Chapter 20 — Monotone Operators
Proofs omitted. Hover over highlighted terms to see their Lean types.
Definition 20.1
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Definition 20.1: a set-valued operator on a real Hilbert space is monotone when any two points of its graph satisfy the inequality 0 ≤ ⟪x - y, u - v⟫_ℝ.
open scoped InnerProductSpace SetRel
universe u v w z
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
abbrev IsMonotone (A : SetValuedOperator H H) : Prop :=
A.graph.IsMonotone
-- Proof sketch: unfold `SetValuedOperator.IsMonotone`, rewrite relation membership in `A.graph`
-- via `SetValuedOperator.mem_graph`, and obtain exactly the textbook pointwise inequality.Proposition 20.2
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 20.2(ii): a set-valued operator is monotone exactly when every graph pair satisfies the accretive norm inequality ‖x - y + α • (u - v)‖ ≥ ‖x - y‖ for all α ∈ [0,1].
open scoped InnerProductSpace
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
theorem isMonotone_iff_accretive_norm_inequality (A : SetValuedOperator H H) :
A.IsMonotone ↔
∀ ⦃x u y v : H⦄, u ∈ A x → v ∈ A y →
∀ α : Set.Icc (0 : ℝ) 1, ‖x - y + (α : ℝ) • (u - v)‖ ≥ ‖x - y‖Example 20.3
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.3: for an ]-∞,+∞]-valued function with nonempty effective domain, the subdifferential is a monotone set-valued operator.
open scoped InnerProductSpace
universe u
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
-- Proof sketch: let `u ∈ (∂ f) x` and `v ∈ (∂ f) y`. Use a finite point of `f` to show that both
-- `x` and `y` lie in the effective domain whenever the corresponding subdifferentials are
-- nonempty. Then apply the subgradient inequality for `u` at `y` and for `v` at `x`, and add the
-- two inequalities to obtain `0 ≤ ⟪x - y, u - v⟫_ℝ`.
theorem subdifferential_isMonotone
(f : H → Set.Ioi (⊥ : EReal)) (hdom : (effectiveDomain f).Nonempty) :
(∂ f).IsMonotoneExample 20.4
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.4: the singleton-valued operator attached to an increasing real function on a subset of ℝ is monotone.
open scoped InnerProductSpace
namespace SetValuedOperator
-- Proof sketch: unfold `SetValuedOperator.IsMonotone`; graph membership for `ofFunction D A`
-- forces `u = A ⟨x, hx⟩` and `v = A ⟨y, hy⟩`. Then the real inner product is ordinary
-- multiplication, and the desired inequality is the canonical mathlib consequence
-- `Monovary.sub_mul_sub_nonneg` for the monotone pair `((↑) : D → ℝ)` and `A`.
theorem ofFunction_isMonotone_of_monotone (D : Set ℝ) (A : D → ℝ) (hA : Monotone A) :
(ofFunction D A).IsMonotoneExample 20.5
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.5: every cocoercive map on a subset of a real Hilbert space is monotone, i.e. for all x, y ∈ D one has 0 ≤ ⟪x - y, T x - T y⟫_ℝ.
open scoped InnerProductSpace
universe u
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
-- Proof sketch: apply the cocoercivity inequality at `x` and `y`; since the parameter `β` is
-- positive, the term `β * ‖T x - T y‖ ^ 2` is nonnegative, so the right-hand side inner product is
-- also nonnegative.
theorem CocoerciveOn.monotone {β : ℝ} {D : Set H} {T : D → H} (hT : CocoerciveOn β D T)
(x y : D) : 0 ≤ ⟪(x : H) - y, T x - T y⟫_ℝExample 20.6
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.6: if T : D → H is α-averaged with α ≤ 1 / 2, then T is monotone, i.e. for all x, y ∈ D one has 0 ≤ ⟪x - y, T x - T y⟫_ℝ.
open SubtypeFirmness
-- Declarations for this item will be appended below by the statement pipeline.
open scoped InnerProductSpace
universe u
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
-- Proof sketch: apply the firm nonexpansiveness inequality
-- `‖T x - T y‖ ^ 2 ≤ ⟪x - y, T x - T y⟫_ℝ` and use that the squared norm is nonnegative.
theorem averagedWith_monotone_of_le_half {D : Set H} {α : ℝ} {T : D → H}
(hT : AveragedWith α T) (hα : α ≤ (1 / 2 : ℝ)) (x y : D) :
0 ≤ ⟪(x : H) - y, T x - T y⟫_ℝExample 20.7
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.7: if T : D → H is nonexpansive and α ∈ [-1,1], then the singleton-valued operator associated with x ↦ x + α • T x is monotone.
open scoped InnerProductSpace
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
-- Proof sketch: apply Proposition 4.11 to the scaled map `x ↦ -α • T x`. Since `|α| ≤ 1`, that
-- scaled map is still nonexpansive, so its residual map is `1 / 2`-cocoercive. But that residual
-- map is exactly `x ↦ x + α • T x`, and Example 20.5 converts cocoercivity of a single-valued map
-- into monotonicity of the associated singleton-valued operator.
theorem ofFunction_id_add_smul_isMonotone_of_nonexpansive
{D : Set H} (T : D → H) (hT : LipschitzWith 1 T) {α : ℝ}
(hα : α ∈ Set.Icc (-1 : ℝ) 1) :
(ofFunction D (fun x : D ↦ (x : H) + α • T x)).IsMonotoneExample 20.8
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.8: a map T : D → H is pseudononexpansive if and only if its residual map A = Id - T defines a monotone singleton-valued operator on D.
open scoped InnerProductSpace
universe u
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
theorem pseudononexpansiveOn_iff_residualMap_isMonotone {D : Set H} {T : D → H} :
PseudononexpansiveOn D T ↔
(SetValuedOperator.ofFunction D (residualMap D T)).IsMonotoneExample 20.9
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.9: the time-derivative operator sends x to the singleton {x'} on the source domain D, where x' is the canonical derivative class, and to ∅ outside D.
open MeasureTheory
open scoped MeasureTheory InnerProductSpace
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [NormedSpace ℝ H]
variable (T : Set.Ioi (0 : ℝ))
local notation "IccT" => Set.Icc (0 : ℝ) (T : ℝ)
local notation "L2T" => MeasureTheory.Lp H 2 (volume : Measure IccT)
local notation "W12T" => SobolevW12 H T
def timeDerivativeOperator (bc : TimeDerivativeBoundaryCondition H) : SetValuedOperator L2T L2T :=
SetValuedOperator.ofFunction (timeDerivativeDomain T bc) (timeDerivative T bc)Proposition 20.10
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 20.10 (1): the inverse of a monotone set-valued operator on a real Hilbert space is monotone.
open scoped Pointwise SetValuedOperator
universe u v
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
-- Proof sketch: rewrite membership in `A⁻¹` with `SetValuedOperator.mem_inverse_iff`. The
-- monotonicity inequality for `A⁻¹` is then exactly the monotonicity inequality for `A`
-- with the graph coordinates exchanged.
theorem IsMonotone.inverse {A : SetValuedOperator H H} (hA : A.IsMonotone) :
A⁻¹.IsMonotoneProposition 20.11
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 20.11: if 𝒜 is monotone, then the operator on L²((Ω,\mathcal F,\mu); H) whose graph is given almost everywhere by gra 𝒜 is monotone.
open MeasureTheory
open MeasureTheory.L2
open scoped MeasureTheory InnerProductSpace
universe u v
namespace SetValuedOperator
variable {Ω : Type u} [MeasurableSpace Ω] {μ : Measure Ω}
variable {H : Type v} [NormedAddCommGroup H]
variable [InnerProductSpace ℝ H]
-- Proof sketch: if `(x, u)` and `(y, v)` belong to the graph of `aeL2Lift 𝒜`, then
-- `h𝒜` gives `0 ≤ ⟪x ω - y ω, u ω - v ω⟫_ℝ` almost everywhere. Example 2.6 identifies the
-- `L²` inner product with the integral of the pointwise inner product, so integrating the
-- almost-everywhere nonnegative function yields `0 ≤ ⟪x - y, u - v⟫_ℝ`.
theorem IsMonotone.aeL2Lift {𝒜 : SetValuedOperator H H} (h𝒜 : 𝒜.IsMonotone)
(μ : Measure Ω) :
(𝒜.aeL2Lift μ).IsMonotoneExample 20.12
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.12: the set-valued projector P[C] onto a subset of a real Hilbert space is monotone.
universe u
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
-- Proof sketch: let `u ∈ P[C] x` and `v ∈ P[C] y`. Unfold
-- `mem_setValuedProjector_iff` to obtain
-- `‖x - u‖ = Metric.infDist x C ≤ ‖x - v‖` and `‖y - v‖ = Metric.infDist y C ≤ ‖y - u‖`, then
-- apply the square-norm characterization of monotonicity from Proposition 20.2 and simplify.
theorem setValuedProjector_isMonotone (C : Set H) :
(P[C]).IsMonotoneExample 20.13
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.13: the negation of the chapter's active farthest-point operator Φ[C] is monotone.
open scoped InnerProductSpace Pointwise
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H]
variable [InnerProductSpace ℝ H]
-- Proof sketch: use the Chapter 20 owner `SetValuedOperator.IsMonotone` together with the
-- canonical four-point reformulation from Proposition 20.2. If `u ∈ (-Φ[C]) x` and
-- `v ∈ (-Φ[C]) y`, then `-u` and `-v` are active farthest points in `C`; comparing the
-- maximizing property of `-u` against `-v` at `x`, and of `-v` against `-u` at `y`, gives the
-- two squared-distance inequalities whose sum is exactly the required four-point inequality.
theorem neg_chebyshevCenterActiveSet_isMonotone (C : Set H) :
(-Φ[C]).IsMonotoneProposition 20.14
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 20.14: for a set-valued operator A, the following are equivalent: (i) A is monotone; (ii) the pairing F(x, u) = ⟪x, u⟫_ℝ satisfies the finite Jensen inequality on graph points of A; (iii) F is convex along two-point combinations of graph points of A.
open scoped BigOperators InnerProductSpace
universe u v
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
-- `source-facing`: Proposition 20.14 is the textbook equivalence between monotonicity, the finite
-- Jensen inequality on `gra A`, and the corresponding two-point convexity statement on `gra A`.
-- `core/canonical`: Lemma 2.14 is the owner identity turning the Jensen defect of
-- `p ↦ ⟪p.1, p.2⟫_ℝ` into a weighted sum of monotonicity pairings.
-- `bridge/view`: a `ConvexOn` statement on `convexHull ℝ (gra A)` is a useful ambient reformulation
-- via `ConvexOn.map_sum_le`, but it is companion-only and not the main source-facing theorem.
-- Proof sketch: Lemma 2.14 rewrites the finite Jensen defect of the pairing on graph points as a
-- half-weighted sum of terms `⟪x i - x j, u i - u j⟫_ℝ`. Clause `(i)` makes each summand
-- nonnegative, yielding clause `(ii)`. The two-point statement in clause `(iii)` is the special
-- case of clause `(ii)`, and expanding its defect gives
-- `-α * (1 - α) * ⟪x - y, u - v⟫_ℝ`, recovering monotonicity.
theorem tfae_isMonotone_graph_pairing_convexity (A : SetValuedOperator H H) :
List.TFAE
[A.IsMonotone,
(∀ {ι : Type v} (s : Finset ι) (α : ι → ℝ) (p : ι → gra A),
(∀ i ∈ s, 0 ≤ α i) →
(∑ i ∈ s, α i) = 1 →
⟪(∑ i ∈ s, α i • (p i : H × H)).1, (∑ i ∈ s, α i • (p i : H × H)).2⟫_ℝ ≤
∑ i ∈ s, α i * ⟪(p i : H × H).1, (p i : H × H).2⟫_ℝ),
(∀ ⦃p q : gra A⦄ ⦃α : ℝ⦄, 0 < α → α < 1 →
⟪(α • (p : H × H) + (1 - α) • (q : H × H)).1,
(α • (p : H × H) + (1 - α) • (q : H × H)).2⟫_ℝ ≤
α * ⟪(p : H × H).1, (p : H × H).2⟫_ℝ +
(1 - α) * ⟪(q : H × H).1, (q : H × H).2⟫_ℝ)]Example 20.15
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.15: when a monotone operator is given by a linear map T, the monotonicity of the associated singleton-valued operator from Definition 20.1 is exactly the linear monotonicity condition of Definition 2.23 (1).
open scoped InnerProductSpace
universe u
namespace LinearMap
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
-- Proof sketch: use the canonical owner `SetValuedOperator.ofFunction_isMonotone_iff` in its
-- `Set.univ` specialization to rewrite monotonicity as the pairwise inequality
-- `0 ≤ ⟪x - y, T x - T y⟫_ℝ`, then use linearity to rewrite `T x - T y = T (x - y)` and
-- `real_inner_comm` to identify this with the quadratic-form condition from Definition 2.23 (1).
theorem toSetValuedOperator_isMonotone_iff (T : H →ₗ[ℝ] H) :
(T : H → H).toSetValuedOperator.IsMonotone ↔ T.IsMonotoneExample 20.16
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.16 (1): clause (i). A bounded operator is monotone if and only if A + A† is monotone.
open scoped InnerProduct InnerProductSpace
universe u
namespace ContinuousLinearMap
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
variable [CompleteSpace H]
-- Proof sketch: compare the quadratic forms of `A` and `A + A†`; in a real Hilbert space,
-- `⟪(A + A†) x, x⟫_ℝ = 2 * ⟪A x, x⟫_ℝ`, so the two monotonicity conditions are equivalent.
theorem isMonotone_iff_add_adjoint_isMonotone (A : H →L[ℝ] H) :
A.toLinearMap.IsMonotone ↔ (A + A†).toLinearMap.IsMonotoneProposition 20.17
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 20.17: if a bounded linear operator on a real Hilbert space is monotone, then its kernel agrees with the kernel of its adjoint, and the closures of the ranges of the operator and of its adjoint coincide.
universe u
open scoped InnerProduct InnerProductSpace
namespace ContinuousLinearMap
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
theorem ker_eq_adjoint_ker_and_closure_range_eq_closure_adjoint_range_of_isMonotone
(A : H →L[ℝ] H) (hA : A.toLinearMap.IsMonotone) :
A.ker = A†.ker ∧
A.range.topologicalClosure = A†.range.topologicalClosureFact 20.18
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Fact 20.18: if A and B are self-adjoint monotone bounded linear operators on a real Hilbert space and commute, then the product AB is monotone.
open scoped InnerProductSpace
universe u v
namespace ContinuousLinearMap
variable {𝕜 : Type u} {E : Type v} [RCLike 𝕜] [NormedAddCommGroup E] [InnerProductSpace 𝕜 E]
variable {H : Type v} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
-- Proof sketch: turn the textbook hypotheses into positivity via
-- `ContinuousLinearMap.isPositive_iff'`, apply the owner-level positivity lemma
-- `ContinuousLinearMap.IsPositive.mul_of_commute`, and then forget positivity back to monotonicity.
theorem isMonotone_mul_of_isSelfAdjoint_of_commute
{A B : H →L[ℝ] H}
(hA_self : IsSelfAdjoint A) (hB_self : IsSelfAdjoint B)
(hA_mono : A.toLinearMap.IsMonotone) (hB_mono : B.toLinearMap.IsMonotone)
(hcomm : Commute A B) :
(A * B).toLinearMap.IsMonotoneExample 20.19
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.19 (1): the operator A = [[1, 1], [1, 1]] is monotone on ℝ².
open ContinuousLinearMap
open Matrix
open scoped EuclideanSpace InnerProductSpace
local notation "ℝ²" => EuclideanSpace ℝ (Fin 2)
local notation "toEuclideanOperator" =>
((Matrix.toEuclideanCLM : Matrix (Fin 2) (Fin 2) ℝ ≃⋆ₐ[ℝ] (ℝ² →L[ℝ] ℝ²)))
theorem allOnesOperator_isMonotone :
allOnesOperator.toLinearMap.IsMonotoneDefinition 20.20
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Definition 20.20: a set-valued operator is maximally monotone exactly when it is maximal, for the canonical pointwise order on set-valued operators, among monotone extensions; this is the specialization Maximal SetValuedOperator.IsMonotone A of the order-theoretic owner Maximal.
open scoped InnerProductSpace
universe u
recall MaximalTheorem 20.21
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Theorem 20.21: every monotone set-valued operator on a real Hilbert space admits an extension Amax with A ≤ Amax and Maximal IsMonotone Amax; equivalently, gra A ⊆ gra Amax and Amax is maximally monotone.
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
theorem exists_isMaximallyMonotone_extension
(A : SetValuedOperator H H) (hA : A.IsMonotone) :
∃ Amax ≥ A, Maximal IsMonotone AmaxProposition 20.22
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 20.22 (1): the inverse of a maximally monotone set-valued operator is maximally monotone.
open scoped Pointwise SetValuedOperator
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
-- Proof sketch: swap the coordinates of the graph. This transports monotonicity to `A⁻¹`,
-- and graph inclusion between monotone extensions is preserved by taking inverses, so maximality
-- transfers as well.
theorem Maximal.inverse {A : SetValuedOperator H H}
(hA : Maximal IsMonotone A) :
Maximal IsMonotone A⁻¹Proposition 20.23
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 20.23: if A and B are maximally monotone, then the componentwise product operator (x, y) ↦ A x × B y is maximally monotone on the Chapter 9 ℓ² Hilbert product structure on the raw pair type H × K.
open scoped InnerProductSpace SetValuedOperator
universe u v
namespace SetValuedOperator
variable {H : Type u} {K : Type v}
variable [NormedAddCommGroup H] [InnerProductSpace ℝ H]
variable [NormedAddCommGroup K] [InnerProductSpace ℝ K]
theorem Maximal.prod
{A : SetValuedOperator H H} {B : SetValuedOperator K K}
(hA : Maximal IsMonotone A) (hB : Maximal IsMonotone B) :
Maximal IsMonotone (A × B)Proposition 20.24
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 20.24: if A is maximally monotone and U is self-adjoint and strongly monotone, then the operator U A = ((U : H → H).toSetValuedOperator.comp A) is maximally monotone on the renormed Hilbert space ContinuousLinearMap.Renormed U hU_self hU_strong whose inner product is (x, y) ↦ ⟪U⁻¹ x, y⟫_ℝ. The companion theorem isMaximallyMonotone_renormed_iff rewrites this canonical maximal-monotonicity statement back into the source pairing formula.
open scoped InnerProductSpace
universe u
namespace SetValuedOperator
open ContinuousLinearMap
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
theorem comp_isMaximallyMonotone_of_isSelfAdjoint_of_isStronglyMonotone
(A : SetValuedOperator H H) (U : H →L[ℝ] H) (α : ℝ)
(hA : Maximal IsMonotone A) (hU_self : IsSelfAdjoint U)
(hU_strong : U.toLinearMap.IsStronglyMonotone α) :
Maximal IsMonotone
(((U : H → H).toSetValuedOperator.comp A).renormed U hU_self hU_strong)Theorem 20.25
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Theorem 20.25: (Moreau) if f ∈ Γ₀(H), then the subdifferential ∂ f is maximally monotone.
open scoped InnerProductSpace
universe u
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
-- Proof sketch: use the graph criterion in Definition 20.20. If `(x, u)` is monotonically related
-- to every point of `gra ∂ f`, then Moreau's proximal construction at `x + u` produces
-- `p = Prox_f (x + u)` with `x + u - p ∈ (∂ f) p`; testing the monotonicity relation against
-- `(p, x + u - p)` forces `x = p` and hence `u ∈ (∂ f) x`.
theorem subdifferential_isMaximallyMonotone_of_mem_gammaZero
{f : H → Set.Ioi (⊥ : EReal)} (hf : f ∈ Γ₀(H)) :
Maximal SetValuedOperator.IsMonotone (∂ f)Example 20.26
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.26: for a nonempty closed convex subset C of a real Hilbert space, the normal cone operator N[C] is maximally monotone.
open scoped InnerProductSpace Pointwise
open ERealFunction
universe u
namespace Set
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
-- Proof sketch: show that the indicator `ι_C` belongs to `Γ₀(H)` for a nonempty closed convex set
-- `C`, apply Moreau's theorem `subdifferential_isMaximallyMonotone_of_mem_gammaZero` to `ι_C`, and
-- then rewrite the resulting operator with `subdifferential_setIndicator_eq_normalCone`.
theorem normalCone_isMaximallyMonotone_of_nonempty_isClosed_convex
{C : Set H} (hC_nonempty : C.Nonempty) (hC_closed : IsClosed C) (hC_convex : Convex ℝ C) :
Maximal SetValuedOperator.IsMonotone (N[C])Proposition 20.27
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 20.27: a monotone hemicontinuous single-valued operator on a real Hilbert space is maximally monotone when viewed as its associated singleton-valued set-valued operator.
open scoped InnerProductSpace
universe u
namespace Function
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
theorem toSetValuedOperator_isMaximallyMonotone_of_monotone_hemicontinuous
(A : H → H) (hA_mono : A.toSetValuedOperator.IsMonotone) (hA_hemi : A.IsHemicontinuous) :
Maximal SetValuedOperator.IsMonotone A.toSetValuedOperatorCorollary 20.28
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Corollary 20.28: a monotone continuous single-valued operator on a real Hilbert space is maximally monotone when regarded as its associated singleton-valued set-valued operator.
open scoped InnerProductSpace
universe u
namespace Function
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
-- Proof sketch: continuity of `A` implies continuity of each affine slice
-- `α ↦ A (x + α • y)`, and composing with the continuous inner-product pairing against `z`
-- yields the required right-continuity on `Set.Ioi 0`.
theorem toSetValuedOperator_isMaximallyMonotone_of_monotone_continuous
(A : H → H) (hA_mono : A.toSetValuedOperator.IsMonotone) (hA_cont : Continuous A) :
Maximal SetValuedOperator.IsMonotone A.toSetValuedOperatorExample 20.29
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.29: if T : H → H is nonexpansive and α ∈ [-1, 1], then the affine perturbation Id + α T, viewed as a singleton-valued set-valued operator, is maximally monotone.
universe u
namespace Function
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
-- Proof sketch: specialize the domain-level owner from Example 20.7 to `Set.univ`, rewriting it
-- through the canonical whole-space owner `Function.toSetValuedOperator`. A `1`-Lipschitz map is
-- continuous, so Corollary 20.28 upgrades monotonicity to maximal monotonicity.
theorem id_add_smul_toSetValuedOperator_isMaximallyMonotone_of_nonexpansive
(T : H → H) (hT : LipschitzWith 1 T) {α : ℝ} (hα : α ∈ Set.Icc (-1 : ℝ) 1) :
Maximal SetValuedOperator.IsMonotone (id + α • T).toSetValuedOperatorExample 20.30
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.30: if T : H → H is α-averaged with α ≤ 1 / 2, then the associated singleton-valued set-valued operator of T is maximally monotone.
open scoped InnerProductSpace
universe u
namespace Function
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
-- Proof sketch: an `α`-averaged self-map with `α ≤ 1 / 2` gives a monotone singleton-valued
-- operator by the canonical owner theorem from Example 20.6, and every averaged map is
-- `1`-Lipschitz, hence continuous. Apply Corollary 20.28 to the associated singleton-valued
-- operator.
theorem toSetValuedOperator_isMaximallyMonotone_of_averaged_le_half
{α : ℝ} {T : H → H}
(hT : Averaged α T) (hα : α ≤ (1 / 2 : ℝ)) :
Maximal SetValuedOperator.IsMonotone T.toSetValuedOperatorExample 20.31
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.31: if T : H → H is β-cocoercive with β ∈ ℝ_{++}, then the associated singleton-valued set-valued operator of T is maximally monotone.
universe u
open scoped Pointwise SetValuedOperator
namespace Function
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
-- Proof sketch: `β`-cocoercivity gives monotonicity by Example 20.5 and Lipschitz continuity by
-- Remark 4.15.1. Apply the canonical owner theorem Corollary 20.28 for continuous monotone
-- singleton-valued operators.
theorem toSetValuedOperator_isMaximallyMonotone_of_cocoerciveOn_univ
{β : ℝ} {T : H → H}
(hT : CocoerciveOn β (Set.univ : Set H) (fun x : (Set.univ : Set H) ↦ T x)) :
Maximal SetValuedOperator.IsMonotone T.toSetValuedOperatorExample 20.32
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.32: the metric projection onto a nonempty closed convex subset of a real Hilbert space is maximally monotone when viewed as its associated singleton-valued set-valued operator.
open scoped InnerProductSpace
universe u
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
variable {C : Set H} (hC_nonempty : C.Nonempty) (hC_closed : IsClosed C) (hC_convex : Convex ℝ C)
local notation "hC" =>
isChebyshev_of_nonempty_isClosed_convex hC_nonempty hC_closed hC_convex
-- Proof sketch: Proposition 4.16 shows that the metric projection `P[C, hC]` onto a nonempty closed
-- convex set is firmly nonexpansive. By Remark 4.34, this is equivalent to `1 / 2`-averagedness,
-- and Example 20.30 then yields maximal monotonicity of the associated singleton-valued operator.
theorem projectionPoint_toSetValuedOperator_isMaximallyMonotone_of_nonempty_isClosed_convex :
Maximal SetValuedOperator.IsMonotone (P[C, hC]).toSetValuedOperatorExample 20.33
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.33: in a finite-dimensional real Hilbert space, the set-valued projector P[C] onto a Chebyshev set is maximally monotone.
universe u
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [FiniteDimensional ℝ H]
variable {C : Set H}
-- Proof sketch: Example 20.12 gives monotonicity of the source-facing projector `P[C]`. The
-- owner theorem `setValuedProjector_eq_singleton_projectionPoint` identifies each value of `P[C]`
-- with the singleton induced by `P[C, hC]`, hence with `(P[C, hC]).toSetValuedOperator`; the
-- continuity of `P[C, hC]` comes from Proposition 3.12, and Corollary 20.28 yields maximal
-- monotonicity.
theorem setValuedProjector_isMaximallyMonotone_of_isChebyshev (hC : IsChebyshev C) :
Maximal SetValuedOperator.IsMonotone (P[C])Example 20.34
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.34: a monotone bounded linear operator on a real inner product space is maximally monotone when viewed as its associated singleton-valued set-valued operator.
universe u
namespace ContinuousLinearMap
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
-- Proof sketch: use the canonical linear-map bridge
-- `LinearMap.toSetValuedOperator_isMonotone_iff` to turn `A.toLinearMap.IsMonotone` into
-- monotonicity of the singleton-valued operator associated with `A`. Since `A` is continuous as
-- a bounded linear operator, Corollary 20.28 applies directly.
theorem toSetValuedOperator_isMaximallyMonotone_of_isMonotone
(A : H →L[ℝ] H) (hA : A.toLinearMap.IsMonotone) :
Maximal SetValuedOperator.IsMonotone A.toSetValuedOperatorExample 20.35
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.35: if A.adjoint = -A, then A, viewed as its associated singleton-valued set-valued operator, is maximally monotone.
universe u
namespace ContinuousLinearMap
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
-- Proof sketch: the owner-level bridge `isMonotone_of_adjoint_eq_neg` turns the skew-adjoint
-- hypothesis into monotonicity of `A`. Example 20.34 then upgrades that monotonicity to maximal
-- monotonicity for the associated singleton-valued set-valued operator.
theorem toSetValuedOperator_isMaximallyMonotone_of_adjoint_eq_neg
(A : H →L[ℝ] H) (hA : A.adjoint = -A) :
Maximal SetValuedOperator.IsMonotone A.toSetValuedOperatorProposition 20.36
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 20.36 (1): for a maximally monotone operator on a real Hilbert space, every value set A x is closed.
open scoped InnerProductSpace
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
theorem Maximal.value_isClosed
{A : SetValuedOperator H H} (hA : Maximal IsMonotone A) (x : H) :
IsClosed (A x)Proposition 20.37
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 20.37 (1): if a bounded net of graph points of a maximally monotone operator converges strongly in the primal variable and weakly in the dual variable, then its limit pair still belongs to the graph.
open Filter
open scoped InnerProductSpace SetValuedOperator Topology
universe u v
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
variable {B : Type v} [Nonempty B] [Preorder B] [IsDirectedOrder B]
theorem Maximal.mem_graph_of_tendsto_of_tendsto_weakly
{A : SetValuedOperator H H} (hA : Maximal IsMonotone A)
{x_b u_b : B → H} {x u : H}
(hgraph : ∀ b, (x_b b, u_b b) ∈ gra A)
(hbounded : Bornology.IsBounded (Set.range fun b ↦ (x_b b, u_b b)))
(hx_b : Tendsto x_b atTop (𝓝 x))
(hu_b : Tendsto (fun b ↦ toWeakSpace ℝ H (u_b b)) atTop (𝓝 (toWeakSpace ℝ H u))) :
(x, u) ∈ gra AProposition 20.38
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 20.38 (1): the graph of a maximally monotone operator is sequentially closed in the mixed topology with strong convergence in the primal variable and weak convergence in the dual variable, encoded as sequential closedness of its image in H × WeakSpace ℝ H.
open Filter
open scoped InnerProductSpace SetValuedOperator Topology
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
variable [CompleteSpace H]
theorem Maximal.graph_isSeqClosed_strongWeak
{A : SetValuedOperator H H} (hA : Maximal IsMonotone A) :
IsSeqClosed ((Prod.map id (toWeakSpace ℝ H)) '' gra A)Example 20.39
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.39: for the closed unit ball of ℓ²(ℕ, ℝ), subtracting the canonical metric projection onto that ball yields a maximally monotone operator whose graph is not sequentially closed in the product weak topology.
open SubtypeFirmness
open Filter
open scoped InnerProductSpace Topology
local notation "L2" => lp (fun _ : ℕ ↦ ℝ) 2
namespace Function
local notation "C" => Metric.closedBall (0 : L2) 1
local notation "hC" =>
isChebyshev_of_nonempty_isClosed_convex
closedUnitBall_nonempty
(Metric.isClosed_closedBall : IsClosed C)
(convex_closedBall (0 : L2) 1)
local notation "W" => Prod.map (toWeakSpace ℝ L2) (toWeakSpace ℝ L2)
-- The standard unit vector `e_n` in `ℓ²(ℕ, ℝ)`.
theorem l2ClosedUnitBallResidualMap_maximallyMonotone_and_graph_not_seqClosed_weakProduct :
Maximal SetValuedOperator.IsMonotone l2ClosedUnitBallResidualMap.toSetValuedOperator ∧
¬ IsSeqClosed
(W '' l2ClosedUnitBallResidualMap.toSetValuedOperator.graph)Proposition 20.40
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 20.40 (1): if A = ofFunction D T is monotone and T is symmetric on D, then the supremal potential f satisfies f + ι_D = h.
open ERealFunction
open scoped InnerProductSpace
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
-- Proof sketch: evaluate the `y = 0` term in the defining supremum; since `0 : D`, the supremum
-- is bounded below by the real value `0`, hence it is strictly above `⊥`.
variable (D : Submodule ℝ H) (T : D →ₗ[ℝ] H)
-- Proof sketch: identify `A` with the graph of `T` on `D`, rewrite monotonicity of `A` as the
-- quadratic inequality for `T`, and compare the supremum defining `f` with the value at `y = x`
-- and with the monotonicity bound obtained from `0 ≤ ⟪x - y, T x - T y⟫`.
theorem supremalPotential_add_indicator_eq_domainQuadraticPotential
(hsymm : ∀ x y : D, ⟪(x : H), T y⟫_ℝ = ⟪T x, (y : H)⟫_ℝ)
(hT_mono : (ofFunction (D : Set H) T).IsMonotone) :
(supremalPotential D T).asEReal + (ι[(D : Set H)]).asEReal =
(domainQuadraticPotential D T).asERealExample 20.41
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.41 states dom A = ran B for the inverse operator A = B⁻¹.
open ERealFunction
open Filter
open scoped InnerProductSpace SetValuedOperator Topology
universe u
namespace HilbertBasis
local notation "L2" => lp (fun _ : ℕ ↦ ℝ) 2
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
-- Proof sketch: use `0 ≤ β n ≤ 1` to bound `‖β n * x n‖^2` by `‖x n‖^2`, then apply the
-- summability characterization of `Memℓp` for `p = 2`.
theorem diagonalInverseOperator_dom_eq_range_diagonalForwardOperator
(b : HilbertBasis ℕ ℝ H) (β : ℕ → NNRealˣ)
(hβ_le_one : ∀ n, (β n : NNReal) ≤ 1) :
(diagonalInverseOperator b β hβ_le_one).dom =
(diagonalForwardOperator b β hβ_le_one).toLinearMap.rangeDefinition 20.42
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Definition 20.42: for a set-valued operator A and a closed linear subspace V, represented in the projection API by the primitive assumption [V.HasOrthogonalProjection], the partial inverse of A with respect to V is the operator whose graph is obtained from gra A by Spingarn's graph transform. The source-facing Lean notation for the textbook partial inverse A_V is A₍V₎.
open scoped InnerProductSpace
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
def partialInverse (A : SetValuedOperator H H) (V : Submodule ℝ H) [V.HasOrthogonalProjection] :
SetValuedOperator H H :=
fun x ↦ { y | partialInverseTransform V (x, y) ∈ gra A }Example 20.43
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.43 (1): the pointwise formula for the partial inverse A₍V₎ is exactly the owner theorem partialInverse_apply from Definition 20.42.
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
open scoped SetValuedOperator
recall partialInverse_applyProposition 20.44
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 20.44 (1): clause (i). The partial inverse with respect to V is the inverse of the partial inverse with respect to Vᗮ.
open scoped InnerProductSpace SetValuedOperator
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
-- Proof sketch: unfold the partial inverse on `V` and on `Vᗮ`, rewrite graph membership through
-- the same transformed pair, and swap the graph coordinates to identify the result with the
-- inverse graph.
theorem partialInverse_eq_inverse_partialInverse_orthogonal
(A : SetValuedOperator H H) (V : Submodule ℝ H) [V.HasOrthogonalProjection] :
A₍V₎ = A₍Vᗮ₎⁻¹Lemma 20.45
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Lemma 20.45: the quadratic kernel G centered at (z, w) has Fenchel conjugate G* = G ∘ L, where L(u, x) = (-x, -u).
open scoped InnerProductSpace
universe u
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
theorem bivariateQuadraticKernel_conjugate_eq_comp_skewSwapNeg (z w : H) :
((bivariateQuadraticKernel z w).toEReal.asEReal)∗ =
((bivariateQuadraticKernel z w) ∘ skewSwapNeg).toEReal.asERealTheorem 20.46
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Theorem 20.46 (1): if F*(u, x) ≥ ⟪x, u⟫ for all x, u, then the pairing-contact operator of the transpose-conjugate (F∗)ᵀ is monotone; equivalently, the equality locus F*(u, x) = ⟪x, u⟫ defines a monotone operator.
open scoped InnerProductSpace
open ERealFunction
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
theorem pairingEqualityOperator_conjugateTranspose_isMonotone
(F : H × H → EReal)
(hFstar_ge : ∀ x u : H, ((⟪x, u⟫_ℝ : ℝ) : EReal) ≤ F∗ (u, x)) :
(pairingEqualityOperator ((F∗)ᵀ)).IsMonotoneCorollary 20.47
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Corollary 20.47: on a real Hilbert space, if F ∈ Γ₀(H × H) is autoconjugate, then the operator defined by gra A = {(x, u) | F (x, u) = ⟪x, u⟫} is maximally monotone.
open scoped InnerProductSpace
open ERealFunction
open WithLp
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
theorem pairingEqualityOperator_isMaximallyMonotone_of_mem_gammaZero_of_autoconjugate
(F : H × H → Set.Ioi (⊥ : EReal)) (hF : F ∈ Γ₀(H × H))
(hauto : autoconjugate F.asEReal) :
Maximal IsMonotone (pairingEqualityOperator F)Theorem 20.48
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Theorem 20.48: the autoconjugate proof route recovers exactly the canonical owner theorem from Theorem 20.25 stating that, for f ∈ Γ₀(ℋ), the subdifferential ∂ f is maximally monotone.
namespace ERealFunction
recall subdifferential_isMaximallyMonotone_of_mem_gammaZeroProposition 20.49
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 20.49: if C is a nonempty closed convex subset of a real Hilbert space and A is monotone and hemicontinuous on C in Rockafellar's sense, then the sum of the restricted singleton-valued operator ofFunction C (fun x ↦ A x) with the normal cone operator N[C] is maximally monotone.
open scoped InnerProductSpace Pointwise Set
universe u
namespace Function
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
variable [CompleteSpace H]
-- Proof sketch: use the maximal-monotonicity criterion from Definition 20.20. A graph point
-- monotonically related to `gra (ofFunction C A + N[C])` first lies over `C` by maximal
-- monotonicity of `N[C]` (Example 20.26); then apply the monotonicity hypothesis on `C` and let
-- the segment parameter tend to `0` via hemicontinuity to show that the residual vector belongs to
-- `N[C]`, hence the point already lies in the graph of the sum.
theorem ofFunction_add_normalCone_isMaximallyMonotone_of_monotoneOn_hemicontinuousOn
{C : Set H} (hC_nonempty : C.Nonempty) (hC_closed : IsClosed C) (hC_convex : Convex ℝ C)
(A : H → H) (hA_mono : (SetValuedOperator.ofFunction C (fun x : C ↦ A x)).IsMonotone)
(hA_hemi : A.IsHemicontinuousOn C) :
Maximal SetValuedOperator.IsMonotone
(SetValuedOperator.ofFunction C (fun x : C ↦ A x) + N[C])Corollary 20.50
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Corollary 20.50: if C is a nonempty closed convex subset of a real Hilbert space, if A is a set-valued operator, and if A coincides on C with a continuous single-valued map whose restricted singleton-valued operator is monotone, then A + N[C] is maximally monotone.
open scoped InnerProductSpace Pointwise Set
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
-- Proof sketch: the singleton hypothesis on `C` identifies `A + N[C]` with the canonical owner
-- `ofFunction C T + N[C]`, since the normal cone is empty off `C`. Extend the continuous subtype
-- map `T : C → H` to an ambient map that agrees with it on `C`, derive Rockafellar
-- hemicontinuity on `C`, apply Proposition 20.49 to the restricted singleton operator, and then
-- rewrite the result back to `A + N[C]`.
theorem add_normalCone_isMaximallyMonotone_of_monotone_of_eq_singleton_continuous
{C : Set H} (hC_nonempty : C.Nonempty) (hC_closed : IsClosed C) (hC_convex : Convex ℝ C)
(A : SetValuedOperator H H) (T : C → H) (hT_mono : (ofFunction C T).IsMonotone)
(hA_eq : ∀ x : C, A x = ({T x} : Set H)) (hT_cont : Continuous T) :
Maximal IsMonotone (A + N[C])Definition 20.51
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Definition 20.51: the Fitzpatrick function attached to a set-valued operator A; for monotone A, this is the textbook Fitzpatrick function. It is the extended-real-valued function on H × H obtained by taking the supremum of ⟪y, u⟫ + ⟪x, v⟫ - ⟪y, v⟫ over graph points (y, v) ∈ gra A. The Lean surface notation for the textbook symbol F_A is F[A].
open scoped InnerProductSpace
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
noncomputable def fitzpatrickFunction (A : SetValuedOperator H H) : H × H → EReal :=
fun (x, u) ↦
⨆ p : gra A, ((⟪p.1.1, u⟫_ℝ + ⟪x, p.1.2⟫_ℝ - ⟪p.1.1, p.1.2⟫_ℝ : ℝ) : EReal)Example 20.52
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.52: for the identity singleton-valued operator, the Fitzpatrick function is the textbook quadratic map (x, u) ↦ (1 / 4) * ‖x + u‖ ^ 2.
open scoped InnerProductSpace SetValuedOperator
universe u
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
theorem fitzpatrickFunction_id_apply (x u : H) :
F[id.toSetValuedOperator] (x, u) = ((1 / 4 : ℝ) * ‖x + u‖ ^ 2 : EReal)Example 20.53
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.53: if A.adjoint = -A, then the Fitzpatrick function of the singleton-valued operator induced by A is the extended-real indicator of its graph.
open scoped InnerProductSpace SetValuedOperator
universe u
namespace ContinuousLinearMap
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
-- Proof sketch: expand the Fitzpatrick supremum for the singleton-valued operator induced by `A`.
-- At a graph point `(y, A y)`, the skew-adjoint hypothesis forces `⟪y, A y⟫_ℝ = 0`, so the
-- supremand becomes `⟪y, u - A x⟫_ℝ`. If `u = A x`, every term is `0`; if `u ≠ A x`, choose `y`
-- along `u - A x` to make the supremum equal `⊤`. This is exactly the graph indicator.
theorem fitzpatrickFunction_eq_indicator_graph_of_adjoint_eq_neg
(A : H →L[ℝ] H) (hA : A.adjoint = -A) :
F[A.toSetValuedOperator] =
(ι[gra A.toSetValuedOperator]).asERealExample 20.54
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.54: for a bounded linear operator A, the Fitzpatrick function of the singleton-valued operator induced by A is 2 q_A^* ((u + A† x) / 2).
open scoped ContinuousLinearMap ERealFunction InnerProduct InnerProductSpace SetValuedOperator
universe u
namespace ContinuousLinearMap
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
variable [CompleteSpace H]
theorem fitzpatrickFunction_eq_two_mul_conjugate_quadraticPotential
(A : H →L[ℝ] H) (x u : H) :
F[A.toSetValuedOperator] (x, u) =
((2 : ℝ) : EReal) * ((q[A]).toEReal.asEReal∗)
(((1 / 2 : ℝ) • u) + ((1 / 2 : ℝ) • (A†) x))Example 20.55
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Example 20.55: if f ∈ Γ₀(H), then F_{∂ f} ≤ f ⊕ f* and dom f × dom f* ⊆ dom F_{∂ f}, where f* is represented by the canonical packaged conjugate f∗[hf].
open scoped InnerProductSpace SetValuedOperator
universe u
namespace ERealFunction
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
variable (f : H → Set.Ioi (⊥ : EReal)) (hf : f ∈ Γ₀(H))
local notation "f⋆" => f∗[hf]
-- Proof sketch: fix `(x, u)` and expand `F_{∂ f}(x, u)`, written in Lean as
-- `F[(∂ f)] (x, u)`, as
-- the supremum over graph points `(y, v)` of `∂ f`. Proposition 16.10 rewrites
-- `v ∈ (∂ f) y` as Fenchel--Young equality `f y + f^*(v) = ⟪y, v⟫`, and Proposition 13.15 gives
-- `⟪y, u⟫ - f y ≤ f^*(u)` together with `⟪x, v⟫ - f^*(v) ≤ f^{**}(x) = f x` by
-- Corollary 13.38. Taking the supremum yields the pointwise bound by `f ⊕ f^*`.
theorem fitzpatrickFunction_subdifferential_le_separableSum_gammaZeroConjugate_and_dom_prod_subset
:
F[(∂ f)] ≤
(f ⊕ f⋆).asEReal ∧
effectiveDomain f ×ˢ effectiveDomain f⋆ ⊆
dom (F[(∂ f)])Proposition 20.56
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 20.56 (1): clause (i). On the graph of a monotone operator, the Fitzpatrick function agrees with the pairing.
open scoped InnerProductSpace Pointwise SetValuedOperator
open ERealFunction
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
theorem fitzpatrickFunction_eq_inner_of_mem_graph
(A : SetValuedOperator H H) (hA_mono : A.IsMonotone) {x u : H} (hxu : (x, u) ∈ gra A) :
F[A] (x, u) = ((⟪x, u⟫_ℝ : ℝ) : EReal)Proposition 20.57
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 20.57 (1): the Fitzpatrick function of the componentwise product operator (x, y) ↦ A x × B y is the sum of the Fitzpatrick functions of A and B on the Chapter 9 raw-pair ℓ² Hilbert product H × K.
open scoped InnerProductSpace SetValuedOperator
open ERealFunction
universe u v
namespace SetValuedOperator
variable {H : Type u} {K : Type v}
variable [NormedAddCommGroup H] [InnerProductSpace ℝ H]
variable [NormedAddCommGroup K] [InnerProductSpace ℝ K]
theorem fitzpatrickFunction_prod_apply
(A : SetValuedOperator H H) (B : SetValuedOperator K K)
(x u : H) (y v : K) :
F[(A × B)] ((x, y), (u, v)) = F[A] (x, u) + F[B] (y, v)Proposition 20.58
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 20.58 (1): the Fitzpatrick function of a maximally monotone operator dominates the pairing everywhere.
open scoped InnerProductSpace SetValuedOperator
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
theorem Maximal.inner_le_fitzpatrickFunction
{A : SetValuedOperator H H} (hA : Maximal IsMonotone A) (x u : H) :
((⟪x, u⟫_ℝ : ℝ) : EReal) ≤ F[A] (x, u)Corollary 20.59
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Corollary 20.59 (1): if graph points (x_n, u_n) of a maximally monotone operator converge weakly to (x, u), then the pairing is weakly lower semicontinuous along the sequence: ⟪x, u⟫ ≤ liminf_n ⟪x_n, u_n⟫.
open Filter
open scoped InnerProductSpace SetValuedOperator Topology
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
variable {A : SetValuedOperator H H} (hA : Maximal IsMonotone A)
variable {xSeq uSeq : ℕ → H} {x u : H}
variable (hgraph : ∀ n, (xSeq n, uSeq n) ∈ gra A)
variable (hxSeq : Tendsto (fun n ↦ toWeakSpace ℝ H (xSeq n)) atTop (𝓝 (toWeakSpace ℝ H x)))
variable (huSeq : Tendsto (fun n ↦ toWeakSpace ℝ H (uSeq n)) atTop (𝓝 (toWeakSpace ℝ H u)))
-- Proof sketch: apply weak lower semicontinuity of the Fitzpatrick function to the sequence
-- `(xSeq n, uSeq n)`, use Proposition 20.58 to bound `⟪x, u⟫` by `F_A(x, u)`, and rewrite
-- `F_A(xSeq n, uSeq n)` as `⟪xSeq n, uSeq n⟫` because every term lies in `gra A`.
theorem Maximal.inner_le_liminf_inner_of_tendsto_weakly_seq
: ⟪x, u⟫_ℝ ≤ liminf (fun n ↦ ⟪xSeq n, uSeq n⟫_ℝ) atTopProposition 20.60
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 20.60: if graph points of a maximally monotone operator converge strongly and their residuals with respect to the canonical metric projectors P[(C : Set H), hC] and P[(D : Set H), hD] onto nonempty closed affine subspaces C and D with D.direction = C.directionᗮ converge strongly to 0, then the pairings converge to ⟪x, u⟫ and the limit pair belongs to ((C : Set H) ×ˢ (D : Set H)) ∩ gra A.
open EuclideanGeometry
open Filter
open scoped InnerProductSpace SetValuedOperator Topology
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H]
variable {C D : AffineSubspace ℝ H}
variable (hC_nonempty : (C : Set H).Nonempty) (hC_closed : IsClosed (C : Set H))
variable (hD_nonempty : (D : Set H).Nonempty) (hD_closed : IsClosed (D : Set H))
local notation "hC" =>
isChebyshev_of_nonempty_isClosed_convex hC_nonempty hC_closed C.convex
local notation "hD" =>
isChebyshev_of_nonempty_isClosed_convex hD_nonempty hD_closed D.convex
theorem
Maximal.tendsto_inner_and_mem_prod_inter_graph_of_strong_graph_seq_of_projection_residual_zero
{A : SetValuedOperator H H} (hA : Maximal IsMonotone A)
(hCD : D.direction = C.directionᗮ)
{xSeq uSeq : ℕ → H} {x u : H}
(hgraph : ∀ n, (xSeq n, uSeq n) ∈ gra A)
(hxSeq : Tendsto xSeq atTop (𝓝 x))
(huSeq : Tendsto uSeq atTop (𝓝 u))
(hCproj : Tendsto (fun n ↦ xSeq n - P[(C : Set H), hC] (xSeq n))
atTop (𝓝 (0 : H)))
(hDproj : Tendsto (fun n ↦ uSeq n - P[(D : Set H), hD] (uSeq n))
atTop (𝓝 (0 : H))) :
Tendsto (fun n ↦ ⟪xSeq n, uSeq n⟫_ℝ) atTop (𝓝 ⟪x, u⟫_ℝ) ∧
(x, u) ∈ ((C : Set H) ×ˢ (D : Set H)) ∩ gra AProposition 20.61
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Proposition 20.61 (1): clause (i). The Fenchel conjugate of the Fitzpatrick function is the Fenchel biconjugate of the inverse-graph indicator plus the pairing on H × H.
open Set
open ERealFunction
open scoped InnerProductSpace SetValuedOperator
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
theorem conjugate_fitzpatrickFunction_eq_biconjugate_inverseGraphIndicator_add_pairing
(A : SetValuedOperator H H) :
(F[A])∗ =
(((ι[gra A⁻¹]).asEReal + fun p : H × H ↦ ((⟪p.1, p.2⟫_ℝ : ℝ) : EReal)))∗∗Remark 20.62
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Remark 20.62: Proposition 20.58 and Proposition 20.61 recover a maximally monotone operator from the Fitzpatrick contact-set descriptions, giving converses to the canonical pairing-equality-operator construction of Theorem 20.46.
namespace SetValuedOperator
recall Maximal.eq_pairingEqualityOperator_conjugateTranspose_fitzpatrickFunctionTheorem 20.63
Lean contributors (project-level attribution): Yifan Bai, Zichen Wang, Zaiwen Wen.
Theorem 20.63: if A is monotone and gra A is nonempty, set G = pav(F_A, F_A^{*T}) and define B by gra B = {(x, u) | G(x, u) = ⟪x, u⟫}. Then fitzpatrickProximalAverageExtension A hA_mono hA_graph is a maximally monotone extension of A.
open scoped InnerProductSpace
open ERealFunction
universe u
namespace SetValuedOperator
variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H]
variable [CompleteSpace H]
theorem fitzpatrickProximalAverageExtension_isMaximallyMonotone_and_extends
(A : SetValuedOperator H H) (hA_mono : A.IsMonotone) (hA_graph : A.graph.Nonempty) :
Maximal IsMonotone (fitzpatrickProximalAverageExtension A hA_mono hA_graph) ∧
A ≤ fitzpatrickProximalAverageExtension A hA_mono hA_graph