) :
intrinsicInterior ℝ C ⊆ euclideanRelativeInterior n C
The affine span of the intrinsic interior equals the affine span of a nonempty convex set.
open scoped Pointwise
lemma affineSpan_intrinsicInterior_eq (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (hC : Convex Real C) (hCne : C.Nonempty) :
affineSpan Real (intrinsicInterior ℝ C) = affineSpan Real C
Theorem 6.2: Let C be a convex set in R^n. Then cl C and ri C are convex sets having the same affine hull (hence the same dimension) as C. In particular, ri C ≠ ∅ if C ≠ ∅.
open scoped Pointwise
theorem euclideanRelativeInterior_closure_convex_affineSpan (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (hC : Convex Real C) :
Convex Real (closure C) ∧
Convex Real (euclideanRelativeInterior n C) ∧
affineSpan Real (closure C) = affineSpan Real C ∧
affineSpan Real (euclideanRelativeInterior n C) = affineSpan Real C ∧
(C.Nonempty → (euclideanRelativeInterior n C).Nonempty)
For a nonempty convex set, cl C is contained in cl (ri C).
open scoped Pointwise
lemma euclidean_closure_subset_closure_relativeInterior_of_nonempty (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (hC : Convex Real C) (hCne : C.Nonempty) :
closure C ⊆ closure (euclideanRelativeInterior n C)
The relative interior is monotone with respect to closure.
open scoped Pointwise
lemma euclideanRelativeInterior_subset_relativeInterior_closure (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) :
euclideanRelativeInterior n C ⊆ euclideanRelativeInterior n (closure C)
For a nonempty convex set, ri (cl C) is contained in ri C.
open scoped Pointwise
lemma euclideanRelativeInterior_closure_subset_relativeInterior (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (hC : Convex Real C) (hCne : C.Nonempty) :
euclideanRelativeInterior n (closure C) ⊆ euclideanRelativeInterior n C
Theorem 6.3: For any convex set C in R^n, cl (ri C) = cl C and ri (cl C) = ri C.
open scoped Pointwise
theorem euclidean_closure_relativeInterior_eq_and_relativeInterior_closure_eq (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (hC : Convex Real C) :
closure (euclideanRelativeInterior n C) = closure C ∧
euclideanRelativeInterior n (closure C) = euclideanRelativeInterior n C
Closure equality for convex sets yields equality of relative interiors.
open scoped Pointwise
lemma euclideanRelativeInterior_eq_of_closure_eq (n : Nat)
(C1 C2 : Set (EuclideanSpace Real (Fin n))) (hC1 : Convex Real C1)
(hC2 : Convex Real C2) :
closure C1 = closure C2 →
euclideanRelativeInterior n C1 = euclideanRelativeInterior n C2
Relative interior equality for convex sets yields equality of closures.
open scoped Pointwise
lemma euclidean_closure_eq_of_relativeInterior_eq (n : Nat)
(C1 C2 : Set (EuclideanSpace Real (Fin n))) (hC1 : Convex Real C1)
(hC2 : Convex Real C2) :
euclideanRelativeInterior n C1 = euclideanRelativeInterior n C2 →
closure C1 = closure C2
Closure equality yields a relative interior and closure sandwich.
open scoped Pointwise
lemma euclidean_between_of_closure_eq (n : Nat)
(C1 C2 : Set (EuclideanSpace Real (Fin n))) (hC1 : Convex Real C1)
(hC2 : Convex Real C2) :
closure C1 = closure C2 →
(euclideanRelativeInterior n C1 ⊆ C2 ∧ C2 ⊆ closure C1)
A relative interior and closure sandwich yields equality of closures.
open scoped Pointwise
lemma euclidean_closure_eq_of_between (n : Nat)
(C1 C2 : Set (EuclideanSpace Real (Fin n))) (hC1 : Convex Real C1) :
(euclideanRelativeInterior n C1 ⊆ C2 ∧ C2 ⊆ closure C1) →
closure C1 = closure C2
Corollary 6.3.1: Let C1 and C2 be convex sets in R^n. Then cl C1 = cl C2 if and only if ri C1 = ri C2. These conditions are equivalent to ri C1 ⊆ C2 ⊆ cl C1.
open scoped Pointwise
theorem euclidean_closure_eq_iff_relativeInterior_eq_and_between (n : Nat)
(C1 C2 : Set (EuclideanSpace Real (Fin n))) (hC1 : Convex Real C1)
(hC2 : Convex Real C2) :
(closure C1 = closure C2 ↔ euclideanRelativeInterior n C1 = euclideanRelativeInterior n C2) ∧
(closure C1 = closure C2 ↔
(euclideanRelativeInterior n C1 ⊆ C2 ∧ C2 ⊆ closure C1))
Corollary 6.3.2. If C is a convex set in R^n, then every open set which meets cl C also meets ri C.
open scoped Pointwise
theorem euclidean_open_meets_closure_meets_relativeInterior (n : Nat)
(C U : Set (EuclideanSpace Real (Fin n))) (hC : Convex Real C)
(hU : IsOpen U) :
(U ∩ closure C).Nonempty → (U ∩ euclideanRelativeInterior n C).Nonempty
A subset of the relative boundary has affine span contained in the ambient affine span.
open scoped Pointwise
lemma affineSpan_le_of_subset_relativeBoundary (n : Nat)
(C1 C2 : Set (EuclideanSpace Real (Fin n)))
(hC1sub : C1 ⊆ euclideanRelativeBoundary n C2) :
affineSpan Real C1 ≤ affineSpan Real C2
Equal direction finrank and nonempty inclusion forces equality of affine spans.
open scoped Pointwise
lemma affineSpan_eq_of_nonempty_of_finrank_eq (n : Nat)
(C1 C2 : Set (EuclideanSpace Real (Fin n))) (hC1ne : C1.Nonempty)
(hspan : affineSpan Real C1 ≤ affineSpan Real C2)
(hfin :
Module.finrank Real (affineSpan Real C1).direction =
Module.finrank Real (affineSpan Real C2).direction) :
affineSpan Real C1 = affineSpan Real C2
A relative interior point cannot lie in the closure of a disjoint relative interior.
open scoped Pointwise
lemma ri_point_not_mem_closure_ri_of_disjoint (n : Nat)
{C1 C2 : Set (EuclideanSpace Real (Fin n))} {x : EuclideanSpace Real (Fin n)}
(hx : x ∈ euclideanRelativeInterior n C1)
(hspan : affineSpan Real C1 = affineSpan Real C2)
(hdisj : Disjoint C1 (euclideanRelativeInterior n C2)) :
x ∉ closure (euclideanRelativeInterior n C2)
Corollary 6.3.3. If C1 is a non-empty convex subset of the relative boundary of a non-empty convex set C2 in R^n, then dim C1 < dim C2.
open scoped Pointwise
theorem euclidean_dim_lt_of_convex_subset_relativeBoundary (n : Nat)
(C1 C2 : Set (EuclideanSpace Real (Fin n))) (hC1 : Convex Real C1)
(hC2 : Convex Real C2) (hC1ne : C1.Nonempty) (hC2ne : C2.Nonempty)
(hC1sub : C1 ⊆ euclideanRelativeBoundary n C2) :
Module.finrank Real (affineSpan Real C1).direction <
Module.finrank Real (affineSpan Real C2).direction
A relative interior point allows extending any segment past it inside C.
open scoped Pointwise
lemma euclideanRelativeInterior_exists_extension_point (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) {z : EuclideanSpace Real (Fin n)}
(hz : z ∈ euclideanRelativeInterior n C) :
∀ x ∈ C, ∃ μ > (1 : Real), (1 - μ) • x + μ • z ∈ C
Theorem 6.4: Let C be a non-empty convex set in R^n. Then z ∈ ri C if and only if, for every x ∈ C, there exists μ > 1 such that (1 - μ) x + μ z belongs to C.
open scoped Pointwise
theorem euclideanRelativeInterior_iff_forall_exists_affine_combination_mem (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (hC : Convex Real C) (hCne : C.Nonempty)
(z : EuclideanSpace Real (Fin n)) :
z ∈ euclideanRelativeInterior n C ↔
∀ x ∈ C, ∃ μ > (1 : Real), (1 - μ) • x + μ • z ∈ C
Interior points allow stepping a positive distance in any direction.
open scoped Pointwise
lemma interior_imp_forall_exists_add_smul_mem (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (z : EuclideanSpace Real (Fin n)) :
z ∈ interior C →
∀ y : EuclideanSpace Real (Fin n), ∃ ε > (0 : Real), z + ε • y ∈ C
Stepping in every direction yields the extension property of Theorem 6.4.
open scoped Pointwise
lemma forall_exists_add_smul_mem_imp_extension (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (z : EuclideanSpace Real (Fin n))
(h : ∀ y : EuclideanSpace Real (Fin n), ∃ ε > (0 : Real), z + ε • y ∈ C) :
∀ x ∈ C, ∃ μ > (1 : Real), (1 - μ) • x + μ • z ∈ C
Stepping in every direction forces the affine span of C to be all of R^n.
open scoped Pointwise
lemma affineSpan_eq_univ_of_forall_exists_add_smul_mem (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (z : EuclideanSpace Real (Fin n))
(h : ∀ y : EuclideanSpace Real (Fin n), ∃ ε > (0 : Real), z + ε • y ∈ C) :
(affineSpan Real C : Set (EuclideanSpace Real (Fin n))) = Set.univ
Corollary 6.4.1: Let C be a convex set in R^n. Then z ∈ int C if and only if, for every y ∈ R^n, there exists ε > 0 such that z + ε y ∈ C.
open scoped Pointwise
theorem euclidean_interior_iff_forall_exists_add_smul_mem (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (hC : Convex Real C)
(z : EuclideanSpace Real (Fin n)) :
z ∈ interior C ↔
∀ y : EuclideanSpace Real (Fin n), ∃ ε > (0 : Real), z + ε • y ∈ C
Points in all closures lie in the closure of the intersection of relative interiors.
open scoped Pointwise
lemma iInter_closure_subset_closure_iInter_relativeInterior (n : Nat) {I : Type*}
(C : I → Set (EuclideanSpace Real (Fin n))) (hC : ∀ i, Convex Real (C i))
(hri : (⋂ i, euclideanRelativeInterior n (C i)).Nonempty) :
(⋂ i, closure (C i)) ⊆ closure (⋂ i, euclideanRelativeInterior n (C i))
Shrinking the extension parameter preserves membership in a convex set.
open scoped Pointwise
lemma affine_combination_mem_of_between (n : Nat) (C : Set (EuclideanSpace Real (Fin n)))
(hC : Convex Real C) {x z : EuclideanSpace Real (Fin n)} (hz : z ∈ C)
{μ0 μ : Real} (hμ0 : 1 < μ0) (hμ1 : 1 ≤ μ) (hμ : μ ≤ μ0)
(hμ0mem : (1 - μ0) • x + μ0 • z ∈ C) :
(1 - μ) • x + μ • z ∈ C
A uniform extension parameter exists in a finite intersection of convex sets.
open scoped Pointwise
lemma exists_mu_gt_one_mem_iInter_of_finite (n : Nat) {I : Type*} [Finite I]
(C : I → Set (EuclideanSpace Real (Fin n))) (hC : ∀ i, Convex Real (C i))
{x z : EuclideanSpace Real (Fin n)}
(hz : z ∈ ⋂ i, euclideanRelativeInterior n (C i)) (hx : x ∈ ⋂ i, C i) :
∃ μ > (1 : Real), (1 - μ) • x + μ • z ∈ ⋂ i, C i
Theorem 6.5: Let C i be convex sets in R^n for i ∈ I, and let the sets ri (C i) have a common point. Then cl (⋂ i, C i) = ⋂ i, cl (C i). If I is finite, then also ri (⋂ i, C i) = ⋂ i, ri (C i).
open scoped Pointwise
theorem euclidean_closure_iInter_eq_iInter_closure_of_common_relativeInterior (n : Nat)
{I : Type*} (C : I → Set (EuclideanSpace Real (Fin n))) (hC : ∀ i, Convex Real (C i))
(hri : (⋂ i, euclideanRelativeInterior n (C i)).Nonempty) :
closure (⋂ i, C i) = ⋂ i, closure (C i)
Theorem 6.5: If I is finite, then ri (⋂ i, C i) = ⋂ i, ri (C i) under the same assumptions as the preceding statement.
open scoped Pointwise
theorem euclideanRelativeInterior_iInter_eq_iInter_relativeInterior_of_finite (n : Nat)
{I : Type*} [Finite I] (C : I → Set (EuclideanSpace Real (Fin n)))
(hC : ∀ i, Convex Real (C i))
(hri : (⋂ i, euclideanRelativeInterior n (C i)).Nonempty) :
euclideanRelativeInterior n (⋂ i, C i) = ⋂ i, euclideanRelativeInterior n (C i)
The intersection of a Fin 2-indexed family is the binary intersection.
open scoped Pointwise
lemma iInter_fin_two_eq_inter {α : Type*} (D : Fin 2 → Set α) :
(⋂ i, D i) = D 0 ∩ D 1
A point of M ∩ ri C gives a point in the intersection of the relative interiors.
open scoped Pointwise
lemma ri_iInter_nonempty_from_hM (n : Nat) (C : Set (EuclideanSpace Real (Fin n)))
(M : AffineSubspace Real (EuclideanSpace Real (Fin n))) (D : Fin 2 → Set _)
(hD0 : D 0 = (M : Set (EuclideanSpace Real (Fin n)))) (hD1 : D 1 = C)
(hM : ((M : Set (EuclideanSpace Real (Fin n))) ∩ euclideanRelativeInterior n C).Nonempty) :
(⋂ i, euclideanRelativeInterior n (D i)).Nonempty
The closure of an affine subspace is itself.
open scoped Pointwise
lemma closure_affineSubspace_eq (n : Nat)
(M : AffineSubspace Real (EuclideanSpace Real (Fin n))) :
closure (M : Set (EuclideanSpace Real (Fin n))) = M
Corollary 6.5.1. Let C be convex and let M be an affine set containing a point of ri C. Then ri (M ∩ C) = M ∩ ri C and cl (M ∩ C) = M ∩ cl C.
open scoped Pointwise
theorem euclideanRelativeInterior_inter_affineSubspace_eq_and_closure_eq (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (hC : Convex Real C)
(M : AffineSubspace Real (EuclideanSpace Real (Fin n)))
(hM : ((M : Set (EuclideanSpace Real (Fin n))) ∩ euclideanRelativeInterior n C).Nonempty) :
euclideanRelativeInterior n ((M : Set (EuclideanSpace Real (Fin n))) ∩ C) =
(M : Set (EuclideanSpace Real (Fin n))) ∩ euclideanRelativeInterior n C ∧
closure ((M : Set (EuclideanSpace Real (Fin n))) ∩ C) =
(M : Set (EuclideanSpace Real (Fin n))) ∩ closure C§6 — Relative Interiors of Convex Sets — Part 4
Lean authors: Zichen Wang, Wanli Ma, Yunxi Duan, Zaiwen Wen.
A set not contained in the relative boundary is nonempty.
open scoped Pointwise
lemma C2_nonempty_of_not_subset_relativeBoundary (n : Nat)
(C1 C2 : Set (EuclideanSpace Real (Fin n)))
(hC2not : ¬ C2 ⊆ euclideanRelativeBoundary n C1) : C2.Nonempty
A relative interior point has a smaller ball in the affine span contained in the relative interior.
open scoped Pointwise
lemma euclideanRelativeInterior_ball_subset (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) {x : EuclideanSpace Real (Fin n)}
(hx : x ∈ euclideanRelativeInterior n C) :
∃ ε : ℝ, 0 < ε ∧
((fun y => x + y) '' (ε • euclideanUnitBall n)) ∩
(affineSpan Real C : Set (EuclideanSpace Real (Fin n))) ⊆
euclideanRelativeInterior n C
Under the corollary hypotheses, the relative interiors of C1 and C2 intersect.
open scoped Pointwise
lemma riC2_inter_riC1_nonempty (n : Nat)
(C1 C2 : Set (EuclideanSpace Real (Fin n)))
(hC2 : Convex Real C2) (hC2sub : C2 ⊆ closure C1)
(hC2not : ¬ C2 ⊆ euclideanRelativeBoundary n C1) :
(euclideanRelativeInterior n C2 ∩ euclideanRelativeInterior n C1).Nonempty
Relative interiors of two convex sets coincide with the relative interior of their intersection when their relative interiors meet.
open scoped Pointwise
lemma ri_intersection_fin_two (n : Nat)
(C1 C2 : Set (EuclideanSpace Real (Fin n))) (hC1 : Convex Real C1)
(hC2 : Convex Real C2)
(hri :
(euclideanRelativeInterior n C2 ∩ euclideanRelativeInterior n (closure C1)).Nonempty) :
euclideanRelativeInterior n (C2 ∩ closure C1) =
euclideanRelativeInterior n C2 ∩ euclideanRelativeInterior n (closure C1)
Corollary 6.5.2. Let C1 be a convex set. Let C2 be a convex set contained in cl C1 but not entirely contained in the relative boundary of C1. Then ri C2 ⊆ ri C1.
open scoped Pointwise
theorem euclideanRelativeInterior_subset_of_subset_closure_not_subset_relativeBoundary (n : Nat)
(C1 C2 : Set (EuclideanSpace Real (Fin n))) (hC1 : Convex Real C1)
(hC2 : Convex Real C2) (hC2sub : C2 ⊆ closure C1)
(hC2not : ¬ C2 ⊆ euclideanRelativeBoundary n C1) :
euclideanRelativeInterior n C2 ⊆ euclideanRelativeInterior n C1
Linear maps send closures into the closure of the image.
open scoped Pointwise
lemma image_closure_subset_closure_image_linearMap (n m : Nat)
(C : Set (EuclideanSpace Real (Fin n)))
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m)) :
A '' closure C ⊆ closure (A '' C)
Linear images of relative interior points stay in the relative interior of the image.
open scoped Pointwise
lemma image_relativeInterior_subset_relativeInterior_image (n m : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (hC : Convex Real C)
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m)) :
A '' euclideanRelativeInterior n C ⊆ euclideanRelativeInterior m (A '' C)
Linear images of a convex set and its relative interior have the same closure.
open scoped Pointwise
lemma closure_image_eq_closure_image_ri (n m : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (hC : Convex Real C)
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m)) :
closure (A '' C) = closure (A '' euclideanRelativeInterior n C)
The relative interior of a linear image lies in the image of the relative interior.
open scoped Pointwise
lemma relativeInterior_image_subset_image_relativeInterior (n m : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (hC : Convex Real C)
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m)) :
euclideanRelativeInterior m (A '' C) ⊆ A '' euclideanRelativeInterior n C
Theorem 6.6: Let C be a convex set in R^n, and let A be a linear transformation from R^n to R^m. Then ri (A C) = A (ri C) and cl (A C) ⊇ A (cl C).
open scoped Pointwise
theorem euclideanRelativeInterior_image_linearMap_eq_and_image_closure_subset (n m : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (hC : Convex Real C)
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m)) :
euclideanRelativeInterior m (A '' C) = A '' euclideanRelativeInterior n C ∧
A '' closure C ⊆ closure (A '' C)
Relative interior commutes with LinearMap.lsmul as a linear image.
open scoped Pointwise
lemma ri_image_lsmul_eq (n : Nat) (C : Set (EuclideanSpace Real (Fin n))) (hC : Convex Real C)
(r : Real) :
euclideanRelativeInterior n
((LinearMap.lsmul Real (EuclideanSpace Real (Fin n)) r) '' C) =
(LinearMap.lsmul Real (EuclideanSpace Real (Fin n)) r) ''
euclideanRelativeInterior n C
The image of a set under LinearMap.lsmul is pointwise scalar multiplication.
open scoped Pointwise
lemma image_lsmul_eq_smul_set (n : Nat) (C : Set (EuclideanSpace Real (Fin n))) (r : Real) :
(LinearMap.lsmul Real (EuclideanSpace Real (Fin n)) r) '' C = r • C
Corollary 6.6.1. For any convex set C and any real number λ, ri (λ C) = λ ri C.
open scoped Pointwise
theorem euclideanRelativeInterior_smul (n : Nat) (C : Set (EuclideanSpace Real (Fin n)))
(hC : Convex Real C) (r : Real) :
euclideanRelativeInterior n (r • C) = r • euclideanRelativeInterior n C§6 — Relative Interiors of Convex Sets — Part 5
Lean authors: Zichen Wang, Wanli Ma, Yunxi Duan, Zaiwen Wen.
The Minkowski sum is the image of the product under addition.
open scoped Pointwise
lemma image_add_prod_eq_add {E : Type*} [Add E] (C1 C2 : Set E) :
(fun xy : E × E => xy.1 + xy.2) '' Set.prod C1 C2 = C1 + C2
The append affine equivalence fixes the origin, hence equals its linear part.
open scoped Pointwise
lemma appendAffineEquiv_eq_linear_toAffineEquiv (m p : Nat) :
(appendAffineEquiv m p :
(EuclideanSpace Real (Fin m) × EuclideanSpace Real (Fin p)) →ᵃ[Real]
EuclideanSpace Real (Fin (m + p))) =
(appendAffineEquiv m p).linear.toAffineEquiv
The addition linear map sends the direct-sum set to the Minkowski sum.
open scoped Pointwise
lemma addLinearMap_image_directSum_eq_add (n : Nat)
(C1 C2 : Set (EuclideanSpace Real (Fin n))) :
let A : EuclideanSpace Real (Fin (n + n)) →ₗ[Real] EuclideanSpace Real (Fin n) :=
((LinearMap.fst (R := Real) (M := EuclideanSpace Real (Fin n))
(M₂ := EuclideanSpace Real (Fin n))) +
(LinearMap.snd (R := Real) (M := EuclideanSpace Real (Fin n))
(M₂ := EuclideanSpace Real (Fin n)))).comp
(appendAffineEquiv n n).symm.linear.toLinearMap
A '' directSumSetEuclidean n n C1 C2 = C1 + C2
Convexity of the direct-sum set.
open scoped Pointwise
lemma convex_directSumSetEuclidean (n : Nat)
(C1 C2 : Set (EuclideanSpace Real (Fin n))) (hC1 : Convex Real C1)
(hC2 : Convex Real C2) :
Convex Real (directSumSetEuclidean n n C1 C2)
Convexity of a direct-sum set in general dimensions.
open scoped Pointwise
lemma convex_directSumSetEuclidean_general (n m : Nat)
(C1 : Set (EuclideanSpace Real (Fin n))) (C2 : Set (EuclideanSpace Real (Fin m)))
(hC1 : Convex Real C1) (hC2 : Convex Real C2) :
Convex Real (directSumSetEuclidean n m C1 C2)
Corollary 6.6.2. For any convex sets C1 and C2 in R^n, ri (C1 + C2) = ri C1 + ri C2, and cl (C1 + C2) ⊇ cl C1 + cl C2.
open scoped Pointwise
theorem euclideanRelativeInterior_add_eq_and_closure_add_superset (n : Nat)
(C1 C2 : Set (EuclideanSpace Real (Fin n))) (hC1 : Convex Real C1)
(hC2 : Convex Real C2) :
euclideanRelativeInterior n (C1 + C2) =
euclideanRelativeInterior n C1 + euclideanRelativeInterior n C2 ∧
closure C1 + closure C2 ⊆ closure (C1 + C2)
Projection of a graph intersected with a direct-sum set recovers the preimage.
open scoped Pointwise
lemma preimage_eq_proj_graph_inter_directSum (n m : Nat)
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m))
(S : Set (EuclideanSpace Real (Fin m))) :
let e := appendAffineEquiv n m
let P : EuclideanSpace Real (Fin (n + m)) →ₗ[Real] EuclideanSpace Real (Fin n) :=
(LinearMap.fst (R := Real) (M := EuclideanSpace Real (Fin n))
(M₂ := EuclideanSpace Real (Fin m))).comp e.symm.linear.toLinearMap
let M : AffineSubspace Real (EuclideanSpace Real (Fin (n + m))) :=
(A.graph).toAffineSubspace.map e.toAffineMap
let D_S : Set (EuclideanSpace Real (Fin (n + m))) :=
directSumSetEuclidean n m Set.univ S
P '' ((M : Set _) ∩ D_S) = A ⁻¹' S
Relative interior and closure of univ ⊕ C.
open scoped Pointwise
lemma ri_closure_directSum_univ_eq (n m : Nat)
(C : Set (EuclideanSpace Real (Fin m))) :
let D := directSumSetEuclidean n m Set.univ C
euclideanRelativeInterior (n + m) D =
directSumSetEuclidean n m Set.univ (euclideanRelativeInterior m C) ∧
closure D = directSumSetEuclidean n m Set.univ (closure C)
A preimage point gives a graph point in the relative interior of the direct-sum set.
open scoped Pointwise
lemma graph_inter_ri_directSum_nonempty (n m : Nat)
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m))
(C : Set (EuclideanSpace Real (Fin m)))
(hri : (A ⁻¹' euclideanRelativeInterior m C).Nonempty) :
let e := appendAffineEquiv n m
let M : AffineSubspace Real (EuclideanSpace Real (Fin (n + m))) :=
(A.graph).toAffineSubspace.map e.toAffineMap
let D : Set (EuclideanSpace Real (Fin (n + m))) :=
directSumSetEuclidean n m Set.univ C
((M : Set _) ∩ euclideanRelativeInterior (n + m) D).Nonempty
Theorem 6.7: Let A be a linear transformation from R^n to R^m. Let C be a convex set in R^m such that A ⁻¹' (ri C) is nonempty. Then ri (A ⁻¹' C) = A ⁻¹' (ri C) and cl (A ⁻¹' C) = A ⁻¹' (cl C).
open scoped Pointwise
theorem euclideanRelativeInterior_preimage_linearMap_eq_and_closure_preimage (n m : Nat)
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m))
(C : Set (EuclideanSpace Real (Fin m))) (hC : Convex Real C)
(hri : (A ⁻¹' euclideanRelativeInterior m C).Nonempty) :
euclideanRelativeInterior n (A ⁻¹' C) = A ⁻¹' euclideanRelativeInterior m C ∧
closure (A ⁻¹' C) = A ⁻¹' closure C
The projection from R^{m+p} onto R^m identifies the section domain.
open scoped Pointwise
lemma section_D_eq_image_projection (m p : Nat)
(C : Set (EuclideanSpace Real (Fin (m + p))))
(y : EuclideanSpace Real (Fin m)) (z : EuclideanSpace Real (Fin p)) :
let e := appendAffineEquiv m p
let P : EuclideanSpace Real (Fin (m + p)) →ₗ[Real] EuclideanSpace Real (Fin m) :=
(LinearMap.fst (R := Real) (M := EuclideanSpace Real (Fin m))
(M₂ := EuclideanSpace Real (Fin p))).comp e.symm.linear.toLinearMap
let Cy : EuclideanSpace Real (Fin m) → Set (EuclideanSpace Real (Fin p)) :=
fun y => {z | e (y, z) ∈ C}
let D : Set (EuclideanSpace Real (Fin m)) := {y | (Cy y).Nonempty}
D = P '' C ∧ P (e (y, z)) = y
The projection fiber is an affine subspace, and its intersection with C is the section.
open scoped Pointwise
lemma section_fiber_affineSubspace_eq (m p : Nat)
(C : Set (EuclideanSpace Real (Fin (m + p))))
(y : EuclideanSpace Real (Fin m)) :
let e := appendAffineEquiv m p
let P : EuclideanSpace Real (Fin (m + p)) →ₗ[Real] EuclideanSpace Real (Fin m) :=
(LinearMap.fst (R := Real) (M := EuclideanSpace Real (Fin m))
(M₂ := EuclideanSpace Real (Fin p))).comp e.symm.linear.toLinearMap
let Cy : EuclideanSpace Real (Fin m) → Set (EuclideanSpace Real (Fin p)) :=
fun y => {z | e (y, z) ∈ C}
let M : AffineSubspace Real (EuclideanSpace Real (Fin (m + p))) :=
AffineSubspace.mk' (e (y, 0)) (LinearMap.ker P)
(M : Set _) = {w | P w = y} ∧
(M : Set _) ∩ C = (fun z => e (y, z)) '' (Cy y)
Relative interior of a section {y} ⊕ C_y is the image of ri C_y.
open scoped Pointwise
lemma ri_fiber_eq_ri_section (m p : Nat)
(C : Set (EuclideanSpace Real (Fin (m + p)))) (hC : Convex Real C)
(y : EuclideanSpace Real (Fin m)) :
let e := appendAffineEquiv m p
let Cy : EuclideanSpace Real (Fin m) → Set (EuclideanSpace Real (Fin p)) :=
fun y => {z | e (y, z) ∈ C}
euclideanRelativeInterior (m + p) (directSumSetEuclidean m p ({y}) (Cy y)) =
(fun z => e (y, z)) '' euclideanRelativeInterior p (Cy y)
Theorem 6.8: Let C be a convex set in R^{m+p}. For each y ∈ R^m, let C_y be the set of vectors z ∈ R^p such that (y, z) ∈ C. Let D = {y | C_y ≠ ∅}. Then (y, z) ∈ ri C if and only if y ∈ ri D and z ∈ ri C_y.
open scoped Pointwise
theorem euclideanRelativeInterior_mem_iff_relativeInterior_section (m p : Nat)
(C : Set (EuclideanSpace Real (Fin (m + p)))) (hC : Convex Real C)
(y : EuclideanSpace Real (Fin m)) (z : EuclideanSpace Real (Fin p)) :
let append :
EuclideanSpace Real (Fin m) → EuclideanSpace Real (Fin p) →
EuclideanSpace Real (Fin (m + p)) :=
fun y z =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (m + p))).symm
((Fin.appendIsometry m p)
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin m)) y,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin p)) z))
let Cy : EuclideanSpace Real (Fin m) → Set (EuclideanSpace Real (Fin p)) :=
fun y => {z | append y z ∈ C}
let D : Set (EuclideanSpace Real (Fin m)) := {y | (Cy y).Nonempty}
append y z ∈ euclideanRelativeInterior (m + p) C ↔
y ∈ euclideanRelativeInterior m D ∧ z ∈ euclideanRelativeInterior p (Cy y)
Convexity of the slice {v | first v = 1 ∧ tail v ∈ C}.
open scoped Pointwise
lemma convex_S_first_tail (n : Nat) (C : Set (EuclideanSpace Real (Fin n)))
(hC : Convex Real C) :
Convex Real
{v : EuclideanSpace Real (Fin (1 + n)) |
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n)) v) 0 = 1 ∧
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n)) v) (Fin.natAdd 1 i)) ∈
C}
Membership in the generated cone in terms of the first and tail coordinates.
open scoped Pointwise
lemma mem_K_iff_first_tail (n : Nat) (C : Set (EuclideanSpace Real (Fin n)))
(hC : Convex Real C) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first v = 1 ∧ tail v ∈ C}
let K : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S : Set (EuclideanSpace Real (Fin (1 + n))))
∀ v, v ∈ K ↔ 0 < first v ∧ tail v ∈ (first v) • C
First and tail coordinates of the append map for m = 1.
open scoped Pointwise
lemma first_tail_append (n : Nat) (y : EuclideanSpace Real (Fin 1))
(z : EuclideanSpace Real (Fin n)) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let append :
EuclideanSpace Real (Fin 1) → EuclideanSpace Real (Fin n) →
EuclideanSpace Real (Fin (1 + n)) :=
fun y z =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))).symm
((Fin.appendIsometry 1 n)
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)) y,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)) z))
first (append y z) = (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1) y) 0 ∧
tail (append y z) = z§6 — Relative Interiors of Convex Sets — Part 6
Lean authors: Zichen Wang, Wanli Ma, Yunxi Duan, Zaiwen Wen.
Corollary 6.8.1: Let C be a non-empty convex set in R^n, and let K be the convex cone in R^{n+1} generated by {(1, x) | x ∈ C}. Then ri K consists of the pairs (λ, x) such that λ > 0 and x ∈ λ ri C.
open scoped Pointwise
theorem euclideanRelativeInterior_convexConeGenerated_eq (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (hC : Convex Real C) (hCne : C.Nonempty) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first v = 1 ∧ tail v ∈ C}
let K : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S : Set (EuclideanSpace Real (Fin (1 + n))))
euclideanRelativeInterior (1 + n) K =
{v | 0 < first v ∧ tail v ∈ (first v) • euclideanRelativeInterior n C}§6 — Relative Interiors of Convex Sets — Part 7
Lean authors: Zichen Wang, Wanli Ma, Yunxi Duan, Zaiwen Wen.
Lifting a point into the cone over C preserves relative interior membership.
open scoped Pointwise
lemma lift_mem_ri_cone_iff (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (hC : Convex Real C) (hCne : C.Nonempty) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first v = 1 ∧ tail v ∈ C}
let K : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S : Set (EuclideanSpace Real (Fin (1 + n))))
let y0 : EuclideanSpace Real (Fin 1) :=
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm (fun _ => 1)
let append :
EuclideanSpace Real (Fin 1) → EuclideanSpace Real (Fin n) →
EuclideanSpace Real (Fin (1 + n)) :=
fun y z =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))).symm
((Fin.appendIsometry 1 n)
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)) y,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)) z))
let lift : EuclideanSpace Real (Fin n) → EuclideanSpace Real (Fin (1 + n)) :=
fun x => append y0 x
∀ x, x ∈ euclideanRelativeInterior n C ↔ lift x ∈ euclideanRelativeInterior (1 + n) K
A finite Minkowski sum of convex sets in Euclidean space is convex.
open scoped Pointwise
lemma convex_sum_finset_set_euclidean {n : ℕ} {ι : Type*} (s : Finset ι)
(A : ι → Set (EuclideanSpace Real (Fin n))) (hA : ∀ i ∈ s, Convex Real (A i)) :
Convex Real (s.sum A)
Choose indices witnessing membership in a union of sets (Euclidean space).
open scoped Pointwise
lemma choose_index_family_of_mem_iUnion_euclidean {n : ℕ} {ι : Type*} {I : Type*}
{C : I → Set (EuclideanSpace Real (Fin n))}
{x : ι → EuclideanSpace Real (Fin n)} (hx : ∀ i, x i ∈ ⋃ j, C j) :
∃ f : ι → I, ∀ i, x i ∈ C (f i)
Relative interior of a finite Minkowski sum of convex sets.
open scoped Pointwise
lemma ri_sum_cones_eq_sum_ri (n m : Nat)
(K : Fin (Nat.succ m) → Set (EuclideanSpace Real (Fin n)))
(hK : ∀ i, Convex Real (K i)) :
euclideanRelativeInterior n (∑ i, K i) =
∑ i, euclideanRelativeInterior n (K i)
A vector in R^{n+1} is determined by its first and tail coordinates.
open scoped Pointwise
lemma eq_of_first_tail_eq (n : Nat) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
∀ u v, first u = first v → tail u = tail v → u = v
The first coordinate scales linearly under scalar multiplication.
open scoped Pointwise
lemma first_smul (n : Nat) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
∀ (a : Real) v, first (a • v) = a * first v
The tail map scales linearly under scalar multiplication.
open scoped Pointwise
lemma tail_smul (n : Nat) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
∀ (a : Real) v, tail (a • v) = a • tail v
The lift map preserves distances.
open scoped Pointwise
lemma dist_lift_eq (n : Nat) :
let y0 : EuclideanSpace Real (Fin 1) :=
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm (fun _ => 1)
let append :
EuclideanSpace Real (Fin 1) → EuclideanSpace Real (Fin n) →
EuclideanSpace Real (Fin (1 + n)) :=
fun y z =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))).symm
((Fin.appendIsometry 1 n)
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)) y,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)) z))
let lift : EuclideanSpace Real (Fin n) → EuclideanSpace Real (Fin (1 + n)) :=
fun x => append y0 x
∀ x y, dist (lift x) (lift y) = dist x y
Weighted sums of lifts with total weight 1 collapse to a single lift.
open scoped Pointwise
lemma sum_smul_lift_eq_lift_sum (n m : Nat)
(w : Fin (Nat.succ m) → Real)
(x_i : Fin (Nat.succ m) → EuclideanSpace Real (Fin n))
(hsum : Finset.univ.sum (fun i => w i) = 1) :
let y0 : EuclideanSpace Real (Fin 1) :=
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm (fun _ => 1)
let append :
EuclideanSpace Real (Fin 1) → EuclideanSpace Real (Fin n) →
EuclideanSpace Real (Fin (1 + n)) :=
fun y z =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))).symm
((Fin.appendIsometry 1 n)
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)) y,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)) z))
let lift : EuclideanSpace Real (Fin n) → EuclideanSpace Real (Fin (1 + n)) :=
fun x => append y0 x
Finset.univ.sum (fun i => w i • lift (x_i i)) =
lift (Finset.univ.sum (fun i => w i • x_i i))
A norm bound for the perturbation of weights in a weighted sum.
open scoped Pointwise
lemma weighted_sum_perturb_bound (n m : Nat)
(w : Fin (Nat.succ m) → Real)
(x_i : Fin (Nat.succ m) → EuclideanSpace Real (Fin n))
{δ : Real} (hδ : 0 ≤ δ) :
let c : Real := 1 + (Nat.succ m : Real) * δ
let w' : Fin (Nat.succ m) → Real := fun i => (w i + δ) / c
‖Finset.univ.sum (fun i => w' i • x_i i) -
Finset.univ.sum (fun i => w i • x_i i)‖ ≤
(δ / c) *
(‖Finset.univ.sum (fun i => x_i i)‖ +
(Nat.succ m : Real) * ‖Finset.univ.sum (fun i => w i • x_i i)‖)
Membership in the sum of the relative interiors of the lifted cones gives positive weights.
open scoped Pointwise
lemma mem_sum_ri_cones_iff_weights (n m : Nat)
(C : Fin (Nat.succ m) → Set (EuclideanSpace Real (Fin n)))
(hCne : ∀ i, (C i).Nonempty) (hCconv : ∀ i, Convex Real (C i)) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S : Fin (Nat.succ m) → Set (EuclideanSpace Real (Fin (1 + n))) :=
fun i => {v | first v = 1 ∧ tail v ∈ C i}
let K : Fin (Nat.succ m) → Set (EuclideanSpace Real (Fin (1 + n))) :=
fun i => (ConvexCone.hull Real (S i) : Set (EuclideanSpace Real (Fin (1 + n))))
∀ v, first v = 1 →
(v ∈ ∑ i, euclideanRelativeInterior (1 + n) (K i) ↔
∃ (w : Fin (Nat.succ m) → Real) (x_i : Fin (Nat.succ m) → EuclideanSpace Real (Fin n)),
(∀ i, 0 < w i) ∧ (Finset.univ.sum (fun i => w i) = 1) ∧
(∀ i, x_i i ∈ euclideanRelativeInterior n (C i)) ∧
tail v = Finset.univ.sum (fun i => w i • x_i i))
A point in the convex hull of a finite union of convex sets admits one point per set.
open scoped Pointwise
lemma convexHull_iUnion_exists_weights_points (n m : Nat)
(C : Fin (Nat.succ m) → Set (EuclideanSpace Real (Fin n)))
(hCne : ∀ i, (C i).Nonempty) (hCconv : ∀ i, Convex Real (C i)) :
∀ x ∈ convexHull Real (⋃ i, C i),
∃ (w : Fin (Nat.succ m) → Real) (x_i : Fin (Nat.succ m) → EuclideanSpace Real (Fin n)),
(∀ i, 0 ≤ w i) ∧ (Finset.univ.sum (fun i => w i) = 1) ∧
(∀ i, x_i i ∈ C i) ∧ x = Finset.univ.sum (fun i => w i • x_i i)§6 — Relative Interiors of Convex Sets — Part 8
Lean authors: Zichen Wang, Wanli Ma, Yunxi Duan, Zaiwen Wen.
The Minkowski sum of the lifted cones lies in the cone over the convex hull.
open scoped Pointwise
lemma sum_cones_subset_K0 (n m : Nat)
(C : Fin (Nat.succ m) → Set (EuclideanSpace Real (Fin n))) :
let C0 : Set (EuclideanSpace Real (Fin n)) := convexHull Real (⋃ i, C i)
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S0 : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first v = 1 ∧ tail v ∈ C0}
let K0 : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S0 : Set (EuclideanSpace Real (Fin (1 + n))))
let S : Fin (Nat.succ m) → Set (EuclideanSpace Real (Fin (1 + n))) :=
fun i => {v | first v = 1 ∧ tail v ∈ C i}
let K : Fin (Nat.succ m) → Set (EuclideanSpace Real (Fin (1 + n))) :=
fun i => (ConvexCone.hull Real (S i) : Set (EuclideanSpace Real (Fin (1 + n))))
(∑ i, K i) ⊆ K0
Base points of the cone over C0 lie in the closure of the sum of cones.
open scoped Pointwise
lemma S0_subset_closure_sum_cones (n m : Nat)
(C : Fin (Nat.succ m) → Set (EuclideanSpace Real (Fin n)))
(hCne : ∀ i, (C i).Nonempty) (hCconv : ∀ i, Convex Real (C i)) :
let C0 : Set (EuclideanSpace Real (Fin n)) := convexHull Real (⋃ i, C i)
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S0 : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first v = 1 ∧ tail v ∈ C0}
let S : Fin (Nat.succ m) → Set (EuclideanSpace Real (Fin (1 + n))) :=
fun i => {v | first v = 1 ∧ tail v ∈ C i}
let K : Fin (Nat.succ m) → Set (EuclideanSpace Real (Fin (1 + n))) :=
fun i => (ConvexCone.hull Real (S i) : Set (EuclideanSpace Real (Fin (1 + n))))
S0 ⊆ closure (∑ i, K i)
The lifted cone over C0 lies in the closure of the sum of cones.
open scoped Pointwise
lemma K0_subset_closure_sum_cones (n m : Nat)
(C : Fin (Nat.succ m) → Set (EuclideanSpace Real (Fin n)))
(hCne : ∀ i, (C i).Nonempty) (hCconv : ∀ i, Convex Real (C i)) :
let C0 : Set (EuclideanSpace Real (Fin n)) := convexHull Real (⋃ i, C i)
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S0 : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first v = 1 ∧ tail v ∈ C0}
let K0 : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S0 : Set (EuclideanSpace Real (Fin (1 + n))))
let S : Fin (Nat.succ m) → Set (EuclideanSpace Real (Fin (1 + n))) :=
fun i => {v | first v = 1 ∧ tail v ∈ C i}
let K : Fin (Nat.succ m) → Set (EuclideanSpace Real (Fin (1 + n))) :=
fun i => (ConvexCone.hull Real (S i) : Set (EuclideanSpace Real (Fin (1 + n))))
K0 ⊆ closure (∑ i, K i)
Theorem 6.9: Let C_1, ..., C_m be non-empty convex sets in R^n, and let C_0 = conv (C_1 ∪ ... ∪ C_m). Then ri C_0 = ⋃ {λ_1 ri C_1 + ... + λ_m ri C_m | λ_i > 0, λ_1 + ... + λ_m = 1}.
open scoped Pointwise
theorem euclideanRelativeInterior_convexHull_iUnion_eq (n m : Nat)
(C : Fin m → Set (EuclideanSpace Real (Fin n))) (hCne : ∀ i, (C i).Nonempty)
(hCconv : ∀ i, Convex Real (C i)) :
euclideanRelativeInterior n (convexHull Real (⋃ i, C i)) =
{x | ∃ (w : Fin m → Real) (x_i : Fin m → EuclideanSpace Real (Fin n)),
(∀ i, 0 < w i) ∧ (Finset.univ.sum (fun i => w i) = 1) ∧
(∀ i, x_i i ∈ euclideanRelativeInterior n (C i)) ∧
x = Finset.univ.sum (fun i => w i • x_i i)}
The tail projection agrees with the linear map obtained from appendAffineEquiv.
open scoped Pointwise
lemma tail_linearMap_apply (n : Nat) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let A : EuclideanSpace Real (Fin (1 + n)) →ₗ[Real] EuclideanSpace Real (Fin n) :=
(LinearMap.snd (R := Real) (M := EuclideanSpace Real (Fin 1))
(M₂ := EuclideanSpace Real (Fin n))).comp
(appendAffineEquiv 1 n).symm.linear.toLinearMap
∀ v, A v = tail v
The tail projection of the lifted cone equals the cone generated by C.
open scoped Pointwise
lemma tail_image_cone_eq_convexCone_hull (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (hC : Convex Real C) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first v = 1 ∧ tail v ∈ C}
let K : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S : Set (EuclideanSpace Real (Fin (1 + n))))
let A : EuclideanSpace Real (Fin (1 + n)) →ₗ[Real] EuclideanSpace Real (Fin n) :=
(LinearMap.snd (R := Real) (M := EuclideanSpace Real (Fin 1))
(M₂ := EuclideanSpace Real (Fin n))).comp
(appendAffineEquiv 1 n).symm.linear.toLinearMap
A '' K = (ConvexCone.hull Real C : Set (EuclideanSpace Real (Fin n)))
The tail projection of the relative interior gives positive scalings of ri C.
open scoped Pointwise
lemma tail_image_ri_cone_eq_smul_ri (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (hC : Convex Real C) (hCne : C.Nonempty) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first v = 1 ∧ tail v ∈ C}
let K : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S : Set (EuclideanSpace Real (Fin (1 + n))))
let A : EuclideanSpace Real (Fin (1 + n)) →ₗ[Real] EuclideanSpace Real (Fin n) :=
(LinearMap.snd (R := Real) (M := EuclideanSpace Real (Fin 1))
(M₂ := EuclideanSpace Real (Fin n))).comp
(appendAffineEquiv 1 n).symm.linear.toLinearMap
A '' euclideanRelativeInterior (1 + n) K =
{v | ∃ r : Real, 0 < r ∧ ∃ x : EuclideanSpace Real (Fin n),
x ∈ euclideanRelativeInterior n C ∧ v = r • x}
Text 6.19: More generally, the relative interior of the convex cone in R^n generated by a non-empty convex set C consists of the vectors of the form λ x with λ > 0 and x ∈ ri C.
open scoped Pointwise
theorem euclideanRelativeInterior_convexConeGenerated_eq_smul (n : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (hC : Convex Real C) (hCne : C.Nonempty) :
let K : Set (EuclideanSpace Real (Fin n)) :=
(ConvexCone.hull Real C : Set (EuclideanSpace Real (Fin n)))
euclideanRelativeInterior n K =
{v | ∃ r : Real, 0 < r ∧ ∃ x : EuclideanSpace Real (Fin n),
x ∈ euclideanRelativeInterior n C ∧ v = r • x}
A convex cone is invariant under positive scalar multiplication.
open scoped Pointwise
lemma smul_convexCone_eq_self (n : Nat)
(K : ConvexCone Real (EuclideanSpace Real (Fin n))) (r : Real) (hr : 0 < r) :
r • (K : Set (EuclideanSpace Real (Fin n))) = (K : Set (EuclideanSpace Real (Fin n)))
A convex set invariant under positive scalings is a convex cone.
open scoped Pointwise
lemma convexCone_of_convex_smul_invariant (n : Nat) (C : Set (EuclideanSpace Real (Fin n)))
(hC : Convex Real C) (hsmul : ∀ r : Real, 0 < r → r • C = C) :
∃ K' : ConvexCone Real (EuclideanSpace Real (Fin n)), (K' : Set _) = C
The relative interior of a convex cone is invariant under positive scalings.
open scoped Pointwise
lemma ri_smul_invariant_of_cone (n : Nat)
(K : ConvexCone Real (EuclideanSpace Real (Fin n))) (r : Real) (hr : 0 < r) :
r • euclideanRelativeInterior n (K : Set (EuclideanSpace Real (Fin n))) =
euclideanRelativeInterior n (K : Set (EuclideanSpace Real (Fin n)))
Text 6.20: The relative interior and the closure of a convex cone are convex cones.
open scoped Pointwise
theorem euclideanRelativeInterior_and_closure_convexCone (n : Nat)
(K : ConvexCone Real (EuclideanSpace Real (Fin n))) :
(∃ K' : ConvexCone Real (EuclideanSpace Real (Fin n)),
(K' : Set (EuclideanSpace Real (Fin n))) =
euclideanRelativeInterior n (K : Set (EuclideanSpace Real (Fin n)))) ∧
(∃ K' : ConvexCone Real (EuclideanSpace Real (Fin n)),
(K' : Set (EuclideanSpace Real (Fin n))) =
closure (K : Set (EuclideanSpace Real (Fin n))))§7 — Closures of Convex Functions — Part 1
Lean authors: Zichen Wang, Wanli Ma, Xinyi Guo, Zaiwen Wen.
Text 7.0.1: Let f : S → [-∞, +∞] with S ⊆ R^n and x ∈ S. Then f is lower semicontinuous at x if for every sequence x_i ⊂ S with x_i → x, f x ≤ liminf (fun i => f (x_i)). Equivalently, f x ≤ liminf_{y → x} f y.
open scoped Topology
def LowerSemicontinuousAtSeq {n : Nat} (S : Set (EuclideanSpace Real (Fin n)))
(f : S → EReal) (x : S) : Prop :=
∀ u : ℕ → S,
Filter.Tendsto (fun i => (u i : EuclideanSpace Real (Fin n))) Filter.atTop
(𝓝 (x : EuclideanSpace Real (Fin n))) →
f x ≤ Filter.liminf (fun i => f (u i)) Filter.atTop
A frequently occurring subset near x yields a sequence in that subset converging to x.
open scoped Topology
lemma exists_seq_tendsto_of_frequently {n : Nat}
{S : Set (EuclideanSpace Real (Fin n))} {x : S} {s : Set S}
(h : ∃ᶠ z in 𝓝 x, z ∈ s) :
∃ u : ℕ → S, (∀ n, u n ∈ s) ∧
Filter.Tendsto (fun i => (u i : EuclideanSpace Real (Fin n))) Filter.atTop
(𝓝 (x : EuclideanSpace Real (Fin n)))
Lower semicontinuity implies the sequential liminf inequality.
open scoped Topology
lemma lowerSemicontinuousAtSeq_of_lowerSemicontinuousAt {n : Nat}
(S : Set (EuclideanSpace Real (Fin n))) (f : S → EReal) (x : S) :
LowerSemicontinuousAt f x → LowerSemicontinuousAtSeq S f x
Sequential lower semicontinuity implies filter lower semicontinuity.
open scoped Topology
lemma lowerSemicontinuousAt_of_lowerSemicontinuousAtSeq {n : Nat}
(S : Set (EuclideanSpace Real (Fin n))) (f : S → EReal) (x : S) :
LowerSemicontinuousAtSeq S f x → LowerSemicontinuousAt f x
The sequential definition agrees with mathlib's LowerSemicontinuousAt on the subtype.
open scoped Topology
theorem lowerSemicontinuousAtSeq_iff_lowerSemicontinuousAt {n : Nat}
(S : Set (EuclideanSpace Real (Fin n))) (f : S → EReal) (x : S) :
LowerSemicontinuousAtSeq S f x ↔ LowerSemicontinuousAt f x
Text 7.0.2: With f, S, x as above, f is upper semicontinuous at x if for every sequence x_i ⊂ S with x_i → x, f x ≥ limsup (fun i => f (x_i)). Equivalently, f x ≥ limsup_{y → x} f y.
open scoped Topology
def UpperSemicontinuousAtSeq {n : Nat} (S : Set (EuclideanSpace Real (Fin n)))
(f : S → EReal) (x : S) : Prop :=
∀ u : ℕ → S,
Filter.Tendsto (fun i => (u i : EuclideanSpace Real (Fin n))) Filter.atTop
(𝓝 (x : EuclideanSpace Real (Fin n))) →
f x ≥ Filter.limsup (fun i => f (u i)) Filter.atTop
Upper semicontinuity implies the sequential limsup inequality.
open scoped Topology
lemma upperSemicontinuousAtSeq_of_upperSemicontinuousAt {n : Nat}
(S : Set (EuclideanSpace Real (Fin n))) (f : S → EReal) (x : S) :
UpperSemicontinuousAt f x → UpperSemicontinuousAtSeq S f x
Sequential upper semicontinuity implies filter upper semicontinuity.
open scoped Topology
lemma upperSemicontinuousAt_of_upperSemicontinuousAtSeq {n : Nat}
(S : Set (EuclideanSpace Real (Fin n))) (f : S → EReal) (x : S) :
UpperSemicontinuousAtSeq S f x → UpperSemicontinuousAt f x
The sequential definition agrees with mathlib's UpperSemicontinuousAt on the subtype.
open scoped Topology
theorem upperSemicontinuousAtSeq_iff_upperSemicontinuousAt {n : Nat}
(S : Set (EuclideanSpace Real (Fin n))) (f : S → EReal) (x : S) :
UpperSemicontinuousAtSeq S f x ↔ UpperSemicontinuousAt f x
Lemma 7.0.3. Let S be a subset of R^n, f : S → [-infty, +infty], and x in S. The following are equivalent: (i) for every sequence x_i in S with x_i -> x, f x >= limsup (fun i => f (x_i)); (ii) f x >= limsup_{y -> x} f y, where limsup_{y -> x} f y := lim_{epsilon -> 0} sup { f y | y in S, |y - x| < epsilon }.
open scoped Topology
lemma upperSemicontinuousAtSeq_iff_limsup_nhds {n : Nat}
(S : Set (EuclideanSpace Real (Fin n))) (f : S → EReal) (x : S) :
UpperSemicontinuousAtSeq S f x ↔
f x ≥ Filter.limsup (fun y : S => f y) (𝓝 x)
A non-bottom EReal exceeds some real number.
open scoped Topology
lemma exists_real_not_le_of_ne_bot {x : EReal} (hx : x ≠ (⊥ : EReal)) :
∃ α : Real, ¬ x ≤ (α : EReal)
Real sublevel sets control the bottom sublevel set.
open scoped Topology
lemma isClosed_sublevel_bot_of_closed_sublevels {n : Nat}
{f : (Fin n → Real) → EReal}
(h : ∀ α : Real, IsClosed {x | f x ≤ (α : EReal)}) :
IsClosed {x | f x ≤ (⊥ : EReal)}
Lower semicontinuity is equivalent to closed real sublevel sets.
open scoped Topology
lemma lowerSemicontinuous_iff_closed_sublevel {n : Nat} (f : (Fin n → Real) → EReal) :
LowerSemicontinuous f ↔ ∀ α : Real, IsClosed {x | f x ≤ (α : EReal)}
Closedness of the epigraph implies closedness of every real sublevel set.
open scoped Topology
lemma closed_sublevel_of_closed_epigraph {n : Nat} {f : (Fin n → Real) → EReal}
(h : IsClosed (epigraph (S := (Set.univ : Set (Fin n → Real))) f)) :
∀ α : Real, IsClosed {x | f x ≤ (α : EReal)}
Closed real sublevel sets imply closedness of the epigraph.
open scoped Topology
lemma closed_epigraph_of_closed_sublevel {n : Nat} {f : (Fin n → Real) → EReal}
(h : ∀ α : Real, IsClosed {x | f x ≤ (α : EReal)}) :
IsClosed (epigraph (S := (Set.univ : Set (Fin n → Real))) f)
Theorem 7.1. Let f : ℝ^n → [-∞, +∞]. The following are equivalent: (a) f is lower semicontinuous on ℝ^n; (b) {x | f x ≤ α} is closed for every α ∈ ℝ; (c) the epigraph of f is closed in ℝ^{n+1}.
open scoped Topology
theorem lowerSemicontinuous_iff_closed_sublevel_iff_closed_epigraph {n : Nat}
(f : (Fin n → Real) → EReal) :
((LowerSemicontinuous f ↔ ∀ α : Real, IsClosed {x | f x ≤ (α : EReal)}) ∧
((∀ α : Real, IsClosed {x | f x ≤ (α : EReal)}) ↔
IsClosed (epigraph (S := (Set.univ : Set (Fin n → Real))) f)))
Text 7.0.4: Given any function f on ℝ^n, there exists a greatest lower semicontinuous function majorized by f; this function is called the lower semicontinuous hull of f. The pointwise supremum of lower semicontinuous minorants of f is lower semicontinuous.
open scoped Topology
lemma lscHull_candidate_lsc {n : Nat} (f : (Fin n → Real) → EReal) :
LowerSemicontinuous (fun x =>
⨆ h :
{h : (Fin n → Real) → EReal // LowerSemicontinuous h ∧ h ≤ f}, h.1 x)
The pointwise supremum of lower semicontinuous minorants of f lies below f.
open scoped Topology
lemma lscHull_candidate_le {n : Nat} (f : (Fin n → Real) → EReal) :
(fun x =>
⨆ h :
{h : (Fin n → Real) → EReal // LowerSemicontinuous h ∧ h ≤ f}, h.1 x) ≤ f
Any lower semicontinuous h ≤ f is bounded above by the candidate hull.
open scoped Topology
lemma lscHull_candidate_max {n : Nat} {f : (Fin n → Real) → EReal}
{h : (Fin n → Real) → EReal} (hh : LowerSemicontinuous h) (hle : h ≤ f) :
h ≤
(fun x =>
⨆ h' :
{h : (Fin n → Real) → EReal // LowerSemicontinuous h ∧ h ≤ f}, h'.1 x)
open scoped Topology
theorem exists_lowerSemicontinuousHull {n : Nat} (f : (Fin n → Real) → EReal) :
∃ g : (Fin n → Real) → EReal,
LowerSemicontinuous g ∧ g ≤ f ∧
∀ h : (Fin n → Real) → EReal, LowerSemicontinuous h → h ≤ f → h ≤ g
Text 7.0.4: The lower semicontinuous hull of f.
open scoped Topology
noncomputable def lowerSemicontinuousHull {n : Nat} (f : (Fin n → Real) → EReal) :
(Fin n → Real) → EReal
Text 7.0.5: The closure of a convex function f is the lower semicontinuous hull when f is never -∞, whereas it is the constant function -∞ when f is an improper convex function with f x = -∞ for some x.
open scoped Topology
noncomputable def convexFunctionClosure {n : Nat} (f : (Fin n → Real) → EReal) :
(Fin n → Real) → EReal
Text 7.0.6: A convex function is called closed if it is lower semicontinuous on ℝ^n.
open scoped Topology
def ClosedConvexFunction {n : Nat} (f : (Fin n → Real) → EReal) : Prop :=
ConvexFunction f ∧ LowerSemicontinuous f
Text 7.0.6: A proper convex function is closed iff it is lower semicontinuous.
open scoped Topology
theorem properConvexFunction_closed_iff_lowerSemicontinuous {n : Nat}
{f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
ClosedConvexFunction f ↔ LowerSemicontinuous f
The constant ⊤ function is closed and improper.
open scoped Topology
lemma closed_improper_const_top {n : Nat} :
ClosedConvexFunction (fun _ : (Fin n → Real) => (⊤ : EReal)) ∧
ImproperConvexFunctionOn (Set.univ : Set (Fin n → Real))
(fun _ : (Fin n → Real) => (⊤ : EReal))
The constant ⊥ function is closed and improper.
open scoped Topology
lemma closed_improper_const_bot {n : Nat} :
ClosedConvexFunction (fun _ : (Fin n → Real) => (⊥ : EReal)) ∧
ImproperConvexFunctionOn (Set.univ : Set (Fin n → Real))
(fun _ : (Fin n → Real) => (⊥ : EReal))
An improper convex function has empty epigraph or attains ⊥ on the set.
open scoped Topology
lemma improperConvexFunctionOn_cases_epigraph_empty_or_bot {n : Nat}
{S : Set (Fin n → Real)} {f : (Fin n → Real) → EReal}
(h : ImproperConvexFunctionOn S f) :
(¬ Set.Nonempty (epigraph S f)) ∨ ∃ x ∈ S, f x = (⊥ : EReal)
If the epigraph is empty, then the function is identically ⊤ on S.
open scoped Topology
lemma epigraph_empty_imp_forall_top {n : Nat} {S : Set (Fin n → Real)}
{f : (Fin n → Real) → EReal} (h : ¬ Set.Nonempty (epigraph S f)) :
∀ x ∈ S, f x = (⊤ : EReal)
If f attains ⊥, then its closure is the constant ⊥ function.
open scoped Topology
lemma convexFunctionClosure_eq_bot_of_exists_bot {n : Nat}
{f : (Fin n → Real) → EReal} (h : ∃ x, f x = (⊥ : EReal)) :
convexFunctionClosure f = (fun _ => (⊥ : EReal))
The closure of the constant ⊤ function is itself.
open scoped Topology
lemma convexFunctionClosure_const_top {n : Nat} :
convexFunctionClosure (fun _ : (Fin n → Real) => (⊤ : EReal)) =
(fun _ : (Fin n → Real) => (⊤ : EReal))
open scoped Topology
theorem closed_improperConvexFunction_eq_top_or_bot {n : Nat}
{f : (Fin n → Real) → EReal} :
(convexFunctionClosure f = f ∧
ImproperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) ↔
f = (fun _ => (⊤ : EReal)) ∨ f = (fun _ => (⊥ : EReal))
Convexity of the reciprocal on (0, ∞) extended by ⊤ outside.
open scoped Topology
lemma convexFunctionOn_inv_pos_aux :
ConvexFunctionOn Set.univ (fun x : Fin 1 → Real =>
if 0 < x 0 then ((1 / x 0 : Real) : EReal) else (⊤ : EReal))
Lower semicontinuity of the reciprocal on (0, ∞) extended by ⊤ outside.
open scoped Topology
lemma lowerSemicontinuous_inv_pos_aux :
LowerSemicontinuous (fun x : Fin 1 → Real =>
if 0 < x 0 then ((1 / x 0 : Real) : EReal) else (⊤ : EReal))
The effective domain of the reciprocal example is the positive half-line.
open scoped Topology
lemma effectiveDomain_inv_pos_eq :
effectiveDomain (Set.univ : Set (Fin 1 → Real))
(fun x : Fin 1 → Real =>
if 0 < x 0 then ((1 / x 0 : Real) : EReal) else (⊤ : EReal)) =
{x : Fin 1 → Real | 0 < x 0}
The positive half-line in Fin 1 → Real is not closed.
open scoped Topology
lemma not_isClosed_effectiveDomain_inv_pos :
¬ IsClosed {x : Fin 1 → Real | 0 < x 0}
Text 7.0.8: It can happen that a closed proper convex function has a domain that is not closed.
open scoped Topology
theorem exists_closed_proper_convexFunction_domain_not_closed :
∃ n : Nat, ∃ f : (Fin n → Real) → EReal,
ClosedConvexFunction f ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f ∧
¬ IsClosed (effectiveDomain (S := (Set.univ : Set (Fin n → Real))) f)
Text 7.0.9: For example, the function f(x) = 1/x for x > 0 and f(x) = +∞ for x ≤ 0 is closed.
open scoped Topology
theorem closedConvexFunction_inv_pos :
ClosedConvexFunction (fun x : (Fin 1 → Real) =>
if 0 < x 0 then ((1 / (x 0) : Real) : EReal) else (⊤ : EReal))
If g is lower semicontinuous and g ≤ f, then closure (epi f) ⊆ epi g.
open scoped Topology
lemma closure_epigraph_subset_epigraph_of_lsc_le {n : Nat}
{f g : (Fin n → Real) → EReal}
(hg : LowerSemicontinuous g) (hle : g ≤ f) :
closure (epigraph (S := (Set.univ : Set (Fin n → Real))) f) ⊆
epigraph (S := (Set.univ : Set (Fin n → Real))) g
A lower semicontinuous minorant lies below the liminf of its majorant.
open scoped Topology
lemma lowerSemicontinuous_le_liminf_of_le {n : Nat}
{f g : (Fin n → Real) → EReal}
(hg : LowerSemicontinuous g) (hle : g ≤ f) :
∀ x, g x ≤ Filter.liminf (fun y : Fin n → Real => f y) (𝓝 x)
The closure of an epigraph is upward closed in the second coordinate.
open scoped Topology
lemma closure_epigraph_upward {n : Nat} {f : (Fin n → Real) → EReal}
{x : Fin n → Real} {μ ν : Real} (hμν : μ ≤ ν) :
(x, μ) ∈ closure (epigraph (S := (Set.univ : Set (Fin n → Real))) f) →
(x, ν) ∈ closure (epigraph (S := (Set.univ : Set (Fin n → Real))) f)
The function obtained by taking the infimum of each vertical slice of the epigraph closure.
open scoped Topology
noncomputable def epigraphClosureInf {n : Nat} (f : (Fin n → Real) → EReal) :
(Fin n → Real) → EReal :=
fun x =>
sInf ((fun μ : Real => (μ : EReal)) '' {μ : Real |
(x, μ) ∈ closure (epigraph (S := (Set.univ : Set (Fin n → Real))) f)})
The closure of an epigraph is the epigraph of epigraphClosureInf.
open scoped Topology
lemma closure_epigraph_eq_epigraph_sInf {n : Nat} (f : (Fin n → Real) → EReal) :
epigraph (S := (Set.univ : Set (Fin n → Real))) (epigraphClosureInf f) =
closure (epigraph (S := (Set.univ : Set (Fin n → Real))) f)
Points in the closure of the epigraph bound the liminf from above.
open scoped Topology
lemma liminf_le_of_mem_closure_epigraph {n : Nat} {f : (Fin n → Real) → EReal}
{x : Fin n → Real} {μ : Real}
(hmem :
(x, μ) ∈ closure (epigraph (S := (Set.univ : Set (Fin n → Real))) f)) :
Filter.liminf (fun y : Fin n → Real => f y) (𝓝 x) ≤ (μ : EReal)
Text 7.0.10: The epigraph of the closure of f is the closure of the epigraph of f. Consequently, (convexFunctionClosure f)(x) = liminf_{y → x} f(y).
open scoped Topology
theorem epigraph_convexFunctionClosure_eq_closure_epigraph {n : Nat}
(f : (Fin n → Real) → EReal) (hbot : ∀ x, f x ≠ (⊥ : EReal)) :
epigraph (S := (Set.univ : Set (Fin n → Real))) (convexFunctionClosure f) =
closure (epigraph (S := (Set.univ : Set (Fin n → Real))) f) ∧
∀ x : Fin n → Real,
convexFunctionClosure f x =
Filter.liminf (fun y : Fin n → Real => f y) (𝓝 x)
A liminf bound yields membership in all closed real sublevel sets above α.
open scoped Topology
lemma liminf_le_mem_iInter_closure_sublevel {n : Nat}
{f : (Fin n → Real) → EReal} {α : Real} {x : Fin n → Real} :
Filter.liminf (fun y : Fin n → Real => f y) (𝓝 x) ≤ (α : EReal) →
x ∈ ⋂ (μ : {μ : Real // μ > α}),
closure {y | f y ≤ ((μ : Real) : EReal)}
Membership in all closed real sublevel sets above α forces the liminf bound.
open scoped Topology
lemma liminf_le_of_mem_iInter_closure_sublevel {n : Nat}
{f : (Fin n → Real) → EReal} {α : Real} {x : Fin n → Real} :
x ∈ ⋂ (μ : {μ : Real // μ > α}),
closure {y | f y ≤ ((μ : Real) : EReal)} →
Filter.liminf (fun y : Fin n → Real => f y) (𝓝 x) ≤ (α : EReal)§7 — Closures of Convex Functions — Part 2
Lean authors: Zichen Wang, Wanli Ma, Xinyi Guo, Zaiwen Wen.
The α-sublevel of the liminf is the intersection of closed μ-sublevel sets above α.
open scoped Topology
lemma sublevel_liminf_eq_iInter_closure_sublevel {n : Nat}
(f : (Fin n → Real) → EReal) (α : Real) :
{x | Filter.liminf (fun y : Fin n → Real => f y) (𝓝 x) ≤ (α : EReal)} =
⋂ (μ : {μ : Real // μ > α}),
closure {x | f x ≤ ((μ : Real) : EReal)}
Text 7.0.11: For each α ∈ ℝ, {x | (convexFunctionClosure f) x ≤ α} = ⋂ (μ > α), closure {x | f x ≤ μ}.
open scoped Topology
theorem sublevel_convexFunctionClosure_eq_iInter_closure_sublevel {n : Nat}
(f : (Fin n → Real) → EReal) (α : Real) (hbot : ∀ x, f x ≠ (⊥ : EReal)) :
{x | convexFunctionClosure f x ≤ (α : EReal)} =
⋂ (μ : {μ : Real // μ > α}),
closure {x | f x ≤ ((μ : Real) : EReal)}
The closure of a function is pointwise below the function itself.
open scoped Topology
lemma convexFunctionClosure_le_self {n : Nat} (f : (Fin n → Real) → EReal) :
convexFunctionClosure f ≤ f
The closure operator is monotone with respect to the pointwise order.
open scoped Topology
lemma convexFunctionClosure_mono {n : Nat} {f1 f2 : (Fin n → Real) → EReal}
(h12 : f1 ≤ f2) : convexFunctionClosure f1 ≤ convexFunctionClosure f2
The infimum of a function equals the infimum of its closure.
open scoped Topology
lemma iInf_convexFunctionClosure_eq {n : Nat} (f : (Fin n → Real) → EReal) :
iInf (fun x => f x) = iInf (fun x => convexFunctionClosure f x)
Text 7.0.12: For any extended-real-valued function f : ℝ^n → [-∞, +∞], one has cl f ≤ f. Moreover, if f₁ ≤ f₂, then cl f₁ ≤ cl f₂. In addition, inf_{x ∈ ℝ^n} f x = inf_{x ∈ ℝ^n} (cl f) x. Here cl f is convexFunctionClosure f.
open scoped Topology
theorem convexFunctionClosure_properties {n : Nat} :
(∀ f : (Fin n → Real) → EReal, convexFunctionClosure f ≤ f) ∧
(∀ f1 f2 : (Fin n → Real) → EReal, f1 ≤ f2 →
convexFunctionClosure f1 ≤ convexFunctionClosure f2) ∧
(∀ f : (Fin n → Real) → EReal,
iInf (fun x => f x) = iInf (fun x => convexFunctionClosure f x))
Points with positive coordinate appear frequently near the origin.
open scoped Topology
lemma frequently_pos_coord_nhds_zero :
∃ᶠ y in 𝓝 (0 : Fin 1 → Real), 0 < y 0
At a point with positive coordinate, the liminf of the step function is 0.
open scoped Topology
lemma liminf_example_origin_pos {x : Fin 1 → Real} (hx : 0 < x 0) :
Filter.liminf (fun y : Fin 1 → Real =>
if 0 < y 0 then (0 : EReal) else (⊤ : EReal)) (𝓝 x) = (0 : EReal)
At a point with negative coordinate, the liminf of the step function is ⊤.
open scoped Topology
lemma liminf_example_origin_neg {x : Fin 1 → Real} (hx : x 0 < 0) :
Filter.liminf (fun y : Fin 1 → Real =>
if 0 < y 0 then (0 : EReal) else (⊤ : EReal)) (𝓝 x) = (⊤ : EReal)
At the origin, the liminf of the step function is 0.
open scoped Topology
lemma liminf_example_origin_zero :
Filter.liminf (fun y : Fin 1 → Real =>
if 0 < y 0 then (0 : EReal) else (⊤ : EReal)) (𝓝 (0 : Fin 1 → Real)) =
(0 : EReal)
Text 7.0.13: If f : ℝ → [-∞, +∞] is defined by f(x) = 0 for x > 0 and f(x) = +∞ for x ≤ 0, then cl f agrees with f except at the origin, where (cl f)(0) = 0 rather than +∞.
open scoped Topology
theorem convexFunctionClosure_example_origin :
convexFunctionClosure (fun x : Fin 1 → Real =>
if 0 < x 0 then (0 : EReal) else (⊤ : EReal)) =
fun x => if x 0 < 0 then (⊤ : EReal) else (0 : EReal)
Points in the closure of the unit ball are frequently in the unit ball.
open scoped Topology
lemma frequently_mem_ball_of_mem_closure {x : Fin 2 → Real}
(hx : x ∈ closure (Metric.ball (0 : Fin 2 → Real) 1)) :
∃ᶠ y in 𝓝 x, y ∈ Metric.ball (0 : Fin 2 → Real) 1
Points outside the closure of the unit ball have a neighborhood outside it.
open scoped Topology
lemma eventually_not_mem_closure_of_not_mem {x : Fin 2 → Real}
(hx : x ∉ closure (Metric.ball (0 : Fin 2 → Real) 1)) :
∀ᶠ y in 𝓝 x, y ∉ closure (Metric.ball (0 : Fin 2 → Real) 1)
On the closure of the unit ball, the liminf of f is 0.
open scoped Topology
lemma liminf_unitDisk_closure_eq_zero
(f : (Fin 2 → Real) → EReal)
(h0 : ∀ x, x ∈ Metric.ball (0 : Fin 2 → Real) 1 → f x = (0 : EReal))
(hnonneg : ∀ x, (0 : EReal) ≤ f x) {x : Fin 2 → Real}
(hx : x ∈ closure (Metric.ball (0 : Fin 2 → Real) 1)) :
Filter.liminf (fun y : Fin 2 → Real => f y) (𝓝 x) = (0 : EReal)
Outside the closure of the unit ball, the liminf of f is ⊤.
open scoped Topology
lemma liminf_unitDisk_outside_eq_top
(f : (Fin 2 → Real) → EReal)
(hInf : ∀ x, x ∉ closure (Metric.ball (0 : Fin 2 → Real) 1) →
f x = (⊤ : EReal)) {x : Fin 2 → Real}
(hx : x ∉ closure (Metric.ball (0 : Fin 2 → Real) 1)) :
Filter.liminf (fun y : Fin 2 → Real => f y) (𝓝 x) = (⊤ : EReal)
Text 7.0.14: If C is the unit disk in ℝ^2 and f(x) = 0 for x ∈ C while f(x) = +∞ for x ∉ C (with arbitrary boundary values), then cl f(x) = 0 for all x ∈ cl C and +∞ elsewhere.
open scoped Topology
theorem convexFunctionClosure_example_unitDisk
(f : (Fin 2 → Real) → EReal)
(h0 : ∀ x, x ∈ Metric.ball (0 : Fin 2 → Real) 1 → f x = (0 : EReal))
(hInf : ∀ x, x ∉ closure (Metric.ball (0 : Fin 2 → Real) 1) → f x = (⊤ : EReal))
(hnonneg : ∀ x, (0 : EReal) ≤ f x) :
(∀ x, x ∈ closure (Metric.ball (0 : Fin 2 → Real) 1) →
convexFunctionClosure f x = (0 : EReal)) ∧
(∀ x, x ∉ closure (Metric.ball (0 : Fin 2 → Real) 1) →
convexFunctionClosure f x = (⊤ : EReal))§7 — Closures of Convex Functions — Part 3
Lean authors: Zichen Wang, Wanli Ma, Xinyi Guo, Zaiwen Wen.
An improper convex function with nonempty effective domain attains ⊥ there.
open scoped Topology
lemma improperConvexFunctionOn_exists_bot_of_effectiveDomain_nonempty {n : Nat}
{f : (Fin n → Real) → EReal}
(hf : ImproperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hne : Set.Nonempty (effectiveDomain (Set.univ : Set (Fin n → Real)) f)) :
∃ u ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) f, f u = (⊥ : EReal)
Inverting an affine combination on a line.
open scoped Topology
lemma affine_combo_inverse {n : Nat} {u x y : EuclideanSpace Real (Fin n)} {μ : Real}
(hμ : μ ≠ 0)
(hy : y = (1 - μ) • u + μ • x) :
x = (1 - μ⁻¹) • u + μ⁻¹ • y
A convex combination with a ⊥ point is ⊥ when the other point is finite.
open scoped Topology
lemma convex_combo_eq_bot_of_bot_point {n : Nat} {f : (Fin n → Real) → EReal}
(hconv : ConvexFunctionOn (Set.univ) f) {u y x : EuclideanSpace Real (Fin n)}
(hu : f u = (⊥ : EReal)) (hy : f y < (⊤ : EReal)) {lam : Real}
(hlam0 : 0 < lam) (hlam1 : lam < 1) (hx : x = (1 - lam) • u + lam • y) :
f x = (⊥ : EReal)
Theorem 7.2: If f is an improper convex function, then f x = -∞ for every x ∈ ri (dom f). Thus an improper convex function is necessarily infinite except perhaps at relative boundary points of its effective domain.
open scoped Topology
theorem improperConvexFunctionOn_eq_bot_on_ri_effectiveDomain {n : Nat}
{f : (Fin n → Real) → EReal}
(hf : ImproperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
∀ x ∈
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f),
f x = (⊥ : EReal)
Lower semicontinuity implies the ⊥-sublevel set is closed.
open scoped Topology
lemma lsc_isClosed_sublevel_bot {n : Nat} {f : (Fin n → Real) → EReal}
(hls : LowerSemicontinuous f) :
IsClosed {x | f x ≤ (⊥ : EReal)}
Lower semicontinuity extends f = ⊥ from ri (dom f) to its closure.
open scoped Topology
lemma improperConvexFunctionOn_eq_bot_on_closure_ri_effectiveDomain {n : Nat}
{f : (Fin n → Real) → EReal}
(hls : LowerSemicontinuous f)
(hf : ImproperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
∀ x ∈
closure
(euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f)),
f x = (⊥ : EReal)
Points of the effective domain lie in the closure of its relative interior.
open scoped Topology
lemma mem_closure_ri_effectiveDomain_of_mem {n : Nat}
{f : (Fin n → Real) → EReal}
(hf : ImproperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
{x : Fin n → Real}
(hx : x ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) f) :
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm x) ∈
closure
(euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f))
Corollary 7.2.1. A lower semi-continuous improper convex function can have no finite values.
open scoped Topology
theorem lowerSemicontinuous_improperConvexFunction_no_finite_values {n : Nat}
{f : (Fin n → Real) → EReal}
(hf : ImproperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hls : LowerSemicontinuous f) :
∀ x, f x = (⊤ : EReal) ∨ f x = (⊥ : EReal)
Corollary 7.2.2. Let f be an improper convex function. Then cl f is a closed improper convex function and agrees with f on ri (dom f).
open scoped Topology
theorem convexFunctionClosure_closed_improperConvexFunctionOn_and_agrees_on_ri
{n : Nat} {f : (Fin n → Real) → EReal}
(hf : ImproperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
(ClosedConvexFunction (convexFunctionClosure f) ∧
ImproperConvexFunctionOn (Set.univ : Set (Fin n → Real))
(convexFunctionClosure f)) ∧
∀ x ∈
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f),
convexFunctionClosure f x = f x
A proper convex function on univ is strictly above ⊥ everywhere.
open scoped Topology
lemma properConvexFunctionOn_univ_imp_bot_lt {n : Nat}
{f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
∀ x, (⊥ : EReal) < f x
Relative openness of the effective domain forces ⊥ on it for improper functions.
open scoped Topology
lemma improperConvexFunctionOn_eq_bot_on_effectiveDomain_of_relOpen {n : Nat}
{f : (Fin n → Real) → EReal}
(hf : ImproperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hopen : euclideanRelativelyOpen n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f)) :
∀ x ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) f, f x = (⊥ : EReal)
Points outside the effective domain on univ must take value ⊤.
open scoped Topology
lemma not_mem_effectiveDomain_univ_imp_eq_top {n : Nat}
{f : (Fin n → Real) → EReal} {x : Fin n → Real}
(hx : x ∉ effectiveDomain (Set.univ : Set (Fin n → Real)) f) :
f x = (⊤ : EReal)
Corollary 7.2.3. If f is a convex function whose effective domain is relatively open (for instance if effectiveDomain Set.univ f = Set.univ), then either f x > -∞ for every x or f x is infinite for every x.
open scoped Topology
theorem convexFunction_relativelyOpen_effectiveDomain_or_infinite {n : Nat}
{f : (Fin n → Real) → EReal}
(hf : ConvexFunction f)
(hopen : euclideanRelativelyOpen n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f)) :
(∀ x, (⊥ : EReal) < f x) ∨ (∀ x, f x = (⊤ : EReal) ∨ f x = (⊥ : EReal))
The closure of an improper convex function agrees with the function on ri (dom f).
open scoped Topology
lemma convexFunctionClosure_agrees_on_ri_of_improper {n : Nat}
{f : (Fin n → Real) → EReal}
(hf : ImproperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
∀ x ∈
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f),
convexFunctionClosure f x = f x
The epigraph of the constant ⊥ function is all of ℝ^n × ℝ.
open scoped Topology
lemma epigraph_const_bot_univ {n : Nat} :
epigraph (S := (Set.univ : Set (Fin n → Real))) (fun _ => (⊥ : EReal)) =
(Set.univ : Set ((Fin n → Real) × Real))
If an improper convex function attains ⊥ and has dense effective domain, then its epigraph has dense closure.
open scoped Topology
lemma closure_epigraph_univ_of_exists_bot {n : Nat} {f : (Fin n → Real) → EReal}
(hf : ImproperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(_hbot : ∃ x, f x = (⊥ : EReal))
(hdense : closure (effectiveDomain (Set.univ : Set (Fin n → Real)) f) = Set.univ) :
closure (epigraph (S := (Set.univ : Set (Fin n → Real))) f) =
(Set.univ : Set ((Fin n → Real) × Real))
Text 7.0.15: Even when a convex function f has -∞ somewhere, its closure cl f is not drastically different: they coincide on ri (dom f), and when dom f is dense their epigraphs have the same closure.
open scoped Topology
theorem convexFunctionClosure_agrees_on_ri_and_same_epigraph_closure {n : Nat}
{f : (Fin n → Real) → EReal}
(hf : ImproperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hdense : closure (effectiveDomain (Set.univ : Set (Fin n → Real)) f) = Set.univ) :
(∀ x ∈
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f),
convexFunctionClosure f x = f x) ∧
closure (epigraph (S := (Set.univ : Set (Fin n → Real)))
(convexFunctionClosure f)) =
closure (epigraph (S := (Set.univ : Set (Fin n → Real))) f)
A lower semicontinuous function equals its lower semicontinuous hull.
open scoped Topology
lemma lowerSemicontinuousHull_eq_of_lsc {n : Nat} {f : (Fin n → Real) → EReal}
(hls : LowerSemicontinuous f) :
lowerSemicontinuousHull f = f
If f is lower semicontinuous and never ⊥, then its closure is itself.
open scoped Topology
lemma convexFunctionClosure_eq_of_no_bot {n : Nat} {f : (Fin n → Real) → EReal}
(hls : LowerSemicontinuous f) (hbot : ∀ x, f x ≠ (⊥ : EReal)) :
convexFunctionClosure f = f
open scoped Topology
theorem convexFunctionClosure_eq_of_closedConvexFunction {n : Nat}
{f : (Fin n → Real) → EReal} (hf : ClosedConvexFunction f)
(hbot : ∀ x, f x ≠ (⊥ : EReal)) :
convexFunctionClosure f = f
Rewriting the line infimum as an infimum over the projection fiber.
open scoped Topology
lemma example_inf_over_line_convex_g_eq_sInf_fiber
{f : (Fin 2 → Real) → EReal} (x : Fin 1 → Real) :
(iInf (fun xi2 : Real =>
f (fun i : Fin 2 => if i = 0 then x 0 else xi2))) =
sInf { z : EReal | ∃ u : Fin 2 → Real,
projectionLinearMap (by decide : 1 ≤ 2) u = x ∧ z = f u }
The fiber infimum of a convex function is convex.
open scoped Topology
lemma example_inf_over_line_convex_g_convex
{f : (Fin 2 → Real) → EReal} (hfconv : ConvexFunction f) :
ConvexFunction (fun x : Fin 1 → Real =>
iInf (fun xi2 : Real =>
f (fun i : Fin 2 => if i = 0 then x 0 else xi2)))
The infimum along a vertical line is strictly below ⊤ when f is finite.
open scoped Topology
lemma example_inf_over_line_convex_g_lt_top
{f : (Fin 2 → Real) → EReal}
(hfinite : ∀ x, f x ≠ (⊤ : EReal) ∧ f x ≠ (⊥ : EReal)) :
∀ x : Fin 1 → Real,
(iInf (fun xi2 : Real =>
f (fun i : Fin 2 => if i = 0 then x 0 else xi2))) < (⊤ : EReal)
The effective domain of the line infimum is all of ℝ.
open scoped Topology
lemma example_inf_over_line_convex_effectiveDomain_univ
{f : (Fin 2 → Real) → EReal}
(hfinite : ∀ x, f x ≠ (⊤ : EReal) ∧ f x ≠ (⊥ : EReal)) :
effectiveDomain (S := (Set.univ : Set (Fin 1 → Real)))
(fun x : Fin 1 → Real =>
iInf (fun xi2 : Real =>
f (fun i : Fin 2 => if i = 0 then x 0 else xi2))) = Set.univ
Corollary 7.2.3 applied to the infimum along vertical lines.
open scoped Topology
lemma example_inf_over_line_convex_dichotomy
{f : (Fin 2 → Real) → EReal}
(hfconv : ConvexFunction f)
(hfinite : ∀ x, f x ≠ (⊤ : EReal) ∧ f x ≠ (⊥ : EReal)) :
(∀ x : Fin 1 → Real,
(iInf (fun xi2 : Real =>
f (fun i : Fin 2 => if i = 0 then x 0 else xi2))) ≠ (⊤ : EReal) ∧
(iInf (fun xi2 : Real =>
f (fun i : Fin 2 => if i = 0 then x 0 else xi2))) ≠ (⊥ : EReal)) ∨
(∀ x : Fin 1 → Real,
(iInf (fun xi2 : Real =>
f (fun i : Fin 2 => if i = 0 then x 0 else xi2))) = (⊥ : EReal))
Bounded below on one vertical line implies bounded below on all vertical lines.
open scoped Topology
lemma example_inf_over_line_convex_bdd_below
{f : (Fin 2 → Real) → EReal}
(hfconv : ConvexFunction f)
(hfinite : ∀ x, f x ≠ (⊤ : EReal) ∧ f x ≠ (⊥ : EReal)) :
(∃ xi1 : Real, ∃ m : Real, ∀ xi2 : Real,
(m : EReal) ≤ f (fun i : Fin 2 => if i = 0 then xi1 else xi2)) →
∀ xi1 : Real, ∃ m : Real, ∀ xi2 : Real,
(m : EReal) ≤ f (fun i : Fin 2 => if i = 0 then xi1 else xi2)
Example 7.0.22: Let f be a finite convex function on ℝ^2 and define g(xi1) = inf_{xi2 ∈ ℝ} f(xi1, xi2). Then g is convex and dom g = ℝ. By Corollary 7.2.3, either g(xi1) is finite for all xi1 or g(xi1) = -∞ for all xi1. Consequently, if f is bounded below on one line parallel to the xi2-axis, then it is bounded below on every such line.
open scoped Topology
theorem example_inf_over_line_convex
{f : (Fin 2 → Real) → EReal}
(hfconv : ConvexFunction f)
(hfinite : ∀ x, f x ≠ (⊤ : EReal) ∧ f x ≠ (⊥ : EReal)) :
let g : (Fin 1 → Real) → EReal :=
fun x =>
iInf (fun xi2 : Real => f (fun i : Fin 2 => if i = 0 then x 0 else xi2))
;
ConvexFunction g ∧
effectiveDomain (S := (Set.univ : Set (Fin 1 → Real))) g = Set.univ ∧
((∀ x, g x ≠ (⊤ : EReal) ∧ g x ≠ (⊥ : EReal)) ∨
(∀ x, g x = (⊥ : EReal))) ∧
((∃ xi1 : Real, ∃ m : Real, ∀ xi2 : Real,
(m : EReal) ≤ f (fun i : Fin 2 => if i = 0 then xi1 else xi2)) →
∀ xi1 : Real, ∃ m : Real, ∀ xi2 : Real,
(m : EReal) ≤ f (fun i : Fin 2 => if i = 0 then xi1 else xi2))
Remark 7.0.23: Theorems 7.2 and later show that comparisons between f and cl f hinge on relative interiors; in particular, the set ri (epi f) plays a key role.
open scoped Topology
def riEpigraph {n : Nat} (f : (Fin n → Real) → EReal) :
Set (EuclideanSpace Real (Fin (n + 1))) :=
euclideanRelativeInterior (n + 1)
((appendAffineEquiv n 1) ''
((fun p : (Fin n → Real) × Real =>
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm p.1,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => p.2))) ''
epigraph (S := (Set.univ : Set (Fin n → Real))) f))§7 — Closures of Convex Functions — Part 4
Lean authors: Zichen Wang, Wanli Ma, Xinyi Guo, Zaiwen Wen.
The vertical section of the embedded epigraph corresponds to the usual inequality.
open scoped Topology
lemma riEpigraph_section_mem_iff {n : Nat} {f : (Fin n → Real) → EReal}
(x : Fin n → Real) (μ : Real) :
let C : Set (EuclideanSpace Real (Fin (n + 1))) :=
(appendAffineEquiv n 1) ''
((fun p : (Fin n → Real) × Real =>
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm p.1,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => p.2))) ''
epigraph (S := (Set.univ : Set (Fin n → Real))) f)
let y : EuclideanSpace Real (Fin n) :=
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm x
let z : EuclideanSpace Real (Fin 1) :=
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm (fun _ : Fin 1 => μ)
let Cy : EuclideanSpace Real (Fin n) → Set (EuclideanSpace Real (Fin 1)) :=
fun y => {z | appendAffineEquiv n 1 (y, z) ∈ C}
z ∈ Cy y ↔ f x ≤ (μ : EReal)§7 — Closures of Convex Functions — Part 5
Lean authors: Zichen Wang, Wanli Ma, Xinyi Guo, Zaiwen Wen.
The base of nonempty vertical sections equals the effective domain.
open scoped Topology
lemma riEpigraph_D_eq_effectiveDomain {n : Nat} {f : (Fin n → Real) → EReal} :
let C : Set (EuclideanSpace Real (Fin (n + 1))) :=
(appendAffineEquiv n 1) ''
((fun p : (Fin n → Real) × Real =>
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm p.1,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => p.2))) ''
epigraph (S := (Set.univ : Set (Fin n → Real))) f)
let Cy : EuclideanSpace Real (Fin n) → Set (EuclideanSpace Real (Fin 1)) :=
fun y => {z | appendAffineEquiv n 1 (y, z) ∈ C}
let D : Set (EuclideanSpace Real (Fin n)) := {y | (Cy y).Nonempty}
D =
(fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f
Relative interior of a vertical section is the strict inequality when f x < ⊤.
open scoped Topology
lemma riEpigraph_section_ri_iff {n : Nat} {f : (Fin n → Real) → EReal}
(x : Fin n → Real) (μ : Real) (hfx : f x < (⊤ : EReal)) :
let C : Set (EuclideanSpace Real (Fin (n + 1))) :=
(appendAffineEquiv n 1) ''
((fun p : (Fin n → Real) × Real =>
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm p.1,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => p.2))) ''
epigraph (S := (Set.univ : Set (Fin n → Real))) f)
let y : EuclideanSpace Real (Fin n) :=
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm x
let z : EuclideanSpace Real (Fin 1) :=
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm (fun _ : Fin 1 => μ)
let Cy : EuclideanSpace Real (Fin n) → Set (EuclideanSpace Real (Fin 1)) :=
fun y => {z | appendAffineEquiv n 1 (y, z) ∈ C}
z ∈ euclideanRelativeInterior 1 (Cy y) ↔ f x < (μ : EReal)
Real points are always strictly below ⊤.
open scoped Topology
lemma riEpigraph_mu_lt_top (μ : Real) : (μ : EReal) < ⊤
Lemma 7.3. For any convex function f, ri (epi f) consists of the pairs (x, μ) such that x ∈ ri (dom f) and f x < μ < ∞.
open scoped Topology
theorem riEpigraph_mem_iff {n : Nat} {f : (Fin n → Real) → EReal}
(hf : ConvexFunction f) (x : Fin n → Real) (μ : Real) :
(appendAffineEquiv n 1)
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm x,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => μ)) ∈ riEpigraph f ↔
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm x) ∈
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f) ∧
f x < (μ : EReal) ∧ (μ : EReal) < ⊤
Find a real height strictly between f x and α.
open scoped Topology
lemma exists_real_between_fx_alpha {n : Nat} {f : (Fin n → Real) → EReal}
{α : Real} {x : Fin n → Real} (h : f x < (α : EReal)) :
∃ μ : Real, f x < (μ : EReal) ∧ (μ : EReal) < (α : EReal)
The embedded epigraph used in riEpigraph is convex.
open scoped Topology
lemma convex_embedded_epigraph {n : Nat} {f : (Fin n → Real) → EReal}
(hf : ConvexFunction f) :
let C : Set (EuclideanSpace Real (Fin (n + 1))) :=
(appendAffineEquiv n 1) ''
((fun p : (Fin n → Real) × Real =>
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm p.1,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => p.2))) ''
epigraph (S := (Set.univ : Set (Fin n → Real))) f)
Convex ℝ C
The open half-space defined by the last coordinate is open.
open scoped Topology
lemma isOpen_openHalfspace_lastCoord {n : Nat} {α : Real} :
IsOpen
{p : EuclideanSpace Real (Fin (n + 1)) |
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)
((appendAffineEquiv n 1).symm p).2) 0 < α}
Build a point of the embedded epigraph lying strictly below height α.
open scoped Topology
lemma exists_point_openHalfspace_inter_embedded_epigraph {n : Nat}
{f : (Fin n → Real) → EReal} {α : Real} (hα : ∃ x, f x < (α : EReal)) :
let C : Set (EuclideanSpace Real (Fin (n + 1))) :=
(appendAffineEquiv n 1) ''
((fun p : (Fin n → Real) × Real =>
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm p.1,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => p.2))) ''
epigraph (S := (Set.univ : Set (Fin n → Real))) f)
let U : Set (EuclideanSpace Real (Fin (n + 1))) :=
{p : EuclideanSpace Real (Fin (n + 1)) |
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1) ((appendAffineEquiv n 1).symm p).2) 0 < α}
(U ∩ C).Nonempty
Corollary 7.3.1: Let α be a real number, and let f be a convex function such that, for some x, f x < α. Then there exists x ∈ ri (dom f) with f x < α.
open scoped Topology
theorem exists_lt_on_ri_effectiveDomain_of_convexFunction {n : Nat}
{f : (Fin n → Real) → EReal} (hf : ConvexFunction f) {α : Real}
(hα : ∃ x, f x < (α : EReal)) :
∃ x ∈
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f),
f x < (α : EReal)
Corollary 7.3.2. Let f be a convex function, and let C be a convex set such that ri C ⊆ dom f. Let α ∈ ℝ be such that f x < α for some x ∈ cl C. Then f x < α for some x ∈ ri C.
open scoped Topology
theorem exists_lt_on_ri_of_convexFunction_convexSet {n : Nat}
{f : (Fin n → Real) → EReal} (hf : ConvexFunction f)
{C : Set (EuclideanSpace Real (Fin n))} (hC : Convex ℝ C)
(hri :
euclideanRelativeInterior n C ⊆
(fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f)
{α : Real} (hα : ∃ x ∈ closure C, f (x : Fin n → Real) < (α : EReal)) :
∃ x ∈ euclideanRelativeInterior n C, f (x : Fin n → Real) < (α : EReal)
If f is finite on C, then ri C lies in the effective domain of f.
open scoped Topology
lemma ri_subset_effectiveDomain_of_finite_on_C {n : Nat}
{f : (Fin n → Real) → EReal} {C : Set (EuclideanSpace Real (Fin n))}
(hfinite : ∀ x ∈ C,
f (x : Fin n → Real) ≠ (⊤ : EReal) ∧ f (x : Fin n → Real) ≠ (⊥ : EReal)) :
euclideanRelativeInterior n C ⊆
(fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f
Corollary 7.3.3. Let f be a convex function on ℝ^n, and let C be a convex set on which f is finite. If f x ≥ α for every x ∈ C, then also f x ≥ α for every x ∈ cl C.
open scoped Topology
theorem convexFunction_ge_on_closure_of_convexSet {n : Nat}
{f : (Fin n → Real) → EReal} (hf : ConvexFunction f)
{C : Set (EuclideanSpace Real (Fin n))} (hC : Convex ℝ C)
(hfinite : ∀ x ∈ C,
f (x : Fin n → Real) ≠ (⊤ : EReal) ∧ f (x : Fin n → Real) ≠ (⊥ : EReal))
{α : Real} (hα : ∀ x ∈ C, f (x : Fin n → Real) ≥ (α : EReal)) :
∀ x ∈ closure C, f (x : Fin n → Real) ≥ (α : EReal)
If a convex function has a bottom value, then it attains ⊥ on ri (dom f).
open scoped Topology
lemma exists_bot_on_ri_effectiveDomain_of_exists_bot {n : Nat}
{f : (Fin n → Real) → EReal} (hf : ConvexFunction f)
(hbot : ∃ x, f x = (⊥ : EReal)) :
∃ y ∈
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f),
f (y : Fin n → Real) = (⊥ : EReal)
Agreement on ri (dom f) yields equality of ri (epi f).
open scoped Topology
lemma riEpigraph_eq_of_agree_on_ri_effectiveDomain {n : Nat}
{f g : (Fin n → Real) → EReal} (hf : ConvexFunction f) (hg : ConvexFunction g)
(hri :
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f) =
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) g))
(hagree :
∀ x ∈
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f),
f (x : Fin n → Real) = g (x : Fin n → Real)) :
riEpigraph f = riEpigraph g
Relative interior equality gives closure equality for embedded epigraphs.
open scoped Topology
lemma closure_embedded_epigraph_eq_of_riEpigraph_eq {n : Nat}
{f g : (Fin n → Real) → EReal} (hf : ConvexFunction f) (hg : ConvexFunction g)
(hri_epi : riEpigraph f = riEpigraph g) :
let C_f : Set (EuclideanSpace Real (Fin (n + 1))) :=
(appendAffineEquiv n 1) ''
((fun p : (Fin n → Real) × Real =>
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm p.1,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => p.2))) ''
epigraph (S := (Set.univ : Set (Fin n → Real))) f)
let C_g : Set (EuclideanSpace Real (Fin (n + 1))) :=
(appendAffineEquiv n 1) ''
((fun p : (Fin n → Real) × Real =>
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm p.1,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => p.2))) ''
epigraph (S := (Set.univ : Set (Fin n → Real))) g)
closure C_f = closure C_g§7 — Closures of Convex Functions — Part 6
Lean authors: Zichen Wang, Wanli Ma, Xinyi Guo, Zaiwen Wen.
Equality of embedded epigraph closures yields equality of epigraph closures.
open scoped Topology
lemma closure_epigraph_eq_of_embedded_closure_eq {n : Nat}
{f g : (Fin n → Real) → EReal} :
let C_f : Set (EuclideanSpace Real (Fin (n + 1))) :=
(appendAffineEquiv n 1) ''
((fun p : (Fin n → Real) × Real =>
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm p.1,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => p.2))) ''
epigraph (S := (Set.univ : Set (Fin n → Real))) f)
let C_g : Set (EuclideanSpace Real (Fin (n + 1))) :=
(appendAffineEquiv n 1) ''
((fun p : (Fin n → Real) × Real =>
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm p.1,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => p.2))) ''
epigraph (S := (Set.univ : Set (Fin n → Real))) g)
closure C_f = closure C_g →
closure (epigraph (S := (Set.univ : Set (Fin n → Real))) f) =
closure (epigraph (S := (Set.univ : Set (Fin n → Real))) g)
Equality of epigraph closures yields equality of convex closures.
open scoped Topology
lemma convexFunctionClosure_eq_of_epigraph_closure_eq {n : Nat}
{f g : (Fin n → Real) → EReal} (hbf : ∀ x, f x ≠ (⊥ : EReal))
(hbg : ∀ x, g x ≠ (⊥ : EReal))
(hcl :
closure (epigraph (S := (Set.univ : Set (Fin n → Real))) f) =
closure (epigraph (S := (Set.univ : Set (Fin n → Real))) g)) :
convexFunctionClosure f = convexFunctionClosure g
Corollary 7.3.4. If f and g are convex functions on ℝ^n such that ri (dom f) = ri (dom g), and f and g agree on the latter set, then cl f = cl g.
open scoped Topology
theorem convexFunctionClosure_eq_of_agree_on_ri_effectiveDomain {n : Nat}
{f g : (Fin n → Real) → EReal}
(hf : ConvexFunction f) (hg : ConvexFunction g)
(hri :
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f) =
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) g))
(hagree :
∀ x ∈
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f),
f (x : Fin n → Real) = g (x : Fin n → Real)) :
convexFunctionClosure f = convexFunctionClosure g
A point of ri (dom f) lifts to a point of ri (epi f) on the vertical line.
open scoped Topology
lemma exists_riEpigraph_point_on_vertical_line {n : Nat}
{f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
{x : EuclideanSpace Real (Fin n)}
(hx :
x ∈
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f)) :
∃ μ : Real,
appendAffineEquiv n 1
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm (x : Fin n → Real),
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => μ)) ∈ riEpigraph f
The vertical section of the embedded epigraph is the inequality in the last coordinate.
open scoped Topology
lemma embedded_epigraph_section_mem_iff {n : Nat} {f : (Fin n → Real) → EReal}
(x : Fin n → Real) :
let C : Set (EuclideanSpace Real (Fin (n + 1))) :=
(appendAffineEquiv n 1) ''
((fun p : (Fin n → Real) × Real =>
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm p.1,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => p.2))) ''
epigraph (S := (Set.univ : Set (Fin n → Real))) f)
let y : EuclideanSpace Real (Fin n) :=
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm x
let Cy : EuclideanSpace Real (Fin n) → Set (EuclideanSpace Real (Fin 1)) :=
fun y => {z | appendAffineEquiv n 1 (y, z) ∈ C}
let coords1 := EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)
let first1 : EuclideanSpace Real (Fin 1) → Real := fun z => coords1 z 0
∀ z, z ∈ Cy y ↔ f x ≤ (first1 z : EReal)
On the relative interior of the effective domain, cl f agrees with f.
open scoped Topology
lemma convexFunctionClosure_eq_on_ri {n : Nat} {f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
∀ x ∈
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f),
convexFunctionClosure f x = f x§7 — Closures of Convex Functions — Part 7
Lean authors: Zichen Wang, Wanli Ma, Xinyi Guo, Zaiwen Wen.
open scoped Topology
theorem convexFunctionClosure_closed_properConvexFunctionOn_and_agrees_on_ri
{n : Nat} {f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
(ClosedConvexFunction (convexFunctionClosure f) ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real))
(convexFunctionClosure f)) ∧
∀ x ∈
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f),
convexFunctionClosure f x = f x§7 — Closures of Convex Functions — Part 8
Lean authors: Zichen Wang, Wanli Ma, Xinyi Guo, Zaiwen Wen.
The effective domain of f is contained in the effective domain of its closure.
open scoped Topology
lemma domf_subset_domcl_of_closure_le {n : Nat} (f : (Fin n → Real) → EReal) :
(fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f ⊆
(fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) (convexFunctionClosure f)
The effective domain of the closure lies in the closure of the effective domain.
open scoped Topology
lemma domcl_subset_closure_domf {n : Nat} {f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
(fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) (convexFunctionClosure f) ⊆
closure
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f)
If domcl lies in closure domf, then it lies in domf or its relative boundary.
open scoped Topology
lemma domcl_subset_domf_union_rbd {n : Nat}
{domf domcl : Set (EuclideanSpace Real (Fin n))} (hsubset : domcl ⊆ closure domf) :
domcl ⊆ domf ∪ euclideanRelativeBoundary n domf
Equal closures for domf and domcl follow from the two inclusions.
open scoped Topology
lemma closure_domcl_eq_domf {n : Nat}
{domf domcl : Set (EuclideanSpace Real (Fin n))} (hdomf : domf ⊆ domcl)
(hdomcl : domcl ⊆ closure domf) :
closure domcl = closure domf
Relative interiors coincide under closure equality for convex sets.
open scoped Topology
lemma ri_domcl_eq_domf {n : Nat}
{domf domcl : Set (EuclideanSpace Real (Fin n))} (hconvf : Convex Real domf)
(hconvcl : Convex Real domcl) (hcl : closure domcl = closure domf) :
euclideanRelativeInterior n domcl = euclideanRelativeInterior n domf
Corollary 7.4.1. If f is a proper convex function, then dom (cl f) differs from dom f at most by including some additional relative boundary points of dom f. In particular, dom (cl f) and dom f have the same closure and relative interior, as well as the same dimension.
open scoped Topology
theorem convexFunctionClosure_effectiveDomain_subset_relativeBoundary_and_same_closure_ri_dim
{n : Nat} {f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
let domf : Set (EuclideanSpace Real (Fin n)) :=
(fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f
let domcl : Set (EuclideanSpace Real (Fin n)) :=
(fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) (convexFunctionClosure f)
domf ⊆ domcl ∧
domcl ⊆ domf ∪ euclideanRelativeBoundary n domf ∧
closure domcl = closure domf ∧
euclideanRelativeInterior n domcl = euclideanRelativeInterior n domf ∧
Module.finrank Real (affineSpan Real domcl).direction =
Module.finrank Real (affineSpan Real domf).direction
Affine sets have full relative interior and empty relative boundary.
open scoped Topology
lemma ri_eq_and_boundary_empty_of_isAffineSet {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))}
(hC :
IsAffineSet n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) '' C)) :
euclideanRelativeInterior n C = C ∧ euclideanRelativeBoundary n C = ∅
For affine effective domain, the effective domain of the closure coincides.
open scoped Topology
lemma domcl_eq_domf_of_affine_effectiveDomain {n : Nat} {f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(haff : IsAffineSet n (effectiveDomain (Set.univ : Set (Fin n → Real)) f)) :
let domf : Set (EuclideanSpace Real (Fin n)) :=
(fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f
let domcl : Set (EuclideanSpace Real (Fin n)) :=
(fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) (convexFunctionClosure f)
domcl = domf
For affine effective domain, the closure agrees with the function everywhere.
open scoped Topology
lemma convexFunctionClosure_eq_of_affine_effectiveDomain {n : Nat} {f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(haff : IsAffineSet n (effectiveDomain (Set.univ : Set (Fin n → Real)) f)) :
convexFunctionClosure f = f
Corollary 7.4.2. If f is a proper convex function such that dom f is an affine set (which holds in particular if f is finite throughout ℝ^n), then f is closed.
open scoped Topology
theorem properConvexFunction_closed_of_affine_effectiveDomain {n : Nat}
{f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(haff : IsAffineSet n (effectiveDomain (Set.univ : Set (Fin n → Real)) f)) :
ClosedConvexFunction f
A lower semicontinuous minorant that agrees at a point transfers lsc.
open scoped Topology
lemma lowerSemicontinuousAt_of_le_of_eq {α β : Type*} [TopologicalSpace α] [Preorder β]
{f g : α → β} {x : α} (hg : LowerSemicontinuousAt g x) (hle : g ≤ f)
(hxeq : g x = f x) : LowerSemicontinuousAt f x
open scoped Topology
theorem convexFunction_lowerSemicontinuousAt_on_ri_effectiveDomain {n : Nat}
{f : (Fin n → Real) → EReal} (hf : ConvexFunction f) :
∀ x ∈
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f),
LowerSemicontinuousAt f x
Proper convex functions are finite on the relative interior of their effective domain.
open scoped Topology
lemma properConvexFunctionOn_ne_top_on_ri_effectiveDomain {n : Nat}
{f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
∀ x ∈
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f),
f x ≠ (⊥ : EReal) ∧ f x ≠ (⊤ : EReal)
The toReal of a proper convex function is convex on its effective domain.
open scoped Topology
lemma convexOn_toReal_on_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)
Lift continuity of toReal ∘ f to continuity of f on finite-valued sets.
open scoped Topology
lemma continuousOn_ereal_of_toReal {α : Type*} [TopologicalSpace α] {f : α → EReal}
{s : Set α} (hcont : ContinuousOn (fun x => (f x).toReal) s)
(hfinite : ∀ x ∈ s, f x ≠ (⊥ : EReal) ∧ f x ≠ (⊤ : EReal)) :
ContinuousOn f s
In full dimension, relative interior equals interior, giving continuity for convex maps.
open scoped Topology
lemma convexOn_continuousOn_ri_of_affineSpan_eq_univ {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} {g : EuclideanSpace Real (Fin n) → ℝ}
(hg : ConvexOn ℝ C g)
(hspan : (affineSpan Real C : Set (EuclideanSpace Real (Fin n))) = Set.univ) :
ContinuousOn g (euclideanRelativeInterior n C)
Affine equivalences transport continuity on relative interiors.
open scoped Topology
lemma continuousOn_ri_of_affineEquiv {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} {g : EuclideanSpace Real (Fin n) → ℝ}
(e : EuclideanSpace Real (Fin n) ≃ᵃ[Real] EuclideanSpace Real (Fin n))
(hcont :
ContinuousOn (fun x => g (e.symm x)) (euclideanRelativeInterior n (e '' C))) :
ContinuousOn g (euclideanRelativeInterior n C)
The coordinate subspace as a submodule.
open scoped Topology
def coordinateSubmodule (n m : Nat) : Submodule Real (EuclideanSpace Real (Fin n))
Coordinate-subspace affine equivalence using extend-by-zero and restriction.
open scoped Topology
noncomputable def coordinateSubspace_affineEquiv {n m : Nat} (hmn : m ≤ n) :
EuclideanSpace Real (Fin m) ≃ᵃ[Real] (coordinateSubmodule n m)
Pulling back a full-dimensional set from the coordinate subspace gives full affine span.
open scoped Topology
lemma affineSpan_preimage_eq_univ {n m : Nat} (hmn : m ≤ n)
{C : Set (EuclideanSpace Real (Fin n))}
(hspan :
(affineSpan Real C : Set (EuclideanSpace Real (Fin n))) = coordinateSubspace n m) :
(affineSpan Real
((coordinateSubspace_affineEquiv (n := n) (m := m) hmn).symm ''
(Subtype.val ⁻¹' C)) : Set (EuclideanSpace Real (Fin m))) = Set.univ
If A has a right inverse on s, then A '' (A ⁻¹' s) = s.
open scoped Topology
lemma image_preimage_eq_of_rightInverse {α β : Type*} {A : α → β} {g : β → α} {s : Set β}
(hright : ∀ y ∈ s, A (g y) = y) : A '' (A ⁻¹' s) = s
For linear affine maps, the value equals the linear part.
open scoped Topology
lemma affineMap_apply_eq_linear {n m : Nat}
(A : EuclideanSpace Real (Fin m) →ᵃ[Real] EuclideanSpace Real (Fin n))
(hA0 : A 0 = 0) : ∀ x, A x = A.linear x
The relative interior of a linear affine image is the image of the relative interior.
open scoped Topology
lemma ri_image_linearMap_eq {n m : Nat}
{D : Set (EuclideanSpace Real (Fin m))} (hconvD : Convex Real D)
(A : EuclideanSpace Real (Fin m) →ᵃ[Real] EuclideanSpace Real (Fin n))
(hA0 : A 0 = 0) :
euclideanRelativeInterior n (A '' D) = A '' euclideanRelativeInterior m D
The coordinate-subspace embedding is surjective onto subsets of the subspace.
open scoped Topology
lemma A_image_eq_C' {n m : Nat} {C' : Set (EuclideanSpace Real (Fin n))}
(e : EuclideanSpace Real (Fin m) ≃ᵃ[Real] coordinateSubmodule n m)
(A : EuclideanSpace Real (Fin m) →ᵃ[Real] EuclideanSpace Real (Fin n))
(hA : A = ((coordinateSubmodule n m).subtype.toAffineMap).comp e.toAffineMap)
(hC' : C' ⊆ coordinateSubspace n m) :
A '' (A ⁻¹' C') = C'
Transfer continuity along a left inverse on an image.
open scoped Topology
lemma continuousOn_of_leftInverse_on_range {α β γ : Type*} [TopologicalSpace α]
[TopologicalSpace β] [TopologicalSpace γ]
{A : α → β} {g : β → α} {f : β → γ} {s : Set α} {t : Set β}
(hcont : ContinuousOn (fun x => f (A x)) s)
(hleft : Function.LeftInverse g A)
(ht : t = A '' s)
(hg : Continuous g) :
ContinuousOn f t
Reduce continuity on relative interior via an affine coordinate subspace model.
open scoped Topology
lemma convexOn_continuousOn_ri_via_coordinateSubspace {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} {g : EuclideanSpace Real (Fin n) → ℝ}
(hg : ConvexOn ℝ C g) : ContinuousOn g (euclideanRelativeInterior n C)§7 — Closures of Convex Functions — Part 9
Lean authors: Zichen Wang, Wanli Ma, Xinyi Guo, Zaiwen Wen.
open scoped Topology
lemma continuousOn_toReal_on_ri_effectiveDomain {n : Nat} {f : (Fin n → Real) → EReal}
(hg : ConvexOn ℝ (effectiveDomain (Set.univ : Set (Fin n → Real)) f) (fun x => (f x).toReal)) :
ContinuousOn (fun x : EuclideanSpace Real (Fin n) => (f x).toReal)
(euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f))
open scoped Topology
lemma convexFunction_continuousOn_ri_effectiveDomain_of_proper {n : Nat}
{f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
ContinuousOn (fun x : EuclideanSpace Real (Fin n) => f x)
(euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f))
Remark 7.0.24: A convex function is continuous relative to ri (dom f).
open scoped Topology
theorem convexFunction_continuousOn_ri_effectiveDomain {n : Nat}
{f : (Fin n → Real) → EReal} (hf : ConvexFunction f) :
ContinuousOn (fun x : EuclideanSpace Real (Fin n) => f x)
(euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f))
The segment map tends to its endpoint as t → 1.
open scoped Topology
lemma tendsto_segment_to_y {n : Nat} (x y : EuclideanSpace Real (Fin n)) :
Filter.Tendsto (fun t : Real => (1 - t) • x + t • y) (𝓝 (1 : Real)) (𝓝 y)
Liminf along the segment bounds the closure from below.
open scoped Topology
lemma liminf_along_segment_ge_closure {n : Nat} {f : (Fin n → Real) → EReal}
{x y : EuclideanSpace Real (Fin n)}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
convexFunctionClosure f y ≤
Filter.liminf (fun t : Real => f ((1 - t) • x + t • y))
(nhdsWithin (1 : Real) (Set.Iio (1 : Real)))
Upper bound for the limsup along a segment via the embedded epigraph.
open scoped Topology
lemma limsup_along_segment_le_beta {n : Nat} {f : (Fin n → Real) → EReal}
{x y : EuclideanSpace Real (Fin n)}
(hx :
x ∈ euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f))
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) {β : Real}
(hβ : convexFunctionClosure f y ≤ (β : EReal)) :
Filter.limsup (fun t : Real => f ((1 - t) • x + t • y))
(nhdsWithin (1 : Real) (Set.Iio (1 : Real))) ≤ (β : EReal)
Theorem 7.5. Let f be a proper convex function, and let x ∈ ri (dom f). Then (cl f)(y) = lim_{λ ↑ 1} f((1 - λ)x + λ y) for every y. (The formula is also valid when f is improper and y ∈ cl (dom f).)
open scoped Topology
theorem convexFunctionClosure_eq_limit_along_segment {n : Nat}
{f : (Fin n → Real) → EReal} {x : EuclideanSpace Real (Fin n)}
(hx :
x ∈ euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f)) :
(ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f →
∀ y : EuclideanSpace Real (Fin n),
Filter.Tendsto
(fun t : Real => f ((1 - t) • x + t • y))
(nhdsWithin (1 : Real) (Set.Iio (1 : Real)))
(𝓝 (convexFunctionClosure f y))) ∧
(ImproperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f →
∀ y : EuclideanSpace Real (Fin n),
y ∈ closure
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f) →
Filter.Tendsto
(fun t : Real => f ((1 - t) • x + t • y))
(nhdsWithin (1 : Real) (Set.Iio (1 : Real)))
(𝓝 (convexFunctionClosure f y)))
Corollary 7.5.1. For a closed proper convex function f, one has f y = lim_{λ ↑ 1} f((1 - λ) x + λ y) for every x ∈ dom f and every y.
open scoped Topology
theorem closedProperConvexFunction_eq_limit_along_segment {n : Nat}
{f : (Fin n → Real) → EReal}
(hfclosed : ClosedConvexFunction f)
(hfproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
{x : EuclideanSpace Real (Fin n)}
(hx :
x ∈
(fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f) :
∀ y : EuclideanSpace Real (Fin n),
Filter.Tendsto
(fun t : Real => f ((1 - t) • x + t • y))
(nhdsWithin (1 : Real) (Set.Iio (1 : Real)))
(𝓝 (f y))§7 — Closures of Convex Functions — Part 10
Lean authors: Zichen Wang, Wanli Ma, Xinyi Guo, Zaiwen Wen.
Convexity of t ↦ -sqrt (1 - t) on (-∞, 1).
open scoped Topology
open scoped Pointwise
lemma convexOn_neg_sqrt_one_sub :
ConvexOn ℝ (Set.Iio (1 : ℝ)) (fun t => -Real.sqrt (1 - t))
Monotonicity of t ↦ -sqrt (1 - t) on (-∞, 1).
open scoped Topology
open scoped Pointwise
lemma monotoneOn_neg_sqrt_one_sub :
MonotoneOn (fun t => -Real.sqrt (1 - t)) (Set.Iio (1 : ℝ))
Convexity of the squared norm on the whole space.
open scoped Topology
open scoped Pointwise
lemma convexOn_norm_sq_univ {n : Nat} :
ConvexOn ℝ (Set.univ : Set (Fin n → ℝ)) (fun x => ‖x‖ ^ (2 : ℕ))
Convexity of x ↦ -sqrt(1 - ‖x‖^2) on the open unit ball.
open scoped Topology
open scoped Pointwise
lemma convexOn_neg_sqrt_one_sub_norm_sq {n : Nat} :
ConvexOn ℝ {x : Fin n → ℝ | ‖x‖ < 1}
(fun x => -Real.sqrt (1 - ‖x‖ ^ (2 : ℕ)))
Remark 7.0.25: Theorem 7.5 can be used to show convexity. For example, the function f(x) = -(1 - |x|^2)^{1/2} on ℝ^n (with f(x) = +∞ for |x| ≥ 1) is convex by verifying the limit relation in Theorem 7.5 at boundary points.
open scoped Topology
open scoped Pointwise
theorem convexFunction_negativeSqrt_unitBall {n : Nat} :
ConvexFunction (fun x : (Fin n → Real) =>
if ‖x‖ < 1 then ((-(Real.sqrt (1 - ‖x‖ ^ 2))) : Real) else (⊤ : EReal))
If iInf f < α, then some point in ri (dom f) satisfies f x < α.
open scoped Topology
open scoped Pointwise
lemma exists_lt_on_ri_effectiveDomain_of_iInf_lt {n : Nat} {f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) {α : Real}
(hα : iInf (fun x => f x) < (α : EReal)) :
∃ x ∈
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f),
f x < (α : EReal)
Horizontal section of the embedded epigraph corresponds to the ≤-sublevel set.
open scoped Topology
open scoped Pointwise
lemma levelSets_horizontal_section_mem_iff {n : Nat} {f : (Fin n → Real) → EReal}
(α : Real) (x : EuclideanSpace Real (Fin n)) :
let C : Set (EuclideanSpace Real (Fin (n + 1))) :=
(appendAffineEquiv n 1) ''
((fun p : (Fin n → Real) × Real =>
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm p.1,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => p.2))) ''
epigraph (S := (Set.univ : Set (Fin n → Real))) f)
let y : EuclideanSpace Real (Fin n) :=
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm (x : Fin n → Real)
let zα : EuclideanSpace Real (Fin 1) :=
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm (fun _ : Fin 1 => α)
appendAffineEquiv n 1 (y, zα) ∈ C ↔ f (x : Fin n → Real) ≤ (α : EReal)
A point in ri (dom f) with f x < α yields a point of ri (epi f) at height α.
open scoped Topology
open scoped Pointwise
lemma exists_riEpigraph_point_at_height {n : Nat} {f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) {α : Real}
(hα : iInf (fun x => f x) < (α : EReal)) :
∃ x ∈
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f),
f x < (α : EReal) ∧
(appendAffineEquiv n 1)
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm (x : Fin n → Real),
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => α)) ∈ riEpigraph f
The horizontal hyperplane at height α meets ri (epi f) when α > inf f.
open scoped Topology
open scoped Pointwise
lemma levelSets_plane_meets_riEpigraph {n : Nat} {f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) {α : Real}
(hα : iInf (fun x => f x) < (α : EReal)) :
let C : Set (EuclideanSpace Real (Fin (n + 1))) :=
(appendAffineEquiv n 1) ''
((fun p : (Fin n → Real) × Real =>
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm p.1,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => p.2))) ''
epigraph (S := (Set.univ : Set (Fin n → Real))) f)
let zα : EuclideanSpace Real (Fin 1) :=
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm (fun _ : Fin 1 => α)
let φ : EuclideanSpace Real (Fin (n + 1)) →ₗ[Real] EuclideanSpace Real (Fin 1) :=
(LinearMap.snd (R := Real) (M := EuclideanSpace Real (Fin n))
(M₂ := EuclideanSpace Real (Fin 1))).comp
(appendAffineEquiv n 1).symm.linear.toLinearMap
let M : AffineSubspace Real (EuclideanSpace Real (Fin (n + 1))) :=
AffineSubspace.mk' (appendAffineEquiv n 1 (0, zα)) (LinearMap.ker φ)
((M : Set _) ∩ euclideanRelativeInterior (n + 1) C).Nonempty
Points of the form appendAffineEquiv n 1 (y, zα) lie in the horizontal plane M.
open scoped Topology
open scoped Pointwise
lemma appendAffineEquiv_mem_horizontal_plane {n : Nat} {α : Real}
(y : EuclideanSpace Real (Fin n)) :
let zα : EuclideanSpace Real (Fin 1) :=
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm (fun _ : Fin 1 => α)
let φ : EuclideanSpace Real (Fin (n + 1)) →ₗ[Real] EuclideanSpace Real (Fin 1) :=
(LinearMap.snd (R := Real) (M := EuclideanSpace Real (Fin n))
(M₂ := EuclideanSpace Real (Fin 1))).comp
(appendAffineEquiv n 1).symm.linear.toLinearMap
let M : AffineSubspace Real (EuclideanSpace Real (Fin (n + 1))) :=
AffineSubspace.mk' (appendAffineEquiv n 1 (0, zα)) (LinearMap.ker φ)
appendAffineEquiv n 1 (y, zα) ∈ (M : Set _)
The horizontal slice of the embedded epigraph is the image of the ≤-sublevel set.
open scoped Topology
open scoped Pointwise
lemma levelSets_horizontal_slice_image {n : Nat} {f : (Fin n → Real) → EReal} {α : Real} :
let sublevel : Set (EuclideanSpace Real (Fin n)) :=
{x | f (x : Fin n → Real) ≤ (α : EReal)}
let C : Set (EuclideanSpace Real (Fin (n + 1))) :=
(appendAffineEquiv n 1) ''
((fun p : (Fin n → Real) × Real =>
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm p.1,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => p.2))) ''
epigraph (S := (Set.univ : Set (Fin n → Real))) f)
let zα : EuclideanSpace Real (Fin 1) :=
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm (fun _ : Fin 1 => α)
let φ : EuclideanSpace Real (Fin (n + 1)) →ₗ[Real] EuclideanSpace Real (Fin 1) :=
(LinearMap.snd (R := Real) (M := EuclideanSpace Real (Fin n))
(M₂ := EuclideanSpace Real (Fin 1))).comp
(appendAffineEquiv n 1).symm.linear.toLinearMap
let M : AffineSubspace Real (EuclideanSpace Real (Fin (n + 1))) :=
AffineSubspace.mk' (appendAffineEquiv n 1 (0, zα)) (LinearMap.ker φ)
let g : EuclideanSpace Real (Fin n) → EuclideanSpace Real (Fin (n + 1)) :=
fun x =>
appendAffineEquiv n 1
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm (x : Fin n → Real), zα)
g '' sublevel = (M : Set _) ∩ C
The horizontal slice map is a homeomorphism onto the plane M.
open scoped Topology
open scoped Pointwise
lemma horizontal_slice_homeomorph {n : Nat} {α : Real} :
let zα : EuclideanSpace Real (Fin 1) :=
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm (fun _ : Fin 1 => α)
let φ : EuclideanSpace Real (Fin (n + 1)) →ₗ[Real] EuclideanSpace Real (Fin 1) :=
(LinearMap.snd (R := Real) (M := EuclideanSpace Real (Fin n))
(M₂ := EuclideanSpace Real (Fin 1))).comp
(appendAffineEquiv n 1).symm.linear.toLinearMap
let M : AffineSubspace Real (EuclideanSpace Real (Fin (n + 1))) :=
AffineSubspace.mk' (appendAffineEquiv n 1 (0, zα)) (LinearMap.ker φ)
let g : EuclideanSpace Real (Fin n) → EuclideanSpace Real (Fin (n + 1)) :=
fun x =>
appendAffineEquiv n 1
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm (x : Fin n → Real), zα)
∃ gM : EuclideanSpace Real (Fin n) ≃ₜ M,
∀ x, (gM x).1 = g x
Pull back closure and relative interior across the horizontal slice homeomorphism.
open scoped Topology
open scoped Pointwise
lemma levelSets_horizontal_slice_preimage_closure_ri {n : Nat} {f : (Fin n → Real) → EReal}
{α : Real} :
let sublevel : Set (EuclideanSpace Real (Fin n)) :=
{x | f (x : Fin n → Real) ≤ (α : EReal)}
let C : Set (EuclideanSpace Real (Fin (n + 1))) :=
(appendAffineEquiv n 1) ''
((fun p : (Fin n → Real) × Real =>
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm p.1,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => p.2))) ''
epigraph (S := (Set.univ : Set (Fin n → Real))) f)
let zα : EuclideanSpace Real (Fin 1) :=
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm (fun _ : Fin 1 => α)
let φ : EuclideanSpace Real (Fin (n + 1)) →ₗ[Real] EuclideanSpace Real (Fin 1) :=
(LinearMap.snd (R := Real) (M := EuclideanSpace Real (Fin n))
(M₂ := EuclideanSpace Real (Fin 1))).comp
(appendAffineEquiv n 1).symm.linear.toLinearMap
let M : AffineSubspace Real (EuclideanSpace Real (Fin (n + 1))) :=
AffineSubspace.mk' (appendAffineEquiv n 1 (0, zα)) (LinearMap.ker φ)
let g : EuclideanSpace Real (Fin n) → EuclideanSpace Real (Fin (n + 1)) :=
fun x =>
appendAffineEquiv n 1
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm (x : Fin n → Real), zα)
Convex Real C →
((M : Set _) ∩ euclideanRelativeInterior (n + 1) C).Nonempty →
closure sublevel = g ⁻¹' ((M : Set _) ∩ closure C) ∧
euclideanRelativeInterior n sublevel =
g ⁻¹' ((M : Set _) ∩ euclideanRelativeInterior (n + 1) C)
The non-strict sublevel lies in the closure of the strict sublevel.
open scoped Topology
open scoped Pointwise
lemma sublevel_subset_closure_strictSublevel {n : Nat} {f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) {α : Real}
(hα : iInf (fun x => f x) < (α : EReal)) :
let sublevel : Set (EuclideanSpace Real (Fin n)) :=
{x | f (x : Fin n → Real) ≤ (α : EReal)}
let strictSublevel : Set (EuclideanSpace Real (Fin n)) :=
{x | f (x : Fin n → Real) < (α : EReal)}
sublevel ⊆ closure strictSublevel
The strict inequality set is relatively open in affineSpan domf.
open scoped Topology
open scoped Pointwise
lemma ri_domf_lt_open {n : Nat} {f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) {α : Real} :
let domf : Set (EuclideanSpace Real (Fin n)) :=
(fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f
let s : Set (EuclideanSpace Real (Fin n)) :=
{x | x ∈ euclideanRelativeInterior n domf ∧ f (x : Fin n → Real) < (α : EReal)}
IsOpen {x : affineSpan Real domf | (x : EuclideanSpace Real (Fin n)) ∈ s}
A nonempty relatively open subset of an affine subspace has full affine span.
open scoped Topology
open scoped Pointwise
lemma affineSpan_eq_of_nonempty_relOpen {n : Nat}
(A : AffineSubspace Real (EuclideanSpace Real (Fin n)))
{s : Set (EuclideanSpace Real (Fin n))}
(hsA : s ⊆ (A : Set (EuclideanSpace Real (Fin n))))
(hsne : s.Nonempty)
(hsopen : IsOpen {x : A | (x : EuclideanSpace Real (Fin n)) ∈ s}) :
affineSpan Real s = A
The closure of the embedded epigraph equals the embedded epigraph of the closure.
open scoped Topology
open scoped Pointwise
lemma closure_embedded_epigraph_eq_convexFunctionClosure {n : Nat}
{f : (Fin n → Real) → EReal} (hbot : ∀ x, f x ≠ (⊥ : EReal)) :
let C : Set (EuclideanSpace Real (Fin (n + 1))) :=
(appendAffineEquiv n 1) ''
((fun p : (Fin n → Real) × Real =>
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm p.1,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => p.2))) ''
epigraph (S := (Set.univ : Set (Fin n → Real))) f)
closure C =
(appendAffineEquiv n 1) ''
((fun p : (Fin n → Real) × Real =>
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm p.1,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => p.2))) ''
epigraph (S := (Set.univ : Set (Fin n → Real))) (convexFunctionClosure f))
Theorem 7.6: Let f be a proper convex function, and let α ∈ ℝ with α > inf f. The convex level sets {x | f x ≤ α} and {x | f x < α} have the same closure and the same relative interior, namely {x | (cl f) x ≤ α} and {x ∈ ri (dom f) | f x < α}. Furthermore, they have the same dimension as dom f.
open scoped Topology
open scoped Pointwise
theorem properConvexFunction_levelSets_same_closure_ri_dim {n : Nat}
{f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
{α : Real} (hα : iInf (fun x => f x) < (α : EReal)) :
let sublevel : Set (EuclideanSpace Real (Fin n)) :=
{x | f (x : Fin n → Real) ≤ (α : EReal)}
let strictSublevel : Set (EuclideanSpace Real (Fin n)) :=
{x | f (x : Fin n → Real) < (α : EReal)}
let domf : Set (EuclideanSpace Real (Fin n)) :=
(fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f
closure sublevel =
{x : EuclideanSpace Real (Fin n) |
convexFunctionClosure f (x : Fin n → Real) ≤ (α : EReal)} ∧
closure strictSublevel =
{x : EuclideanSpace Real (Fin n) |
convexFunctionClosure f (x : Fin n → Real) ≤ (α : EReal)} ∧
euclideanRelativeInterior n sublevel =
{x | x ∈ euclideanRelativeInterior n domf ∧
f (x : Fin n → Real) < (α : EReal)} ∧
euclideanRelativeInterior n strictSublevel =
{x | x ∈ euclideanRelativeInterior n domf ∧
f (x : Fin n → Real) < (α : EReal)} ∧
Module.finrank Real (affineSpan Real sublevel).direction =
Module.finrank Real (affineSpan Real domf).direction ∧
Module.finrank Real (affineSpan Real strictSublevel).direction =
Module.finrank Real (affineSpan Real domf).direction
The infimum of the constant zero EReal function is zero.
open scoped Topology
open scoped Pointwise
lemma iInf_const_zero_ereal {n : Nat} :
iInf (fun _ : (Fin n → Real) => (0 : EReal)) = (0 : EReal)
The strict sublevel set for the constant zero function at zero is empty.
open scoped Topology
open scoped Pointwise
lemma strictSublevel_const_zero_eq_empty {n : Nat} :
{x : EuclideanSpace Real (Fin n) | (0 : EReal) < (0 : EReal) ∧ x = x} =
(∅ : Set (EuclideanSpace Real (Fin n)))
The right-hand side of the strict-sublevel closure formula is nonempty for the zero function.
open scoped Topology
open scoped Pointwise
lemma rhs_nonempty_via_closure_le_self {n : Nat} :
∃ x : EuclideanSpace Real (Fin n),
convexFunctionClosure (fun _ : (Fin n → Real) => (0 : EReal)) (x : Fin n → Real) ≤
(0 : EReal)
Text 7.0.17: The closure and relative interior formulas in Theorem 7.6 can fail when α = inf f.
open scoped Topology
open scoped Pointwise
theorem properConvexFunction_levelSets_formulas_fail_at_inf :
∃ (n : Nat) (f : (Fin n → Real) → EReal) (α : Real),
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f ∧
(α : EReal) = iInf (fun x => f x) ∧
let sublevel : Set (EuclideanSpace Real (Fin n)) :=
{x | f (x : Fin n → Real) ≤ (α : EReal)}
let strictSublevel : Set (EuclideanSpace Real (Fin n)) :=
{x | f (x : Fin n → Real) < (α : EReal)}
let domf : Set (EuclideanSpace Real (Fin n)) :=
(fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f
¬(closure sublevel =
{x : EuclideanSpace Real (Fin n) |
convexFunctionClosure f (x : Fin n → Real) ≤ (α : EReal)} ∧
closure strictSublevel =
{x : EuclideanSpace Real (Fin n) |
convexFunctionClosure f (x : Fin n → Real) ≤ (α : EReal)} ∧
euclideanRelativeInterior n sublevel =
{x | x ∈ euclideanRelativeInterior n domf ∧
f (x : Fin n → Real) < (α : EReal)} ∧
euclideanRelativeInterior n strictSublevel =
{x | x ∈ euclideanRelativeInterior n domf ∧
f (x : Fin n → Real) < (α : EReal)})
If α < iInf f, then α < f x for every x.
open scoped Topology
open scoped Pointwise
lemma lt_f_of_lt_iInf {n : Nat} {f : (Fin n → Real) → EReal} {α : Real}
(hα : (α : EReal) < iInf (fun x => f x)) (x : Fin n → Real) :
(α : EReal) < f x
A strict lower bound excludes membership in the ≤-sublevel set.
open scoped Topology
open scoped Pointwise
lemma not_mem_sublevel_of_lt {n : Nat} {f : (Fin n → Real) → EReal} {α : Real}
{x : Fin n → Real} (h : (α : EReal) < f x) :
¬ f x ≤ (α : EReal)
If no point satisfies the sublevel predicate, the sublevel set is empty.
open scoped Topology
open scoped Pointwise
lemma sublevel_eq_empty_of_forall_not_mem {n : Nat} {f : (Fin n → Real) → EReal} {α : Real}
(h : ∀ x : Fin n → Real, ¬ f x ≤ (α : EReal)) :
{x : Fin n → Real | f x ≤ (α : EReal)} = (∅ : Set (Fin n → Real))
Text 7.0.18: If α < inf f then {x | f x ≤ α} is empty and the formulas are trivial. The case α = inf f is more subtle and can also lead to failure of the formulas; see the example above.
open scoped Topology
open scoped Pointwise
theorem sublevel_eq_empty_of_lt_inf {n : Nat} (f : (Fin n → Real) → EReal) {α : Real}
(hα : (α : EReal) < iInf (fun x => f x)) :
{x : Fin n → Real | f x ≤ (α : EReal)} = (∅ : Set (Fin n → Real))
If α < ⊤, the strict sublevel already lies in the effective domain.
open scoped Topology
open scoped Pointwise
lemma strictSublevel_eq_domf_inter_of_lt_top {n : Nat} {f : (Fin n → Real) → EReal} {α : Real}
(hαtop : (α : EReal) < ⊤) :
{x : EuclideanSpace Real (Fin n) | f (x : Fin n → Real) < (α : EReal)} =
{x : EuclideanSpace Real (Fin n) |
x ∈ (fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f ∧
f (x : Fin n → Real) < (α : EReal)}
Relative openness means the relative interior agrees with the set.
open scoped Topology
open scoped Pointwise
lemma ri_domf_eq_of_relOpen {n : Nat} {domf : Set (EuclideanSpace Real (Fin n))}
(hopen : euclideanRelativelyOpen n domf) :
euclideanRelativeInterior n domf = domf
Closedness lets us replace cl f with f in sublevel sets.
open scoped Topology
open scoped Pointwise
lemma sublevel_eq_set_of_convexFunctionClosure_of_closed {n : Nat}
{f : (Fin n → Real) → EReal} {α : Real} (hfclosed : ClosedConvexFunction f)
(hbot : ∀ x, f x ≠ (⊥ : EReal)) :
{x : EuclideanSpace Real (Fin n) |
convexFunctionClosure f (x : Fin n → Real) ≤ (α : EReal)} =
{x : EuclideanSpace Real (Fin n) | f (x : Fin n → Real) ≤ (α : EReal)}
Corollary 7.6.1: If f is a closed proper convex function whose effective domain is relatively open (in particular if effectiveDomain Set.univ f is an affine set), then for inf f < α < +∞ one has ri {x | f x ≤ α} = {x | f x < α} and cl {x | f x < α} = {x | f x ≤ α}.
open scoped Topology
open scoped Pointwise
theorem closedProperConvexFunction_ri_sublevel_eq_strictSublevel_and_closure {n : Nat}
{f : (Fin n → Real) → EReal}
(hfclosed : ClosedConvexFunction f)
(hfproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hopen :
euclideanRelativelyOpen n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f))
{α : Real} (hαinf : iInf (fun x => f x) < (α : EReal))
(hαtop : (α : EReal) < ⊤) :
let sublevel : Set (EuclideanSpace Real (Fin n)) :=
{x | f (x : Fin n → Real) ≤ (α : EReal)}
let strictSublevel : Set (EuclideanSpace Real (Fin n)) :=
{x | f (x : Fin n → Real) < (α : EReal)}
euclideanRelativeInterior n sublevel = strictSublevel ∧
closure strictSublevel = sublevel
Auxiliary proof for Theorem 7.6.
open scoped Topology
open scoped Pointwise
lemma properConvexFunction_levelSets_same_closure_ri_dim_aux {n : Nat}
{f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
{α : Real} (hα : iInf (fun x => f x) < (α : EReal)) :
let sublevel : Set (EuclideanSpace Real (Fin n)) :=
{x | f (x : Fin n → Real) ≤ (α : EReal)}
let strictSublevel : Set (EuclideanSpace Real (Fin n)) :=
{x | f (x : Fin n → Real) < (α : EReal)}
let domf : Set (EuclideanSpace Real (Fin n)) :=
(fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f
closure sublevel =
{x : EuclideanSpace Real (Fin n) |
convexFunctionClosure f (x : Fin n → Real) ≤ (α : EReal)} ∧
closure strictSublevel =
{x : EuclideanSpace Real (Fin n) |
convexFunctionClosure f (x : Fin n → Real) ≤ (α : EReal)} ∧
euclideanRelativeInterior n sublevel =
{x | x ∈ euclideanRelativeInterior n domf ∧
f (x : Fin n → Real) < (α : EReal)} ∧
euclideanRelativeInterior n strictSublevel =
{x | x ∈ euclideanRelativeInterior n domf ∧
f (x : Fin n → Real) < (α : EReal)} ∧
Module.finrank Real (affineSpan Real sublevel).direction =
Module.finrank Real (affineSpan Real domf).direction ∧
Module.finrank Real (affineSpan Real strictSublevel).direction =
Module.finrank Real (affineSpan Real domf).direction§8 — Recession Cones and Unboundedness — Part 1
Lean authors: Zichen Wang, Wanli Ma, Shu Miao, Zaiwen Wen.
Definition 8.0.1. Let C be a nonempty convex set in ℝ^n. The set C recedes in the direction of a nonzero vector y if x + λ • y ∈ C for all x ∈ C and all λ ≥ 0.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
def Set.RecedesInDirection {n : Nat} (C : Set (EuclideanSpace Real (Fin n)))
(y : EuclideanSpace Real (Fin n)) : Prop :=
y ≠ 0 ∧ ∀ ⦃x⦄, x ∈ C → ∀ ⦃t : Real⦄, 0 ≤ t → x + t • y ∈ C
Definition 8.0.2. Let C be a non-empty convex set. The recession cone of C is the set 0^+ C = { y | x + λ • y ∈ C for all x ∈ C and all λ ≥ 0 }.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
def Set.recessionCone {E : Type*} [AddCommGroup E] [Module Real E] (C : Set E) : Set E :=
{ y | ∀ ⦃x⦄, x ∈ C → ∀ ⦃t : Real⦄, 0 ≤ t → x + t • y ∈ C }
Definition 8.0.3. Let C be a non-empty convex set in R^n. A direction D of R^n is called a direction of recession of C if C recedes in the direction D.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
def Set.IsDirectionOfRecession {n : Nat} (C : Set (EuclideanSpace Real (Fin n)))
(D : EuclideanSpace Real (Fin n)) : Prop :=
Set.RecedesInDirection C D
If C is closed under translation by y, then all integer translates stay in C.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
lemma add_mem_of_add_mem_nat_smul {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
{y : EuclideanSpace Real (Fin n)} (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 RealInnerProductSpace
open scoped Pointwise Topology
lemma mem_recessionCone_of_add_mem {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hconv : Convex Real C) {y : EuclideanSpace Real (Fin n)}
(hadd : ∀ x ∈ C, x + y ∈ C) : y ∈ Set.recessionCone C
Characterization of the recession cone via unit translation.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
lemma recessionCone_eq_add_mem {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hconv : Convex Real C) :
Set.recessionCone C = { y | ∀ x ∈ C, x + y ∈ C }
The recession cone of a convex set is convex.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
lemma recessionCone_convex {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hconv : Convex Real C) : Convex Real (Set.recessionCone C)
Positive scaling preserves membership in the recession cone.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
lemma recessionCone_smul_pos {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
{y : EuclideanSpace Real (Fin n)} (hy : y ∈ Set.recessionCone C) {t : Real} (ht : 0 < t) :
t • y ∈ Set.recessionCone C
Theorem 8.1. Let C be a non-empty convex set. The recession cone 0^+ C is a convex cone containing the origin, and it equals the set of vectors y such that C + y ⊆ C.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
theorem recessionCone_convexCone_and_eq {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(_hC : C.Nonempty) (hconv : Convex Real C) :
Convex Real (Set.recessionCone C) ∧
(∀ y ∈ Set.recessionCone C, ∀ t : Real, 0 < t → t • y ∈ Set.recessionCone C) ∧
(0 : EuclideanSpace Real (Fin n)) ∈ Set.recessionCone C ∧
Set.recessionCone C = { y | ∀ x ∈ C, x + y ∈ C }
A direction vector of an affine subspace belongs to its recession cone.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
lemma direction_subset_recessionCone_affineSubspace {n : Nat}
(M : AffineSubspace Real (EuclideanSpace Real (Fin n)))
{y : EuclideanSpace Real (Fin n)} (hy : y ∈ M.direction) :
y ∈ Set.recessionCone (M : Set (EuclideanSpace Real (Fin n)))
Any vector in the recession cone of a nonempty affine subspace lies in its direction.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
lemma recessionCone_subset_direction_affineSubspace {n : Nat}
(M : AffineSubspace Real (EuclideanSpace Real (Fin n)))
(hM : (M : Set (EuclideanSpace Real (Fin n))).Nonempty)
{y : EuclideanSpace Real (Fin n)}
(hy : y ∈ Set.recessionCone (M : Set (EuclideanSpace Real (Fin n)))) :
y ∈ M.direction
Corollary 8.1.1. If M ⊆ ℝ^n is a non-empty affine set and L is the linear subspace parallel to M, then 0^+ M = L.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
theorem recessionCone_affineSubspace_eq_direction {n : Nat}
(M : AffineSubspace Real (EuclideanSpace Real (Fin n)))
(hM : (M : Set (EuclideanSpace Real (Fin n))).Nonempty)
(L : Submodule Real (EuclideanSpace Real (Fin n)))
(hL : L = M.direction) :
Set.recessionCone (M : Set (EuclideanSpace Real (Fin n))) =
(L : Set (EuclideanSpace Real (Fin n)))
Inner product with a translated point splits into a sum.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
lemma inner_add_smul_eq {n : Nat} (x y b : EuclideanSpace Real (Fin n)) (t : ℝ) :
⟪x + t • y, b⟫ = ⟪x, b⟫ + t * ⟪y, b⟫
If a linear function is bounded below on t ≥ 0, its slope is nonnegative.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
lemma nonneg_of_forall_add_smul_ge {a b c : ℝ} (h : ∀ t ≥ 0, a + t * b ≥ c) : 0 ≤ b
Vectors in the recession cone of a nonempty intersection of half-spaces have nonnegative inner products with the defining normals.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
lemma recessionCone_subset_inner_ge_zero {n : Nat} {ι : Type*} (I : Set ι)
(b : ι → EuclideanSpace Real (Fin n)) (β : ι → Real)
(hC :
({x : EuclideanSpace Real (Fin n) | ∀ i, i ∈ I → ⟪x, b i⟫ ≥ β i} : Set _).Nonempty)
{y : EuclideanSpace Real (Fin n)}
(hy :
y ∈ Set.recessionCone
({x : EuclideanSpace Real (Fin n) | ∀ i, i ∈ I → ⟪x, b i⟫ ≥ β i} : Set _)) :
∀ i, i ∈ I → ⟪y, b i⟫ ≥ 0
The recession cone of a closed set is closed.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
lemma recessionCone_isClosed_of_closed {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hC : IsClosed C) : IsClosed (Set.recessionCone C)
Corollary 8.1.2. Let C = {x ∈ ℝ^n | ⟪x, b_i⟫ ≥ β_i, ∀ i ∈ I} be nonempty. Then 0^+ C = {y ∈ ℝ^n | ⟪y, b_i⟫ ≥ 0, ∀ i ∈ I}.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
theorem recessionCone_eq_inner_ge_zero {n : Nat} {ι : Type*} (I : Set ι)
(b : ι → EuclideanSpace Real (Fin n)) (β : ι → Real)
(hC :
({x : EuclideanSpace Real (Fin n) | ∀ i, i ∈ I → ⟪x, b i⟫ ≥ β i} : Set _).Nonempty) :
Set.recessionCone
({x : EuclideanSpace Real (Fin n) | ∀ i, i ∈ I → ⟪x, b i⟫ ≥ β i} : Set _) =
{y : EuclideanSpace Real (Fin n) | ∀ i, i ∈ I → ⟪y, b i⟫ ≥ 0}
Elements of the recession cone give limits with decreasing coefficients.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
lemma recessionCone_subset_seq_limits {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCne : C.Nonempty) {y : EuclideanSpace Real (Fin n)} (hy : y ∈ Set.recessionCone C) :
∃ (x : ℕ → EuclideanSpace Real (Fin n)) (r : ℕ → ℝ),
(∀ n, x n ∈ C) ∧ Antitone r ∧ Filter.Tendsto r Filter.atTop (𝓝 (0 : ℝ)) ∧
Filter.Tendsto (fun n => r n • x n) Filter.atTop (𝓝 y)
If a point in closure K has positive first coordinate, then it lies in K.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
lemma mem_K_of_mem_closure_K_first_pos {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} (hCclosed : IsClosed C) (hCconv : Convex Real C) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first v = 1 ∧ tail v ∈ C}
let K : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S : Set (EuclideanSpace Real (Fin (1 + n))))
∀ v, v ∈ closure K → 0 < first v → v ∈ K
Points in closure K with zero first coordinate yield recession directions.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
lemma tail_mem_recessionCone_of_mem_closure_K_first_zero {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} (hCclosed : IsClosed C) (hCconv : Convex Real C) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first v = 1 ∧ tail v ∈ C}
let K : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S : Set (EuclideanSpace Real (Fin (1 + n))))
∀ v, v ∈ closure K → first v = 0 → tail v ∈ Set.recessionCone C
The closure of K adds precisely the recession directions at first = 0.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
lemma closure_cone_eq_union_recession {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} (hCne : C.Nonempty) (hCclosed : IsClosed C)
(hCconv : Convex Real C) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first v = 1 ∧ tail v ∈ C}
let K : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S : Set (EuclideanSpace Real (Fin (1 + n))))
closure K = K ∪ {v | first v = 0 ∧ tail v ∈ Set.recessionCone C}
Reverse inclusion for the sequence characterization of the recession cone.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
lemma seq_limits_subset_recessionCone {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} (hCne : C.Nonempty) (hCclosed : IsClosed C)
(hCconv : Convex Real C) {y : EuclideanSpace Real (Fin n)}
(hy :
∃ (x : ℕ → EuclideanSpace Real (Fin n)) (r : ℕ → ℝ),
(∀ n, x n ∈ C) ∧ Antitone r ∧ Filter.Tendsto r Filter.atTop (𝓝 (0 : ℝ)) ∧
Filter.Tendsto (fun n => r n • x n) Filter.atTop (𝓝 y)) :
y ∈ Set.recessionCone C§8 — Recession Cones and Unboundedness — Part 2
Lean authors: Zichen Wang, Wanli Ma, Shu Miao, Zaiwen Wen.
Theorem 8.2. Let C be a non-empty closed convex set in ℝ^n. Then 0^+ C is closed, and it consists of all possible limits of sequences of the form λ_1 x_1, λ_2 x_2, ..., where x_i ∈ C and λ_i decreases to 0. In fact, for the convex cone K in ℝ^(n+1) generated by {(1, x) | x ∈ C} one has cl K = K ∪ {(0, x) | x ∈ 0^+ C}.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
theorem recessionCone_closed_and_limits_and_closure_cone {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} (hCne : C.Nonempty) (hCclosed : IsClosed C)
(hCconv : Convex Real C) :
IsClosed (Set.recessionCone C) ∧
Set.recessionCone C =
{y : EuclideanSpace Real (Fin n) |
∃ (x : ℕ → EuclideanSpace Real (Fin n)) (r : ℕ → ℝ),
(∀ n, x n ∈ C) ∧ Antitone r ∧ Filter.Tendsto r Filter.atTop (𝓝 (0 : ℝ)) ∧
Filter.Tendsto (fun n => r n • x n) Filter.atTop (𝓝 y)} ∧
(let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n));
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0;
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i));
let S : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first v = 1 ∧ tail v ∈ C};
let K : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S : Set (EuclideanSpace Real (Fin (1 + n))));
closure K = K ∪ {v | first v = 0 ∧ tail v ∈ Set.recessionCone C})
Sequence data associated with a half-line inside C.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma halfline_seq_data {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
{x0 y : EuclideanSpace Real (Fin n)}
(hx0 : ∀ t : Real, 0 ≤ t → x0 + t • y ∈ C) :
∃ (x : ℕ → EuclideanSpace Real (Fin n)) (r : ℕ → ℝ),
(∀ n, x n ∈ C) ∧ Antitone r ∧ Filter.Tendsto r Filter.atTop (𝓝 (0 : ℝ)) ∧
Filter.Tendsto (fun n => r n • x n) Filter.atTop (𝓝 y)
A half-line contained in C yields a recession direction.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma halfline_mem_recessionCone {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCne : C.Nonempty) (hCclosed : IsClosed C) (hCconv : Convex Real C)
{x0 y : EuclideanSpace Real (Fin n)}
(hx0 : ∀ t : Real, 0 ≤ t → x0 + t • y ∈ C) :
y ∈ Set.recessionCone C
Rays from points in the relative interior stay in the relative interior.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma ri_ray_mem {n : Nat} {C : Set (EuclideanSpace Real (Fin n))} (hCconv : Convex Real C)
{y : EuclideanSpace Real (Fin n)} (hy : y ∈ Set.recessionCone C)
{x : EuclideanSpace Real (Fin n)} (hx : x ∈ euclideanRelativeInterior n C) :
∀ t : Real, 0 ≤ t → x + t • y ∈ euclideanRelativeInterior n C
A ray property in ri C yields membership in 0^+ (ri C).
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionCone_ri_of_ri_ray {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
{y : EuclideanSpace Real (Fin n)}
(hri : ∀ x ∈ euclideanRelativeInterior n C, ∀ t : Real, 0 ≤ t →
x + t • y ∈ euclideanRelativeInterior n C) :
y ∈ Set.recessionCone (euclideanRelativeInterior n C)
Theorem 8.3. Let C be a non-empty closed convex set, and let y ≠ 0. If there exists x such that the half-line {x + λ y | λ ≥ 0} is contained in C, then the same holds for every x ∈ C, i.e. y ∈ 0^+ C. Moreover, for each x ∈ ri C, the half-line {x + λ y | λ ≥ 0} is contained in ri C, so y ∈ 0^+ (ri C).
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
theorem recessionCone_of_exists_halfline {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCne : C.Nonempty) (hCclosed : IsClosed C) (hCconv : Convex Real C)
{y : EuclideanSpace Real (Fin n)} (_hy : y ≠ 0)
(hex : ∃ x, ∀ t : Real, 0 ≤ t → x + t • y ∈ C) :
y ∈ Set.recessionCone C ∧
(∀ x ∈ euclideanRelativeInterior n C, ∀ t : Real, 0 ≤ t →
x + t • y ∈ euclideanRelativeInterior n C) ∧
y ∈ Set.recessionCone (euclideanRelativeInterior n C)
Characterization of recession directions of closure C via rays from a point in ri C.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionCone_closure_iff_ray_in_C {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} (hCne : C.Nonempty) (hCconv : Convex Real C)
{x : EuclideanSpace Real (Fin n)} (hx : x ∈ euclideanRelativeInterior n C) :
∀ y : EuclideanSpace Real (Fin n),
(y ∈ Set.recessionCone (closure C) ↔
∀ t : Real, 0 < t → x + t • y ∈ C)
Corollary 8.3.1. For any non-empty convex set C, one has 0^+ (ri C) = 0^+ (cl C). In fact, given any x ∈ ri C, one has y ∈ 0^+ (cl C) if and only if x + λ • y ∈ C for every λ > 0.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
theorem recessionCone_ri_eq_cl_and_characterization {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} (hCne : C.Nonempty) (hCconv : Convex Real C) :
Set.recessionCone (euclideanRelativeInterior n C) =
Set.recessionCone (closure C) ∧
(∀ x ∈ euclideanRelativeInterior n C, ∀ y : EuclideanSpace Real (Fin n),
(y ∈ Set.recessionCone (closure C) ↔
∀ t : Real, 0 < t → x + t • y ∈ C))
Elements of the recession cone scale back into C from the origin.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionCone_subset_inv_smul_set {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hC0 : (0 : EuclideanSpace Real (Fin n)) ∈ C) :
∀ y, y ∈ Set.recessionCone C → ∀ ε : ℝ, 0 < ε → (ε⁻¹) • y ∈ C
If all inverse scalings of y lie in C, then y is a recession direction.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma inv_smul_set_subset_recessionCone {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCclosed : IsClosed C) (hCconv : Convex Real C)
(hC0 : (0 : EuclideanSpace Real (Fin n)) ∈ C) :
∀ y, (∀ ε : ℝ, 0 < ε → (ε⁻¹) • y ∈ C) → y ∈ Set.recessionCone C
The inverse-scaling characterization equals the intersection of all positive scalings.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma inv_smul_set_eq_iInter_smul {n : Nat} {C : Set (EuclideanSpace Real (Fin n))} :
{ y : EuclideanSpace Real (Fin n) | ∀ ε : ℝ, 0 < ε → (ε⁻¹) • y ∈ C } =
⋂ (ε : ℝ) (_ : 0 < ε), (ε • C)
Corollary 8.3.2. If C is a closed convex set containing the origin, then 0^+ C = { y | ε⁻¹ • y ∈ C, ∀ ε > 0 } = ⋂ (ε > 0), ε • C.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
theorem recessionCone_eq_inv_smul_set_and_iInter {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} (hCclosed : IsClosed C) (hCconv : Convex Real C)
(hC0 : (0 : EuclideanSpace Real (Fin n)) ∈ C) :
Set.recessionCone C =
{ y : EuclideanSpace Real (Fin n) | ∀ ε : ℝ, 0 < ε → (ε⁻¹) • y ∈ C } ∧
{ y : EuclideanSpace Real (Fin n) | ∀ ε : ℝ, 0 < ε → (ε⁻¹) • y ∈ C } =
⋂ (ε : ℝ) (_ : 0 < ε), (ε • C)
A recession direction for the intersection yields half-lines in each set.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionCone_iInter_halfline {n : Nat} {ι : Type*}
{C : ι → Set (EuclideanSpace Real (Fin n))}
{x0 y : EuclideanSpace Real (Fin n)}
(hy : y ∈ Set.recessionCone (⋂ i, C i)) (hx0 : x0 ∈ ⋂ i, C i) :
∀ i, ∀ t : Real, 0 ≤ t → x0 + t • y ∈ C i
The recession cone of an intersection is contained in the intersection of recession cones.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionCone_iInter_subset {n : Nat} {ι : Type*}
(C : ι → Set (EuclideanSpace Real (Fin n)))
(hCclosed : ∀ i, IsClosed (C i)) (hCconv : ∀ i, Convex Real (C i))
(hCne : (⋂ i, C i).Nonempty) :
Set.recessionCone (⋂ i, C i) ⊆ ⋂ i, Set.recessionCone (C i)
Recession directions common to all sets are recession directions of the intersection.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma iInter_recessionCone_subset {n : Nat} {ι : Type*}
(C : ι → Set (EuclideanSpace Real (Fin n))) :
(⋂ i, Set.recessionCone (C i)) ⊆ Set.recessionCone (⋂ i, C i)
Corollary 8.3.3. If {C_i | i ∈ I} is an arbitrary collection of closed convex sets in ℝ^n whose intersection is not empty, then 0^+ (⋂ i, C_i) = ⋂ i, 0^+ C_i.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
theorem recessionCone_iInter_eq_iInter {n : Nat} {ι : Type*}
(C : ι → Set (EuclideanSpace Real (Fin n)))
(hCclosed : ∀ i, IsClosed (C i)) (hCconv : ∀ i, Convex Real (C i))
(hCne : (⋂ i, C i).Nonempty) :
Set.recessionCone (⋂ i, C i) = ⋂ i, Set.recessionCone (C i)
Nonempty preimage implies nonempty target set.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma preimage_nonempty_implies_C_nonempty {n m : Nat}
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m))
{C : Set (EuclideanSpace Real (Fin m))} :
(A ⁻¹' C).Nonempty → C.Nonempty
Recession cone of the preimage lies in the preimage of the recession cone.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionCone_preimage_linearMap_subset {n m : Nat}
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m))
{C : Set (EuclideanSpace Real (Fin m))} (hCclosed : IsClosed C) (hCconv : Convex Real C)
(hCne : (A ⁻¹' C).Nonempty) :
Set.recessionCone (A ⁻¹' C) ⊆ A ⁻¹' (Set.recessionCone C)
Preimage of the recession cone lies in the recession cone of the preimage.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma preimage_recessionCone_subset_recessionCone_preimage {n m : Nat}
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m))
{C : Set (EuclideanSpace Real (Fin m))} :
A ⁻¹' (Set.recessionCone C) ⊆ Set.recessionCone (A ⁻¹' C)
Corollary 8.3.4. Let A be a linear transformation from ℝ^n to ℝ^m, and let C be a closed convex set in ℝ^m such that A ⁻¹' C is nonempty. Then 0^+ (A ⁻¹' C) = A ⁻¹' (0^+ C).
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
theorem recessionCone_preimage_linearMap {n m : Nat}
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m))
{C : Set (EuclideanSpace Real (Fin m))} (hCclosed : IsClosed C) (hCconv : Convex Real C)
(hCne : (A ⁻¹' C).Nonempty) :
Set.recessionCone (A ⁻¹' C) = A ⁻¹' (Set.recessionCone C)
Bounded sets have only the zero recession direction.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionCone_subset_singleton_of_bounded {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} (hCne : C.Nonempty)
(hCbdd : Bornology.IsBounded C) :
Set.recessionCone C ⊆ {0}
An unbounded set admits a sequence with strictly increasing norms.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma exists_strictMono_norm_seq_of_unbounded {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} (hCunbdd : ¬ Bornology.IsBounded C) :
∃ x : ℕ → EuclideanSpace Real (Fin n),
(∀ n, x n ∈ C) ∧ StrictMono (fun n => ‖x n‖) ∧
(∀ n : ℕ, (n : ℝ) + 1 ≤ ‖x n‖)
An unbounded closed convex set has a nonzero recession direction.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma exists_nonzero_mem_recessionCone_of_unbounded {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} (hCne : C.Nonempty) (hCclosed : IsClosed C)
(hCconv : Convex Real C) (hCunbdd : ¬ Bornology.IsBounded C) :
∃ y, y ≠ 0 ∧ y ∈ Set.recessionCone C
Theorem 8.4. A non-empty closed convex set C in ℝ^n is bounded if and only if its recession cone 0^+ C consists of the zero vector alone.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
theorem bounded_iff_recessionCone_eq_singleton_zero {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} (hCne : C.Nonempty) (hCclosed : IsClosed C)
(hCconv : Convex Real C) :
Bornology.IsBounded C ↔ Set.recessionCone C = {0}
Recession cone of a nonempty intersection of two closed convex sets.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionCone_inter_eq {n : Nat}
{A B : Set (EuclideanSpace Real (Fin n))}
(hAclosed : IsClosed A) (hBclosed : IsClosed B)
(hAconv : Convex Real A) (hBconv : Convex Real B)
(hABne : (A ∩ B).Nonempty) :
Set.recessionCone (A ∩ B) = Set.recessionCone A ∩ Set.recessionCone B
Parallel nonempty affine subspaces have the same recession cone.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionCone_affine_parallel_eq {n : Nat}
{M M' : AffineSubspace Real (EuclideanSpace Real (Fin n))}
(hM : (M : Set (EuclideanSpace Real (Fin n))).Nonempty)
(hM' : (M' : Set (EuclideanSpace Real (Fin n))).Nonempty)
(hparallel : M'.direction = M.direction) :
Set.recessionCone (M' : Set (EuclideanSpace Real (Fin n))) =
Set.recessionCone (M : Set (EuclideanSpace Real (Fin n)))
An empty intersection is bounded.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma bounded_empty_intersection {n : Nat} {A B : Set (EuclideanSpace Real (Fin n))}
(h : ¬ (A ∩ B).Nonempty) : Bornology.IsBounded (A ∩ B)
Boundedness transfer via recession cones for parallel affine subspaces.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma boundedness_via_recessionCone_inter {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCclosed : IsClosed C) (hCconv : Convex Real C)
{M M' : AffineSubspace Real (EuclideanSpace Real (Fin n))}
(hMCne : ((M : Set (EuclideanSpace Real (Fin n))) ∩ C).Nonempty)
(hMCbdd : Bornology.IsBounded ((M : Set (EuclideanSpace Real (Fin n))) ∩ C))
(hM'Cne : ((M' : Set (EuclideanSpace Real (Fin n))) ∩ C).Nonempty)
(hparallel : M'.direction = M.direction) :
Set.recessionCone ((M' : Set (EuclideanSpace Real (Fin n))) ∩ C) = {0}
Corollary 8.4.1. Let C be a closed convex set, and let M be an affine set such that M ∩ C is non-empty and bounded. Then M' ∩ C is bounded for every affine set M' parallel to M.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
theorem bounded_inter_of_parallel_affine {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCclosed : IsClosed C) (hCconv : Convex Real C)
(M : AffineSubspace Real (EuclideanSpace Real (Fin n)))
(hMCne : ((M : Set (EuclideanSpace Real (Fin n))) ∩ C).Nonempty)
(hMCbdd : Bornology.IsBounded ((M : Set (EuclideanSpace Real (Fin n))) ∩ C)) :
∀ (M' : AffineSubspace Real (EuclideanSpace Real (Fin n))),
M'.direction = M.direction →
Bornology.IsBounded ((M' : Set (EuclideanSpace Real (Fin n))) ∩ C)§8 — Recession Cones and Unboundedness — Part 3
Lean authors: Zichen Wang, Wanli Ma, Shu Miao, Zaiwen Wen.
If every x ∈ C satisfies x + y ∈ C, then the translation of C by y stays in C.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma image_add_subset_of_add_mem {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
{y : EuclideanSpace Real (Fin n)} (h : ∀ x ∈ C, x + y ∈ C) :
Set.image (fun x => x + y) C ⊆ C
If every x ∈ C satisfies x - y ∈ C, then C is contained in its translation by y.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma subset_image_add_of_sub_mem {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
{y : EuclideanSpace Real (Fin n)} (h : ∀ x ∈ C, x + (-y) ∈ C) :
C ⊆ Set.image (fun x => x + y) C
If C is invariant under translation by y, then C is closed under adding y and -y.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma add_mem_of_image_eq {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
{y : EuclideanSpace Real (Fin n)}
(hEq : Set.image (fun x => x + y) C = C) :
(∀ x ∈ C, x + y ∈ C) ∧ (∀ x ∈ C, x + (-y) ∈ C)
Theorem 8.4.5. For any non-empty convex set C, the lineality space lin(C) = (-0^+ C) ∩ 0^+ C equals { y ∈ ℝ^n | C + y = C }.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
theorem linealitySpace_eq_translation {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCne : C.Nonempty) (hCconv : Convex Real C) :
(-Set.recessionCone C) ∩ Set.recessionCone C =
{ y : EuclideanSpace Real (Fin n) | Set.image (fun x => x + y) C = C }
Definiton 8.4.6. Let C be a non-empty convex set with lineality space L = lin(C). The rank of C is defined by rank(C) = dim(C) - dim(L).
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
noncomputable def Set.rank {n : Nat} (C : Set (EuclideanSpace Real (Fin n))) : Nat :=
Module.finrank Real (affineSpan Real C).direction -
Module.finrank Real (Submodule.span Real ((-Set.recessionCone C) ∩ Set.recessionCone C))
Proper convexity implies the domain is nonempty.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma properConvexFunctionOn_nonempty {n : Nat} {C : Set (Fin n → Real)}
{f : (Fin n → Real) → EReal} (hf : ProperConvexFunctionOn C f) : C.Nonempty
The difference-quotient set used in the recession formula is nonempty.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma diffQuotient_set_nonempty {n : Nat} {C : Set (Fin n → Real)}
{f : (Fin n → Real) → Real}
(hf : ProperConvexFunctionOn C (fun x => (f x : EReal))) :
∀ y : Fin n → Real,
( { r : EReal | ∃ x ∈ C, r = ((f (x + y) - f x : Real) : EReal) } ).Nonempty
The recession formula forces f0_plus 0 = 0.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionFunction_zero {n : Nat} {C : Set (Fin n → Real)}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn C (fun x => (f x : EReal)))
(hf0_plus :
∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal | ∃ x ∈ C, r = ((f (x + y) - f x : Real) : EReal) }) :
f0_plus (0 : Fin n → Real) = (0 : EReal)
The recession formula prevents f0_plus from taking the value ⊥.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionFunction_ne_bot {n : Nat} {C : Set (Fin n → Real)}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn C (fun x => (f x : EReal)))
(hf0_plus :
∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal | ∃ x ∈ C, r = ((f (x + y) - f x : Real) : EReal) }) :
∀ y : Fin n → Real, f0_plus y ≠ (⊥ : EReal)
The recession formula gives a pointwise inequality characterization.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionFunction_le_iff {n : Nat} {C : Set (Fin n → Real)}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → EReal}
(hf0_plus :
∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal | ∃ x ∈ C, r = ((f (x + y) - f x : Real) : EReal) }) :
∀ y : Fin n → Real, ∀ v : Real,
f0_plus y ≤ (v : EReal) ↔
∀ x ∈ C, ((f (x + y) - f x : Real) : EReal) ≤ (v : EReal)
The epigraph of the recession function is nonempty, witnessed at the origin.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionFunction_epigraph_nonempty {n : Nat} {C : Set (Fin n → Real)}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn C (fun x => (f x : EReal)))
(hf0_plus :
∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal | ∃ x ∈ C, r = ((f (x + y) - f x : Real) : EReal) }) :
Set.Nonempty (epigraph (Set.univ : Set (Fin n → Real)) f0_plus)
Proper convexity on the whole space gives real-valued convexity on Set.univ.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma convexOn_real_of_convexFunctionOn_ereal_univ {n : Nat} {C : Set (Fin n → Real)}
{f : (Fin n → Real) → Real} (hC : C = Set.univ)
(hf : ProperConvexFunctionOn C (fun x => (f x : EReal))) :
ConvexOn ℝ (Set.univ : Set (Fin n → Real)) f
A convex real function with bounded unit increments along y has bounded increments at all positive scales.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionFunction_ray_bound_real {n : Nat} {C : Set (Fin n → Real)}
{f : (Fin n → Real) → Real} (hC : C = Set.univ)
(hf : ProperConvexFunctionOn C (fun x => (f x : EReal))) :
∀ {y : Fin n → Real} {v : Real},
(∀ x, f (x + y) - f x ≤ v) →
∀ t > 0, ∀ x, f (x + t • y) - f x ≤ t * v
The ray bound rewritten in EReal form for the recession function.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionFunction_ray_bound {n : Nat} {C : Set (Fin n → Real)}
{f : (Fin n → Real) → Real} (hC : C = Set.univ)
(hf : ProperConvexFunctionOn C (fun x => (f x : EReal))) :
∀ {y : Fin n → Real} {v : Real},
(∀ x, ((f (x + y) - f x : Real) : EReal) ≤ (v : EReal)) →
∀ t > 0, ∀ x,
((f (x + t • y) - f x : Real) : EReal) ≤ ((t * v : Real) : EReal)
A pointwise bound on f0_plus is equivalent to bounds on all ray difference quotients.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionFunction_le_iff_ray {n : Nat} {C : Set (Fin n → Real)}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → EReal}
(hC : C = Set.univ)
(hf : ProperConvexFunctionOn C (fun x => (f x : EReal)))
(hf0_plus :
∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal | ∃ x ∈ C, r = ((f (x + y) - f x : Real) : EReal) }) :
∀ y : Fin n → Real, ∀ v : Real,
f0_plus y ≤ (v : EReal) ↔
∀ x : Fin n → Real, ∀ t : Real, 0 < t →
((f (x + t • y) - f x) / t : Real) ≤ v
Subadditivity of the recession formula when the domain is the whole space.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionFunction_subadditive_univ {n : Nat} {C : Set (Fin n → Real)}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → EReal}
(hC : C = Set.univ)
(hf0_plus :
∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal | ∃ x ∈ C, r = ((f (x + y) - f x : Real) : EReal) }) :
∀ y1 y2 : Fin n → Real, f0_plus (y1 + y2) ≤ f0_plus y1 + f0_plus y2
Convexity of the recession function on Set.univ from subadditivity and pos. homogeneity.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionFunction_convex_univ {n : Nat} {C : Set (Fin n → Real)}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → EReal}
(hC : C = Set.univ)
(hf : ProperConvexFunctionOn C (fun x => (f x : EReal)))
(hf0_plus :
∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal | ∃ x ∈ C, r = ((f (x + y) - f x : Real) : EReal) })
(hpos : PositivelyHomogeneous f0_plus) :
ConvexFunctionOn (Set.univ : Set (Fin n → Real)) f0_plus
Along a fixed ray, the difference quotient is monotone in the step length.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma differenceQuotient_monotone {n : Nat} {C : Set (Fin n → Real)}
{f : (Fin n → Real) → Real} (hC : C = Set.univ)
(hf : ProperConvexFunctionOn C (fun x => (f x : EReal))) :
∀ x y : Fin n → Real,
Monotone (fun t : {t : ℝ // 0 < t} => (f (x + (t : ℝ) • y) - f x) / (t : ℝ))
Rewrite the recession formula as an iSup over base points.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionFunction_iSup_formula {n : Nat} {C : Set (Fin n → Real)}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → EReal}
(hC : C = Set.univ)
(hf0_plus :
∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal | ∃ x ∈ C, r = ((f (x + y) - f x : Real) : EReal) }) :
∀ y : Fin n → Real,
f0_plus y =
⨆ x : Fin n → Real, ((f (x + y) - f x : Real) : EReal)
Closedness of f implies lower semicontinuity of the recession function.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionFunction_lsc_of_closed {n : Nat} {C : Set (Fin n → Real)}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → EReal}
(hC : C = Set.univ)
(hf0_plus :
∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal | ∃ x ∈ C, r = ((f (x + y) - f x : Real) : EReal) })
(hclosed : ClosedConvexFunction (fun x => (f x : EReal))) :
LowerSemicontinuous f0_plus
The embedded epigraph is closed when f is lower semicontinuous.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma closed_embedded_epigraph {n : Nat} {f : (Fin n → Real) → EReal}
(hls : LowerSemicontinuous f) :
let C : Set (EuclideanSpace Real (Fin (n + 1))) :=
(appendAffineEquiv n 1) ''
((fun p : (Fin n → Real) × Real =>
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm p.1,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => p.2))) ''
epigraph (S := (Set.univ : Set (Fin n → Real))) f)
IsClosed C
The embedded epigraph map sends affine rays to affine rays.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma embedded_epigraph_add_smul {n : Nat} :
let eN : (Fin n → Real) ≃ᵃ[Real] EuclideanSpace Real (Fin n) :=
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm.toLinearEquiv.toAffineEquiv
let e1 : Real ≃ᵃ[Real] EuclideanSpace Real (Fin 1) :=
((ContinuousLinearEquiv.funUnique (ι := Fin 1) (R := Real)
(M := Real)).symm.toLinearEquiv.trans
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm.toLinearEquiv).toAffineEquiv
let g : (Fin n → Real) × Real ≃ᵃ[Real]
(EuclideanSpace Real (Fin n) × EuclideanSpace Real (Fin 1)) :=
AffineEquiv.prodCongr eN e1
let embed : (Fin n → Real) × Real → EuclideanSpace Real (Fin (n + 1)) :=
fun p => appendAffineEquiv n 1 (g p)
let dir : (Fin n → Real) × Real → EuclideanSpace Real (Fin (n + 1)) :=
fun q => (appendAffineEquiv n 1).linear (g.linear q)
∀ p q : (Fin n → Real) × Real, ∀ t : Real,
embed (p + t • q) = embed p + t • dir q
A ray bound at one base point extends to all base points for closed convex f.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma ray_bound_extend_closed {n : Nat} {f : (Fin n → Real) → Real}
{x0 y : Fin n → Real} {v : Real}
(hclosed : ClosedConvexFunction (fun x => (f x : EReal)))
(hbound : ∀ t > 0, ((f (x0 + t • y) - f x0) / t : Real) ≤ v) :
∀ x : Fin n → Real, ∀ t : Real, 0 < t →
((f (x + t • y) - f x) / t : Real) ≤ v
Closed convexity makes the ray supremum independent of the base point.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma ray_sSup_eq_of_closed {n : Nat} {C : Set (Fin n → Real)}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → EReal}
(hC : C = Set.univ)
(hf : ProperConvexFunctionOn C (fun x => (f x : EReal)))
(hf0_plus :
∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal | ∃ x ∈ C, r = ((f (x + y) - f x : Real) : EReal) })
(hpos : PositivelyHomogeneous f0_plus)
(hclosed : ClosedConvexFunction (fun x => (f x : EReal))) :
∀ x y : Fin n → Real,
f0_plus y =
sSup { r : EReal |
∃ t : Real, 0 < t ∧ r = (((f (x + t • y) - f x) / t : Real) : EReal) }
Along a fixed ray, the difference quotient converges to f0_plus.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma ray_limit_monotone {n : Nat} {C : Set (Fin n → Real)}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → EReal}
(hC : C = Set.univ)
(hf : ProperConvexFunctionOn C (fun x => (f x : EReal)))
(hf0_plus :
∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal | ∃ x ∈ C, r = ((f (x + y) - f x : Real) : EReal) })
(hpos : PositivelyHomogeneous f0_plus)
(hclosed : ClosedConvexFunction (fun x => (f x : EReal))) :
∀ x y : Fin n → Real,
Filter.Tendsto
(fun t : Real => (((f (x + t • y) - f x) / t : Real) : EReal))
Filter.atTop (𝓝 (f0_plus y))§8 — Recession Cones and Unboundedness — Part 4
Lean authors: Zichen Wang, Wanli Ma, Shu Miao, Zaiwen Wen.
Definition 8.4.2. Let C be a non-empty convex set. The set lin(C) := (-0^+ C) ∩ 0^+ C is called the lineality space of C.
open scoped Pointwise
def Set.linealitySpace {n : Nat} (C : Set (EuclideanSpace Real (Fin n))) :
Set (EuclideanSpace Real (Fin n)) :=
(-Set.recessionCone C) ∩ Set.recessionCone C
Definition 8.4.3. The directions of the vectors y in the lineality space of C are called directions in which C is linear.
open scoped Pointwise
def Set.IsLinearDirection {n : Nat} (C : Set (EuclideanSpace Real (Fin n)))
(y : EuclideanSpace Real (Fin n)) : Prop :=
y ∈ Set.linealitySpace C
Definition 8.4.4. The dimension of the lineality space lin(C) is called the lineality of C.
open scoped Pointwise
def Set.lineality {n : Nat} (C : Set (EuclideanSpace Real (Fin n))) : Nat :=
Module.finrank Real (Submodule.span Real (Set.linealitySpace C))
The lineality space of a convex set is a submodule.
open scoped Pointwise
lemma linealitySpace_isSubmodule {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCconv : Convex Real C) :
∃ L : Submodule Real (EuclideanSpace Real (Fin n)), (L : Set _) = Set.linealitySpace C
The lineality space is contained in the direction of the affine span.
open scoped Pointwise
lemma linealitySpace_subset_direction_affineSpan {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCne : C.Nonempty) (hCconv : Convex Real C) :
Set.linealitySpace C ⊆ (affineSpan Real C).direction
Rank zero forces the affine span direction to be the lineality submodule.
open scoped Pointwise
lemma direction_eq_linealitySpace_of_rank_eq_zero {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCne : C.Nonempty) (hCconv : Convex Real C) (hrank : Set.rank C = 0) :
∃ L : Submodule Real (EuclideanSpace Real (Fin n)),
(L : Set _) = Set.linealitySpace C ∧ (affineSpan Real C).direction = L
Rank zero implies the set is a translate of its lineality space.
open scoped Pointwise
lemma C_eq_translate_linealitySpace {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCne : C.Nonempty) (hCconv : Convex Real C) (hrank : Set.rank C = 0) :
∃ x0 ∈ C, C = (fun v => x0 + v) '' Set.linealitySpace C
An affine subspace has rank zero.
open scoped Pointwise
lemma rank_affineSubspace_eq_zero {n : Nat}
(M : AffineSubspace Real (EuclideanSpace Real (Fin n)))
(hM : (M : Set (EuclideanSpace Real (Fin n))).Nonempty) :
Set.rank (M : Set (EuclideanSpace Real (Fin n))) = 0
Theorem 8.4.7. Let C be a non-empty convex set. Then C has rank 0 if and only if C is an affine set.
open scoped Pointwise
theorem rank_eq_zero_iff_isAffineSet {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCne : C.Nonempty) (hCconv : Convex Real C) :
Set.rank C = 0 ↔
IsAffineSet n
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n) :
EuclideanSpace Real (Fin n) → Fin n → Real)
'' C)
A line is nonempty.
open scoped Pointwise
lemma nonempty_of_isLine {n : Nat} {L : Set (Fin n → Real)} (hL : IsLine n L) :
Set.Nonempty L
Rank equals dimension iff the lineality span has finrank zero.
open scoped Pointwise
lemma rank_eq_dim_iff_lineality_finrank_zero {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCne : C.Nonempty) (hCconv : Convex Real C) :
Set.rank C = Module.finrank Real (affineSpan Real C).direction ↔
Module.finrank Real (Submodule.span Real (Set.linealitySpace C)) = 0
A translate of the span of a nonzero vector is a line.
open scoped Pointwise
lemma isLine_of_translate_span_singleton {n : Nat} {y : Fin n → Real} (hy : y ≠ 0)
(x0 : Fin n → Real) :
IsLine n (SetTranslate n (Submodule.span Real {y} : Set (Fin n → Real)) x0)
A nonzero lineality direction yields a line contained in the image.
open scoped Pointwise
lemma exists_line_of_nonzero_lineality {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCne : C.Nonempty) (hlin : ∃ y, y ≠ 0 ∧ y ∈ Set.linealitySpace C) :
∃ L : Set (Fin n → Real), IsLine n L ∧
L ⊆
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n) :
EuclideanSpace Real (Fin n) → Fin n → Real)
'' C)
A line contained in the image yields a nonzero lineality direction.
open scoped Pointwise
lemma nonzero_lineality_of_line {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCne : C.Nonempty) (hCclosed : IsClosed C) (hCconv : Convex Real C)
(hline :
∃ L : Set (Fin n → Real), IsLine n L ∧
L ⊆
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n) :
EuclideanSpace Real (Fin n) → Fin n → Real)
'' C)) :
∃ y : EuclideanSpace Real (Fin n), y ≠ 0 ∧ y ∈ Set.linealitySpace C
Theorem 8.4.8. The rank of a closed convex set equals its dimension if and only if the set contains no lines.
open scoped Pointwise
theorem rank_eq_dim_iff_no_lines {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCclosed : IsClosed C) (hCconv : Convex Real C) :
Set.rank C = Module.finrank Real (affineSpan Real C).direction ↔
¬ ∃ L : Set (Fin n → Real), IsLine n L ∧
L ⊆
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n) :
EuclideanSpace Real (Fin n) → Fin n → Real)
'' C)§8 — Recession Cones and Unboundedness — Part 5
Lean authors: Zichen Wang, Wanli Ma, Shu Miao, Zaiwen Wen.
Theorem 8.5. Let f be a proper convex function. The recession function f0^+ of f is a positively homogeneous proper convex function. For every vector y, one has (f0^+)(y) = sup { f(x + y) - f(x) | x ∈ dom f }. If f is closed, f0^+ is closed too, and for any x ∈ dom f, `(f0^+)(y) = sup_{λ > 0} (f(x + λ y) - f(x)) / λ = lim_{λ → ∞} (f(x + λ y) - f(x)) / λ`.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
theorem recessionFunction_properties {n : Nat}
{C : Set (Fin n → Real)}
{f : (Fin n → Real) → Real}
{f0_plus : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn C (fun x => (f x : EReal)))
(hC : C = Set.univ)
(hf0_plus :
∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal | ∃ x ∈ C, r = ((f (x + y) - f x : Real) : EReal) }) :
PositivelyHomogeneous f0_plus ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f0_plus ∧
(∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal | ∃ x ∈ C, r = ((f (x + y) - f x : Real) : EReal) }) ∧
(ClosedConvexFunction (fun x => (f x : EReal)) →
ClosedConvexFunction f0_plus ∧
∀ x ∈ C, ∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal |
∃ t : Real, 0 < t ∧ r = (((f (x + t • y) - f x) / t : Real) : EReal) } ∧
Filter.Tendsto
(fun t : Real => (((f (x + t • y) - f x) / t : Real) : EReal))
Filter.atTop (𝓝 (f0_plus y)))§8 — Recession Cones and Unboundedness — Part 6
Lean authors: Zichen Wang, Wanli Ma, Shu Miao, Zaiwen Wen.
Each recession difference is bounded above by f0_plus.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionFunction_diff_le_sup {n : Nat}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → EReal}
(hf0_plus :
∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal |
∃ x ∈ (Set.univ : Set (Fin n → Real)),
r = ((f (x + y) - f x : Real) : EReal) }) :
∀ x y : Fin n → Real, ((f (x + y) - f x : Real) : EReal) ≤ f0_plus y
The recession function controls increments of f.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionFunction_additive_majorant {n : Nat}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → EReal}
(hf0_plus :
∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal |
∃ x ∈ (Set.univ : Set (Fin n → Real)),
r = ((f (x + y) - f x : Real) : EReal) }) :
∀ x y : Fin n → Real, (f (x + y) : EReal) ≤ (f x : EReal) + f0_plus y
Any additive majorant bounds the recession function pointwise.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionFunction_le_of_additive_majorant {n : Nat}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → EReal}
(hf0_plus :
∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal |
∃ x ∈ (Set.univ : Set (Fin n → Real)),
r = ((f (x + y) - f x : Real) : EReal) })
(hne : ∀ x : Fin n → Real, (f x : EReal) ≠ (⊥ : EReal)) :
∀ h : (Fin n → Real) → EReal,
(∀ z x : Fin n → Real, (f z : EReal) ≤ (f x : EReal) + h (z - x)) →
f0_plus ≤ h
Corollary 8.5.1. Let f be a proper convex function. Then f0^+ is the least of the functions h such that f(z) ≤ f(x) + h(z - x) for all z and all x. The recession function of f can be viewed in terms of a closure construction when f is a closed proper convex function; the source text continues with "Let g be the positively ..." (truncated).
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
theorem recessionFunction_least_additive_majorant {n : Nat}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (fun x => (f x : EReal)))
(hf0_plus :
∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal |
∃ x ∈ (Set.univ : Set (Fin n → Real)),
r = ((f (x + y) - f x : Real) : EReal) }) :
(∀ z x : Fin n → Real, (f z : EReal) ≤ (f x : EReal) + f0_plus (z - x)) ∧
(∀ h : (Fin n → Real) → EReal,
(∀ z x : Fin n → Real, (f z : EReal) ≤ (f x : EReal) + h (z - x)) →
f0_plus ≤ h)
Convert a right-sided limit at 0 into a limit at +∞ via inversion.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma tendsto_rightScalarMultiple_comp_inv {n : Nat} {f : (Fin n → Real) → EReal}
{y : Fin n → Real} {l : Filter EReal} :
Filter.Tendsto (fun t : ℝ => rightScalarMultiple f t y) (𝓝[>] (0 : ℝ)) l ↔
Filter.Tendsto (fun t : ℝ => rightScalarMultiple f (t⁻¹) y) Filter.atTop l
Positive right scalar multiples satisfy the usual scaling formula.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma rightScalarMultiple_pos_rewrite {n : Nat} {f : (Fin n → Real) → EReal}
(hf : ConvexFunction f) {lam : Real} (hlam : 0 < lam) (x : Fin n → Real) :
rightScalarMultiple f lam x = ((lam : Real) : EReal) * f (lam⁻¹ • x)
A ray difference-quotient limit yields the inverse right scalar multiple limit.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma tendsto_rightScalarMultiple_inv_of_ray_limit {n : Nat}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → EReal} {y : Fin n → Real}
(hfconv : ConvexFunction (fun x => (f x : EReal)))
(hlim :
Filter.Tendsto
(fun t : Real => (((f (t • y) - f 0) / t : Real) : EReal))
Filter.atTop (𝓝 (f0_plus y))) :
Filter.Tendsto
(fun t : Real => rightScalarMultiple (fun x => (f x : EReal)) (t⁻¹) y)
Filter.atTop (𝓝 (f0_plus y))
Corollary 8.5.2. If f is a closed proper convex function, then (f0^+)(y) = lim_{λ ↑ 0} (f λ)(y) for every y ∈ dom f. If 0 ∈ dom f, this formula holds for every y ∈ ℝ^n.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
theorem recessionFunction_limit_rightScalarMultiple {n : Nat}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → EReal}
(hfclosed : ClosedConvexFunction (fun x => (f x : EReal)))
(hfproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (fun x => (f x : EReal)))
(hf0_plus :
∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal |
∃ x ∈ (Set.univ : Set (Fin n → Real)),
r = ((f (x + y) - f x : Real) : EReal) }) :
(∀ y ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) (fun x => (f x : EReal)),
Filter.Tendsto
(fun t : ℝ => rightScalarMultiple (fun x => (f x : EReal)) t y)
(𝓝[>] (0 : ℝ)) (𝓝 (f0_plus y))) ∧
((0 : Fin n → Real) ∈
effectiveDomain (Set.univ : Set (Fin n → Real)) (fun x => (f x : EReal)) →
∀ y : Fin n → Real,
Filter.Tendsto
(fun t : ℝ => rightScalarMultiple (fun x => (f x : EReal)) t y)
(𝓝[>] (0 : ℝ)) (𝓝 (f0_plus y)))
Antitone along all base points is equivalent to nonpositive forward steps.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma ray_antitone_all_iff_nonpos_step {n : Nat} {f : (Fin n → Real) → EReal} {y : Fin n → Real} :
(∀ x : Fin n → Real, Antitone (fun t : ℝ => f (x + t • y))) ↔
∀ x : Fin n → Real, ∀ t : ℝ, 0 ≤ t → f (x + t • y) ≤ f x
Antitone along all base points is equivalent to nonpositive forward steps (real-valued).
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma ray_antitone_all_iff_nonpos_step_real {n : Nat} {f : (Fin n → Real) → Real}
{y : Fin n → Real} :
(∀ x : Fin n → Real, Antitone (fun t : ℝ => f (x + t • y))) ↔
∀ x : Fin n → Real, ∀ t : ℝ, 0 ≤ t → f (x + t • y) ≤ f x
Antitone along all base points is equivalent to nonpositive ray difference quotients.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma ray_antitone_all_iff_diffquotient_le_zero_real {n : Nat}
{f : (Fin n → Real) → Real} {y : Fin n → Real} :
(∀ x : Fin n → Real, Antitone (fun t : ℝ => f (x + t • y))) ↔
∀ x : Fin n → Real, ∀ t : ℝ, 0 < t →
((f (x + t • y) - f x) / t : Real) ≤ 0
Finite liminf along a ray forces antitone behavior of the ray.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma ray_antitone_of_liminf_lt_top_real {n : Nat} {f : (Fin n → Real) → Real}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (fun x => (f x : EReal)))
{x y : Fin n → Real} :
Filter.liminf (fun t : ℝ => (f (x + t • y) : EReal)) Filter.atTop < (⊤ : EReal) →
Antitone (fun t : ℝ => f (x + t • y))
All ray functions are antitone iff the recession function is nonpositive.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma ray_antitone_all_iff_f0plus_le_zero_real {n : Nat}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (fun x => (f x : EReal)))
(hf0_plus :
∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal | ∃ x ∈ (Set.univ : Set (Fin n → Real)),
r = ((f (x + y) - f x : Real) : EReal) }) {y : Fin n → Real} :
(∀ x : Fin n → Real, Antitone (fun t : ℝ => f (x + t • y))) ↔ f0_plus y ≤ 0
For closed convex f, a ray antitone at one base point extends to all base points.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma ray_antitone_extend_closed_real {n : Nat} {f : (Fin n → Real) → Real} {y : Fin n → Real}
(hclosed : ClosedConvexFunction (fun x => (f x : EReal)))
(hx :
∃ x ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) (fun x => (f x : EReal)),
Antitone (fun t : ℝ => f (x + t • y))) :
∀ x : Fin n → Real, Antitone (fun t : ℝ => f (x + t • y))
Theorem 8.6. Let f be a proper convex function, and let y be a vector. If lim_{λ → +∞} inf f (x + λ • y) < +∞ for a given x, then f (x + λ • y) is non-increasing in λ for all real λ. This property holds for every x iff (f0^+)(y) ≤ 0. When f is closed, this property holds for every x if it holds for some x ∈ dom f.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
theorem recessionFunction_ray_antitone_iff {n : Nat}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (fun x => (f x : EReal)))
(hf0_plus :
∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal | ∃ x ∈ (Set.univ : Set (Fin n → Real)),
r = ((f (x + y) - f x : Real) : EReal) }) {y : Fin n → Real} :
(∀ x : Fin n → Real,
Filter.liminf (fun t : ℝ => (f (x + t • y) : EReal)) Filter.atTop < (⊤ : EReal) →
Antitone (fun t : ℝ => f (x + t • y))) ∧
((∀ x : Fin n → Real, Antitone (fun t : ℝ => f (x + t • y))) ↔ f0_plus y ≤ 0) ∧
(ClosedConvexFunction (fun x => (f x : EReal)) →
(∃ x ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) (fun x => (f x : EReal)),
Antitone (fun t : ℝ => f (x + t • y))) →
∀ x : Fin n → Real, Antitone (fun t : ℝ => f (x + t • y)))
Antitone on the reverse ray yields monotone on the forward ray.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma ray_monotone_of_antitone_neg {n : Nat} {f : (Fin n → Real) → Real} {y : Fin n → Real} :
(∀ x : Fin n → Real, Antitone (fun t : ℝ => f (x + t • (-y)))) →
∀ x : Fin n → Real, Monotone (fun t : ℝ => f (x + t • y))
Constancy along rays is equivalent to antitone behavior for y and -y.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma ray_const_iff_antitone_antitone_neg {n : Nat} {f : (Fin n → Real) → Real} {y : Fin n → Real} :
(∀ x : Fin n → Real, ∀ s t : ℝ, f (x + s • y) = f (x + t • y)) ↔
(∀ x : Fin n → Real, Antitone (fun t : ℝ => f (x + t • y))) ∧
(∀ x : Fin n → Real, Antitone (fun t : ℝ => f (x + t • (-y))))
A uniform upper bound on a ray forces a finite liminf.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma liminf_lt_top_of_ray_bounded_above {n : Nat} {f : (Fin n → Real) → Real}
{x y : Fin n → Real} {α : Real}
(hbound : ∀ t : ℝ, (f (x + t • y) : EReal) ≤ (α : EReal)) :
Filter.liminf (fun t : ℝ => (f (x + t • y) : EReal)) Filter.atTop < (⊤ : EReal)
Corollary 8.6.1. Let f be a proper convex function and y a vector. The function λ ↦ f (x + λ • y) is constant on ℝ for every x if and only if (f0^+)(y) ≤ 0 and (f0^+)(-y) ≤ 0. When f is closed, this condition holds if there exists x and α such that f (x + λ • y) ≤ α for all real λ.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
theorem recessionFunction_ray_const_iff {n : Nat}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (fun x => (f x : EReal)))
(hf0_plus :
∀ y : Fin n → Real,
f0_plus y =
sSup { r : EReal | ∃ x ∈ (Set.univ : Set (Fin n → Real)),
r = ((f (x + y) - f x : Real) : EReal) }) {y : Fin n → Real} :
((∀ x : Fin n → Real, ∀ s t : ℝ, f (x + s • y) = f (x + t • y)) ↔
f0_plus y ≤ 0 ∧ f0_plus (-y) ≤ 0) ∧
(ClosedConvexFunction (fun x => (f x : EReal)) →
(∃ x : Fin n → Real, ∃ α : Real, ∀ t : Real, f (x + t • y) ≤ (α : EReal)) →
f0_plus y ≤ 0 ∧ f0_plus (-y) ≤ 0)
The toReal of a convex EReal function is convex on any convex set where it is finite.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma convexOn_toReal_on_affine_of_finite {n : Nat} {f : (Fin n → Real) → EReal}
(hf : ConvexFunction f) {M : Set (Fin n → Real)} (hMconv : Convex ℝ M)
(hfinite : ∀ x ∈ M, f x ≠ (⊥ : EReal) ∧ f x ≠ (⊤ : EReal)) :
ConvexOn ℝ M (fun x => (f x).toReal)
A convex real function bounded above on ℝ is antitone.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma convexOn_real_antitone_of_bddAbove {g : ℝ → ℝ}
(hconv : ConvexOn ℝ Set.univ g) (hbounded : ∃ α, ∀ t, g t ≤ α) :
Antitone g
A convex real function bounded above on ℝ is constant.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma convexOn_real_const_of_bddAbove {g : ℝ → ℝ}
(hconv : ConvexOn ℝ Set.univ g) (hbounded : ∃ α, ∀ t, g t ≤ α) :
∀ s t, g s = g t
A bounded-above convex function is constant along any line in an affine set where it is finite.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma line_const_of_bddAbove_on_affine {n : Nat} {f : (Fin n → Real) → EReal}
(hf : ConvexFunction f) (M : AffineSubspace Real (Fin n → Real))
(hfinite :
∀ x ∈ (M : Set (Fin n → Real)), f x ≠ (⊥ : EReal) ∧ f x ≠ (⊤ : EReal))
(hbounded : ∃ α : Real, ∀ x ∈ (M : Set (Fin n → Real)), f x ≤ (α : EReal))
{x0 x : Fin n → Real} (hx0 : x0 ∈ (M : Set (Fin n → Real)))
(hx : x ∈ (M : Set (Fin n → Real))) :
∀ s t : ℝ, f (AffineMap.lineMap x0 x s) = f (AffineMap.lineMap x0 x t)§8 — Recession Cones and Unboundedness — Part 7
Lean authors: Zichen Wang, Wanli Ma, Shu Miao, Zaiwen Wen.
Corollary 8.6.2. A convex function f is constant on any affine set M where it is finite and bounded above.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
theorem convexFunction_const_on_affine_of_finite_boundedAbove {n : Nat}
{f : (Fin n → Real) → EReal} (hf : ConvexFunction f)
(M : AffineSubspace Real (Fin n → Real))
(hfinite :
∀ x ∈ (M : Set (Fin n → Real)), f x ≠ (⊥ : EReal) ∧ f x ≠ (⊤ : EReal))
(hbounded : ∃ α : Real, ∀ x ∈ (M : Set (Fin n → Real)), f x ≤ (α : EReal)) :
∃ c : EReal, ∀ x ∈ (M : Set (Fin n → Real)), f x = c
Definiton 8.7.0. Let f be a proper convex function on ℝ^n and let f0^+ be its recession function. The constancy space of f is the set `{ y ∈ ℝ^n | (f0^+)(y) ≤ 0 and (f0^+)(-y) ≤ 0 }`.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
def Function.constancySpace {n : Nat} (f0_plus : EuclideanSpace Real (Fin n) → EReal) :
Set (EuclideanSpace Real (Fin n)) :=
{ y | f0_plus y ≤ 0 ∧ f0_plus (-y) ≤ 0 }
If all rays are antitone, then any sublevel set recedes in direction y.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionCone_sublevel_of_antitone {n : Nat} {f : EuclideanSpace Real (Fin n) → EReal}
{r : Real} {y : EuclideanSpace Real (Fin n)} :
(∀ x : EuclideanSpace Real (Fin n), Antitone (fun t : ℝ => f (x + t • y))) →
y ∈ Set.recessionCone {x : EuclideanSpace Real (Fin n) | f x ≤ (r : EReal)}
If the recession cone is given by f0_plus y ≤ 0, its lineality space is the constancy space of f0_plus.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma lineality_eq_constancySpace_of_recessionCone_eq {n : Nat}
{S : Set (EuclideanSpace Real (Fin n))}
{f0_plus : EuclideanSpace Real (Fin n) → EReal}
(hEq : Set.recessionCone S = {y : EuclideanSpace Real (Fin n) | f0_plus y ≤ 0}) :
(-Set.recessionCone S) ∩ Set.recessionCone S = Function.constancySpace f0_plus
If x is in a sublevel set and y in its recession cone, the ray liminf is finite.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma liminf_lt_top_of_recessionCone_sublevel {n : Nat}
{f : EuclideanSpace Real (Fin n) → Real} {r : Real}
{x y : EuclideanSpace Real (Fin n)}
(hx : (f x : EReal) ≤ (r : EReal))
(hy :
y ∈ Set.recessionCone {x : EuclideanSpace Real (Fin n) | (f x : EReal) ≤ (r : EReal)}) :
Filter.liminf (fun t : ℝ => (f (x + t • y) : EReal)) Filter.atTop < (⊤ : EReal)
Membership in a sublevel recession cone forces all rays to be antitone.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma antitone_all_of_mem_recessionCone_sublevel {n : Nat}
{f : EuclideanSpace Real (Fin n) → Real} {f0_plus : EuclideanSpace Real (Fin n) → EReal}
(hfclosed :
ClosedConvexFunction
(fun x => (f ((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm x) : EReal)))
(hfproper :
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real))
(fun x => (f ((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm x) : EReal)))
(hf0_plus :
∀ y : Fin n → Real,
f0_plus ((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm y) =
sSup { r : EReal | ∃ x ∈ (Set.univ : Set (Fin n → Real)),
r =
((f ((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm (x + y)) -
f ((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm x) : Real) :
EReal) })
{r : Real}
(hnonempty :
({x : EuclideanSpace Real (Fin n) | (f x : EReal) ≤ (r : EReal)} : Set _).Nonempty)
{y : EuclideanSpace Real (Fin n)}
(hy :
y ∈ Set.recessionCone {x : EuclideanSpace Real (Fin n) | (f x : EReal) ≤ (r : EReal)}) :
∀ x : EuclideanSpace Real (Fin n), Antitone (fun t : ℝ => f (x + t • y))
Nonempty sublevel sets share recession cone {y | f0_plus y ≤ 0}.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma recessionCone_sublevel_eq_f0plus_le_zero {n : Nat}
{f : EuclideanSpace Real (Fin n) → Real} {f0_plus : EuclideanSpace Real (Fin n) → EReal}
(hfclosed :
ClosedConvexFunction
(fun x => (f ((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm x) : EReal)))
(hfproper :
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real))
(fun x => (f ((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm x) : EReal)))
(hf0_plus :
∀ y : Fin n → Real,
f0_plus ((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm y) =
sSup { r : EReal | ∃ x ∈ (Set.univ : Set (Fin n → Real)),
r =
((f ((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm (x + y)) -
f ((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm x) : Real) :
EReal) })
{r : Real}
(hnonempty :
({x : EuclideanSpace Real (Fin n) | (f x : EReal) ≤ (r : EReal)} : Set _).Nonempty) :
Set.recessionCone {x : EuclideanSpace Real (Fin n) | (f x : EReal) ≤ (r : EReal)} =
{y : EuclideanSpace Real (Fin n) | f0_plus y ≤ 0}
Theorem 8.7. Let f be a closed proper convex function. Then all non-empty level sets {x | f(x) ≤ λ} have the same recession cone and the same lineality space, namely {y | (f0^+)(y) ≤ 0} and the constancy space of f, respectively.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
theorem levelSet_recessionCone_and_lineality_eq_constancySpace {n : Nat}
{f : EuclideanSpace Real (Fin n) → Real}
{f0_plus : EuclideanSpace Real (Fin n) → EReal}
(hfclosed :
ClosedConvexFunction
(fun x => (f ((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm x) : EReal)))
(hfproper :
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real))
(fun x => (f ((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm x) : EReal)))
(hf0_plus :
∀ y : Fin n → Real,
f0_plus ((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm y) =
sSup { r : EReal | ∃ x ∈ (Set.univ : Set (Fin n → Real)),
r =
((f ((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm (x + y)) -
f ((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm x) : Real) :
EReal) }) :
∀ {r : Real},
({x : EuclideanSpace Real (Fin n) | (f x : EReal) ≤ (r : EReal)} : Set _).Nonempty →
Set.recessionCone {x : EuclideanSpace Real (Fin n) | (f x : EReal) ≤ (r : EReal)} =
{y : EuclideanSpace Real (Fin n) | f0_plus y ≤ 0} ∧
(-Set.recessionCone {x : EuclideanSpace Real (Fin n) | (f x : EReal) ≤ (r : EReal)}) ∩
Set.recessionCone {x : EuclideanSpace Real (Fin n) | (f x : EReal) ≤ (r : EReal)} =
Function.constancySpace f0_plus
Sublevel sets of a closed convex function are closed and convex.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma sublevel_isClosed_convex_of_closedConvex {n : Nat}
{f : EuclideanSpace Real (Fin n) → Real}
(hfclosed :
ClosedConvexFunction
(fun x => (f ((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm x) : EReal)))
(r : Real) :
IsClosed ({x : EuclideanSpace Real (Fin n) | (f x : EReal) ≤ (r : EReal)}) ∧
Convex Real ({x : EuclideanSpace Real (Fin n) | (f x : EReal) ≤ (r : EReal)})
Corollary 8.7.1. Let f be a closed proper convex function. If the level set {x | f(x) ≤ λ} is non-empty and bounded for one λ, it is bounded for every λ.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
theorem levelSet_bounded_of_bounded_one {n : Nat}
{f : EuclideanSpace Real (Fin n) → Real}
(hfclosed :
ClosedConvexFunction
(fun x => (f ((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm x) : EReal)))
(hfproper :
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real))
(fun x => (f ((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm x) : EReal)))
{r0 : Real}
(hlevel :
({x : EuclideanSpace Real (Fin n) | f x ≤ (r0 : EReal)} : Set _).Nonempty)
(hbounded :
Bornology.IsBounded
({x : EuclideanSpace Real (Fin n) | f x ≤ (r0 : EReal)} : Set _)) :
∀ r : Real,
Bornology.IsBounded ({x : EuclideanSpace Real (Fin n) | f x ≤ (r : EReal)} : Set _)
Two-sided ray inequalities force affine behavior along the direction.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma ray_eq_of_ineq_pair {n : Nat} {f : (Fin n → Real) → Real} {y : Fin n → Real} {v : Real}
(hpos : ∀ x : Fin n → Real, ∀ t : Real, 0 ≤ t → f (x + t • y) ≤ f x + t * v)
(hneg : ∀ x : Fin n → Real, ∀ t : Real, 0 ≤ t → f (x + t • (-y)) ≤ f x + t * (-v)) :
∀ x : Fin n → Real, ∀ t : Real, f (x + t • y) = f x + t * v
Negation commutes with the embedded epigraph map.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma embedded_epigraph_neg {n : Nat} :
let eN : (Fin n → Real) ≃ᵃ[Real] EuclideanSpace Real (Fin n) :=
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm.toLinearEquiv.toAffineEquiv
let e1 : Real ≃ᵃ[Real] EuclideanSpace Real (Fin 1) :=
((ContinuousLinearEquiv.funUnique (ι := Fin 1) (R := Real) (M := Real)).symm.toLinearEquiv.trans
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm.toLinearEquiv).toAffineEquiv
let g : (Fin n → Real) × Real ≃ᵃ[Real]
(EuclideanSpace Real (Fin n) × EuclideanSpace Real (Fin 1)) :=
AffineEquiv.prodCongr eN e1
∀ y : Fin n → Real, ∀ v : Real,
(appendAffineEquiv n 1) (g (-y, -v)) = - (appendAffineEquiv n 1) (g (y, v))
Membership in the embedded epigraph recession cone is equivalent to ray inequalities.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma mem_recessionCone_embedded_epigraph_iff_ray_ineq {n : Nat}
{f : (Fin n → Real) → Real} {y : Fin n → Real} {v : Real} :
(let eN : (Fin n → Real) ≃ᵃ[Real] EuclideanSpace Real (Fin n) :=
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm.toLinearEquiv.toAffineEquiv
let e1 : Real ≃ᵃ[Real] EuclideanSpace Real (Fin 1) :=
((ContinuousLinearEquiv.funUnique (ι := Fin 1) (R := Real) (M := Real)).symm.toLinearEquiv.trans
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm.toLinearEquiv).toAffineEquiv
let g : (Fin n → Real) × Real ≃ᵃ[Real]
(EuclideanSpace Real (Fin n) × EuclideanSpace Real (Fin 1)) :=
AffineEquiv.prodCongr eN e1
let epi : Set (EuclideanSpace Real (Fin (n + 1))) :=
(appendAffineEquiv n 1) ''
(g '' epigraph (S := (Set.univ : Set (Fin n → Real)))
(fun x => (f x : EReal)))
let yv : EuclideanSpace Real (Fin (n + 1)) :=
(appendAffineEquiv n 1) (g (y, v))
yv ∈ Set.recessionCone epi) ↔
∀ x : Fin n → Real, ∀ t : Real, 0 ≤ t → f (x + t • y) ≤ f x + t * v
An affine ray determines the recession function value.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
lemma f0plus_eq_slope_of_affine_ray {n : Nat}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → Real}
(hf0_plus :
∀ y : Fin n → Real,
(f0_plus y : EReal) =
sSup { r : EReal | ∃ x ∈ (Set.univ : Set (Fin n → Real)),
r = ((f (x + y) - f x : Real) : EReal) })
(hclosed : ClosedConvexFunction (fun x => (f x : EReal))) :
∀ {x : Fin n → Real}
(_hx :
x ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) (fun x => (f x : EReal)))
{y : Fin n → Real} {a b : Real},
(∀ t : Real, f (x + t • y) = a * t + b) →
f0_plus y = a ∧ f0_plus (-y) = -a
Theorem 8.8. For a proper convex function f, the following conditions on a vector y and a real number v are equivalent: (a) f (x + λ • y) = f x + λ v for all x and λ ∈ ℝ; (b) (y, v) belongs to the lineality space of epi f; (c) -(f0^+)(-y) = (f0^+)(y) = v. When f is closed, y satisfies these conditions with v = (f0^+)(y) if there is some x ∈ dom f such that λ ↦ f (x + λ • y) is affine.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
theorem properConvexFunction_lineality_equiv {n : Nat}
{f : (Fin n → Real) → Real} {f0_plus : (Fin n → Real) → Real}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (fun x => (f x : EReal)))
(hf0_plus :
∀ y : Fin n → Real,
(f0_plus y : EReal) =
sSup { r : EReal | ∃ x ∈ (Set.univ : Set (Fin n → Real)),
r = ((f (x + y) - f x : Real) : EReal) }) :
(∀ y : Fin n → Real, ∀ v : Real,
((∀ x : Fin n → Real, ∀ t : Real, f (x + t • y) = f x + t * v) ↔
(let epi : Set (EuclideanSpace Real (Fin (n + 1))) :=
(appendAffineEquiv n 1) ''
((fun p : (Fin n → Real) × Real =>
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm p.1,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => p.2))) ''
epigraph (S := (Set.univ : Set (Fin n → Real)))
(fun x => (f x : EReal)));
let yv : EuclideanSpace Real (Fin (n + 1)) :=
(appendAffineEquiv n 1)
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm y,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => v));
yv ∈ (-Set.recessionCone epi) ∩ Set.recessionCone epi)) ∧
((let epi : Set (EuclideanSpace Real (Fin (n + 1))) :=
(appendAffineEquiv n 1) ''
((fun p : (Fin n → Real) × Real =>
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm p.1,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => p.2))) ''
epigraph (S := (Set.univ : Set (Fin n → Real)))
(fun x => (f x : EReal)));
let yv : EuclideanSpace Real (Fin (n + 1)) :=
(appendAffineEquiv n 1)
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm y,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => v));
yv ∈ (-Set.recessionCone epi) ∩ Set.recessionCone epi) ↔
(-(f0_plus (-y)) = f0_plus y ∧ f0_plus y = v))) ∧
(∀ y : Fin n → Real,
ClosedConvexFunction (fun x => (f x : EReal)) →
(∃ x ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) (fun x => (f x : EReal)),
∃ a b : Real, ∀ t : Real, f (x + t • y) = a * t + b) →
(∀ x : Fin n → Real, ∀ t : Real, f (x + t • y) = f x + t * f0_plus y) ∧
(let epi : Set (EuclideanSpace Real (Fin (n + 1))) :=
(appendAffineEquiv n 1) ''
((fun p : (Fin n → Real) × Real =>
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm p.1,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => p.2))) ''
epigraph (S := (Set.univ : Set (Fin n → Real)))
(fun x => (f x : EReal)));
let yv : EuclideanSpace Real (Fin (n + 1)) :=
(appendAffineEquiv n 1)
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm y,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => f0_plus y));
yv ∈ (-Set.recessionCone epi) ∩ Set.recessionCone epi) ∧
(-(f0_plus (-y)) = f0_plus y))
Definition 8.9.0. Let f be a proper convex function on ℝ^n with recession function f0^+. The set of vectors y such that (f0^+)(-y) = -(f0^+)(y) is called the lineality space of the proper convex function f.
open scoped RealInnerProductSpace
open scoped Pointwise Topology
open Metric
def Function.linealitySpace {n : Nat} (f0_plus : EuclideanSpace Real (Fin n) → EReal) :
Set (EuclideanSpace Real (Fin n)) :=
{ y | f0_plus (-y) = -(f0_plus y) }§8 — Recession Cones and Unboundedness — Part 8
Lean authors: Zichen Wang, Wanli Ma, Shu Miao, Zaiwen Wen.
Definition 8.9.0. The lineality space of f0_plus on ℝ^n.
open scoped Pointwise
def Function.linealitySpace' {n : Nat} (f0_plus : (Fin n → Real) → EReal) :
Set (Fin n → Real) :=
{ y | f0_plus (-y) = -(f0_plus y) }
Definition 8.9.1. The directions of the vectors in the lineality space of f are called directions in which f is affine.
open scoped Pointwise
def Function.IsAffineDirection {n : Nat} (f0_plus : (Fin n → Real) → EReal)
(y : Fin n → Real) : Prop :=
y ∈ Function.linealitySpace' f0_plus
Definition 8.9.2. The dimension of the lineality space of f is the lineality of f.
open scoped Pointwise
def Function.lineality {n : Nat} (f0_plus : (Fin n → Real) → EReal) : Nat :=
Module.finrank Real (Submodule.span Real (Function.linealitySpace' f0_plus))
Definition 8.9.2. The rank of f is defined to be the dimension of f minus the lineality of f.
open scoped Pointwise
def Function.rank {n : Nat} (f0_plus : (Fin n → Real) → EReal)
(dim_f : Nat) : Nat :=
dim_f - Function.lineality f0_plus
Rank zero forces dim_f to be at most the lineality.
open scoped Pointwise
lemma rank_eq_zero_le_lineality {n : Nat}
{f0_plus : (Fin n → Real) → EReal} {dim_f : Nat}
(h : Function.rank f0_plus dim_f = 0) :
dim_f ≤ Function.lineality f0_plus
Proper convexity implies the effective domain is nonempty.
open scoped Pointwise
lemma properConvexFunctionOn_effectiveDomain_nonempty {n : Nat}
{f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
Set.Nonempty (effectiveDomain (Set.univ : Set (Fin n → Real)) f)
Points outside the affine span of the effective domain are not in the effective domain.
open scoped Pointwise
lemma not_mem_effectiveDomain_of_not_mem_affineSpan {n : Nat}
{f : (Fin n → Real) → EReal} {x : Fin n → Real}
(hx :
x ∉ affineSpan Real (effectiveDomain (Set.univ : Set (Fin n → Real)) f)) :
x ∉ effectiveDomain (Set.univ : Set (Fin n → Real)) f
Values of the recession function are finite on its lineality space.
open scoped Pointwise
lemma linealitySpace_value_real {n : Nat}
{f0_plus : (Fin n → Real) → EReal}
(hf0_plus : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f0_plus)
{y : Fin n → Real} (hy : y ∈ Function.linealitySpace' f0_plus) :
∃ v : ℝ, f0_plus y = (v : EReal)
Rank zero identifies the affine-span direction with the span of the lineality space, provided the inclusion is available.
open scoped Pointwise
lemma lineality_span_eq_affineSpan_direction_of_rank_zero {n : Nat}
{f : (Fin n → Real) → EReal} {f0_plus : (Fin n → Real) → EReal}
{dim_f : Nat}
(hdim :
dim_f =
Module.finrank Real
(affineSpan Real (effectiveDomain (Set.univ : Set (Fin n → Real)) f)).direction)
(hrank : Function.rank f0_plus dim_f = 0)
(hspan :
Submodule.span Real (Function.linealitySpace' f0_plus) ≤
(affineSpan Real (effectiveDomain (Set.univ : Set (Fin n → Real)) f)).direction) :
Submodule.span Real (Function.linealitySpace' f0_plus) =
(affineSpan Real (effectiveDomain (Set.univ : Set (Fin n → Real)) f)).direction
If x0 and x0 + y are in the effective domain, then y is a direction of its affine span.
open scoped Pointwise
lemma mem_direction_affineSpan_of_mem_effectiveDomain_add {n : Nat}
{f : (Fin n → Real) → EReal} {x0 y : Fin n → Real}
(hx0 : x0 ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) f)
(hx0y : x0 + y ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) f) :
y ∈
(affineSpan Real (effectiveDomain (Set.univ : Set (Fin n → Real)) f)).direction
Additive majorants keep lineality directions inside the effective domain.
open scoped Pointwise
lemma lineality_mem_effectiveDomain_of_additive_majorant {n : Nat}
{f : (Fin n → Real) → EReal} {f0_plus : (Fin n → Real) → EReal}
(hf0_plus : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f0_plus)
(hmaj : ∀ x y : Fin n → Real, f (x + y) ≤ f x + f0_plus y)
{x0 : Fin n → Real}
(hx0 : x0 ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) f) :
∀ y ∈ Function.linealitySpace' f0_plus,
x0 + y ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) f
If lineality directions preserve the effective domain, then their span lies in the direction of the affine span of the effective domain.
open scoped Pointwise
lemma lineality_span_le_direction_affineSpan {n : Nat}
{f : (Fin n → Real) → EReal} {f0_plus : (Fin n → Real) → EReal}
{x0 : Fin n → Real}
(hx0 : x0 ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) f)
(hmem :
∀ y ∈ Function.linealitySpace' f0_plus,
x0 + y ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) f) :
Submodule.span Real (Function.linealitySpace' f0_plus) ≤
(affineSpan Real (effectiveDomain (Set.univ : Set (Fin n → Real)) f)).direction
Membership in the affine span is equivalent to membership in the translated direction.
open scoped Pointwise
lemma mem_affineSpan_iff_vsub_mem_direction {n : Nat}
{s : Set (Fin n → Real)} {x0 x : Fin n → Real}
(hx0 : x0 ∈ affineSpan Real s) :
x ∈ affineSpan Real s ↔ x -ᵥ x0 ∈ (affineSpan Real s).direction
Lineality is symmetric under negation.
open scoped Pointwise
lemma lineality_neg {n : Nat} {f0_plus : (Fin n → Real) → EReal}
{y : Fin n → Real}
(hy : y ∈ Function.linealitySpace' f0_plus) :
(-y) ∈ Function.linealitySpace' f0_plus
Lineality directions give additivity of f0_plus.
open scoped Pointwise
lemma lineality_add {n : Nat} {f0_plus : (Fin n → Real) → EReal}
(hf0_plus : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f0_plus)
(hpos : PositivelyHomogeneous f0_plus)
{y1 y2 : Fin n → Real}
(hy1 : y1 ∈ Function.linealitySpace' f0_plus) :
f0_plus (y1 + y2) = f0_plus y1 + f0_plus y2
If the lineality space is nonempty, then f0_plus 0 = 0.
open scoped Pointwise
lemma lineality_zero_of_nonempty {n : Nat} {f0_plus : (Fin n → Real) → EReal}
(hf0_plus : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f0_plus)
(hpos : PositivelyHomogeneous f0_plus)
(hne : Set.Nonempty (Function.linealitySpace' f0_plus)) :
f0_plus (0 : Fin n → Real) = 0
The lineality relation extends to the linear span of the lineality space.
open scoped Pointwise
lemma lineality_span_neg_eq {n : Nat} {f0_plus : (Fin n → Real) → EReal}
(hf0_plus : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f0_plus)
(hpos : PositivelyHomogeneous f0_plus)
(hne : Set.Nonempty (Function.linealitySpace' f0_plus)) :
∀ y ∈ Submodule.span Real (Function.linealitySpace' f0_plus),
f0_plus (-y) = - f0_plus y
On the span of its lineality space, f0_plus agrees with a linear functional.
open scoped Pointwise
lemma linealitySpace_linear_on_span {n : Nat} {f0_plus : (Fin n → Real) → EReal}
(hf0_plus : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f0_plus)
(hpos : PositivelyHomogeneous f0_plus)
(hne : Set.Nonempty (Function.linealitySpace' f0_plus)) :
∃ aL :
(Submodule.span Real (Function.linealitySpace' f0_plus)) →ₗ[Real] Real,
∀ y : Submodule.span Real (Function.linealitySpace' f0_plus),
f0_plus (y : Fin n → Real) = (aL y : EReal)
Rank zero should force affine behavior on the affine span of the effective domain.
open scoped Pointwise
lemma properConvexFunction_rank_zero_affine_on_affineSpan {n : Nat}
{f : (Fin n → Real) → EReal} {f0_plus : (Fin n → Real) → EReal}
{dim_f : Nat}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hf0_plus : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f0_plus)
(hpos : PositivelyHomogeneous f0_plus)
(hmaj : ∀ x y : Fin n → Real, f (x + y) ≤ f x + f0_plus y)
(hdim :
dim_f =
Module.finrank Real
(affineSpan Real (effectiveDomain (Set.univ : Set (Fin n → Real)) f)).direction)
(hrank : Function.rank f0_plus dim_f = 0) :
∃ g : (Fin n → Real) →ᵃ[Real] Real,
∀ x ∈
(affineSpan Real (effectiveDomain (Set.univ : Set (Fin n → Real)) f) :
Set (Fin n → Real)),
f x = (g x : EReal)
Theorem 8.9.3. The proper convex functions of rank 0 are the partial affine functions, i.e. the functions which agree with an affine function along a certain affine set and are ⊤ elsewhere.
open scoped Pointwise
theorem properConvexFunction_rank_zero_partialAffine {n : Nat}
{f : (Fin n → Real) → EReal} {f0_plus : (Fin n → Real) → EReal}
{dim_f : Nat}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hf0_plus : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f0_plus)
(hpos : PositivelyHomogeneous f0_plus)
(hmaj : ∀ x y : Fin n → Real, f (x + y) ≤ f x + f0_plus y)
(hdim :
dim_f =
Module.finrank Real
(affineSpan Real (effectiveDomain (Set.univ : Set (Fin n → Real)) f)).direction)
(hrank : Function.rank f0_plus dim_f = 0) :
∃ (S : AffineSubspace Real (Fin n → Real)) (g : (Fin n → Real) →ᵃ[Real] Real),
(∀ x ∈ (S : Set (Fin n → Real)), f x = (g x : EReal)) ∧
(∀ x ∉ (S : Set (Fin n → Real)), f x = (⊤ : EReal))
Rank equals dimension iff the lineality is zero, provided the lineality is bounded.
open scoped Pointwise
lemma rank_eq_dim_iff_lineality_eq_zero {n : Nat}
{f0_plus : (Fin n → Real) → EReal} {dim_f : Nat}
(hle : Function.lineality f0_plus ≤ dim_f) :
Function.rank f0_plus dim_f = dim_f ↔ Function.lineality f0_plus = 0
Lineality zero iff there is no nonzero affine direction.
open scoped Pointwise
lemma lineality_eq_zero_iff_no_nonzero_affineDirection {n : Nat}
{f0_plus : (Fin n → Real) → EReal} :
Function.lineality f0_plus = 0 ↔
¬ ∃ y : Fin n → Real, y ≠ 0 ∧ Function.IsAffineDirection f0_plus y
Theorem 8.9.4. A closed proper convex function f has rank f = dim f if and only if it is not affine along any line in dom f.
open scoped Pointwise
theorem closedProperConvexFunction_rank_eq_dim_iff_not_affineLine {n : Nat}
{f : (Fin n → Real) → EReal} {f0_plus : (Fin n → Real) → EReal}
{dim_f : Nat}
(_hfclosed :
ClosedConvexFunction
(fun x => f ((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm x)))
(_hfproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hle : Function.lineality f0_plus ≤ dim_f) :
Function.rank f0_plus dim_f = dim_f ↔
¬ ∃ y : Fin n → Real, y ≠ 0 ∧ Function.IsAffineDirection f0_plus y
The lineality space of an indicator function consists of points in the set and its negation.
open scoped Pointwise
lemma indicatorFunction_linealitySpace'_iff {n : Nat} {S : Set (Fin n → Real)}
{y : Fin n → Real} :
y ∈ Function.linealitySpace' (indicatorFunction S) ↔ y ∈ S ∧ -y ∈ S
If the recession cone equals the set, then the lineality space is C ∩ -C.
open scoped Pointwise
lemma Set.linealitySpace_eq_inter_of_recessionCone_eq {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} (hcone : Set.recessionCone C = C) :
Set.linealitySpace C = C ∩ -C
The lineality space of the indicator function of the image of C is the image of C ∩ -C.
open scoped Pointwise
lemma linealitySpace_indicatorFunction_image_eq {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} :
Function.linealitySpace'
(indicatorFunction (Set.image (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)) C)) =
Set.image (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)) (C ∩ -C)
Finrank of the span is preserved under a linear equivalence.
open scoped Pointwise
lemma finrank_span_image_linearEquiv {V W : Type*}
[AddCommGroup V] [Module Real V] [AddCommGroup W] [Module Real W]
[FiniteDimensional Real V] [FiniteDimensional Real W]
(e : V ≃ₗ[Real] W) (S : Set V) :
Module.finrank Real (Submodule.span Real (Set.image e S)) =
Module.finrank Real (Submodule.span Real S)
Theorem 8.9.5. The rank of a convex set C equals the rank of its indicator function.
open scoped Pointwise
theorem convexSet_rank_eq_indicatorFunction_rank {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} (_hC : Convex Real C)
(hcone : Set.recessionCone C = C) :
Set.rank C =
Function.rank
(fun x =>
indicatorFunction
(C := Set.image (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)) C)
x)
(Module.finrank Real (affineSpan Real C).direction)§9 — Some Closedness Criteria — Part 1
Lean authors: Zichen Wang, Wanli Ma, Qiming Dai, Zaiwen Wen.
Proposition 9.0.0.1. For a convex set C and linear map A, one has ri (A '' C) = A '' ri C, while in general only closure (A '' C) ⊇ A '' closure C (Theorem 6.6); the text asks when equality holds and when the image of a closed convex set is closed.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
theorem linearMap_relativeInterior_image_eq_and_image_closure_subset (n m : Nat)
(C : Set (EuclideanSpace Real (Fin n))) (hC : Convex Real C)
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m)) :
A '' euclideanRelativeInterior n C = euclideanRelativeInterior m (A '' C) ∧
A '' closure C ⊆ closure (A '' C)
The projected epigraph is not closed, and the induced Ah is not lsc at 0.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma image_epigraph_not_closed_and_Ah_not_lsc :
let h : (Fin 2 → Real) → EReal :=
fun x =>
if 0 ≤ x 0 ∧ 0 ≤ x 1 then
(Real.exp (-(Real.sqrt (x 0 * x 1))) : EReal)
else
(⊤ : EReal)
let A : (Fin 2 → Real) → Real := fun x => x 0
let B : (Fin 2 → Real) × Real → Real × Real := fun p => (A p.1, p.2)
let Ah : Real → EReal :=
fun xi1 =>
if 0 < xi1 then (0 : EReal) else if xi1 = 0 then (1 : EReal) else (⊤ : EReal)
¬ IsClosed (Set.image B (epigraph (Set.univ : Set (Fin 2 → Real)) h)) ∧
¬ LowerSemicontinuousAt Ah 0
Image of the nonnegative orthant under x ↦ -sqrt(x0*x1) is (-∞, 0].
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma image_neg_sqrt_mul_nonneg :
(fun x : Fin 2 → Real => -(Real.sqrt (x 0 * x 1))) '' {x | 0 ≤ x 0 ∧ 0 ≤ x 1} =
Set.Iic (0 : Real)
Convexity of x ↦ exp(-sqrt(x0*x1)) on the nonnegative orthant.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma convexOn_exp_neg_sqrt_mul_nonneg :
ConvexOn ℝ {x : Fin 2 → Real | 0 ≤ x 0 ∧ 0 ≤ x 1}
(fun x => Real.exp (-(Real.sqrt (x 0 * x 1))))
Convexity of the extended-value function h on Set.univ.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma h_convexFunctionOn_univ :
ConvexFunctionOn (Set.univ : Set (Fin 2 → Real))
(fun x =>
if 0 ≤ x 0 ∧ 0 ≤ x 1 then
(Real.exp (-(Real.sqrt (x 0 * x 1))) : EReal)
else
(⊤ : EReal))
Properness of the function h.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma h_proper :
ProperConvexFunctionOn (Set.univ : Set (Fin 2 → Real))
(fun x =>
if 0 ≤ x 0 ∧ 0 ≤ x 1 then
(Real.exp (-(Real.sqrt (x 0 * x 1))) : EReal)
else
(⊤ : EReal))
Lower semicontinuity of the function h.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma h_lowerSemicontinuous :
LowerSemicontinuous
(fun x : Fin 2 → Real =>
if 0 ≤ x 0 ∧ 0 ≤ x 1 then
(Real.exp (-(Real.sqrt (x 0 * x 1))) : EReal)
else
(⊤ : EReal))
Example 9.0.0.2. Let h : R^2 -> (-infty, +infty] be given by h(x) = exp[-(xi1 * xi2)^(1/2)] for x = (xi1, xi2) ≥ 0, and h(x) = +infty otherwise. For the projection A (xi1, xi2) = xi1 and B (x, mu) = (A x, mu), the image of epi h under B need not be closed (even though h is closed proper convex), and the image function (Ah) satisfies (Ah)(xi1) = 0 for xi1 > 0, (Ah)(0) = 1, and (Ah)(xi1) = +infty for xi1 < 0, so (Ah) is not lower semicontinuous at 0.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
theorem projection_epigraph_not_closed_example :
let h : (Fin 2 → Real) → EReal :=
fun x =>
if 0 ≤ x 0 ∧ 0 ≤ x 1 then
(Real.exp (-(Real.sqrt (x 0 * x 1))) : EReal)
else
(⊤ : EReal)
let A : (Fin 2 → Real) → Real := fun x => x 0
let B : (Fin 2 → Real) × Real → Real × Real := fun p => (A p.1, p.2)
let Ah : Real → EReal :=
fun xi1 =>
if 0 < xi1 then (0 : EReal) else if xi1 = 0 then (1 : EReal) else (⊤ : EReal)
ClosedConvexFunction h ∧
ProperConvexFunctionOn (Set.univ : Set (Fin 2 → Real)) h ∧
¬ IsClosed (Set.image B (epigraph (Set.univ : Set (Fin 2 → Real)) h)) ∧
¬ LowerSemicontinuousAt Ah 0
If x0 minimizes h on the fiber A x = y, then the vertical section of F is the closed half-line starting at h x0.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma verticalLine_intersection_eq_closedHalfLine_of_minimizer
{X Y : Type*} {A : X → Y} {h : X → Real} {y : Y} {F : Set (Y × Real)}
(hF : F = {p | ∃ x, A x = p.1 ∧ h x ≤ p.2})
{x0 : X} (hx0 : A x0 = y ∧ ∀ z, A z = y → h x0 ≤ h z) :
(let verticalLine : Set (Y × Real) := {p | p.1 = y}
let closedHalfLine : Real → Set (Y × Real) := fun mu0 => {p | p.1 = y ∧ mu0 ≤ p.2}
verticalLine ∩ F = closedHalfLine (h x0))
If the vertical section is a closed half-line, then h attains its infimum on the fiber A x = y.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma minimizer_of_verticalLine_eq_closedHalfLine
{X Y : Type*} {A : X → Y} {h : X → Real} {y : Y} {F : Set (Y × Real)}
(hF : F = {p | ∃ x, A x = p.1 ∧ h x ≤ p.2}) {mu0 : Real}
(hline :
(let verticalLine : Set (Y × Real) := {p | p.1 = y}
let closedHalfLine : Real → Set (Y × Real) := fun mu0 => {p | p.1 = y ∧ mu0 ≤ p.2}
verticalLine ∩ F = closedHalfLine mu0)) :
∃ x, A x = y ∧ ∀ z, A z = y → h x ≤ h z
Proposition 9.0.0.3. The value (A h)(y) is the infimum of h on the affine set {x | A x = y}; the infimum is attained iff the vertical line {(y, mu) | mu ∈ Real} meets F in a closed half-line (or is empty), which would hold if F is closed and has no downward direction of recession.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma infimum_attained_iff_verticalLine_intersection_closedHalfLine
{X Y : Type*} (A : X → Y) (h : X → Real) (y : Y) (F : Set (Y × Real))
(hF : F = {p | ∃ x, A x = p.1 ∧ h x ≤ p.2}) (hne : ∃ x, A x = y) :
(∃ x, A x = y ∧ ∀ z, A z = y → h x ≤ h z) ↔
(let verticalLine : Set (Y × Real) := {p | p.1 = y}
let closedHalfLine : Real → Set (Y × Real) :=
fun mu0 => {p | p.1 = y ∧ mu0 ≤ p.2}
(∃ mu0, verticalLine ∩ F = closedHalfLine mu0) ∨ verticalLine ∩ F = ∅)
The projection of C is (0, +∞).
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma image_A_C_eq_Ioi :
Set.image (fun x : Fin 2 → Real => x 0)
{x : Fin 2 → Real | 0 < x 0 ∧ x 1 ≥ (x 0)⁻¹} = Set.Ioi (0 : Real)
Convexity of the set {x | 0 < x 0 ∧ x 1 ≥ (x 0)⁻¹}.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma convex_C :
Convex Real {x : Fin 2 → Real | 0 < x 0 ∧ x 1 ≥ (x 0)⁻¹}
Sequential closedness of the set {x | 0 < x 0 ∧ x 1 ≥ (x 0)⁻¹}.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma seqClosed_C :
IsSeqClosed {x : Fin 2 → Real | 0 < x 0 ∧ x 1 ≥ (x 0)⁻¹}
Example 9.0.0.4. The closed convex set C = {(xi1, xi2) | xi1 > 0, xi2 ≥ xi1^{-1}} in ℝ^2 has nonclosed projection A (xi1, xi2) = xi1; the difficulty is that C is asymptotic to a vertical line, and the recession cone condition rules out directions (0,1) and (0,-1).
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
theorem projection_closed_convex_not_closed_image_example :
let C : Set (Fin 2 → Real) := {x | 0 < x 0 ∧ x 1 ≥ (x 0)⁻¹}
let A : (Fin 2 → Real) → Real := fun x => x 0
IsClosed C ∧ Convex Real C ∧ ¬ IsClosed (Set.image A C)
Theorem 9.1. Let C be a non-empty convex set in ℝ^n, and let A be a linear transformation from ℝ^n to ℝ^m. Assume that every non-zero vector z ∈ 0+ (cl C) satisfying Az = 0 belongs to the lineality space of cl C. Then cl (A C) = A (cl C) and 0+ (A (cl C)) = A (0+ (cl C)). In particular, if C is closed, and z = 0 is the only z ∈ 0+ C such that Az = 0, then A C is closed.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma lineality_inter_kernel_nonempty
{n m : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m)) :
(Set.linealitySpace (closure C) ∩ (LinearMap.ker A : Set _)).Nonempty
Kernel intersection of the recession cone equals the kernel intersection of the lineality space under the kernel-lineality hypothesis.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma recessionCone_inter_kernel_eq_lineality_inter_kernel
{n m : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m))
(hlineal :
∀ z, z ≠ 0 → z ∈ Set.recessionCone (closure C) → A z = 0 →
z ∈ Set.linealitySpace (closure C)) :
Set.recessionCone (closure C) ∩ (LinearMap.ker A : Set _) =
Set.linealitySpace (closure C) ∩ (LinearMap.ker A : Set _)
Projecting along lin(closure C) ∩ ker A preserves the image of closure C.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma image_closure_eq_image_Lperp_inter
{n m : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCne : Set.Nonempty C) (hCconv : Convex Real C)
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m))
(L0 : Submodule Real (EuclideanSpace Real (Fin n)))
(hL0 : (L0 : Set _) = Set.linealitySpace (closure C)) :
let L : Submodule Real (EuclideanSpace Real (Fin n)) := L0 ⊓ LinearMap.ker A
A '' closure C = A '' ((Lᗮ : Set _) ∩ closure C)
The recession cone of a submodule is the submodule itself.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma recessionCone_submodule_eq {n : Nat}
(L : Submodule Real (EuclideanSpace Real (Fin n))) :
Set.recessionCone (L : Set (EuclideanSpace Real (Fin n))) =
(L : Set (EuclideanSpace Real (Fin n)))
The approximation sets near a closure point of the image are closed and bounded.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma Ceps_nonempty_closed_bounded
{n m : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCne : Set.Nonempty C) (hCconv : Convex Real C)
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m))
(L0 : Submodule Real (EuclideanSpace Real (Fin n)))
(hL0 : (L0 : Set _) = Set.linealitySpace (closure C))
(hlineal :
∀ z, z ≠ 0 → z ∈ Set.recessionCone (closure C) → A z = 0 →
z ∈ Set.linealitySpace (closure C))
{y : EuclideanSpace Real (Fin m)} (hy : y ∈ closure (A '' C))
{ε : ℝ} (hε : 0 < ε) :
let L : Submodule Real (EuclideanSpace Real (Fin n)) := L0 ⊓ LinearMap.ker A
let Dε : Set (EuclideanSpace Real (Fin n)) := A ⁻¹' Metric.closedBall y ε
let Cε : Set (EuclideanSpace Real (Fin n)) := ((Lᗮ : Set _) ∩ closure C) ∩ Dε
Cε.Nonempty ∧ IsClosed Cε ∧ Bornology.IsBounded Cε
Closure of the image equals the image of the closure under the kernel-lineality hypothesis.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma closure_image_eq_image_closure_of_kernel_lineality
{n m : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCne : Set.Nonempty C) (hCconv : Convex Real C)
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m))
(hlineal :
∀ z, z ≠ 0 → z ∈ Set.recessionCone (closure C) → A z = 0 →
z ∈ Set.linealitySpace (closure C)) :
closure (A '' C) = A '' closure C
The recession cone of a convex set is closed under addition.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma recessionCone_add_mem {n : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCconv : Convex Real C) {y₁ y₂ : EuclideanSpace Real (Fin n)}
(hy₁ : y₁ ∈ Set.recessionCone C) (hy₂ : y₂ ∈ Set.recessionCone C) :
y₁ + y₂ ∈ Set.recessionCone C§9 — Some Closedness Criteria — Part 2
Lean authors: Zichen Wang, Wanli Ma, Qiming Dai, Zaiwen Wen.
The lifted linear map preserves first coordinate and applies A to the tail.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma lifted_linearMap_first_tail {n m : Nat}
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m)) :
let coords₁ : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first₁ : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords₁ v 0
let tail₁ : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords₁ v (Fin.natAdd 1 i))
let coords₂ : EuclideanSpace Real (Fin (1 + m)) → (Fin (1 + m) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + m))
let first₂ : EuclideanSpace Real (Fin (1 + m)) → Real := fun v => coords₂ v 0
let tail₂ : EuclideanSpace Real (Fin (1 + m)) → EuclideanSpace Real (Fin m) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin m)).symm
(fun i => coords₂ v (Fin.natAdd 1 i))
let B : EuclideanSpace Real (Fin (1 + n)) →ₗ[Real] EuclideanSpace Real (Fin (1 + m)) :=
let lift :
(EuclideanSpace Real (Fin 1) × EuclideanSpace Real (Fin n)) →ₗ[Real]
(EuclideanSpace Real (Fin 1) × EuclideanSpace Real (Fin m)) :=
(LinearMap.fst (R := Real) (M := EuclideanSpace Real (Fin 1))
(M₂ := EuclideanSpace Real (Fin n))).prod
(A.comp (LinearMap.snd (R := Real) (M := EuclideanSpace Real (Fin 1))
(M₂ := EuclideanSpace Real (Fin n))))
(appendAffineEquiv 1 m).linear.toLinearMap.comp
(lift.comp (appendAffineEquiv 1 n).symm.linear.toLinearMap)
∀ v, first₂ (B v) = first₁ v ∧ tail₂ (B v) = A (tail₁ v)
The lifted map sends the cone over C to the cone over A '' C.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma image_cone_eq_cone_image {n m : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCconv : Convex Real C)
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m)) :
let coords₁ : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first₁ : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords₁ v 0
let tail₁ : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords₁ v (Fin.natAdd 1 i))
let coords₂ : EuclideanSpace Real (Fin (1 + m)) → (Fin (1 + m) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + m))
let first₂ : EuclideanSpace Real (Fin (1 + m)) → Real := fun v => coords₂ v 0
let tail₂ : EuclideanSpace Real (Fin (1 + m)) → EuclideanSpace Real (Fin m) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin m)).symm
(fun i => coords₂ v (Fin.natAdd 1 i))
let S : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first₁ v = 1 ∧ tail₁ v ∈ C}
let K : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S : Set (EuclideanSpace Real (Fin (1 + n))))
let S' : Set (EuclideanSpace Real (Fin (1 + m))) :=
{v | first₂ v = 1 ∧ tail₂ v ∈ A '' C}
let K' : Set (EuclideanSpace Real (Fin (1 + m))) :=
(ConvexCone.hull Real S' : Set (EuclideanSpace Real (Fin (1 + m))))
let B : EuclideanSpace Real (Fin (1 + n)) →ₗ[Real] EuclideanSpace Real (Fin (1 + m)) :=
let lift :
(EuclideanSpace Real (Fin 1) × EuclideanSpace Real (Fin n)) →ₗ[Real]
(EuclideanSpace Real (Fin 1) × EuclideanSpace Real (Fin m)) :=
(LinearMap.fst (R := Real) (M := EuclideanSpace Real (Fin 1))
(M₂ := EuclideanSpace Real (Fin n))).prod
(A.comp (LinearMap.snd (R := Real) (M := EuclideanSpace Real (Fin 1))
(M₂ := EuclideanSpace Real (Fin n))))
(appendAffineEquiv 1 m).linear.toLinearMap.comp
(lift.comp (appendAffineEquiv 1 n).symm.linear.toLinearMap)
B '' K = K'
Kernel-lineality for the lifted cone over closure C.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma kernel_lineality_for_cone {n m : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCne : Set.Nonempty C) (hCconv : Convex Real C)
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m))
(hlineal :
∀ z, z ≠ 0 → z ∈ Set.recessionCone (closure C) → A z = 0 →
z ∈ Set.linealitySpace (closure C)) :
let Cbar : Set (EuclideanSpace Real (Fin n)) := closure C
let coords₁ : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first₁ : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords₁ v 0
let tail₁ : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords₁ v (Fin.natAdd 1 i))
let S : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first₁ v = 1 ∧ tail₁ v ∈ Cbar}
let K : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S : Set (EuclideanSpace Real (Fin (1 + n))))
let B : EuclideanSpace Real (Fin (1 + n)) →ₗ[Real] EuclideanSpace Real (Fin (1 + m)) :=
let lift :
(EuclideanSpace Real (Fin 1) × EuclideanSpace Real (Fin n)) →ₗ[Real]
(EuclideanSpace Real (Fin 1) × EuclideanSpace Real (Fin m)) :=
(LinearMap.fst (R := Real) (M := EuclideanSpace Real (Fin 1))
(M₂ := EuclideanSpace Real (Fin n))).prod
(A.comp (LinearMap.snd (R := Real) (M := EuclideanSpace Real (Fin 1))
(M₂ := EuclideanSpace Real (Fin n))))
(appendAffineEquiv 1 m).linear.toLinearMap.comp
(lift.comp (appendAffineEquiv 1 n).symm.linear.toLinearMap)
∀ v, v ≠ 0 → v ∈ Set.recessionCone (closure K) → B v = 0 →
v ∈ Set.linealitySpace (closure K)
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
theorem linearMap_closure_image_eq_image_closure_of_recessionCone_kernel_lineality
{n m : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCne : Set.Nonempty C) (hCconv : Convex Real C)
(A : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m))
(hlineal :
∀ z, z ≠ 0 → z ∈ Set.recessionCone (closure C) → A z = 0 →
z ∈ Set.linealitySpace (closure C)) :
closure (A '' C) = A '' closure C ∧
Set.recessionCone (A '' closure C) = A '' Set.recessionCone (closure C) ∧
(IsClosed C →
(∀ z, z ∈ Set.recessionCone C → A z = 0 → z = 0) →
IsClosed (A '' C))§9 — Some Closedness Criteria — Part 3
Lean authors: Zichen Wang, Wanli Ma, Qiming Dai, Zaiwen Wen.
The coordinate-sum linear map on Fin m-indexed tuples.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
def sumLinearMap {n m : Nat} :
(Fin m → EuclideanSpace Real (Fin n)) →ₗ[Real] EuclideanSpace Real (Fin n)
The image of the product set under the sum map is the Minkowski sum.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma image_sumLinearMap_pi_eq_sum {n m : Nat}
(C : Fin m → Set (EuclideanSpace Real (Fin n))) :
sumLinearMap (n := n) (m := m) '' (Set.pi Set.univ C) = ∑ i, C i
Linear equivalence between EuclideanSpace Real (Fin (m * n)) and block vectors.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
noncomputable def euclideanSpace_equiv_prod_block {n m : Nat} :
EuclideanSpace Real (Fin (m * n)) ≃ₗ[Real] (Fin m → EuclideanSpace Real (Fin n))
The block product set in ℝ^{m n} associated to C.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
def blockSet {n m : Nat} (C : Fin m → Set (EuclideanSpace Real (Fin n))) :
Set (EuclideanSpace Real (Fin (m * n))) :=
(euclideanSpace_equiv_prod_block (n := n) (m := m)) ⁻¹' Set.pi Set.univ C
The block sum linear map on EuclideanSpace Real (Fin (m * n)).
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
def blockSumLinearMap {n m : Nat} :
EuclideanSpace Real (Fin (m * n)) →ₗ[Real] EuclideanSpace Real (Fin n) :=
(sumLinearMap (n := n) (m := m)).comp
(euclideanSpace_equiv_prod_block (n := n) (m := m)).toLinearMap
The block equivalence sends blockSet to the product set.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma image_equiv_block_eq_pi {n m : Nat}
(C : Fin m → Set (EuclideanSpace Real (Fin n))) :
(euclideanSpace_equiv_prod_block (n := n) (m := m)) '' blockSet (n := n) (m := m) C =
Set.pi Set.univ C
Membership in the closure of blockSet is coordinatewise closure.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma mem_closure_block_iff {n m : Nat}
(C : Fin m → Set (EuclideanSpace Real (Fin n)))
(x : EuclideanSpace Real (Fin (m * n))) :
x ∈ closure (blockSet (n := n) (m := m) C) ↔
∀ i, (euclideanSpace_equiv_prod_block (n := n) (m := m) x) i ∈ closure (C i)
Recession cone of the block set is coordinatewise recession.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma mem_recessionCone_block_iff {n m : Nat}
(C : Fin m → Set (EuclideanSpace Real (Fin n))) (hCne : ∀ i, Set.Nonempty (C i))
(z : EuclideanSpace Real (Fin (m * n))) :
z ∈ Set.recessionCone (closure (blockSet (n := n) (m := m) C)) ↔
∀ i,
(euclideanSpace_equiv_prod_block (n := n) (m := m) z) i ∈
Set.recessionCone (closure (C i))
The block equivalence sends the block recession cone to the product of recession cones.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma image_equiv_recessionCone_block_eq_pi {n m : Nat}
(C : Fin m → Set (EuclideanSpace Real (Fin n))) (hCne : ∀ i, Set.Nonempty (C i)) :
(euclideanSpace_equiv_prod_block (n := n) (m := m)) ''
Set.recessionCone (closure (blockSet (n := n) (m := m) C)) =
Set.pi Set.univ (fun i => Set.recessionCone (closure (C i)))
Corollary 9.1.1. Let C1, ..., Cm be non-empty convex sets in R^n. Assume that whenever z1, ..., zm satisfy zi ∈ 0^+ (cl Ci) and z1 + ... + zm = 0, then each zi lies in the lineality space of cl Ci. Then cl (C1 + ... + Cm) = cl C1 + ... + cl Cm and 0^+ (cl (C1 + ... + Cm)) = 0^+ (cl C1) + ... + 0^+ (cl Cm). In particular, if each Ci is closed then C1 + ... + Cm is closed.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
theorem closure_sum_eq_sum_closure_of_recessionCone_sum_zero_lineality
{n m : Nat} (C : Fin m → Set (EuclideanSpace Real (Fin n)))
(hCne : ∀ i, Set.Nonempty (C i)) (hCconv : ∀ i, Convex Real (C i))
(hlineal :
∀ z : Fin m → EuclideanSpace Real (Fin n),
(∀ i, z i ∈ Set.recessionCone (closure (C i))) →
(∑ i, z i) = 0 →
∀ i, z i ∈ Set.linealitySpace (closure (C i))) :
closure (∑ i, C i) = ∑ i, closure (C i) ∧
Set.recessionCone (closure (∑ i, C i)) =
∑ i, Set.recessionCone (closure (C i)) ∧
((∀ i, IsClosed (C i)) → IsClosed (∑ i, C i))
Corollary 9.1.1.1. Under the hypotheses of Corollary 9.1.1, if all C_i are closed, then C_1 + ... + C_m is closed.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
theorem closed_sum_of_closed_sets_of_recessionCone_sum_zero_lineality
{n m : Nat} (C : Fin m → Set (EuclideanSpace Real (Fin n)))
(hCne : ∀ i, Set.Nonempty (C i)) (hCconv : ∀ i, Convex Real (C i))
(hlineal :
∀ z : Fin m → EuclideanSpace Real (Fin n),
(∀ i, z i ∈ Set.recessionCone (closure (C i))) →
(∑ i, z i) = 0 →
∀ i, z i ∈ Set.linealitySpace (closure (C i)))
(hCclosed : ∀ i, IsClosed (C i)) :
IsClosed (∑ i, C i)
Zero belongs to the lineality space of any set.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma zero_mem_linealitySpace {n : Nat} (C : Set (EuclideanSpace Real (Fin n))) :
(0 : EuclideanSpace Real (Fin n)) ∈ Set.linealitySpace C
In the Fin 2 case, no-opposite-recession implies lineality.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma lineality_of_no_opposite_recession_fin2
{n : Nat} (C : Fin 2 → Set (EuclideanSpace Real (Fin n)))
(hC0closed : IsClosed (C 0)) (hC1closed : IsClosed (C 1))
(hopp :
∀ z,
z ∈ Set.recessionCone (C 0) → -z ∈ Set.recessionCone (C 1) → z = 0) :
∀ z : Fin 2 → EuclideanSpace Real (Fin n),
(∀ i, z i ∈ Set.recessionCone (closure (C i))) →
(∑ i, z i) = 0 →
∀ i, z i ∈ Set.linealitySpace (closure (C i))
Corollary 9.1.2. Let C1 and C2 be non-empty closed convex sets in R^n. Assume there is no direction of recession of C1 whose opposite is a direction of recession of C2 (in particular, this holds if either C1 or C2 is bounded). Then C1 + C2 is closed and 0^+ (C1 + C2) = 0^+ C1 + 0^+ C2.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
theorem closed_add_recessionCone_eq_of_no_opposite_recession
{n : Nat} {C1 C2 : Set (EuclideanSpace Real (Fin n))}
(hC1ne : Set.Nonempty C1) (hC2ne : Set.Nonempty C2)
(hC1conv : Convex Real C1) (hC2conv : Convex Real C2)
(hC1closed : IsClosed C1) (hC2closed : IsClosed C2)
(hopp :
∀ z,
z ∈ Set.recessionCone C1 → -z ∈ Set.recessionCone C2 → z = 0) :
IsClosed (C1 + C2) ∧
Set.recessionCone (C1 + C2) = Set.recessionCone C1 + Set.recessionCone C2
Convexity is preserved when transporting a convex cone through EuclideanSpace.equiv.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma convex_of_IsConvexCone_image_equiv {n : Nat}
{K : Set (EuclideanSpace Real (Fin n))}
(hKcone :
IsConvexCone n
(Set.image (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)) K)) :
Convex Real K
Positive scaling in K inherited from the cone structure of its coordinate image.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma pos_smul_mem_of_IsConvexCone_image_equiv {n : Nat}
{K : Set (EuclideanSpace Real (Fin n))}
(hKcone :
IsConvexCone n
(Set.image (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)) K)) :
∀ x ∈ K, ∀ t : Real, 0 < t → t • x ∈ K
A nonempty set closed under positive scaling has 0 in its closure.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma zero_mem_closure_of_pos_smul_closed {n : Nat}
{K : Set (EuclideanSpace Real (Fin n))}
(hKne : K.Nonempty)
(hsmul : ∀ x ∈ K, ∀ t : Real, 0 < t → t • x ∈ K) :
(0 : EuclideanSpace Real (Fin n)) ∈ closure K
The recession cone of the closure of a convex cone is the closure itself.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma recessionCone_closure_eq_of_convexCone {n : Nat}
{K : Set (EuclideanSpace Real (Fin n))}
(hKne : K.Nonempty)
(hKcone :
IsConvexCone n
(Set.image (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)) K)) :
Set.recessionCone (closure K) = closure K
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
theorem closure_sum_eq_sum_closure_of_convexCone_sum_zero_lineality
{n m : Nat} (K : Fin m → Set (EuclideanSpace Real (Fin n)))
(hKne : ∀ i, Set.Nonempty (K i))
(hKcone :
∀ i,
IsConvexCone n
(Set.image (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)) (K i)))
(hlineal :
∀ z : Fin m → EuclideanSpace Real (Fin n),
(∀ i, z i ∈ closure (K i)) →
(∑ i, z i) = 0 →
∀ i, z i ∈ Set.linealitySpace (closure (K i))) :
closure (∑ i, K i) = ∑ i, closure (K i)
The parabola set {x | x1 ≥ x0^2} is closed.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma isClosed_parabola :
IsClosed {x : EuclideanSpace Real (Fin 2) | x 1 ≥ (x 0) ^ 2}
The projection of the parabola set onto the first coordinate is all of ℝ.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma image_parabola_eq_univ :
let C : Set (EuclideanSpace Real (Fin 2)) := {x | x 1 ≥ (x 0) ^ 2}
let A : EuclideanSpace Real (Fin 2) → EuclideanSpace Real (Fin 1) :=
fun x =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm (fun _ : Fin 1 => x 0)
A '' C = (Set.univ : Set (EuclideanSpace Real (Fin 1)))
The recession cone of the parabola is the vertical ray {(0, z2) | z2 ≥ 0}.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma recessionCone_parabola_eq :
Set.recessionCone {x : EuclideanSpace Real (Fin 2) | x 1 ≥ (x 0) ^ 2} =
{z : EuclideanSpace Real (Fin 2) | z 0 = 0 ∧ 0 ≤ z 1}
The image of the recession cone under the projection is {0}.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma image_recessionCone_parabola_eq_singleton :
let C : Set (EuclideanSpace Real (Fin 2)) := {x | x 1 ≥ (x 0) ^ 2}
let A : EuclideanSpace Real (Fin 2) → EuclideanSpace Real (Fin 1) :=
fun x =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm (fun _ : Fin 1 => x 0)
A '' Set.recessionCone C = ({0} : Set (EuclideanSpace Real (Fin 1)))
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma recessionCone_image_ne_image_recessionCone_parabola :
let C : Set (EuclideanSpace Real (Fin 2)) := {x | x 1 ≥ (x 0) ^ 2}
let A : EuclideanSpace Real (Fin 2) → EuclideanSpace Real (Fin 1) :=
fun x =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm (fun _ : Fin 1 => x 0)
IsClosed C ∧ IsClosed (A '' C) ∧
Set.recessionCone (A '' C) ≠ A '' Set.recessionCone C
The fiber infimum under a linear map is convex when the original function is convex.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma convexFunction_linearMap_infimum {n m : Nat}
{h : (Fin n → Real) → EReal}
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real))
(hconv : ConvexFunction h) :
ConvexFunction
(fun y => sInf { z : EReal | ∃ x : Fin n → Real, A x = y ∧ z = h x })
The epigraph of the fiber infimum is nonempty if the original epigraph is nonempty.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma nonempty_epigraph_linearMap_infimum {n m : Nat}
{h : (Fin n → Real) → EReal}
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real))
(hne : Set.Nonempty (epigraph (Set.univ : Set (Fin n → Real)) h)) :
Set.Nonempty
(epigraph (Set.univ : Set (Fin m → Real))
(fun y => sInf { z : EReal | ∃ x : Fin n → Real, A x = y ∧ z = h x }))
The lifted linear map on product spaces sending (x, μ) to (A x, μ).
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
def linearMap_prod {n m : Nat} (A : (Fin n → Real) →ₗ[Real] (Fin m → Real)) :
(Fin n → Real) × Real →ₗ[Real] (Fin m → Real) × Real
Linear equivalence between (Fin n → ℝ) × ℝ and ℝ^{n+1} via append.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
noncomputable def prodLinearEquiv_append {n : Nat} :
((Fin n → Real) × Real) ≃ₗ[Real] EuclideanSpace Real (Fin (n + 1))
Conjugate linearMap_prod under the append equivalence.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
noncomputable def linearMap_prod_embedded {n m : Nat}
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real)) :
EuclideanSpace Real (Fin (n + 1)) →ₗ[Real] EuclideanSpace Real (Fin (m + 1)) :=
LinearMap.comp (prodLinearEquiv_append (n := m)).toLinearMap
(LinearMap.comp (linearMap_prod A) (prodLinearEquiv_append (n := n)).symm.toLinearMap)
The embedded image matches the image under the conjugated linear map.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma image_linearMap_prod_embedded {n m : Nat} {h : (Fin n → Real) → EReal}
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real)) :
(prodLinearEquiv_append (n := m)) ''
((linearMap_prod A) '' epigraph (Set.univ : Set (Fin n → Real)) h) =
(linearMap_prod_embedded (A := A)) ''
((prodLinearEquiv_append (n := n)) '' epigraph (Set.univ : Set (Fin n → Real)) h)§9 — Some Closedness Criteria — Part 4
Lean authors: Zichen Wang, Wanli Ma, Qiming Dai, Zaiwen Wen.
Membership in the kernel of linearMap_prod is componentwise.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma mem_ker_linearMap_prod_iff {n m : Nat}
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real)) (p : (Fin n → Real) × Real) :
p ∈ LinearMap.ker (linearMap_prod A) ↔ A p.1 = 0 ∧ p.2 = 0
Recession cones commute with linear equivalences.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma recessionCone_image_linearEquiv {E F : Type*} [AddCommGroup E] [Module Real E]
[AddCommGroup F] [Module Real F] (e : E ≃ₗ[Real] F) (C : Set E) :
Set.recessionCone (e '' C) = e '' Set.recessionCone C
The recession cone of the embedded epigraph is the embedded epigraph of h0_plus.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma recessionCone_embedded_epigraph_eq_epigraph_h0_plus {n : Nat}
{h h0_plus : (Fin n → Real) → EReal}
(hrec :
Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) h) =
epigraph (Set.univ : Set (Fin n → Real)) h0_plus) :
Set.recessionCone
((prodLinearEquiv_append (n := n)) ''
epigraph (Set.univ : Set (Fin n → Real)) h) =
(prodLinearEquiv_append (n := n)) ''
epigraph (Set.univ : Set (Fin n → Real)) h0_plus
Corollary 4.7.2 specialized to h0_plus.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma h0_plus_neg_ge_neg_of_posHom_proper {n : Nat}
{h0_plus : (Fin n → Real) → EReal}
(hpos : PositivelyHomogeneous h0_plus)
(hproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) h0_plus) :
∀ z : Fin n → Real, h0_plus (-z) ≥ -h0_plus z
Precomposition with a surjective linear map preserves proper convexity on Set.univ.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma properConvexFunctionOn_precomp_linearMap_surjective {n m : Nat}
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real))
(hA : Function.Surjective A)
{f : (Fin m → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin m → Real)) f) :
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (fun x => f (A x))
Kernel directions in the embedded recession cone are lineality directions.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma lineality_of_kernel_recession_embedded_epigraph {n m : Nat}
{h h0_plus : (Fin n → Real) → EReal}
(hclosed : ClosedConvexFunction h)
(hrec :
Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) h) =
epigraph (Set.univ : Set (Fin n → Real)) h0_plus)
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real))
(hA :
∀ z : Fin n → Real,
h0_plus z ≤ (0 : EReal) → h0_plus (-z) > (0 : EReal) → A z ≠ 0) :
∀ z, z ≠ 0 →
z ∈ Set.recessionCone
(closure ((prodLinearEquiv_append (n := n)) ''
epigraph (Set.univ : Set (Fin n → Real)) h)) →
(linearMap_prod_embedded (A := A)) z = 0 →
z ∈ Set.linealitySpace
(closure ((prodLinearEquiv_append (n := n)) ''
epigraph (Set.univ : Set (Fin n → Real)) h))
The image epigraph has no downward vertical recession directions under h0_plus.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma no_downward_vertical_recession_image_epigraph {n m : Nat}
{h h0_plus : (Fin n → Real) → EReal}
(hclosed : ClosedConvexFunction h)
(hproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) h)
(hrec :
Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) h) =
epigraph (Set.univ : Set (Fin n → Real)) h0_plus)
(hpos : PositivelyHomogeneous h0_plus)
(hproper0 : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) h0_plus)
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real))
(hA :
∀ z : Fin n → Real,
h0_plus z ≤ (0 : EReal) → h0_plus (-z) > (0 : EReal) → A z ≠ 0) :
¬ ∃ μ < 0,
(0, μ) ∈ Set.recessionCone
((linearMap_prod A) '' epigraph (Set.univ : Set (Fin n → Real)) h)
The image of the epigraph under the lifted linear map.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma image_epigraph_linearMap_eq {n m : Nat}
{h : (Fin n → Real) → EReal}
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real)) :
(linearMap_prod A) '' epigraph (Set.univ : Set (Fin n → Real)) h =
{p : (Fin m → Real) × Real |
∃ x : Fin n → Real, A x = p.1 ∧ h x ≤ (p.2 : EReal)}
The projected epigraph is contained in the epigraph of the fiber infimum.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma image_epigraph_linearMap_subset_epigraph_infimum {n m : Nat}
{h : (Fin n → Real) → EReal}
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real)) :
(linearMap_prod A) '' epigraph (Set.univ : Set (Fin n → Real)) h ⊆
epigraph (Set.univ : Set (Fin m → Real))
(fun y => sInf { z : EReal | ∃ x : Fin n → Real, A x = y ∧ z = h x })
If the fiber infimum is not ⊤, then the fiber is nonempty.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma exists_preimage_of_sInf_ne_top {n m : Nat} {h : (Fin n → Real) → EReal}
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real)) {y : Fin m → Real}
(hy :
sInf { z : EReal | ∃ x : Fin n → Real, A x = y ∧ z = h x } ≠ (⊤ : EReal)) :
∃ x : Fin n → Real, A x = y
If the fiber infimum is strictly below a real bound, there is a fiber point below that bound.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma exists_preimage_of_sInf_lt {n m : Nat} {h : (Fin n → Real) → EReal}
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real)) {y : Fin m → Real} {μ : Real}
(hμ :
sInf { z : EReal | ∃ x : Fin n → Real, A x = y ∧ z = h x } < (μ : EReal)) :
∃ x : Fin n → Real, A x = y ∧ h x ≤ (μ : EReal)
Closed image of the epigraph identifies the epigraph of the fiber infimum.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma epigraph_infimum_eq_image_of_closed_image {n m : Nat}
{h : (Fin n → Real) → EReal}
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real))
(hclosed :
IsClosed ((linearMap_prod A) '' epigraph (Set.univ : Set (Fin n → Real)) h)) :
epigraph (Set.univ : Set (Fin m → Real))
(fun y => sInf { z : EReal | ∃ x : Fin n → Real, A x = y ∧ z = h x }) =
(linearMap_prod A) '' epigraph (Set.univ : Set (Fin n → Real)) h
If a fiber infimum is ⊥, the projected epigraph has a negative vertical recession direction.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma exists_negative_vertical_recession_of_bot {n m : Nat} {h : (Fin n → Real) → EReal}
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real))
(hproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) h)
(himage_closed :
IsClosed ((linearMap_prod A) '' epigraph (Set.univ : Set (Fin n → Real)) h))
(hEq_epi :
epigraph (Set.univ : Set (Fin m → Real))
(fun y => sInf { z : EReal | ∃ x : Fin n → Real, A x = y ∧ z = h x }) =
(linearMap_prod A) '' epigraph (Set.univ : Set (Fin n → Real)) h)
{y : Fin m → Real}
(hbot :
sInf { z : EReal | ∃ x : Fin n → Real, A x = y ∧ z = h x } = (⊥ : EReal)) :
∃ μ < 0, (0, μ) ∈ Set.recessionCone
((linearMap_prod A) '' epigraph (Set.univ : Set (Fin n → Real)) h)
Theorem 9.2. Let h be a closed proper convex function on R^n, and let A be a linear transformation from R^n to R^m. Assume that A z ≠ 0 for every z such that (h0^+)(z) ≤ 0 and (h0^+)(-z) > 0. Then the function Ah, where (Ah)(y) = inf { h(x) | A x = y }, is a closed proper convex function, and (Ah)0^+ = A(h0^+). Moreover, for each y such that (Ah)(y) ≠ +infty, the infimum in the definition of (Ah)(y) is attained for some x.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
theorem linearMap_infimum_closed_proper_convex_recession
{n m : Nat} {h h0_plus : (Fin n → Real) → EReal}
(hclosed : ClosedConvexFunction h)
(hproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) h)
(hrec :
Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) h) =
epigraph (Set.univ : Set (Fin n → Real)) h0_plus)
(hpos : PositivelyHomogeneous h0_plus)
(hproper0 : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) h0_plus)
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real))
(hA :
∀ z : Fin n → Real,
h0_plus z ≤ (0 : EReal) → h0_plus (-z) > (0 : EReal) → A z ≠ 0) :
let Ah : (Fin m → Real) → EReal :=
fun y => sInf { z : EReal | ∃ x : Fin n → Real, A x = y ∧ z = h x }
let Ah0_plus : (Fin m → Real) → EReal :=
fun y => sInf { z : EReal | ∃ x : Fin n → Real, A x = y ∧ z = h0_plus x }
ClosedConvexFunction Ah ∧
ProperConvexFunctionOn (Set.univ : Set (Fin m → Real)) Ah ∧
Ah0_plus =
(fun y => sInf { z : EReal | ∃ x : Fin n → Real, A x = y ∧ z = h0_plus x }) ∧
(∀ y : Fin m → Real, Ah y ≠ (⊤ : EReal) →
∃ x : Fin n → Real, A x = y ∧ Ah y = h x)
If h0_plus is strictly positive away from zero, then h0_plus z ≤ 0 forces z = 0.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma h0_plus_le_zero_imp_zero {n : Nat} {h0_plus : (Fin n → Real) → EReal}
(hno : ∀ z : Fin n → Real, z ≠ 0 → h0_plus z > (0 : EReal)) :
∀ z : Fin n → Real, h0_plus z ≤ (0 : EReal) → z = 0
If z = 0 and h0_plus 0 = 0, then h0_plus (-z) = 0.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma h0_plus_neg_eq_zero_of_z_eq_zero {n : Nat} {h0_plus : (Fin n → Real) → EReal}
(hzero : h0_plus (0 : Fin n → Real) = (0 : EReal)) :
∀ z : Fin n → Real, z = 0 → h0_plus (-z) = (0 : EReal)
The coordinate-sum linear map on function blocks.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
def sumLinearMapFun {n m : Nat} :
(Fin m → (Fin n → Real)) →ₗ[Real] (Fin n → Real)
Linear equivalence between Fin (m * n) → Real and block functions.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
noncomputable def blockEquivFun {n m : Nat} :
(Fin (m * n) → Real) ≃ₗ[Real] (Fin m → (Fin n → Real))
Block-sum linear map on function spaces.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
noncomputable def blockSumLinearMapFun {n m : Nat} :
(Fin (m * n) → Real) →ₗ[Real] (Fin n → Real) :=
(sumLinearMapFun (n := n) (m := m)).comp
(blockEquivFun (n := n) (m := m)).toLinearMap
Remark 9.2.0.2. The hypothesis of Theorem 9.2 concerning h0_plus is trivially satisfied if h has no directions of recession, in particular if dom h is bounded; the example at the beginning of this section violates this hypothesis.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma linearMap_infimum_hypothesis_trivial_of_no_recession_direction
{n m : Nat} {h0_plus : (Fin n → Real) → EReal}
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real))
(hno : ∀ z : Fin n → Real, z ≠ 0 → h0_plus z > (0 : EReal))
(hzero : h0_plus (0 : Fin n → Real) = (0 : EReal)) :
∀ z : Fin n → Real, h0_plus z ≤ (0 : EReal) → h0_plus (-z) > (0 : EReal) → A z ≠ 0§9 — Some Closedness Criteria — Part 5
Lean authors: Zichen Wang, Wanli Ma, Qiming Dai, Zaiwen Wen.
The infimal convolution family is the linear-map infimum for the block sum.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma infimalConvolutionFamily_eq_Ah {n m : Nat}
(f : Fin m → (Fin n → Real) → EReal) :
let e := blockEquivFun (n := n) (m := m)
let A := (sumLinearMapFun (n := n) (m := m)).comp e.toLinearMap
let h : (Fin (m * n) → Real) → EReal :=
fun x => Finset.univ.sum (fun i => f i ((e x) i))
let Ah : (Fin n → Real) → EReal :=
fun y => sInf { z : EReal | ∃ x : Fin (m * n) → Real, A x = y ∧ z = h x }
Ah = infimalConvolutionFamily f
The h0_plus kernel condition for the block sum follows from the corollary hypothesis.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma hA_from_hzero {n m : Nat}
{f0_plus : Fin m → (Fin n → Real) → EReal}
(hzero :
∀ z : Fin m → (Fin n → Real),
(Finset.univ.sum (fun i => f0_plus i (z i)) ≤ (0 : EReal)) →
(Finset.univ.sum (fun i => f0_plus i (-(z i))) > (0 : EReal)) →
(Finset.univ.sum (fun i => z i) ≠ (0 : Fin n → Real))) :
let e := blockEquivFun (n := n) (m := m)
let A := (sumLinearMapFun (n := n) (m := m)).comp e.toLinearMap
let h0_plus : (Fin (m * n) → Real) → EReal :=
fun x => Finset.univ.sum (fun i => f0_plus i ((e x) i))
∀ z : Fin (m * n) → Real,
h0_plus z ≤ (0 : EReal) →
h0_plus (-z) > (0 : EReal) → A z ≠ 0
Linear map sending (xᵢ, μᵢ) to (xᵢ, ∑ μᵢ).
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
noncomputable def sumMuLinearMap {n m : Nat} :
(Fin m → (Fin n → Real) × Real) →ₗ[Real] (Fin m → (Fin n → Real)) × Real
Recession cone of a nonempty product set is the product of recession cones.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma recessionCone_pi_eq_pi_recessionCone {m : Nat} {E : Type*} [AddCommGroup E] [Module Real E]
(C : Fin m → Set E) (hC : ∀ i, (C i).Nonempty) :
Set.recessionCone (Set.pi Set.univ C) =
Set.pi Set.univ (fun i => Set.recessionCone (C i))
Sum of real coefficients coerced to EReal is the coercion of the real sum.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma sum_coe_ereal {α : Type*} (s : Finset α) (f : α → Real) :
s.sum (fun i => (f i : EReal)) = ((s.sum f : Real) : EReal)
A finite sum of EReal terms is not ⊥ if none of the terms is ⊥.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma sum_ne_bot_of_ne_bot' {α : Type*} (s : Finset α) (β : α → EReal)
(h : ∀ i ∈ s, β i ≠ (⊥ : EReal)) : s.sum β ≠ (⊥ : EReal)
A finite sum is ⊤ if one term is ⊤ and none are ⊥.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma sum_eq_top_of_eq_top {α : Type*} (s : Finset α) (β : α → EReal)
(i0 : α) (hi0 : i0 ∈ s) (htop : β i0 = (⊤ : EReal))
(hbot : ∀ i ∈ s, β i ≠ (⊥ : EReal)) :
s.sum β = (⊤ : EReal)
Split a real upper bound on an EReal sum into componentwise real bounds.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma split_real_sum_of_le {m : Nat} (hm : 0 < m) {β : Fin m → EReal} {α : Real}
(hbot : ∀ i, β i ≠ (⊥ : EReal))
(hβ : (Finset.univ.sum (fun i => β i)) ≤ (α : EReal)) :
∃ a : Fin m → Real,
(∀ i, β i ≤ (a i : EReal)) ∧
(Finset.univ.sum (fun i => a i) = α)
A strict inequality on a finite EReal sum is equivalent to the strict real inequality.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma g0_plus_gt_toReal_sum {m : Nat} {β : Fin m → EReal} {α : Real}
(hnot_top : ∀ i, β i ≠ (⊤ : EReal))
(hnot_bot : ∀ i, β i ≠ (⊥ : EReal)) :
(Finset.univ.sum (fun i => β i) > (α : EReal)) ↔
(Finset.univ.sum (fun i => (β i).toReal) > α)
Split a strict inequality for a real sum by lowering one component.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma split_real_sum_of_gt {m : Nat} (hm : 0 < m) {r : Fin m → Real} {α : Real} :
(Finset.univ.sum (fun i => r i) > α) →
∃ i0 : Fin m,
∃ a : Fin m → Real,
a i0 < r i0 ∧
(∀ i ≠ i0, a i = r i) ∧
(Finset.univ.sum (fun i => a i) = α)
Split a strict sum inequality into strictly lower components with the same total.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma split_real_sum_of_gt_strict_all {m : Nat} (hm : 0 < m) {r : Fin m → Real} {α : Real}
(hgt : Finset.univ.sum (fun i => r i) > α) :
∃ a : Fin m → Real, (∀ i, a i < r i) ∧ Finset.univ.sum (fun i => a i) = α
A witness for failing recession membership in an epigraph.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma not_mem_recessionCone_epigraph_witness {n : Nat}
{f : (Fin n → Real) → EReal} {v : Fin n → Real} {a : Real}
(hnot :
(v, a) ∉ Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) f)) :
∃ x μ t,
0 ≤ t ∧
f (x + t • v) > ((μ + t * a : Real) : EReal) ∧
f x ≤ (μ : EReal)
Epigraph nonemptiness from recession-epigraph equality and non-bottomness.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma nonempty_epigraph_of_hrec {n m : Nat}
{f f0_plus : Fin m → (Fin n → Real) → EReal}
(hnotbot0 : ∀ i : Fin m, ∀ x, f0_plus i x ≠ (⊥ : EReal))
(hrec_i :
∀ i : Fin m,
Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) (f i)) =
epigraph (Set.univ : Set (Fin n → Real)) (f0_plus i)) :
∀ i : Fin m, Set.Nonempty (epigraph (Set.univ : Set (Fin n → Real)) (f i))
A fixed-step translation in a convex set yields a recession direction.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma mem_recessionCone_of_add_mem_fixed_t {E : Type*} [AddCommGroup E] [Module Real E]
{C : Set E} (hconv : Convex Real C) {v : E} {t : Real} (ht : 0 < t)
(hadd : ∀ x ∈ C, x + t • v ∈ C) : v ∈ Set.recessionCone C
Approximate a recession slope at a fixed time with slack.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma recCone_tight_approx_at_t {n : Nat}
{f f0_plus : (Fin n → Real) → EReal}
(hrec :
Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) f) =
epigraph (Set.univ : Set (Fin n → Real)) f0_plus)
(hconv : Convex Real (epigraph (Set.univ : Set (Fin n → Real)) f))
{v : Fin n → Real} {a : Real} (ha : ((a : Real) : EReal) < f0_plus v)
{t : Real} (ht : 0 < t) {ε : Real} (hε : 0 < ε) :
∃ x : Fin n → Real, ∃ μ : Real, f x ≤ (μ : EReal) ∧
((μ + t * a - ε : Real) : EReal) ≤ f (x + t • v)
A strict gap at one component plus slack bounds on the others yields a strict sum.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma strict_sum_from_one_component_with_slack {m : Nat} {β b ε : Fin m → Real} {i0 : Fin m}
(hgt : β i0 > b i0)
(hslack : ∀ i ≠ i0, β i ≥ b i - ε i)
(hε : (Finset.univ.erase i0).sum (fun i => ε i) < β i0 - b i0) :
Finset.univ.sum (fun i => β i) > Finset.univ.sum (fun i => b i)
A strict gap with slack on EReal components yields a strict sum inequality.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma ereal_strict_sum_from_one_component_with_slack {m : Nat} {β : Fin m → EReal}
{b ε : Fin m → Real} {i0 : Fin m}
(hnot_top : ∀ i, β i ≠ (⊤ : EReal)) (hnot_bot : ∀ i, β i ≠ (⊥ : EReal))
(hgt : β i0 > (b i0 : EReal))
(hslack : ∀ i ≠ i0, β i ≥ ((b i - ε i : Real) : EReal))
(hε : (Finset.univ.erase i0).sum (fun i => ε i) < (β i0).toReal - b i0) :
Finset.univ.sum (fun i => β i) > (Finset.univ.sum (fun i => b i) : EReal)
Finite-case contradiction in the ⊤ branch of recessionCone_sum_epigraph_prod.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma recessionCone_sum_epigraph_prod_top_contra {n m : Nat}
{f f0_plus : Fin m → (Fin n → Real) → EReal}
{g : (Fin m → (Fin n → Real)) → EReal}
(hg : g = fun x => Finset.univ.sum (fun i => f i (x i)))
(hnotbot : ∀ i : Fin m, ∀ x, f i x ≠ (⊥ : EReal))
(_hnotbot0 : ∀ i : Fin m, ∀ x, f0_plus i x ≠ (⊥ : EReal))
(_hrec_i :
∀ i : Fin m,
Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) (f i)) =
epigraph (Set.univ : Set (Fin n → Real)) (f0_plus i))
{p : (Fin m → (Fin n → Real)) × Real} {i0 : Fin m}
(_htop : f0_plus i0 (p.1 i0) = (⊤ : EReal))
{a : Fin m → Real}
{x0 : Fin n → Real} {μ0 t : Real}
(hgt0 : f i0 (x0 + t • p.1 i0) > ((μ0 + t * a i0 : Real) : EReal))
(_hx0 : f i0 x0 ≤ (μ0 : EReal))
(ha_sum : Finset.univ.sum (fun i => a i) = p.2)
{x : Fin m → (Fin n → Real)} {μ : Fin m → Real}
(hxi0 : x i0 = x0) (hμi0 : μ i0 = μ0)
(_hxle : ∀ i, f i (x i) ≤ (μ i : EReal))
{ε : Fin m → Real}
(hslack :
∀ i ≠ i0,
((μ i + t * a i - ε i : Real) : EReal) ≤ f i (x i + t • p.1 i))
(hε :
(Finset.univ.erase i0).sum (fun i => ε i) <
(f i0 (x0 + t • p.1 i0)).toReal - (μ0 + t * a i0))
(hmem''' :
g (x + t • p.1) ≤ (Finset.univ.sum (fun i => μ i) + t * p.2 : EReal))
(htopg : ¬ ∃ i, f i (x i + t • p.1 i) = (⊤ : EReal)) :
False§9 — Some Closedness Criteria — Part 6
Lean authors: Zichen Wang, Wanli Ma, Qiming Dai, Zaiwen Wen.
Finite-case contradiction in the non-⊤ branch of recessionCone_sum_epigraph_prod.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma recessionCone_sum_epigraph_prod_finite_contra {n m : Nat}
{f f0_plus : Fin m → (Fin n → Real) → EReal}
{g : (Fin m → (Fin n → Real)) → EReal}
(hg : g = fun x => Finset.univ.sum (fun i => f i (x i)))
(hnotbot : ∀ i : Fin m, ∀ x, f i x ≠ (⊥ : EReal))
(_hnotbot0 : ∀ i : Fin m, ∀ x, f0_plus i x ≠ (⊥ : EReal))
(_hrec_i :
∀ i : Fin m,
Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) (f i)) =
epigraph (Set.univ : Set (Fin n → Real)) (f0_plus i))
{p : (Fin m → (Fin n → Real)) × Real} {i0 : Fin m} {a : Fin m → Real}
{x0 : Fin n → Real} {μ0 t : Real}
(hgt0 : f i0 (x0 + t • p.1 i0) > ((μ0 + t * a i0 : Real) : EReal))
(_hx0 : f i0 x0 ≤ (μ0 : EReal))
(ha_sum : Finset.univ.sum (fun i => a i) = p.2)
{x : Fin m → (Fin n → Real)} {μ : Fin m → Real}
(hxi0 : x i0 = x0) (hμi0 : μ i0 = μ0)
(_hxle : ∀ i, f i (x i) ≤ (μ i : EReal))
{ε : Fin m → Real}
(hslack :
∀ i ≠ i0,
((μ i + t * a i - ε i : Real) : EReal) ≤ f i (x i + t • p.1 i))
(hε :
(Finset.univ.erase i0).sum (fun i => ε i) <
(f i0 (x0 + t • p.1 i0)).toReal - (μ0 + t * a i0))
(hmem''' :
g (x + t • p.1) ≤ (Finset.univ.sum (fun i => μ i) + t * p.2 : EReal))
(htopg : ¬ ∃ i, f i (x i + t • p.1 i) = (⊤ : EReal)) :
False
A strict inequality at one component gives a strict inequality for the total sum.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma sum_lt_sum_of_one_lt {m : Nat} {a b : Fin m → Real} {i0 : Fin m}
(hlt : a i0 < b i0) (hEq : ∀ i ≠ i0, a i = b i) :
Finset.univ.sum (fun i => a i) < Finset.univ.sum (fun i => b i)
A strict inequality at one component of a finite EReal sum gives a strict inequality for the total sum (when the other components agree with the real bounds).
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma ereal_sum_lt_sum_of_one_lt {m : Nat} {β : Fin m → EReal} {b : Fin m → Real} {i0 : Fin m}
(hbot : ∀ i, β i ≠ (⊥ : EReal)) (htop : ∀ i, β i ≠ (⊤ : EReal))
(hlt : (b i0 : EReal) < β i0) (hEq : ∀ i ≠ i0, β i = (b i : EReal)) :
Finset.univ.sum (fun i => β i) > (Finset.univ.sum (fun i => b i) : EReal)
Componentwise recession directions give a recession direction for the summed epigraph.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma recessionCone_sum_epigraph_prod_of_components {n m : Nat}
{f : Fin m → (Fin n → Real) → EReal}
{g : (Fin m → (Fin n → Real)) → EReal}
(hg : g = fun x => Finset.univ.sum (fun i => f i (x i)))
(hm : 0 < m) {v : Fin m → (Fin n → Real)} {a : Fin m → Real}
(hnotbot : ∀ i : Fin m, ∀ x, f i x ≠ (⊥ : EReal))
(hrec : ∀ i, (v i, a i) ∈
Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) (f i))) :
(v, Finset.univ.sum (fun i => a i)) ∈
Set.recessionCone {p : (Fin m → (Fin n → Real)) × Real | g p.1 ≤ (p.2 : EReal)}
Recession cone of the sum-epigraph on the product space.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma recessionCone_sum_epigraph_prod {n m : Nat}
{f f0_plus : Fin m → (Fin n → Real) → EReal}
(hnotbot : ∀ i : Fin m, ∀ x, f i x ≠ (⊥ : EReal))
(hnotbot0 : ∀ i : Fin m, ∀ x, f0_plus i x ≠ (⊥ : EReal))
(hconv_i :
∀ i : Fin m,
Convex Real (epigraph (Set.univ : Set (Fin n → Real)) (f i)))
(hrec_i :
∀ i : Fin m,
Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) (f i)) =
epigraph (Set.univ : Set (Fin n → Real)) (f0_plus i)) :
let g : (Fin m → (Fin n → Real)) → EReal :=
fun x => Finset.univ.sum (fun i => f i (x i))
let g0_plus : (Fin m → (Fin n → Real)) → EReal :=
fun x => Finset.univ.sum (fun i => f0_plus i (x i))
Set.recessionCone
{p : (Fin m → (Fin n → Real)) × Real | g p.1 ≤ (p.2 : EReal)} =
{p : (Fin m → (Fin n → Real)) × Real | g0_plus p.1 ≤ (p.2 : EReal)}§9 — Some Closedness Criteria — Part 7
Lean authors: Zichen Wang, Wanli Ma, Qiming Dai, Zaiwen Wen.
The recession cone of the block-sum epigraph is the epigraph of the block-sum recession.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma recessionCone_epigraph_sum_block {n m : Nat}
{f f0_plus : Fin m → (Fin n → Real) → EReal}
(hnotbot : ∀ i : Fin m, ∀ x, f i x ≠ (⊥ : EReal))
(hnotbot0 : ∀ i : Fin m, ∀ x, f0_plus i x ≠ (⊥ : EReal))
(hconv_i :
∀ i : Fin m,
Convex Real (epigraph (Set.univ : Set (Fin n → Real)) (f i)))
(hrec_i :
∀ i : Fin m,
Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) (f i)) =
epigraph (Set.univ : Set (Fin n → Real)) (f0_plus i)) :
let e := blockEquivFun (n := n) (m := m)
let h : (Fin (m * n) → Real) → EReal :=
fun x => Finset.univ.sum (fun i => f i ((e x) i))
let h0_plus : (Fin (m * n) → Real) → EReal :=
fun x => Finset.univ.sum (fun i => f0_plus i ((e x) i))
Set.recessionCone (epigraph (Set.univ : Set (Fin (m * n) → Real)) h) =
epigraph (Set.univ : Set (Fin (m * n) → Real)) h0_plus
The block-sum h0_plus is positively homogeneous and proper under componentwise assumptions.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma h0_plus_sum_posHom_proper {n m : Nat}
{f0_plus : Fin m → (Fin n → Real) → EReal}
(hpos_i : ∀ i : Fin m, PositivelyHomogeneous (f0_plus i))
(hproper_i :
∀ i : Fin m,
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (f0_plus i)) :
let e := blockEquivFun (n := n) (m := m)
let h0_plus : (Fin (m * n) → Real) → EReal :=
fun x => Finset.univ.sum (fun i => f0_plus i ((e x) i))
PositivelyHomogeneous h0_plus ∧
ProperConvexFunctionOn (Set.univ : Set (Fin (m * n) → Real)) h0_plus
Attainment in the ⊤ case for infimal convolution families when m > 0.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma attainment_top_case_infimalConvolutionFamily {n m : Nat}
(hm : 0 < m) {f : Fin m → (Fin n → Real) → EReal} :
let fInf : (Fin n → Real) → EReal := infimalConvolutionFamily f
∀ x : Fin n → Real, fInf x = (⊤ : EReal) →
∃ x' : Fin m → (Fin n → Real),
Finset.univ.sum (fun i => x' i) = x ∧
fInf x = Finset.univ.sum (fun i => f i (x' i))
Corollary 9.2.1. Let f₁, …, fₘ be closed proper convex functions on ℝ^n. Assume that z₁ + … + zₘ ≠ 0 for every choice of vectors z₁, …, zₘ such that (f₁ 0+)(z₁) + … + (fₘ 0+)(zₘ) ≤ 0 and (f₁ 0+)(-z₁) + … + (fₘ 0+)(-zₘ) > 0. Then the infimal convolute f₁ □ … □ fₘ is a closed proper convex function on ℝ^n, and the infimum in the definition of (f₁ □ … □ fₘ)(x) is attained for each x. Moreover, (f₁ □ … □ fₘ)0+ = f₁ 0+ □ … □ fₘ 0+.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
theorem infimalConvolutionFamily_closed_proper_convex_recession
{n m : Nat}
{f f0_plus : Fin m → (Fin n → Real) → EReal}
(hclosed : ∀ i : Fin m, ClosedConvexFunction (f i))
(hproper : ∀ i : Fin m, ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (f i))
(hrec_i :
∀ i : Fin m,
Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) (f i)) =
epigraph (Set.univ : Set (Fin n → Real)) (f0_plus i))
(hpos_i : ∀ i : Fin m, PositivelyHomogeneous (f0_plus i))
(hproper0_i :
∀ i : Fin m,
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (f0_plus i))
(hzero :
∀ z : Fin m → (Fin n → Real),
(Finset.univ.sum (fun i => f0_plus i (z i)) ≤ (0 : EReal)) →
(Finset.univ.sum (fun i => f0_plus i (-(z i))) > (0 : EReal)) →
(Finset.univ.sum (fun i => z i) ≠ (0 : Fin n → Real)))
(hm : 0 < m) :
let fInf : (Fin n → Real) → EReal := infimalConvolutionFamily f
let fInf0_plus : (Fin n → Real) → EReal := infimalConvolutionFamily f0_plus
ClosedConvexFunction fInf ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) fInf ∧
(∀ x : Fin n → Real,
∃ x' : Fin m → (Fin n → Real),
Finset.univ.sum (fun i => x' i) = x ∧
fInf x = Finset.univ.sum (fun i => f i (x' i))) ∧
fInf0_plus = infimalConvolutionFamily f0_plus§9 — Some Closedness Criteria — Part 8
Lean authors: Zichen Wang, Wanli Ma, Qiming Dai, Zaiwen Wen.
In the two-function case, the zero-sum condition in the family corollary follows from the positivity hypothesis on f1_0_plus and f2_0_plus.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma hzero_two
{n : Nat} {f1_0_plus f2_0_plus : (Fin n → Real) → EReal}
(hpos : ∀ z : Fin n → Real, z ≠ 0 → f1_0_plus z + f2_0_plus (-z) > (0 : EReal)) :
∀ z : Fin 2 → (Fin n → Real),
(Finset.univ.sum (fun i =>
(if i = 0 then f1_0_plus else f2_0_plus) (z i)) ≤ (0 : EReal)) →
(Finset.univ.sum (fun i =>
(if i = 0 then f1_0_plus else f2_0_plus) (-(z i))) > (0 : EReal)) →
(Finset.univ.sum (fun i => z i) ≠ (0 : Fin n → Real))
For Fin 2, a decomposition x = x' 0 + x' 1 rewrites the sum over the family into the binary form.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma extract_two_attainer
{n : Nat} {f1 f2 : (Fin n → Real) → EReal}
{x : Fin n → Real} {x' : Fin 2 → (Fin n → Real)}
(hsum : Finset.univ.sum (fun i => x' i) = x) :
Finset.univ.sum (fun i => (if i = 0 then f1 else f2) (x' i)) =
f1 (x - x' 1) + f2 (x' 1)
Corollary 9.2.2. Let f1 and f2 be closed proper convex functions on ℝ^n such that (f1 0+)(z) + (f2 0+)(-z) > 0 for all z ≠ 0. Then infimalConvolution f1 f2 is a closed proper convex function, and for each x the infimum in (f1 square f2)(x) = inf_y { f1 (x - y) + f2 y } is attained by some y.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
theorem infimalConvolution_closed_proper_convex_recession
{n : Nat} {f1 f2 f1_0_plus f2_0_plus : (Fin n → Real) → EReal}
(hclosed1 : ClosedConvexFunction f1) (hclosed2 : ClosedConvexFunction f2)
(hproper1 : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f1)
(hproper2 : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f2)
(hpos :
∀ z : Fin n → Real, z ≠ 0 → f1_0_plus z + f2_0_plus (-z) > (0 : EReal))
(hrec1 :
Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) f1) =
epigraph (Set.univ : Set (Fin n → Real)) f1_0_plus)
(hrec2 :
Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) f2) =
epigraph (Set.univ : Set (Fin n → Real)) f2_0_plus)
(hpos1 : PositivelyHomogeneous f1_0_plus)
(hpos2 : PositivelyHomogeneous f2_0_plus)
(hproper0_1 : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f1_0_plus)
(hproper0_2 : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f2_0_plus) :
ClosedConvexFunction (infimalConvolution f1 f2) ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (infimalConvolution f1 f2) ∧
(∀ x : Fin n → Real,
∃ y : Fin n → Real, infimalConvolution f1 f2 x = f1 (x - y) + f2 y)
Rewrite the infimal convolution with an indicator as an infimum over the translate.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma infimalConvolution_indicator_neg_eq_sInf_translate
{n : Nat} {f : (Fin n → Real) → EReal} {C : Set (Fin n → Real)}
(hproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
∀ x : Fin n → Real,
infimalConvolution (indicatorFunction (-C)) f x =
sInf { z : EReal |
∃ y : Fin n → Real, y ∈ Set.image (fun c => c + x) C ∧ z = f y }
The indicator of a closed convex nonempty set is closed and proper.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma closedConvexFunction_indicator_neg
{n : Nat} {C : Set (Fin n → Real)}
(hCne : Set.Nonempty C) (hCclosed : IsClosed C) (hCconvex : Convex Real C) :
ClosedConvexFunction (indicatorFunction (-C)) ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (indicatorFunction (-C))
Positive scaling preserves membership in the recession cone.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma recessionCone_smul_pos_fin {n : Nat} {C : Set (Fin n → Real)} {y : Fin n → Real}
(hy : y ∈ Set.recessionCone C) {t : Real} (ht : 0 < t) :
t • y ∈ Set.recessionCone C
The recession cone of a convex set is convex (finite-dimensional version).
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma recessionCone_convex_fin {n : Nat} {C : Set (Fin n → Real)}
(hCconvex : Convex Real C) : Convex Real (Set.recessionCone C)
The indicator of the recession cone is positively homogeneous and proper.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma indicator_recession_posHom_proper {n : Nat} {C : Set (Fin n → Real)}
(hCconvex : Convex Real C) :
PositivelyHomogeneous (indicatorFunction (-Set.recessionCone C)) ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real))
(indicatorFunction (-Set.recessionCone C))
The recession cone of the indicator epigraph is the epigraph of the recession indicator.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma recessionCone_epigraph_indicator_neg {n : Nat} {C : Set (Fin n → Real)}
(hCne : Set.Nonempty C) :
Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) (indicatorFunction (-C))) =
epigraph (Set.univ : Set (Fin n → Real))
(indicatorFunction (-Set.recessionCone C))
Positivity from a recession-direction separation hypothesis.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma indicator_recession_hpos_from_noCommon
{n : Nat} {f0_plus : (Fin n → Real) → EReal} {C : Set (Fin n → Real)}
(hproper0 : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f0_plus)
(hnoCommon :
∀ z : Fin n → Real, z ≠ 0 →
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm z ∈
Set.recessionCone
(Set.image (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm C) →
f0_plus z > (0 : EReal)) :
∀ z : Fin n → Real, z ≠ 0 →
indicatorFunction (-Set.recessionCone C) z + f0_plus (-z) > (0 : EReal)
Example 9.2.2.2. For f = f₂ closed proper convex and f₁ the indicator function of -C with C nonempty closed convex, (f₁ □ f)(x) = inf { δ(x - y | -C) + f(y) | y ∈ ℝ^n } and equals inf { f(y) | y ∈ (C + x) }. If f and C have no common direction of recession, then the infimum over the translate C + x is attained for each x, and the resulting function of x is lower semicontinuous and convex.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
theorem infimalConvolution_indicator_neg_example
{n : Nat} {f f0_plus : (Fin n → Real) → EReal} {C : Set (Fin n → Real)}
(hclosed : ClosedConvexFunction f)
(hproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hrec2 :
Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) f) =
epigraph (Set.univ : Set (Fin n → Real)) f0_plus)
(hpos2 : PositivelyHomogeneous f0_plus)
(hproper0_2 : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f0_plus)
(hCne : Set.Nonempty C) (hCclosed : IsClosed C) (hCconvex : Convex Real C)
(hnoCommon :
∀ z : Fin n → Real, z ≠ 0 →
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm z ∈
Set.recessionCone
(Set.image (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm C) →
f0_plus z > (0 : EReal)) :
infimalConvolution (indicatorFunction (-C)) f =
(fun x =>
sInf { z : EReal |
∃ y : Fin n → Real, y ∈ Set.image (fun c => c + x) C ∧ z = f y }) ∧
ClosedConvexFunction (infimalConvolution (indicatorFunction (-C)) f) ∧
(∀ x : Fin n → Real,
∃ y : Fin n → Real,
y ∈ Set.image (fun c => c + x) C ∧
infimalConvolution (indicatorFunction (-C)) f x = f y)
The set of values of f above a point x in the coordinatewise order.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
def coordinatewiseInfSet {n : Nat} (f : (Fin n → Real) → EReal) (x : Fin n → Real) :
Set EReal :=
{ z : EReal | ∃ y : Fin n → Real, x ≤ y ∧ z = f y }
A larger base point yields a smaller coordinatewise infimum set.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma coordinatewiseInfSet_mono {n : Nat} {f : (Fin n → Real) → EReal}
{x x' : Fin n → Real} (hxx' : x ≤ x') :
coordinatewiseInfSet f x' ⊆ coordinatewiseInfSet f x
The coordinatewise infimum is bounded above by f.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma coordinatewiseInf_le_self {n : Nat} {f : (Fin n → Real) → EReal} (x : Fin n → Real) :
sInf (coordinatewiseInfSet f x) ≤ f x
The coordinatewise infimum is monotone.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma coordinatewiseInf_monotone {n : Nat} {f : (Fin n → Real) → EReal} :
Monotone (fun x => sInf (coordinatewiseInfSet f x))
Any monotone minorant of f lies below the coordinatewise infimum.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma coordinatewiseInf_greatest {n : Nat} {f : (Fin n → Real) → EReal}
{h : (Fin n → Real) → EReal}
(hle : ∀ x : Fin n → Real, h x ≤ f x) (hmono : Monotone h) :
∀ x : Fin n → Real, h x ≤ sInf (coordinatewiseInfSet f x)
The non-negative orthant is nonempty, closed, and convex.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma nonnegOrthant_closed_convex_nonempty {n : Nat} :
Set.Nonempty (Set.Ici (0 : Fin n → Real)) ∧
IsClosed (Set.Ici (0 : Fin n → Real)) ∧
Convex Real (Set.Ici (0 : Fin n → Real))
Translating the non-negative orthant by x gives the coordinatewise upper set.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma image_add_nonnegOrthant_eq_ge {n : Nat} (x : Fin n → Real) :
Set.image (fun c => c + x) (Set.Ici (0 : Fin n → Real)) =
{ y : Fin n → Real | x ≤ y }
The recession cone of the non-negative orthant lies in the orthant.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma recessionCone_nonnegOrthant_subset {n : Nat} :
Set.recessionCone (Set.Ici (0 : Fin n → Real)) ⊆ Set.Ici (0 : Fin n → Real)
Nonnegative recession directions yield the no-common-recession hypothesis.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma hnoCommon_of_hnoNonneg {n : Nat} {f0_plus : (Fin n → Real) → EReal}
(hnoNonneg :
∀ z : Fin n → Real, z ≠ 0 → (∀ i, 0 ≤ z i) → f0_plus z > (0 : EReal)) :
∀ z : Fin n → Real, z ≠ 0 →
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm z ∈
Set.recessionCone
(Set.image (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(Set.Ici (0 : Fin n → Real))) →
f0_plus z > (0 : EReal)
The coordinatewise infimum set is the translate-based infimum set.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma coordinatewiseInfSet_eq_translate {n : Nat} {f : (Fin n → Real) → EReal}
(x : Fin n → Real) :
coordinatewiseInfSet f x =
{ z : EReal |
∃ y : Fin n → Real,
y ∈ Set.image (fun c => c + x) (Set.Ici (0 : Fin n → Real)) ∧ z = f y }
Example 9.2.2.3. Taking C to be the non-negative orthant of ℝ^n, we have C + x = { y | y ≥ x } for each x. If f is a closed proper convex function on ℝ^n whose recession cone contains no non-negative non-zero vectors, then the infimum in g(x) = inf { f(y) | y ≥ x } is attained for each x, and g is a closed proper convex function. Moreover, g is the greatest function with g ≤ f and g coordinatewise non-decreasing.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
theorem coordinatewiseInfimum_closed_proper_convex
{n : Nat} {f f0_plus : (Fin n → Real) → EReal}
(hclosed : ClosedConvexFunction f)
(hproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hrec2 :
Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) f) =
epigraph (Set.univ : Set (Fin n → Real)) f0_plus)
(hpos2 : PositivelyHomogeneous f0_plus)
(hproper0_2 : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f0_plus)
(hnoNonneg :
∀ z : Fin n → Real, z ≠ 0 → (∀ i, 0 ≤ z i) → f0_plus z > (0 : EReal)) :
let g : (Fin n → Real) → EReal :=
fun x => sInf (coordinatewiseInfSet f x)
ClosedConvexFunction g ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) g ∧
(∀ x : Fin n → Real, ∃ y : Fin n → Real, x ≤ y ∧ g x = f y) ∧
(∀ x : Fin n → Real, g x ≤ f x) ∧
Monotone g ∧
(∀ h : (Fin n → Real) → EReal,
(∀ x : Fin n → Real, h x ≤ f x) → Monotone h →
∀ x : Fin n → Real, h x ≤ g x)
Non-⊥ is preserved by finite sums.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma finset_sum_ne_bot_of_forall {α : Type*} (s : Finset α)
(f : α → EReal) (h : ∀ a ∈ s, f a ≠ (⊥ : EReal)) : s.sum f ≠ (⊥ : EReal)
Non-⊤ is preserved by finite sums.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma finset_sum_ne_top_of_forall {α : Type*} (s : Finset α)
(f : α → EReal) (h : ∀ a ∈ s, f a ≠ (⊤ : EReal)) : s.sum f ≠ (⊤ : EReal)
The effective domain of a finite sum is the intersection of the effective domains, assuming no summand takes the value ⊥.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma effectiveDomain_sum_eq_iInter_univ
{n m : Nat} {f : Fin m → (Fin n → Real) → EReal}
(hnotbot : ∀ i : Fin m, ∀ x, f i x ≠ (⊥ : EReal)) :
effectiveDomain (Set.univ : Set (Fin n → Real))
(fun x => Finset.univ.sum (fun i => f i x)) =
⋂ i, effectiveDomain (Set.univ : Set (Fin n → Real)) (f i)§9 — Some Closedness Criteria — Part 9
Lean authors: Zichen Wang, Wanli Ma, Qiming Dai, Zaiwen Wen.
The relative interior of the effective domain of a finite sum is the intersection of the relative interiors.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma ri_effectiveDomain_sum_eq_iInter
{n m : Nat} {f : Fin m → (Fin n → Real) → EReal}
(hproper : ∀ i : Fin m, ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (f i)) :
let e := (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n))
let fSum : (Fin n → Real) → EReal := fun x => Finset.univ.sum (fun i => f i x)
(⋂ i,
euclideanRelativeInterior n
(Set.image e.symm
(effectiveDomain (Set.univ : Set (Fin n → Real)) (f i)))).Nonempty →
euclideanRelativeInterior n
(Set.image e.symm (effectiveDomain (Set.univ : Set (Fin n → Real)) fSum)) =
⋂ i,
euclideanRelativeInterior n
(Set.image e.symm
(effectiveDomain (Set.univ : Set (Fin n → Real)) (f i)))
A finite sum of closed proper convex functions is closed.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma closedConvexFunction_sum_of_closed
{n m : Nat} {f : Fin m → (Fin n → Real) → EReal}
(hclosed : ∀ i : Fin m, ClosedConvexFunction (f i))
(hproper : ∀ i : Fin m, ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (f i)) :
ClosedConvexFunction (fun x => Finset.univ.sum (fun i => f i x))
A finite sum of proper convex functions is proper if it is finite somewhere.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma properConvexFunctionOn_sum_of_exists_ne_top
{n m : Nat} {f : Fin m → (Fin n → Real) → EReal}
(hproper : ∀ i : Fin m, ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (f i))
(hexists : ∃ x : Fin n → Real, Finset.univ.sum (fun i => f i x) ≠ (⊤ : EReal)) :
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real))
(fun x => Finset.univ.sum (fun i => f i x))
Under a common relative-interior point, the closure of a finite sum is the sum of closures.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma convexFunctionClosure_sum_eq_sum_closure
{n m : Nat} {f : Fin m → (Fin n → Real) → EReal}
(hproper : ∀ i : Fin m, ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (f i)) :
let e := (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n))
let fSum : (Fin n → Real) → EReal := fun x => Finset.univ.sum (fun i => f i x)
(∃ x : EuclideanSpace Real (Fin n),
∀ i : Fin m,
x ∈
euclideanRelativeInterior n
(Set.image e.symm
(effectiveDomain (Set.univ : Set (Fin n → Real)) (f i)))) →
convexFunctionClosure fSum =
fun x => Finset.univ.sum (fun i => convexFunctionClosure (f i) x)
Theorem 9.3. Let f₁, …, fₘ be proper convex functions on ℝ^n. If every f_i is closed and f₁ + ··· + fₘ is not identically +∞, then the sum is a closed proper convex function and (f₁ + ··· + fₘ)0^+ = f₁ 0^+ + ··· + fₘ 0^+. If the f_i are not all closed, but there exists a point common to every ri (dom f_i), then cl (f₁ + ··· + fₘ) = cl f₁ + ··· + cl fₘ.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
theorem sum_closed_proper_convex_recession_and_closure
{n m : Nat} {f f0_plus : Fin m → (Fin n → Real) → EReal}
(hproper : ∀ i : Fin m, ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (f i)) :
let e := (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n))
let fSum : (Fin n → Real) → EReal := fun x => Finset.univ.sum (fun i => f i x)
let fSum0_plus : (Fin n → Real) → EReal := fun y => Finset.univ.sum (fun i => f0_plus i y)
((∀ i : Fin m, ClosedConvexFunction (f i)) →
(∃ x : Fin n → Real, fSum x ≠ (⊤ : EReal)) →
ClosedConvexFunction fSum ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) fSum ∧
fSum0_plus = fun y => Finset.univ.sum (fun i => f0_plus i y)) ∧
((∃ x : EuclideanSpace Real (Fin n),
∀ i : Fin m,
x ∈
euclideanRelativeInterior n
(Set.image e.symm
(effectiveDomain (Set.univ : Set (Fin n → Real)) (f i)))) →
convexFunctionClosure fSum =
fun x => Finset.univ.sum (fun i => convexFunctionClosure (f i) x))
The pointwise supremum of closed convex functions is closed.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma closedConvexFunction_iSup_of_closed
{n : Nat} {I : Type _} {f : I → (Fin n → Real) → EReal}
(hclosed : ∀ i, ClosedConvexFunction (f i)) :
ClosedConvexFunction (fun x => iSup (fun i => f i x))
A pointwise supremum of proper convex functions is proper if it is finite somewhere.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma properConvexFunctionOn_iSup_of_exists
{n : Nat} {I : Type _} {f : I → (Fin n → Real) → EReal}
(hproper : ∀ i : I, ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (f i)) :
let fSup : (Fin n → Real) → EReal := fun x => iSup (fun i => f i x)
(∃ x : Fin n → Real, fSup x ≠ (⊤ : EReal) ∧ fSup x ≠ (⊥ : EReal)) →
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) fSup
A common relative-interior point yields a point in all ri (epi f i) above the supremum.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
lemma mem_iInter_riEpigraph_of_common_ri_dom
{n : Nat} {I : Type _} {f : I → (Fin n → Real) → EReal}
(hconv : ∀ i : I, ConvexFunction (f i)) :
let e := (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n))
let fSup : (Fin n → Real) → EReal := fun x => iSup (fun i => f i x)
∀ {x0 : EuclideanSpace Real (Fin n)},
(∀ i : I,
x0 ∈
euclideanRelativeInterior n
(Set.image e.symm
(effectiveDomain (Set.univ : Set (Fin n → Real)) (f i)))) →
fSup (e.symm x0) ≠ (⊤ : EReal) →
fSup (e.symm x0) ≠ (⊥ : EReal) →
(appendAffineEquiv n 1
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
((EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)) x0),
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 1)).symm
(fun _ : Fin 1 => (fSup (e.symm x0)).toReal + 1)))
∈ ⋂ i, riEpigraph (f i)
Theorem 9.4. Let f_i be a proper convex function on ℝ^n for i ∈ I, and let f = sup { f_i | i ∈ I }. If f is finite somewhere and every f_i is closed, then f is closed and proper and f0^+ = sup { f_i0^+ | i ∈ I }. If the f_i are not all closed, but there exists a point common to every ri (dom f_i) such that f is finite there, then cl f = sup { cl f_i | i ∈ I }.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
theorem iSup_closed_proper_convex_recession_and_closure
{n : Nat} {I : Type _} {f f0_plus : I → (Fin n → Real) → EReal}
(hproper : ∀ i : I, ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (f i)) :
let e := (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n))
let fSup : (Fin n → Real) → EReal := fun x => iSup (fun i => f i x)
let fSup0_plus : (Fin n → Real) → EReal := fun y => iSup (fun i => f0_plus i y)
((∀ i : I, ClosedConvexFunction (f i)) →
(∃ x : Fin n → Real, fSup x ≠ (⊤ : EReal) ∧ fSup x ≠ (⊥ : EReal)) →
ClosedConvexFunction fSup ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) fSup ∧
fSup0_plus = fun y => iSup (fun i => f0_plus i y)) ∧
((∃ x : EuclideanSpace Real (Fin n),
(∀ i : I,
x ∈
euclideanRelativeInterior n
(Set.image e.symm
(effectiveDomain (Set.univ : Set (Fin n → Real)) (f i)))) ∧
fSup (e.symm x) ≠ (⊤ : EReal) ∧ fSup (e.symm x) ≠ (⊥ : EReal)) →
convexFunctionClosure fSup =
fun x => iSup (fun i => convexFunctionClosure (f i) x))§9 — Some Closedness Criteria — Part 10
Lean authors: Zichen Wang, Wanli Ma, Qiming Dai, Zaiwen Wen.
Precomposition with a linear map preserves closedness of convex functions.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma closedConvexFunction_precomp_linearMap {n m : Nat} {g : (Fin m → Real) → EReal}
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real)) :
ClosedConvexFunction g → ClosedConvexFunction (fun x => g (A x))
Effective domain of a precomposition is a preimage.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma effectiveDomain_precomp_linearMap {n m : Nat} {g : (Fin m → Real) → EReal}
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real)) :
effectiveDomain (Set.univ : Set (Fin n → Real)) (fun x => g (A x)) =
A ⁻¹' effectiveDomain (Set.univ : Set (Fin m → Real)) g
Relative interior of the effective domain under a linear preimage.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma ri_effectiveDomain_preimage_linearMap
{n m : Nat} {g : (Fin m → Real) → EReal}
(hgproper : ProperConvexFunctionOn (Set.univ : Set (Fin m → Real)) g)
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real))
(hri :
∃ x : Fin n → Real,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin m)).symm (A x) ∈
euclideanRelativeInterior m
(Set.image (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin m)).symm
(effectiveDomain (Set.univ : Set (Fin m → Real)) g))) :
let e_n := (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n))
let e_m := (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin m))
let A_e : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m) :=
(e_m.symm.toLinearMap).comp (A.comp e_n.toLinearMap)
euclideanRelativeInterior n
(Set.image e_n.symm
(effectiveDomain (Set.univ : Set (Fin n → Real)) (fun x => g (A x)))) =
A_e ⁻¹' euclideanRelativeInterior m
(Set.image e_m.symm (effectiveDomain (Set.univ : Set (Fin m → Real)) g))
Relative interior of the effective domain in preimage form under a linear map.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma ri_effectiveDomain_preimage_linearMap_preimage
{n m : Nat} {g : (Fin m → Real) → EReal}
(hgproper : ProperConvexFunctionOn (Set.univ : Set (Fin m → Real)) g)
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real))
(hri :
∃ x : Fin n → Real,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin m)).symm (A x) ∈
euclideanRelativeInterior m
(Set.image (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin m)).symm
(effectiveDomain (Set.univ : Set (Fin m → Real)) g))) :
let e_n := (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n))
let e_m := (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin m))
let A_e : EuclideanSpace Real (Fin n) →ₗ[Real] EuclideanSpace Real (Fin m) :=
(e_m.symm.toLinearMap).comp (A.comp e_n.toLinearMap)
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) (fun x => g (A x))) =
A_e ⁻¹' euclideanRelativeInterior m
((fun x : EuclideanSpace Real (Fin m) => (x : Fin m → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin m → Real)) g)
Relative interior of the effective domain is preserved by convex closure.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma ri_effectiveDomain_closure_eq {m : Nat} {g : (Fin m → Real) → EReal}
(hgproper : ProperConvexFunctionOn (Set.univ : Set (Fin m → Real)) g) :
euclideanRelativeInterior m
((fun x : EuclideanSpace Real (Fin m) => (x : Fin m → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin m → Real)) (convexFunctionClosure g)) =
euclideanRelativeInterior m
((fun x : EuclideanSpace Real (Fin m) => (x : Fin m → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin m → Real)) g)
Closure commutes with linear precomposition under a relative-interior preimage point.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma convexFunctionClosure_precomp_linearMap_eq
{n m : Nat} {g : (Fin m → Real) → EReal}
(hgproper : ProperConvexFunctionOn (Set.univ : Set (Fin m → Real)) g)
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real))
(hri :
∃ x : Fin n → Real,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin m)).symm (A x) ∈
euclideanRelativeInterior m
(Set.image (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin m)).symm
(effectiveDomain (Set.univ : Set (Fin m → Real)) g))) :
convexFunctionClosure (fun x => g (A x)) = fun x => convexFunctionClosure g (A x)
Theorem 9.5. Let A be a linear transformation from ℝ^n to ℝ^m, and let g be a proper convex function on ℝ^m such that g ∘ A is not identically +∞. If g is closed, then g ∘ A is closed and (g ∘ A)0^+ = (g0^+) ∘ A. If g is not closed, but A x lies in ri (dom g) for some x, then cl (g ∘ A) = (cl g) ∘ A.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
theorem linearMap_comp_closed_recession_and_closure
{n m : Nat} {g g0_plus : (Fin m → Real) → EReal}
(hgproper : ProperConvexFunctionOn (Set.univ : Set (Fin m → Real)) g)
(A : (Fin n → Real) →ₗ[Real] (Fin m → Real))
(_hfinite : ∃ x : Fin n → Real, g (A x) ≠ (⊤ : EReal)) :
let gA : (Fin n → Real) → EReal := fun x => g (A x)
let gA0_plus : (Fin n → Real) → EReal := fun x => g0_plus (A x)
(ClosedConvexFunction g →
ClosedConvexFunction gA ∧ gA0_plus = fun x => g0_plus (A x)) ∧
((¬ ClosedConvexFunction g) →
(∃ x : Fin n → Real,
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin m)).symm (A x) ∈
euclideanRelativeInterior m
(Set.image (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin m)).symm
(effectiveDomain (Set.univ : Set (Fin m → Real)) g))) →
convexFunctionClosure gA = fun x => convexFunctionClosure g (A x))
In the lifted cone, a closure point with zero tail must be the origin.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma tail_kernel_closure_cone_zero {n : Nat} {C : Set (Fin n → Real)}
(hCne : Set.Nonempty C) (hCclosed : IsClosed C) (hCconv : Convex Real C)
(hC0 : (0 : Fin n → Real) ∉ C) :
let e := (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n))
let C' : Set (EuclideanSpace Real (Fin n)) := Set.image e.symm C
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first v = 1 ∧ tail v ∈ C'}
let K : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S : Set (EuclideanSpace Real (Fin (1 + n))))
∀ v, v ∈ closure K → tail v = 0 → v = 0
The tail image of the lifted cone closes by adjoining recession directions.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma closure_tail_image_cone_eq_union_recession {n : Nat} {C : Set (Fin n → Real)}
(hCne : Set.Nonempty C) (hCclosed : IsClosed C) (hCconv : Convex Real C)
(hC0 : (0 : Fin n → Real) ∉ C) :
let e := (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n))
let C' : Set (EuclideanSpace Real (Fin n)) := Set.image e.symm C
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first v = 1 ∧ tail v ∈ C'}
let K : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S : Set (EuclideanSpace Real (Fin (1 + n))))
let A : EuclideanSpace Real (Fin (1 + n)) →ₗ[Real] EuclideanSpace Real (Fin n) :=
(LinearMap.snd (R := Real) (M := EuclideanSpace Real (Fin 1))
(M₂ := EuclideanSpace Real (Fin n))).comp
(appendAffineEquiv 1 n).symm.linear.toLinearMap
closure (A '' K) = (A '' K) ∪ Set.recessionCone C'
The generated cone equals positive scalings, up to union with a set containing 0.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma convexConeGenerated_union_recession_eq_iUnion_pos {n : Nat}
{C : Set (Fin n → Real)} (hCconv : Convex Real C) {recC : Set (Fin n → Real)}
(hrec0 : (0 : Fin n → Real) ∈ recC) :
let K : Set (Fin n → Real) := convexConeGenerated n C
K ∪ recC = (⋃ (t : Real) (_ : 0 < t), (t • C)) ∪ recC
Theorem 9.6. Let C be a non-empty closed convex set not containing the origin, and let K be the convex cone generated by C. Then closure K = K ∪ 0^+ C, and this set equals ⋃ { λ C | λ > 0 or λ = 0^+ }.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
theorem closure_convexConeGenerated_eq_union_recessionCone
{n : Nat} {C : Set (Fin n → Real)}
(hCne : Set.Nonempty C) (hCclosed : IsClosed C) (hCconv : Convex Real C)
(hC0 : (0 : Fin n → Real) ∉ C) :
let e := (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n))
let C' : Set (EuclideanSpace Real (Fin n)) := Set.image e.symm C
let recC : Set (Fin n → Real) := Set.image e (Set.recessionCone C')
let K : Set (Fin n → Real) := convexConeGenerated n C
closure K = K ∪ recC ∧
K ∪ recC =
(⋃ (t : Real) (_ : 0 < t), (t • C)) ∪ recC
Boundedness forces the recession cone of the Euclidean image to be {0}.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma recCone_eq_singleton_zero_of_bounded_image {n : Nat} {C : Set (Fin n → Real)}
(hCne : Set.Nonempty C) (hCclosed : IsClosed C) (hCconv : Convex Real C)
(hCbdd : Bornology.IsBounded C) :
let e := (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n))
let C' : Set (EuclideanSpace Real (Fin n)) := Set.image e.symm C
Set.recessionCone C' = {0}
Corollary 9.6.1. If C is a non-empty closed bounded convex set not containing the origin, then the convex cone K generated by C is closed.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
theorem closed_convexConeGenerated_of_bounded {n : Nat} {C : Set (Fin n → Real)}
(hCne : Set.Nonempty C) (hCclosed : IsClosed C) (hCbdd : Bornology.IsBounded C)
(hCconv : Convex Real C) (hC0 : (0 : Fin n → Real) ∉ C) :
IsClosed (convexConeGenerated n C)
The vertical line through ![1, 0] is the set of points with first coordinate 1.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma lineSet_eq_first_coord :
{x : Fin 2 → Real | ∃ t : Real, x = ![1, 0] + t • ![0, 1]} =
{x : Fin 2 → Real | x 0 = 1}
The vertical line {(1, t)} is closed.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma closed_lineSet :
IsClosed {x : Fin 2 → Real | ∃ t : Real, x = ![1, 0] + t • ![0, 1]}
The vertical line {(1, t)} is convex.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma convex_lineSet :
Convex Real {x : Fin 2 → Real | ∃ t : Real, x = ![1, 0] + t • ![0, 1]}
Nonzero points in the cone over the vertical line have positive first coordinate.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma cone_line_pos_first {x : Fin 2 → Real}
(hx : x ∈ convexConeGenerated 2 {x : Fin 2 → Real | ∃ t : Real, x = ![1, 0] + t • ![0, 1]})
(hx0 : x ≠ 0) : 0 < x 0
Points with positive first coordinate lie in the cone over the vertical line.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma pos_first_mem_cone_line {x : Fin 2 → Real} (hx : 0 < x 0) :
x ∈ convexConeGenerated 2 {x : Fin 2 → Real | ∃ t : Real, x = ![1, 0] + t • ![0, 1]}
The point (0, 1) lies in the closure of the cone over the vertical line.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma closure_cone_line_has_point :
(![0, 1] : Fin 2 → Real) ∈
closure (convexConeGenerated 2
{x : Fin 2 → Real | ∃ t : Real, x = ![1, 0] + t • ![0, 1]})§9 — Some Closedness Criteria — Part 11
Lean authors: Zichen Wang, Wanli Ma, Qiming Dai, Zaiwen Wen.
A pointwise true predicate holds eventually.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma eventually_of_forall {α} {l : Filter α} {p : α → Prop} (hp : ∀ x, p x) :
Filter.Eventually p l
The vertical line {(1, t)} is unbounded.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma lineSet_unbounded :
¬ Bornology.IsBounded {x : Fin 2 → Real | ∃ t : Real, x = ![1, 0] + t • ![0, 1]}
The recession cone of the closed ball centered at ![1, 0] is {0}.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma recC_closedBall_eq_zero :
let C : Set (Fin 2 → Real) :=
Metric.closedBall (![1, 0]) ‖(![1, 0] : Fin 2 → Real)‖
let e := (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin 2))
let C' : Set (EuclideanSpace Real (Fin 2)) := Set.image e.symm C
let recC : Set (Fin 2 → Real) := Set.image e (Set.recessionCone C')
recC = ({0} : Set (Fin 2 → Real))
Nonzero points in the cone over the closed ball centered at ![1, 0] have positive first coordinate.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma cone_closedBall_imp_pos_first {x : Fin 2 → Real}
(hx :
x ∈ convexConeGenerated 2
(Metric.closedBall (![1, 0]) ‖(![1, 0] : Fin 2 → Real)‖))
(hx0 : x ≠ 0) : 0 ≤ x 0
Points with positive first coordinate lie in the cone over the closed ball centered at ![1, 0].
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma pos_first_mem_cone_closedBall {x : Fin 2 → Real} (hx : 0 < x 0) :
x ∈ convexConeGenerated 2
(Metric.closedBall (![1, 0]) ‖(![1, 0] : Fin 2 → Real)‖)
The point (0, 1) lies in the closure of the cone over the closed ball centered at ![1, 0].
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma closure_cone_closedBall_has_point :
(![0, 1] : Fin 2 → Real) ∈
closure (convexConeGenerated 2
(Metric.closedBall (![1, 0]) ‖(![1, 0] : Fin 2 → Real)‖))
Remark 9.6.1.2. The need for the condition 0 ∉ C in Theorem 9.6 and Corollary 9.6.1 is shown by the case where C is a closed ball with the origin on its boundary. The need for the boundedness assumption in Corollary 9.6.1 is shown by the case where C is a line not passing through the origin.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma counterexample_origin_boundary_or_unbounded_line :
(∃ (n : Nat) (x0 : Fin n → Real) (r : Real),
let C : Set (Fin n → Real) := Metric.closedBall x0 r
(0 : Fin n → Real) ∈ Metric.sphere x0 r ∧
IsClosed C ∧ Convex Real C ∧ True) ∧
(∃ (n : Nat) (x0 v : Fin n → Real),
let C : Set (Fin n → Real) := {x | ∃ t : Real, x = x0 + t • v}
(0 : Fin n → Real) ∉ C ∧ IsClosed C ∧ Convex Real C ∧
(¬ Bornology.IsBounded C) ∧ ¬ IsClosed (convexConeGenerated n C))
Coordinate version of prodLinearEquiv_append landing in Fin (n + 1) → Real.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
def prodLinearEquiv_append_coord (n : Nat) :
((Fin n → Real) × Real) ≃ₗ[Real] (Fin (n + 1) → Real) :=
(prodLinearEquiv_append (n := n)).trans
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (n + 1))).toLinearEquiv
The epigraph of a positive right scalar multiple is the scaled epigraph.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma epigraph_rightScalarMultiple_eq_smul {n : Nat} {f : (Fin n → Real) → EReal}
(hf : ConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
{lam : Real} (hlam : 0 < lam) :
epigraph (Set.univ : Set (Fin n → Real)) (rightScalarMultiple f lam) =
(fun p : (Fin n → Real) × Real => lam • p) ''
epigraph (Set.univ : Set (Fin n → Real)) f
Scaling the embedded epigraph corresponds to the right scalar multiple.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma smul_embedded_epigraph_eq_embedded_epigraph_rightScalarMultiple {n : Nat}
{f : (Fin n → Real) → EReal}
(hf : ConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
{lam : Real} (hlam : 0 < lam) :
(lam • ((prodLinearEquiv_append_coord (n := n)) ''
epigraph (Set.univ : Set (Fin n → Real)) f)) =
(prodLinearEquiv_append_coord (n := n)) ''
epigraph (Set.univ : Set (Fin n → Real)) (rightScalarMultiple f lam)
The closure of the cone over the embedded epigraph is the union of embedded right-scalar multiples and the embedded recession epigraph.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma closure_convexConeGenerated_embedded_epigraph_eq_union {n : Nat}
{f f0_plus : (Fin n → Real) → EReal}
(hf : ConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hrec :
Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) f) =
epigraph (Set.univ : Set (Fin n → Real)) f0_plus)
(hCne :
Set.Nonempty
((prodLinearEquiv_append_coord (n := n)) ''
epigraph (Set.univ : Set (Fin n → Real)) f))
(hCclosed :
IsClosed
((prodLinearEquiv_append_coord (n := n)) ''
epigraph (Set.univ : Set (Fin n → Real)) f))
(hCconv :
Convex Real
((prodLinearEquiv_append_coord (n := n)) ''
epigraph (Set.univ : Set (Fin n → Real)) f))
(hC0 :
(0 : Fin (n + 1) → Real) ∉
((prodLinearEquiv_append_coord (n := n)) ''
epigraph (Set.univ : Set (Fin n → Real)) f)) :
let C : Set (Fin (n + 1) → Real) :=
(prodLinearEquiv_append_coord (n := n)) '' epigraph (Set.univ : Set (Fin n → Real)) f
closure (convexConeGenerated (n + 1) C) =
(⋃ (lam : Real) (_ : 0 < lam),
(prodLinearEquiv_append_coord (n := n)) ''
epigraph (Set.univ : Set (Fin n → Real)) (rightScalarMultiple f lam)) ∪
(prodLinearEquiv_append_coord (n := n)) ''
epigraph (Set.univ : Set (Fin n → Real)) f0_plus
The embedded image of the generated epigraph cone equals the generated cone of the embedded epigraph.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma convexConeGenerated_embedded_epigraph_eq_image_convexConeGeneratedEpigraph {n : Nat}
{f : (Fin n → Real) → EReal}
(hf : ConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
let C : Set (Fin (n + 1) → Real) :=
(prodLinearEquiv_append_coord (n := n)) '' epigraph (Set.univ : Set (Fin n → Real)) f
(prodLinearEquiv_append_coord (n := n)) '' convexConeGeneratedEpigraph f =
convexConeGenerated (n + 1) C
The generated epigraph cone lies in the epigraph of its inf-section.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma convexConeGeneratedEpigraph_subset_epigraph_posHom {n : Nat}
{f : (Fin n → Real) → EReal} :
let k := positivelyHomogeneousConvexFunctionGenerated f
convexConeGeneratedEpigraph f ⊆ epigraph (Set.univ : Set (Fin n → Real)) k
Remove the prodLinearEquiv_append_coord embedding from the closure/union formula.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma closure_convexConeGeneratedEpigraph_eq_union {n : Nat}
{f f0_plus : (Fin n → Real) → EReal}
(hclosure_union' :
closure ((prodLinearEquiv_append_coord (n := n)) '' convexConeGeneratedEpigraph f) =
(⋃ (lam : Real) (_ : 0 < lam),
(prodLinearEquiv_append_coord (n := n)) ''
epigraph (Set.univ : Set (Fin n → Real)) (rightScalarMultiple f lam)) ∪
(prodLinearEquiv_append_coord (n := n)) ''
epigraph (Set.univ : Set (Fin n → Real)) f0_plus) :
closure (convexConeGeneratedEpigraph f) =
(⋃ (lam : Real) (_ : 0 < lam),
epigraph (Set.univ : Set (Fin n → Real)) (rightScalarMultiple f lam)) ∪
epigraph (Set.univ : Set (Fin n → Real)) f0_plus
The positively homogeneous hull is convex and has a nonempty epigraph.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma posHomGenerated_convex_and_nonempty {n : Nat} {f : (Fin n → Real) → EReal}
(hfconv : ConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
let k := positivelyHomogeneousConvexFunctionGenerated f
ConvexFunctionOn (Set.univ : Set (Fin n → Real)) k ∧
Set.Nonempty (epigraph (Set.univ : Set (Fin n → Real)) k)
If 0 lies in the effective domain, the positive-scaling infimum formula holds.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma posHomGenerated_formula_pos {n : Nat} {f : (Fin n → Real) → EReal}
(hfconv : ConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hfinite : ∃ x, f x ≠ (⊤ : EReal))
(hdom0 : (0 : Fin n → Real) ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) f) :
let k := positivelyHomogeneousConvexFunctionGenerated f
∀ x : Fin n → Real,
k x = sInf { z : EReal | ∃ lam : Real, 0 < lam ∧ z = rightScalarMultiple f lam x }
A proper epigraph has no negative vertical recession direction.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
lemma no_neg_vertical_recession_epigraph {n : Nat} {f : (Fin n → Real) → EReal}
(hproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) {μ : Real}
(hμ : μ < 0) :
(0, μ) ∉ Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) f)§9 — Some Closedness Criteria — Part 12
Lean authors: Zichen Wang, Wanli Ma, Qiming Dai, Zaiwen Wen.
Theorem 9.7. Let f be a closed proper convex function on ℝ^n with f(0) > 0, and let k be the positively homogeneous convex function generated by f. Then k is proper and (convexFunctionClosure k)(x) is the infimum of (f λ)(x) over λ > 0 together with λ = 0^+ (encoded here as f0_plus), the infimum being attained for each x. If 0 ∈ dom f, then k is closed and the λ = 0^+ term can be omitted (but the infimum might not be attained).
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
theorem closure_posHomogeneousGenerated_infimum_rightScalarMultiple {n : Nat}
{f f0_plus : (Fin n → Real) → EReal}
(hfclosed : ClosedConvexFunction f)
(hfproper : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(hf0 : f (0 : Fin n → Real) > (0 : EReal))
(hrec :
Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) f) =
epigraph (Set.univ : Set (Fin n → Real)) f0_plus) :
let k := positivelyHomogeneousConvexFunctionGenerated f
(ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) k) ∧
(∀ x : Fin n → Real,
convexFunctionClosure k x =
sInf
(Set.insert (f0_plus x)
{ z : EReal | ∃ lam : Real, 0 < lam ∧ z = rightScalarMultiple f lam x })) ∧
(∀ x : Fin n → Real,
∃ z ∈
Set.insert (f0_plus x)
{ z : EReal | ∃ lam : Real, 0 < lam ∧ z = rightScalarMultiple f lam x },
convexFunctionClosure k x = z) ∧
((0 : Fin n → Real) ∈ effectiveDomain (Set.univ : Set (Fin n → Real)) f →
ClosedConvexFunction k ∧
(∀ x : Fin n → Real,
k x =
sInf { z : EReal | ∃ lam : Real, 0 < lam ∧ z = rightScalarMultiple f lam x }))§9 — Some Closedness Criteria — Part 13
Lean authors: Zichen Wang, Wanli Ma, Qiming Dai, Zaiwen Wen.
Corollary 9.7.1. Let C be a closed convex set in ℝ^n containing 0. The gauge function gamma(· | C) of C is closed. One has {x | gamma(x | C) ≤ λ} = λ C for any λ > 0, and {x | gamma(x | C) = 0} = 0^+ C.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
open scoped RealInnerProductSpace
open Filter
theorem gaugeFunction_closed_and_level_sets {n : Nat} {C : Set (Fin n → Real)}
(hCclosed : IsClosed C) (hCconv : Convex Real C) (hC0 : (0 : Fin n → Real) ∈ C)
(hCabs : ∀ x : Fin n → Real, ∃ lam : Real, 0 ≤ lam ∧ x ∈ (fun y => lam • y) '' C) :
let e := (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n))
let C' : Set (EuclideanSpace Real (Fin n)) := Set.image e.symm C
let recC : Set (Fin n → Real) := Set.image e (Set.recessionCone C')
ClosedConvexFunction (fun x => (gaugeFunction C x : EReal)) ∧
(∀ lam : Real, 0 < lam → {x : Fin n → Real | gaugeFunction C x ≤ lam} = lam • C) ∧
{x : Fin n → Real | gaugeFunction C x = 0} = recC§9 — Some Closedness Criteria — Part 14
Lean authors: Zichen Wang, Wanli Ma, Qiming Dai, Zaiwen Wen.
Closure of the lifted cone over convexHull equals the closure of the sum of lifted cones when the index type is Fin (Nat.succ m).
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma closure_cone_over_convexHull_eq_closure_sum_cones_succ (n m : Nat)
(C : Fin (Nat.succ m) → Set (EuclideanSpace Real (Fin n)))
(hCne : ∀ i, Set.Nonempty (C i)) (hCconv : ∀ i, Convex Real (C i)) :
let C0 : Set (EuclideanSpace Real (Fin n)) := convexHull Real (⋃ i, C i)
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S0 : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first v = 1 ∧ tail v ∈ C0}
let K0 : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S0 : Set (EuclideanSpace Real (Fin (1 + n))))
let S : Fin (Nat.succ m) → Set (EuclideanSpace Real (Fin (1 + n))) :=
fun i => {v | first v = 1 ∧ tail v ∈ C i}
let K : Fin (Nat.succ m) → Set (EuclideanSpace Real (Fin (1 + n))) :=
fun i => (ConvexCone.hull Real (S i) : Set (EuclideanSpace Real (Fin (1 + n))))
closure K0 = closure (∑ i, K i)
Points in the closure of a lifted cone have nonnegative first coordinate.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma first_nonneg_of_mem_closure_lifted_cone {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} (hCne : C.Nonempty) (hCclosed : IsClosed C)
(hCconv : Convex Real C) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first v = 1 ∧ tail v ∈ C}
let K : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S : Set (EuclideanSpace Real (Fin (1 + n))))
∀ v, v ∈ closure K → 0 ≤ first v
The lifted cone over a nonempty set is nonempty.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma lifted_cone_nonempty {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} (hCne : C.Nonempty) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first v = 1 ∧ tail v ∈ C}
let K : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S : Set (EuclideanSpace Real (Fin (1 + n))))
Set.Nonempty K
The coordinate image of a lifted cone is a convex cone.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma lifted_cone_isConvexCone {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first v = 1 ∧ tail v ∈ C}
let K : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S : Set (EuclideanSpace Real (Fin (1 + n))))
IsConvexCone (1 + n)
(Set.image (EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))) K)
Zero-sum families in closures of lifted cones are lineality directions.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma lineality_of_zero_sum_closure_lifted_cones {n m : Nat}
(C : Fin m → Set (EuclideanSpace Real (Fin n)))
(hCne : ∀ i, Set.Nonempty (C i)) (hCclosed : ∀ i, IsClosed (C i))
(hCconv : ∀ i, Convex Real (C i))
(hlineal :
∀ z : Fin m → EuclideanSpace Real (Fin n),
(∀ i, z i ∈ Set.recessionCone (C i)) →
(∑ i, z i) = 0 →
∀ i, z i ∈ Set.linealitySpace (C i)) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S : Fin m → Set (EuclideanSpace Real (Fin (1 + n))) :=
fun i => {v | first v = 1 ∧ tail v ∈ C i}
let K : Fin m → Set (EuclideanSpace Real (Fin (1 + n))) :=
fun i => (ConvexCone.hull Real (S i) : Set (EuclideanSpace Real (Fin (1 + n))))
∀ z : Fin m → EuclideanSpace Real (Fin (1 + n)),
(∀ i, z i ∈ closure (K i)) →
(∑ i, z i) = 0 →
∀ i, z i ∈ Set.linealitySpace (closure (K i))
Closure of the sum of lifted cones equals the sum of their closures.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma closure_sum_cones_eq_sum_closure_cones {n m : Nat}
(C : Fin m → Set (EuclideanSpace Real (Fin n)))
(hCne : ∀ i, Set.Nonempty (C i)) (hCclosed : ∀ i, IsClosed (C i))
(hCconv : ∀ i, Convex Real (C i))
(hlineal :
∀ z : Fin m → EuclideanSpace Real (Fin n),
(∀ i, z i ∈ Set.recessionCone (C i)) →
(∑ i, z i) = 0 →
∀ i, z i ∈ Set.linealitySpace (C i)) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S : Fin m → Set (EuclideanSpace Real (Fin (1 + n))) :=
fun i => {v | first v = 1 ∧ tail v ∈ C i}
let K : Fin m → Set (EuclideanSpace Real (Fin (1 + n))) :=
fun i => (ConvexCone.hull Real (S i) : Set (EuclideanSpace Real (Fin (1 + n))))
closure (∑ i, K i) = ∑ i, closure (K i)
The closure of the lifted cone over C0 agrees with the closure of the lifted cone over closure C0.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma closure_lifted_cone_eq_closure_lifted_cone_closure {n : Nat}
(C0 : Set (EuclideanSpace Real (Fin n))) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S0 : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first v = 1 ∧ tail v ∈ C0}
let K0 : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S0 : Set (EuclideanSpace Real (Fin (1 + n))))
let S0bar : Set (EuclideanSpace Real (Fin (1 + n))) :=
{v | first v = 1 ∧ tail v ∈ closure C0}
let K0bar : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S0bar : Set (EuclideanSpace Real (Fin (1 + n))))
closure K0 = closure K0bar
The first = 1 slice of the closure of the lifted cone over C0 is closure C0.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma closure_C0_eq_first_one_slice_closure_lifted_cone {n : Nat}
{C0 : Set (EuclideanSpace Real (Fin n))} (hC0conv : Convex Real C0) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S0 : Set (EuclideanSpace Real (Fin (1 + n))) := {v | first v = 1 ∧ tail v ∈ C0}
let K0 : Set (EuclideanSpace Real (Fin (1 + n))) :=
(ConvexCone.hull Real S0 : Set (EuclideanSpace Real (Fin (1 + n))))
closure C0 =
{x | ∃ v, v ∈ closure K0 ∧ first v = 1 ∧ tail v = x}
The first = 1 slice of the sum of closures of lifted cones.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma mem_sum_closure_cones_first_eq_one_iff {n m : Nat}
(C : Fin m → Set (EuclideanSpace Real (Fin n)))
(hCne : ∀ i, Set.Nonempty (C i)) (hCclosed : ∀ i, IsClosed (C i))
(hCconv : ∀ i, Convex Real (C i)) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S : Fin m → Set (EuclideanSpace Real (Fin (1 + n))) :=
fun i => {v | first v = 1 ∧ tail v ∈ C i}
let K : Fin m → Set (EuclideanSpace Real (Fin (1 + n))) :=
fun i => (ConvexCone.hull Real (S i) : Set (EuclideanSpace Real (Fin (1 + n))))
∀ v, first v = 1 →
(v ∈ ∑ i, closure (K i) ↔
∃ (lam : Fin m → ℝ), (∀ i, 0 ≤ lam i) ∧ (∑ i, lam i) = 1 ∧
tail v ∈ ∑ i, (if lam i = 0 then Set.recessionCone (C i) else lam i • (C i)))
The first = 0 slice of the sum of closures of lifted cones.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma mem_sum_closure_cones_first_eq_zero_iff {n m : Nat}
(C : Fin m → Set (EuclideanSpace Real (Fin n)))
(hCne : ∀ i, Set.Nonempty (C i)) (hCclosed : ∀ i, IsClosed (C i))
(hCconv : ∀ i, Convex Real (C i)) :
let coords : EuclideanSpace Real (Fin (1 + n)) → (Fin (1 + n) → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin (1 + n))
let first : EuclideanSpace Real (Fin (1 + n)) → Real := fun v => coords v 0
let tail : EuclideanSpace Real (Fin (1 + n)) → EuclideanSpace Real (Fin n) :=
fun v =>
(EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)).symm
(fun i => coords v (Fin.natAdd 1 i))
let S : Fin m → Set (EuclideanSpace Real (Fin (1 + n))) :=
fun i => {v | first v = 1 ∧ tail v ∈ C i}
let K : Fin m → Set (EuclideanSpace Real (Fin (1 + n))) :=
fun i => (ConvexCone.hull Real (S i) : Set (EuclideanSpace Real (Fin (1 + n))))
∀ v, first v = 0 →
(v ∈ ∑ i, closure (K i) ↔ tail v ∈ ∑ i, Set.recessionCone (C i))§9 — Some Closedness Criteria — Part 15
Lean authors: Zichen Wang, Wanli Ma, Qiming Dai, Zaiwen Wen.
Theorem 9.8. Let C1, ..., Cm be nonempty closed convex sets in ℝ^n such that whenever z1, ..., zm satisfy zi ∈ 0^+ Ci and z1 + ... + zm = 0, each zi lies in the lineality space of Ci. Let C = conv (C1 ∪ ... ∪ Cm). Then cl C = ⋃ {lam1 C1 + ... + lamm Cm | lam_i ≥ 0^+, lam1 + ... + lamm = 1}, where lam_i ≥ 0^+ means lam_i Ci is interpreted as 0^+ Ci when lam_i = 0. Moreover, 0^+ (cl C) = 0^+ C1 + ... + 0^+ Cm.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
theorem closure_convexHull_iUnion_eq_iUnion_weighted_sum_and_recessionCone
{n m : Nat} (C : Fin m → Set (EuclideanSpace Real (Fin n)))
(hCne : ∀ i, Set.Nonempty (C i)) (hCclosed : ∀ i, IsClosed (C i))
(hCconv : ∀ i, Convex Real (C i))
(hlineal :
∀ z : Fin m → EuclideanSpace Real (Fin n),
(∀ i, z i ∈ Set.recessionCone (C i)) →
(∑ i, z i) = 0 →
∀ i, z i ∈ Set.linealitySpace (C i))
(hm : 0 < m) :
let C0 : Set (EuclideanSpace Real (Fin n)) := convexHull Real (⋃ i, C i)
closure C0 =
⋃ (lam : Fin m → ℝ)
(_ : (∀ i, 0 ≤ lam i) ∧ (∑ i, lam i) = 1),
(∑ i, (if lam i = 0 then Set.recessionCone (C i) else lam i • (C i))) ∧
Set.recessionCone (closure C0) = ∑ i, Set.recessionCone (C i)
Finite sums of recession cone elements stay in the recession cone.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma recessionCone_sum_mem_finset {n m : Nat} {C : Set (EuclideanSpace Real (Fin n))}
(hCconv : Convex Real C) {s : Finset (Fin m)}
(z : Fin m → EuclideanSpace Real (Fin n))
(hz : ∀ i ∈ s, z i ∈ Set.recessionCone C) :
(Finset.sum s (fun i => z i)) ∈ Set.recessionCone C
Zero-sum vectors in a common recession cone give lineality directions.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma linealitySpace_of_sum_zero_same_recession {n m : Nat}
(C : Fin m → Set (EuclideanSpace Real (Fin n))) (K : Set (EuclideanSpace Real (Fin n)))
(hCconv : ∀ i, Convex Real (C i)) (hK : ∀ i, Set.recessionCone (C i) = K) :
∀ z : Fin m → EuclideanSpace Real (Fin n),
(∀ i, z i ∈ K) → (∑ i, z i) = 0 → ∀ i, z i ∈ Set.linealitySpace (C i)
A positive weight absorbs the recession cone.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma absorb_recessionCone_into_positive_weight {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} {K : Set (EuclideanSpace Real (Fin n))}
(hK : Set.recessionCone C = K) {t : ℝ} (ht : 0 < t) :
K + t • C = t • C
Weighted Minkowski sums lie in the convex hull of the union.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma weighted_sum_subset_convexHull {n m : Nat}
(C : Fin m → Set (EuclideanSpace Real (Fin n))) {lam : Fin m → ℝ}
(hlam : ∀ i, 0 ≤ lam i) (hsum : ∑ i, lam i = 1) :
(∑ i, lam i • C i) ⊆ convexHull Real (⋃ i, C i)
A weighted recession-cone sum can be reduced to the usual weighted sum.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma mem_sum_if_recessionCone {n m : Nat}
(C : Fin m → Set (EuclideanSpace Real (Fin n))) (K : Set (EuclideanSpace Real (Fin n)))
(hCne : ∀ i, Set.Nonempty (C i)) (hCconv : ∀ i, Convex Real (C i))
(hK : ∀ i, Set.recessionCone (C i) = K) {lam : Fin m → ℝ}
(hlam : ∀ i, 0 ≤ lam i) (hsum : ∑ i, lam i = 1)
{x : EuclideanSpace Real (Fin n)}
(hx : x ∈ ∑ i, (if lam i = 0 then K else lam i • C i)) :
x ∈ ∑ i, lam i • C i
The sum of a nonempty family of equal recession cones is the cone.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma sum_recessionCone_eq_K {n m : Nat}
(C : Fin m → Set (EuclideanSpace Real (Fin n))) (K : Set (EuclideanSpace Real (Fin n)))
(hCconv : ∀ i, Convex Real (C i)) (hK : ∀ i, Set.recessionCone (C i) = K) (hm : 0 < m) :
(∑ i, Set.recessionCone (C i)) = K
Corollary 9.8.1. If C1, ..., Cm are nonempty closed convex sets in ℝ^n all having the same recession cone K, then C = conv (C1 ∪ ... ∪ Cm) is closed and has K as its recession cone.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
theorem convexHull_iUnion_closed_and_recessionCone_eq {n m : Nat}
(C : Fin m → Set (EuclideanSpace Real (Fin n))) (K : Set (EuclideanSpace Real (Fin n)))
(hCne : ∀ i, Set.Nonempty (C i)) (hCclosed : ∀ i, IsClosed (C i))
(hCconv : ∀ i, Convex Real (C i))
(hK : ∀ i, Set.recessionCone (C i) = K) (hm : 0 < m) :
let C0 : Set (EuclideanSpace Real (Fin n)) := convexHull Real (⋃ i, C i)
IsClosed C0 ∧ Set.recessionCone C0 = K
Replace empty members of a finite family by a fixed set without changing the union.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma iUnion_replace_empty_eq {n m : Nat}
(C : Fin m → Set (EuclideanSpace Real (Fin n))) (i0 : Fin m)
[DecidablePred fun i => (C i).Nonempty] :
(⋃ i, (if (C i).Nonempty then C i else C i0)) = ⋃ i, C i
Bounded closed convex sets have recession cone {0}.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma recessionCone_eq_singleton_zero_of_bounded {n : Nat}
{C : Set (EuclideanSpace Real (Fin n))} (hCne : C.Nonempty) (hCclosed : IsClosed C)
(hCconv : Convex Real C) (hCbdd : Bornology.IsBounded C) :
Set.recessionCone C = {0}
If some member of a finite family is nonempty, the convex hull of the union is nonempty.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma convexHull_iUnion_nonempty {n m : Nat}
(C : Fin m → Set (EuclideanSpace Real (Fin n))) (hne : ∃ i, (C i).Nonempty) :
(convexHull Real (⋃ i, C i)).Nonempty
Corollary 9.8.2 9.8.2.1. If C1, ..., Cm are closed bounded convex sets in ℝ^n, then conv (C1 ∪ ... ∪ Cm) is closed and bounded.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
theorem convexHull_iUnion_closed_and_bounded {n m : Nat}
(C : Fin m → Set (EuclideanSpace Real (Fin n)))
(hCclosed : ∀ i, IsClosed (C i)) (hCbdd : ∀ i, Bornology.IsBounded (C i))
(hCconv : ∀ i, Convex Real (C i)) :
IsClosed (convexHull Real (⋃ i, C i)) ∧
Bornology.IsBounded (convexHull Real (⋃ i, C i))
Closed proper convex functions have closed, convex, nonempty epigraphs.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma epigraph_family_closed_convex_nonempty {n m : Nat}
{f : Fin m → (Fin n → Real) → EReal}
(hclosed : ∀ i, ClosedConvexFunction (f i))
(hproper : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (f i)) :
∀ i,
Set.Nonempty (epigraph (Set.univ : Set (Fin n → Real)) (f i)) ∧
IsClosed (epigraph (Set.univ : Set (Fin n → Real)) (f i)) ∧
Convex ℝ (epigraph (Set.univ : Set (Fin n → Real)) (f i))
The convex hull of a union of epigraphs lies in the epigraph of any convex minorant.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma convexHull_union_epigraph_subset_epigraph_of_minorant {n m : Nat}
{f : Fin m → (Fin n → Real) → EReal} {h : (Fin n → Real) → EReal}
(hh : ConvexFunctionOn (Set.univ : Set (Fin n → Real)) h)
(hle : ∀ i, h ≤ f i) :
convexHull ℝ (⋃ i, epigraph (Set.univ : Set (Fin n → Real)) (f i)) ⊆
epigraph (Set.univ : Set (Fin n → Real)) h
Recession cones of the epigraph family agree with the common recession function.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma recessionCone_epigraph_eq_epigraph_k {n m : Nat}
{f : Fin m → (Fin n → Real) → EReal} {k : (Fin n → Real) → EReal}
(hconv : ∀ i, Convex ℝ (epigraph (Set.univ : Set (Fin n → Real)) (f i)))
(hproper : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (f i))
(hk :
∀ (i : Fin m) (y : Fin n → Real),
k y = sSup { r : EReal | ∃ x : Fin n → Real, r = f i (x + y) - f i x }) :
∀ i,
Set.recessionCone (epigraph (Set.univ : Set (Fin n → Real)) (f i)) =
epigraph (Set.univ : Set (Fin n → Real)) k
Closedness and recession cone of the convex hull of a finite epigraph family in product space.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma convexHull_union_epigraph_closed_recession_prod {n m : Nat}
{f : Fin m → (Fin n → Real) → EReal} {k : (Fin n → Real) → EReal}
(hclosed : ∀ i, ClosedConvexFunction (f i))
(hproper : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (f i))
(hk :
∀ (i : Fin m) (y : Fin n → Real),
k y = sSup { r : EReal | ∃ x : Fin n → Real, r = f i (x + y) - f i x })
(hm : 0 < m) :
let C0 : Set ((Fin n → Real) × Real) :=
convexHull ℝ (⋃ i, epigraph (Set.univ : Set (Fin n → Real)) (f i))
IsClosed C0 ∧
Set.recessionCone C0 = epigraph (Set.univ : Set (Fin n → Real)) k§9 — Some Closedness Criteria — Part 16
Lean authors: Zichen Wang, Wanli Ma, Qiming Dai, Zaiwen Wen.
The common recession function vanishes at the origin.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma k_zero_eq_zero {n m : Nat}
{f : Fin m → (Fin n → Real) → EReal} {k : (Fin n → Real) → EReal}
(hproper : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (f i))
(hk :
∀ (i : Fin m) (y : Fin n → Real),
k y = sSup { r : EReal | ∃ x : Fin n → Real, r = f i (x + y) - f i x })
(hm : 0 < m) :
k 0 = (0 : EReal)
The vertical direction (0,1) lies in the recession cone of the convex hull.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma zero_one_mem_recessionCone_convexHull_union_epigraph {n m : Nat}
{f : Fin m → (Fin n → Real) → EReal} {k : (Fin n → Real) → EReal}
(hclosed : ∀ i, ClosedConvexFunction (f i))
(hproper : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (f i))
(hk :
∀ (i : Fin m) (y : Fin n → Real),
k y = sSup { r : EReal | ∃ x : Fin n → Real, r = f i (x + y) - f i x })
(hm : 0 < m) :
let C0 : Set ((Fin n → Real) × Real) :=
convexHull ℝ (⋃ i, epigraph (Set.univ : Set (Fin n → Real)) (f i))
(0, (1 : Real)) ∈ Set.recessionCone C0
The convex-hull epigraph is an actual epigraph for the convex-hull function family.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma epigraph_convexHullFunctionFamily_eq_convexHull {n m : Nat}
{f : Fin m → (Fin n → Real) → EReal}
(hsubset :
convexHull ℝ (⋃ i, epigraph (Set.univ : Set (Fin n → Real)) (f i)) ⊆
epigraph (Set.univ : Set (Fin n → Real)) (convexHullFunctionFamily f))
(hclosed :
IsClosed (convexHull ℝ (⋃ i, epigraph (Set.univ : Set (Fin n → Real)) (f i))))
(hrec :
(0, (1 : Real)) ∈
Set.recessionCone
(convexHull ℝ (⋃ i, epigraph (Set.univ : Set (Fin n → Real)) (f i)))) :
epigraph (Set.univ : Set (Fin n → Real)) (convexHullFunctionFamily f) =
convexHull ℝ (⋃ i, epigraph (Set.univ : Set (Fin n → Real)) (f i))
The convex-hull function family is never ⊥.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
lemma convexHullFunctionFamily_ne_bot {n m : Nat}
{f : Fin m → (Fin n → Real) → EReal} {k : (Fin n → Real) → EReal}
(hclosed : ∀ i, ClosedConvexFunction (f i))
(hproper : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (f i))
(hk :
∀ (i : Fin m) (y : Fin n → Real),
k y = sSup { r : EReal | ∃ x : Fin n → Real, r = f i (x + y) - f i x })
(hm : 0 < m) :
let fconv : (Fin n → Real) → EReal := convexHullFunctionFamily f
∀ x : Fin n → Real, fconv x ≠ (⊥ : EReal)
Corollary 9.8.3 9.8.3.1. Let f₁, …, fₘ be closed proper convex functions on ℝ^n all having the same recession function k. Then f = conv {f₁, …, fₘ} is closed and proper and likewise has k as its recession function. In the formula for f(x) in Theorem 5.6, the infimum is attained for each x by some convex combination.
open scoped Pointwise
open scoped BigOperators
open scoped Topology
theorem convexHullFunctionFamily_closed_proper_recession_and_attained {n m : Nat}
{f : Fin m → (Fin n → Real) → EReal} {k : (Fin n → Real) → EReal}
(hclosed : ∀ i, ClosedConvexFunction (f i))
(hproper : ∀ i, ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) (f i))
(hk :
∀ (i : Fin m) (y : Fin n → Real),
k y = sSup { r : EReal | ∃ x : Fin n → Real, r = f i (x + y) - f i x })
(hm : 0 < m) :
let fconv : (Fin n → Real) → EReal := convexHullFunctionFamily f
ClosedConvexFunction fconv ∧
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) fconv ∧
(∀ y : Fin n → Real,
k y =
sSup { r : EReal | ∃ x : Fin n → Real, r = fconv (x + y) - fconv x }) ∧
(∀ x : Fin n → Real,
∃ (lam : Fin m → Real) (x' : Fin m → (Fin n → Real)),
(∀ i, 0 ≤ lam i) ∧
(Finset.univ.sum (fun i => lam i) = 1) ∧
(Finset.univ.sum (fun i => lam i • x' i) = x) ∧
fconv x =
Finset.univ.sum (fun i => ((lam i : Real) : EReal) * f i (x' i)))§10 — Continuity of Convex Functions — Part 1
Lean authors: Zichen Wang, Wanli Ma, Guangxuan Pan, Siyuan Shao, Zaiwen Wen.
Definition 10.0.1. Let f : ℝ^n → ℝ be a function and let S ⊆ ℝ^n. The function f is continuous relative to S if the restriction of f to S (denoted f|_S) is continuous on S.
open scoped BigOperators
def Function.ContinuousRelativeTo {n : Nat} (f : EuclideanSpace Real (Fin n) → Real)
(S : Set (EuclideanSpace Real (Fin n))) : Prop :=
ContinuousOn f S
If a sequence in a subset S converges in the ambient space, then it also converges in the topology nhdsWithin _ S.
open scoped BigOperators
lemma Function.tendsto_nhdsWithin_of_forall_mem {n : Nat} {S : Set (EuclideanSpace Real (Fin n))}
{x : EuclideanSpace Real (Fin n)} {y : ℕ → EuclideanSpace Real (Fin n)}
(hyS : ∀ k, y k ∈ S) (hy : Filter.Tendsto y Filter.atTop (nhds x)) :
Filter.Tendsto y Filter.atTop (nhdsWithin x S)
From ContinuousRelativeTo f S (i.e. ContinuousOn f S), every sequence in S converging to x ∈ S has f (y k) → f x.
open scoped BigOperators
lemma Function.continuousRelativeTo_imp_seq_tendsto {n : Nat}
{f : EuclideanSpace Real (Fin n) → Real} {S : Set (EuclideanSpace Real (Fin n))} :
Function.ContinuousRelativeTo f S →
∀ x ∈ S,
∀ y : ℕ → EuclideanSpace Real (Fin n),
(∀ k, y k ∈ S) →
Filter.Tendsto y Filter.atTop (nhds x) →
Filter.Tendsto (fun k => f (y k)) Filter.atTop (nhds (f x))
A sequence in a subtype S converging to x yields an ambient sequence converging to x.
open scoped BigOperators
lemma Function.seq_in_subtype_to_seq_in_ambient {n : Nat} {S : Set (EuclideanSpace Real (Fin n))}
{u : ℕ → S} {x : S} (hu : Filter.Tendsto u Filter.atTop (nhds x)) :
Filter.Tendsto (fun k => (u k : EuclideanSpace Real (Fin n))) Filter.atTop (nhds (x : _))
If every ambient sequence in S converging to x ∈ S satisfies f (y k) → f x, then f is continuous relative to S.
open scoped BigOperators
lemma Function.seq_tendsto_imp_continuousRelativeTo_via_restrict {n : Nat}
{f : EuclideanSpace Real (Fin n) → Real} {S : Set (EuclideanSpace Real (Fin n))}
(hseq :
∀ x ∈ S,
∀ y : ℕ → EuclideanSpace Real (Fin n),
(∀ k, y k ∈ S) →
Filter.Tendsto y Filter.atTop (nhds x) →
Filter.Tendsto (fun k => f (y k)) Filter.atTop (nhds (f x))) :
Function.ContinuousRelativeTo f S
Theorem 10.0.2. Let f : ℝ^n → ℝ and S ⊆ ℝ^n. Then f is continuous relative to S if and only if for every x ∈ S and every sequence (y_k) ⊆ S with y_k → x, one has f(y_k) → f(x).
open scoped BigOperators
theorem Function.continuousRelativeTo_iff_seq_tendsto {n : Nat}
(f : EuclideanSpace Real (Fin n) → Real) (S : Set (EuclideanSpace Real (Fin n))) :
Function.ContinuousRelativeTo f S ↔
∀ x ∈ S,
∀ y : ℕ → EuclideanSpace Real (Fin n),
(∀ k, y k ∈ S) →
Filter.Tendsto y Filter.atTop (nhds x) →
Filter.Tendsto (fun k => f (y k)) Filter.atTop (nhds (f x))
Theorem 10.1. A convex function f on ℝ^n is continuous relative to any relatively open convex set C in its effective domain, in particular relative to ri (dom f).
open scoped BigOperators
theorem convexFunctionOn_continuousOn_of_relOpen_effectiveDomain {n : Nat}
{f : (Fin n → Real) → EReal} (hf : ConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
{C : Set (Fin n → Real)} (hCconv : Convex ℝ C)
(hCsub : C ⊆ effectiveDomain (Set.univ : Set (Fin n → Real)) f)
(hCrelOpen : euclideanRelativelyOpen n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹' C)) :
ContinuousOn f C
Theorem 10.1 (in particular). A convex function f on ℝ^n is continuous relative to ri (dom f), where dom f is the effective domain.
open scoped BigOperators
theorem convexFunctionOn_continuousOn_relativeInterior_effectiveDomain {n : Nat}
{f : (Fin n → Real) → EReal} (hf : ConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) :
ContinuousOn (fun x : EuclideanSpace Real (Fin n) => f (x : Fin n → Real))
(euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f))
For a real-valued function, coercion to EReal is finite everywhere, hence its effective domain over univ is univ.
open scoped BigOperators
lemma effectiveDomain_univ_coe_real {n : Nat} (f : (Fin n → Real) → Real) :
effectiveDomain (Set.univ : Set (Fin n → Real)) (fun x => (f x : EReal)) = Set.univ
The whole Euclidean space is relatively open in itself (ri univ = univ).
open scoped BigOperators
lemma euclideanRelativelyOpen_univ (n : Nat) :
euclideanRelativelyOpen n (Set.univ : Set (EuclideanSpace Real (Fin n)))
A real-valued map is continuous iff its coercion to EReal is continuous.
open scoped BigOperators
lemma continuous_coe_ereal_iff_continuous {α : Type*} [TopologicalSpace α] {f : α → Real} :
(Continuous fun a => (f a : EReal)) ↔ Continuous f
Corollary 10.1.1. A convex function finite on all of ℝ^n is necessarily continuous.
open scoped BigOperators
theorem convexFunctionOn_continuous_of_real {n : Nat} {f : (Fin n → Real) → Real}
(hf : ConvexFunctionOn (Set.univ : Set (Fin n → Real)) (fun x => (f x : EReal))) :
Continuous f
Any element of a bounded-above range is bounded above by its sSup.
open scoped BigOperators
lemma le_csSup_range_apply {n : Nat} {T : Type*} [Nonempty T]
(f : (Fin n → Real) → T → Real) (hbdd : ∀ x, BddAbove (Set.range fun t : T => f x t))
(x : Fin n → Real) (t : T) :
f x t ≤ sSup (Set.range fun t : T => f x t)
The pointwise supremum of a family of convex functions is convex.
open scoped BigOperators
lemma convexOn_sSup_range_of_convexOn {n : Nat} {T : Type*} [Nonempty T]
(f : (Fin n → Real) → T → Real)
(hconv : ∀ t, ConvexOn ℝ (Set.univ : Set (Fin n → Real)) (fun x => f x t))
(hbdd : ∀ x, BddAbove (Set.range fun t : T => f x t)) :
ConvexOn ℝ (Set.univ : Set (Fin n → Real))
(fun x => sSup (Set.range fun t : T => f x t))
A convex real-valued function on all of ℝ^n is continuous.
open scoped BigOperators
lemma continuous_of_convexOn_univ {n : Nat} {h : (Fin n → Real) → Real}
(hh : ConvexOn ℝ (Set.univ : Set (Fin n → Real)) h) : Continuous h
Theorem 10.1.2. Let T be an arbitrary index set and let f : ℝ^n × T → ℝ be a function such that: - for each t ∈ T, the function x ↦ f(x,t) is convex on ℝ^n; - for each x ∈ ℝ^n, the function t ↦ f(x,t) is bounded above on T. Define h(x) := sup { f(x,t) | t ∈ T }. Then h is a finite convex function on ℝ^n, and h(x) depends continuously on x.
open scoped BigOperators
theorem convexOn_continuous_sSup_range {n : Nat} {T : Type*} [Nonempty T]
(f : (Fin n → Real) → T → Real)
(hconv : ∀ t, ConvexOn ℝ (Set.univ : Set (Fin n → Real)) (fun x => f x t))
(hbdd : ∀ x, BddAbove (Set.range fun t : T => f x t)) :
(let h : (Fin n → Real) → Real := fun x => sSup (Set.range fun t : T => f x t)
ConvexOn ℝ (Set.univ : Set (Fin n → Real)) h ∧ Continuous h)
The inf-set in Theorem 10.1.3 is nonempty.
open scoped BigOperators
lemma sInf_image_add_nonempty {n : Nat} {f : (Fin n → Real) → Real} {C : Set (Fin n → Real)}
(hCne : C.Nonempty) (x : Fin n → Real) :
(f '' ((fun y => y + x) '' C)).Nonempty
For ε > 0, there is a point in C whose translate is ε-close above the infimum.
open scoped BigOperators
lemma exists_mem_C_lt_sInf_add {n : Nat} {f : (Fin n → Real) → Real} {C : Set (Fin n → Real)}
(hCne : C.Nonempty) (x : Fin n → Real) {ε : Real} (hε : 0 < ε) :
∃ y ∈ C, f (y + x) < sInf (f '' ((fun y => y + x) '' C)) + ε
If all the inf-sets are bounded below, the pointwise infimum of translates of a convex function is convex.
open scoped BigOperators
lemma convexOn_sInf_image_add_of_bddBelow {n : Nat} {f : (Fin n → Real) → Real}
(hf : ConvexOn ℝ (Set.univ : Set (Fin n → Real)) f) {C : Set (Fin n → Real)}
(hCne : C.Nonempty) (hCconv : Convex ℝ C)
(hbdd : ∀ x, BddBelow (f '' ((fun y => y + x) '' C))) :
ConvexOn ℝ (Set.univ : Set (Fin n → Real))
(fun x => sInf (f '' ((fun y => y + x) '' C)))
Theorem 10.1.3. Let f : ℝ^n → ℝ be a convex function finite on all of ℝ^n, and let C ⊆ ℝ^n be a nonempty convex set. For each x ∈ ℝ^n, define h(x) := inf {f(y) | y ∈ C + x}. Then h is convex on ℝ^n, and h depends continuously on x.
open scoped BigOperators
theorem convexOn_continuous_sInf_image_add {n : Nat} {f : (Fin n → Real) → Real}
(hf : ConvexOn ℝ (Set.univ : Set (Fin n → Real)) f) {C : Set (Fin n → Real)}
(hCne : C.Nonempty) (hCconv : Convex ℝ C)
(hbdd : ∀ x, BddBelow (f '' ((fun y => y + x) '' C))) :
(let h : (Fin n → Real) → Real := fun x => sInf (f '' ((fun y => y + x) '' C))
ConvexOn ℝ (Set.univ : Set (Fin n → Real)) h ∧ Continuous h)
Theorem 10.1.4. Define f : ℝ^2 → (-∞, +∞] by - f(ξ₁, ξ₂) = ξ₂^2 / (2 ξ₁) if ξ₁ > 0, - f(0, 0) = 0, - f(ξ₁, ξ₂) = +∞ otherwise. Then f is convex with effective domain {(ξ₁, ξ₂) | ξ₁ > 0} ∪ {(0,0)}. Moreover, f is the support function of the convex set C = {(η₁, η₂) | η₁ + η₂^2 / 2 ≤ 0} (i.e. the supremum of ⟪ξ, η⟫ over η ∈ C). The function is continuous at every point of its effective domain except at (0,0), where it is only lower semicontinuous. In particular, for any α > 0, lim_{t → 0} f(t^2/(2α), t) = α, while lim_{t ↓ 0} f(t • x) = 0 for every x with x₁ > 0.
open scoped BigOperators
noncomputable def quadraticOverLinearEReal (ξ : Fin 2 → Real) : EReal :=
if ξ 0 > 0 then ((ξ 1) ^ 2 / (2 * ξ 0) : Real)
else if ξ 0 = 0 ∧ ξ 1 = 0 then (0 : EReal)
else (⊤ : EReal)
The convex set C = {(η₁, η₂) | η₁ + η₂^2/2 ≤ 0} appearing in Theorem 10.1.4.
open scoped BigOperators
def quadraticOverLinearSupportSet : Set (Fin 2 → Real) :=
{η | η 0 + η 1 ^ 2 / 2 ≤ 0}
Expand dotProduct on Fin 2 as a two-term sum.
open scoped BigOperators
lemma dotProduct_fin2 (ξ η : Fin 2 → Real) :
dotProduct ξ η = ξ 0 * η 0 + ξ 1 * η 1
The set of dot products ⟪ξ,η⟫ with η ∈ quadraticOverLinearSupportSet, seen in EReal.
open scoped BigOperators
def quadraticOverLinearSupportValues (ξ : Fin 2 → Real) : Set EReal :=
{z : EReal | ∃ η ∈ quadraticOverLinearSupportSet, z = ((dotProduct ξ η : Real) : EReal)}
The quadratic-over-linear function is nonnegative everywhere (as an EReal-valued map).
open scoped BigOperators
lemma zero_le_quadraticOverLinearEReal (ξ : Fin 2 → Real) :
(0 : EReal) ≤ quadraticOverLinearEReal ξ
A completing-the-square bound for a concave quadratic.
open scoped BigOperators
lemma concave_quadratic_bound {a b t : Real} (ha : 0 < a) :
(- (a / 2) * t ^ 2 + b * t) ≤ b ^ 2 / (2 * a)
Upper bound on ⟪ξ,η⟫ when η ∈ quadraticOverLinearSupportSet and ξ₁ > 0.
open scoped BigOperators
lemma quadraticOverLinear_supportSet_dotProduct_le_quadraticBound {ξ η : Fin 2 → Real}
(hξ : 0 < ξ 0) (hη : η ∈ quadraticOverLinearSupportSet) :
(dotProduct ξ η : Real) ≤ (ξ 1) ^ 2 / (2 * ξ 0)
For ξ₁ > 0, there is η ∈ quadraticOverLinearSupportSet attaining the quadratic bound.
open scoped BigOperators
lemma exists_supportSet_attains_quadraticBound {ξ : Fin 2 → Real} (hξ : 0 < ξ 0) :
∃ η ∈ quadraticOverLinearSupportSet,
(dotProduct ξ η : Real) = (ξ 1) ^ 2 / (2 * ξ 0)
ξ = 0 in Fin 2 → ℝ iff both coordinates are zero.
open scoped BigOperators
lemma fin2_eq_zero_iff (ξ : Fin 2 → Real) : ξ = 0 ↔ ξ 0 = 0 ∧ ξ 1 = 0
If b < ⊤ in EReal, it is below some real coercion.
open scoped BigOperators
lemma exists_lt_coe_of_lt_top {b : EReal} (hb : b < (⊤ : EReal)) : ∃ r : Real, b < (r : EReal)
If ξ₁ > 0, the support supremum equals the quadratic bound.
open scoped BigOperators
lemma sSup_quadraticOverLinearSupportValues_of_pos {ξ : Fin 2 → Real} (hξ : 0 < ξ 0) :
sSup (quadraticOverLinearSupportValues ξ) = (( (ξ 1) ^ 2 / (2 * ξ 0) : Real) : EReal)
For ξ = 0, the support supremum equals 0.
open scoped BigOperators
lemma sSup_quadraticOverLinearSupportValues_zero :
sSup (quadraticOverLinearSupportValues (0 : Fin 2 → Real)) = (0 : EReal)
If ξ₁ < 0, the support supremum is ⊤ (unbounded above).
open scoped BigOperators
lemma sSup_quadraticOverLinearSupportValues_eq_top_of_neg {ξ : Fin 2 → Real} (hξ : ξ 0 < 0) :
sSup (quadraticOverLinearSupportValues ξ) = (⊤ : EReal)
If ξ₁ = 0 and ξ₂ ≠ 0, the support supremum is ⊤ (unbounded above).
open scoped BigOperators
lemma sSup_quadraticOverLinearSupportValues_eq_top_of_zero_ne {ξ : Fin 2 → Real}
(hξ0 : ξ 0 = 0) (hξ1 : ξ 1 ≠ 0) :
sSup (quadraticOverLinearSupportValues ξ) = (⊤ : EReal)
Compute the support supremum in closed form.
open scoped BigOperators
lemma sSup_quadraticOverLinearSupportValues (ξ : Fin 2 → Real) :
sSup (quadraticOverLinearSupportValues ξ) =
if ξ 0 > 0 then (( (ξ 1) ^ 2 / (2 * ξ 0) : Real) : EReal)
else if ξ 0 = 0 ∧ ξ 1 = 0 then (0 : EReal)
else (⊤ : EReal)
Support function representation for quadraticOverLinearEReal (as a pointwise equality).
open scoped BigOperators
theorem quadraticOverLinearEReal_eq_supportFunction_aux :
quadraticOverLinearEReal =
fun ξ => sSup (quadraticOverLinearSupportValues ξ)
For fixed η, the map ξ ↦ ⟪ξ,η⟫ (coerced to EReal) is convex on ℝ^2.
open scoped BigOperators
lemma convexFunctionOn_ereal_dotProduct_fixed (η : Fin 2 → Real) :
ConvexFunctionOn (Set.univ : Set (Fin 2 → Real))
(fun ξ => ((dotProduct ξ η : Real) : EReal))
The EReal support function ξ ↦ sSup {⟪ξ,η⟫ | η ∈ C} is convex.
open scoped BigOperators
lemma convexFunctionOn_supportFunctionEReal (C : Set (Fin 2 → Real)) :
ConvexFunctionOn (Set.univ : Set (Fin 2 → Real))
(fun ξ =>
sSup {z : EReal | ∃ η ∈ C, z = ((dotProduct ξ η : Real) : EReal)})
Theorem 10.1.4 (convexity): the quadratic-over-linear extended-valued function is convex.
open scoped BigOperators
theorem convexFunctionOn_quadraticOverLinearEReal :
ConvexFunctionOn (Set.univ : Set (Fin 2 → Real)) quadraticOverLinearEReal
Theorem 10.1.4 (effective domain): dom f = {ξ | ξ₁ > 0} ∪ {(0,0)}.
open scoped BigOperators
theorem effectiveDomain_quadraticOverLinearEReal :
effectiveDomain (Set.univ : Set (Fin 2 → Real)) quadraticOverLinearEReal =
{ξ : Fin 2 → Real | ξ 0 > 0} ∪ {0}
Theorem 10.1.4 (convex set): C = {(η₁, η₂) | η₁ + η₂^2/2 ≤ 0} is convex.
open scoped BigOperators
theorem convex_quadraticOverLinearSupportSet :
Convex ℝ quadraticOverLinearSupportSet
Theorem 10.1.4 (support function): f is the support function of C.
open scoped BigOperators
theorem quadraticOverLinearEReal_eq_supportFunction :
quadraticOverLinearEReal =
fun ξ =>
sSup
{z : EReal |
∃ η ∈ quadraticOverLinearSupportSet, z = ((dotProduct ξ η : Real) : EReal)}
Theorem 10.1.4 (continuity away from (0,0)): f is continuous at every nonzero point of its effective domain.
open scoped BigOperators
theorem continuousAt_quadraticOverLinearEReal_of_mem_effectiveDomain {x : Fin 2 → Real}
(hx : x ∈ effectiveDomain (Set.univ : Set (Fin 2 → Real)) quadraticOverLinearEReal)
(hx0 : x ≠ 0) :
ContinuousAt quadraticOverLinearEReal x§10 — Continuity of Convex Functions — Part 2
Lean authors: Zichen Wang, Wanli Ma, Guangxuan Pan, Siyuan Shao, Zaiwen Wen.
Theorem 10.1.4 (lower semicontinuity at (0,0)).
open scoped BigOperators
theorem lowerSemicontinuousAt_quadraticOverLinearEReal_zero :
LowerSemicontinuousAt quadraticOverLinearEReal (0 : Fin 2 → Real)
Theorem 10.1.4 (path limit, auxiliary): for α > 0, along the parabola and excluding t = 0, the values tend to α.
open scoped BigOperators
lemma tendsto_quadraticOverLinearEReal_along_parabola_aux {α : Real} (hα : 0 < α) :
Filter.Tendsto
(fun t : Real =>
quadraticOverLinearEReal
(![t ^ 2 / (2 * α), t] : Fin 2 → Real))
(nhdsWithin (0 : Real) ({0}ᶜ)) (nhds (α : EReal))
Theorem 10.1.4 (failure of continuity at (0,0)).
open scoped BigOperators
theorem not_continuousAt_quadraticOverLinearEReal_zero :
¬ ContinuousAt quadraticOverLinearEReal (0 : Fin 2 → Real)
Theorem 10.1.4 (path limit): for any α > 0, lim_{t → 0} f(t^2/(2α), t) = α.
open scoped BigOperators
theorem tendsto_quadraticOverLinearEReal_along_parabola {α : Real} (hα : 0 < α) :
Filter.Tendsto
(fun t : Real =>
quadraticOverLinearEReal
(![t ^ 2 / (2 * α), t] : Fin 2 → Real))
(nhdsWithin (0 : Real) ({0}ᶜ)) (nhds (α : EReal))
Theorem 10.1.4 (path limit): for x₁ > 0, lim_{t ↓ 0} f(t • x) = 0.
open scoped BigOperators
theorem tendsto_quadraticOverLinearEReal_smul_of_pos {x : Fin 2 → Real} (hx : x 0 > 0) :
Filter.Tendsto (fun t : Real => quadraticOverLinearEReal (t • x))
(nhdsWithin (0 : Real) (Set.Ioi 0)) (nhds (0 : EReal))
Definition 10.1.5. A subset S ⊆ ℝ^n is locally simplicial if for each x ∈ S there exist finitely many simplices S₁, …, Sₘ ⊆ S and a neighborhood U of x such that U ∩ (S₁ ∪ ⋯ ∪ Sₘ) = U ∩ S.
open scoped BigOperators
def Set.LocallySimplicial (n : ℕ) (S : Set (Fin n → Real)) : Prop :=
∀ x ∈ S,
∃ (𝒮 : Set (Set (Fin n → Real))) (U : Set (Fin n → Real)),
𝒮.Finite ∧
U ∈ nhds x ∧
(∀ P ∈ 𝒮, ∃ m : ℕ, IsSimplex n m P ∧ P ⊆ S) ∧
U ∩ (⋃₀ 𝒮) = U ∩ S
Replacing b j by x produces exactly the vertex set {x} ∪ {b (succAbove j i) | i : Fin m}.
open scoped BigOperators
lemma range_update_eq_insert_range_succAbove {n m : ℕ} (b : Fin (m + 1) → Fin n → Real)
(j : Fin (m + 1)) (x : Fin n → Real) :
Set.range (Function.update b j x) =
Set.insert x (Set.range fun i : Fin m => b (Fin.succAbove j i))
Given two barycentric weight functions μ and ν with μ ≥ 0 and ∑ μ = 1, choose an index j with μ j > 0 minimizing ν i / μ i among μ i > 0. The resulting j satisfies the cross-multiplied inequalities ν j * μ i ≤ ν i * μ j.
open scoped BigOperators
lemma exists_index_min_ratio_barycentric {m : ℕ} {μ ν : Fin (m + 1) → Real}
(hμ0 : ∀ i, 0 ≤ μ i) (hμ1 : (∑ i, μ i) = 1) (hν0 : ∀ i, 0 ≤ ν i) :
∃ j : Fin (m + 1), 0 < μ j ∧ ∀ i, ν j * μ i ≤ ν i * μ j
If x is an affine combination of an affinely independent family b with a positive weight at j, then x does not lie in the affine span of the other vertices.
open scoped BigOperators
lemma not_mem_affineSpan_facet_of_barycentric_weight_pos {n m : ℕ}
{b : Fin (m + 1) → Fin n → Real} (hb : AffineIndependent Real b) {x : Fin n → Real}
{μ : Fin (m + 1) → Real} (hμ1 : (∑ i, μ i) = 1) (hx : x = ∑ i, μ i • b i)
{j : Fin (m + 1)} (hμj : 0 < μ j) :
x ∉ affineSpan Real (b '' { i | i ≠ j })
If x and y have barycentric coordinates μ and ν in a simplex, and j is chosen by the min-ratio condition, then y lies in the convex hull of x and the facet opposite j.
open scoped BigOperators
lemma mem_cone_facet_of_min_ratio {n m : ℕ} {b : Fin (m + 1) → Fin n → Real} {x y : Fin n → Real}
{μ ν : Fin (m + 1) → Real} (hμ1 : (∑ i, μ i) = 1) (hν0 : ∀ i, 0 ≤ ν i)
(hν1 : (∑ i, ν i) = 1) (hx : x = ∑ i, μ i • b i) (hy : y = ∑ i, ν i • b i)
{j : Fin (m + 1)} (hμj : 0 < μ j) (hcross : ∀ i, ν j * μ i ≤ ν i * μ j) :
y ∈
convexHull Real
(Set.insert x (Set.range fun i : Fin m => b (Fin.succAbove j i)))
The convex hull of x and a facet of an m-simplex is again an m-simplex, provided x is not in the affine span of that facet.
open scoped BigOperators
lemma isSimplex_cone_over_facet {n m : ℕ} {b : Fin (m + 1) → Fin n → Real} (hb : AffineIndependent Real b)
{x : Fin n → Real} {μ : Fin (m + 1) → Real} (hμ1 : (∑ i, μ i) = 1)
(hx : x = ∑ i, μ i • b i) {j : Fin (m + 1)} (hμj : 0 < μ j) :
let P : Set (Fin n → Real) :=
convexHull Real (Set.insert x (Set.range fun i : Fin m => b (Fin.succAbove j i)))
IsSimplex n m P
Theorem 10.1.6. Let C ⊆ ℝ^n be a simplex with vertices x₀, x₁, …, x_m, and let x ∈ C. Then C can be triangulated into finitely many simplices having x as a vertex. More precisely, for every y ∈ C there exists a simplex P ⊆ C whose vertices consist of x together with m of the m+1 vertices of C, and such that y ∈ P.
open scoped BigOperators
theorem simplex_exists_subsimplex_through_point {n m : ℕ} {C : Set (Fin n → Real)}
{x : Fin n → Real} (hC : IsSimplex n m C) (hx : x ∈ C) :
∃ b : Fin (m + 1) → Fin n → Real,
AffineIndependent Real b ∧
C = convexHull Real (Set.range b) ∧
∀ y ∈ C,
∃ j : Fin (m + 1),
let P : Set (Fin n → Real) :=
convexHull Real
(Set.insert x (Set.range fun i : Fin m => b (Fin.succAbove j i)))
P ⊆ C ∧ IsSimplex n m P ∧ x ∈ P ∧ y ∈ P
Upper semicontinuity within a finite union (binary case).
open scoped BigOperators
lemma upperSemicontinuousWithinAt_union {α : Type*} [TopologicalSpace α] {β : Type*} [Preorder β]
{f : α → β} {s t : Set α} {x : α} :
UpperSemicontinuousWithinAt f s x → UpperSemicontinuousWithinAt f t x →
UpperSemicontinuousWithinAt f (s ∪ t) x
Upper semicontinuity within a sUnion of finitely many sets.
open scoped BigOperators
lemma upperSemicontinuousWithinAt_sUnion_of_finite {α : Type*} [TopologicalSpace α] {β : Type*}
[Preorder β] {f : α → β} {x : α} {𝒮 : Set (Set α)} (h𝒮 : 𝒮.Finite)
(husc : ∀ s ∈ 𝒮, UpperSemicontinuousWithinAt f s x) :
UpperSemicontinuousWithinAt f (⋃₀ 𝒮) x
Upper semicontinuity withinAt is invariant under local equality of the underlying set.
open scoped BigOperators
lemma upperSemicontinuousWithinAt_congr_of_local_eq {α : Type*} [TopologicalSpace α] {β : Type*}
[Preorder β] {f : α → β} {s t U : Set α} {x : α} (hU : U ∈ nhds x)
(hEq : U ∩ s = U ∩ t) :
UpperSemicontinuousWithinAt f s x ↔ UpperSemicontinuousWithinAt f t x
A finite affine-combination inequality from epigraph convexity: if f (x i) ≤ μ i for i ∈ s, then f at the affine combination is bounded above by the real affine combination of the μ i.
open scoped BigOperators
lemma Section10.convexFunctionOn_le_affineCombination_real {n : ℕ} {ι : Type*}
{f : (Fin n → Real) → EReal} (hf : ConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
(s : Finset ι) (x : ι → Fin n → Real) (μ w : ι → Real)
(hμ : ∀ i ∈ s, f (x i) ≤ (μ i : EReal)) (hw0 : ∀ i ∈ s, 0 ≤ w i) (hw1 : s.sum w = 1) :
f (s.affineCombination ℝ x w) ≤ ((∑ i ∈ s, w i * μ i : Real) : EReal)
A simplex contained in the effective domain admits a uniform *real* upper bound.
open scoped BigOperators
lemma Section10.simplex_real_upper_bound_of_dom {n m : ℕ} {f : (Fin n → Real) → EReal}
(hf : ConvexFunctionOn (Set.univ : Set (Fin n → Real)) f) {P : Set (Fin n → Real)}
(hP : IsSimplex n m P) (hPdom : P ⊆ effectiveDomain (Set.univ : Set (Fin n → Real)) f) :
∃ M : Real, ∀ z ∈ P, f z ≤ (M : EReal)
Given an affinely independent family of vertices, build an AffineBasis for their affine span.
open scoped BigOperators
noncomputable def Section10.simplex_affineBasis_on_affineSpan {n m : ℕ}
{b : Fin (m + 1) → Fin n → Real} (hb : AffineIndependent Real b) :
AffineBasis (Fin (m + 1)) Real (affineSpan Real (Set.range b))
In the relative topology of a simplex, the barycentric coordinate at a vertex tends to 1.
open scoped BigOperators
lemma Section10.vertex_coord_eventually_gt {n m : ℕ} {b : Fin (m + 1) → Fin n → Real}
(hb : AffineIndependent Real b) (j : Fin (m + 1)) {δ : Real} (hδ : 0 < δ) :
∀ᶠ z : {z // z ∈ convexHull Real (Set.range b)} in
nhds
(⟨b j,
(subset_convexHull (𝕜 := Real) (s := Set.range b) ⟨j, rfl⟩)⟩ :
{z // z ∈ convexHull Real (Set.range b)}),
(1 - δ) <
(Section10.simplex_affineBasis_on_affineSpan (b := b) hb).coord j
⟨z.1, convexHull_subset_affineSpan (𝕜 := Real) (s := Set.range b) z.2⟩§10 — Continuity of Convex Functions — Part 3
Lean authors: Zichen Wang, Wanli Ma, Guangxuan Pan, Siyuan Shao, Zaiwen Wen.
Upper semicontinuity at a vertex of an m-simplex contained in dom f.
open scoped BigOperators
lemma Section10.upperSemicontinuousWithinAt_of_isSimplex_vertex {n m : ℕ}
{f : (Fin n → Real) → EReal} (hf : ConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
{b : Fin (m + 1) → Fin n → Real} (hb : AffineIndependent Real b)
(hdom : convexHull Real (Set.range b) ⊆ effectiveDomain (Set.univ : Set (Fin n → Real)) f)
(j : Fin (m + 1)) :
UpperSemicontinuousWithinAt f (convexHull Real (Set.range b)) (b j)
Upper semicontinuity within an m-simplex contained in dom f.
open scoped BigOperators
lemma Section10.upperSemicontinuousWithinAt_of_isSimplex {n m : ℕ}
{f : (Fin n → Real) → EReal} (hf : ConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
{P : Set (Fin n → Real)} {x : Fin n → Real} (hP : IsSimplex n m P) (hxP : x ∈ P)
(hPdom : P ⊆ effectiveDomain (Set.univ : Set (Fin n → Real)) f) :
UpperSemicontinuousWithinAt f P x
Theorem 10.2. Let f be a convex function on ℝ^n, and let S be any locally simplicial subset of dom f. Then f is upper semicontinuous relative to S, so that if f is closed, then f is continuous relative to S.
open scoped BigOperators
theorem convexFunctionOn_upperSemicontinuousOn_of_locallySimplicial {n : ℕ}
{f : (Fin n → Real) → EReal} (hf : ConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
{S : Set (Fin n → Real)} (hS : Set.LocallySimplicial n S)
(hSdom : S ⊆ effectiveDomain (Set.univ : Set (Fin n → Real)) f) :
UpperSemicontinuousOn f S ∧ (ClosedConvexFunction f → ContinuousOn f S)
Extending a finite convex function by ⊤ outside a nonempty set yields a proper convex function on Set.univ.
open scoped BigOperators
lemma Section10.properConvexFunctionOn_univ_extendTopOutside_ri {n : ℕ}
{riCE : Set (EuclideanSpace Real (Fin n))} (hri : riCE.Nonempty)
(f : EuclideanSpace Real (Fin n) → Real) (hfconv : ConvexOn ℝ riCE f)
(e : EuclideanSpace Real (Fin n) ≃L[Real] (Fin n → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)) :
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real))
(fun x : Fin n → Real => by
classical
exact if e.symm x ∈ riCE then (f (e.symm x) : EReal) else (⊤ : EReal))
For the ⊤-extension F outside the relative interior riCE, the effective domain of F pulled back to EuclideanSpace is exactly riCE; hence its relative interior is riCE.
open scoped BigOperators
lemma Section10.preimage_effectiveDomain_extendTopOutside_ri {n : ℕ} {C : Set (Fin n → Real)}
(f : EuclideanSpace Real (Fin n) → Real) :
let CE : Set (EuclideanSpace Real (Fin n)) :=
(fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹' C
let riCE : Set (EuclideanSpace Real (Fin n)) := euclideanRelativeInterior n CE
let e : EuclideanSpace Real (Fin n) ≃L[Real] (Fin n → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n)
let F : (Fin n → Real) → EReal :=
fun x : Fin n → Real => by
classical
exact if e.symm x ∈ riCE then (f (e.symm x) : EReal) else (⊤ : EReal)
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) F = riCE) ∧
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) F) = riCE
Boundedness above on bounded subsets of riCE forces the convex closure of the ⊤-extension to be finite at any point in the closure of riCE.
open scoped BigOperators
lemma Section10.convexFunctionClosure_ne_top_of_mem_closure_ri_of_bddAbove {n : ℕ}
{riCE : Set (EuclideanSpace Real (Fin n))} {f : EuclideanSpace Real (Fin n) → Real}
(hf_bddAbove : ∀ s, s ⊆ riCE → Bornology.IsBounded s → BddAbove (f '' s))
(e : EuclideanSpace Real (Fin n) ≃L[Real] (Fin n → Real) :=
EuclideanSpace.equiv (𝕜 := Real) (ι := Fin n))
(y : Fin n → Real) (hy : (e.symm y : EuclideanSpace Real (Fin n)) ∈ closure riCE) :
convexFunctionClosure (fun x : Fin n → Real => by
classical
exact if e.symm x ∈ riCE then (f (e.symm x) : EReal) else (⊤ : EReal)) y ≠ (⊤ : EReal)
If two functions are continuous relative to CE and agree on riCE, then they agree on CE provided riCE ⊆ CE and CE ⊆ closure riCE.
open scoped BigOperators
lemma Section10.unique_extension_on_C_of_continuous_eqOn_ri {n : ℕ}
{CE riCE : Set (EuclideanSpace Real (Fin n))} (hri : riCE ⊆ CE)
(hclosure : CE ⊆ closure riCE) {g g' : EuclideanSpace Real (Fin n) → Real}
(hg : Function.ContinuousRelativeTo g CE) (hg' : Function.ContinuousRelativeTo g' CE)
(hEq : ∀ x ∈ riCE, g x = g' x) :
∀ x ∈ CE, g x = g' x§10 — Continuity of Convex Functions — Part 4
Lean authors: Zichen Wang, Wanli Ma, Guangxuan Pan, Siyuan Shao, Zaiwen Wen.
Theorem 10.3. Let C be a locally simplicial convex set, and let f be a finite convex function on ri C which is bounded above on every bounded subset of ri C. Then f can be extended in one and only one way to a continuous finite convex function on the whole of C.
open scoped BigOperators Pointwise
theorem convexOn_exists_unique_continuousExtension_of_locallySimplicial {n : ℕ}
{C : Set (Fin n → Real)} (hCconv : Convex ℝ C) (hCloc : Set.LocallySimplicial n C)
(f : EuclideanSpace Real (Fin n) → Real)
(hfconv :
ConvexOn ℝ
(euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹' C))
f)
(hf_bddAbove :
∀ s,
s ⊆
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹' C) →
Bornology.IsBounded s → BddAbove (f '' s)) :
∃ g : EuclideanSpace Real (Fin n) → Real,
(ConvexOn ℝ ((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹' C) g ∧
Function.ContinuousRelativeTo g
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹' C) ∧
(∀ x ∈
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹' C),
g x = f x)) ∧
∀ g' : EuclideanSpace Real (Fin n) → Real,
(ConvexOn ℝ ((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹' C) g' ∧
Function.ContinuousRelativeTo g'
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹' C) ∧
(∀ x ∈
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹' C),
g' x = f x)) →
∀ x ∈ ((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹' C),
g' x = g x
Boundedness above on bounded subsets of the positive orthant, from coordinatewise monotonicity.
open scoped BigOperators Pointwise
lemma Section10.bddAbove_image_of_monotoneOn_of_isBounded_posOrthant {n : ℕ}
{f : (Fin n → Real) → Real} (hfmono : MonotoneOn f {x : Fin n → Real | ∀ i, 0 < x i})
{s : Set (Fin n → Real)} (hs : s ⊆ {x : Fin n → Real | ∀ i, 0 < x i})
(hsbdd : Bornology.IsBounded s) :
BddAbove (f '' s)
The relative interior of the nonnegative orthant in EuclideanSpace is the positive orthant.
open scoped BigOperators Pointwise
lemma Section10.euclideanRelativeInterior_preimage_nonnegOrthant {n : ℕ} :
let C : Set (Fin n → Real) := {y : Fin n → Real | ∀ i, 0 ≤ y i}
let CE : Set (EuclideanSpace Real (Fin n)) :=
(fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹' C
euclideanRelativeInterior n CE =
{x : EuclideanSpace Real (Fin n) | ∀ i, 0 < (x : Fin n → Real) i}
The coordinate singleton vectors Pi.single i M are linearly independent when M ≠ 0.
open scoped BigOperators Pointwise
lemma Section10.linearIndependent_piSingle {n : ℕ} {M : Real} (hM : M ≠ 0) :
LinearIndependent Real (fun i : Fin n => (Pi.single i M : Fin n → Real))
The family of vertices {0} ∪ {Pi.single i M} is affinely independent when M ≠ 0.
open scoped BigOperators Pointwise
lemma Section10.affineIndependent_zero_piSingle {n : ℕ} {M : Real} (hM : M ≠ 0) :
let b : Fin (n + 1) → Fin n → Real := Fin.cases 0 (fun i : Fin n => Pi.single i M)
AffineIndependent Real b
The convex hull of {0} ∪ {Pi.single i M} is an n-simplex when M ≠ 0.
open scoped BigOperators Pointwise
lemma Section10.isSimplex_convexHull_zero_piSingle {n : ℕ} {M : Real} (hM : M ≠ 0) :
let b : Fin (n + 1) → Fin n → Real := Fin.cases 0 (fun i : Fin n => Pi.single i M)
IsSimplex n n (convexHull Real (Set.range b))
A nonnegative vector with sum bounded by M lies in the simplex convexHull {0, M e_i}.
open scoped BigOperators Pointwise
lemma Section10.mem_convexHull_zero_piSingle_of_nonneg_sum_le {n : ℕ} {M : Real} (hM : 0 < M)
(y : Fin n → Real) (hy0 : ∀ i, 0 ≤ y i) (hysum : (∑ i : Fin n, y i) ≤ M) :
let b : Fin (n + 1) → Fin n → Real := Fin.cases 0 (fun i : Fin n => Pi.single i M)
y ∈ convexHull Real (Set.range b)
The nonnegative orthant is locally simplicial.
open scoped BigOperators Pointwise
lemma Section10.locallySimplicial_nonnegOrthant {n : ℕ} :
Set.LocallySimplicial n {y : Fin n → Real | ∀ i, 0 ≤ y i}
Theorem 10.3.1. Let C ⊆ ℝ^n be the nonnegative orthant, C = {x = (ξ₁, …, ξₙ) | ξⱼ ≥ 0, j = 1, …, n}. Assume f is a finite convex function on the positive orthant int C and is non-decreasing in each coordinate. Then f is bounded above on every bounded subset of the positive orthant, and hence f admits a unique extension to a finite continuous convex function on the whole nonnegative orthant C.
open scoped BigOperators Pointwise
theorem convexOn_bddAbove_and_exists_unique_extension_nonnegOrthant_of_monotoneOn {n : ℕ}
(f : (Fin n → Real) → Real)
(hfconv : ConvexOn ℝ {x : Fin n → Real | ∀ i, 0 < x i} f)
(hfmono : MonotoneOn f {x : Fin n → Real | ∀ i, 0 < x i}) :
(∀ s,
s ⊆ {x : Fin n → Real | ∀ i, 0 < x i} →
Bornology.IsBounded s → BddAbove (f '' s)) ∧
(∃ g : (Fin n → Real) → Real,
(ConvexOn ℝ {x : Fin n → Real | ∀ i, 0 ≤ x i} g ∧
ContinuousOn g {x : Fin n → Real | ∀ i, 0 ≤ x i} ∧
(∀ x ∈ {x : Fin n → Real | ∀ i, 0 < x i}, g x = f x)) ∧
∀ g' : (Fin n → Real) → Real,
(ConvexOn ℝ {x : Fin n → Real | ∀ i, 0 ≤ x i} g' ∧
ContinuousOn g' {x : Fin n → Real | ∀ i, 0 ≤ x i} ∧
(∀ x ∈ {x : Fin n → Real | ∀ i, 0 < x i}, g' x = f x)) →
∀ x ∈ {x : Fin n → Real | ∀ i, 0 ≤ x i}, g' x = g x)
Definition 10.3.2. Let S ⊆ ℝ^n. A real-valued function f defined on S is called Lipschitzian relative to S if there exists a real number α ≥ 0 such that |f(y) - f(x)| ≤ α |y - x| for all x ∈ S and y ∈ S.
open scoped BigOperators Pointwise
def Function.LipschitzRelativeTo {n : ℕ} (f : EuclideanSpace Real (Fin n) → Real)
(S : Set (EuclideanSpace Real (Fin n))) : Prop :=
∃ K : NNReal, LipschitzOnWith K f S
A function that is Lipschitz on a set is uniformly continuous on that set.
open scoped BigOperators Pointwise
lemma Function.LipschitzRelativeTo.uniformContinuousOn {n : ℕ}
{f : EuclideanSpace Real (Fin n) → Real} {S : Set (EuclideanSpace Real (Fin n))}
(hf : Function.LipschitzRelativeTo f S) :
UniformContinuousOn f S
Theorem 10.3.3. Let S ⊆ ℝ^n. If a real-valued function f is Lipschitzian relative to S, then f is uniformly continuous relative to S.
open scoped BigOperators Pointwise
theorem Function.uniformContinuousOn_of_lipschitzRelativeTo {n : ℕ}
{f : EuclideanSpace Real (Fin n) → Real} {S : Set (EuclideanSpace Real (Fin n))}
(hf : Function.LipschitzRelativeTo f S) :
UniformContinuousOn f S
A compact set included in the relative interior admits a uniform radius whose translated scaled unit ball (intersected with the affine span) stays in the relative interior.
open scoped BigOperators Pointwise
lemma Section10.exists_pos_eps_uniform_ball_subset_ri {n : ℕ}
{C S : Set (EuclideanSpace Real (Fin n))} (hS : IsCompact S)
(hSsubset : S ⊆ euclideanRelativeInterior n C) :
∃ ε : ℝ, 0 < ε ∧
∀ x ∈ S,
((fun u => x + u) '' (ε • euclideanUnitBall n)) ∩
(affineSpan Real C : Set (EuclideanSpace Real (Fin n))) ⊆
euclideanRelativeInterior n C§10 — Continuity of Convex Functions — Part 5
Lean authors: Zichen Wang, Wanli Ma, Guangxuan Pan, Siyuan Shao, Zaiwen Wen.
Theorem 10.4. Let f be a proper convex function, and let S be any closed bounded subset of ri (dom f). Then f is Lipschitzian relative to S.
open scoped BigOperators Pointwise
theorem properConvexFunctionOn_lipschitzRelativeTo_of_isClosed_isBounded_subset_ri_effectiveDomain
{n : ℕ} {f : (Fin n → Real) → EReal}
(hf : ProperConvexFunctionOn (Set.univ : Set (Fin n → Real)) f)
{S : Set (EuclideanSpace Real (Fin n))} (hSclosed : IsClosed S)
(hSbdd : Bornology.IsBounded S)
(hSsubset :
S ⊆
euclideanRelativeInterior n
((fun x : EuclideanSpace Real (Fin n) => (x : Fin n → Real)) ⁻¹'
effectiveDomain (Set.univ : Set (Fin n → Real)) f)) :
Function.LipschitzRelativeTo
(fun x : EuclideanSpace Real (Fin n) => EReal.toReal (f (x : Fin n → Real))) S
WithLp.toLp is a two-sided coercion between Fin n → ℝ and the Euclidean space ℝ^n = EuclideanSpace ℝ (Fin n); converting to a function and back gives the same point.
open scoped BigOperators Pointwise
lemma Section10.toLp_coeFn_eq {n : ℕ} (x : EuclideanSpace Real (Fin n)) :
WithLp.toLp (p := 2) (x : Fin n → Real) = x
A Lipschitz function on a set remains Lipschitz when restricting to a subset.
open scoped BigOperators Pointwise
lemma Function.LipschitzRelativeTo.mono {n : ℕ} {f : EuclideanSpace Real (Fin n) → Real}
{S T : Set (EuclideanSpace Real (Fin n))}
(hT : Function.LipschitzRelativeTo f T) (hST : S ⊆ T) :
Function.LipschitzRelativeTo f S
The relative interior of the whole Euclidean space is the whole space.
open scoped BigOperators Pointwise
lemma Section10.euclideanRelativeInterior_univ (n : ℕ) :
euclideanRelativeInterior n (Set.univ : Set (EuclideanSpace Real (Fin n))) = Set.univ
A finite convex function on ℝ^n yields a proper convex EReal-valued function on univ by coercing to EReal (after viewing Fin n → ℝ as ℝ^n via WithLp.toLp).
open scoped BigOperators Pointwise
lemma Section10.properConvexFunctionOn_univ_coe_comp_toLp_of_convexOn {n : ℕ}
{f : EuclideanSpace Real (Fin n) → Real}
(hf : ConvexOn ℝ (Set.univ : Set (EuclideanSpace Real (Fin n))) f) :
ProperConvexFunctionOn (Set.univ : Set (Fin n → Real))
(fun x : Fin n → Real => (f (WithLp.toLp (p := 2) x) : EReal))
Theorem 10.4.1. Let f : ℝ^n → ℝ be a finite convex function. Then f is uniformly continuous, and even Lipschitzian, relative to every bounded subset of ℝ^n.
open scoped BigOperators Pointwise
theorem convexOn_uniformContinuousOn_and_lipschitzRelativeTo_of_isBounded {n : ℕ}
{f : EuclideanSpace Real (Fin n) → Real}
(hf : ConvexOn ℝ (Set.univ : Set (EuclideanSpace Real (Fin n))) f) :
∀ S : Set (EuclideanSpace Real (Fin n)),
Bornology.IsBounded S → UniformContinuousOn f S ∧ Function.LipschitzRelativeTo f S
The recession function f₀⁺ of a real-valued function f on ℝ^n, defined by f₀⁺(y) = sup_x (f(x + y) - f(x)) (as an extended real).
open scoped BigOperators Pointwise
noncomputable def Function.recessionFunction {n : ℕ} (f : EuclideanSpace Real (Fin n) → Real) :
EuclideanSpace Real (Fin n) → EReal :=
fun y =>
sSup
{r : EReal |
∃ x : EuclideanSpace Real (Fin n), r = ((f (x + y) - f x : Real) : EReal)}
The recession function f₀⁺ is finite everywhere (i.e. takes values in ℝ).
open scoped BigOperators Pointwise
def Function.RecessionFunctionFiniteEverywhere {n : ℕ}
(f : EuclideanSpace Real (Fin n) → Real) : Prop :=
∀ y, ∃ r : Real, Function.recessionFunction f y = (r : EReal)
Any increment f(x+y)-f x is bounded above by the recession function f₀⁺(y).
open scoped BigOperators Pointwise
lemma Section10.recessionFunction_diff_le {n : ℕ}
{f : EuclideanSpace Real (Fin n) → Real} (x y : EuclideanSpace Real (Fin n)) :
((f (x + y) - f x : Real) : EReal) ≤ Function.recessionFunction f y
The recession function f₀⁺ is zero at the origin.
open scoped BigOperators Pointwise
lemma Section10.recessionFunction_zero {n : ℕ}
{f : EuclideanSpace Real (Fin n) → Real} :
Function.recessionFunction f 0 = (0 : EReal)
The recession function never takes the value ⊥.
open scoped BigOperators Pointwise
lemma Section10.recessionFunction_ne_bot {n : ℕ}
(f : EuclideanSpace Real (Fin n) → Real) (y : EuclideanSpace Real (Fin n)) :
Function.recessionFunction f y ≠ (⊥ : EReal)
Uniform continuity of f on univ forces its recession function to be finite everywhere.
open scoped BigOperators Pointwise
lemma Section10.recessionFunctionFiniteEverywhere_of_uniformContinuousOn_univ {n : ℕ}
{f : EuclideanSpace Real (Fin n) → Real}
(hu : UniformContinuousOn f (Set.univ : Set (EuclideanSpace Real (Fin n)))) :
Function.RecessionFunctionFiniteEverywhere f
If the recession function is finite everywhere, a convex function is Lipschitz on univ.
open scoped BigOperators Pointwise
lemma Section10.lipschitzRelativeTo_univ_of_recessionFunctionFiniteEverywhere {n : ℕ}
{f : EuclideanSpace Real (Fin n) → Real}
(hf : ConvexOn ℝ (Set.univ : Set (EuclideanSpace Real (Fin n))) f)
(hfin : Function.RecessionFunctionFiniteEverywhere f) :
Function.LipschitzRelativeTo f (Set.univ : Set (EuclideanSpace Real (Fin n)))
Theorem 10.5. Let f be a finite convex function on ℝ^n. In order that f be uniformly continuous relative to ℝ^n, it is necessary and sufficient that the recession function f₀⁺ of f be finite everywhere. In this event, f is actually Lipschitzian relative to ℝ^n.
open scoped BigOperators Pointwise
theorem convexOn_uniformContinuousOn_univ_iff_recessionFunctionFiniteEverywhere {n : ℕ}
{f : EuclideanSpace Real (Fin n) → Real}
(hf : ConvexOn ℝ (Set.univ : Set (EuclideanSpace Real (Fin n))) f) :
(UniformContinuousOn f (Set.univ : Set (EuclideanSpace Real (Fin n))) ↔
Function.RecessionFunctionFiniteEverywhere f) ∧
(Function.RecessionFunctionFiniteEverywhere f →
Function.LipschitzRelativeTo f (Set.univ : Set (EuclideanSpace Real (Fin n))))
A finite convex function on ℝ^n yields a closed convex EReal-valued function on univ after composing with WithLp.toLp and coercing to EReal.
open scoped BigOperators Pointwise
lemma Section10.closedConvexFunction_coe_comp_toLp_of_convexOn {n : ℕ}
{f : EuclideanSpace Real (Fin n) → Real}
(hf : ConvexOn ℝ (Set.univ : Set (EuclideanSpace Real (Fin n))) f) :
ClosedConvexFunction (fun x : Fin n → Real => (f (WithLp.toLp (p := 2) x) : EReal))
Along a ray, the difference quotient (f(t•y) - f(0))/t converges to the recession function. This is Theorem 8.5 specialized to x = 0 and rewritten in terms of Function.recessionFunction.
open scoped BigOperators Pointwise
lemma Section10.tendsto_div_sub_to_recessionFunction {n : ℕ}
{f : EuclideanSpace Real (Fin n) → Real}
(hf : ConvexOn ℝ (Set.univ : Set (EuclideanSpace Real (Fin n))) f) :
∀ y : EuclideanSpace Real (Fin n),
Filter.Tendsto (fun t : ℝ => (((f (t • y) - f 0) / t : Real) : EReal)) Filter.atTop
(nhds (Function.recessionFunction f y))
If liminf_{t→∞} f(t•y)/t is finite for all y, then the recession function is finite everywhere.
open scoped BigOperators Pointwise
lemma Section10.recessionFunctionFiniteEverywhere_of_liminf_div_lt_top {n : ℕ}
{f : EuclideanSpace Real (Fin n) → Real}
(hf : ConvexOn ℝ (Set.univ : Set (EuclideanSpace Real (Fin n))) f)
(hlim :
∀ y : EuclideanSpace Real (Fin n),
Filter.liminf (fun t : ℝ => ((f (t • y) / t : Real) : EReal)) Filter.atTop < (⊤ : EReal)) :
Function.RecessionFunctionFiniteEverywhere f
Corollary 10.5.1. A finite convex function f is Lipschitzian relative to ℝ^n if liminf_{λ → ∞} f(λ • y) / λ < ∞ for all y.
open scoped BigOperators Pointwise
theorem convexOn_lipschitzRelativeTo_univ_of_liminf_div_lt_top {n : ℕ}
{f : EuclideanSpace Real (Fin n) → Real}
(hf : ConvexOn ℝ (Set.univ : Set (EuclideanSpace Real (Fin n))) f)
(hlim :
∀ y : EuclideanSpace Real (Fin n),
Filter.liminf (fun t : ℝ => ((f (t • y) / t : Real) : EReal)) Filter.atTop <
(⊤ : EReal)) :
Function.LipschitzRelativeTo f (Set.univ : Set (EuclideanSpace Real (Fin n)))
If u ≤ v eventually along a filter, then liminf u ≤ liminf v.
open scoped BigOperators Pointwise
lemma Section10.liminf_le_liminf_of_eventually_le {α : Type*} {l : Filter α}
{u v : α → EReal} (h : ∀ᶠ a in l, u a ≤ v a) :
Filter.liminf u l ≤ Filter.liminf v l§10 — Continuity of Convex Functions — Part 6
Lean authors: Zichen Wang, Wanli Ma, Guangxuan Pan, Siyuan Shao, Zaiwen Wen.
A Lipschitz function on univ has at most linear growth along rays, hence liminf_{t → ∞} g (t • y) / t < ⊤.
open scoped BigOperators Pointwise
lemma Section10.liminf_div_lt_top_of_lipschitzRelativeTo_univ {n : ℕ}
{g : EuclideanSpace Real (Fin n) → Real}
(hg : Function.LipschitzRelativeTo g (Set.univ : Set (EuclideanSpace Real (Fin n)))) :
∀ y : EuclideanSpace Real (Fin n),
Filter.liminf (fun t : ℝ => ((g (t • y) / t : Real) : EReal)) Filter.atTop < (⊤ : EReal)
If f ≤ g and g is Lipschitz on univ, then the growth condition liminf_{t → ∞} f (t • y) / t < ⊤ also holds for f.
open scoped BigOperators Pointwise
lemma Section10.liminf_div_lt_top_of_le_and_lipschitzRelativeTo_univ {n : ℕ}
{f g : EuclideanSpace Real (Fin n) → Real}
(hglip : Function.LipschitzRelativeTo g (Set.univ : Set (EuclideanSpace Real (Fin n))))
(hfg : ∀ x, f x ≤ g x) :
∀ y : EuclideanSpace Real (Fin n),
Filter.liminf (fun t : ℝ => ((f (t • y) / t : Real) : EReal)) Filter.atTop < (⊤ : EReal)
Corollary 10.5.2. Let g be any finite convex function Lipschitzian relative to ℝ^n (for instance, g(x) = α ‖x‖ + β with α > 0). Then every finite convex function f such that f ≤ g is likewise Lipschitzian relative to ℝ^n.
open scoped BigOperators Pointwise
theorem convexOn_lipschitzRelativeTo_univ_of_le_of_lipschitzRelativeTo {n : ℕ}
{f g : EuclideanSpace Real (Fin n) → Real}
(hgconv : ConvexOn ℝ (Set.univ : Set (EuclideanSpace Real (Fin n))) g)
(hglip : Function.LipschitzRelativeTo g (Set.univ : Set (EuclideanSpace Real (Fin n))))
(hfconv : ConvexOn ℝ (Set.univ : Set (EuclideanSpace Real (Fin n))) f)
(hfg : ∀ x, f x ≤ g x) :
Function.LipschitzRelativeTo f (Set.univ : Set (EuclideanSpace Real (Fin n)))
Definition 10.5.3. Let S ⊆ ℝ^n and let {f i | i ∈ I} be a family of real-valued functions defined on S. The family is *equi-Lipschitzian relative to S* if there exists a constant α ≥ 0 such that |f i y - f i x| ≤ α * ‖y - x‖ for all x ∈ S, y ∈ S, and all indices i.
open scoped BigOperators Pointwise
def Function.EquiLipschitzRelativeTo {n : ℕ} {I : Type*}
(f : I → EuclideanSpace Real (Fin n) → Real) (S : Set (EuclideanSpace Real (Fin n))) : Prop :=
∃ K : NNReal, ∀ i : I, LipschitzOnWith K (f i) S§10 — Continuity of Convex Functions — Part 7
Lean authors: Zichen Wang, Wanli Ma, Guangxuan Pan, Siyuan Shao, Zaiwen Wen.
Definition 10.5.4. Let S ⊆ ℝ^n and let {f i | i ∈ I} be a family of real-valued functions defined on S. The family is *uniformly equicontinuous relative to S* if for every ε > 0 there exists δ > 0 such that for all x ∈ S, y ∈ S, and all indices i, if ‖y - x‖ ≤ δ then |f i y - f i x| ≤ ε.
open scoped BigOperators Pointwise
def Function.UniformlyEquicontinuousRelativeTo {n : ℕ} {I : Type*}
(f : I → EuclideanSpace Real (Fin n) → Real) (S : Set (EuclideanSpace Real (Fin n))) : Prop :=
UniformEquicontinuousOn f S
Theorem 10.5.5. Let S ⊆ ℝ^n and let {f i | i ∈ I} be a collection of real-valued functions on S. If the family is equi-Lipschitzian relative to S, then it is uniformly equicontinuous relative to S.
open scoped BigOperators Pointwise
theorem Function.uniformlyEquicontinuousRelativeTo_of_equiLipschitzRelativeTo {n : ℕ} {I : Type*}
{f : I → EuclideanSpace Real (Fin n) → Real} {S : Set (EuclideanSpace Real (Fin n))}
(hf : Function.EquiLipschitzRelativeTo f S) :
Function.UniformlyEquicontinuousRelativeTo f S
Definition 10.5.6. Let S ⊆ ℝ^n and let {f i | i ∈ I} be a collection of real-valued functions on S. The collection {f i | i ∈ I} is *pointwise bounded on S* if for each x ∈ S the set of real numbers {f i x | i ∈ I} is bounded.
open scoped BigOperators Pointwise
def Function.PointwiseBoundedOn {n : ℕ} {I : Type*}
(f : I → EuclideanSpace Real (Fin n) → Real) (S : Set (EuclideanSpace Real (Fin n))) : Prop :=
∀ x ∈ S, Bornology.IsBounded (Set.range fun i : I => f i x)
Definition 10.5.7. Let S ⊆ ℝ^n and let {f i | i ∈ I} be a collection of real-valued functions on S. The collection {f i | i ∈ I} is *uniformly bounded on S* if there exist real numbers α₁ and α₂ such that α₁ ≤ f i x ≤ α₂ for all x ∈ S and all indices i.
open scoped BigOperators Pointwise
def Function.UniformlyBoundedOn {n : ℕ} {I : Type*}
(f : I → EuclideanSpace Real (Fin n) → Real) (S : Set (EuclideanSpace Real (Fin n))) : Prop :=
∃ α₁ α₂ : Real, ∀ i x, x ∈ S → α₁ ≤ f i x ∧ f i x ≤ α₂
A closed and bounded subset of ℝ^n is compact.
open scoped BigOperators Pointwise
lemma Section10.isCompact_of_isClosed_isBounded {n : ℕ}
{S : Set (EuclideanSpace Real (Fin n))} (hSclosed : IsClosed S)
(hSbdd : Bornology.IsBounded S) : IsCompact S
A uniform two-sided bound implies a uniform absolute bound.
open scoped BigOperators Pointwise
lemma Section10.exists_abs_le_of_uniformlyBoundedOn {n : ℕ} {I : Type*}
{f : I → EuclideanSpace Real (Fin n) → Real} {S : Set (EuclideanSpace Real (Fin n))}
(h : Function.UniformlyBoundedOn f S) :
∃ M : Real, ∀ i x, x ∈ S → |f i x| ≤ M
If all functions in a convex family are uniformly bounded by M on a closed thickening of S contained in C, then the family is uniformly bounded on S and equi-Lipschitzian relative to S.
open scoped BigOperators Pointwise
lemma Section10.uniformlyBoundedOn_and_equiLipschitzRelativeTo_of_abs_le_cthickening
{n : ℕ} {I : Type*} {C S : Set (EuclideanSpace Real (Fin n))}
{f : I → EuclideanSpace Real (Fin n) → Real} (hfconv : ∀ i, ConvexOn ℝ C (f i))
{δ M : Real} (hδ : 0 < δ) (hcthick : Metric.cthickening δ S ⊆ C)
(hM : ∀ i x, x ∈ Metric.cthickening δ S → |f i x| ≤ M) :
Function.UniformlyBoundedOn f S ∧ Function.EquiLipschitzRelativeTo f S
For a convex set S in ℝ^n, taking the closure does not create new interior points: interior (closure S) ⊆ S.
open scoped BigOperators Pointwise
lemma Section10.interior_closure_subset_of_convex {n : ℕ}
(S : Set (EuclideanSpace Real (Fin n))) (hS : Convex ℝ S) :
interior (closure S) ⊆ S
Under hypothesis (a) from Theorem 10.6, the family is uniformly bounded above on any compact subset K ⊆ C.
open scoped BigOperators Pointwise
lemma Section10.exists_uniform_upper_bound_on_compact_of_exists_subset
{n : ℕ} {I : Type*} {C : Set (EuclideanSpace Real (Fin n))}
(hCopen : IsOpen C) (hCconv : Convex ℝ C) (f : I → EuclideanSpace Real (Fin n) → Real)
(hfconv : ∀ i, ConvexOn ℝ C (f i))
{C' : Set (EuclideanSpace Real (Fin n))} (hC'sub : C' ⊆ C)
(hC'hull : C ⊆ convexHull ℝ (closure C'))
(hC'bdAbove : ∀ x ∈ C', BddAbove (Set.range fun i : I => f i x))
{K : Set (EuclideanSpace Real (Fin n))} (hKcomp : IsCompact K) (hKsubset : K ⊆ C) :
∃ M : ℝ, ∀ i x, x ∈ K → f i x ≤ M
Under hypothesis (b) from Theorem 10.6 and the uniform upper bound on a compact neighborhood of x₀, the family is uniformly bounded below on any compact subset K ⊆ C.
open scoped BigOperators Pointwise
lemma Section10.exists_uniform_lower_bound_on_compact_of_point_bddBelow
{n : ℕ} {I : Type*} {C : Set (EuclideanSpace Real (Fin n))}
(hCopen : IsOpen C) (f : I → EuclideanSpace Real (Fin n) → Real)
(hfconv : ∀ i, ConvexOn ℝ C (f i))
(hexists_bddBelow : ∃ x₀ ∈ C, BddBelow (Set.range fun i : I => f i x₀))
(hUpper :
∀ {K : Set (EuclideanSpace Real (Fin n))},
IsCompact K → K ⊆ C → ∃ M : ℝ, ∀ i x, x ∈ K → f i x ≤ M)
{K : Set (EuclideanSpace Real (Fin n))} (hKcomp : IsCompact K) (hKsubset : K ⊆ C) :
∃ m : ℝ, ∀ i x, x ∈ K → m ≤ f i x
Under hypotheses (a) and (b) from Theorem 10.6, the family is uniformly bounded on any compact subset K ⊆ C.
open scoped BigOperators Pointwise
lemma Section10.uniformlyBoundedOn_on_compact_of_exists_subset
{n : ℕ} {I : Type*} {C : Set (EuclideanSpace Real (Fin n))}
(hCopen : IsOpen C) (hCconv : Convex ℝ C) (f : I → EuclideanSpace Real (Fin n) → Real)
(hfconv : ∀ i, ConvexOn ℝ C (f i))
{C' : Set (EuclideanSpace Real (Fin n))} (hC'sub : C' ⊆ C)
(hC'hull : C ⊆ convexHull ℝ (closure C'))
(hC'bdAbove : ∀ x ∈ C', BddAbove (Set.range fun i : I => f i x))
(hexists_bddBelow : ∃ x ∈ C, BddBelow (Set.range fun i : I => f i x))
{K : Set (EuclideanSpace Real (Fin n))} (hKcomp : IsCompact K) (hKsubset : K ⊆ C) :
Function.UniformlyBoundedOn f K
Theorem 10.6 (variant, auxiliary proof): reduce uniform boundedness + equi-Lipschitz to uniform boundedness on a compact thickening of S.
open scoped BigOperators Pointwise
lemma Section10.convexOn_family_uniformlyBoundedOn_and_equiLipschitzRelativeTo_of_exists_subset_aux
{n : ℕ} {I : Type*} {C : Set (EuclideanSpace Real (Fin n))}
(hCopen : IsOpen C) (hCconv : Convex ℝ C) (f : I → EuclideanSpace Real (Fin n) → Real)
(hfconv : ∀ i, ConvexOn ℝ C (f i))
{C' : Set (EuclideanSpace Real (Fin n))} (hC'sub : C' ⊆ C)
(hC'hull : C ⊆ convexHull ℝ (closure C'))
(hC'bdAbove : ∀ x ∈ C', BddAbove (Set.range fun i : I => f i x))
(hexists_bddBelow : ∃ x ∈ C, BddBelow (Set.range fun i : I => f i x))
{S : Set (EuclideanSpace Real (Fin n))} (hSclosed : IsClosed S) (hSbdd : Bornology.IsBounded S)
(hSsubset : S ⊆ C) :
Function.UniformlyBoundedOn f S ∧ Function.EquiLipschitzRelativeTo f S
Theorem 10.6. Let C be a relatively open convex set, and let {f i | i ∈ I} be an arbitrary collection of convex functions finite and pointwise bounded on C. Let S be any closed bounded subset of C. Then {f i | i ∈ I} is uniformly bounded on S and equi-Lipschitzian relative to S. The conclusion remains valid if the pointwise boundedness assumption is weakened to the following pair of assumptions: (a) There exists a subset C' of C such that conv (cl C') ⊇ C and sup {f i x | i ∈ I} is finite for every x ∈ C'; (b) There exists at least one x ∈ C such that inf {f i x | i ∈ I} is finite.
open scoped BigOperators Pointwise
theorem convexOn_family_uniformlyBoundedOn_and_equiLipschitzRelativeTo_of_pointwiseBoundedOn
{n : ℕ} {I : Type*} {C : Set (EuclideanSpace Real (Fin n))}
(hCopen : IsOpen C) (hCconv : Convex ℝ C) (f : I → EuclideanSpace Real (Fin n) → Real)
(hfconv : ∀ i, ConvexOn ℝ C (f i)) (hfpt : Function.PointwiseBoundedOn f C)
{S : Set (EuclideanSpace Real (Fin n))} (hSclosed : IsClosed S) (hSbdd : Bornology.IsBounded S)
(hSsubset : S ⊆ C) :
Function.UniformlyBoundedOn f S ∧ Function.EquiLipschitzRelativeTo f S
Theorem 10.6 (variant). The same conclusion under the weaker assumptions (a) and (b) stated in Theorem 10.6.
open scoped BigOperators Pointwise
theorem convexOn_family_uniformlyBoundedOn_and_equiLipschitzRelativeTo_of_exists_subset
{n : ℕ} {I : Type*} {C : Set (EuclideanSpace Real (Fin n))}
(hCopen : IsOpen C) (hCconv : Convex ℝ C) (f : I → EuclideanSpace Real (Fin n) → Real)
(hfconv : ∀ i, ConvexOn ℝ C (f i))
{C' : Set (EuclideanSpace Real (Fin n))} (hC'sub : C' ⊆ C)
(hC'hull : C ⊆ convexHull ℝ (closure C'))
(hC'bdAbove : ∀ x ∈ C', BddAbove (Set.range fun i : I => f i x))
(hexists_bddBelow : ∃ x ∈ C, BddBelow (Set.range fun i : I => f i x))
{S : Set (EuclideanSpace Real (Fin n))} (hSclosed : IsClosed S) (hSbdd : Bornology.IsBounded S)
(hSsubset : S ⊆ C) :
Function.UniformlyBoundedOn f S ∧ Function.EquiLipschitzRelativeTo f S§10 — Continuity of Convex Functions — Part 8
Lean authors: Zichen Wang, Wanli Ma, Guangxuan Pan, Siyuan Shao, Zaiwen Wen.
If C is open and x₀ ∈ C, then some closed ball around x₀ is contained in C.
open scoped BigOperators
lemma Section10.exists_closedBall_subset_of_isOpen {n : ℕ}
{C : Set (EuclideanSpace Real (Fin n))} (hCopen : IsOpen C) {x0 : EuclideanSpace Real (Fin n)}
(hx0 : x0 ∈ C) : ∃ r : ℝ, 0 < r ∧ Metric.closedBall x0 r ⊆ C
In a locally compact space, every point admits a compact neighborhood.
open scoped BigOperators
lemma Section10.exists_compact_mem_nhds_of_locallyCompact {T : Type*} [TopologicalSpace T]
[LocallyCompactSpace T] (t0 : T) : ∃ K : Set T, IsCompact K ∧ K ∈ nhds t0
A continuous real-valued function is bounded on a compact set (rephrased over a subtype).
open scoped BigOperators
lemma Section10.isBounded_range_subtype_of_continuous_compact {T : Type*} [TopologicalSpace T]
{K : Set T} (hK : IsCompact K) {g : T → ℝ} (hg : Continuous g) :
Bornology.IsBounded (Set.range fun t : {t // t ∈ K} => g t.1)
Theorem 10.7. Let C be a relatively open convex set in ℝ^n, and let T be a locally compact topological space. Let f : ℝ^n × T → ℝ be such that: - for each t ∈ T, the function x ↦ f (x,t) is convex on C; - for each x ∈ C, the function t ↦ f (x,t) is continuous on T. Then f is continuous on C × T (i.e. jointly continuous in x and t).
open scoped BigOperators
theorem convexOn_continuousOn_prod_of_locallyCompact {n : ℕ} {T : Type*} [TopologicalSpace T]
[LocallyCompactSpace T] {C : Set (EuclideanSpace Real (Fin n))} (hCopen : IsOpen C)
(hCconv : Convex ℝ C) (f : EuclideanSpace Real (Fin n) × T → Real)
(hfconv : ∀ t, ConvexOn ℝ C (fun x => f (x, t)))
(hfcont : ∀ x ∈ C, Continuous fun t => f (x, t)) :
ContinuousOn f (C ×ˢ (Set.univ : Set T))
Theorem 10.7 (variant). The same conclusion holds if the continuity-in-t assumption is weakened as follows: there exists C' ⊆ C with closure C' ⊇ C such that t ↦ f (x,t) is continuous for each x ∈ C'.
open scoped BigOperators
theorem convexOn_continuousOn_prod_of_locallyCompact_of_exists_subset_closure {n : ℕ} {T : Type*}
[TopologicalSpace T] [LocallyCompactSpace T] {C : Set (EuclideanSpace Real (Fin n))}
(hCopen : IsOpen C) (hCconv : Convex ℝ C) (f : EuclideanSpace Real (Fin n) × T → Real)
(hfconv : ∀ t, ConvexOn ℝ C (fun x => f (x, t))) {C' : Set (EuclideanSpace Real (Fin n))}
(hC'sub : C' ⊆ C) (hCclosure : C ⊆ closure C')
(hfcont : ∀ x ∈ C', Continuous fun t => f (x, t)) :
ContinuousOn f (C ×ˢ (Set.univ : Set T))
If x ∈ interior S' and x ∈ closure C', then for every δ > 0 there exists a point z ∈ C' ∩ S' within distance δ of x.
open scoped BigOperators
lemma Section10.exists_point_mem_C'_mem_S'_dist_lt_of_mem_interior_of_mem_closure {n : ℕ}
{C' S' : Set (EuclideanSpace Real (Fin n))} {x : EuclideanSpace Real (Fin n)}
(hxS' : x ∈ interior S') (hxC' : x ∈ closure C') :
∀ δ : ℝ, 0 < δ → ∃ z, z ∈ C' ∧ z ∈ S' ∧ dist z x < δ
A compact set admits a finite δ-net with points in C' ∩ S', provided C' is dense around points of S and S sits inside interior S'.
open scoped BigOperators
lemma Section10.exists_finset_net_in_C'_inter_S'_cover {n : ℕ}
{C' S S' : Set (EuclideanSpace Real (Fin n))} (hScomp : IsCompact S)
(hSint : S ⊆ interior S') (hSclosure : S ⊆ closure C') :
∀ δ : ℝ, 0 < δ →
∃ t : Finset (EuclideanSpace Real (Fin n)),
(∀ z ∈ t, z ∈ C' ∧ z ∈ S') ∧ ∀ x ∈ S, ∃ z ∈ t, dist x z < δ
A finite family of convergent real sequences is uniformly Cauchy on the index set.
open scoped BigOperators
lemma Section10.exists_tail_cauchy_on_finset_of_pointwise_tendsto {n : ℕ}
{f : ℕ → EuclideanSpace Real (Fin n) → Real} (t : Finset (EuclideanSpace Real (Fin n)))
(hpoint :
∀ z ∈ t, ∃ l : Real, Filter.Tendsto (fun i => f i z) Filter.atTop (nhds l)) :
∀ ε : Real, 0 < ε → ∃ N : ℕ, ∀ i ≥ N, ∀ j ≥ N, ∀ z ∈ t, ‖f i z - f j z‖ < ε
Triangle inequality estimate upgrading control on a finite δ-net to control on all of S under an equi-Lipschitz hypothesis.
open scoped BigOperators
lemma Section10.uniformCauchyOn_of_equiLipschitz_and_finset_net {n : ℕ}
{f : ℕ → EuclideanSpace Real (Fin n) → Real} {S S' : Set (EuclideanSpace Real (Fin n))}
{t : Finset (EuclideanSpace Real (Fin n))} {L : NNReal} (hL : ∀ i, LipschitzOnWith L (f i) S')
(hSsub : S ⊆ S') (htsub : ∀ z ∈ t, z ∈ S') {ε δ : Real}
(hδbound : (((L : Real) + 1) * δ) ≤ ε / 3) {N : ℕ}
(hfin : ∀ i ≥ N, ∀ j ≥ N, ∀ z ∈ t, dist (f i z) (f j z) < ε / 3)
(hcover : ∀ x ∈ S, ∃ z ∈ t, dist x z < δ) :
∀ i ≥ N, ∀ j ≥ N, ∀ x ∈ S, dist (f i x) (f j x) < ε
If convex functions converge pointwise on a convex set, the pointwise limit is convex.
open scoped BigOperators
lemma Section10.convexOn_lim_of_pointwise_tendsto {n : ℕ}
{C : Set (EuclideanSpace Real (Fin n))} (hCconv : Convex ℝ C)
{f : ℕ → EuclideanSpace Real (Fin n) → Real} {g : EuclideanSpace Real (Fin n) → Real}
(hg : ∀ x ∈ C, Filter.Tendsto (fun i => f i x) Filter.atTop (nhds (g x)))
(hfconv : ∀ i, ConvexOn ℝ C (f i)) : ConvexOn ℝ C g
Theorem 10.8. Let C be a relatively open convex set, and let f 0, f 1, … be a sequence of finite convex functions on C. Suppose that there exists C' ⊆ C with closure C' ⊇ C such that for each x ∈ C' the limit lim_{i → ∞} f i x exists (as a finite real number). Then the limit exists for every x ∈ C, the resulting pointwise limit function f is finite and convex on C, and the sequence f 0, f 1, … converges to f uniformly on each closed bounded subset of C.
open scoped BigOperators
theorem convexOn_seq_exists_convex_limit_of_tendstoOn_dense {n : ℕ}
{C : Set (EuclideanSpace Real (Fin n))} (hCopen : IsOpen C) (hCconv : Convex ℝ C)
(f : ℕ → EuclideanSpace Real (Fin n) → Real) (hfconv : ∀ i, ConvexOn ℝ C (f i))
{C' : Set (EuclideanSpace Real (Fin n))} (hC'sub : C' ⊆ C) (hCclosure : C ⊆ closure C')
(hpoint :
∀ x ∈ C', ∃ l : Real, Filter.Tendsto (fun i => f i x) Filter.atTop (nhds l)) :
∃ f_lim : EuclideanSpace Real (Fin n) → Real,
(∀ x ∈ C, Filter.Tendsto (fun i => f i x) Filter.atTop (nhds (f_lim x))) ∧
ConvexOn ℝ C f_lim ∧
∀ S : Set (EuclideanSpace Real (Fin n)),
IsClosed S → Bornology.IsBounded S → S ⊆ C →
∀ ε : Real, 0 < ε → ∃ N : ℕ, ∀ i ≥ N, ∀ x ∈ S, ‖f i x - f_lim x‖ < ε
The pointwise maximum of two convex functions is convex.
open scoped BigOperators
lemma Section10.convexOn_max {n : ℕ} {C : Set (EuclideanSpace Real (Fin n))}
{f g : EuclideanSpace Real (Fin n) → Real} (hf : ConvexOn ℝ C f) (hg : ConvexOn ℝ C g) :
ConvexOn ℝ C (fun x => max (f x) (g x))
If limsup (u i) ≤ a, then max (u i) a tends to a.
open scoped BigOperators
lemma Section10.tendsto_max_of_limsup_le {u : ℕ → Real} {a : Real}
(h : Filter.limsup (fun i : ℕ => (u i : EReal)) Filter.atTop ≤ (a : EReal)) :
Filter.Tendsto (fun i => max (u i) a) Filter.atTop (nhds a)
If g is uniformly within ε of f on S, and f ≤ g, then g ≤ f + ε on S.
open scoped BigOperators
lemma Section10.le_add_of_uniform_norm_sub_lt_of_le {X : Type*}
{g : ℕ → X → Real} {f : X → Real} {S : Set X} {ε : Real} {N : ℕ}
(hU : ∀ i ≥ N, ∀ x ∈ S, ‖g i x - f x‖ < ε) (hle : ∀ i x, f x ≤ g i x) :
∀ i ≥ N, ∀ x ∈ S, g i x ≤ f x + ε
Corollary 10.8.1. Let f be a finite convex function on a relatively open convex set C. Let f₁, f₂, … be a sequence of finite convex functions on C such that limsup_{i → ∞} fᵢ(x) ≤ f(x) for all x ∈ C. Then for each closed bounded subset S of C and each ε > 0, there exists i₀ such that fᵢ(x) ≤ f(x) + ε for all i ≥ i₀ and all x ∈ S.
open scoped BigOperators
theorem convexOn_eventually_le_add_of_limsup_le {n : ℕ}
{C : Set (EuclideanSpace Real (Fin n))} (hCopen : IsOpen C) (hCconv : Convex ℝ C)
{f : EuclideanSpace Real (Fin n) → Real} (hfconv : ConvexOn ℝ C f)
(fseq : ℕ → EuclideanSpace Real (Fin n) → Real) (hfseqconv : ∀ i, ConvexOn ℝ C (fseq i))
(hlimsup :
∀ x ∈ C, Filter.limsup (fun i : ℕ => (fseq i x : EReal)) Filter.atTop ≤ (f x : EReal)) :
∀ S : Set (EuclideanSpace Real (Fin n)),
IsClosed S → Bornology.IsBounded S → S ⊆ C →
∀ ε : Real, 0 < ε → ∃ i0 : ℕ, ∀ i ≥ i0, ∀ x ∈ S, fseq i x ≤ f x + ε
Any subset of a separable space contains a countable subset with closure containing the set.
open scoped BigOperators
lemma Section10.exists_countable_subset_closure_superset {n : ℕ}
(C' : Set (EuclideanSpace Real (Fin n))) :
∃ C'' : Set (EuclideanSpace Real (Fin n)),
C'' ⊆ C' ∧ C''.Countable ∧ C' ⊆ closure C''
A diagonal compactness argument: if each coordinate sequence i ↦ f i (x j) is bounded in ℝ, then some subsequence converges at every point in the countable range {x j}.
open scoped BigOperators
lemma Section10.exists_strictMono_subseq_pointwise_tendsto_on_countable_range {α : Type*}
(x : ℕ → α) (f : ℕ → α → ℝ)
(hbounded : ∀ j, Bornology.IsBounded (Set.range fun i : ℕ => f i (x j))) :
∃ φ : ℕ → ℕ, StrictMono φ ∧ ∀ j, ∃ l : ℝ,
Filter.Tendsto (fun i => f (φ i) (x j)) Filter.atTop (nhds l)
Turn convergence along an enumeration x : ℕ → α into convergence on Set.range x.
open scoped BigOperators
lemma Section10.pointwise_tendsto_on_countable_subset_from_range {α : Type*} {f : ℕ → α → ℝ}
{x : ℕ → α} {φ : ℕ → ℕ}
(hpoint : ∀ j, ∃ l : ℝ, Filter.Tendsto (fun i => f (φ i) (x j)) Filter.atTop (nhds l)) :
∀ z ∈ Set.range x, ∃ l : ℝ, Filter.Tendsto (fun i => f (φ i) z) Filter.atTop (nhds l)
Theorem 10.9. Let C be a relatively open convex set, and let f₁, f₂, … be a sequence of finite convex functions on C. Suppose that for each x ∈ C the real sequence f₁ x, f₂ x, … is bounded (or merely for each x in some dense subset C' of C). Then there exists a subsequence of f₁, f₂, … which converges uniformly on each closed bounded subset of C to a finite convex function f.
open scoped BigOperators
theorem convexOn_seq_exists_subseq_uniformlyConvergentOn_of_boundedOn_dense {n : ℕ}
{C : Set (EuclideanSpace Real (Fin n))} (hCopen : IsOpen C) (hCconv : Convex ℝ C)
(f : ℕ → EuclideanSpace Real (Fin n) → Real) (hfconv : ∀ i, ConvexOn ℝ C (f i))
(hbounded :
∃ C' : Set (EuclideanSpace Real (Fin n)),
C' ⊆ C ∧ C ⊆ closure C' ∧
∀ x ∈ C', Bornology.IsBounded (Set.range fun i : ℕ => f i x)) :
∃ φ : ℕ → ℕ,
StrictMono φ ∧
∃ f_lim : EuclideanSpace Real (Fin n) → Real,
ConvexOn ℝ C f_lim ∧
∀ S : Set (EuclideanSpace Real (Fin n)),
IsClosed S → Bornology.IsBounded S → S ⊆ C →
∀ ε : Real, 0 < ε →
∃ N : ℕ, ∀ i ≥ N, ∀ x ∈ S, ‖f (φ i) x - f_lim x‖ < ε