Chapter 18 — Further Differentiability Results
Proofs omitted. Hover over highlighted terms to see their Lean types.
Proposition 18.1
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 18.1: for a convex ]-∞,+∞]-valued function on a real Hilbert space, ambient continuity at x in the source sense x ∈ cont f is equivalent to Fréchet differentiability of the finite-valued representative at x together with the source symmetric second-difference estimate at every positive tolerance.
universe u
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
-- Proof sketch: for the forward implication, `x ∈ cont f` supplies a ball on which the
-- finite-valued representative is an ambient convex real function, so Fréchet differentiability at
-- `x` gives a first-order expansion whose error is `o (‖h‖)`; evaluating it at `h = ± η y` and
-- adding cancels the linear part and yields the source-facing symmetric second-difference estimate
-- for every unit vector `y`. For the reverse implication, Proposition 16.17 applies at the
-- Chapter 16 bridge point `ContinuousAtOnEffectiveDomain.of_mem_cont hxcont` to provide a
-- subgradient at `x`; combine the subgradient inequality with the symmetric estimate and the
-- secant-slope monotonicity theorem in the radial direction to obtain the uniform little-o
-- remainder required for Fréchet differentiability.
theorem differentiableAt_toReal_iff_forall_pos_exists_pos_symmetricSecondDifference_lt
(f : H → Set.Ioi (⊥ : EReal)) (hconv : ConvexOn f (effectiveDomain f))
{x : H} (hxcont : x ∈ cont f) :
DifferentiableAt ℝ (fun y ↦ (f y : EReal).toReal) x ↔
∀ ε : Set.Ioi (0 : ℝ), ∃ η : Set.Ioi (0 : ℝ), ∀ y : H, ‖y‖ = 1 →
(f (x + (η : ℝ) • y) : EReal) + (f (x - (η : ℝ) • y) : EReal) -
2 * (f x : EReal) <
(((η : ℝ) * (ε : ℝ) : ℝ) : EReal)Proposition 18.2
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 18.2: for a convex ]-∞,+∞]-valued function, the set S_ε of continuity points x ∈ cont f admitting some positive symmetric second-difference radius whose unit-sphere supremum is strictly below ε is open.
universe u
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [NormedSpace ℝ H]
-- Proof sketch: for `x ∈ S_ε`, fix a witness `η > 0` with symmetric second-difference supremum
-- `σ < ε`. The source continuity datum `x ∈ cont f` already provides a ball contained in
-- `effectiveDomain f` and ambient continuity of the finite-valued representative there, so the
-- convex continuity theorem gives a Lipschitz bound on a smaller ball. Then Proposition 9.27
-- controls the same symmetric second-difference quotient at nearby points by `σ` plus a Lipschitz
-- error term, so a smaller ball around `x` remains in `S_ε`.
theorem isOpen_symmetricSecondDifferenceSublevelSet
(f : H → Set.Ioi (⊥ : EReal))
(hconv : ConvexOn f (effectiveDomain f)) (ε : Set.Ioi (0 : ℝ)) :
IsOpen (symmetricSecondDifferenceSublevelSet f ε)Theorem 18.3
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Theorem 18.3: if a convex ]-∞,+∞]-valued function on a real Hilbert space has a nonempty continuity set cont f, then the points where its finite-valued representative is Fréchet differentiable in the source sense x ∈ cont f form a dense Gδ subset of closure (effectiveDomain f).
universe u
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
theorem dense_isGδ_differentiableAt_toReal_in_closure_effectiveDomain_of_exists_continuityPoint
(f : H → Set.Ioi (⊥ : EReal)) (hconv : ConvexOn f (effectiveDomain f))
(hcont : (cont f).Nonempty) :
Dense {x : closure (effectiveDomain f) |
(x : H) ∈ cont f ∧ DifferentiableAt ℝ (fun y ↦ (f y : EReal).toReal) x} ∧
IsGδ {x : closure (effectiveDomain f) |
(x : H) ∈ cont f ∧ DifferentiableAt ℝ (fun y ↦ (f y : EReal).toReal) x}Theorem 18.5
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Theorem 18.5: for a nonempty finite family of convex ]-∞,+∞]-valued functions, the subdifferential of the pointwise supremum at a common source continuity point is the convex hull of the union of the subdifferentials of the active functions at that point.
open scoped InnerProductSpace
universe u v
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
variable {ι : Type v} [Finite ι] [Nonempty ι]
-- Proof sketch: the pointwise supremum of finitely many convex functions is again convex, and at
-- a common point of the source continuity sets `cont (f i)` the directional derivative of the
-- supremum is the pointwise supremum of the directional derivatives of the active functions. Apply
-- Theorem 17.18 to the maximum and to each active summand, then identify support functions under
-- finite convex hull and weak compactness.
theorem subdifferential_pointwiseSup_eq_convexHull_activeSubdifferentials
(f : ι → H → Set.Ioi (⊥ : EReal))
(hconv : ∀ i, ConvexOn (f i) (effectiveDomain (f i)))
{x : H} (hxcont : ∀ i, x ∈ cont (f i)) :
(∂ pointwiseSup f) x =
convexHull ℝ (⋃ i ∈ activeIndices f x, (∂ f i) x)Proposition 18.6
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 18.6: if x belongs to the domain of the subdifferential of the infimal convolution f □ g, the infimal convolution is exact at x with minimizer y, and f has Gâteaux gradient gradf at y, then the subdifferential of f □ g at x is the singleton {gradf}.
open scoped InnerProductSpace
open InnerProductSpace
universe u
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
-- Proof sketch: Proposition 16.61 rewrites the subdifferential of the exact infimal convolution at
-- `x` as `(∂ f) y ∩ (∂ g) (x - y)`. The Gâteaux derivative hypothesis identifies `(∂ f) y` with
-- the singleton `{gradf}` via Proposition 17.31 once exactness and domain membership of the
-- left-hand side place `y` in `effectiveDomain f`. Rewriting that domain membership as pointwise
-- nonemptiness makes the intersection nonempty, hence it must equal `{gradf}`.
theorem subdifferential_infimalConvolution_eq_singleton_of_mem_dom_of_value_eq_of_gateauxDerivative
(f g : H → Set.Ioi (⊥ : EReal)) (hf : f ∈ Γ₀(H)) (hg : g ∈ Γ₀(H))
(x y gradf : H)
(hx : x ∈ SetValuedOperator.dom (∂ (f □ g)))
(hEq : (f □ g) x = (f y : EReal) + (g (x - y) : EReal))
(hgrad :
HasGateauxDerivativeAt (fun z ↦ (f z : EReal).toReal) (toDual ℝ H gradf) y) :
(∂ (f □ g)) x = ({gradf} : Set H)Proposition 18.7
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 18.7: if f, g ∈ Γ₀(H), if the raw infimal convolution f □ g is proper and lies in Γ(H), if it is exact at x with minimizer y, and if f has Gâteaux gradient gradf at y, then the finite real representative of f □ g inherits the same Gâteaux gradient at x.
open scoped InnerProductSpace
open InnerProductSpace
universe u
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
variable {f g : H → Set.Ioi (⊥ : EReal)}
variable {x y gradf : H}
variable [CompleteSpace H]
theorem hasGateauxDerivativeAt_infimalConvolution_of_value_eq_of_hasGateauxDerivativeAt
(hf : f ∈ Γ₀(H)) (hg : g ∈ Γ₀(H))
(hproper : IsProper (f □ g)) (hgamma : (f □ g) ∈ Γ(H))
(hEq : (f □ g) x = (f y : EReal) + (g (x - y) : EReal))
(hgrad :
HasGateauxDerivativeAt (fun z ↦ (f z : EReal).toReal) (toDualMap ℝ H gradf) y) :
HasGateauxDerivativeAt
(fun z ↦ ((f □ g) z).toReal) (toDualMap ℝ H gradf) xCorollary 18.8
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Corollary 18.8: on a finite-dimensional real Hilbert space, if f, g ∈ Γ₀(H), if f is finite everywhere, supercoercive, and Fréchet differentiable on H, then the finite-valued representative of f □ g is Fréchet differentiable on H.
open InnerProductSpace
open scoped Gradient
open scoped InnerProductSpace
universe u
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
variable [FiniteDimensional ℝ H]
theorem differentiable_infimalConvolution_toReal_of_mem_gammaZero_of_supercoercive_of_differentiable
(f g : H → Set.Ioi (⊥ : EReal)) (hf : f ∈ Γ₀(H)) (hg : g ∈ Γ₀(H))
(hfin : ∀ x : H, (f x : EReal) < ⊤)
(hsuper : Supercoercive f.asEReal)
(hdiff : Differentiable ℝ (fun x ↦ (f x : EReal).toReal)) :
Differentiable ℝ (fun x ↦ ((f □ g) x).toReal)Proposition 18.9
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 18.9: if f ∈ Γ₀(H) and the Fenchel conjugate f*, represented by f∗[hf], is strictly convex on every nonempty convex subset of the domain of its subdifferential, then the finite-valued representative of f is Gâteaux differentiable on interior (effectiveDomain f).
open scoped InnerProductSpace
universe u
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
-- Proof sketch: fix `x ∈ interior (effectiveDomain f)`. Corollary 8.39 gives continuity of `f`
-- on the effective domain at `x`, so Proposition 17.31 (2) will apply once the subdifferential is
-- shown to be a singleton. If `u₁, u₂ ∈ (∂ f) x`, then Corollary 16.30 moves the whole segment
-- `segment ℝ u₁ u₂` into `dom (∂ f*)`, where the hypothesis gives strict convexity of `f*` on
-- that segment. Proposition 16.37 (1) also makes `f*` affine on the segment, forcing `u₁ = u₂`;
-- then Proposition 17.31 (2) yields Gâteaux differentiability at `x`.
theorem
gateauxDifferentiableOn_interior_effectiveDomain_of_strictlyConvexOn_every_nonempty_convex_subset_subdifferentialDom_gammaZeroConjugate
{f : H → Set.Ioi (⊥ : EReal)} (hf : f ∈ Γ₀(H))
(hstrict :
∀ ⦃C : Set H⦄, C.Nonempty → Convex ℝ C →
C ⊆ SetValuedOperator.dom (∂ (f∗[hf])) →
StrictlyConvexOn (f∗[hf]) C) :
GateauxDifferentiableOn (fun x ↦ (f x : EReal).toReal) (interior (effectiveDomain f))Proposition 18.10
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 18.10: if f ∈ Γ₀(H) and gradf is a Gâteaux gradient field of the finite-valued representative of f on interior (effectiveDomain f), then the Fenchel conjugate f*, represented by f∗[hf], is strictly convex on every nonempty convex subset of the gradient image gradf '' interior (effectiveDomain f).
open InnerProductSpace
open scoped InnerProductSpace
universe u
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
variable {f : H → Set.Ioi (⊥ : EReal)} (hf : f ∈ Γ₀(H)) (gradf : H → H)
theorem gammaZeroConjugate_strictlyConvexOn_of_hasGateauxDerivativeOn
(hgrad :
HasGateauxDerivativeOn (fun x ↦ (f x : EReal).toReal)
(fun x ↦ toDualMap ℝ H (gradf x))
(interior (effectiveDomain f)))
{C : Set H} (hC_nonempty : C.Nonempty) (hC_convex : Convex ℝ C)
(hC_subset : C ⊆ gradf '' interior (effectiveDomain f)) :
StrictlyConvexOn (f∗[hf]) CCorollary 18.11
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Corollary 18.11: if f ∈ Γ₀(H) and dom (∂ f) = interior (effectiveDomain f), then the finite-valued representative of f is Gâteaux differentiable on interior (effectiveDomain f) if and only if the Fenchel conjugate f*, represented by f∗[hf], is strictly convex on every nonempty convex subset of the domain of its subdifferential.
open InnerProductSpace
open scoped InnerProductSpace
universe u
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
-- Proof sketch: apply Proposition 18.9 for the reverse implication. For the forward implication,
-- use Proposition 18.10 on the source-facing gradient image and then rewrite
-- `SetValuedOperator.dom (∂ (f∗[hf]))` as that gradient image via
-- `subdifferentialDom_gammaZeroConjugate_eq_gradientImage_of_hasGateauxDerivativeOn`; the extra
-- hypothesis `dom (∂ f) = interior (effectiveDomain f)` is exactly what supplies that bridge.
theorem
gateauxDifferentiableOn_interior_effectiveDomain_iff_strictlyConvexOn_every_nonempty_convex_subset_subdifferentialDom_gammaZeroConjugate
{f : H → Set.Ioi (⊥ : EReal)} (hf : f ∈ Γ₀(H))
(hdom : SetValuedOperator.dom (∂ f) = interior (effectiveDomain f)) :
GateauxDifferentiableOn (fun x ↦ (f x : EReal).toReal) (interior (effectiveDomain f)) ↔
∀ ⦃C : Set H⦄, C.Nonempty → Convex ℝ C →
C ⊆ SetValuedOperator.dom (∂ (f∗[hf])) →
StrictlyConvexOn (f∗[hf]) CCorollary 18.12
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Corollary 18.12 (1): if f ∈ Γ₀(H) and both f and its Fenchel conjugate f*, represented by f∗[hf], satisfy dom (∂ ·) = interior (effectiveDomain ·), then the finite-valued representative of f is Gâteaux differentiable on interior (effectiveDomain f) if and only if f* is strictly convex on interior (effectiveDomain f*).
open InnerProductSpace
open scoped InnerProductSpace
universe u
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
-- Proof sketch: apply Corollary 18.11 to convert Gâteaux differentiability of `f` into strict
-- convexity of `f∗[hf]` on every nonempty convex subset of `dom (∂ f*)`, then use `hdom_conj` to
-- identify that domain with `interior (effectiveDomain (f∗[hf]))`.
-- The reverse implication restricts strict convexity on the whole interior effective domain to the
-- convex subsets required by Corollary 18.11.
theorem
gateauxDifferentiableOn_interior_effectiveDomain_iff_strictlyConvexOn_interior_effectiveDomain_gammaZeroConjugate
{f : H → Set.Ioi (⊥ : EReal)} (hf : f ∈ Γ₀(H))
(hdom : SetValuedOperator.dom (∂ f) = interior (effectiveDomain f))
(hdom_conj :
SetValuedOperator.dom (∂ (f∗[hf])) = interior (effectiveDomain (f∗[hf]))) :
GateauxDifferentiableOn (fun x ↦ (f x : EReal).toReal) (interior (effectiveDomain f)) ↔
StrictlyConvexOn (f∗[hf]) (interior (effectiveDomain (f∗[hf])))Theorem 18.13
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Theorem 18.13 (1): the pointwise gradient norm bound by ψ implies the scalar inner-product bound by φ.
open InnerProductSpace
open scoped ERealFunction Gradient InnerProductSpace
universe u
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
variable (φ : ℝ → ℝ)
-- Proof sketch: apply Cauchy--Schwarz to bound the inner product by
-- `‖x - y‖ * ‖∇ f x - ∇ f y‖`, then invoke the assumed `ψ`-control and the identity
-- `ψ(r) * r = φ(r)` for `r = ‖x - y‖`.
theorem gradient_inner_le_phi_of_gradient_norm_bound_by_psi
(gradf : H → H)
(hφ_zero_nonneg : 0 ≤ φ 0)
(hi : ∀ x y : H, ‖gradf x - gradf y‖ ≤ ψ φ ‖x - y‖) (x y : H) :
⟪x - y, gradf x - gradf y⟫_ℝ ≤ φ ‖x - y‖Corollary 18.14
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Corollary 18.14 (1): if the gradient of a convex Fréchet differentiable function is β-Hölder of order p, then ⟪x - y, ∇f(x) - ∇f(y)⟫ ≤ β ‖x - y‖^(p+1).
open InnerProductSpace
open scoped InnerProductSpace
universe u
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
-- Proof sketch: specialize Theorem 18.13 (1) to `φ(s) = β * |s|^(p+1)`. The Cauchy--Schwarz
-- reduction built into Theorem 18.13 then gives the stated bound
-- `⟪x - y, ∇f(x) - ∇f(y)⟫ ≤ β ‖x - y‖^(p+1)`.
theorem gradient_inner_le_power_bound_of_gradient_norm_le_power_bound
(f : H → ℝ) (gradf : H → H) (β p : ℝ)
(hgrad : ∀ x : H, HasFDerivAt f (toDual ℝ H (gradf x)) x)
(hconv : ConvexOn ℝ (Set.univ : Set H) f)
(hβ : 0 < β) (hp₀ : 0 < p) (hp₁ : p ≤ 1)
(hi : ∀ x y : H, ‖gradf x - gradf y‖ ≤ β * ‖x - y‖ ^ p) :
∀ x y : H, ⟪x - y, gradf x - gradf y⟫_ℝ ≤ β * ‖x - y‖ ^ (p + 1)Theorem 18.15
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Theorem 18.15: for a continuous convex function f : H → ℝ, a positive parameter β, and h = f* - β⁻¹ q with q(x) = ‖x‖² / 2, the standard smoothness, descent, cocoercivity, convexity, and proximal formulations of β-Lipschitz differentiability are equivalent.
open scoped Gradient InnerProductSpace
universe u
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
theorem frechetDifferentiable_tfae_lipschitz_gradient
(f : H → ℝ) (hcont : Continuous f) (hconv : _root_.ConvexOn ℝ Set.univ f)
(β : Set.Ioi (0 : ℝ)) :
List.TFAE
[ HasLipschitzGradient f β,
HasGradientInnerQuadraticUpperBound f β,
HasQuadraticDescentEstimate f β,
HasConjugateGradientQuadraticLowerBound f β,
HasCocoerciveGradient f β,
HasHalfSquaredNormSubConvexity f β,
HasShiftedConjugateConvexity f β,
HasMoreauEnvelopeRepresentation f β,
HasProximalGradientRepresentation f β ]Remark 18.16
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Remark 18.16: for a convex function with β-Lipschitz gradient, the affine model at x minorizes f and the quadratic model with curvature β majorizes f.
open scoped Gradient InnerProductSpace
open InnerProductSpace
universe u
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
theorem gradient_models_sandwich_at
(f : H → ℝ) (β : Set.Ioi (0 : ℝ)) (x : H)
(hDiff : Differentiable ℝ f) (hconv : ConvexOn ℝ Set.univ f)
(hLip : LipschitzWith (Real.toNNReal (β : ℝ)) (∇ f)) :
gradientAffineModelAt f x ≤ f ∧ f ≤ gradientQuadraticModelAt f β xCorollary 18.17
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Corollary 18.17: for a Fréchet differentiable convex function on a real Hilbert space, the gradient is β-Lipschitz if and only if it is 1 / β-cocoercive.
open SubtypeFirmness
-- Declarations for this item will be appended below by the statement pipeline.
open scoped Gradient InnerProductSpace
universe u
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
-- Proof sketch: apply Theorem 18.15 to identify clause `(i)` with clause `(v)` for the TFAE list
-- attached to a continuous convex function. The hypothesis `Differentiable ℝ f` supplies the
-- differentiability component already built into both clauses, so the equivalence reduces to the
-- raw gradient statements.
theorem gradient_lipschitz_iff_cocoercive_of_differentiable_convex
(f : H → ℝ) (hdiff : Differentiable ℝ f) (hconv : _root_.ConvexOn ℝ Set.univ f)
(β : Set.Ioi (0 : ℝ)) :
LipschitzWith (Real.toNNReal (β : ℝ)) (∇ f) ↔
CocoerciveOn (1 / (β : ℝ)) (Set.univ : Set H) (fun x : Set.univ ↦ ∇ f x)Corollary 18.18
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Corollary 18.18: if a bounded operator L : H →L[ℝ] H is self-adjoint and monotone, then ‖L‖ * ⟪L x, x⟫_ℝ ≥ ‖L x‖^2 for every x.
open scoped InnerProductSpace
universe u
namespace ContinuousLinearMap
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
-- Proof sketch: set `f y = ⟪L y, y⟫_ℝ / 2`. Example 2.57 and self-adjointness identify
-- `∇ f = L`, while monotonicity gives convexity of `f`. Apply implication `(i) -> (v)` from
-- Theorem 18.15 with `β = ‖L‖`, then specialize the resulting cocoercivity inequality at `y = 0`.
theorem norm_mul_inner_apply_ge_sq_norm_of_isSelfAdjoint_of_isMonotone
(L : H →L[ℝ] H) (hL_self : IsSelfAdjoint L) (hL_mono : L.toLinearMap.IsMonotone) (x : H) :
‖L‖ * ⟪L x, x⟫_ℝ ≥ ‖L x‖ ^ (2 : ℕ)Corollary 18.19
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Corollary 18.19: for f ∈ Γ₀(H) viewed through its real-valued representative and β ∈ ℝ_{++}, f is Fréchet differentiable on H with β-Lipschitz gradient if and only if f is the Moreau envelope of parameter 1 / β of the canonical function (f^* - β⁻¹ q)^*, where q(x) = ‖x‖² / 2.
open scoped Gradient InnerProductSpace
universe u
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
-- Proof sketch: use `hf : f.toEReal ∈ Γ₀(H)` to recover the continuity and convexity
-- hypotheses needed for Theorem 18.15, then apply the equivalence between clause `(i)` and clause
-- `(viii)`.
theorem hasLipschitzGradient_iff_hasMoreauEnvelopeRepresentation_of_mem_gammaZero
(f : H → ℝ) (hf : f.toEReal ∈ Γ₀(H)) (β : Set.Ioi (0 : ℝ)) :
HasLipschitzGradient f β ↔ HasMoreauEnvelopeRepresentation f βCorollary 18.20
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Corollary 18.20: for a finite family (f_i) in Γ₀(H) and nonnegative coefficients α i with ∑ i, α i = 1, the convex combination ∑ i, α i Prox_{f_i} is the proximity operator of weightedProxGenerator f hf α, that is, of h = (∑ i, α i (f_i^* □ q))^* - q.
open scoped BigOperators InnerProductSpace
universe u v
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
variable {I : Type v} [Fintype I]
theorem weightedSum_proximityOperator_eq_proximityOperator_weightedProxGenerator
(f : I → H → Set.Ioi (⊥ : EReal)) (hf : ∀ i, f i ∈ Γ₀(H)) (α : I → ℝ)
(hα_nonneg : ∀ i, 0 ≤ α i) (hα_sum : ∑ i, α i = 1) :
(fun x : H ↦
∑ i, α i • Prox[f i, hf i] x) =
Prox[weightedProxGenerator f hf α,
weightedProxGenerator_mem_gammaZero f hf α hα_nonneg hα_sum]Remark 18.21
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Remark 18.21: for two functions f₁, f₂ ∈ Γ₀(H), the function h from (18.40) specialized to the equal weights α₁ = α₂ = 1 / 2 is exactly the proximal average pav(f₁, f₂).
open scoped InnerProductSpace
universe u
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
-- Proof sketch: specialize the Chapter 18 generator `h = (Σ i, α i (fᵢ* □ q))^* - q` to the two
-- weights `α₁ = α₂ = 1 / 2`, rewrite the sum as the average of the two unit Moreau envelopes of
-- `f₁*` and `f₂*`, and then apply the identification of that equal-weight case with the proximal
-- average from Corollary 14.8(4).
theorem equalWeight_conjugateSubInvHalfSquaredNorm_eq_proximalAverage
(f₁ f₂ : H → Set.Ioi (⊥ : EReal)) (hf₁ : f₁ ∈ Γ₀(H)) (hf₂ : f₂ ∈ Γ₀(H)) :
(fun x : H ↦
(conjugateSubInvHalfSquaredNorm
(fun y : H ↦
(1 / 2 : ℝ) *
(({}^[⟨(1 : ℝ), Set.mem_Ioi.2 zero_lt_one⟩] (gammaZeroConjugate f₁ hf₁)) y).toReal +
(1 / 2 : ℝ) *
(({}^[⟨(1 : ℝ), Set.mem_Ioi.2 zero_lt_one⟩] (gammaZeroConjugate f₂ hf₂)) y).toReal)
⟨(1 : ℝ), Set.mem_Ioi.2 zero_lt_one⟩ x : EReal)) =
pav(f₁, f₂)Proposition 18.22
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 18.22: for a closed convex subset C of a real Hilbert space and x ∈ C, the following are equivalent: the distance function to C has Gâteaux derivative 0 at x, the point x is not a support point of C, and the tangent cone T[C] x is the whole space.
open InnerProductSpace
open ERealFunction
open scoped InnerProductSpace Pointwise
universe u
namespace Set
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
theorem distanceToSet_zeroGateauxDerivative_tfae_not_mem_supportPoints_tangentCone_eq_univ
{C : Set H}
(hC_closed : IsClosed C) (hC_convex : Convex ℝ C)
{x : H} (hx : x ∈ C) :
List.TFAE
[ HasGateauxDerivativeAt
(fun y ↦ Metric.infDist y C)
(toDual ℝ H (0 : H)) x,
x ∉ spts C,
T[C] x = (univ : Set H) ]Proposition 18.23
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 18.23 (1): clause (i). At an interior point of C, the distance function to C is Fréchet differentiable with gradient 0.
open InnerProductSpace
open ERealFunction
open scoped Gradient InnerProductSpace Pointwise
open scoped Set
universe u
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
-- Proof sketch: if `x ∈ interior C`, then some open ball around `x` lies in `C`, so
-- `Metric.infDist · C` vanishes on a neighborhood of `x`. The distance function is therefore
-- locally constant at `x`, hence Fréchet differentiable there with gradient `0`.
theorem distanceToSet_hasGradientAt_zero_of_mem_interior
{C : Set H} {x : H} (hx : x ∈ interior C) :
HasGradientAt (fun y ↦ Metric.infDist y C) (0 : H) x