Chapter 11 — Convex Minimization Problems

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

Proposition 11.1

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Proposition 11.1 (i): a lower semicontinuous function has the same supremum on a set and on its closure.

open scoped InnerProductSpace Topology universe u namespace ERealFunction variable {H : Type u} [TopologicalSpace H] theorem sSup_image_closure_eq_of_lowerSemicontinuous {f : H → EReal} (hf : LowerSemicontinuous f) (C : Set H) : sSup (f '' closure C) = sSup (f '' C)

Definition 11.3

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Definition 11.3: argminOn f C is the set of minimizers of f over the constraint set C, written Argmin[C] f. The global argmin set is the special case argminOn f Set.univ, written Argmin f.

universe u namespace ERealFunction variable {X : Type u} def argminOn (f : X → EReal) (C : Set X) : Set X := {x | x ∈ C ∧ IsMinOn f C x}

Proposition 11.4

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Proposition 11.4: a finite-valued local minimizer of a convex ]-∞,+∞]-valued function is a global minimizer.

universe u namespace ERealFunction variable {H : Type u} [TopologicalSpace H] [AddCommGroup H] [Module ℝ H] [IsTopologicalAddGroup H] [ContinuousSMul ℝ H] theorem mem_argmin_of_isLocalMin_of_convexOn_effectiveDomain (f : H → Set.Ioi (⊥ : EReal)) (hconv : ConvexOn f (effectiveDomain f)) {x : H} (hx : x ∈ effectiveDomain f) (hlocal : IsLocalMin f.asEReal x) : x ∈ Argmin f.asEReal

Proposition 11.5

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Proposition 11.5: if x minimizes a convex extended-real-valued function over a set C, lies in the effective domain, and belongs to interior C, then x is a global minimizer.

universe u namespace ERealFunction variable {H : Type u} [TopologicalSpace H] [AddCommGroup H] [Module ℝ H] [IsTopologicalAddGroup H] [ContinuousSMul ℝ H] theorem mem_argmin_of_mem_argminOn_of_mem_interior_of_convexOn_effectiveDomain (f : H → Set.Ioi (⊥ : EReal)) (hconv : ConvexOn f (effectiveDomain f)) {C : Set H} {x : H} (hx : x ∈ effectiveDomain f) (hargmin : x ∈ Argmin[C] f.asEReal) (hxint : x ∈ interior C) : x ∈ Argmin f.asEReal

Proposition 11.6

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Proposition 11.6: the set of global minimizers of a quasiconvex ]-∞,+∞]-valued function is convex.

universe u namespace ERealFunction variable {𝕜 : Type*} [Semiring 𝕜] [PartialOrder 𝕜] variable {H : Type u} [AddCommMonoid H] [SMul 𝕜 H] theorem convex_argmin_of_quasiconvexOn_univ {f : H → EReal} (hf_quasi : QuasiconvexOn 𝕜 Set.univ f) : Convex 𝕜 (Argmin f)

Proposition 11.7

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Proposition 11.7 (1): an even convex ]-∞,+∞]-valued function on ℝ attains its global minimum at 0.

open Set namespace ERealFunction variable (φ : ℝ → Set.Ioi (⊥ : EReal)) theorem zero_mem_argmin_of_even_convexOn (hconv : ConvexOn φ (effectiveDomain φ)) (heven : Function.Even φ.asEReal) : (0 : ℝ) ∈ Argmin φ.asEReal

Proposition 11.8

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Proposition 11.8 (1): if the indicator-augmented function f + ι_C is strictly quasiconvex, then f has at most one minimizer on C.

universe u namespace ERealFunction variable {H : Type u} variable [AddCommMonoid H] [Module ℝ H] theorem argminOn_subsingleton_of_indicator_strictlyQuasiconvex {f : H → EReal} {C : Set H} (hstrict : StrictlyQuasiconvex (f + (ι[C]).asEReal)) : (Argmin[C] f).Subsingleton

Corollary 11.9

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Corollary 11.9: a strictly convex ]-∞,+∞]-valued function with nonempty effective domain has at most one global minimizer.

universe u namespace ERealFunction variable {H : Type u} [AddCommGroup H] [Module ℝ H] theorem argmin_subsingleton_of_nonempty_effectiveDomain_of_strictlyConvex {f : H → Set.Ioi (⊥ : EReal)} (hdom : (effectiveDomain f).Nonempty) (hstrict : StrictlyConvex f) : (Argmin f.asEReal).Subsingleton

Theorem 11.10

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Theorem 11.10: if a lower semicontinuous quasiconvex ]-∞,+∞]-valued function on a real Hilbert space has a nonempty bounded intersection of a closed convex constraint set with one of its lower level sets, then its constrained argmin set is nonempty.

open scoped InnerProductSpace Topology universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] theorem argminOn_nonempty_of_quasiconvexOn_univ_of_nonempty_bounded_inter_lowerLevelSet {f : H → EReal} {C : Set H} (hf_quasi : QuasiconvexOn ℝ Set.univ f) (hf_lsc : LowerSemicontinuous f) (hC_closed : IsClosed C) (hC_convex : Convex ℝ C) (ξ : ℝ) (hlevel_nonempty : (C ∩ lowerLevelSet f ξ).Nonempty) (hlevel_bounded : Bornology.IsBounded (C ∩ lowerLevelSet f ξ)) : (Argmin[C] f).Nonempty

Definition 11.11

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Definition 11.11 (1): an extended-real-valued function is coercive when its values tend to +∞ along every net for which ‖x‖ → +∞.

universe u namespace ERealFunction variable {H : Type u} [SeminormedAddCommGroup H] abbrev Coercive (f : H → EReal) : Prop := Filter.Tendsto f (Bornology.cobounded H) (nhds (⊤ : EReal))

Proposition 11.12

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Proposition 11.12: an extended-real-valued function is coercive if and only if each of its real lower level sets is bounded.

universe u namespace ERealFunction variable {H : Type u} [SeminormedAddCommGroup H] -- Proof sketch: if some lower level set is unbounded, choose a sequence in it with norms tending -- to `+∞`; along that sequence the function values stay below the same real threshold, so `f` is -- not coercive. Conversely, if every real lower level set is bounded and `‖xₙ‖ → +∞`, then for -- each real threshold `ξ` the tail of `xₙ` eventually leaves `lowerLevelSet f ξ`, which forces -- `f xₙ → +∞`. theorem coercive_iff_bounded_lowerLevelSet (f : H → EReal) : Coercive f ↔ ∀ ξ : ℝ, Bornology.IsBounded (lowerLevelSet f ξ)

Proposition 11.13

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Proposition 11.13: for a function f ∈ Γ₀(H) on a finite-dimensional real normed space, coercivity is equivalent to the existence of a nonempty bounded real lower level set.

universe u namespace ERealFunction open Filter variable {H : Type u} [NormedAddCommGroup H] [NormedSpace ℝ H] [FiniteDimensional ℝ H] theorem coercive_iff_exists_nonempty_bounded_lowerLevelSet_of_mem_gammaZero (f : H → Set.Ioi (⊥ : EReal)) (hf : f ∈ Γ₀(H)) : Coercive f.asEReal ↔ ∃ ξ : ℝ, (lowerLevelSet f.asEReal ξ).Nonempty ∧ Bornology.IsBounded (lowerLevelSet f.asEReal ξ)

Proposition 11.14

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Proposition 11.14: if f ∈ Γ₀(H) and g is supercoercive, then the canonical underlying extended-real-valued pointwise sum (f + g).asEReal is supercoercive.

universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] -- Proof sketch: use Theorem 9.20 to minorize `f` by a continuous affine functional -- `x ↦ ⟪x, u⟫ + η`, divide by `‖x‖`, bound the affine term below by `-‖u‖ + η / ‖x‖`, and then -- combine this lower bound with the supercoercive growth of `g`. The `Γ₀(H)`-valued wrapper -- `pointwiseAdd f g` is only a bridge; the supercoercive owner lives on the underlying -- `EReal`-valued sum. theorem pointwiseAdd_supercoercive_of_mem_gammaZero (f g : H → Set.Ioi (⊥ : EReal)) (hf : f ∈ Γ₀(H)) (hg_super : Supercoercive g.asEReal) : Supercoercive (f + g).asEReal

Proposition 11.15

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Proposition 11.15: if f ∈ Γ₀(H), if C is a nonempty closed convex set, and if either (i) the coerced function x ↦ (f x : EReal) is coercive or (ii) C is bounded, then f attains a minimum over C.

open scoped InnerProductSpace Topology universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] -- Proof sketch: this is a bridge from the source-facing `Γ₀(H)` hypothesis to the core bounded -- lower-level-set existence theorem. If `C ∩ effectiveDomain f` is empty, then `f.asEReal = ⊤` -- everywhere on `C`, so every point of the nonempty constraint set is a constrained minimizer. If -- `C ∩ effectiveDomain f` is nonempty, choose a feasible point to define a finite level `ξ`, -- bound `C ∩ lowerLevelSet (fun y ↦ (f y : EReal)) ξ` either from coercivity via Proposition -- 11.12 or directly from boundedness of `C`, and apply Theorem 11.10. theorem argminOn_nonempty_of_mem_gammaZero_of_coercive_or_bounded {f : H → Set.Ioi (⊥ : EReal)} (hf : f ∈ Γ₀(H)) {C : Set H} (hC_closed : IsClosed C) (hC_convex : Convex ℝ C) (hC_nonempty : C.Nonempty) (hcase : Coercive f.asEReal ∨ Bornology.IsBounded C) : (Argmin[C] f.asEReal).Nonempty

Corollary 11.16

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Corollary 11.16 (1): if g ∈ Γ₀(H) and if either (i) f is supercoercive or (ii) f is coercive while g is bounded below, then f + g is coercive.

open scoped InnerProductSpace Topology universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] variable {f g : H → Set.Ioi (⊥ : EReal)} -- Proof sketch: in case (i), commute the summands and apply Proposition 11.14 to `g + f`, then -- pass from supercoercive to coercive. In case (ii), use the zero-slope affine minorant for `g` -- together with coercivity of `f` and -- `pointwiseAdd_coercive_of_coercive_of_zeroSlopeAffineMinorant`. The `Γ₀(H)` hypothesis on `f` -- and the textbook domain-intersection hypothesis are both redundant for this coercivity -- conclusion, so the Lean statement keeps only the assumptions that are actually used. theorem pointwiseAdd_coercive_of_supercoercive_or_coercive_bddBelow (hg : g ∈ Γ₀(H)) (hcase : Supercoercive f.asEReal ∨ (Coercive f.asEReal ∧ BddBelow (Set.range g))) : Coercive (f + g).asEReal

Corollary 11.17

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Corollary 11.17 (1): a lower-semicontinuous strongly convex function is supercoercive.

universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] theorem supercoercive_of_lowerSemicontinuous_of_stronglyConvex {f : H → Set.Ioi (⊥ : EReal)} (hlower : LowerSemicontinuous f.asEReal) {β : ℝ} (hstrong : StronglyConvex f β) : Supercoercive f.asEReal

Proposition 11.18

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Proposition 11.18 (1): clause (i). For a bounded sequence, the real-valued bridge of the asymptotic-center objective is strongly convex on the whole Hilbert space with constant 2.

open ERealFunction Filter open scoped Topology universe u variable {H : Type u} [NormedAddCommGroup H] variable [InnerProductSpace ℝ H] -- Proof sketch: boundedness makes the canonical `EReal` owner finite everywhere, so its `toReal` -- bridge satisfies the pointwise identity -- `‖α x + (1 - α) y - zₙ‖² = α ‖x - zₙ‖² + (1 - α) ‖y - zₙ‖² - α (1 - α) ‖x - y‖²`; -- passing to `Filter.limsup` yields the real-valued source clause. theorem asymptoticCenterObjective_strongConvexOn_univ {zₙ : ℕ → H} (hzₙ_bdd : Bornology.IsBounded (Set.range zₙ)) : StrongConvexOn (Set.univ : Set H) (2 : ℝ) (fun x ↦ (asymptoticCenterObjective zₙ x : EReal).toReal)

Corollary 11.19

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Corollary 11.19: for a nonexpansive self-map of a nonempty closed convex subset of a real Hilbert space, nonemptiness of the fixed-point set, boundedness of every Picard orbit, and boundedness of some Picard orbit are equivalent.

universe u variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] -- Proof sketch: `(i) → (ii)` comes from Example 5.3.2 and Proposition 5.4(1), since every -- Picard orbit of a nonexpansive self-map is Fejér monotone with respect to the ambient -- fixed-point set. -- `(ii) → (iii)` is immediate. For `(iii) → (i)`, apply Proposition 11.18(6) to the bounded -- orbit to see that its asymptotic center relative to `C` belongs to -- `Subtype.val '' fixedPoints T`, hence `Function.fixedPoints T` is nonempty. theorem fixedPoints_nonempty_bounded_picard_orbits_tfae_of_nonexpansive {C : Set H} (hC_nonempty : C.Nonempty) (hC_closed : IsClosed C) (hC_convex : Convex ℝ C) {T : C → C} (hT : LipschitzWith 1 T) : List.TFAE [ (Function.fixedPoints T).Nonempty, ∀ z₀ : C, Bornology.IsBounded (Set.range fun n ↦ ((T^[n]) z₀ : H)), ∃ z₀ : C, Bornology.IsBounded (Set.range fun n ↦ ((T^[n]) z₀ : H)) ]

Proposition 11.20

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Proposition 11.20: every minimizing sequence of a coercive extended-real-valued function is bounded.

open Filter universe u namespace ERealFunction variable {H : Type u} [SeminormedAddCommGroup H] -- Proof sketch: choose a real level `ξ` so that the minimizing-sequence values eventually satisfy -- `f (x n) ≤ ξ`; Proposition 11.12 makes `lowerLevelSet f ξ` bounded, hence the tail of the -- sequence is bounded, and adjoining finitely many initial terms preserves boundedness. theorem isBounded_range_of_isMinimizingSequence_of_coercive {f : H → EReal} (hf : Coercive f) {x : ℕ → H} (hx : IsMinimizingSequence f x) : Bornology.IsBounded (Set.range x)

Proposition 11.21

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Proposition 11.21: the weak limit of a weakly convergent minimizing sequence of a lower semicontinuous quasiconvex ]-∞,+∞]-valued function on a real Hilbert space is a global minimizer.

open Filter open scoped InnerProductSpace Topology universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] theorem mem_argmin_of_isMinimizingSequence_of_tendsto_weakly_of_quasiconvexOn_univ {f : H → EReal} (hf_quasi : QuasiconvexOn ℝ Set.univ f) (hf_lsc : LowerSemicontinuous f) {xₙ : ℕ → H} (hxₙ : IsMinimizingSequence f xₙ) {x : H} (hweak : Tendsto (fun n ↦ toWeakSpace ℝ H (xₙ n)) atTop (𝓝 (toWeakSpace ℝ H x))) : x ∈ Argmin f

Remark 11.22

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Remark 11.22 (1): clause (i), real case. For a function in Γ₀(ℝ), any sequence in the effective domain that converges strongly to a global minimizer is a minimizing sequence.

open Filter open scoped InnerProductSpace Topology universe u namespace ERealFunction theorem isMinimizingSequence_of_tendsto_of_mem_argmin_real {f : ℝ → Set.Ioi (⊥ : EReal)} (hf : f ∈ Γ₀(ℝ)) {xₙ : ℕ → ℝ} (hxdom : ∀ n, xₙ n ∈ effectiveDomain f) {x : ℝ} (hxmin : x ∈ Argmin f.asEReal) (hstrong : Tendsto xₙ atTop (𝓝 x)) : IsMinimizingSequence f.asEReal xₙ

Example 11.23

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Example 11.23: the closed quadratic perspective is exactly the specialization p = 2 of the canonical Γ₀ theorem normPowerPerspectiveAtOrigin_mem_gammaZero from Example 9.43.

open Filter open Set namespace ERealFunction local notation "q" => (fun y : ℝ × ℝ ↦ (normPowerPerspectiveAtOrigin 2 y : EReal)) #check (normPowerPerspectiveAtOrigin_mem_gammaZero 2 (by norm_num) : normPowerPerspectiveAtOrigin 2 ∈ Γ₀(ℝ × ℝ)) -- Proof sketch: show that the global minimizers are exactly the points where the quadratic -- perspective takes its minimal value `0`, namely the nonnegative horizontal axis.

Example 11.24

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Example 11.24: for f(ξ₁, ξ₂) = ξ₁ + ‖(ξ₁, ξ₂)‖, the sequence xₙ = (-n, 1) is minimizing, while every term stays at distance 1 from Argmin f.

open Filter namespace ERealFunction local notation "f" => example11_24Function.toEReal.asEReal theorem example11_24_minimizingSequence_and_infDist_argmin : IsMinimizingSequence f example11_24Sequence ∧ ∀ n : ℕ, Metric.infDist (example11_24Sequence n) (Argmin f) = 1

Example 11.25

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Example 11.25: evaluating closedScalarRelativeEntropy gives the three-branch textbook formula: the logarithmic expression on the positive orthant, the vertical branch {0} × ℝ₊, and +∞ otherwise.

open Filter namespace ERealFunction local notation "f" => closedScalarRelativeEntropy.asEReal @[simp] theorem closedScalarRelativeEntropy_apply (ξ₁ ξ₂ : ℝ) : (closedScalarRelativeEntropy (ξ₁, ξ₂) : EReal) = if 0 < ξ₁ ∧ 0 < ξ₂ then ((ξ₁ * Real.log (ξ₁ / ξ₂) - ξ₁ + ξ₂ : ℝ) : EReal) else if ξ₁ = 0 ∧ 0 ≤ ξ₂ then ξ₂ else ⊤

Example 11.26

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Example 11.26 (1): the alternating sequence xₙ = (-1)^n x is a minimizing sequence for the indicator of the symmetric segment [-x,x].

open Filter open Set open scoped InnerProductSpace Topology universe u namespace ERealFunction variable {H : Type u} variable [AddCommGroup H] [Module ℝ H] variable (x : H) local notation "xₙ" => fun n : ℕ ↦ ((-1 : ℝ) ^ n) • x -- Proof sketch: each term of `n ↦ ((-1 : ℝ) ^ n) • x` is one of the endpoints `x` or `-x`, -- hence lies in `segment ℝ (-x) x`; therefore the indicator of that segment vanishes along the -- whole sequence, so the function values converge to the infimum `0`. theorem alternatingSequence_isMinimizing_symmetricSegmentIndicator (x : H) : IsMinimizingSequence (ι[segment ℝ (-x) x]).asEReal xₙ

Example 11.27

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Example 11.27: if ωₙ → 0, then the Hilbert basis itself is a minimizing sequence for the weighted squared-coordinate series, and it converges weakly to 0.

open Filter open scoped InnerProductSpace Topology universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] theorem weightedHilbertBasisSquareSeries_basis_isMinimizingSequence_and_tendsto_weakly (ω : ℕ → NNReal) (b : HilbertBasis ℕ ℝ H) (hω_tendsto : Tendsto (fun n ↦ (ω n : ℝ)) atTop (𝓝 0)) : IsMinimizingSequence (weightedHilbertBasisSquareSeries ω b).asEReal b ∧ Tendsto (fun n ↦ toWeakSpace ℝ H (b n)) atTop (𝓝 (0 : WeakSpace ℝ H))

Example 11.28

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Example 11.28: for every iterate after the initial one, alternating minimization produces the diagonal point (ξ₂,₀, ξ₂,₀) and the function value there is exactly ξ₂,₀.

open Filter namespace ERealFunction local notation "f" => example11_28Function.asEReal -- Proof sketch: write the function as the sum of the positive-orthant indicator and the maximum -- of two affine forms. Lower semicontinuity and convexity come from these two pieces, while the -- explicit formula shows the function is proper. theorem alternatingMinimization_eq_diagonal_and_value (x0 : ℝ × ℝ) (hx0_2 : 0 < x0.2) {n : ℕ} (hn : 1 ≤ n) : example11_28Recurrence x0 n = (x0.2, x0.2) ∧ f (example11_28Recurrence x0 n) = x0.2

Proposition 11.29

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Proposition 11.29 (1): clause (i). If some lower level set strictly above the infimum of f is bounded, then every minimizing sequence of f has a weak sequential cluster point.

open Filter open scoped InnerProductSpace Topology universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] variable [CompleteSpace H] -- Proof sketch: `hxₙ.tendsto` makes the minimizing sequence eventually enter every lower level set -- strictly above `sInf (Set.range f)`. Once the tail lies in the bounded set `lowerLevelSet f ξ`, -- put that set inside a closed ball and apply -- `exists_subsequence_tendsto_weakly_mem_of_bounded_isClosed_convex` to obtain a weakly -- convergent subsequence, hence a weak sequential cluster point. theorem IsMinimizingSequence.exists_weakSequentialClusterPoint_of_bounded_lowerLevelSet {f : H → EReal} {xₙ : ℕ → H} (hxₙ : IsMinimizingSequence f xₙ) {ξ : ℝ} (hξ : sInf (Set.range f) < (ξ : EReal)) (hlevel_bounded : Bornology.IsBounded (lowerLevelSet f ξ)) : ∃ x : H, IsSequentialClusterPt (fun n ↦ toWeakSpace ℝ H (xₙ n)) (toWeakSpace ℝ H x)

Corollary 11.30

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Corollary 11.30 (1): clause (i). Every minimizing sequence of a coercive ]-∞,+∞]-valued function has a weak sequential cluster point.

open Filter open scoped InnerProductSpace Topology universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] variable [CompleteSpace H] -- Proof sketch: Proposition 11.20 makes every minimizing sequence of a coercive function -- bounded. Equivalently, Proposition 11.12 makes every real lower level set bounded, so -- Proposition 11.29 (1) yields a weak sequential cluster point. No `Γ₀(H)` structure is used. theorem exists_weakSequentialClusterPoint_of_coercive {f : H → Set.Ioi (⊥ : EReal)} (hf_coe : Coercive f.asEReal) {xₙ : ℕ → H} (hxₙ : IsMinimizingSequence f.asEReal xₙ) : ∃ x : H, IsSequentialClusterPt (fun n ↦ toWeakSpace ℝ H (xₙ n)) (toWeakSpace ℝ H x)

Proposition 11.31

Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.

Source header · Attribution source

Proposition 11.31: if f ∈ Γ₀(H), C is bounded, closed, disjoint from Argmin f, and uniformly convex in the source-facing midpoint-ball sense (11.13), then every minimizing sequence of f + ι_C converges strongly to the unique minimizer of f over C.

open Filter open scoped Topology universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] theorem existsUnique_mem_argminOn_and_tendsto_of_isMinimizingSequence_add_indicator {f : H → Set.Ioi (⊥ : EReal)} (hf : f ∈ Γ₀(H)) {C : Set H} (hC_bounded : Bornology.IsBounded C) (hC_closed : IsClosed C) (hC_disjoint : Disjoint C (Argmin f.asEReal)) (hC_uniformlyConvex : Set.UniformlyConvex C) {xₙ : ℕ → H} (hxₙ : IsMinimizingSequence (f.asEReal + (ι[C]).asEReal) xₙ) : ∃! x : H, x ∈ Argmin[C] f.asEReal ∧ Tendsto xₙ atTop (𝓝 x)