Chapter 03 — Nonsmooth Convex Optimization
Proofs omitted. Hover over highlighted terms to see their Lean types.
Algorithm 3.1
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Algorithm 3.1: the resisting oracle for f_k sends x ∈ ℝⁿ to the pair consisting of the Nemirovski hard-instance value f_k(x) and the vector γ e_{i*} + μ x, where i* is the least index among 1, ..., k attaining the maximum prefix coordinate.
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
local notation "e[" i "]" => EuclideanSpace.single i (1 : ℝ)
def resistingOracle
(k : ℕ) (hk : 0 < k) (hkn : k ≤ n) (μ γ : ℝ) : E → ℝ × E := fun x ↦
(f_k n k μ γ x, resistingSubgradient k hk hkn μ γ x)
Definition 3.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1 is a source-facing recall in the convex constrained minimization domain. Primary domain: - finite-dimensional convex minimization with an extended-real-valued objective and finitely many extended-real-valued inequality constraints. Sampled owner-style declarations: - GeneralMinimizationProblem in Chap01/Definition_1_1_3, the earlier project owner for a basic feasible set together with finitely many scalar constraints and comparison senses. - SetConstrainedMinimizationProblem in Chap01/Definition_1_3_3, the zero-constraint owner for an ambient feasible set and a real-valued objective. - GeneralConvexMinimizationProblem in Definition_3_1_1_1, the chapter owner matching the exact extended-valued convex data of Definition 3.1. - GeneralConvexMinimizationProblem.IsFeasible, the derived feasibility predicate attached to that owner. Best owner abstraction: - GeneralConvexMinimizationProblem X m, with textbook specialization X = EuclideanSpace ℝ (Fin n) Primitive data: - feasibleSet : Set X - objective : X → WithTop ℝ - constraints : Fin m → X → WithTop ℝ - feasibleSet_closed - feasibleSet_convex - objective_convex - constraints_convex Derived API: - the coercion to the ambient objective function - GeneralConvexMinimizationProblem.IsFeasible Source/core/bridge triage: - source-facing: the textbook general convex minimization problem with convex inequality constraints fᵢ(x) ≤ 0. - core/canonical: GeneralConvexMinimizationProblem X m. - bridge/view: the textbook Euclidean specialization X = EuclideanSpace ℝ (Fin n), together with the coercion to the ambient objective and the feasibility predicate GeneralConvexMinimizationProblem.IsFeasible. This file therefore recalls the canonical owner abstraction directly and keeps no parallel public wrapper such as ConvexOptimizationProblem or a separate feasibility package.
universe u
variable {X : Type u} [TopologicalSpace X] [AddCommMonoid X] [Module ℝ X]
recall GeneralConvexMinimizationProblem
{X : Type u} [TopologicalSpace X] [AddCommMonoid X] [Module ℝ X] (m : ℕ) :
Type uLemma 3.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1 lies in the finite Jensen / convex-combination domain. Primary mathematical domain: - Jensen's inequality for convex functions on finite weighted sums Relevant owner-style declarations sampled before refinement: - ConvexOn.map_centerMass_le - ConvexOn.map_sum_le - Convex.sum_mem - ConvexOn.map_add_sum_le Best owner abstraction: - ConvexOn.map_sum_le Primitive data: - a convex-on-set witness hf : ConvexOn ℝ s f - a finite family x : Fin m → E with x i ∈ s - a coefficient family α : Fin m → ℝ with 0 ≤ α i and ∑ i, α i = 1 Derived API: - Jensen's inequality at the weighted sum ∑ i, α i • x i - the separate domain-membership consequence from Convex.sum_mem - any simplex-packaged or convex-combination-packaged presentation as downstream bridge API, not as the owner theorem itself Source/core/bridge triage: - source-facing: the explicit finite Jensen inequality from the text - core/canonical: ConvexOn.map_sum_le - bridge/view: Convex.sum_mem for domain membership, and later chapter simplex packaging only as a presentation layer for the same finite weighted-sum data The textbook item is Jensen's inequality itself. Since mathlib already owns that statement in the correct nonnegative-weight form, this file centers the canonical owner declaration directly. The domain-membership consequence belongs to Convex.sum_mem and is not bundled into the main item.
recall ConvexOn.map_sum_leProposition 3.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.1: if a β-valued convex function is defined on a set containing a finite nonempty vertex set t, then its supremum on convexHull 𝕜 (t : Set E) equals its supremum on the vertex set itself. Since t is finite and nonempty, this is exactly the finite-maximum statement behind the textbook real-valued simplex maximum principle; specializing 𝕜 = β = ℝ recovers the usual formulation.
universe u v w
variable {𝕜 : Type u} [Field 𝕜] [LinearOrder 𝕜] [IsStrictOrderedRing 𝕜]
variable {E : Type v} [AddCommGroup E] [Module 𝕜 E]
variable {β : Type w} [ConditionallyCompleteLinearOrder β] [AddCommGroup β]
[IsOrderedAddMonoid β] [Module 𝕜 β] [IsStrictOrderedModule 𝕜 β]
theorem convexOn_sSup_image_convexHull_eq_sSup_image_vertices
(t : Finset E) (ht : t.Nonempty) {C : Set E} {f : E → β} (hf : ConvexOn 𝕜 C f)
(htC : (t : Set E) ⊆ C) :
sSup (f '' convexHull 𝕜 (t : Set E)) = sSup (f '' (t : Set E))Theorem 3.1
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1, owner form: if x is a convex combination of points and all those points lie in C, then a convex function on C takes at x a value bounded above by one of the endpoint values.
open scoped BigOperators
universe u v w z
variable {𝕜 : Type u} [Field 𝕜] [LinearOrder 𝕜] [IsStrictOrderedRing 𝕜]
variable {ι : Type v}
variable {E : Type w} [AddCommGroup E] [Module 𝕜 E]
variable {β : Type z} [AddCommGroup β] [LinearOrder β] [IsOrderedAddMonoid β]
[Module 𝕜 β] [IsStrictOrderedModule 𝕜 β]
local notation "convexCombination" =>
@ConvexSpace.convexCombination 𝕜 E inferInstance inferInstance inferInstance inferInstance
theorem convexOn_exists_ge_of_convex_combination
{C : Set E} {f : E → β} (hf : ConvexOn 𝕜 C f) {points : ι → E} {x : E}
(hx : is_convex_combination_of 𝕜 points x) (hpoints : Set.range points ⊆ C) :
∃ i, f x ≤ f (points i)
Algorithm 3.2
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Algorithm 3.2: a simple-set subgradient method for minimizing f over a simple closed convex set Q consists of a feasible starting point x₀ and a stepsize sequence; the objective, feasible set, and first-order oracle are inherited from the owner first-order convex minimization problem. The output sequence is the recursively defined iteration obtained by applying the owner normalized subgradient step with stepsize h_k at the current iterate, i.e. x_{k+1} = π_Q (x_k - h_k • (g(x_k) / ‖g(x_k)‖)), where g(x_k) is the owner-oracle subgradient at x_k. The textbook admissibility conditions on (h_k) are recorded separately by IsAdmissibleSubgradientStepsizeSequence.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
structure SimpleSetSubgradientMethod
(problem : FirstOrderConvexMinimizationProblem E) where
/-- The prescribed initial point `x₀`. -/
x0 : E
/-- The initial point is feasible. -/
x0_mem : x0 ∈ problem.feasibleSet
/-- The stepsize sequence `h₀, h₁, ...`. -/
stepsize : ℕ → ℝ
Definition 3.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.2 is a source-facing recall of the finite-value domain of an extended-real-valued function together with the standing assumption that this domain is nonempty. Primary domain: - convex analysis of EReal-valued functions through their finite-value domain. Relevant declarations and owner-style recall sampled before refinement: - extendedRealEffectiveDomain - extendedRealEffectiveDomain_nonempty_iff - mem_extendedRealEffectiveDomain_iff - (dom f).Nonempty - Set.Nonempty Best owner abstraction: - extendedRealEffectiveDomain Primitive data: - the owner set dom f Derived API: - (dom f).Nonempty - extendedRealEffectiveDomain_nonempty_iff - mem_extendedRealEffectiveDomain_iff - the existential expansion of Set.Nonempty Source/core/bridge triage: - source-facing: the domain object dom f together with its standing nonemptiness assumption - core/canonical: extendedRealEffectiveDomain - bridge/view: mem_extendedRealEffectiveDomain_iff, extendedRealEffectiveDomain_nonempty_iff This item therefore targets the source-facing owner layer directly: it recalls extendedRealEffectiveDomain as the main entry and keeps the nonemptiness condition only as the companion standing assumption. Since neither declaration uses Euclidean structure, the public surface stays at the owner's arbitrary-domain level.
open scoped ConvexAnalysis
recall extendedRealEffectiveDomain
{X : Type _} (f : X → EReal) :
Set XLemma 3.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.2: if the epigraph of f over Q is a closed convex subset of X × ℝ and Q₁ ⊆ Q is closed and convex, then the epigraph of the restriction of f to Q₁ is also closed and convex.
universe u
variable {X : Type u} [TopologicalSpace X] [AddCommMonoid X] [Module ℝ X]
theorem isClosed_convex_constrainedEpigraph_restrict
{Q Q₁ : Set X} {f : X → WithTop ℝ}
(hf_closed : IsClosed (constrainedEpigraph Q f))
(hf_convex : Convex ℝ (constrainedEpigraph Q f))
(hQ₁_subset : Q₁ ⊆ Q)
(hQ₁_closed : IsClosed Q₁)
(hQ₁_convex : Convex ℝ Q₁) :
IsClosed (constrainedEpigraph Q₁ f) ∧ Convex ℝ (constrainedEpigraph Q₁ f)Proposition 3.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.2, generalized from the textbook ℝⁿ setting: if the finite real part of f : X → ℝ ∪ {±∞} is continuous on its finite-value domain dom f, and dom f is closed, then the effective epigraph of f is a closed subset of X × ℝ.
universe u
open scoped ConvexAnalysis
variable {X : Type u} [TopologicalSpace X]
theorem isClosed_effectiveEpigraph_of_continuousOn_of_isClosed_dom
{f : X → EReal}
(hf_cont : ContinuousOn (extendedRealRealPart f) (dom f))
(hdom_closed : IsClosed (dom f)) :
IsClosed (effectiveEpigraph f)Theorem 3.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.2: for a function considered on a set s, the usual two-point convexity inequality along segments in s is equivalent to the forward affine-ray secant lower bound. Specializing to 𝕜 = ℝ and E = EuclideanSpace ℝ (Fin n) recovers the textbook statement on ℝⁿ.
variable {𝕜 E : Type*} [Field 𝕜] [LinearOrder 𝕜] [IsStrictOrderedRing 𝕜]
[AddCommGroup E] [Module 𝕜 E]
theorem segment_inequality_iff_affine_ray_inequality
(s : Set E) (f : E → 𝕜) :
(∀ ⦃x y : E⦄, x ∈ s → y ∈ s →
∀ ⦃α : 𝕜⦄, α ∈ Set.Icc (0 : 𝕜) 1 →
α • x + (1 - α) • y ∈ s →
f (α • x + (1 - α) • y) ≤ α * f x + (1 - α) * f y) ↔
(∀ ⦃x y : E⦄, x ∈ s → y ∈ s →
∀ ⦃β : 𝕜⦄, 0 ≤ β →
y + β • (y - x) ∈ s →
f (y + β • (y - x)) ≥ f y + β * (f y - f x))Algorithm 3.3
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Algorithm 3.3: a subgradient method with functional constraints is determined by a one-constraint first-order problem, a feasible starting point x₀ ∈ Q, and a fixed accuracy parameter ε. At iterate x_k, if the unique constraint value problem.constraints 0 x_k = f̄(x_k) satisfies problem.constraints 0 x_k ≤ ε, the method takes the objective step x_{k+1} = π_Q (x_k - (ε / ‖g(x_k)‖²) • g(x_k)); otherwise it takes the aggregate-constraint correction x_{k+1} = π_Q (x_k - (f̄(x_k) / ‖ḡ(x_k)‖²) • ḡ(x_k)).
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
structure FunctionalConstraintSubgradientMethod
(problem : ProjectedMultipleConstraintFirstOrderProblem E 1) where
/-- The prescribed initial point `x₀`. -/
x0 : E
/-- The initial point lies in the simple feasible set `Q`. -/
x0_mem : x0 ∈ problem.feasibleSet
/-- The target accuracy parameter `ε`. -/
ε : ℝ
Definition 3.3
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.3 (1), generalized from the textbook ℝⁿ setting: an ℝ ∪ {+∞}-valued function is convex when its effective domain is convex and its finite-value part satisfies the Jensen inequality there.
universe u
open scoped WithTopConvexAnalysis
variable {X : Type u} [AddCommMonoid X] [Module ℝ X]
variable (f : X → WithTop ℝ)
#check ConvexOn ℝ (dom f) (withTopRealPart f)Lemma 3.3
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.3 lies in the chapter's support-function / convex-hull domain. Sampled owner-style declarations: - supportFunction from Definition_3_9 - supportFunction_apply - supportFunction_convexHull_eq - mathlib sSup_union Best owner abstraction: - the chapter owner supportFunction together with its canonical convex-hull invariance theorem supportFunction_convexHull_eq Primitive data: - two sets Q₁ Q₂ : Set E Derived API: - the textbook pointwise specialization ξ[convexHull ℝ (Q₁ ∪ Q₂)] x = max (ξ[Q₁] x) (ξ[Q₂] x) Source/core/bridge triage: - source-facing: supportFunction_convexHull_union_eq_max - core/canonical: supportFunction and supportFunction_convexHull_eq - bridge/view: supportFunction_apply and sSup_union
universe u
variable {E : Type u} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
open scoped SupportFunction
recall supportFunctionProposition 3.3
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.3 (1): the absolute value function on ℝ is convex on all of ℝ.
theorem abs_convexOn_univ :
ConvexOn ℝ Set.univ (fun x : ℝ ↦ |x|)Theorem 3.3
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.3: an ℝ ∪ {+∞}-valued function is convex on its effective domain exactly when its epigraph over that domain is a convex subset of the ambient product space.
universe u
open scoped WithTopConvexAnalysis
variable {X : Type u} [AddCommMonoid X] [Module ℝ X]
theorem convexOn_iff_convex_effective_epigraph
(f : X → WithTop ℝ) :
ConvexOn ℝ (dom f) (withTopRealPart f) ↔
Convex ℝ (WithTopConvexAnalysis.effectiveEpigraph f)Algorithm 3.4
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Algorithm 3.4: a subgradient switching strategy for approximate Lagrange multipliers is determined by a constrained first-order problem with feasible set Q, a feasible initial point x₀, a scalar h, and a stepwise choice sequence j_k. Here j_k = none exactly on objective iterations, while j_k = some j records the chosen active constraint j ∈ 𝒥_k on correction iterations. This matches the textbook's arbitrary per-iteration choices j_k ∈ 𝒥_k without packaging an off-trajectory selector. To keep the run owner below completeness, the iterate sequence x_k is stored directly as primitive run data, while the textbook projection recursion is recorded separately by step_rule exactly on the completeness-dependent layer where the projection map is available. Positivity consequences for h are kept as separate theorem hypotheses rather than as primitive owner data.
universe u
open NormedSpace
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable {m : ℕ}
structure ApproximateLagrangeMultiplierSwitchingMethod
(problem : ProjectedMultipleConstraintFirstOrderProblem E m) where
/-- The prescribed initial point `x₀`. -/
x0 : E
/-- The initial point lies in the feasible set `Q`. -/
x0_mem : x0 ∈ problem.feasibleSet
/-- The constant scalar `h` from the textbook rule defining the active set. -/
h : ℝ
/-- The chosen branch at time `k`, recorded as `none` exactly on objective steps and as
`some j` on selected active-constraint steps. -/
selectedIndexAt : ℕ → Option (Fin m)
/-- The iterate sequence `x₀, x₁, ...` of the run. -/
iterates : ℕ → E
/-- The zeroth iterate is the prescribed starting point `x₀`. -/
iterates_zero : iterates 0 = x0
/-- Along the run, the chosen branch is `none` exactly when the active set is empty and
otherwise records an index from that active set. -/
selectedIndexAt_valid (k : ℕ) :
match selectedIndexAt k with
| none => problem.switchingActiveSet h (iterates k) = ∅
| some j => j ∈ problem.switchingActiveSet h (iterates k)
/-- Whenever the projection operator is available, the iterates satisfy the textbook switching
recursion. -/
step_rule [CompleteSpace E] (k : ℕ) :
iterates (k + 1) = problem.switchingStep h (iterates k) (selectedIndexAt k)
Definition 3.4
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.4: a point is a strict convex combination of a finite family of points when it is the canonical convex combination associated to a strict simplex weight vector.
open scoped BigOperators
open ConvexSpace
universe u v w
-- Declarations for this item will be appended below by the statement pipeline.
variable (R : Type u) [PartialOrder R] [Semiring R] [IsStrictOrderedRing R]
variable {ι : Type v}
variable {E : Type w} [ConvexSpace R E]
def is_strict_convex_combination_of (points : ι → E) (x : E) : Prop :=
∃ w : StdSimplex.Strict R ι, x = convexCombination (w.1.map points)
Lemma 3.4
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.4: any univariate closed convex function is continuous on its effective domain.
open Filter
open scoped Topology
open scoped WithTopConvexAnalysis
theorem ClosedConvexFunction.continuousOn_effectiveDomain_one_dimensional
{f : ℝ → WithTop ℝ} (hf : ClosedConvexFunction f) :
ContinuousOn (withTopRealPart f) (dom f)Proposition 3.4
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.4 lies in the chapter's closed-convex WithTop-valued convex-analysis domain. Primary domain: - continuous convex real-valued functions viewed through the owner predicate ClosedConvexFunction. Sampled owner-style declarations: - closedConvexFunction_coe_of_convexOn_continuous in Proposition_3_1_1_3, the existing chapter theorem with the same mathematical content at the intrinsic real topological-module level; - ClosedConvexFunction and ClosedConvexOn in Definition_3_1_1_5, the source-facing owners for closed convex extended-real-valued functions; - mathlib ConvexOn.convex_epigraph; - mathlib IsClosed.epigraph. Best owner abstraction: - closedConvexFunction_coe_of_convexOn_continuous. Primitive data: - a real-valued function f; - its convexity on Set.univ; - its continuity. Derived API: - the closed-convex owner statement for the WithTop ℝ coercion of f. Source/core/bridge triage: - source-facing: Proposition 3.4 as the Euclidean ℝⁿ statement; - core/canonical: ClosedConvexFunction together with ConvexOn and closed epigraphs; - bridge/view: the already-proved intrinsic theorem closedConvexFunction_coe_of_convexOn_continuous. This file is therefore recall-only. Keeping a second Euclidean theorem shell here would duplicate an existing owner-level declaration instead of reusing the chapter's canonical abstraction.
recall closedConvexFunction_coe_of_convexOn_continuousTheorem 3.4
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.4: if f is convex on its effective domain, then each constrained sublevel set of f over that domain is convex.
universe u
open scoped WithTopConvexAnalysis
namespace ConvexOn
variable {X : Type u} [AddCommMonoid X] [Module ℝ X]
variable {f : X → WithTop ℝ}
theorem convex_constrainedSublevelSet
(hf : ConvexOn ℝ (dom f) (withTopRealPart f)) (β : ℝ) :
Convex ℝ (constrainedSublevelSet (dom f) f β)Algorithm 3.5
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Algorithm 3.5: the resisting oracle for the feasibility problem starts from the initialized box [-R \bar e_n, R \bar e_n], returns any feasibility separator g satisfying ⟪g, x - z⟫ ≥ 0 for all z ∈ B₀ when the query point x lies outside that box, reuses the stored separator g_k for the maximal active box B_k with k < m, and otherwise splits the current box B_m at the midpoint of the selected coordinate, returning e_i when the query lies in the upper half and -e_i when it lies in the lower half, then retaining the opposite half-box and updating the state by incrementing m and cycling the coordinate index.
local notation "E(" n ")" => EuclideanSpace ℝ (Fin n)
open FeasibilityResistingOracleState
def FeasibilityResistingOracleStep
{n : ℕ} (R : ℝ) (hn : 0 < n) (state : FeasibilityResistingOracleState n)
(x g : E(n)) (state' : FeasibilityResistingOracleState n) : PropDefinition 3.5
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.5 is a source-facing recall in the chapter's closed-convex-function domain. Primary domain: - closed convex WithTop ℝ-valued functions on real modules, specializing to the textbook Euclidean setting. Sampled owner-style declarations: - constrainedEpigraph - ConvexOn - ClosedConvexOn - ClosedConvexFunction Best owner abstraction: - the source-facing owner predicate ClosedConvexOn Q f, with ClosedConvexFunction f as the exact owner-derived special case ClosedConvexOn (dom f) f. Primitive data: - dom f - constrainedEpigraph Q f Derived API: - ClosedConvexFunction - ClosedConvexOn.subset_withTopEffectiveDomain, projecting the owner field Q ⊆ dom f - ClosedConvexOn.isClosed_constrainedEpigraph - ClosedConvexOn.convex_constrainedEpigraph - ClosedConvexOn.convexOn_withTopRealPart Source/core/bridge triage: - source-facing: ClosedConvexOn, ClosedConvexFunction - core/canonical: dom f, constrainedEpigraph, ConvexOn - bridge/view: the projection lemmas out of ClosedConvexOn and the textbook specialization X = EuclideanSpace ℝ (Fin n) This file therefore recalls the upstream owner declarations directly and introduces no parallel chapter-local reformulation of closed convexity.
recall ClosedConvexOnLemma 3.5
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.5 is a recall-only surface in the chapter's finite directional-derivative domain. Primary domain: - finite one-sided directional derivatives of convex ℝ ∪ {+∞}-valued functions at interior points of the effective domain. Sampled owner-style declarations: - convexDirectionalDerivative in Theorem_3_21, the chapter owner for the extended-valued directional derivative; - convexDirectionalDerivativeReal_convexOn_univ_of_mem_interior in Theorem_3_21, the owner convexity theorem for the finite theorem-surface view; - convexDirectionalDerivativeReal_posHomOn_univ_of_mem_interior and convexDirectionalDerivativeReal_affine_support_of_mem_interior in Lemma_3_1_3_1, the source-facing real-valued consequences already consolidated by Lemma_3_1_5. Best owner abstraction: - the finite directional-derivative owner f′[hx], with the recalled consequence surface already gathered upstream by Lemma_3_1_5. Primitive data: - none in this file; all directional-derivative data is already owned upstream. Derived API: - the recalled convexity, positive-homogeneity, and affine-support consequences at an interior point. Source/core/bridge triage: - source-facing: the three consequences asserted in Lemma 3.5 for the finite directional derivative; - core/canonical: convexDirectionalDerivative; - bridge/view: this numbered-item recall surface, reusing Lemma_3_1_5. The previous version repeated the same three recalls with a second local domain summary. This file now reuses the earlier chapter recall surface directly and keeps only the numbered-item layer.
recall convexDirectionalDerivativeReal_convexOn_univ_of_mem_interior
Proposition 3.5
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.5 (1): the reciprocal function x ↦ 1 / x is convex on (0, ∞).
open scoped WithTopConvexAnalysis
open WithTopConvexAnalysis
theorem convexOn_one_div_Ioi_zero :
ConvexOn ℝ (Set.Ioi 0) (fun x : ℝ ↦ 1 / x)Theorem 3.5
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.5 is a source-facing recall in the chapter's closed-convex-function domain. Primary domain: - closed convex WithTop ℝ-valued functions and their real sublevel sets. Sampled owner-style declarations: - ClosedConvexOn - ClosedConvexFunction - ClosedConvexFunction.isClosed_convex_sublevelSet Best owner abstraction: - ClosedConvexFunction Primitive data: - the effective-domain closed-convex owner hypothesis ClosedConvexFunction f Derived API: - ClosedConvexFunction.isClosed_convex_sublevelSet Source/core/bridge triage: - source-facing: Theorem 3.5, the sublevel-set consequence of closed convexity - core/canonical: ClosedConvexFunction - bridge/view: ClosedConvexOn and the constrained-epigraph API behind it This file therefore reuses the canonical owner theorem directly, without introducing a second sublevel-set wrapper or theorem-shaped alias.
recall ClosedConvexFunction.isClosed_convex_sublevelSetAlgorithm 3.6
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Algorithm 3.6: a general cutting-plane scheme for minimizing f over Q consists of bounded initial localization data E₀ ⊇ Q, a query sequence yₖ ∈ Eₖ, the cutting vectors returned by the owner separation oracle at those query points, and successor localization sets Eₖ₊₁ containing the retained slice {x ∈ Eₖ | ⟪gₖ, yₖ - x⟫ ≥ 0}. On feasible queries the owner oracle returns a subgradient of f; on infeasible queries it returns a strict separating vector.
universe u
variable {E : Type u} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
structure GeneralCuttingPlaneScheme (problem : ConvexMinimizationWithSeparationOracle E) where
/-- The localization sets `E₀, E₁, E₂, ...`. -/
localizer : ℕ → Set E
/-- The query points `y₀, y₁, y₂, ...`. -/
queryPoint : ℕ → E
/-- The initial localization set `E₀` is bounded. -/
initial_bounded : Bornology.IsBounded (localizer 0)
/-- The feasible set `Q` is contained in the initial localization set `E₀`. -/
feasibleSet_subset_initial : problem.feasibleSet ⊆ localizer 0
/-- Each query point `yₖ` is chosen from the current localization set `Eₖ`. -/
query_mem (k : ℕ) : queryPoint k ∈ localizer k
/-- The next localization set contains the half-space slice of the current localization set cut
out by the current cutting vector at the current query point. -/
next_localizer_contains (k : ℕ) :
localizer k ∩ cuttingHalfspace (queryPoint k) (problem.oracle (queryPoint k)) ⊆
localizer (k + 1)Definition 3.6
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.6: the unit ball is the canonical owner specialization p.closedBall 0 1.
universe u
variable {𝕜 E : Type u} [SeminormedRing 𝕜] [AddCommGroup E] [SMul 𝕜 E]
variable (p : Seminorm 𝕜 E)
#check p.closedBall 0 1Lemma 3.6
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.6 (1), generalized from the textbook ℝⁿ setting: if Q is convex and every constrained subdifferential ∂_Q f(x) with x ∈ Q is nonempty, then the finite-valued representative of f is convex on Q. The nonemptiness hypothesis already forces f to be finite on Q.
open scoped Topology WithTopConvexAnalysis
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable (Q : Set E) (f : E → WithTop ℝ)
theorem convexOn_of_constrainedSubdifferential_nonempty
(hQ_convex : Convex ℝ Q)
(hsub : ∀ x ∈ Q, (∂[Q] f(x)).Nonempty) :
ConvexOn ℝ Q (withTopRealPart f)Proposition 3.6
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.6, stated at the intrinsic owner level: on any finite-dimensional real normed space, every seminorm defines a closed convex WithTop ℝ-valued function. The textbook ℝⁿ statement is the specialization to E = EuclideanSpace ℝ (Fin n), and separation still plays no role in convexity or closedness of the epigraph.
universe u
variable {E : Type u} [NormedAddCommGroup E] [NormedSpace ℝ E] [FiniteDimensional ℝ E]
namespace Seminorm
theorem closedConvexFunction
(p : Seminorm ℝ E) :
ClosedConvexFunction (fun x : E ↦ (p x : WithTop ℝ))Theorem 3.6
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.6 is recall-only in the chapter's closed-convex WithTop ℝ-valued function API. Primary domain: - closure properties of closed convex extended-real-valued functions on feasible sets, specialized in the source to the textbook ℝⁿ model and refined here to the owner ambient real module. Sampled owner-style declarations in this domain: - ClosedConvexOn from Definition_3_1_1_5, the chapter owner for closed convexity on a feasible set; - ClosedConvexOn.nonneg_smul from Theorem_3_1_5, the owner nonnegative-scalar closure rule; - ClosedConvexOn.add_inter from Theorem_3_1_5, the owner sum closure rule on intersected feasible sets; - ClosedConvexOn.max_inter from Theorem_3_1_5, the owner pointwise-maximum closure rule on intersected feasible sets. Best owner abstraction: - ClosedConvexOn. Primitive data: - the owner witnesses hf, hf₁, hf₂; - the scalar β together with the nonnegativity hypothesis 0 ≤ β. Derived API: - ClosedConvexOn.nonneg_smul; - ClosedConvexOn.add_inter; - ClosedConvexOn.max_inter. Source/core/bridge triage: - source-facing: the three closure properties recorded under Theorem 3.6; - core/canonical: the owner namespace ClosedConvexOn; - bridge/view: this Euclidean specialization file, which should recall the owner theorems directly rather than keep a second public vocabulary closedConvexOn_smul_nonneg, closedConvexOn_add, and closedConvexOn_max. The earlier file Theorem_3_1_5 already owns these exact closure operations with the correct chapter-level names and general ambient assumptions. This file therefore reuses those owner entries directly instead of exporting parallel theorem shells in the ℝⁿ presentation.
universe u
variable {X : Type u} [TopologicalSpace X] [AddCommMonoid X] [Module ℝ X]
recall ClosedConvexOn.nonneg_smul
Algorithm 3.7
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Algorithm 3.7 viewed as the chapter's general cutting-plane owner abstraction.
open MeasureTheory
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
[FiniteDimensional ℝ E]
namespace CenterOfGravityMethod
variable (problem : ConvexMinimizationWithSeparationOracle E)
def toGeneralCuttingPlaneScheme : GeneralCuttingPlaneScheme problemDefinition 3.7
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.7: the coordinate ℓ_p seminorm on ℝ^n, obtained by pulling back the owner L^p norm on WithLp (p : ENNReal) (Fin n → ℝ) along the Euclidean coordinate map. The induced Seminorm.IsNorm instance recovers the textbook norm property, its value on a vector is written ‖x‖_[p] below, and its closed balls are the textbook ℓ_p balls.
open scoped BigOperators
open WithLp
namespace EuclideanSpace
def lpSeminorm (n : ℕ) (p : LpExponent) : Seminorm ℝ (EuclideanSpace ℝ (Fin n)) :=
Seminorm.comp
(normSeminorm ℝ (WithLp (p : ENNReal) (Fin n → ℝ)))
(coordinateLpEquiv n p).toLinearMapLemma 3.7
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.7 lies in the chapter's extended-valued convex-analysis / subdifferential-calculus domain. Sampled owner-style declarations: - dom and withTopRealPart in Definition_3_3, the chapter owners for the effective domain and finite-value representative; - IsSubgradientAt and subdifferential in Definition_3_1_5, the chapter owners for extended-valued subgradients; - subdifferential_eq_singleton_gradient in Lemma_3_1_7, already stated on those owner abstractions. Best owner abstraction: - subdifferential_eq_singleton_gradient from Lemma_3_1_7. Primitive data: - none in this file; the theorem already lives upstream on the canonical owner surface. Derived API: - this recall-only source-facing entry point. Source/core/bridge triage: - source-facing: Lemma 3.7's singleton-subdifferential statement; - core/canonical: dom, withTopRealPart, and subdifferential; - bridge/view: this recall surface. The previous file duplicated the effective-domain, finite-real-part, convexity, subgradient, and subdifferential owners even though the exact theorem already exists in Lemma_3_1_7. This refinement removes that parallel local API and reuses the canonical chapter theorem directly.
recall subdifferential_eq_singleton_gradientProposition 3.7
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.7: for 0 < n and r ≥ 0, the closed ℓ₁-ball centered at x₀ in ℝⁿ is the convex hull of the signed standard-basis vertices x₀ ± r eᵢ.
open scoped BigOperators
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
local notation "e[" i "]" => EuclideanSpace.single i (1 : ℝ)
theorem l1_ball_eq_convexHull_signed_standard_basis_prop
(hn : 0 < n) (x₀ : E) (r : ℝ) (hr : 0 ≤ r) :
(EuclideanSpace.l1Seminorm n).closedBall x₀ r =
convexHull ℝ
(Set.range (fun i : Fin n ↦ x₀ + r • e[i]) ∪
Set.range (fun i : Fin n ↦ x₀ - r • e[i]))Theorem 3.7
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.7 is recall-only in the chapter's affine pullback calculus. Primary domain: - closed convex WithTop ℝ-valued functions on Euclidean spaces and their affine pullbacks. Sampled owner-style declarations in this domain: - ClosedConvexOn from Definition_3_1_1_5 - ClosedConvexOn.comp_continuousAffineMap from Theorem_3_1_2_2 - ClosedConvexOn.comp_affineMap from Theorem_3_1_2_2 - mathlib ConvexOn.comp_affineMap Best owner abstraction: - ClosedConvexOn Primitive data: - the owner witness hφ : ClosedConvexOn S φ - the affine map g : Eₙ →ᵃ[ℝ] Eₘ Derived API: - ClosedConvexOn.comp_affineMap Source/core/bridge triage: - source-facing: the affine-preimage closed-convexity statement - core/canonical: ClosedConvexOn.comp_affineMap - bridge/view: the earlier continuous-affine owner theorem ClosedConvexOn.comp_continuousAffineMap The earlier owner file already proves the exact Euclidean affine-map statement. The Nonempty and bounded hypotheses previously carried here are mathematically redundant for this theorem and do not belong in the public API. This file therefore recalls the canonical owner theorem directly instead of keeping a parallel wrapper name.
variable {m n : ℕ}
local notation "Eₙ" => EuclideanSpace ℝ (Fin n)
local notation "Eₘ" => EuclideanSpace ℝ (Fin m)
recall ClosedConvexOn.comp_affineMap
{m n : ℕ}
{S : Set (EuclideanSpace ℝ (Fin m))}
{φ : EuclideanSpace ℝ (Fin m) → WithTop ℝ}
(hφ : ClosedConvexOn S φ)
(g : EuclideanSpace ℝ (Fin n) →ᵃ[ℝ] EuclideanSpace ℝ (Fin m)) :
ClosedConvexOn (g ⁻¹' S) (φ ∘ g)Algorithm 3.8
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Algorithm 3.8: the ellipsoid-method center sequence y₀, y₁, ... attached to an initial center y₀ and radius R, where the update uses the recursively defined shape matrices and the cutting vectors returned by the separation oracle at the current center.
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
open scoped EllipsoidNotation
namespace EllipsoidMethod
variable (problem : ConvexMinimizationWithSeparationOracle (EuclideanSpace ℝ (Fin n)))
local notation "Mat" => Matrix (Fin n) (Fin n) ℝ
def center
(problem : ConvexMinimizationWithSeparationOracle (EuclideanSpace ℝ (Fin n)))
(initialCenter : E) (radius : ℝ) : ℕ → EDefinition 3.8
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.8: the textbook Fenchel dual is the chapter owner fenchelDual, written φ⋆.
open scoped ConvexAnalysis WithTopConvexAnalysis
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable (φ : E → WithTop ℝ)
#check fenchelDualLemma 3.8
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.8, subdifferential clause: at each x ∈ interior (dom ψ), the subdifferential of the monotone convex composition is the convex hull of the products λ • g with λ ∈ ∂ φ(withTopRealPart ψ x) and g ∈ ∂ ψ(x).
open scoped Pointwise WithTopConvexAnalysis
open Set
universe u
variable {V : Type u} [SeminormedAddCommGroup V] [InnerProductSpace ℝ V]
variable {ψ : V → WithTop ℝ} {φ : ℝ → WithTop ℝ}
theorem subdifferential_monotoneConvexComp_eq_convexHull {x : V}
(hψ_convex : ConvexOn ℝ (dom ψ) (withTopRealPart ψ))
(hφ_convex : ConvexOn ℝ (dom φ) (withTopRealPart φ))
(hφ_mono : MonotoneOn φ (withTopRealPart ψ '' dom ψ))
(hx : x ∈ interior (dom ψ)) :
∂ (monotoneConvexComp φ ψ)(x) =
convexHull ℝ (∂ φ((withTopRealPart ψ x)) • ∂ ψ(x))Proposition 3.8
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.8 on the Euclidean unit circle: for a nonnegative function on the unit sphere in EuclideanSpace ℝ (Fin 2), the associated extended-real-valued unit-disk boundary extension is convex in the chapter owner sense, and its effective domain is exactly the closed unit disk.
open scoped Topology WithTopConvexAnalysis
local notation "E" => EuclideanSpace ℝ (Fin 2)
local notation "S" => Metric.sphere (0 : E) 1
local notation "B" => Metric.closedBall (0 : E) 1
theorem unitDiskBoundaryExtension_convex_and_effectiveDomain
(φ : S → ℝ) (hφ_nonneg : ∀ z : S, 0 ≤ φ z) :
ConvexOn ℝ (dom (unitDiskBoundaryExtension φ))
(withTopRealPart (unitDiskBoundaryExtension φ)) ∧
dom (unitDiskBoundaryExtension φ) = BTheorem 3.8
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.8: if Q ⊆ X × Y is convex and φ : X × Y → ℝ ∪ {+∞} is convex on its effective domain, then the constrained partial infimum of the canonical EReal view withTopToEReal ∘ φ is convex in the chapter's EReal sense. Internally one restricts to Q ∩ dom φ to reuse the real-valued owner theorem, but that restriction is a proof device rather than a public hypothesis.
universe u v
open scoped ConvexAnalysis WithTopConvexAnalysis
variable {X : Type u} {Y : Type v}
variable [AddCommMonoid X] [Module ℝ X]
variable [AddCommMonoid Y] [Module ℝ Y]
theorem partialInfProjection_convexOn_of_convexWithTop
{Q : Set (X × Y)} {φ : X × Y → WithTop ℝ}
(hQ : Convex ℝ Q)
(hφ : ConvexOn ℝ (dom φ) (withTopRealPart φ)) :
ConvexOn ℝ (dom (partialInfProjection Q (withTopToEReal ∘ φ)))
(extendedRealRealPart (partialInfProjection Q (withTopToEReal ∘ φ)))
Algorithm 3.9
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Algorithm 3.9: for a constrained problem min_{x ∈ Q} f(x), Kelley's method is given by an initial feasible point x₀, a sampled subgradient sequence g_k ∈ ∂f(x_k), and an iterate sequence whose successor x_{k+1} minimizes the canonical Kelley model \hat f_k(X; ·) = nonsmoothModel f x g k over the feasible set Q.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
open scoped ConstrainedArgmin WithTopConvexAnalysis
structure KelleyMethod (problem : SetConstrainedMinimizationProblem E) where
/-- The iterate sequence `x₀, x₁, x₂, ...`. -/
iterates : ℕ → E
/-- The chosen sampled vector `g_k`. -/
subgradient : ℕ → E
/-- The initialization chooses a feasible point `x₀ ∈ Q`. -/
x0_mem : iterates 0 ∈ problem.feasibleSet
/-- Each sampled vector is a genuine subgradient of the owner objective at the current iterate.
-/
subgradient_spec (k : ℕ) :
IsSubgradientAt (fun x ↦ (problem x : WithTop ℝ)) (iterates k) (subgradient k)
/-- Each new iterate `x_{k+1}` belongs to the argmin set of the current Kelley model on `Q`. -/
nextIterate_mem_argmin (k : ℕ) :
iterates (k + 1) ∈
argmin[problem.feasibleSet] (nonsmoothModel problem iterates subgradient k)Definition 3.9
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.9: for a set Q ⊆ ℝⁿ, the support function of Q is the extended-real-valued function sending x to the supremum of the inner products ⟪g, x⟫ over g ∈ Q. The textbook ℝⁿ statement is the specialization E = EuclideanSpace ℝ (Fin n).
universe u
variable {E : Type u} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
def supportFunction (Q : Set E) : E → EReal :=
pointwiseSupremumOn Q fun x g ↦ ↑(inner ℝ g x)Lemma 3.9
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.9: generalized from the textbook ℝⁿ × ℝᵐ setting, at an interior point of the effective domain of a convex function on the intrinsic product WithLp 2 (E₁ × E₂) of real inner-product spaces, if the first-variable slice admits the displayed partial gradient in a neighborhood of the base point zBar, then the full subdifferential at zBar is exactly the image of the product of the singleton containing the first partial gradient with the second partial subdifferential under the canonical L² product map.
open Filter
open scoped WithTopConvexAnalysis
open scoped Topology
universe u v
variable {E₁ : Type u} {E₂ : Type v}
local notation "Z" => WithLp 2 (E₁ × E₂)
local notation "toZ" => WithLp.toLp 2
variable [NormedAddCommGroup E₁] [NormedAddCommGroup E₂]
variable [InnerProductSpace ℝ E₁] [CompleteSpace E₁] [InnerProductSpace ℝ E₂]
theorem subdifferential_eq_image_partialGradientFst_partialSubdifferentialSnd_of_nhds
{f : Z → WithTop ℝ}
(hf : ConvexOn ℝ (dom f) (withTopRealPart f))
{zBar : Z}
(hzBar : zBar ∈ interior (dom f))
(hgrad : ∀ᶠ z in 𝓝 zBar,
HasGradientAt
(withTopRealPart fun x ↦ f (toZ (x, z.snd)))
(partialGradientFst f z) z.fst)
(hcont : ContinuousAt (partialGradientFst f) zBar) :
∂ f(zBar) =
toZ ''
(({partialGradientFst f zBar} : Set E₁) ×ˢ partialSubdifferentialSnd f zBar)Proposition 3.9
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.9 reuses the source-facing weighted pointwise-supremum bridge theorem already established in Proposition 3.1.2.1.
recall closedConvexFunction_pointwiseSupremumOn_nonneg_weighted
Theorem 3.9
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.9: for a nonempty index type, if each slice x ↦ φ x y is closed and convex on Q, then the pointwise supremum over all indices is closed and convex on its canonical effective domain.
open Set
open scoped WithTopConvexAnalysis
universe u v
variable {ι : Type u} {X : Type v}
variable [TopologicalSpace X] [AddCommMonoid X] [Module ℝ X] [Nonempty ι]
variable {Q : Set X} {φ : X → ι → WithTop ℝ}
#check
(show (∀ y : ι, ClosedConvexOn Q (fun x ↦ φ x y)) →
ClosedConvexOn (pointwiseSupremumOnEffectiveDomain Q univ φ)
(pointwiseSupremumOn univ φ) from
fun hφ ↦ ClosedConvexOn.pointwise_sSup univ_nonempty (fun y _ ↦ hφ y))Algorithm 3.10
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Algorithm 3.10: for a fixed constrained problem problem : SetConstrainedMinimizationProblem E with feasible set Q ⊆ E and objective f, a complete-data level method records the iterate sequence x₀, x₁, ..., the approximate models \hat f_k(X; ·), the scalar parameters ε and α, and the textbook step condition that each successor x_{k+1} is a projection point of x_k onto the level set 𝓛_k(α) = {x ∈ Q | \hat f_k(X; x) ≤ (1 - α)\hat f_k^* + α f_k^*}. The stopping rule is the derived owner predicate history.shouldStop ε k, i.e. the gap test δ_k ≤ ε; whenever later results use the source side conditions ε > 0 or α ∈ (0, 1), those conditions are imposed in the theorem statement rather than stored as primitive fields of the run owner. The run owner itself is metric-level: downstream normed or inner-product estimates are separate companions, not part of the primitive algorithm data.
universe u
open scoped LevelMethodNotation
structure CompleteLevelMethod {E : Type u} [PseudoMetricSpace E]
(problem : SetConstrainedMinimizationProblem E) where
/-- The approximate model `\hat f_k(X; ·)` used at iteration `k`. -/
approximateObjective : ℕ → E → ℝ
/-- The explicit real lower value `\hat f_k^*` attached to the `k`-th model. -/
approximateOptimalValue : ℕ → ℝ
/-- The supplied real lower value agrees with the canonical exact `EReal` model minimum. -/
approximateOptimalValue_eq_optimalValue (k : ℕ) :
((approximateOptimalValue k : ℝ) : EReal) =
(levelMethodApproximateProblem problem.feasibleSet approximateObjective k).optimalValue
/-- The prescribed initial point `x₀`. -/
initialPoint : E
/-- The initial point lies in the feasible set `Q`. -/
initialPoint_mem : initialPoint ∈ problem.feasibleSet
/-- The accuracy parameter `ε`. -/
epsilon : ℝ
/-- The level coefficient `α`. -/
levelCoefficient : ℝ
/-- The iterate sequence `x₀, x₁, ...`. -/
iterate : ℕ → E
/-- The iterate sequence starts from the prescribed point `x₀`. -/
iterate_zero : iterate 0 = initialPoint
/-- The update step sends `x_k` to a projection point `x_{k+1}` on the current level set
`𝓛_k(α)`. -/
step_isProjectionPointOn (k : ℕ) :
let history :=
levelMethodHistoryFromApproximateValues
approximateOptimalValue
problem
iterate
IsProjectionPointOn
(𝓛[problem.feasibleSet, approximateObjective, history](levelCoefficient, k))
(iterate k)
(iterate (k + 1))Definition 3.10
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.10 lies in the chapter's convex-analysis/minimax-linearization domain. Sampled owner-style declarations: - mathlib unitInterval - mathlib AffineMap.lineMap - IsMinimaxLinearizationParameter - isMinimaxLinearizationParameter_iff Best owner abstraction: - the upstream source-facing owner IsMinimaxLinearizationParameter Primitive data: - a domain Q - two functions f₁ f₂ : Q → ℝ - a parameter lam : unitInterval Derived API: - the companion specification theorem isMinimaxLinearizationParameter_iff Source/core/bridge triage: - source-facing: the textbook notion of a minimax linearization parameter - core/canonical: the earlier owner predicate IsMinimaxLinearizationParameter from Definition_3_1_2_3 - bridge/view: the displayed EReal-infimum equality recalled by isMinimaxLinearizationParameter_iff Definition 3.10 adds no new mathematical data beyond that owner predicate, so this file recalls the owner declarations directly instead of keeping a parallel local wrapper.
recall IsMinimaxLinearizationParameterLemma 3.10
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.10: if f : E → ℝ ∪ {+∞} is convex, x₀ lies in the interior of its effective domain, and a subgradient selection g is continuous at x₀, then the finite real representative withTopRealPart f has gradient g x₀ at x₀. This is the canonical owner form of the textbook statement that f is differentiable at x₀ with ∇f(x₀) = g(x₀).
open scoped Gradient Topology WithTopConvexAnalysis
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E] [FiniteDimensional ℝ E]
variable {f : E → WithTop ℝ} {x0 : E} {g : E → E}
theorem hasGradientAt_withTopRealPart_of_continuous_subgradient_selection
(hf : ConvexOn ℝ (dom f) (withTopRealPart f))
(hx0 : x0 ∈ interior (dom f))
(hsub : ∀ᶠ x in 𝓝 x0, g x ∈ ∂ f(x))
(hcont : ContinuousAt g x0) :
HasGradientAt (withTopRealPart f) (g x0) x0Proposition 3.10
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.10: the support function of a nonempty set in a real inner-product space is positively homogeneous with respect to every bundled nonnegative scalar. The textbook ℝⁿ statement is the specialization E = EuclideanSpace ℝ (Fin n).
universe u
variable {E : Type u} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
open scoped SupportFunction
theorem supportFunction_smul (Q : Set E) (hQ : Q.Nonempty) (x : E) (τ : NNReal) :
ξ[Q] (τ • x) = (τ : EReal) * ξ[Q] xTheorem 3.10
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.10: if ψ is convex on domψ, φ is convex on a set I containing ψ '' domψ, and φ is nondecreasing on ψ '' domψ, then φ ∘ ψ is convex on domψ.
universe u
namespace ConvexOn
variable {X : Type u} [AddCommMonoid X] [Module ℝ X]
variable {domψ : Set X} {I : Set ℝ} {ψ : X → ℝ} {φ : ℝ → ℝ}
theorem comp_of_monotoneOn_image
(hφ : ConvexOn ℝ I φ) (hψ : ConvexOn ℝ domψ ψ)
(hφ_mono : MonotoneOn φ (ψ '' domψ)) (himage : ψ '' domψ ⊆ I) :
ConvexOn ℝ domψ (φ ∘ ψ)Algorithm 3.11
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Algorithm 3.11: the recursive outer constrained level method generated from the primitive stage-local data method. At master step k, the next state equals the current state if the stage already satisfies the global exact-value stopping rule; otherwise the method performs the continuing-step update by moving to the selected produced iterate and the selected real threshold t_{j(k)}^*(X_k). The public real parameter is read from the finite-threshold core recursion above.
universe u
variable {E : Type u} [PseudoMetricSpace E]
noncomputable def ConstrainedLevelMethod
(method : ConstrainedLevelMethodInput E)
(hrelative : method.RelativeStoppingExists)
(hfinite : method.SelectedThresholdFinite hrelative) :
ℕ → E × ℝ :=
fun k ↦
let statek := constrainedLevelMethodState method hrelative hfinite k
(statek.1, finiteThresholdReal statek.2)Definition 3.11
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.11 is a source-facing recall in the chapter's one-sided directional-derivative domain. Primary mathematical domain: - one-sided directional derivatives of EReal-valued functions on real modules. Sampled owner-style declarations: - HasDerivWithinAt - HasDirectionalDerivAt - DirectionallyDifferentiableAt - directionallyDifferentiableAt_iff_exists_hasDirectionalDerivAt Best owner abstraction: - HasDirectionalDerivAt Primitive data: - none in this file; the owner predicate lives upstream in Definition_3_1_3_1, and the earlier chapter recall surface already lives in Definition_3_1_3. Derived API: - the source-facing owner predicate HasDirectionalDerivAt - its existential wrapper DirectionallyDifferentiableAt - the thin specification theorem directionallyDifferentiableAt_iff_exists_hasDirectionalDerivAt Source/core/bridge triage: - source-facing: HasDirectionalDerivAt, DirectionallyDifferentiableAt - core/canonical: the earlier chapter recall Definition_3_1_3, backed by HasDerivWithinAt on the scalar slice - bridge/view: directionallyDifferentiableAt_iff_exists_hasDirectionalDerivAt Definition 3.11 adds no new mathematical data beyond the earlier chapter recall surface in Definition_3_1_3. This file therefore reuses that surface directly instead of restating the full signatures a second time.
recall HasDirectionalDerivAtLemma 3.11
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.11 lies in the chapter's extended-valued convex-analysis / affine subdifferential calculus on Euclidean spaces. Sampled owner declarations: - mathlib ConvexOn.comp_affineMap - ClosedConvexOn.comp_affineMap - IsSubgradientAt.comp_affineMap - subdifferential - mem_subdifferential_iff Best owner abstractions: - ClosedConvexOn for the closed-convex affine-pullback statement - IsSubgradientAt and subdifferential from Definition_3_1_5 for the affine subgradient calculus statement Primitive data: - the affine map g : Eₙ →ᵃ[ℝ] Eₘ - the ambient closed-convex hypothesis ClosedConvexOn S φ - the owner-level subgradient predicate IsSubgradientAt f (g x) h Derived API: - the recalled owner theorem ClosedConvexOn.comp_affineMap - the Euclidean bridge theorem subdifferential_comp_affineMap_image_adjoint_subset Source/core/bridge triage: - source-facing: Lemma 3.11's two affine-pullback consequences, namely closed convexity on g ⁻¹' S and the adjoint-image inclusion for the subdifferential - core/canonical: ClosedConvexOn, IsSubgradientAt, and subdifferential - bridge/view: ConvexOn.comp_affineMap, which underlies the closed-convex owner theorem, and the finite-dimensional set-level subdifferential inclusion derived from IsSubgradientAt.comp_affineMap This file therefore recalls the assumption-free closed-convex owner theorem directly and keeps only the thin affine subdifferential bridge theorem that specializes IsSubgradientAt.comp_affineMap to the source-facing subdifferential statement.
open scoped WithTopConvexAnalysis
recall ClosedConvexOn.comp_affineMap
{m n : ℕ}
{S : Set (EuclideanSpace ℝ (Fin m))}
{φ : EuclideanSpace ℝ (Fin m) → WithTop ℝ}
(hφ : ClosedConvexOn S φ)
(g : EuclideanSpace ℝ (Fin n) →ᵃ[ℝ] EuclideanSpace ℝ (Fin m)) :
ClosedConvexOn (g ⁻¹' S) (φ ∘ g)Proposition 3.11
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.11: if Q is a nonempty bounded subset of a real inner-product space E, then the domain dom ξ_Q of its support function is all of E; equivalently, ξ_Q(u) is finite for every u ∈ E. The textbook ℝⁿ statement is the specialization E = EuclideanSpace ℝ (Fin n).
universe u
variable {E : Type u} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
open scoped ConvexAnalysis SupportFunction
theorem supportFunction_dom_eq_univ_of_nonempty_bounded
(Q : Set E) (hQ_nonempty : Q.Nonempty) (hQ_bounded : Bornology.IsBounded Q) :
dom ξ[Q] = Set.univTheorem 3.11
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.11: if f₁ and f₂ are real-valued functions whose canonical WithTop lifts are closed and convex on a feasible set Q, and every constrained sublevel set of the pointwise maximum x ↦ max (f₁ x) (f₂ x) on Q is bounded, then there exists a parameter λ* ∈ [0, 1] for which the constrained minimum value of x ↦ max (f₁ x) (f₂ x) on Q equals that of the convex combination x ↦ λ* f₁ x + (1 - λ*) f₂ x; Lean records this conclusion by the owner predicate IsMinimaxLinearizationParameter on the subtype Q.
recall exists_minimax_parameter_of_bounded_constrainedSublevelSetsDefinition 3.12
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.12 (1): two sets are separable when some nonzero normal vector and offset define a nesterovHyperplane whose opposite closed half-spaces contain the two sets.
universe u
variable {E : Type u} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
local notation "HPlane" => AffineHyperplane E
def AreSeparable (Q₁ Q₂ : Set E) : Prop :=
∃ H : HPlane, H.SeparatesSets Q₁ Q₂Lemma 3.12
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.12 lies in the chapter's weighted-sum / subdifferential calculus for closed convex WithTop ℝ-valued functions on the intrinsic ambient spaces already used by the chapter owners. Relevant sampled declarations in this domain: - withTopEffectiveDomain_nonneg_weighted_add_eq_inter_of_pos from Lemma_3_1_12 - ClosedConvexFunction.nonneg_weighted_add from Lemma_3_1_12 - interior_effectiveDomain_nonneg_weighted_add_eq_of_pos from Lemma_3_1_12 - subdifferential_nonneg_weighted_add_eq_of_pos from Lemma_3_1_12 - withTopEffectiveDomain from Definition_3_3 - ClosedConvexFunction from Definition_3_1_1_5 - subdifferential from Definition_3_1_5 Best owner abstraction: - the canonical pointwise weighted sum ((α₁ : WithTop ℝ) • f₁ + (α₂ : WithTop ℝ) • f₂) - the ambient effective-domain, closed-convex, and subdifferential notions remain owned by the earlier chapter files sampled above - this file is therefore a bridge/view recalling the weighted-sum theorems stated directly on that canonical owner Primitive data: - none in this recall file; the imported owner surface already carries the only primitive mathematical inputs, namely the scalars α₁, α₂ and the summands f₁, f₂ Derived API: - withTopEffectiveDomain_nonneg_weighted_add_eq_inter_of_pos - ClosedConvexFunction.nonneg_weighted_add - interior_effectiveDomain_nonneg_weighted_add_eq_of_pos - subdifferential_nonneg_weighted_add_eq_of_pos Source/core/bridge triage: - source-facing: the three weighted-sum consequences in Lemma 3.12, now stated on the canonical pointwise weighted sum - core/canonical: withTopEffectiveDomain, ClosedConvexFunction, subdifferential - bridge/view: the positive-weight effective-domain identity and this recall-only file This bridge now recalls the weighted-sum conclusions directly on the canonical pointwise weighted sum. The incorrect auxiliary wrapper has been removed, so the exported surface no longer shrinks domains at zero weights.
recall withTopEffectiveDomain_nonneg_weighted_add_eq_inter_of_posProposition 3.12
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.12: for f(x) = (x)_+, the subdifferential at 0 is exactly the interval [0, 1].
open scoped WithTopConvexAnalysis
theorem real_posPart_subdifferential_at_zero_eq_Icc :
∂ (fun x : ℝ ↦ (x⁺ : ℝ))(0) = Set.Icc (0 : ℝ) 1
Theorem 3.12
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.12 is a recall-only item in the chapter's convex-analysis/minimax-linearization domain. Sampled owner-style declarations: - ClosedConvexOn - maxTypeObjective - constrainedSublevelSet - exists_stdSimplex_minimax_linearization_of_bounded_familyMaximumSublevelSets Best owner abstraction: - source-facing: the bounded-level-set minimax-linearization statement of Theorem 3.12 - core/canonical: maxTypeObjective fs, the bounded feasible sublevel owner constrainedSublevelSet Q (fun x ↦ ((maxTypeObjective fs x : ℝ) : WithTop ℝ)) α, and the simplex coefficient owner StdSimplex ℝ ι - bridge/view: the weighted-sum expression ∑ i, coeffs.weights i * fs i x, derived from the canonical simplex data Primitive data: - a nonempty finite family fs : ι → E → ℝ - closed convexity of each component on Q - boundedness of the constrained sublevel sets of maxTypeObjective fs Derived API: - the simplex coefficient vector coeffs : StdSimplex ℝ ι - the equality of constrained EReal infima between maxTypeObjective fs and the simplex-weighted objective The previous version introduced a second public theorem name with exactly the same interface as the upstream source-facing theorem exists_stdSimplex_minimax_linearization_of_bounded_familyMaximumSublevelSets. Since that theorem already has the correct statement and owner-level data, this file now recalls it directly instead of keeping a duplicate wrapper. In particular, the old file-level [FiniteDimensional ℝ E] assumption was redundant and is removed here because the canonical theorem does not use it.
recall exists_stdSimplex_minimax_linearization_of_bounded_familyMaximumSublevelSets
Definition 3.13
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.13: IsSubgradientAt f x0 g is the chapter's canonical predicate saying that g is a subgradient of the extended-real-valued function f at x0; the corresponding subdifferential ∂ f(x0) and constrained subdifferential ∂[Q] f(x0) are the derived sets of all such supporting vectors.
recall IsSubgradientAtLemma 3.13
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.13: for a nonempty finite family of closed convex WithTop ℝ-valued functions on a finite-dimensional real inner-product space, the canonical owner f := pointwiseSupremumOn (Set.univ : Set ι) φ is closed and convex, its effective-domain interior is the intersection of the slice interiors, and at each interior point its subdifferential is the convex hull of the subdifferentials of the active slices. Specializing V = EuclideanSpace ℝ (Fin n) recovers the textbook ℝⁿ statement.
open Set
open scoped Topology WithTopConvexAnalysis
universe u v
variable {ι : Type u}
variable {V : Type v} [SeminormedAddCommGroup V] [InnerProductSpace ℝ V]
theorem pointwiseSupremumOn_univ_closedConvex_interior_dom_subdifferential
[Finite ι] [Nonempty ι] [FiniteDimensional ℝ V] {φ : V → ι → WithTop ℝ}
(hφ : ∀ i, ClosedConvexFunction (fun y ↦ φ y i)) :
let f : V → WithTop ℝ := pointwiseSupremumOn (Set.univ : Set ι) φ
ClosedConvexFunction f ∧
interior (dom f) = ⋂ i : ι, interior (dom (fun y ↦ φ y i)) ∧
∀ ⦃x : V⦄, x ∈ interior (dom f) →
∂ f(x) =
convexHull ℝ
{g | ∃ i : ι,
i ∈ activePointwiseSupremumOnIndices (Set.univ : Set ι) φ x ∧
g ∈ ∂ (fun y ↦ φ y i)(x)}
Proposition 3.13
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.13: for f(x) = (x)_+ = max x 0, the subdifferential at 0 is exactly the interval [0, 1].
recall real_posPart_subdifferential_at_zero_eq_Icc
Theorem 3.13
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.13 is already the chapter's canonical source-facing ℓ₁ local regularity theorem, recorded upstream as ConvexOn.exists_l1_ball_subset_effectiveDomain_and_abs_sub_le_of_mem_interior in Nesterov.Chap03.Theorem_3_1_3_1. This file reuses that owner theorem directly and keeps only the bounded-image consequence as additional derived API.
open scoped Topology
open scoped WithTopConvexAnalysis
open EuclideanSpace
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
namespace ConvexOn
recall exists_l1_ball_subset_effectiveDomain_and_abs_sub_le_of_mem_interior
{f : E → WithTop ℝ}
(hf : ConvexOn ℝ (dom f) (withTopRealPart f))
{x0 : E} (hx0 : x0 ∈ interior (dom f)) :
∃ ε > 0, ∃ L > 0,
(l1Seminorm n).ball x0 ε ⊆ dom f ∧
∀ ⦃y : E⦄, y ∈ (l1Seminorm n).ball x0 ε →
|withTopRealPart f y - withTopRealPart f x0| ≤
L * l1Seminorm n (y - x0)Definition 3.14
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.14: the textbook real positive-part function x ↦ (x)_+ is the real specialization of the canonical owner posPart.
#check ((·⁺) : ℝ → ℝ)Lemma 3.14
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.14 [Chapter3_2.json:8] (1): if Δ is nonempty and every slice x ↦ φ(x, y) with y ∈ Δ is closed and convex on Q, then the pointwise supremum is closed and convex on its effective domain pointwiseSupremumOnEffectiveDomain Q Δ φ.
open Set
open scoped WithTopConvexAnalysis
universe u v
variable {ι : Type u} {X : Type v}
recall ClosedConvexOn.pointwise_sSup
[TopologicalSpace X] [AddCommMonoid X] [Module ℝ X]
{Q : Set X} {Δ : Set ι} {φ : X → ι → WithTop ℝ}
(hΔ : Δ.Nonempty)
(hφ : ∀ y ∈ Δ, ClosedConvexOn Q (fun x ↦ φ x y)) :
ClosedConvexOn (pointwiseSupremumOnEffectiveDomain Q Δ φ) (pointwiseSupremumOn Δ φ)Proposition 3.14
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.14: for f(x) = ∑_{i ∈ s} |⟪aᵢ, x⟫|, the subdifferential at x is the signed sum of the active vectors with positive and negative inner products, translated by the Minkowski sum of the symmetric line segments [-aᵢ, aᵢ] = segment ℝ (-aᵢ) aᵢ over the zero inner-product indices.
open Set
open scoped BigOperators Pointwise RealInnerProductSpace WithTopConvexAnalysis
universe u v
variable {ι : Type u} {V : Type v} [NormedAddCommGroup V] [InnerProductSpace ℝ V]
theorem subdifferential_sum_abs_inner_eq_signed_sum_add_zero_segments
(s : Finset ι) (a : ι → V) (x : V) :
∂ (fun y ↦ ((∑ i ∈ s, |⟪a i, y⟫|) : ℝ))(x) =
({(s.filter fun i ↦ 0 < ⟪a i, x⟫).sum a -
(s.filter fun i ↦ ⟪a i, x⟫ < 0).sum a} : Set V) +
(s.filter fun i ↦ ⟪a i, x⟫ = 0).sum (fun i ↦ segment ℝ (-a i) (a i))
Theorem 3.14
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.14 lies in the chapter's convex directional-derivative domain. Sampled owner-style declarations: - HasDirectionalDerivAt in Nesterov.Chap03.Definition_3_1_3_1, the source-facing owner for one-sided directional derivatives of extended-real-valued functions; - exists_tendsto_right_directionalSlope_of_convexOn_of_mem_interior_effectiveDomain in Nesterov.Chap03.Theorem_3_1_12, the canonical secant-slope existence theorem for convex WithTop ℝ-valued functions at interior points; - ConvexOn.slope_mono and bddBelow_slope_lt_of_mem_interior in mathlib Mathlib/Analysis/Convex/Deriv.lean, the one-variable slope monotonicity and interior-point boundedness lemmas underlying the owner theorem. Best owner abstraction: - source-facing: Theorem 3.14's one-sided directional-slope limit statement at an interior point of the effective domain; - core/canonical: the stronger chapter theorem exists_tendsto_right_directionalSlope_of_convexOn_of_mem_interior_effectiveDomain; - bridge/view: the EReal-valued limit presentation obtained by coercing the owner theorem's finite real limit. Primitive data: - a convexity witness hf : ConvexOn ℝ {y : E | f y < ⊤} (fun y ↦ (f y).untopD 0); - an interior point hx : x ∈ interior {y : E | f y < ⊤}. Derived API: - the eventual finiteness of the ray α ↦ x + α • p, already bundled in the owner theorem; - the finite real right limit of the secant slopes, also bundled in the owner theorem; - the weaker EReal-valued limit form stated by the textbook item. The previous file reintroduced a second public theorem name for a weaker consequence of the already-canonical chapter theorem in Theorem_3_1_12. Since the stronger owner theorem is the correct reusable API and there are no direct downstream uses of the duplicate local name, this file now recalls the owner theorem directly instead of preserving a parallel wrapper.
recall exists_tendsto_right_directionalSlope_of_convexOn_of_mem_interior_effectiveDomainDefinition 3.15
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.15 is a bridge/view item in the chapter's convex-analysis normal-cone API. Primary domain: - convex analysis of extended-real-valued functions on Euclidean space. Relevant owner-style declarations sampled before refinement: - extendedRealEffectiveDomain - the source-facing sublevel set {x ∈ dom f | f x ≤ f x0} - normalCone - level_set_inequality_at_iff Best owner abstraction: - normalCone Primitive data: - the finite-value domain extendedRealEffectiveDomain f - the level set {x ∈ dom f | f x ≤ f x0} Derived API: - the bridge theorem level_set_inequality_at_iff Source/core/bridge triage: - source-facing: the textbook level-set inequality at x0 - core/canonical: normalCone - bridge/view: level_set_inequality_at_iff The normal-cone and sublevel-set vocabulary is already owned upstream, so this file recalls only the numbered bridge statement rather than re-recalling that upstream surface.
recall level_set_inequality_at_iffLemma 3.15
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.15: for a positively 1-homogeneous function on a real inner-product space, the subdifferential at x is exactly the intersection of the origin subdifferential with the touching condition ⟪g, x⟫ = f x.
open scoped WithTopConvexAnalysis
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
theorem subdifferential_eq_subdifferential_zero_of_posHomogeneous
{f : E → ℝ} (hf_hom : IsPositivelyHomogeneousOn 1 Set.univ f) (x : E) :
∂ (fun y ↦ (f y : WithTop ℝ))(x) =
∂ (fun y ↦ (f y : WithTop ℝ))(0) ∩ {g | inner ℝ g x = f x}Proposition 3.15
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.15: the convex subdifferential of the coordinatewise maximum x ↦ max_{1 ≤ i ≤ n} x^{(i)} is the convex hull of the standard basis vectors indexed by the active coordinates of x.
open Set
open EuclideanSpace
open scoped WithTopConvexAnalysis
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
local notation "coordinateFamily" =>
(fun y : E ↦ fun i ↦ (y i : WithTop ℝ))
variable (hn : 0 < n)
theorem subdifferential_coordinatewiseMaximum_eq_convexHull_activeBasis
(hn : 0 < n)
(x : E) :
∂ (pointwiseSupremumOn (Set.univ : Set (Fin n)) coordinateFamily)(x) =
convexHull ℝ
((fun i : Fin n ↦ single i (1 : ℝ)) ''
activePointwiseSupremumOnIndices (Set.univ : Set (Fin n)) coordinateFamily x)
Theorem 3.15
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.15: if Q₁, Q₂ ⊆ ℝⁿ are nonempty closed convex sets with empty intersection, and at least one of them is bounded, then there exist a nonzero vector g and a scalar γ such that ⟪g, x⟫ < γ for every x ∈ Q₂ and γ < ⟪g, y⟫ for every y ∈ Q₁.
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
theorem areStronglySeparable_of_disjoint_closed_convex_of_one_bounded
(Q₁ Q₂ : Set E) (hQ₁_nonempty : Q₁.Nonempty) (hQ₂_nonempty : Q₂.Nonempty)
(hQ₁_closed : IsClosed Q₁) (hQ₂_closed : IsClosed Q₂)
(hQ₁_convex : Convex ℝ Q₁) (hQ₂_convex : Convex ℝ Q₂)
(hdisj : Disjoint Q₁ Q₂)
(hbounded : Bornology.IsBounded Q₁ ∨ Bornology.IsBounded Q₂) :
∃ (g : E) (_ : g ≠ 0) (γ : ℝ),
(∀ x ∈ Q₂, inner ℝ g x < γ) ∧
∀ y ∈ Q₁, γ < inner ℝ g yDefinition 3.16
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.16: for a nonempty closed convex set X ⊆ dom f, the epigraph facet of X with respect to f is the common regular subdifferential ∂̂ f(X) = ⋂ x ∈ X, ∂ f(x).
open scoped WithTopConvexAnalysis
universe u
recall commonRegularSubdifferential
{V : Type u} [SeminormedAddCommGroup V] [InnerProductSpace ℝ V]
(f : V → WithTop ℝ) (X : Set V) : Set VLemma 3.16
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.16 (2): for a bounded convex nonnegative weight set, the subdifferential of the WithTop bridge of the support-function composition x ↦ ξ[Λ] (vectorMap fs x) is the convex hull of the image of the closure active-multiplier face under the owner weighted-gradient map.
open Set
open scoped Gradient SupportFunction WithTopConvexAnalysis EuclideanOrthant
universe u
variable {m : ℕ} {E : Type u}
variable [NormedAddCommGroup E] [InnerProductSpace ℝ E] [CompleteSpace E]
local notation "Eₘ" => EuclideanSpace ℝ (Fin m)
theorem
subdifferential_supportFunction_comp_vectorMap_eq_convexHull_image_activeMultipliers_of_bounded
{Λ : Set Eₘ} {fs : Fin m → E → ℝ}
(hΛ_bounded : Bornology.IsBounded Λ) (hΛ_nonempty : Λ.Nonempty)
(hΛ_nonneg : Λ ⊆ ℝ₊^m) (hΛ_convex : Convex ℝ Λ)
(hfs_convex : ∀ i, ConvexOn ℝ Set.univ (fs i)) (x : E)
(hfs_grad : ∀ i, HasGradientAt (fs i) (∇ (fs i) x) x) :
∂ (supportFunctionCompWithTop Λ fs)(x) =
convexHull ℝ
((weightedGradientCombination fs x) ''
activeSupportFunctionMultipliers (closure Λ) fs x)
Proposition 3.16
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.16: the subdifferential of the norm is the closed unit ball at the origin and the singleton {‖x‖⁻¹ • x} at every nonzero point.
open scoped WithTopConvexAnalysis
variable {E : Type*} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
theorem subdifferential_norm_eq_closedBall_or_singleton
(x : E) :
∂ (fun y : E ↦ ‖y‖)(x) =
if x = 0 then
Metric.closedBall 0 1
else
{‖x‖⁻¹ • x}Theorem 3.16
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.16: if Q is a nonempty closed convex subset of a real inner-product space and x ∉ Q, then there exist a ≠ 0 and b : ℝ with ⟪a, x⟫ > b ≥ sup_{y ∈ Q} ⟪a, y⟫; equivalently, in the local Chapter 3 API, Q and the singleton {x} are strongly separable by a nesterovHyperplane.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E] [CompleteSpace E]
theorem areStronglySeparable_singleton_of_nonmem_closed_convex
(Q : Set E) (hQ_nonempty : Q.Nonempty) {x : E} (hQ_closed : IsClosed Q)
(hQ_convex : Convex ℝ Q) (hx : x ∉ Q) :
AreStronglySeparable Q ({x} : Set E)Definition 3.17
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.17 (Intersection of images of a set-valued mapping): the common-value set of all images 𝒮 x with x ∈ X.
universe u v
open Set
variable {α : Type u} {β : Type v}
variable (𝒮 : α → Set β) (X : Set α) (y : β)
def intersectionOfImages : Set β :=
⋂ x ∈ X, 𝒮 xLemma 3.17
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.17 (2): under the same hypotheses, the real-valued relative subdifferential ∂[Q] (F ∘ vectorMap fs) (x) equals the weighted sum of the coordinate relative subdifferentials ∑ i, g i • ∂[Q] (fs i) (x) for a gradient witness HasGradientAt F g (vectorMap fs x).
open scoped BigOperators Gradient Pointwise WithTopConvexAnalysis
universe u
variable {m : ℕ} {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
local notation "Y" => EuclideanSpace ℝ (Fin m)
variable {Q : Set E} {fs : Fin m → E → ℝ} {F : Y → ℝ}
variable (hF_convex : ConvexOn ℝ Set.univ F)
variable (hQ_convex : Convex ℝ Q)
variable (hfs_convex : ∀ i, ConvexOn ℝ Q (fs i))
variable (hF_mono : Monotone (F ∘ (EuclideanSpace.equiv (Fin m) ℝ).symm))
theorem subdifferentialWithin_comp_coordinatewiseMonotone_eq_weighted_sum
{x : E} (_hx : x ∈ interior Q) {g : Y}
(_hF_grad : HasGradientAt F g (vectorMap fs x)) :
∂[Q] (F ∘ vectorMap fs) (x) =
∑ i,
(g i) • ∂[Q] (fs i) (x)
Proposition 3.17
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.17: the subdifferential of the ℓ₁-norm on ℝⁿ is the signed sum of the active standard basis vectors, translated by the Minkowski sum of the symmetric line segments [-e_i, e_i] = segment ℝ (-e_i) e_i over the zero coordinates.
open scoped BigOperators Pointwise WithTopConvexAnalysis
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
local notation "e[" i "]" => EuclideanSpace.single i (1 : ℝ)
theorem subdifferential_l1Seminorm_eq_signedBasis_add_segmentSum
(x : E) :
∂ (fun y : E ↦ ((EuclideanSpace.l1Seminorm n) y : WithTop ℝ))(x) =
({(Finset.univ.filter fun i ↦ 0 < x i).sum (fun i ↦ e[i]) -
(Finset.univ.filter fun i ↦ x i < 0).sum (fun i ↦ e[i])} : Set E) +
(Finset.univ.filter fun i ↦ x i = 0).sum (fun i ↦ segment ℝ (-e[i]) e[i])
Theorem 3.17
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.17 (1): if Q₂ is a nonempty closed convex subset of a real Hilbert space and the support function of Q₁ is bounded above by that of Q₂ at every vector in dom ξ[Q₂], then Q₁ ⊆ Q₂.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E] [CompleteSpace E]
open scoped ConvexAnalysis SupportFunction
theorem subset_of_supportFunction_le_on_domain
(Q₁ Q₂ : Set E) (hQ₂_nonempty : Q₂.Nonempty)
(hQ₂_closed : IsClosed Q₂) (hQ₂_convex : Convex ℝ Q₂)
(hξ : ∀ g ∈ dom ξ[Q₂], ξ[Q₁] g ≤ ξ[Q₂] g) :
Q₁ ⊆ Q₂Definition 3.18
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.18, generalized from the textbook Euclidean setting: the polar set of Q consists of the vectors whose pairing with every point of Q is at most 1. Specializing to E = EuclideanSpace ℝ (Fin n) recovers the textbook Qᵒ ⊆ ℝⁿ.
universe u
variable {E : Type u} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
open scoped SupportFunction
def polarSet (Q : Set E) : Set E :=
{g | ξ[Q] g ≤ (1 : EReal)}Lemma 3.18
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.18 is a source-facing recall in the chapter's tangent-cone domain. Primary domain: - tangent cones of convex subsets of real normed spaces. Relevant owner-style declarations sampled before refinement: - mathlib posTangentConeAt, the ambient positive tangent-cone owner; - mathlib mem_posTangentConeAt_of_segment_subset, the canonical segment-based entry lemma for feasible displacements; - mathlib PointedCone.hull, the canonical cone-hull owner for feasible directions; - mathlib tangentConeAt NNReal, the underlying owner behind posTangentConeAt; - mathlib Set.vsub_singleton, the canonical displacement-set view. Best owner abstraction: - posTangentConeAt Q xBar - PointedCone.hull ℝ (Q -ᵥ ({xBar} : Set E)) Primitive data: - the feasible set Q; - the base point xBar. Derived API: - Q -ᵥ ({xBar} : Set E); - PointedCone.hull ℝ (Q -ᵥ ({xBar} : Set E)); - the owner-shaped closure characterization posTangentConeAt_eq_closure_pointedConeHull_vsub_singleton. Source/core/bridge triage: - source-facing: the textbook boundary-point statement that the tangent cone is the closure of the conical hull of feasible displacements; - core/canonical: mathlib's posTangentConeAt Q xBar; - core/canonical: mathlib's PointedCone.hull ℝ (Q -ᵥ ({xBar} : Set E)); - bridge/view: the direct displacement expression Q -ᵥ ({xBar} : Set E). The exact source-facing theorem is already owned upstream by Lemma_3_1_18.lean, exported in the atomic owner-shaped equality form with hypotheses Convex ℝ Q and xBar ∈ Q; it is stated for the positive tangent cone posTangentConeAt, whose underlying mathlib owner is tangentConeAt NNReal. The textbook closed boundary-point case is a direct specialization. This file therefore recalls the upstream owner theorem directly instead of staging local binders only to #check an applied term, and it maintains no parallel displacementSet/conicalHull wrappers or bouligandTangentCone layer.
recall posTangentConeAt_eq_closure_pointedConeHull_vsub_singletonProposition 3.18
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.18 (2): for a closed convex set Q containing 0 in a finite-dimensional real inner-product space, the subdifferential of the Minkowski functional at a finite point x is the active parameter set for the polar-pairing pointwise supremum over Qᵒ.
open scoped Pointwise PolarSet
open scoped MinkowskiFunctional
open scoped WithTopConvexAnalysis
universe u
variable {E : Type u} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
variable (Q : Set E)
local notation "polarPairingFamily" =>
(fun y : E ↦ fun g ↦ ((inner ℝ g y : ℝ) : WithTop ℝ))
variable {x : E}
theorem subdifferential_minkowskiFunctional_eq_activePolarIndices
[FiniteDimensional ℝ E]
(hQ_closed : IsClosed Q)
(hQ_convex : Convex ℝ Q)
(h0Q : (0 : E) ∈ Q)
(hx : x ∈ dom (ψ[Q])) :
∂ ψ[Q](x) =
activePointwiseSupremumOnIndices Qᵒ polarPairingFamily x
Theorem 3.18
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.18 lives in the chapter's nesterovHyperplane-support domain. Primary domain: - affine hyperplanes and supporting hyperplanes for closed convex sets in finite-dimensional real inner-product spaces. Relevant sampled declarations: - AffineHyperplane - IsSupportingHyperplane - Set.IsExposed - exists_supporting_hyperplane_at_boundary_point_of_closed_convex Source-facing layer: - existence of a supporting nesterovHyperplane through a boundary point of a closed convex set. Core/canonical owner: - AffineHyperplane from Definition_3_1_4_1, whose primitive data are the nonzero normal vector and offset. Bridge/view: - nesterovHyperplane g γ and IsSupportingHyperplane Q g γ, which spell the owner object in coordinate form. The current file's former theorem was a third copy of the same source-facing result already recorded upstream as exists_supporting_hyperplane_at_boundary_point_of_closed_convex. The extra hypothesis Q.Nonempty is redundant here, since x₀ ∈ frontier Q implies x₀ ∈ closure Q, while closure ∅ = ∅. This file therefore reuses the earlier theorem directly instead of keeping a parallel local shell with a strictly longer hypothesis list.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
recall exists_supporting_hyperplane_at_boundary_point_of_closed_convex
[FiniteDimensional ℝ E] (Q : Set E) (hQ_closed : IsClosed Q) (hQ_convex : Convex ℝ Q)
{x₀ : E}
(hx₀ : x₀ ∈ frontier Q) :
∃ g : E, ∃ γ : ℝ, x₀ ∈ nesterovHyperplane g γ ∧ IsSupportingHyperplane Q g γDefinition 3.19
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.19: an ℝ ∪ {+∞}-valued function is positively homogeneous of degree p when its effective domain is closed under nonnegative scaling and its finite real part scales with degree p on that domain.
open scoped WithTopConvexAnalysis
universe u
variable {E : Type u} [SMul NNReal E]
variable (p : ℝ) (f : E → WithTop ℝ)
abbrev IsPositivelyHomogeneous (p : ℝ) (f : E → WithTop ℝ) : Prop :=
IsPositivelyHomogeneousOn p (dom f) (withTopRealPart f)Lemma 3.19
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.19: if xStar minimizes f on the convex feasible set Q, then the intrinsic within-set directional-derivative map lineDerivWithin ℝ f Q xStar is nonnegative on the tangent cone 𝒯[Q] xStar.
open scoped Pointwise TangentCone Topology
universe u
variable {E : Type u} [NormedAddCommGroup E] [NormedSpace ℝ E]
theorem directionalDerivative_nonneg_on_posTangentCone_of_isMinOn
{Q : Set E} {f : E → ℝ} {xStar : E}
(hxStar : xStar ∈ Q) (hQ_convex : Convex ℝ Q) (hmin : IsMinOn f Q xStar)
(hline_cont : Continuous (lineDerivWithin ℝ f Q xStar))
(hline_hom : IsPositivelyHomogeneousOn 1 Set.univ (lineDerivWithin ℝ f Q xStar))
(hline :
∀ d : E,
(∀ᶠ β : ℝ in 𝓝[>] (0 : ℝ), xStar + β • d ∈ Q) →
LineDifferentiableWithinAt ℝ f Q xStar d)
(p : E) (hp : p ∈ 𝒯[Q] xStar) :
0 ≤ lineDerivWithin ℝ f Q xStar pProposition 3.19
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.19: for a convex positively 1-homogeneous function on a finite-dimensional real inner-product space, the value at x is the maximum of the pairings ⟪g, x⟫ over all subgradients g ∈ ∂f(0), expressed as an IsGreatest statement for the image of the origin subdifferential.
open scoped WithTopConvexAnalysis
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
theorem isGreatest_pairing_image_subdifferential_zero_of_convex_posHomogeneous
{f : E → ℝ} (hf_convex : ConvexOn ℝ Set.univ f)
(hf_hom : IsPositivelyHomogeneousOn 1 Set.univ f)
[FiniteDimensional ℝ E] (x : E) :
IsGreatest ((fun g : E ↦ inner ℝ g x) '' ∂ (fun y ↦ (f y : WithTop ℝ))(0)) (f x)Theorem 3.19
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.19 is a recall-only surface in the chapter's extended-valued convex-subdifferential domain. Primary domain: - convex analysis of ℝ ∪ {+∞}-valued functions on Euclidean space. Sampled owner-style declarations: - dom and withTopRealPart in Definition_3_3, the chapter owners for the effective domain and finite-value representative; - IsSubgradientAt and subdifferential in Definition_3_1_5, the chapter owners for extended-valued subgradients; - subdifferential_nonempty_and_isBounded_of_convexOn_effectiveDomain_of_mem_interior in Theorem_3_1_15, the canonical chapter theorem on that owner surface. Best owner abstraction: - subdifferential_nonempty_and_isBounded_of_convexOn_effectiveDomain_of_mem_interior. Primitive data: - none in this file; the theorem already lives upstream on the canonical chapter owners. Derived API: - this recall-only source-facing entry point. Source/core/bridge triage: - source-facing: Theorem 3.19's nonempty-and-bounded subdifferential statement; - core/canonical: the upstream theorem in Theorem_3_1_15 on dom, withTopRealPart, and subdifferential; - bridge/view: this recall surface. The previous file rebuilt local copies of the effective domain, subgradient predicate, and subdifferential set even though those notions are already owned upstream, and it duplicated the same theorem already present in Theorem_3_1_15. This refinement removes that parallel local API and reuses the canonical chapter theorem directly.
recall subdifferential_nonempty_and_isBounded_of_convexOn_effectiveDomain_of_mem_interiorDefinition 3.20
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.20: the dual norm attached to a norm on ℝⁿ is recalled through the canonical Chapter 2 owner Seminorm.dualNorm, whose value at g is the maximum of ⟪g, x⟫ over the closed unit ball of the primal norm.
universe u
open scoped SeminormDualNorm
namespace Seminorm
recall dualNorm
{E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
(p : Seminorm ℝ E) [FiniteDimensional ℝ E] [p.IsNorm] :
E → ℝLemma 3.20
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.20: for a symmetric positive-definite matrix A, the linear functional x ↦ ⟪c, x⟫ attains its maximum on the centered ellipsoid affineEllipsoid A⁻¹ 0, equivalently on {x | ⟪A x, x⟫ ≤ 1}, and that maximum is √⟪c, A⁻¹ c⟫.
open Module LinearMap
open scoped BInducedNorm
open scoped InnerProduct
open scoped EllipsoidNotation
open scoped SeminormDualNorm
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
local notation "Mat" => Matrix (Fin n) (Fin n) ℝ
theorem isGreatest_inner_image_spdEllipsoid
(A : Mat) (hA : A.PosDef) (c : E) :
IsGreatest ((fun x : E ↦ inner ℝ c x) '' E(A⁻¹, (0 : E)))
(Real.sqrt (inner ℝ c ((A⁻¹).toEuclideanLin c)))Proposition 3.20
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.20 (2), intrinsic owner form: the subdifferential of the seminorm at the origin consists exactly of the vectors whose pairing with every y is bounded by p y.
open scoped SupportFunction
open scoped SeminormDualNorm
open scoped WithTopConvexAnalysis
universe u
open Seminorm
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable (p : Seminorm ℝ E)
theorem subdifferential_seminorm_at_zero_eq_inner_le :
∂ (fun y : E ↦ (p y : WithTop ℝ))(0) = {g | ∀ y : E, inner ℝ g y ≤ p y}Theorem 3.20
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.20 (1): in the Euclidean specialization E = EuclideanSpace ℝ (Fin n), the Fenchel bidual is bounded above by the original value at every point of dom f.
open scoped ConvexAnalysis WithTopConvexAnalysis
variable {n : ℕ}
variable {f : EuclideanSpace ℝ (Fin n) → WithTop ℝ}
variable {x : EuclideanSpace ℝ (Fin n)}
#check
(fenchelBidual_le_of_mem_dom :
x ∈ dom f → (f⋆⋆) x ≤ withTopToEReal (f x))Definition 3.21
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.21, generalized from the textbook ℝⁿ setting: a composite convex minimization problem consists of a closed feasible set Q, a smooth convex part f that belongs to ConvexC1On Q, and a closed convex term Ψ on Q, representing the objective x ↦ f x + Ψ x minimized over Q. The ambient smooth pair (Q, f) is owned canonically by SetConstrainedMinimizationProblem, the smooth regularity/convexity package by ConvexC1On, and the nonsmooth term by ClosedConvexOn. The textbook ℝⁿ case is the specialization CompositeConvexMinimizationProblem (EuclideanSpace ℝ (Fin n)).
universe u
variable {X : Type u}
structure CompositeConvexMinimizationProblem (E : Type u) [NormedAddCommGroup E] [NormedSpace ℝ E]
extends SetConstrainedMinimizationProblem E where
/-- The feasible set `Q` is closed. -/
feasibleSet_closed : IsClosed feasibleSet
/-- The smooth term belongs to the canonical Chapter 2 owner `ConvexC1On Q`. -/
smoothPart_convexC1 : ConvexC1On feasibleSet objective
/-- The closed convex term `Ψ : ℝⁿ → ℝ ∪ {+∞}`. -/
nonsmoothPart : E → WithTop ℝ
/-- The nonsmooth term is closed and convex on `Q`. -/
nonsmoothPart_closedConvex : ClosedConvexOn feasibleSet nonsmoothPartLemma 3.21
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.21: if xStar minimizes the Lagrangian associated to lambdaStar and complementary slackness holds at xStar, then the objective gap at xBar dominates the multiplier-weighted constraint violation.
open scoped BigOperators
universe u
variable {Q : Type u} {m : ℕ}
local notation "Λ" => EuclideanSpace ℝ (Fin m)
lemma objective_gap_ge_weighted_constraint_violation_of_lagrangian_minimizer
(problem : LagrangianProblem Q m) {xStar xBar : Q} {lambdaStar : Λ}
(h_lagrangian_min : xStar ∈ problem.lagrangianMinimizers lambdaStar)
(h_complementary_slackness :
∀ i : Fin m, lambdaStar i * problem.constraints i xStar = 0)
:
problem xBar - problem xStar ≥
∑ i, (-(problem.constraints i xBar)) * lambdaStar iProposition 3.21
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.21: the log-sum-exp of a nonempty finite family of convex functions is convex on their common domain.
open scoped BigOperators
variable {ι E : Type*}
variable [AddCommMonoid E] [Module ℝ E]
theorem convexOn_log_sum_exp_of_convexOn
(s : Set E) {t : Finset ι} {f : ι → E → ℝ}
(ht : t.Nonempty) (hf : ∀ i ∈ t, ConvexOn ℝ s (f i)) :
ConvexOn ℝ s (fun x ↦ Real.log (∑ i ∈ t, Real.exp (f i x)))Theorem 3.21
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.21 (1): if f is a convex proper extended-real-valued function on a real topological module and x₀ lies in the interior of its effective domain, then the directional derivative p ↦ f'(x₀; p) is a finite convex function on all directions.
open Filter
open scoped WithTopConvexAnalysis
open scoped Topology
universe u
variable {E : Type u} [TopologicalSpace E] [AddCommGroup E] [Module ℝ E]
[IsTopologicalAddGroup E] [ContinuousSMul ℝ E]
theorem convexDirectionalDerivativeReal_convexOn_univ_of_mem_interior
{f : E → WithTop ℝ} (hf : ConvexOn ℝ (dom f) (withTopRealPart f))
{x0 : E} (hx0 : x0 ∈ interior (dom f)) :
ConvexOn ℝ Set.univ (f′[hx0])Definition 3.22
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.22: the normal cone to Q at xBar, realized as the inner dual cone of the displacement set Q - xBar.
open ProperCone
universe u
variable {V : Type u} [NormedAddCommGroup V] [InnerProductSpace ℝ V] [CompleteSpace V]
abbrev normalCone (Q : Set V) (xBar : V) : ProperCone ℝ V :=
innerDual (Q -ᵥ ({xBar} : Set V))Lemma 3.22
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.22: if every primal slice x ↦ Ψ(x, u) is closed and convex on P with bounded real sublevel sets, every dual slice u ↦ Ψ(x, u) is closed and concave on S, and uStar is a dual-feasible maximizer of the lower value u ↦ sInf ((fun x ↦ Ψ x u) '' P) on S, then for every u ∈ S the two-slice maximum x ↦ max (Ψ(x, u)) (Ψ(x, uStar)) has a feasible minimizer xBar ∈ P, and that minimum value is sInf ((fun x ↦ Ψ x uStar) '' P). The convexity of S is recovered from the dual closed-concavity owner by ClosedConvexOn.convex. This owner-level statement lives on the same proper real normed-space layer as the chapter's minimax-linearization existence theorem.
open scoped WithTopConvexAnalysis
universe u v
variable {X : Type u} {U : Type v}
variable [NormedAddCommGroup X] [NormedSpace ℝ X] [ProperSpace X]
variable [TopologicalSpace U] [AddCommMonoid U] [Module ℝ U]
theorem exists_isMinOn_parametricMaximumObjective_eq_valueFunction_of_valueFunction_maximizer
{P : Set X} (hP_nonempty : P.Nonempty)
{S : Set U} {Ψ : X → U → ℝ}
(hΨ_primal_closedConvex :
∀ ⦃u : U⦄, u ∈ S → ClosedConvexOn P (fun x ↦ (Ψ x u : WithTop ℝ)))
(hlevel_bounded :
∀ ⦃u : U⦄, u ∈ S → ∀ α : ℝ,
Bornology.IsBounded
(constrainedSublevelSet P (fun x ↦ (Ψ x u : WithTop ℝ)) α))
(hΨ_dual_closedConcave :
∀ ⦃x : X⦄, x ∈ P →
ClosedConvexOn S (fun v ↦ (-Ψ x v : WithTop ℝ)))
{uStar u : U}
(huStar_mem : uStar ∈ S)
(huStar_max : IsMaxOn (fun u ↦ sInf ((fun x ↦ Ψ x u) '' P)) S uStar)
(hu : u ∈ S) :
∃ xBar ∈ P,
IsMinOn (fun x ↦ max (Ψ x u) (Ψ x uStar)) P xBar ∧
max (Ψ xBar u) (Ψ xBar uStar) = sInf ((fun x ↦ Ψ x uStar) '' P)Proposition 3.22
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.22 (1), intrinsic owner-level form: at a feasible point of the affine level set {x | L x = b} of a continuous linear constraint map, the textbook tangent cone is exactly ker L.
open Matrix
open scoped NormalCone
open scoped TangentCone
open scoped Topology
variable {E F : Type*}
variable [NormedAddCommGroup E] [NormedSpace ℝ E]
variable [NormedAddCommGroup F] [NormedSpace ℝ F]
theorem posTangentConeAt_linearLevelSet
(L : E →ₗ[ℝ] F) (hL : Continuous L) (b : F) {xBar : E} (hxBar : L xBar = b) :
𝒯[{x | L x = b}] xBar =
L.kerTheorem 3.22
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.22: every subgradient g ∈ ∂f(x₀) is a supporting vector to the level set {x | f x ≤ f x₀} at x₀, in the sense that ⟪g, x - x₀⟫ ≤ 0 for every point of that sublevel set. Specializing E to EuclideanSpace ℝ (Fin n) recovers the textbook statement on ℝⁿ.
open scoped WithTopConvexAnalysis
universe u
variable {E : Type u} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
theorem subgradient_nonpos_on_sublevelSet_of_mem_subdifferential
{f : E → WithTop ℝ} {x0 g : E} (hg : g ∈ ∂ f(x0)) {x : E} (hx : f x ≤ f x0) :
inner ℝ g (x - x0) ≤ 0Definition 3.23
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.23: for a convex set Q in a complete real inner-product space and xBar ∈ Q, membership in the textbook tangent cone 𝒯[Q] xBar is exactly the polarity condition against every normal vector g ∈ N[Q] xBar. The notation 𝒯[Q] xBar is the source-facing surface for the canonical owner posTangentConeAt Q xBar, and the textbook ℝⁿ statement is the corresponding specialization.
open ProperCone
open scoped Pointwise Topology
open scoped NormalCone
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E] [CompleteSpace E]
open scoped TangentCone
theorem mem_tangentCone_iff {Q : Set E} {xBar p : E}
(hQ_convex : Convex ℝ Q) (hxBar : xBar ∈ Q) :
p ∈ 𝒯[Q] xBar ↔
∀ g ∈ N[Q] xBar, 0 ≤ inner ℝ g pLemma 3.23
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.23: if x_N satisfies the residual estimate against the textbook affine model l_N = ∑_{k=0}^N α_k [Ψ(y_k, u(y_k)) + ⟪g(y_k), · - y_k⟫] built from the sampled selections u(y_k) and g(y_k) ∈ ∂[P] (fun x ↦ Ψ x (u(y_k))) (y_k), then the raw primal-dual gap against the canonical lower value at the averaged sampled dual point \hat u_N = (Finset.univ).centerMass α (u ∘ y) is bounded above by r_N.
open Finset
open scoped BigOperators WithTopConvexAnalysis
universe uX uU
variable {N : ℕ}
variable {X : Type uX} [SeminormedAddCommGroup X] [InnerProductSpace ℝ X]
variable {U : Type uU} [AddCommGroup U] [Module ℝ U]
variable {P : Set X} {S : Set U}
variable {f : X → ℝ}
variable {Ψ : X → U → ℝ}
variable {α : Fin (N + 1) → ℝ} {y : Fin (N + 1) → X}
variable {xN : X}
variable {modelMin rN fStar φStar : ℝ}
local notation "lowerValue" => fun u' : U ↦ sInf ((fun x ↦ Ψ x u') '' P)
variable {u : X → U} {g : X → X}
local notation "uHatN" => Finset.univ.centerMass α (u ∘ y)
local notation "samplePoint[" k "]" => y k
local notation "sampledSlice[" k "]" => (((fun x ↦ Ψ x (u (y k))) : X → ℝ))
local notation "lN" =>
(∑ k, α k • sampledAffineMinorant (y k) (g (y k)) (Ψ (y k) (u (y k))) : X →ᵃ[ℝ] ℝ)
theorem duality_gap_le_residual_of_averaged_affine_lower_model
(hα_nonneg : ∀ k, 0 ≤ α k)
(hα_sum_one : ∑ k, α k = 1)
(hu_mem : ∀ k, u (y k) ∈ S)
(h_subgrad :
∀ k : Fin (N + 1), g (y k) ∈ ∂[P] (sampledSlice[k]) (samplePoint[k]))
(h_concave_u : ∀ x ∈ P, ConcaveOn ℝ S (fun v ↦ Ψ x v))
(h_modelMin : IsLeast (lN '' P) modelMin)
(h_residual : f xN - modelMin ≤ rN) :
f xN - lowerValue uHatN ≤ rNProposition 3.23
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.23, affine normal-cone form: for a convex objective on a real inner-product space, a feasible point minimizes f on the affine level set {x | L x = b} exactly when some subgradient at that point lies in the normal cone of the level set.
open Matrix
open scoped ConstrainedArgmin
open scoped NormalCone
open scoped WithTopConvexAnalysis
variable {E Λ : Type*}
variable [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable [NormedAddCommGroup Λ] [InnerProductSpace ℝ Λ]
theorem isMinOn_linearLevelSet_iff_exists_subgradient_mem_normalCone
[FiniteDimensional ℝ E]
{f : E → ℝ} (hf_conv : ConvexOn ℝ Set.univ f)
(L : E →ₗ[ℝ] Λ) (b : Λ) {xStar : E}
(hxStar : L xStar = b) :
IsMinOn f {x | L x = b} xStar ↔
∃ gStar : E,
gStar ∈ ∂ (fun x : E ↦ (f x : WithTop ℝ))(xStar) ∧
gStar ∈ N[{x | L x = b}] xStarTheorem 3.23
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.23 lies in the chapter's extended-valued convex-analysis / subdifferential domain. Relevant owner-style declarations sampled before refinement: - IsSubgradientAt - subdifferential - subgradient_nonneg_on_sublevelSet_of_mem_subdifferential - subgradient_inner_sub_nonneg_of_isMinOn Best owner abstraction: - the intrinsic subdifferential owner subdifferential together with the minimizer theorem subgradient_inner_sub_nonneg_of_isMinOn Primitive data: - a feasible set Q, an extended-real objective f, points x0, xStar, g - the feasibility hypothesis hx0 : x0 ∈ Q - the minimizing hypothesis hxStar : IsMinOn f Q xStar - the owner-membership hypothesis hg : g ∈ subdifferential f x0 Derived API: - the nonnegative pairing inequality 0 ≤ inner ℝ g (x0 - xStar) Source/core/bridge triage: - source-facing: this textbook minimizer-pairing theorem - core/canonical: subdifferential, IsSubgradientAt, IsMinOn, and the theorem subgradient_inner_sub_nonneg_of_isMinOn - bridge/view: the former Euclidean/closed/convex wrapper, now deleted because its extra hypotheses were redundant and created a duplicate theorem surface The earlier version of this file reintroduced a second public theorem with the same name and mathematical content, but on an over-concrete Euclidean wrapper API carrying unused convexity, closedness, nonemptiness, and domain hypotheses. The intrinsic owner theorem already exists in Theorem_3_1_5_6 with the exact source-facing conclusion, so this item is now a direct recall instead of a parallel duplicate.
recall subgradient_inner_sub_nonneg_of_isMinOnDefinition 3.24
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.24: the set of minimizers X^* of f on the feasible set Q is the canonical constrained argmin set argmin[Q] f.
open scoped ConstrainedArgmin
universe u v
variable {X : Type u} (Q : Set X) (f : X → ℝ) (x : X)
abbrev minimizerSet : Set X :=
argmin[Q] fLemma 3.24
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.24: if the weighted primal-dual gap is controlled by the attained certificate maximum max_{x ∈ P} δ_N(x) and these maxima are bounded above by r_N, then for every stage N the source decomposition (\hat f_N - f^*) + (\phi^* - inf_{x ∈ P} ψ(x, \hat u_N)) with \hat f_N = dotProduct (α N) (fun k ↦ ψ (y N k) (u N k)) and \hat u_N = (Finset.univ).centerMass (α N) (u N) lies in the interval [0, \hat f_N - inf_{x ∈ P} ψ(x, \hat u_N)], and that gap is bounded above by r_N.
open scoped BigOperators
open Filter
universe u v
variable {E : Type u} {U : Type v}
variable [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable [AddCommGroup U] [Module ℝ U]
variable {P : Set E} {ψ : E → U → ℝ} {g : E → E}
variable (y : ∀ N : ℕ, Fin (N + 1) → E)
variable (u : ∀ N : ℕ, Fin (N + 1) → U)
variable (α : ∀ N : ℕ, Fin (N + 1) → ℝ)
variable (r : ℕ → ℝ)
local notation "hatf" =>
fun N ↦ dotProduct (α N) (fun k ↦ ψ (y N k) (u N k))
local notation "uHat" =>
fun N ↦ Finset.univ.centerMass (α N) (u N)
local notation "lowerValue" =>
fun u' : U ↦ sInf ((fun x ↦ ψ x u') '' P)
local notation "gap" =>
fun N ↦ hatf N - lowerValue (uHat N)
local notation "δ" =>
fun N ↦ gapFunctionCertificate (y N) (α N) g
variable
(hP_nonempty : P.Nonempty)
(hlower_bddBelow :
∀ N : ℕ, BddBelow ((fun x ↦ ψ x ((Finset.univ).centerMass (α N) (u N))) '' P))
(hα_nonneg : ∀ N : ℕ, ∀ k, 0 ≤ α N k)
(hα_sum_one : ∀ N : ℕ, ∑ k, α N k = 1)
(hsubgradient_uHat :
∀ N : ℕ, ∀ (k : Fin (N + 1)) (x : E) (_hx : x ∈ P),
ψ (y N k) ((Finset.univ).centerMass (α N) (u N)) -
ψ x ((Finset.univ).centerMass (α N) (u N)) ≤
inner ℝ (g (y N k)) (y N k - x))
(haggregate :
∀ N : ℕ,
dotProduct (α N) (fun k ↦ ψ (y N k) (u N k)) ≤
∑ k, α N k * ψ (y N k) ((Finset.univ).centerMass (α N) (u N)))
include hP_nonempty hlower_bddBelow hα_nonneg hα_sum_one hsubgradient_uHat haggregate
variable (fStar φStar : ℝ)
variable
(h_primal_lower :
∀ N : ℕ, fStar ≤ dotProduct (α N) (fun k ↦ ψ (y N k) (u N k)))
(h_dual_upper_uHat :
∀ N : ℕ,
sInf ((fun x ↦ ψ x ((Finset.univ).centerMass (α N) (u N))) '' P) ≤ φStar)
(h_weak_duality : φStar ≤ fStar)
include h_primal_lower h_dual_upper_uHat h_weak_duality
local notation "decomposition" =>
fun N ↦ (hatf N - fStar) + (φStar - lowerValue (uHat N))
theorem primal_dual_decomposition_mem_Icc_and_gap_le_of_gapFunctionCertificate_max
(δMax : ℕ → ℝ)
(hδMax : ∀ N, IsGreatest (δ N '' P) (δMax N))
(hδMax_le : ∀ N, δMax N ≤ r N) :
∀ N,
decomposition N ∈ Set.Icc 0 (gap N) ∧ gap N ≤ r NProposition 3.24
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.24: encode a function on Q₁ by an ambient extension f with dom f ⊆ Q₁. If Q₁ is convex, f is proper, and every constrained subdifferential ∂[Q₁] f(x) on Q₁ is presented as convexHull ℝ (Y x) for a nonempty family Y, then the finite real part of f is convex and lower semicontinuous on Q₁.
open scoped WithTopConvexAnalysis
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
theorem closedConvexFunction_of_subdifferential_eq_convexHull
{Q₁ : Set E} {f : E → WithTop ℝ}
(Y : E → Set E)
(hQ₁_convex : Convex ℝ Q₁)
(hdom : dom f ⊆ Q₁)
(hproper : (dom f).Nonempty)
(hY_nonempty : ∀ x ∈ Q₁, (Y x).Nonempty)
(hsubdiff : ∀ x ∈ Q₁, ∂[Q₁] f(x) = convexHull ℝ (Y x)) :
ConvexOn ℝ Q₁ (withTopRealPart f) ∧
LowerSemicontinuousOn (withTopRealPart f) Q₁Theorem 3.24
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.24 (1): every common subgradient gives the affine increment formula on X.
recall eq_add_inner_of_mem_commonRegularSubdifferentialDefinition 3.25
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.25: a smooth minimization problem with functional constraints is a smooth minimization problem whose constraints are all of the form fⱼ(x) ≤ 0 and whose basic feasible set Q is closed and convex.
namespace GeneralMinimizationProblem
variable {n m : ℕ}
variable (problem : GeneralMinimizationProblem n m)
#check (
problem.IsSmooth ∧
problem.HasLeConstraints ∧
IsClosed problem.basicFeasibleSet ∧
Convex ℝ problem.basicFeasibleSet
)Lemma 3.25
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.25 lies in the chapter's subgradient-localization / pointwise-growth domain. Primary mathematical domain: - real-valued convex analysis on ℝⁿ, organized around subgradient-based localization radii and growth functions. Sampled owner-style declarations: - IsSubgradientAt in Definition_3_1_5, the chapter owner predicate for subgradients via the canonical WithTop-valued formulation; - subgradientLocalizationMeasure in Lemma_3_2_1, with source-facing surface v[g; xBar] x for the localization radius attached to a chosen subgradient selection; - pointwiseGrowthFunction in Lemma_3_2_1, with source-facing surface ω[f; xBar] t for the supremal ball-growth profile; - sub_le_lipschitz_mul_max_localizationMeasure in Lemma_3_2_1, the owner theorem for the Lipschitz growth bound. Best owner abstraction: - subgradientLocalizationMeasure - pointwiseGrowthFunction - the owner comparison theorems built from them in Lemma_3_2_1 Primitive data: - a real-valued function f : V → ℝ on a real inner product space; - a chosen subgradient selection g : V → V; - the base point xBar and evaluation point x. Derived API: - the growth comparison sub_le_pointwiseGrowthFunction_of_localizationMeasure; - the Lipschitz refinement sub_le_lipschitz_mul_max_localizationMeasure. Source/core/bridge triage: - source-facing: Lemma 3.25's comparison of f x - f xBar with the canonical growth function and its Lipschitz specialization; - core/canonical: IsSubgradientAt, subgradientLocalizationMeasure, and pointwiseGrowthFunction; - bridge/view: none beyond the coercion fun y ↦ (f y : WithTop ℝ) already absorbed by the owner theorems. The previous version duplicated a real-valued subgradient predicate, its subdifferential, the localization measure, the growth function, and two theorem wrappers carrying an unused convexity hypothesis. The chapter already centers this domain on Lemma_3_2_1, so this file now recalls the owner theorems directly instead of keeping a second local API surface.
recall sub_le_pointwiseGrowthFunction_of_localizationMeasureProposition 3.25
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.25: for a nonempty compact convex nonnegative weight set, the subdifferential of the thin WithTop bridge of the support-function composition x ↦ ξ[Λ] (vectorMap fs x) at x is the convex hull of the image of the active-multiplier face under weightedGradientCombination fs x. This is the representation (3.1.80) at x.
open Set
open scoped Gradient WithTopConvexAnalysis EuclideanOrthant
universe u
variable {m : ℕ} {E : Type u}
variable [NormedAddCommGroup E] [InnerProductSpace ℝ E] [CompleteSpace E]
theorem subdifferential_supportFunction_comp_vectorMap_eq_convexHull_image_activeMultipliers
{Λ : Set (EuclideanSpace ℝ (Fin m))} {fs : Fin m → E → ℝ}
(hΛ_nonempty : Λ.Nonempty) (hΛ_compact : IsCompact Λ)
(hΛ_convex : Convex ℝ Λ) (hΛ_nonneg : Λ ⊆ ℝ₊^m)
(hfs_convex : ∀ i, ConvexOn ℝ Set.univ (fs i))
(x : E) (hfs_grad : ∀ i, HasGradientAt (fs i) (∇ (fs i) x) x) :
∂ (supportFunctionCompWithTop Λ fs)(x) =
convexHull ℝ
((weightedGradientCombination fs x) '' activeSupportFunctionMultipliers Λ fs x)
Theorem 3.25
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.25: a set is contained in arg min_{x ∈ dom f} f(x) if and only if the zero vector lies in the common regular subdifferential on that set.
open scoped ConstrainedArgmin WithTopConvexAnalysis
universe u
variable {V : Type u} [SeminormedAddCommGroup V] [InnerProductSpace ℝ V]
recall subset_constrainedArgmin_effectiveDomain_iff_zero_mem_commonRegularSubdifferential
{f : V → WithTop ℝ} {XStar : Set V} :
XStar ⊆ argmin[dom f] f ↔ (0 : V) ∈ ∂̂ f(XStar)Definition 3.26
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.26: a vector λ* is a vector of optimal dual (Lagrange) multipliers if it is nonnegative and maximizes the dual function on ℝ₊^m.
universe u
namespace LagrangianProblem
variable {Q : Type u} {m : ℕ}
local notation "Λ" => EuclideanSpace ℝ (Fin m)
open scoped EuclideanOrthant
def IsOptimalDualMultiplier (problem : LagrangianProblem Q m) (lamStar : Λ) : Prop :=
lamStar ∈ ℝ₊^m ∧
IsMaxOn problem.dualFunction (ℝ₊^m) lamStarLemma 3.26
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.26: for any reference point xStar ∈ Q, the best objective-value gap above f xStar among the sample points x₀, …, x_k is bounded by the pointwise growth function of the restricted objective, evaluated at the best sampled distance to xStar. The growth function is WithTop ℝ-valued so that unbounded ball-growth is represented by ⊤. In particular, this recovers the textbook minimizer formulation when xStar is optimal on Q.
open scoped PointwiseGrowthFunction
universe u
variable {X : Type u} [MetricSpace X] {Q : Set X}
theorem bestFunctionValueUpTo_sub_le_pointwiseGrowthFunction_at_bestRadius
(f : ↥Q → ℝ) (xSeq : ℕ → ↥Q) (xStar : ↥Q) (k : ℕ) :
bestFunctionValueUpTo (fun i ↦ f (xSeq i)) k - f xStar ≤
ω[f; xStar] (bestRadiusUpTo (fun i ↦ dist (xSeq i) xStar) k)Proposition 3.26
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.26 (1): the aggregated affine lower model is dominated by the original convex function.
universe uE uU uι uR
open scoped BigOperators
open scoped WithTopConvexAnalysis
variable {ι : Type uι}
variable {E : Type uE} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
theorem sum_smul_sampledAffineMinorant_le
[Fintype ι] {f : E → ℝ} (y g : ι → E)
(hsubgrad : ∀ i, g i ∈ ∂ (fun z ↦ (f z : WithTop ℝ))((y i)))
(α : ι → ℝ) (hα_nonneg : ∀ i, 0 ≤ α i)
(hα_sum : ∑ i, α i = 1) (x : E) :
(∑ i, α i • sampledAffineMinorant (y i) (g i) (f (y i)) : E →ᵃ[ℝ] ℝ) x ≤ f xTheorem 3.26
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.26: if an ℝ ∪ {+∞}-valued function is homogeneous of degree p on its effective domain, then every subgradient at x pairs with x to give p times the finite real part of the function value at x.
open scoped WithTopConvexAnalysis
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
recall euler_homogeneous_function_theorem
{f : E → WithTop ℝ} {p : ℝ}
(hhom : IsPositivelyHomogeneousOn p (dom f) (withTopRealPart f))
{x g : E} (hg : g ∈ ∂ f(x)) :
inner ℝ g x = p * withTopRealPart f x
Definition 3.27
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.27: a convex optimization problem with linear equality constraints consists of an ambient closed convex set Q, a convex objective f on Q, a linear map A, and a right-hand side b, encoding the program min_{x ∈ Q} {f x : A x = b}. In the textbook Euclidean realization, A is represented by a matrix; full row rank is an additional theorem-level assumption on that matrix presentation, not primitive owner data.
universe u v
structure LinearEqualityConstrainedConvexProblem
(E : Type u) (Λ : Type v) [TopologicalSpace E] [AddCommMonoid E] [Module ℝ E]
[AddCommMonoid Λ] [Module ℝ Λ]
extends PrimalEqualityConstrainedProblem E Λ where
/-- The ambient set `Q` is closed. -/
feasibleSet_closed : IsClosed feasibleSet
/-- The objective is convex on the ambient set `Q`. -/
objective_convex : ConvexOn ℝ feasibleSet objectiveLemma 3.27
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.27 is the direct owner recall of the quadratic lower-support estimate already formalized in Lemma_3_2_3.
recall StrongConvexOn.lower_tangent_derivWithin_of_mem_interiorProposition 3.27
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.27: if affine minorants ℓ i lie below f on the feasible set P, then any point whose objective value is within rN of a minimizer of their weighted sum is within rN of the infimum of f over P.
open scoped BigOperators
universe u v
variable {ι : Type v} [Fintype ι]
variable {E : Type u} [AddCommGroup E] [Module ℝ E]
theorem approximate_optimality_of_sum_smul_affine
{P : Set E} {f : E → ℝ}
(α : ι → ℝ) (ℓ : ι → E →ᵃ[ℝ] ℝ)
(hα_nonneg : ∀ i, 0 ≤ α i) (hα_sum : ∑ i, α i = 1)
(h_lower : ∀ i x, x ∈ P → ℓ i x ≤ f x)
{xAgg xN : E} {rN : ℝ}
(hagg_min : IsLeast (((∑ i, α i • ℓ i : E →ᵃ[ℝ] ℝ) '' P)) ((∑ i, α i • ℓ i) xAgg))
(happrox : f xN ≤ (∑ i, α i • ℓ i) xAgg + rN) :
f xN ≤ sInf (f '' P) + rNTheorem 3.27
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.27 [Chapter3_2.json:22]: for bounded convex sets Q₁ and Q₂ with nonempty interior intersection in a finite-dimensional real inner-product space, the support function of Q₁ ∩ Q₂ at x is the minimum over translated support-function sums ξ[Q₁] (x + y) + ξ[Q₂] (-y).
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
open scoped SupportFunction
theorem supportFunction_inter_isLeast_add_supportFunction
[FiniteDimensional ℝ E]
{Q₁ Q₂ : Set E}
(hQ₁_bounded : Bornology.IsBounded Q₁) (hQ₁_convex : Convex ℝ Q₁)
(hQ₂_bounded : Bornology.IsBounded Q₂) (hQ₂_convex : Convex ℝ Q₂)
(hQ_int : (interior (Q₁ ∩ Q₂)).Nonempty) (x : E) :
IsLeast
(Set.range fun y : E ↦ ξ[Q₁] (x + y) + ξ[Q₂] (-y))
(ξ[Q₁ ∩ Q₂] x)Definition 3.28
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.28, generalized from the textbook ℝⁿ × ℝᵐ setting: a convex optimization problem with a max-representation consists of a convex primal owner on E, a convex parameter set S ⊆ U, and a real-valued kernel Ψ : E → U → ℝ such that each primal slice x ↦ Ψ(x, u) is closed and convex on P, each dual slice u ↦ Ψ(x, u) is closed and concave on S, and for every x ∈ P the objective value f(x) is the maximum of u ↦ Ψ(x, u) over S. The textbook P ⊆ ℝⁿ, S ⊆ ℝᵐ form is recovered by specializing E and U to Euclidean spaces.
universe u v
open scoped WithTopConvexAnalysis
structure MaxRepresentationPrimalDualProblem
(E : Type u) (U : Type v)
[TopologicalSpace E] [AddCommMonoid E] [Module ℝ E]
[TopologicalSpace U] [AddCommMonoid U] [Module ℝ U]
extends SetConstrainedMinimizationProblem E where
/-- The parameter set `S ⊆ U` indexing the max-representation. -/
dualSet : Set U
/-- The parameter set `S` is convex. -/
dualSet_convex : Convex ℝ dualSet
/-- The real-valued kernel `Ψ(x, u)`. -/
kernel : E → U → ℝ
/-- Each primal slice `x ↦ Ψ(x, u)` is closed and convex on `P`. -/
kernel_primal_closedConvex {u : U} (_ : u ∈ dualSet) :
ClosedConvexOn feasibleSet (fun x ↦ (kernel x u : WithTop ℝ))
/-- Each dual slice `u ↦ Ψ(x, u)` is closed and concave on `S`, encoded by closed convexity of
`u ↦ -Ψ(x, u)`. -/
kernel_dual_closedConcave {x : E} (_ : x ∈ feasibleSet) :
ClosedConvexOn dualSet (fun u ↦ (-kernel x u : WithTop ℝ))
/-- On `P`, the objective value is the maximum of `u ↦ Ψ(x, u)` over `S`. -/
objective_isGreatest {x : E} (_ : x ∈ feasibleSet) :
IsGreatest (kernel x '' dualSet) (objective x)Lemma 3.28
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.28 lies in the Euclidean closed-ball / midpoint-bisection-box domain. Sampled owner-style declarations: - mathlib Metric.closedBall - project FeasibilityResistingOracleState.currentCenter - project FeasibilityResistingOracleState.currentBox - project FeasibilityResistingOracleState.closedBall_subset_currentBox Best owner abstraction: - the earlier chapter owner theorem FeasibilityResistingOracleState.closedBall_subset_currentBox. Primitive data: - a radius parameter R - the positive dimension witness hn - the resisting-oracle transcript state Derived API: - the textbook-radius closed-ball inclusion in the current realized midpoint-bisection box. Source/core/bridge triage: - source-facing: the textbook-radius ball inclusion in the current realized box - core/canonical: the midpoint-bisection box owner in Algorithm_3_5 - bridge/view: the earlier chapter theorem FeasibilityResistingOracleState.closedBall_subset_currentBox, which already expresses exactly this source-level consequence This item is recall-only: Lemma_3_28 duplicated the earlier chapter theorem exactly, so the file keeps the canonical recall instead of a second parallel public theorem.
recall FeasibilityResistingOracleState.closedBall_subset_currentBox
Proposition 3.28
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.28: for nonnegative parameters μ and γ, the subdifferential of f_k n k μ γ, i.e. the Nemirovski hard instance x ↦ (μ / 2) ‖x‖² + γ max_{1 ≤ i ≤ k} xᵢ, is the affine image of the convex hull of the active standard basis vectors among the first k coordinates.
open Set
open EuclideanSpace
open scoped Pointwise WithTopConvexAnalysis
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
local notation "e[" i "]" => EuclideanSpace.single i (1 : ℝ)
variable (k : ℕ)
theorem subdifferential_f_k_eq_affineImage_convexHull_activeBasis
(μ γ : ℝ) (hk : 0 < k) (hkn : k ≤ n) (hμ : 0 ≤ μ) (hγ : 0 ≤ γ) (x : E) :
∂ (fun y : E ↦ (f_k n k μ γ y : WithTop ℝ))(x) =
(fun v : E ↦ μ • x + γ • v) ''
convexHull ℝ
((fun i : FirstKIndex n k ↦ e[i.1]) ''
activePointwiseSupremumOnIndices
(Set.univ : Set (FirstKIndex n k)) (firstKCoordinateFamily n k) x)
Theorem 3.28
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.28 lies in the chapter's convex composite minimization / first-order optimality domain. Sampled owner-style declarations: - ConvexOn.isMinOn_iff_variational_inequality_of_hasGradientAt in Chap02/Theorem_2_29 - ConvexOn.isMinOn_add_iff_neg_hasGradient_mem_constrainedSubdifferential in Chap03/Theorem_3_1_23 - ConvexOn.isMinOn_add_iff_variational_inequality_of_hasGradientAt in Chap03/Theorem_3_1_23 - isMinOn_add_convex_iff_forall_inner_gradient_add_ge in Chap03/Theorem_3_1_23 Best owner abstraction: - ConvexOn.isMinOn_add_iff_neg_hasGradient_mem_constrainedSubdifferential, the primitive gradient-witness owner theorem for convex composite minimization on a real inner-product space, phrased through the chapter's constrained-subdifferential owner. Primitive data: - ConvexOn ℝ Q f - ConvexOn ℝ Q Ψ - xStar ∈ Q - HasGradientAt f g xStar Derived API: - the raw variational-inequality bridge ConvexOn.isMinOn_add_iff_variational_inequality_of_hasGradientAt - the source-facing gradient specialization isMinOn_add_convex_iff_forall_inner_gradient_add_ge Source/core/bridge triage: - source-facing: Theorem 3.28's variational inequality for minimizing x ↦ f x + Ψ x on Q - core/canonical: ConvexOn.isMinOn_add_iff_neg_hasGradient_mem_constrainedSubdifferential - bridge/view: ConvexOn.isMinOn_add_iff_variational_inequality_of_hasGradientAt and isMinOn_add_convex_iff_forall_inner_gradient_add_ge This file is recall-only. The earlier duplicate local theorem has already been removed, and the upstream owner file now carries the right abstraction layer: general real inner-product spaces, no redundant Convex ℝ Q binder, and only pointwise differentiability at xStar for the gradient-based view. Accordingly, Theorem 3.28 recalls that source-facing corollary directly rather than rebuilding a parallel local bridge.
recall isMinOn_add_convex_iff_forall_inner_gradient_add_geDefinition 3.29
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.29: for a max-representation problem, an oracle chooses at each feasible query point x ∈ P a feasible maximizer u(x) ∈ S and a relative subgradient of the active slice Ψ(·, u(x)) at x. The equality f(x) = Ψ(x, u(x)) is derived from maximizer optimality via the owner problem.
universe u v
open scoped WithTopConvexAnalysis
variable {E : Type u} {U : Type v}
variable [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable [TopologicalSpace U] [AddCommMonoid U] [Module ℝ U]
structure MaxRepresentationOracle
(problem : MaxRepresentationPrimalDualProblem E U) where
/-- The oracle reply at the feasible query point `x`, consisting of the selected maximizer
`u(x)` together with the selected relative subgradient `g(x)`. -/
answer : problem.feasibleSet → problem.dualSet × E
/-- At feasible query points, the selected `u(x)` maximizes `Ψ(x, ·)` over the parameter set. -/
maximizer_spec :
∀ x : problem.feasibleSet,
IsMaxOn (problem.kernel x) problem.dualSet (answer x).1
/-- At feasible query points, the selected `g(x)` lies in the relative subdifferential of the
active slice `Ψ(·, u(x))` over the primal set. -/
subgradient_spec :
∀ x : problem.feasibleSet,
(answer x).2 ∈
∂[problem.feasibleSet]
(((fun y ↦ problem.kernel y ((answer x).1 : U)) : E → ℝ))
((x : E))Lemma 3.29
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.29, pairing form: the propagated certificate belongs to the common regular subdifferential and satisfies the normal-cone pairing inequalities on the whole constrained minimizer set.
open scoped ConstrainedArgmin
open scoped NormalCone
open scoped WithTopConvexAnalysis
recall subgradient_mem_commonRegularSubdifferential_and_nonneg_pairing_of_mem_constrainedArgminProposition 3.29
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.29: if every point of the ball Metric.ball xStar ρ lies in the effective domain of f and every subgradient on that ball satisfies ‖g‖ ≤ μ * ‖y‖ + γ, then for x, y ∈ Metric.ball xStar ρ and any g ∈ ∂ f(y) one has the local estimate withTopRealPart f y - withTopRealPart f x ≤ (μ * ‖xStar‖ + μ * ρ + γ) * ‖y - x‖.
open scoped WithTopConvexAnalysis
universe u
variable {E : Type u} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
theorem subgradient_upper_difference_bound_on_ball
{f : E → WithTop ℝ} {xStar x y g : E} {ρ μ γ : ℝ}
(hμ : 0 ≤ μ)
(hball_dom : Metric.ball xStar ρ ⊆ dom f)
(hsubgrad_bound :
∀ ⦃z s : E⦄, z ∈ Metric.ball xStar ρ → s ∈ ∂ f(z) → ‖s‖ ≤ μ * ‖z‖ + γ)
(hx : x ∈ Metric.ball xStar ρ) (hy : y ∈ Metric.ball xStar ρ)
(hg : g ∈ ∂ f(y)) :
withTopRealPart f y - withTopRealPart f x ≤
(μ * ‖xStar‖ + μ * ρ + γ) * ‖y - x‖Theorem 3.29
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.29 [Chapter3_2.json:34]: a feasible point xStar ∈ Q belongs to the constrained optimal set argmin[Q] f if and only if there exists a subgradient certificate with nonnegative pairing on Q; every such certificate then belongs to the common regular subdifferential and the common normal cone of the whole minimizer set.
open scoped ConstrainedArgmin
open scoped NormalCone
open scoped WithTopConvexAnalysis
universe u
variable {V : Type u} [NormedAddCommGroup V] [InnerProductSpace ℝ V] [CompleteSpace V]
theorem mem_constrainedArgmin_iff_exists_subgradient_nonneg_pairing_with_common_certificate
[FiniteDimensional ℝ V]
{Q : Set V} (hQ_convex : Convex ℝ Q)
{f : V → ℝ} (hf_conv : ConvexOn ℝ Set.univ f)
{xStar : V} (hxStar : xStar ∈ Q) :
(xStar ∈ argmin[Q] f ↔
∃ gStar : V,
gStar ∈ ∂ (fun x : V ↦ (f x : WithTop ℝ))(xStar) ∧
∀ x ∈ Q, 0 ≤ inner ℝ gStar (x - xStar)) ∧
∀ {gStar : V},
gStar ∈ ∂ (fun x : V ↦ (f x : WithTop ℝ))(xStar) →
(∀ x ∈ Q, 0 ≤ inner ℝ gStar (x - xStar)) →
gStar ∈
∂̂ (fun x : V ↦ (f x : WithTop ℝ))((argmin[Q] f)) ∩
NormalCone.common (argmin[Q] f)Definition 3.30
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.30: the aggregated linear model attached to sample points y₀, …, y_N and a normalized coefficient vector is the direct simplex specialization of the canonical affine-map sum of the sampled affine minorants. The textbook additionally assumes these weights are strictly positive, but the defining expression depends only on the normalized simplex data.
universe uE uι
variable {E : Type uE} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
variable {N : ℕ}
variable (weights : StdSimplex ℝ (Fin (N + 1))) (y : Fin (N + 1) → E)
variable (f : E → ℝ) (g : E → E) (x : E)
#check (∑ i, weights.weights i • sampledAffineMinorant (y i) (g (y i)) (f (y i)) : E →ᵃ[ℝ] ℝ)Lemma 3.30
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.30: let y_k = center problem initialCenter radius k be the ellipsoid-method query sequence, let i(k) = Nat.count (fun j ↦ y j ∈ problem.feasibleSet) k be the canonical selected-feasible counter from Definition 3.53, and let X = feasibleSubsequence problem.feasibleSet y be the corresponding feasible subsequence. Then the selected recursive localization stage S_(i(k)) built from the feasible queried centers and their oracle cuts is contained in the raw stage-k associated ellipsoid E_k.
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
open EllipsoidMethod
namespace EllipsoidMethod
theorem selectedLocalizationSets_subset_associatedEllipsoid
(problem : ConvexMinimizationWithSeparationOracle (EuclideanSpace ℝ (Fin n)))
(initialCenter : E) (radius : ℝ)
(hn : 1 < n)
(hcut_nonzero : ∀ k : ℕ, cuttingVector problem initialCenter radius k ≠ 0)
(hE0_cover :
problem.feasibleSet ⊆ associatedEllipsoid problem initialCenter radius 0)
(hshape_pos : ∀ k : ℕ, (shape problem initialCenter radius k).PosDef)
(k : ℕ) :
let y : ℕ → E := center problem initialCenter radius
let X : ℕ → E := feasibleSubsequence problem.feasibleSet y
let S : ℕ → Set E := localizationSets problem.feasibleSet X (problem.oracle ∘ X)
let i : ℕ → ℕ := Nat.count (fun j ↦ y j ∈ problem.feasibleSet)
S (i k) ⊆ associatedEllipsoid problem initialCenter radius kProposition 3.30
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.30: if 0 < k ≤ n, μ > 0, and γ ≥ 0, then the explicit vector f_k_minimizer n k μ γ is a global minimizer of f_k n k μ γ.
variable (n k : ℕ) (μ γ : ℝ)
theorem isMinOn_f_k_minimizer (hk : 0 < k) (hkn : k ≤ n) (hμ : 0 < μ) (hγ : 0 ≤ γ) :
IsMinOn (f_k n k μ γ) Set.univ (f_k_minimizer n k μ γ)
Theorem 3.30
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.30: the unconstrained infimal projection lies below the real level λ exactly on the intersection, over all ε > 0, of the unions of the (λ + ε)-sublevel sets of the slices x ↦ F (x, y). This faithful EReal statement also covers fibers where the infimal projection equals ⊥.
universe u v
open scoped ConvexAnalysis
variable {X : Type u} {Y : Type v}
theorem partialInfProjection_univ_sublevelSet_eq_iInter_iUnion
(F : X × Y → EReal) (level : ℝ) :
{x | partialInfProjection Set.univ F x ≤ (level : EReal)} =
⋂ ε > 0, ⋃ y : Y, {x | F (x, y) ≤ (level + ε : EReal)}Definition 3.31
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.31: the gap function certificate attached to test points y_k, scaling coefficients α_k, and a map g is the function δ_N(x) = ∑_{k=0}^N α_k ⟪g(y_k), y_k - x⟫.
open scoped BigOperators
universe u
recall gapFunctionCertificate
{E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
{N : ℕ}
(y : Fin (N + 1) → E) (α : Fin (N + 1) → ℝ) (g : E → E) :
E → ℝLemma 3.31
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.31 at the intrinsic owner level: if S ⊆ E is bounded and convex with nonempty interior in a finite-dimensional real inner-product space, then for any nonzero g : E the centroid cut S₊ = S ∩ cuttingHalfspace (⨍ x in S, x) g = {x ∈ S | ⟪g, (⨍ x in S, x) - x⟫_ℝ ≥ 0} has relative volume at most 1 - 1 / e. Specializing to E = EuclideanSpace ℝ (Fin n) recovers the textbook ℝⁿ statement.
open MeasureTheory
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable [FiniteDimensional ℝ E]
theorem centerOfGravityCut_volumeRatio_le_one_sub_inv_e_of_bounded_convex_nonemptyInterior
(S : Set E) (g : E) (hS_bounded : Bornology.IsBounded S) (hS_convex : Convex ℝ S)
(hS_int : (interior S).Nonempty) (hg : g ≠ 0) :
(volume (S ∩ cuttingHalfspace (⨍ x in S, x) g)).toReal / (volume S).toReal ≤
1 - 1 / Real.exp 1
Proposition 3.31
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.31: coordinate-support growth under the resisting oracle. If a deterministic first-order method on ℝⁿ starts at 0 and the resisting oracle returns, at each stage i < k, a vector in ℝ^{i+1,n} whenever the current iterate lies in ℝ^{i,n}, then every iterate xᵢ with i ≤ k lies in ℝ^{i,n}. The iterate process itself is expressed through the chapter's canonical span-condition owner SatisfiesLinearSpanCondition.
open scoped CoordinateSubspace
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
theorem iterates_mem_coordinateSubspace_under_resistingOracle
{k : ℕ} {g : E → E} {xSeq : ℕ → E}
(hxSeq : SatisfiesLinearSpanCondition (0 : E) g xSeq k)
(hresist : ∀ i < k, xSeq i ∈ ℝ^{i,n} → g (xSeq i) ∈ ℝ^{i + 1,n})
(i : ℕ) (hi : i ≤ k) :
xSeq i ∈ ℝ^{i,n}
Theorem 3.31
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.31 is recall-only in the chapter's partial-infimal-projection / subdifferential-transfer domain. Primary domain: - partial infimal projection and subdifferential transfer for convex extended-real objectives. Sampled owner-style declarations: - partialInfProjection_convexOn_of_convexWithTop in Theorem_3_8, the upstream convexity owner for the canonical partialInfProjection; - partialInfProjection_realPart_convexOn_of_convexWithTop in Theorem_3_1_25, the canonical convexity theorem on the owner partial infimal projection; - mem_subdifferentialWithin_partialInfProjection_realPart_of_mem_argmin_of_subgradient in Theorem_3_1_25, the source-faithful subgradient-transfer theorem. Best owner abstraction: - the two chapter owner-level recall targets in Theorem_3_1_25. Primitive data: - none in this file. Derived API: - this numbered recall surface. Source/core/bridge triage: - source-facing: Theorem 3.31's convexity and subgradient-transfer statements for partial infimal projection; - core/canonical: the owner declarations in Theorem_3_1_25; - bridge/view: this recall surface. The previous file kept parallel theorem wrappers around declarations that now already live on the correct owner abstractions in Theorem_3_1_25. The redundant wrappers are deleted here in favor of direct recall/use.
recall partialInfProjection_realPart_convexOn_of_convexWithTop
Definition 3.32
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.32 is a recall-only item in the chapter's finite weighted-sum domain. Primary domain: - finite weighted sums of sampled scalar values. Sampled owner-style declarations: - mathlib dotProduct, the canonical owner for finite sums of entrywise products; - mathlib notation α ⬝ᵥ fy, the idiomatic display form of that owner; - hatf in Chap03/Lemma_3_24, the local weighted-sum notation built from dotProduct for the chapter's sampled primal values; - project gapFunctionCertificate_apply in Chap03/Lemma_3_24, the nearby textbook-style expansion from the dotProduct owner back to an explicit finite sum. Best owner abstraction: - dotProduct α (f ∘ y), equivalently α ⬝ᵥ (f ∘ y). Primitive data: - a horizon N : ℕ; - sample points y : Fin (N + 1) → X; - coefficients α : Fin (N + 1) → ℝ; - the sampled scalar family f ∘ y : Fin (N + 1) → ℝ. Derived API: - the textbook display α ⬝ᵥ (f ∘ y) = ∑ k, α k * f (y k), which is just the defining expansion of dotProduct; - nearby sampled weighted-gap owners built from the same dotProduct abstraction. Source/core/bridge triage: - source-facing: the textbook quantity \hat f_N = ∑_{k=0}^N α_k f(y_k); - core/canonical: dotProduct α (f ∘ y); - bridge/view: the explicit finite-sum display of that dot product. This file therefore recalls the mathlib owner directly instead of keeping a parallel local weighted-sum wrapper.
open scoped BigOperators
universe u
variable {X : Type u}
recall dotProductLemma 3.32
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.32 is a downstream recall of the chapter's ellipsoid-update owner API. Sampled owner-style declarations: - affineEllipsoid in Lemma_3_2_7, the source-facing owner of the textbook ellipsoid E(H, x̄); - centerCutEllipsoid in Lemma_3_2_7, the source-facing owner of the center cut E₊; - updatedEllipsoidCenter and updatedEllipsoidMatrix in Lemma_3_2_7, the textbook ellipsoid update formulas; - centerCutEllipsoid_subset_updatedEllipsoid_and_volume_le in Lemma_3_2_7, the owner theorem for the containment and volume decrease. Best owner abstraction: - source-facing/core owner: the ellipsoid-update API already introduced in Lemma_3_2_7; - bridge/view: this file is recall-only. Primitive data: - H : Mat; - xBar g : E; - the valid-update hypotheses H.PosDef, g ≠ 0, and 1 < n for the containment theorem. Derived API: - the ellipsoid affineEllipsoid H xBar; - the center cut centerCutEllipsoid H xBar g; - the updated center and shape matrix; - the canonical containment and volume-decrease theorem. Source/core/bridge triage: - source-facing: the textbook ellipsoid update and its containment/volume statement; - core/canonical: the existing owner declarations from Lemma_3_2_7; - bridge/view: this recall file. Accordingly, this file no longer keeps parallel local copies of the ellipsoid-update definitions or their main theorem. The owner objects already live upstream, so this file stays theorem-recall-only and reuses the canonical containment/volume statement directly.
open MeasureTheory
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
local notation "Mat" => Matrix (Fin n) (Fin n) ℝ
open scoped EllipsoidNotation
recall centerCutEllipsoid_subset_updatedEllipsoid_and_volume_le
(H : Mat) (hH : H.PosDef) (xBar g : E) (hg : g ≠ 0) (hn : 1 < n) :
E₊(H, xBar, g) ⊆ E(H₊(H, g), x̄₊(H, xBar, g)) ∧
((volume (E(H₊(H, g), x̄₊(H, xBar, g)))).toReal ≤
Real.rpow (1 - 1 / (((n : ℝ) + 1) ^ (2 : ℕ))) ((n : ℝ) / 2) *
(volume (E(H, xBar))).toReal)Proposition 3.32
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.32 [Chapter3_2.json:70]: if each iterate x_i with 0 ≤ i ≤ k - 1 satisfies max_{1 ≤ j ≤ k} x_i^(j) = 0, then along the first k - 1 nonzero steps one has f_k(x_i) ≥ γ max_{1 ≤ j ≤ k} x_i^(j) = 0; equivalently, f_k (x_{i+1}) ≥ 0 for every i < k - 1.
open scoped CoordinateSubspace
variable {k : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin k)
theorem no_decrease_f_k_first_steps
(μ γ : ℝ) (hμ : 0 ≤ μ) (x : ℕ → E)
(hmax : ∀ i : ℕ, i ≤ k - 1 → first_k_coordinate_max k k (x i) = 0)
{i : ℕ} (hi : i < k - 1) :
0 ≤ f_k k k μ γ (x (i + 1))
Theorem 3.32
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.32: (Karush--Kuhn--Tucker) under convexity, differentiability, and a Slater point for the inequality constraints on Q, a point x* solves min {f₀(x) | x ∈ Q, fᵢ(x) ≤ 0} if and only if there exists a nonnegative multiplier vector whose Lagrangian gradient pairing is nonnegative on Q and which satisfies complementary slackness at x*. This numbered item is already formalized on the canonical Chapter 3 owner surface in isMinOn_iff_exists_karush_kuhn_tucker_multiplier, so this file reuses that theorem directly instead of introducing a parallel local KKT API.
recall isMinOn_iff_exists_karush_kuhn_tucker_multiplierDefinition 3.33
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.33: an unconstrained convex minimization problem on ℝⁿ is a real-valued objective f : ℝⁿ → ℝ together with the canonical whole-space convexity predicate ConvexOn ℝ Set.univ f.
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
variable (f : E → ℝ) (xStar : E)
abbrev IsUnconstrainedConvexMinimizationProblem (f : E → ℝ) : Prop :=
ConvexOn ℝ Set.univ fLemma 3.33
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.33 is the direct recall of the owner interval-monotonicity theorem.
#check LevelMethodHistory.levelValue_ge_approximateOptimalValue_of_intervalMonotonicityProposition 3.33
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.33 lies in the chapter's extended-valued convex-analysis / subdifferential domain. Sampled owner-style declarations: - dom in Definition_3_3, the chapter owner for effective domains of WithTop ℝ-valued functions; - subdifferential and mem_subdifferential_iff in Definition_3_1_5, the source-facing owner API for extended-valued subgradients; - subgradient_nonneg_on_sublevelSet_of_mem_subdifferential in Theorem_3_1_18, the canonical sublevel-set inequality behind the minimizer pairing result; - subgradient_inner_sub_nonneg_of_isMinOn in Theorem_3_1_5_6, the chapter owner theorem for pairing a subgradient at a feasible point with the displacement to a minimizer. Best owner abstraction: - core/canonical: subgradient_inner_sub_nonneg_of_isMinOn; - bridge/view: the codomain coercion from a real-valued objective f : E → ℝ to the WithTop ℝ-valued owner surface fun y ↦ (f y : WithTop ℝ). Primitive data: - an inner-product space E, a feasible set Q, a real-valued objective f, points x, xStar and a subgradient g; - the feasibility hypothesis x ∈ Q; - the minimizing hypothesis IsMinOn f Q xStar; - the owner-membership hypothesis g ∈ ∂ (fun y ↦ (f y : WithTop ℝ))(x). Derived API: - the nonnegativity inequality 0 ≤ inner ℝ g (x - xStar). Source/core/bridge triage: - source-facing: the textbook real-valued spelling of the minimizer-pairing inequality; - core/canonical: subgradient_inner_sub_nonneg_of_isMinOn; - bridge/view: the coercion from real-valued objectives to the chapter's extended-valued subdifferential owner. This proposition contributes no new public owner beyond the existing theorem subgradient_inner_sub_nonneg_of_isMinOn: its former local theorem was only the codomain-coercion specialization of that owner theorem, and its second theorem was a verbatim alias of the first. The file therefore keeps the canonical owner theorem directly instead of preserving a parallel specialized wrapper API.
recall subgradient_inner_sub_nonneg_of_isMinOnTheorem 3.33
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.33: this item contains only the cross-reference “Theorem 3.1.26.”, so its source-faithful formalization is a direct recall of the existing Chapter 3 Karush--Kuhn--Tucker optimality theorem rather than a second local theorem shell. The recalled owner theorem is isMinOn_iff_exists_karush_kuhn_tucker_multiplier.
recall isMinOn_iff_exists_karush_kuhn_tucker_multiplierDefinition 3.34
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.34: relative to a chosen minimizer x*, a point x̄ is an ε-approximate solution of the unconstrained objective f when its objective gap above f(x*) is at most ε.
universe u
variable {E : Type u}
def IsApproximateSolution (f : E → ℝ) (xStar : E) (ε : ℝ) (xBar : E) : Prop :=
f xBar - f xStar ≤ εLemma 3.34
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.34: if the one-step affine cut generated by g_k satisfies f(x_k) + ⟪g_k, x_{k+1} - x_k⟫ ≤ f_k^* - (1 - α) δ_k, if ‖g_k‖ ≤ M_f, and if f_k^* ≤ f(x_k), then the step length satisfies ‖x_{k+1} - x_k‖ ≥ ((1 - α) δ_k) / M_f.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
theorem step_norm_lower_bound_of_level_method_cutting_plane_bound
{f : E → ℝ} {xk xNext gk : E} {α δk fkStar Mf : ℝ}
(hMf_pos : 0 < Mf)
(hgk_bound : ‖gk‖ ≤ Mf)
(hcutting_plane :
f xk + inner ℝ gk (xNext - xk) ≤ fkStar - (1 - α) * δk)
(hfkStar_le : fkStar ≤ f xk) :
‖xNext - xk‖ ≥ ((1 - α) * δk) / MfProposition 3.34
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.34 (1): the pointwise growth function ω_f(xBar; t) vanishes for every nonpositive radius t.
universe u
open scoped PointwiseGrowthFunction
variable {X : Type u} [MetricSpace X]
theorem pointwiseGrowthFunction_eq_zero_of_nonpos
(f : X → ℝ) (xBar : X) (t : ℝ) (ht : t ≤ 0) :
ω[f; xBar] t = 0Theorem 3.34
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.34 is the intrinsic linear-map theorem isMinOn_linearEqualityFeasibleSet_iff_exists_subgradient_multiplier_with_bound; the textbook matrix form is the downstream specialization via A.toEuclideanLin and the standard transpose/adjoint identification.
recall isMinOn_linearEqualityFeasibleSet_iff_exists_subgradient_multiplier_with_bound
Definition 3.35
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.35: the Nemirovski hard-instance objective f_k(x) = (μ / 2) ‖x‖^2 + γ max{x^(1), …, x^(k)} on ℝ^n.
def f_k (n k : ℕ) (μ γ : ℝ) (x : EuclideanSpace ℝ (Fin n)) : ℝ := (μ / 2) * ‖x‖ ^ 2 + γ * first_k_coordinate_max n k x
Lemma 3.35
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.35: if the later gap satisfies δ_p ≥ (1 - α) δ_k, then the gap drop δ_k - δ_p is bounded above by α δ_k.
open scoped LevelMethodNotation
namespace LevelMethodHistory
lemma gap_drop_le_alpha_mul_gap_of_gap_ge_one_sub_alpha_mul_gap
(history : LevelMethodHistory) {k p : ℕ} {α : ℝ}
(hgap : δ[history](p) ≥ (1 - α) * δ[history](k)) :
δ[history](k) - δ[history](p) ≤ α * δ[history](k)Proposition 3.35
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.35: for R > 0, the function Δ_N attains its minimum on the positive orthant at the constant choice h_i = R / √(N + 1). The closed-form value is recorded separately by deltaN_constantChoice.
open scoped BigOperators EuclideanOrthant
local notation "E" N => EuclideanSpace ℝ (Fin (N + 1))
open scoped DeltaN
theorem deltaN_constantChoice_minimizes_positiveOrthant
(N : ℕ) {R : ℝ} (hR : 0 < R) :
IsMinOn (Δ[N; R]) (ℝ₊₊^(N + 1)) (deltaNConstantChoice N R)Theorem 3.35
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.35 (1): under convexity of f on its effective domain plus convexity of Q, the finite real part of the canonical affine-fiber infimal projection is convex on its finite-value domain.
open scoped ConvexAnalysis WithTopConvexAnalysis
universe u v
variable {E : Type u} {Λ : Type v}
variable [AddCommMonoid E] [Module ℝ E]
variable [AddCommMonoid Λ] [Module ℝ Λ]
variable {f : E → WithTop ℝ} {Q : Set E} (A : E →ₗ[ℝ] Λ)
theorem affinePartialInfProjection_realPart_convexOn
(hf : ConvexOn ℝ (dom f) (withTopRealPart f))
(hQ_convex : Convex ℝ Q) :
ConvexOn ℝ
(dom
(partialInfProjection
{p : Λ × E | p.2 ∈ linearEqualityFeasibleSet Q A p.1}
(withTopToEReal ∘ f ∘ Prod.snd)))
(extendedRealRealPart
(partialInfProjection
{p : Λ × E | p.2 ∈ linearEqualityFeasibleSet Q A p.1}
(withTopToEReal ∘ f ∘ Prod.snd)))
Definition 3.36
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.36: a convex minimization problem is a set-constrained minimization problem min_{x ∈ Q} f(x) whose feasible set Q is nonempty, closed, and convex, and whose objective f is convex on the whole ambient space.
universe u
open scoped ConstrainedArgmin
variable {X : Type u} [TopologicalSpace X] [AddCommMonoid X] [Module ℝ X]
structure ConvexMinimizationProblem (X : Type u) [TopologicalSpace X] [AddCommMonoid X]
[Module ℝ X] extends SetConstrainedMinimizationProblem X where
/-- The feasible set `Q` is nonempty. -/
feasibleSet_nonempty : feasibleSet.Nonempty
/-- The feasible set `Q` is closed. -/
feasibleSet_closed : IsClosed feasibleSet
/-- The feasible set `Q` is convex. -/
feasibleSet_convex : Convex ℝ feasibleSet
/-- The objective `f` is convex on the whole ambient space. -/
objective_convex : ConvexOn ℝ Set.univ objectiveLemma 3.36
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.36: if the feasible objective values attain a least element m, then t_k^*(X) equals that minimum value.
universe u v w
open scoped ConstrainedArgmin ConstrainedThreshold
variable {Index : Type u} {Param : Type v} {Decision : Type w}
variable (Q : Set Decision) (hatFn checkFn : Index → Param → Decision → ℝ)
variable (k : Index) (X : Param)
lemma constrainedThreshold_eq_minimum_of_isLeast
{m : ℝ}
(hmin : IsLeast ((hatFn k X) '' (Q ∩ {x | checkFn k X x ≤ 0})) m) :
t*[Q; hatFn; checkFn](k, X) = (m : EReal)Proposition 3.36
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.36: if f is convex on the ambient space, x_{i+1} = x_i - (ε / M²) • g_i, each g_i is a subgradient of f at x_i with ‖g_i‖ ≤ M, and a chosen minimizer xStar satisfies ‖x₀ - xStar‖ ≤ R, then the best sampled objective value bestFunctionValueUpTo (fun i ↦ f (xSeq i)) N among x₀, …, x_N satisfies the textbook constant-stepsize estimate M^2 * R^2 / (2 * ε * N) + ε / 2 for every integer N ≥ 1.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
open scoped BigOperators DeltaN WithTopConvexAnalysis
variable (f : E → ℝ) (xStar : E) (R M ε : ℝ)
variable (xSeq g : ℕ → E)
theorem subgradientMethod_bestObjectiveValue_sub_le_of_constant_stepsize
(hM : 0 < M) (hε : 0 < ε)
(h_convex : ConvexOn ℝ Set.univ f)
(hxStar_min : IsMinOn f Set.univ xStar)
(hxSeq_zero_dist : ‖xSeq 0 - xStar‖ ≤ R)
(hxSeq_succ : ∀ i, xSeq (i + 1) = xSeq i - (ε / M ^ (2 : ℕ)) • g i)
(h_subgradient : ∀ i, g i ∈ ∂[Set.univ] f((xSeq i)))
(h_subgradient_bound : ∀ i, ‖g i‖ ≤ M)
(N : ℕ) (hN : 1 ≤ N) :
bestFunctionValueUpTo (fun i ↦ f (xSeq i)) N - f xStar ≤
M ^ (2 : ℕ) * R ^ (2 : ℕ) / (2 * ε * N) + ε / 2Theorem 3.36
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.36 lies in the chapter's unrestricted minimax / saddle-value domain. Primary domain: - unrestricted minimax for a parametric payoff, already organized in the project around the set-based saddle-point and pointwise-supremum owners. Relevant sampled declarations: - minimax_eq_of_unique_slice_argmin_and_attained_dual_max - isSaddlePointOn_of_unique_slice_argmin_and_attained_dual_max - isMinOn_objective_of_unique_slice_argmin_and_attained_dual_max - pointwiseSupremumOn Best owner abstraction: - source-facing: this numbered recall surface; - core/canonical: the chapter owner minimax_eq_of_unique_slice_argmin_and_attained_dual_max from Theorem_3_1_29, phrased through pointwiseSupremumOn, IsSaddlePointOn, IsMinOn, and IsMaxOn; - bridge/view: the unrestricted case is the specialization P = Set.univ, S = Set.univ. Primitive data: - none in this file; the owner theorem already carries the full mathematical data. Derived API: - the recalled minimax theorem, together with the saddle-point and primal-minimizer companions in Theorem_3_1_29. This file no longer introduces a second public lower-envelope theorem or companion bridge lemmas. The same textbook item is already canonicalized upstream in Theorem_3_1_29, so the correct surface here is direct reuse of that owner rather than a parallel sInf (Set.range ...) wrapper.
recall minimax_eq_of_unique_slice_argmin_and_attained_dual_maxDefinition 3.37
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.37 lies in the cutting-plane localization-measure domain. Primary mathematical domain: - subgradient-based localization radii around a reference point in finite-dimensional Euclidean cutting-plane analysis. Sampled owner-style declarations: - subgradientLocalizationMeasure in Lemma_3_2_1, the chapter owner for v_f(xBar; x) - subgradientLocalizationMeasure_eq_zero_of_eq_zero in Lemma_3_2_1, the owner-level zero-branch simplification - subgradientLocalizationMeasure_eq_inner_div_norm_of_ne_zero in Lemma_3_2_1, the owner-level nonzero-branch simplification - sub_le_pointwiseGrowthFunction_of_localizationMeasure in Lemma_3_2_1, the first comparison theorem built from the same owner Best owner abstraction for this file: - subgradientLocalizationMeasure Primitive data: - a reference point xBar - a chosen subgradient selection g - an evaluation point x Derived API: - the zero-case simplification theorem subgradientLocalizationMeasure_eq_zero_of_eq_zero - the nonzero-case simplification theorem subgradientLocalizationMeasure_eq_inner_div_norm_of_ne_zero - the comparison sub_le_pointwiseGrowthFunction_of_localizationMeasure Source/core/bridge triage: - source-facing: the pointwise localization measure subgradientLocalizationMeasure g xBar x - core/canonical: subgradientLocalizationMeasure - bridge/view: the zero/nonzero branch simplifications and the growth-function comparison obtained by unfolding the owner The owner declaration subgradientLocalizationMeasure in Lemma_3_2_1 already carries the exact mathematical content of Definition 3.37. This file therefore recalls that owner directly instead of keeping a parallel local copy. The branch simplification theorems stay with the owner file rather than being re-recalled from this definition-only item.
recall subgradientLocalizationMeasureLemma 3.37
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.37: if t₀ < t₁ ≤ τ, the scalar model value \hat f_k^*(X; t₁) is positive, \hat f_k^*(X; τ) = 0, and t ↦ \hat f_k^*(X; t) is convex on (-∞, τ], then τ > t₁ and the displayed secant lower bound holds.
universe u v
variable {Index : Type u} {Param : Type v}
variable (estimatedValue : Index → Param → ℝ → ℝ)
variable (k : Index) (X : Param)
local notation "value" => estimatedValue k X
theorem estimatedValue_strict_lt_right_and_secant_lower_bound_of_eq_zero
{t0 t1 τ : ℝ}
(ht01 : t0 < t1)
(ht1_le_right : t1 ≤ τ)
(hpositive : 0 < value t1)
(hright_zero : value τ = 0)
(hconvex : ConvexOn ℝ (Set.Iic τ) value) :
τ > t1 ∧
value t0 ≥
value t1 +
((t1 - t0) / (τ - t1)) * value t1Proposition 3.37
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.37: every index in 𝒜(N) is a Case A iteration, hence its iterate satisfies f̄(x_k) ≤ ε.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E] [CompleteSpace E]
namespace FunctionalConstraintSubgradientMethod
variable {problem : ProjectedMultipleConstraintFirstOrderProblem E 1}
theorem constraint_le_epsilon_of_mem_admissibleIndices
(method : FunctionalConstraintSubgradientMethod problem) (N : ℕ) {k : Fin (N + 1)}
(hk : k ∈ method.admissibleIndices N) :
problem.constraints 0 (method k) ≤ method.ε
Theorem 3.37
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.37: for a continuous convex-concave payoff on compact nonempty sets P and S, the infimum of the upper slice-value function x ↦ sSup ((fun u ↦ Ψ x u) '' S) over P equals the supremum of the lower slice-value function u ↦ sInf ((fun x ↦ Ψ x u) '' P) over S.
universe u v
variable {X : Type u} {U : Type v}
variable [TopologicalSpace X] [AddCommGroup X] [Module ℝ X]
[IsTopologicalAddGroup X] [ContinuousSMul ℝ X]
variable [TopologicalSpace U] [AddCommGroup U] [Module ℝ U]
[IsTopologicalAddGroup U] [ContinuousSMul ℝ U]
theorem compact_convex_concave_minimax
{P : Set X} {S : Set U} (hP_nonempty : P.Nonempty) (hS_nonempty : S.Nonempty)
(hP_compact : IsCompact P) (hS_compact : IsCompact S)
{Ψ : X → U → ℝ}
(hΨ_cont : ContinuousOn (fun z : X × U ↦ Ψ z.1 z.2) (Set.prod P S))
(hΨ_convex : ∀ u ∈ S, ConvexOn ℝ P (fun x ↦ Ψ x u))
(hΨ_concave : ∀ x ∈ P, ConcaveOn ℝ S (fun u ↦ Ψ x u)) :
sInf ((fun x ↦ sSup ((fun u ↦ Ψ x u) '' S)) '' P) =
sSup ((fun u ↦ sInf ((fun x ↦ Ψ x u) '' P)) '' S)Definition 3.38
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.38 lies in the Chapter 3 pointwise-growth / local-modulus domain. Primary domain: - supremal growth profiles of a real-valued function around a base point on ℝⁿ. Sampled owner-style declarations: - pointwiseGrowthFunction in Lemma_3_2_1, the chapter owner for ω_f(xBar; t) - pointwiseGrowthFunction_eq_zero_of_neg in Lemma_3_2_1, the owner-level negative-radius branch simplification - sub_le_pointwiseGrowthFunction_of_localizationMeasure in Lemma_3_2_1, the first comparison theorem built from the same owner - pointwiseGrowthFunction_monotone in Proposition_3_34, the radius-monotonicity theorem for the same owner profile Best owner abstraction: - pointwiseGrowthFunction Primitive data: - a real-valued function f - a base point xBar - a radius parameter t Derived API: - the owner-level negative-radius simplification pointwiseGrowthFunction_eq_zero_of_neg - the localization-measure comparison sub_le_pointwiseGrowthFunction_of_localizationMeasure - the monotonicity and radius-evaluation consequences developed downstream in Proposition_3_34 Source/core/bridge triage: - source-facing: the textbook growth function ω_f(xBar; t) - core/canonical: pointwiseGrowthFunction - bridge/view: the branch simplifications and comparison lemmas obtained by unfolding the owner The owner declaration pointwiseGrowthFunction in Lemma_3_2_1 already carries the exact mathematical content of Definition 3.38. This file therefore recalls that owner directly instead of keeping a parallel local copy such as pointwise_growth_function. The negative-radius branch theorem stays with the owner file instead of being re-exported from this definition-only recall.
recall pointwiseGrowthFunctionLemma 3.38
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.38, exact-value part: direct reuse of the chapter owner shift theorem.
universe u
recall parametricValueFunction_sub_le_shiftProposition 3.38
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.38: a nonnegative linear combination of two strongly convex functions on the same convex domain is strongly convex, with strong-convexity parameter α₁ * μ₁ + α₂ * μ₂.
universe u
variable {E : Type u} [NormedAddCommGroup E] [NormedSpace ℝ E]
namespace StrongConvexOn
theorem nonneg_weighted_add
{Q : Set E} {f₁ f₂ : E → ℝ} {μ₁ μ₂ α₁ α₂ : ℝ}
(hf₁ : StrongConvexOn Q μ₁ f₁)
(hf₂ : StrongConvexOn Q μ₂ f₂)
(hα₁ : 0 ≤ α₁) (hα₂ : 0 ≤ α₂) :
StrongConvexOn Q (α₁ * μ₁ + α₂ * μ₂) (α₁ • f₁ + α₂ • f₂)Theorem 3.38
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.38: if f : E → ℝ is a real-valued objective on P whose WithTop ℝ lift agrees with the faithful upper-envelope owner pointwiseSupremumOn S (fun x u ↦ (Ψ x u : WithTop ℝ)), if xStar minimizes f on P, if a relative subgradient gStar ∈ ∂[P] f(xStar) annihilates every feasible displacement x - xStar, and if gStar is a simplex-weighted combination of relative subgradients g i ∈ ∂[P] (fun x ↦ Ψ x (u i)) (xStar) of active slices x ↦ Ψ(x, u i) at xStar whose weighted slice values can be aggregated by one parameter uBar ∈ S, then that aggregated parameter is a saddle point of Ψ on P × S with first coordinate xStar; equivalently, it realizes the same lower-envelope value as the primal optimum at xStar.
open scoped BigOperators WithTopConvexAnalysis
universe u v w
variable {E : Type u} {U : Type v} {ι : Type w}
variable [Fintype ι]
variable [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
theorem exists_saddle_parameter_of_active_subgradient_representation
{P : Set E} {S : Set U} {Ψ : E → U → ℝ} {f : E → ℝ}
{xStar gStar : E}
(hf_eq :
∀ ⦃x : E⦄, x ∈ P →
(f x : WithTop ℝ) = pointwiseSupremumOn S (fun x' u ↦ (Ψ x' u : WithTop ℝ)) x)
(hxStar_min : IsMinOn f P xStar)
(hgStar_mem : gStar ∈ ∂[P] f(xStar))
(horth : ∀ ⦃x : E⦄, x ∈ P → inner ℝ gStar (x - xStar) = 0)
(weights : StdSimplex ℝ ι)
(u : ι → U) (g : ι → E)
(hu_active :
∀ i : ι,
u i ∈ activePointwiseSupremumOnIndices S
(fun x u ↦ (Ψ x u : WithTop ℝ)) xStar)
(hg_mem :
∀ i : ι, g i ∈ ∂[P] ((fun x ↦ Ψ x (u i)) : E → ℝ) (xStar))
(hgStar_repr : gStar = ∑ i, weights.weights i • g i)
(haggregate :
∃ uBar ∈ S,
∀ ⦃x : E⦄, x ∈ P →
∑ i, weights.weights i * Ψ x (u i) ≤ Ψ x uBar) :
∃ uBar ∈ S, IsSaddlePointOn P S Ψ xStar uBar
Definition 3.39
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.39 [Chapter3_3.json:138]: for a feasible region Q, a cut map g, a query sequence xSeq, and a reference point xStar, the textbook localization set S_k = {x ∈ Q | ⟪g(x_i), x_i - x⟫ ≥ 0 for all i = 0, ..., k} is the chapter owner localizationSet Q xSeq (g ∘ xSeq) k, the pointwise quantities v_i are the localization measures v[g; xStar] (xSeq i), and the best localization radius v_k^* = min_{0 ≤ i ≤ k} v_i is the owner localization_radius xStar g xSeq k. The equivalent closed-ball characterization of v_k^* is already provided by the companion localization-set inclusion theorems, so this item remains a pure canonical recall with no parallel local wrapper.
recall localization_radiusLemma 3.39
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.39 lies in the complete-data level-method geometric-decay domain. Sampled owner-style declarations: - selected_exactValue_le_initial_gap_div_one_sub_epsilon_mul_geometric_decay in Lemma_3_3_7.lean, the earlier chapter theorem with the same geometric-decay content and the sharper Lean interface keeping only the mathematically effective hypotheses; - HasGeometricRateOfConvergence in Chap01/Definition_1_2_6.lean, the scalar owner predicate for geometric decay; - HasGeometricRateOfConvergence.of_step_bound in Chap01/Definition_1_2_6.lean, the canonical one-step-to-geometric bridge used throughout the project; - constrainedMinimizationInternalGap_hasGeometricRateOfConvergence in Chap02/Proposition_2_30.lean, the chapter style example of exposing the source-facing theorem through that scalar owner abstraction. Best owner abstraction: - source-facing owner for this exact statement: selected_exactValue_le_initial_gap_div_one_sub_epsilon_mul_geometric_decay; - core/canonical owner beneath it: HasGeometricRateOfConvergence. Primitive data: - the scalar sequences t and j; - the exact and estimated value families; - the comparison, initialization, and step-size hypotheses. Derived API: - the geometric upper bound for the selected exact values. Source/core/bridge triage: - source-facing: Lemma 3.39 itself, stated in the complete-data exactValue / estimatedValue notation; - core/canonical: HasGeometricRateOfConvergence and of_step_bound; - bridge/view: the earlier chapter theorem from Lemma_3_3_7, whose parameter name ε is only a binder-name variation of the present source notation α, and whose public header already drops the redundant textbook-side hypotheses. The former version of this file duplicated both the source-facing theorem and its local helper chain. Since Lemma_3_3_7 already owns the exact public statement, this file stays recall-only and introduces no parallel theorem shell; the direct owner use below is a recall of that owner theorem rather than a second declaration.
recall selected_exactValue_le_initial_gap_div_one_sub_epsilon_mul_geometric_decayProposition 3.39
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.39: the pointwise maximum of two strongly convex functions on the same feasible set is strongly convex with parameter min μ₁ μ₂.
universe u
variable {E : Type u} [NormedAddCommGroup E] [NormedSpace ℝ E]
namespace StrongConvexOn
theorem sup
{Q : Set E} {f₁ f₂ : E → ℝ} {μ₁ μ₂ : ℝ}
(hf₁ : StrongConvexOn Q μ₁ f₁)
(hf₂ : StrongConvexOn Q μ₂ f₂) :
StrongConvexOn Q (min μ₁ μ₂) (f₁ ⊔ f₂)Theorem 3.39
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.39 [Chapter3_2.json:71]: for k + 1 ≤ n, the canonical Chapter 3 hard-instance theorem exists_problem_with_nonsmooth_firstOrder_lower_bound gives a convex Lipschitz objective in 𝒫(x₀, R, M) such that every valid first-order oracle satisfying the prefix-support-growth condition, and every iterate sequence satisfying the linear-span condition for that oracle, has objective gap at least MR / (2 (2 + √(k + 1))) at step k. The source inequalities R > 0, M > 0, and 0 ≤ k ≤ n - 1 are absorbed by the project-facing parameters R M : NNReal and k + 1 ≤ n.
open scoped LipschitzConvexProblemClass recall exists_problem_with_nonsmooth_firstOrder_lower_bound
Definition 3.40
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.40: a convex minimization problem over a simple closed convex feasible set, with its first-order oracle for the objective. On complete spaces, the canonical Euclidean projection onto the feasible set and projected normalized subgradient step are derived below.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
open NormedSpace
structure FirstOrderConvexMinimizationProblem (E : Type u) [NormedAddCommGroup E]
[InnerProductSpace ℝ E]
extends SetConstrainedMinimizationProblem E where
/-- The feasible set `Q` is nonempty. -/
feasibleSet_nonempty : feasibleSet.Nonempty
/-- The feasible set `Q` is closed. -/
feasibleSet_closed : IsClosed feasibleSet
/-- The feasible set `Q` is convex. -/
feasibleSet_convex : Convex ℝ feasibleSet
/-- A first-order oracle for the objective. -/
oracle : FirstOrderOracle objective
Proposition 3.40
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.40: in a real inner product space, μ-strong convexity on Q is exactly convexity of the quadratic-corrected objective.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
recall strongConvexOn_iff_convex
{Q : Set E} {μ : ℝ} {f : E → ℝ} :
StrongConvexOn Q μ f ↔ ConvexOn ℝ Q (fun x ↦ f x - μ / (2 : ℝ) * ‖x‖ ^ 2)Theorem 3.40
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.40: for a simple-set subgradient method on an owner first-order convex minimization problem, the best objective value among the first k + 1 iterates satisfies the standard M * Δ_k error bound, written with the chapter owners bestFunctionValueUpTo (fun i ↦ problem (method i)) k for the sampled minimum f_k^* and Δ[k; R] (method.stepsizePrefix k) for the finite stepsize prefix; the chosen oracle subgradient is assumed nonzero at every iterate different from xStar, as in the source.
open scoped BigOperators DeltaN SubgradientLocalizationMeasure
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E] [CompleteSpace E]
namespace SimpleSetSubgradientMethod
variable {problem : FirstOrderConvexMinimizationProblem E}
theorem bestFunctionValueUpTo_sub_le_of_projected_normalized_subgradient_method
(method : SimpleSetSubgradientMethod problem) (xStar : E) (R M : NNReal)
(hxStar_min : IsMinOn problem problem.feasibleSet xStar)
(hxStar_mem : xStar ∈ problem.feasibleSet)
(hf_lipschitz : LipschitzOnWith M problem (Metric.closedBall xStar R))
(hx0_ball : method.x0 ∈ Metric.closedBall xStar R)
(hstepsize_pos : ∀ i : ℕ, 0 < method.stepsize i)
(hsubgradient_nonzero :
∀ i : ℕ, method i ≠ xStar → problem.oracle.subgradient (method i) ≠ 0)
(k : ℕ) :
bestFunctionValueUpTo (fun i ↦ problem (method i)) k - problem xStar ≤
(M : ℝ) * Δ[k; R] (method.stepsizePrefix k)
Definition 3.41
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.41 reuses the chapter owner for the finite stepsize scalar directly.
open scoped BigOperators
open scoped DeltaN
local notation "E" N => EuclideanSpace ℝ (Fin (N + 1))
recall deltaN (N : ℕ) (R : ℝ) (h : E N) : ℝProposition 3.41
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.41: if μ > 0, M ≥ 0, the quadratic lower-gap bound at x₀, and the linear upper-gap bound at x₀ all hold relative to xStar, then ‖x₀ - xStar‖ ≤ 2 M / μ.
universe u
variable {E : Type u} [NormedAddCommGroup E]
theorem norm_sub_le_two_mul_div_of_quadratic_growth_le_lipschitz
{f : E → ℝ} {μ M : ℝ} {x₀ xStar : E}
(hμ : 0 < μ)
(hM : 0 ≤ M)
(hquad : (μ / 2) * ‖x₀ - xStar‖ ^ (2 : ℕ) ≤ f x₀ - f xStar)
(hLip : f x₀ - f xStar ≤ M * ‖x₀ - xStar‖) :
‖x₀ - xStar‖ ≤ 2 * M / μTheorem 3.41
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.41: if the first N + 1 stepsizes of a simple-set subgradient method are chosen constantly as h_i = R / √(N + 1), then the best sampled objective gap among the first N + 1 iterates is bounded by M R / √(N + 1).
universe u
open scoped DeltaN
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E] [CompleteSpace E]
namespace SimpleSetSubgradientMethod
variable {problem : FirstOrderConvexMinimizationProblem E}
theorem bestFunctionValueUpTo_sub_le_of_constant_stepsizes
(method : SimpleSetSubgradientMethod problem) (xStar : E) (R M : NNReal)
(hxStar_min : IsMinOn problem problem.feasibleSet xStar)
(hf_lipschitz : LipschitzOnWith M problem (Metric.closedBall xStar R))
(hx0_ball : method.x0 ∈ Metric.closedBall xStar R)
(N : ℕ)
(h_stepsize :
∀ i : Fin (N + 1), method.stepsize i = (R : ℝ) / Real.sqrt (N + 1 : ℝ)) :
bestFunctionValueUpTo (fun i ↦ problem (method i)) N - problem xStar ≤
(M : ℝ) * ((R : ℝ) / Real.sqrt (N + 1 : ℝ))
Definition 3.42
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.42: the textbook aggregate constraint function \bar f is exactly the univ finite specialization of the chapter owner pointwiseSupremumOn for a WithTop ℝ-valued constraint family.
open Set
universe u v
variable {ι : Type u} {X : Type v}
#check (fun constraints : ι → X → WithTop ℝ ↦
pointwiseSupremumOn (univ : Set ι) (fun x i ↦ constraints i x) :
(ι → X → WithTop ℝ) → X → WithTop ℝ)
Proposition 3.42
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.42: for projected subgradient descent on a problem in the strongly convex class 𝒫_s(x₀, μ, M), the sampled-prefix owner value bestFunctionValueUpTo (fun i ↦ f (xSeq i)) N is at most f(x*) + ε once the iteration budget satisfies the parameter-only lower bound N ≥ (M² / (μ ε)) log (2 M² / (μ ε)). This is exactly the earlier strong-convexity owner theorem with the instance-dependent factor ‖x₀ - x*‖ replaced by the class-level estimate 2 M / μ from Proposition 3.41.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
open scoped WithTopConvexAnalysis StronglyConvexProblemClass
variable {x0 xStar : E} {μ M : NNReal} {f : E → ℝ}
local notation "Q" => (Metric.closedBall xStar ‖x0 - xStar‖ : Set E)
open IsInStronglyConvexProblemClass
theorem bestFunctionValueUpTo_le_optimalValue_add_eps_of_parameter_log_budget
(hf : 𝒫_s(x0, μ, M) f xStar)
{ε : ℝ} (hε : 0 < ε)
(projQ : E → E)
(hproj : ∀ x : E, IsProjectionPointOn Q x (projQ x))
(xSeq g : ℕ → E) (hxSeq_zero : xSeq 0 = x0)
(hsubgrad : ∀ k : ℕ, g k ∈ (∂[Q] f((xSeq k)) : Set E))
(hsubgrad_norm : ∀ k : ℕ, ‖g k‖ ≤ M)
(hxSeq_succ :
∀ k : ℕ,
xSeq (k + 1) =
projQ (xSeq k - ((2 * ε) / ‖g k‖ ^ (2 : ℕ)) • g k))
{N : ℕ}
(hN :
((M : ℝ) ^ (2 : ℕ)) / ((μ : ℝ) * ε) *
Real.log (((2 : ℝ) * (M : ℝ) ^ (2 : ℕ)) / ((μ : ℝ) * ε)) ≤
(N : ℝ)) :
bestFunctionValueUpTo (fun i ↦ f (xSeq i)) N ≤ f xStar + ε
Theorem 3.42
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.42: for the subgradient iteration with normalized stepsizes αₖ = ηₖ / ‖g(xₖ)‖², where ηₖ = ε whenever f(xₖ) - f(x*) ≥ ε, if every chosen search direction gₖ is a whole-space subgradient of f at xₖ and every such subgradient on the reference closed ball has norm at most M, then the chapter owner bestFunctionValueUpTo (fun i ↦ f (x i)) N among x₀, …, x_N is at most f(x*) + ε once N ≥ (M² / ε²) ‖x₀ - x*‖².
universe u
open scoped WithTopConvexAnalysis
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable {f : E → ℝ} {xStar : E} {x g : ℕ → E} {η : ℕ → ℝ} {ε M : ℝ}
theorem subgradient_method_bestFunctionValueUpTo_le_minimizer_add_of_stepsize_bound
(hxStar : IsMinOn f Set.univ xStar)
(hε : 0 < ε)
(hsubgrad : ∀ k : ℕ, g k ∈ ∂[Set.univ] f((x k)))
(hsubgrad_norm :
∀ ⦃y v : E⦄, y ∈ Metric.closedBall xStar ‖x 0 - xStar‖ →
v ∈ ∂[Set.univ] f(y) → ‖v‖ ≤ M)
(hη : ∀ k : ℕ, ε ≤ f (x k) - f xStar → η k = ε)
(hstep : ∀ k : ℕ, x (k + 1) = x k - (η k / ‖g k‖ ^ (2 : ℕ)) • g k)
{N : ℕ}
(hN :
(M ^ (2 : ℕ) / ε ^ (2 : ℕ)) * ‖x 0 - xStar‖ ^ (2 : ℕ) ≤ (N : ℝ)) :
bestFunctionValueUpTo (fun i ↦ f (x i)) N ≤ f xStar + ε
Definition 3.43
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.43: for a domain Q, a family of constraint functions fⱼ : Q → α, and a threshold ε : α, the extended feasible set consists of the points x : Q such that every constraint value satisfies fⱼ(x) ≤ ε. Its Lean surface notation is 𝓕[constraints](ε).
universe u v w
variable {Q : Type u} {ι : Type v} {α : Type w}
variable [Preorder α]
variable (constraints : ι → Q → α) (ε : α)
def extendedFeasibleSet : Set Q :=
{x | ∀ j : ι, constraints j x ≤ ε}Proposition 3.43
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.43 (1): each midpoint bisection step divides the volume of the resulting axis-aligned box by two.
open MeasureTheory
theorem generated_box_volume_eq_half_of_midpoint_bisection
{n : ℕ} {a b a' b' : Fin n → ℝ} {i : Fin n}
(hbox : a ≤ b)
(hstep : IsMidpointCoordinateBisectionStep a b a' b' i) :
(volume (Set.Icc a' b')).toReal = (1 / 2 : ℝ) * (volume (Set.Icc a b)).toRealTheorem 3.43
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.43 lies in the chapter's approximate-Lagrange-multiplier switching-method domain. Sampled owner-style declarations: - positive_inactiveConstraintCount_of_large_iteration_count - maxTypeObjective and maxTypeObjective_le_iff - constraintMaximumAt_le_sampleMaxSubgradientNorm_mul_h_of_mem_inactiveConstraintIndices - delta_le_sampleMaxSubgradientNorm_mul_h_of_large_iteration_count Best owner abstraction: - the owner run ApproximateLagrangeMultiplierSwitchingMethod problem - the chapter finite-family maximum owner maxTypeObjective Primitive data: - the switching-method run method - the time index t, the objective-step iterate k ∈ A₀(t), and the stepsize positivity 0 < h Derived API: - the source-facing residual maximum maxTypeObjective problem.constraints (method k) - the componentwise owner theorem constraintMaximumAt_le_sampleMaxSubgradientNorm_mul_h_of_mem_inactiveConstraintIndices - the gap bound δ_t ≤ M h Source/core/bridge triage: - source-facing: Theorem 3.43's positivity statement, residual-maximum bound, and gap estimate - core/canonical: the run owner theorems in Theorem_3_2_4 and the finite-max owner maxTypeObjective - bridge/view: the finite-max reformulation of the componentwise residual bound The first and third clauses are exact owner recalls. The middle clause in the source is the finite residual maximum max_{1 ≤ j ≤ m} f_j(x_k) ≤ M h, and this theorem surface now lives directly in Theorem_3_2_4, so this file recalls it rather than reproving a parallel copy.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E] [CompleteSpace E]
variable {m : ℕ}
open ApproximateLagrangeMultiplierSwitchingMethod
namespace ApproximateLagrangeMultiplierSwitchingMethod
variable {problem : ProjectedMultipleConstraintFirstOrderProblem E m}
open scoped ApproximateLagrangeMultiplierSwitchingNotation
recall positive_inactiveConstraintCount_of_large_iteration_count
Definition 3.44
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.44: the textbook Lagrangian is the canonical owner problem.lagrangian.
universe u
open scoped BigOperators EuclideanOrthant
namespace LagrangianProblem
variable {Q : Type u} {m : ℕ}
local notation "Λ" => EuclideanSpace ℝ (Fin m)
recall lagrangian (problem : LagrangianProblem Q m) (x : Q) (l : Λ) : ℝProposition 3.44
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.44 at the intrinsic owner level: if Q ⊆ E satisfies the radius-ρ interior-ball condition in a finite-dimensional real normed space E, if 0 < Module.finrank ℝ E, if R ≥ 0, and if a stage set E_k has Haar/Lebesgue measure bounded by δ^(k * Module.finrank ℝ E / 2) * vol(B₂(x0, R)), then the normalized Module.finrank ℝ E-th-root volume ratio of E_k satisfies the corresponding geometric-decay estimate and the resulting exponential upper bound (R / ρ) * exp (-k / (2 (Module.finrank ℝ E + 1)^2)). Specializing to E = EuclideanSpace ℝ (Fin n) recovers the textbook ℝⁿ statement.
universe u
open MeasureTheory
variable {E : Type u} [NormedAddCommGroup E] [NormedSpace ℝ E]
variable [FiniteDimensional ℝ E] [MeasurableSpace E] [BorelSpace E]
local notation "dim" => Module.finrank ℝ E
local notation "δ" => 1 - 1 / (((dim : ℝ) + 1) ^ (2 : ℕ))
theorem volume_ratio_rpow_decay_under_interior_ball_condition
(hdim : 0 < dim) {μ : Measure E} [μ.IsAddHaarMeasure]
{Q : Set E} {ρ : ℝ} (hQ : Q.SatisfiesInteriorBallCondition ρ)
{x0 : E} {R : ℝ} (hR : 0 ≤ R) (k : ℕ) {Ek : Set E}
(hvol_decay :
μ Ek ≤
ENNReal.ofReal (Real.rpow δ (((k * dim : ℕ) : ℝ) / 2)) *
μ (Metric.closedBall x0 R)) :
Real.rpow (μ.real Ek / μ.real Q) (1 / (dim : ℝ)) ≤
Real.rpow δ ((k : ℝ) / 2) *
Real.rpow
(μ.real (Metric.closedBall x0 R) / μ.real Q)
(1 / (dim : ℝ)) ∧
Real.rpow δ ((k : ℝ) / 2) *
Real.rpow
(μ.real (Metric.closedBall x0 R) / μ.real Q)
(1 / (dim : ℝ)) ≤
(R / ρ) * Real.exp (-((k : ℝ) / (2 * (((dim : ℝ) + 1) ^ (2 : ℕ)))))Theorem 3.44
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.44: if f is μ-strongly convex on Q, then every feasible subgradient g ∈ ∂[Q] f(x) gives the quadratic affine lower bound f y ≥ f x + ⟪g, y - x⟫ + (μ / 2) * ‖y - x‖^2 for every y ∈ Q.
open scoped WithTopConvexAnalysis
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
namespace StrongConvexOn
theorem lower_bound_of_mem_subdifferentialWithin
{Q : Set E} {μ : ℝ} {f : E → ℝ} {x y g : E}
(hf : StrongConvexOn Q μ f)
(hg : g ∈ ∂[Q] f(x)) (hy : y ∈ Q) :
f y ≥ f x + inner ℝ g (y - x) + (μ / 2) * ‖y - x‖ ^ (2 : ℕ)Definition 3.45
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.45: for a run of Algorithm 3.4, the approximate dual multiplier vector λ_t is the canonical multiplier vector in EuclideanSpace ℝ (Fin m) whose jth coordinate is the textbook real ratio λ_t^{(j)} = σ_t⁻¹ ∑_{k ∈ A_j(t)} h_k, where A₀(t) = {k ≤ t : 𝒥_k = ∅}, A_j(t) = {k ≤ t : j_k = j}, N(t) = |A₀(t)|, S_t = ∑_{k ∈ A₀(t)} 1 / ‖g(x_k)‖, and σ_t = h S_t. On A_j(t), the chapter owner scalar (problem.constraintOracle j).correctionStepsize (method k) is exactly the textbook h_k. The owner is defined only under the full denominator regime hdenom, so all ratios are genuine textbook ratios rather than totalized zero-division artifacts. Orthant membership and positivity consequences belong to separate companion theorems.
open scoped BigOperators EuclideanOrthant
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable {m : ℕ}
namespace ApproximateLagrangeMultiplierSwitchingMethod
variable {problem : ProjectedMultipleConstraintFirstOrderProblem E m}
open scoped ApproximateLagrangeMultiplierSwitchingNotation
open scoped ApproximateLagrangeMultiplierSwitchingNotation
open scoped ApproximateLagrangeMultiplierSwitchingNotation
open scoped ApproximateLagrangeMultiplierSwitchingNotation
def approximateDualMultiplier
(method : ApproximateLagrangeMultiplierSwitchingMethod problem) (t : ℕ)
(hdenom : method.HasApproximateDualMultiplierDenominators t) :
EuclideanSpace ℝ (Fin m) :=
WithLp.toLp 2 fun j ↦
(Finset.sum (A[method](t, j)).attach
(selectedCorrectionWeight method t hdenom.selected j)) /
↑(Units.mk0 (σ[method](t; hdenom.objective))
hdenom.normalizingFactor_ne_zero : ℝˣ)
Proposition 3.45
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.45 as the source-facing selected-index bridge: let i(k) = Nat.count (fun j ↦ querySeq j ∈ Q) k be the canonical selected feasible index attached to the query sequence querySeq. If Q satisfies the radius-ρ interior-ball condition in a finite-dimensional real inner-product space E, if 0 < Module.finrank ℝ E, if Q lies in the outer ball B₂(x0, R), if the selected localization stage has volume at most vol(E_k), if E_k satisfies the standard ellipsoid-style stagewise volume decay, and if k > 2 (Module.finrank ℝ E + 1)^2 log (R / ρ), then i(k) > 0. The intrinsic normed-space core is volume_drop_of_log_threshold_under_interior_ball_condition; the extra inner-product-space assumption here comes only from the chapter owner localizationSets.
universe u
open MeasureTheory
variable {E : Type u} [NormedAddCommGroup E] [NormedSpace ℝ E]
variable [FiniteDimensional ℝ E] [MeasurableSpace E] [BorelSpace E]
local notation "μ" => (Measure.addHaar : Measure E)
local notation "finDim" => Module.finrank ℝ E
local notation "δ" => 1 - 1 / (((finDim : ℝ) + 1) ^ (2 : ℕ))
variable [InnerProductSpace ℝ E]
theorem selected_index_pos_of_log_threshold_under_interior_ball_condition
(hdim : 0 < finDim)
{Q : Set E} {ρ : ℝ} (hQ : Q.SatisfiesInteriorBallCondition ρ)
{querySeq : ℕ → E} {g : E → E} {Ek : Set E} {x0 : E} {R : ℝ} (k : ℕ)
(hQ_subset : Q ⊆ Metric.closedBall x0 R)
(hstage :
μ
(localizationSets
Q
(feasibleSubsequence Q querySeq)
(g ∘ feasibleSubsequence Q querySeq)
(Nat.count (fun j ↦ querySeq j ∈ Q) k)) ≤
μ Ek)
(hvol_decay :
μ Ek ≤ ENNReal.ofReal (Real.rpow δ (((k * finDim : ℕ) : ℝ) / 2)) *
μ (Metric.closedBall x0 R))
(hk :
(2 : ℝ) * (((finDim + 1 : ℕ) : ℝ) ^ (2 : ℕ)) * Real.log (R / ρ) < (k : ℝ)) :
0 < Nat.count (fun j ↦ querySeq j ∈ Q) kTheorem 3.45
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.45: if Q ⊂ ℝ^n is nonempty and f ∈ 𝒮^0_μ(Q), then every nonempty constrained level set L_α = {x ∈ Q : f x ≤ α} is bounded.
universe u
open scoped WithTopConvexAnalysis
namespace StrongConvexOn
open scoped StrongConvex
variable {n : ℕ}
local notation "Eₙ" => EuclideanSpace ℝ (Fin n)
theorem boundedSublevels_of_mem_S0On
{Q : Set Eₙ} {f : Eₙ → ℝ} {μ : ℝ}
(hf : f ∈ 𝒮^0_μ(Q)) (hQ_nonempty : Q.Nonempty) :
∀ α : ℝ,
(constrainedSublevelSet Q (((↑) : ℝ → WithTop ℝ) ∘ f) α).Nonempty →
Bornology.IsBounded
(constrainedSublevelSet Q (((↑) : ℝ → WithTop ℝ) ∘ f) α)Definition 3.46
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.46: for the feasible-set Lagrangian bridge attached to a run of Algorithm 3.4, the primal--dual gap quantity δ[method](t; hdenom) is the difference between the owner weighted primal average over A₀[method](t) and the canonical dual value at the approximate multiplier λ[method](t; hdenom). Under the full denominator regime hdenom from Definition 3.45, this is the textbook quantity S_t⁻¹ ∑_{k ∈ A₀[method](t)} f(x_k) / ‖g(x_k)‖ - φ(λ_t). The value is taken in EReal because LagrangianProblem.dualFunction is extended-real-valued.
open scoped BigOperators
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable {m : ℕ}
namespace ApproximateLagrangeMultiplierSwitchingMethod
variable {problem : ProjectedMultipleConstraintFirstOrderProblem E m}
open scoped ApproximateLagrangeMultiplierSwitchingNotation
local notation "φ" => problem.toLagrangianProblem.dualFunction
def primalDualGapQuantity
(method : ApproximateLagrangeMultiplierSwitchingMethod problem) (t : ℕ)
(hdenom : method.HasApproximateDualMultiplierDenominators t) : EReal :=
(((A₀[method](t)).centerMass
(fun k ↦ ‖method.objectiveSubgradient k‖⁻¹)
(fun k ↦ problem.objective (method k)) : ℝ) : EReal) -
φ (λ[method](t; hdenom))
Proposition 3.46
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.46: let X = feasibleSubsequence Q querySeq be the canonical selected feasible subsequence of a raw query sequence querySeq, and let i(k) = Nat.count (fun j ↦ querySeq j ∈ Q) k be the corresponding selected-feasible prefix count. If querySeq k is feasible, if f is M-Lipschitz on B₂(xStar, R), if xStar realizes the infimum of f on that ball, and if the raw feasible query querySeq k satisfies the standard distance estimate to xStar, then the best sampled objective value among the first i(k) + 1 selected feasible points satisfies the same decay estimate relative to sInf (f '' Metric.closedBall xStar R).
universe v
open MeasureTheory
variable {E : Type v} [NormedAddCommGroup E] [NormedSpace ℝ E]
variable [FiniteDimensional ℝ E] [MeasurableSpace E]
local notation "dim" => Module.finrank ℝ E
theorem selected_feasible_bestObjectiveValue_sub_sInf_image_closedBall_le_lipschitz_radius_mul_geometricDecay_volumeRatio_of_feasible
(μ : Measure E)
{f : E → ℝ} {Q : Set E} {querySeq : ℕ → E} {xStar x0 : E} {R : ℝ} {M : NNReal}
(k : ℕ)
(hf_lipschitz : LipschitzOnWith M f (Metric.closedBall xStar R))
(hxStar_opt : IsMinOn f (Metric.closedBall xStar R) xStar)
(hk : querySeq k ∈ Q)
(hquery_mem : querySeq k ∈ Metric.closedBall xStar R)
(hquery_dist :
‖querySeq k - xStar‖ ≤
R *
Real.rpow (1 - 1 / (((dim : ℝ) + 1) ^ (2 : ℕ))) ((k : ℝ) / 2) *
Real.rpow
(Measure.real μ (Metric.closedBall x0 R) / Measure.real μ Q)
(1 / (dim : ℝ))) :
bestFunctionValueUpTo
(fun j ↦ f (feasibleSubsequence Q querySeq j))
(Nat.count (fun j ↦ querySeq j ∈ Q) k) -
sInf (f '' Metric.closedBall xStar R) ≤
(M : ℝ) * R *
Real.rpow (1 - 1 / (((dim : ℝ) + 1) ^ (2 : ℕ))) ((k : ℝ) / 2) *
Real.rpow
(Measure.real μ (Metric.closedBall x0 R) / Measure.real μ Q)
(1 / (dim : ℝ))Theorem 3.46
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.46 (quadratic-growth part): this is exactly the owner theorem StrongConvexOn.quadratic_growth_of_isMinOn_of_mem.
universe u
namespace StrongConvexOn
variable {E : Type u} [NormedAddCommGroup E] [NormedSpace ℝ E]
recall StrongConvexOn.quadratic_growth_of_isMinOn_of_memDefinition 3.47
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.47 is a source-facing recall in the chapter's fixed-parameter strong-convexity domain. Primary domain: * fixed-parameter strong convexity of real-valued functions on a feasible set Sampled owner-style declarations: * StrongConvexOnWith and strongConvexOnWith_normSeminorm_iff in Nesterov.Chap02.Definition_2_14 * mathlib StrongConvexOn * exists_pos_strongConvexOn_iff_forall_segment_upper_bound in Nesterov.Chap03.Definition_3_2_2 Best owner abstraction: * source-facing: the fixed-parameter class 𝒮^0_μ(Q) as f ∈ 𝒮^0_μ(Q) * core/canonical: StrongConvexOn Q μ f * bridge/view: the Euclidean StrongConvexOnWith specialization from Definition_2_14 Primitive data: * a feasible set Q * a modulus μ * an objective f Derived API: * positivity of the fixed modulus via membership in 𝒮^0_μ(Q) * convexity and the quadratic Jensen inequality carried by StrongConvexOn * the positive-existential class ∃ μ > 0, StrongConvexOn Q μ f from Definition_3_2_2 Source/core/bridge triage: * source-facing: the textbook fixed-parameter class 𝒮^0_μ(Q) * core/canonical: StrongConvexOn Q μ f * bridge/view: the Euclidean reformulation from Definition_2_14 and the positive-existential restatement from Definition_3_2_2 The source semantics include μ > 0, so the public center here is the source-facing membership surface f ∈ 𝒮^0_μ(Q). The bare owner predicate StrongConvexOn Q μ f is retained only as the canonical core companion.
universe u
variable {E : Type u} [NormedAddCommGroup E] [NormedSpace ℝ E]
variable {Q : Set E} {μ : ℝ} {f : E → ℝ}
open scoped StrongConvex
#check (f ∈ 𝒮^0_μ(Q))Proposition 3.47
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.47 in source-facing form: under the logarithmic oracle budget, one of the queried ellipsoid centers is an ε-approximate minimizer of the constrained problem.
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
open EllipsoidMethod
open MeasureTheory
open scoped ConstrainedArgmin
variable
(problem : ConvexMinimizationWithSeparationOracle (EuclideanSpace ℝ (Fin n)))
(initialCenter xStar : E) (M : NNReal) {R ρ ε : ℝ}
(hε : 0 < ε) (hε_le : ε ≤ (M : ℝ) * R)
(hf_lipschitz : LipschitzOnWith M problem (Metric.closedBall xStar R))
(hfeasible_ball : problem.feasibleSet.SatisfiesInteriorBallCondition ρ)
(hxStar : xStar ∈ argmin[problem.feasibleSet] problem)
(hfeasible_subset_ball : problem.feasibleSet ⊆ Metric.closedBall xStar R)
(hn : 1 < n)
(hcut_nonzero : ∀ k : ℕ, cuttingVector problem initialCenter R k ≠ 0)
(hshape_pos : ∀ k : ℕ, (shape problem initialCenter R k).PosDef)
(hE0_cover : problem.feasibleSet ⊆ associatedEllipsoid problem initialCenter R 0)
local notation "y" => center problem initialCenter R
local notation "Y" => feasibleSubsequence problem.feasibleSet y
include initialCenter xStar hε hε_le hf_lipschitz hfeasible_ball hxStar
hfeasible_subset_ball hn hcut_nonzero hshape_pos hE0_cover
theorem ellipsoid_method_oracle_complexity_of_interior_ball
{N : ℕ}
(hN :
2 * (((n + 1 : ℕ) : ℝ) ^ (2 : ℕ)) * Real.log ((M : ℝ) * R ^ (2 : ℕ) / (ρ * ε)) ≤
(N : ℝ)) :
∃ k ≤ N, problem.IsApproximateMinimizer ε (y k)Theorem 3.47
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.47: for every k with 0 ≤ k ≤ n - 1, there exists a problem in the strongly convex class 𝒫_s(x₀, μ, M) such that every span-based first-order method whose valid subgradient oracle satisfies HasCoordinateSupportGrowth has objective gap at least M² / (2 μ (2 + √(k + 1))²) at its k-th iterate.
universe u
open scoped StrongConvex
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
open scoped StronglyConvexProblemClass CoordinateSubspace WithTopConvexAnalysis
theorem exists_stronglyConvexLipschitzProblem_with_span_method_lower_bound
(x0 : E) (μ M : NNReal)
(hμ : 0 < μ) (hM : 0 < M) {k : ℕ} (hk : k + 1 ≤ n) :
∃ f : E → ℝ,
∃ xStar : E,
𝒫_s(x0, μ, M) f xStar ∧
∀ oracle : FirstOrderOracle f,
HasCoordinateSupportGrowth x0 oracle.subgradient k →
∀ xSeq : ℕ → E, SatisfiesLinearSpanCondition x0 oracle.subgradient xSeq k →
f (xSeq k) - f xStar ≥
((M : ℝ) ^ (2 : ℕ)) /
((2 : ℝ) * (μ : ℝ) *
(((2 : ℝ) + Real.sqrt (((k + 1 : ℕ) : ℝ))) ^ (2 : ℕ)))
Definition 3.48
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.48 recalls the source-facing strongly convex problem class 𝒫_s(x₀, μ, M).
universe u
variable {V : Type u} [NormedAddCommGroup V] [NormedSpace ℝ V]
open scoped StronglyConvexProblemClass
variable (x0 : V) (μ M : NNReal)
#check (𝒫_s(x0, μ, M) : (V → ℝ) → V → Prop)
Proposition 3.48
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.48: for the problem min {f(y, x) | (y, x) ∈ Q} with f(y, x) = max {|y|, ‖x‖²} and Q = {(y, x) | y² + ‖x‖² ≤ 1}, the unique optimal solution is the origin, i.e. the minimizer set is {(0, 0)}. Specializing X = EuclideanSpace ℝ (Fin n) recovers the textbook ℝ × ℝⁿ instance.
universe u
variable {X : Type u} [NormedAddCommGroup X]
local notation "Z" => ℝ × X
open scoped ConstrainedArgmin
theorem kelleyCompleteArgmin_eq_singleton_origin :
(argmin[kelleyCompleteFeasibleSet] kelleyCompleteObjective : Set Z) =
({(0 : Z)} : Set Z)Theorem 3.48
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.48: if f belongs to the strongly convex class 𝒫_s(x₀, μ, M) with minimizer x*, π_Q is a projection map on the controlling ball Q = B₂(x*, ‖x₀ - x*‖), every selected subgradient g_k lies in ∂[Q] f(x_k), satisfies ‖g_k‖ ≤ M, and the projected subgradient iteration x_{k+1} = π_Q (x_k - (2 ε / ‖g_k‖²) • g_k) is run for a budget N ≥ (M² / (μ ε)) log (M ‖x₀ - x*‖ / ε), then the sampled best value bestFunctionValueUpTo (fun i ↦ f (xSeq i)) N is at most f x* + ε.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
open scoped WithTopConvexAnalysis StronglyConvexProblemClass
variable {x0 xStar : E} {μ M : NNReal} {f : E → ℝ}
local notation "Q" => (Metric.closedBall xStar ‖x0 - xStar‖ : Set E)
open IsInStronglyConvexProblemClass
theorem bestFunctionValueUpTo_le_optimalValue_add_eps_of_log_budget_projected_subgradient_method
(hf : 𝒫_s(x0, μ, M) f xStar)
{ε : ℝ} (hε : 0 < ε)
(projQ : E → E)
(hproj : ∀ x : E, IsProjectionPointOn Q x (projQ x))
(xSeq g : ℕ → E) (hxSeq_zero : xSeq 0 = x0)
(hsubgrad : ∀ k : ℕ, g k ∈ ∂[Q] f((xSeq k)))
(hsubgrad_norm : ∀ k : ℕ, ‖g k‖ ≤ M)
(hxSeq_succ :
∀ k : ℕ,
xSeq (k + 1) =
projQ (xSeq k - ((2 * ε) / ‖g k‖ ^ (2 : ℕ)) • g k))
{N : ℕ}
(hN :
(N : ℝ) ≥
((M : ℝ) ^ (2 : ℕ)) / ((μ : ℝ) * ε) *
Real.log ((M : ℝ) * ‖x0 - xStar‖ / ε)) :
bestFunctionValueUpTo (fun i ↦ f (xSeq i)) N ≤ f xStar + ε
Definition 3.49
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.49: a feasibility problem with separation oracle consists of a nonempty closed convex set S in a real inner-product space together with a separation oracle that, for every query point xBar, either certifies xBar ∈ S or, when xBar ∉ S, returns a nonzero vector g that separates xBar from S by the retained-cut inequality ⟪g, xBar - x⟫ ≥ 0 for all x ∈ S. Specializing E = EuclideanSpace ℝ (Fin n) recovers the textbook ℝⁿ formulation. Solving the feasibility problem means producing a point of S.
universe u
variable (E : Type u) [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
structure FeasibilityProblemWithSeparationOracle where
/-- The feasible set `S` in the ambient space. -/
feasibleSet : Set E
/-- The feasible set is nonempty. -/
feasibleSet_nonempty : feasibleSet.Nonempty
/-- The feasible set is closed. -/
feasibleSet_closed : IsClosed feasibleSet
/-- The feasible set is convex. -/
feasibleSet_convex : Convex ℝ feasibleSet
/-- A separation oracle for the feasible set. -/
oracle : SeparationOracle feasibleSetProposition 3.49
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.49 at the intrinsic owner level: the canonical Euclidean surface measure μHE[sphereRank] of a spherical cap of the unit sphere is bounded by the measure of the corresponding hemisphere times (1 - α^2)^(sphereDim / 2) for every α ∈ [0, 1]. The Kelley-specific lower bound 1 / 2 ≤ α belongs only to a later specialization, not to the source-facing spherical-cap estimate itself. No extra lower bound on Module.finrank ℝ E belongs to the public cap estimate: once ‖d‖ = 1, the statement already makes sense and stays valid even in the one-dimensional case. Specializing E = EuclideanSpace ℝ (Fin n) recovers the textbook ℝⁿ formulation used in the Kelley-method cut-count argument. The raw Hausdorff measure μH[sphereDim] is only the scaled bridge/view behind this canonical surface-measure owner.
open MeasureTheory
open scoped MeasureTheory RealInnerProductSpace
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable [FiniteDimensional ℝ E] [MeasurableSpace E] [BorelSpace E]
local notation "sphereRank" => Module.finrank ℝ E - 1
local notation "sphereDim" => ((sphereRank : ℕ) : ℝ)
theorem euclideanHausdorffMeasure_sphericalCap_le_hemisphere_mul_rpow
{d : E} (hd : ‖d‖ = 1) {α : ℝ} (hα_left : 0 ≤ α) (hα_right : α ≤ 1) :
μHE[sphereRank] (sphericalCap d α) ≤
μHE[sphereRank] (sphericalCap d 0) *
ENNReal.ofReal (Real.rpow (1 - α ^ 2) (sphereDim / 2))Theorem 3.49
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.49: fix R > 0 and ε ∈ (0, R / 2). Then for every deterministic separation-oracle algorithm and every oracle budget below n log (R / (2 ε)), there exists a feasibility problem with separation oracle, whose feasible set has nonempty interior and is contained in the textbook ball B∞(0, R), on which that algorithm cannot correctly output either a feasible point or a certificate that the feasible set misses B∞(0, ε) within that many oracle calls.
universe u
variable {n : ℕ}
open EuclideanSpace
local notation "E" => EuclideanSpace ℝ (Fin n)
theorem exists_hard_feasibility_problem_for_short_separation_oracle_algorithms
(m : ℕ)
(algorithm : SeparationOracleAlgorithm (EuclideanSpace ℝ (Fin m)))
{R ε : ℝ} (hR : 0 < R) (hε : 0 < ε) (hεR : ε < R / 2) {k : ℕ}
(hk : (k : ℝ) < (m : ℝ) * Real.log (R / (2 * ε))) :
∃ problem : FeasibilityProblemWithSeparationOracle (EuclideanSpace ℝ (Fin m)),
IsHardFeasibilityInstanceFor algorithm R ε k problem
Definition 3.50
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.50: a set satisfies the interior ball assumption exactly when it has a nonempty interior. The textbook Euclidean-ball formulation is recorded below as a bridge theorem, while the main owner expression is the canonical predicate (interior S).Nonempty.
open Set Metric
variable {E : Type*} [PseudoMetricSpace E]
variable (S : Set E)
#check (interior S).NonemptyProposition 3.50
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.50, canonical form: the source-faithful stepwise order chain is stated on the exact lower-value owner in EReal, so no extra attainment hypothesis is needed to express the model lower values faithfully.
universe u
open Set
open scoped LevelMethodNotation
variable {α : Type u}
variable {Q : Set α} {model : ℕ → α → ℝ}
variable (f : α → ℝ) (xSeq : ℕ → α)
theorem levelMethodApproximateProblem_optimalValue_stepwiseBounds_of_monotoneUnderestimators
(hxSeq : ∀ k : ℕ, xSeq k ∈ Q)
(hunder : ∀ k : ℕ, ∀ x ∈ Q, model k x ≤ f x)
(hmono : ∀ k : ℕ, ∀ x ∈ Q, model k x ≤ model (k + 1) x)
(k : ℕ) :
(levelMethodApproximateProblem Q model k).optimalValue ≤
(levelMethodApproximateProblem Q model (k + 1)).optimalValue ∧
(levelMethodApproximateProblem Q model (k + 1)).optimalValue ≤
(bestFunctionValueUpTo (fun i ↦ f (xSeq i)) (k + 1) : EReal) ∧
(bestFunctionValueUpTo (fun i ↦ f (xSeq i)) (k + 1) : EReal) ≤
bestFunctionValueUpTo (fun i ↦ f (xSeq i)) kTheorem 3.50
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.50: let n ≥ 1, R > 0, M > 0, and ε > 0. If a possibly randomized value-oracle method with at most T oracle calls uniformly guarantees, for every nonempty convex Q ⊆ B∞(0, R) and every f ∈ 𝓕_{M}^{0,0}(B∞(0, R)), an ε-approximate minimizer of the induced set-constrained problem, then the query budget satisfies the lower bound n * log (M R / (8 ε)) ≤ T. Here the possible randomization is represented by a nonempty family of deterministic transcript rules indexed by an internal seed.
universe v
open EuclideanSpace
open scoped ConvexLipschitz
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
local notation "coordEquiv" => EuclideanSpace.equiv (Fin n) ℝ
theorem value_oracle_query_lower_bound_of_uniform_epsilon_guarantee
(hn : 1 ≤ n) {Ω : Type v} [Nonempty Ω]
{R M : NNReal} {ε : ℝ} {T : ℕ}
(method : Ω → Set E → DeterministicValueOracleMethod E)
(hR : 0 < (R : ℝ)) (hM : 0 < (M : ℝ)) (hε : 0 < ε)
(hmethod :
DeterministicValueOracleMethod.SeededSolvesLinftyConstrainedProblemClassWithin
method R M ε T) :
(n : ℝ) * Real.log ((M : ℝ) * R / (8 * ε)) ≤ (T : ℝ)
Definition 3.51
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.51: a convex minimization problem with set constraint and separation oracle consists of a convex real-valued objective on a real inner-product space, a bounded closed convex feasible set with nonempty interior, and an oracle returning subgradients on feasible points and strict separating vectors outside the feasible set.
open Set
universe u
variable {E : Type u} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
structure ConvexMinimizationWithSeparationOracle
(E : Type u) [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
extends SetConstrainedMinimizationProblem E where
/-- The objective is convex on the whole ambient space. -/
objective_convex : ConvexOn ℝ univ objective
/-- The feasible set is bounded. -/
feasibleSet_bounded : Bornology.IsBounded feasibleSet
/-- The feasible set is closed. -/
feasibleSet_closed : IsClosed feasibleSet
/-- The feasible set is convex. -/
feasibleSet_convex : Convex ℝ feasibleSet
/-- The feasible set has nonempty interior. -/
feasibleSet_interior_nonempty : (interior feasibleSet).Nonempty
/-- A separation oracle for the constrained minimization problem. -/
oracle : ConvexMinimizationSeparationOracle feasibleSet objectiveProposition 3.51
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.51, owner form: if a complete level-method run satisfies the standard estimate f_N^* - f(x*) ≤ 2 M_f ‖x₀ - x*‖ / √N on the owner sampled-prefix value f_N^* = fstar(method.history, N), then every budget N above 4 M_f² ‖x₀ - x*‖² / ε² forces f_N^* ≤ f(x*) + ε.
universe u
open scoped ConstrainedArgmin LevelMethodNotation
variable {E : Type u} [NormedAddCommGroup E]
namespace CompleteLevelMethod
variable {problem : SetConstrainedMinimizationProblem E}
theorem optimalValue_le_comparison_add_of_complexity_estimate
(method : CompleteLevelMethod problem) {Mf : ℝ} {xStar : E} {ε : ℝ} {N : ℕ}
(h_level_estimate :
fstar(method.history, N) - problem xStar ≤
2 * Mf * ‖method.initialPoint - xStar‖ / Real.sqrt (N : ℝ))
(hε : 0 < ε)
(hN :
(4 * Mf ^ (2 : ℕ) * ‖method.initialPoint - xStar‖ ^ (2 : ℕ)) / ε ^ (2 : ℕ) ≤ (N : ℝ)) :
fstar(method.history, N) ≤ problem xStar + εTheorem 3.51
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.51, stated at the intrinsic owner level: if 0 < Module.finrank ℝ E, if a convex set Q of positive μ-volume in a finite-dimensional real normed space is contained in the closed ball B(xStar, D) with 0 < D, if Sk ⊆ Q is measurable, and if B(xStar, vkStar) ∩ Q ⊆ Sk, then under the chapter's standing center context xStar ∈ Q and the standing radius bound vkStar ≤ D coming from the earlier localization setup, the inner radius vkStar is bounded by D times the dim-th root of the Haar-measure ratio μ.real Sk / μ.real Q. The finite-volume side condition on Sk is derived internally from Sk ⊆ Q ⊆ B(xStar, D). Specializing to the canonical choice μ = Measure.addHaar gives the chapter owner used downstream, and in Euclidean space this differs from textbook Lebesgue volume only by a global positive normalization factor, so the displayed ratio is unchanged.
open MeasureTheory
universe u
variable {E : Type u} [NormedAddCommGroup E] [NormedSpace ℝ E] [FiniteDimensional ℝ E]
variable [MeasurableSpace E] [BorelSpace E]
variable {μ : Measure E} [μ.IsAddHaarMeasure]
local notation "dim" => Module.finrank ℝ E
theorem inner_ball_radius_le_outer_radius_mul_volume_ratio_rpow_of_convex
(hdim : 0 < dim)
{Q Sk : Set E} {xStar : E} {D vkStar : ℝ}
(hQ_convex : Convex ℝ Q) (hxStar : xStar ∈ Q)
(hQ_pos : 0 < μ.real Q)
(hD_pos : 0 < D)
(hvkStar_le : vkStar ≤ D)
(hQ_subset : Q ⊆ Metric.closedBall xStar D)
(hSk_meas : MeasurableSet Sk)
(hSk_subset : Sk ⊆ Q)
(hball : Metric.closedBall xStar vkStar ∩ Q ⊆ Sk) :
vkStar ≤
D *
Real.rpow
(μ.real Sk / μ.real Q)
(1 / (dim : ℝ))Definition 3.52
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.52: given an initial set Q, a sequence X = (x_k), and associated vectors g_k, the localization sets are defined recursively by S₀(X) = Q and S_{k+1}(X) = {x ∈ S_k(X) | ⟪g_k, x_k - x⟫ ≥ 0}. In the textbook this is specialized to Q ⊆ ℝⁿ.
universe u
variable {E : Type u} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
def localizationSets (Q : Set E) (xSeq gSeq : ℕ → E) : ℕ → Set EProposition 3.52
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.52: if xStar ∈ argmin[Q ∩ {x | fBar x ≤ 0}] f and every root of the owner value function parametricValueFunction Q f fBar is attained by a minimizer of the corresponding owner objective setConstrainedParametricObjective f fBar t, then f xStar is the smallest root of parametricValueFunction Q f fBar.
universe u
open scoped ConstrainedArgmin ConstrainedThreshold
theorem optimalValue_is_smallest_root_of_parametricValueFunction
{X : Type u} {Q : Set X} {f fBar : X → ℝ} {xStar : X}
(hxStar : xStar ∈ argmin[Q ∩ {x | fBar x ≤ 0}] f)
(hroot_attain : ∀ ⦃t : ℝ⦄, parametricValueFunction Q f fBar t = 0 →
(argmin[Q] (setConstrainedParametricObjective f fBar t)).Nonempty) :
IsLeast {t : ℝ | parametricValueFunction Q f fBar t = (0 : EReal)} (f xStar)Theorem 3.52
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.52 (1): let i(k) = Nat.count (fun j ↦ querySeq j ∈ Q) k be the canonical selected feasible index and let X = feasibleSubsequence Q querySeq be the corresponding selected sequence. If i(k) is positive, if the stage-i(k) recursive localization set of X has volume at most the ambient comparison volume vol_n Ell_k, and if the chapter owner bound for the localization radius is available, then the selected localization radius is bounded first by the selected localization-stage volume ratio and then by the ambient comparison volume ratio.
open MeasureTheory
universe u
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
variable {Q : Set E} {xStar : E} {g : E → E} {querySeq : ℕ → E} {Ell : ℕ → Set E}
variable {D : ℝ} {k : ℕ}
local notation "X" => feasibleSubsequence Q querySeq
local notation "m" => Nat.count (fun j ↦ querySeq j ∈ Q) k
local notation "S" => localizationSets Q X (g ∘ X)
theorem selected_radius_bound_of_positive_index
(hn : 0 < n)
(hQ_convex : Convex ℝ Q) (hxStar : xStar ∈ Q)
(hQ_pos : 0 < volume.real Q)
(hQ_subset : Q ⊆ Metric.closedBall xStar D)
(hEll_finite : volume (Ell k) ≠ ⊤)
(hvol : volume (S m) ≤ volume (Ell k))
(hik : 0 < m) :
(localization_radius xStar g X (m - 1) : ℝ) ≤
D * Real.rpow (volume.real (S m) / volume.real Q) ((1 : ℝ) / n) ∧
D *
Real.rpow (volume.real (S m) / volume.real Q) ((1 : ℝ) / n) ≤
D * Real.rpow (volume.real (Ell k) / volume.real Q) ((1 : ℝ) / n)Definition 3.53
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.53: the textbook counter i(k) is the canonical prefix count below.
universe u
variable {α : Type u}
variable (Q : Set α) (y : ℕ → α) (k : ℕ)
#check Nat.count (fun j ↦ y j ∈ Q) kProposition 3.53
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.53: if the model objective and model constraint are pointwise dominated by the exact objective and exact constraint on Q, then the model threshold t_k^*(X) is bounded above by the exact feasible optimum value f xStar.
universe u v w
open scoped ConstrainedArgmin ConstrainedThreshold
variable {Index : Type u} {Param : Type v} {Decision : Type w}
variable {Q : Set Decision} {hatFn checkFn : Index → Param → Decision → ℝ}
variable {k : Index} {X : Param} {f fBar : Decision → ℝ}
theorem constrainedThreshold_le_of_pointwise_le
{xStar : Decision}
(hxStar : xStar ∈ argmin[Q ∩ {x | fBar x ≤ 0}] f)
(hhat_le : ∀ ⦃x : Decision⦄, x ∈ Q → hatFn k X x ≤ f x)
(hcheck_le : ∀ ⦃x : Decision⦄, x ∈ Q → checkFn k X x ≤ fBar x) :
t*[Q; hatFn; checkFn](k, X) ≤ f xStarTheorem 3.53
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.53: along the center-of-gravity iterates, the best sampled objective value satisfies the geometric decay estimate already recorded by the chapter owner theorem in Theorem_3_2_10.
universe u recall CenterOfGravityMethod.bestFunctionValueUpTo_sub_optimalValue_le_lipschitz_geometricDecay
Definition 3.54
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.54: the center of gravity of a set S ⊆ ℝⁿ is the canonical Lebesgue set-average construction specialized to the identity map on S.
open MeasureTheory
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
variable (S : Set E)
#check (⨍ x in S, x ∂volume : E)Proposition 3.54
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.54: if the global stop condition holds, then some point of the finite set X attains the finite minimum of x ↦ max (f x - tk) (fBar x) and this value is at most ε.
universe u
variable {α : Type u}
open scoped ConstrainedLevelMinimum
theorem global_stop_exists_minimizer (X : Finset α) [Fact X.Nonempty] (tk ε : ℝ)
(f fBar : α → ℝ) (hstop : f*[X; f; fBar](tk) ≤ ε) :
∃ xStar ∈ X,
setConstrainedParametricObjective f fBar tk xStar = f*[X; f; fBar](tk) ∧
setConstrainedParametricObjective f fBar tk xStar ≤ εTheorem 3.54
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.54: let μ be the comparison measure used in the selected-point decay estimate, let xSeq be the feasible subsequence produced by the cutting-plane iterates, let i(k) be the corresponding feasible index, and let f_{i(k)}^* be the best objective value bestFunctionValueUpTo (fun j ↦ f (xSeq j)) (i k) among the first i(k) + 1 feasible points. If f is M-Lipschitz on B₂(xStar, R), if xStar realizes the infimum f^* on that ball, and if the selected feasible iterate xSeq (i k) satisfies the standard cutting-plane distance estimate ‖xSeq (i k) - xStar‖ ≤ R (1 - 1 / (dim + 1)^2)^(k / 2) (vol(B₂(x0, R)) / vol(Q))^(1 / dim), where dim = Module.finrank ℝ E and vol = Measure.real μ, then the same decay estimate holds for the objective gap f_{i(k)}^* - f^*. Taking μ = volume and specializing to E = EuclideanSpace ℝ (Fin n) recovers the textbook ℝⁿ statement.
universe u v
open MeasureTheory
variable {X : Type u} [PseudoMetricSpace X]
variable {E : Type v} [NormedAddCommGroup E] [NormedSpace ℝ E]
variable [FiniteDimensional ℝ E] [MeasurableSpace E]
local notation "dim" => Module.finrank ℝ E
theorem selected_feasible_bestObjectiveValue_sub_sInf_image_closedBall_le_lipschitz_radius_mul_geometricDecay_volumeRatio
(μ : Measure E)
{f : E → ℝ} {xStar x0 : E} {R : ℝ} {M : NNReal} {Q : Set E}
{xSeq : ℕ → E} {i : ℕ → ℕ}
(k : ℕ)
(hf_lipschitz : LipschitzOnWith M f (Metric.closedBall xStar R))
(hxStar_opt : IsMinOn f (Metric.closedBall xStar R) xStar)
(hselected_mem : xSeq (i k) ∈ Metric.closedBall xStar R)
(hselected_dist :
‖xSeq (i k) - xStar‖ ≤
R *
Real.rpow (1 - 1 / (((dim : ℝ) + 1) ^ (2 : ℕ))) ((k : ℝ) / 2) *
Real.rpow
(Measure.real μ (Metric.closedBall x0 R) / Measure.real μ Q)
(1 / (dim : ℝ))) :
bestFunctionValueUpTo (fun j ↦ f (xSeq j)) (i k) -
sInf (f '' Metric.closedBall xStar R) ≤
(M : ℝ) * R *
Real.rpow (1 - 1 / (((dim : ℝ) + 1) ^ (2 : ℕ))) ((k : ℝ) / 2) *
Real.rpow
(Measure.real μ (Metric.closedBall x0 R) / Measure.real μ Q)
(1 / (dim : ℝ))Definition 3.55
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.55: the best sampled objective value up to step k, namely the infimum of the sampled values indexed by 0, ..., k.
abbrev bestFunctionValueUpTo (values : ℕ → ℝ) (k : ℕ) : ℝ :=
⨅ j : Fin (k + 1), values jProposition 3.55
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.55 (1): if the relative termination rule fhat(history, k) ≥ (1 - κ) fstar(history, k) holds at a selected full step, then the owner gap δ[history](k) = fstar(history, k) - fhat(history, k) is bounded by κ fstar(history, k).
namespace LevelMethodHistory
theorem gap_le_kappa_mul_optimalValue_of_termination_rule
(history : LevelMethodHistory) {κ : ℝ} (k : ℕ)
(htermination :
history.approximateOptimalValue k ≥ (1 - κ) * history.optimalValue k) :
history.gap k ≤ κ * history.optimalValue kTheorem 3.55
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.55: at a full master step, the canonical internal iteration count ConstrainedLevelMethod.stoppingIndex method hrelative hfinite k = j(k) - j(k - 1) is bounded by the chapter owner levelMethodIterationCap evaluated at the uniform tolerance χ ε.
universe u
open scoped LevelMethodNotation
variable {E : Type u} [NormedAddCommGroup E]
recall ConstrainedLevelMethod.full_step_increment_le_uniform_internal_iteration_boundDefinition 3.56
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.56 belongs to the chapter's Euclidean ellipsoid API. Sampled owner-style declarations: - affineEllipsoid in Lemma_3_2_7, the earlier chapter owner of the textbook ellipsoid E(H, x̄); - mem_affineEllipsoid_iff in Lemma_3_2_7, the exact membership companion theorem; - centerCutEllipsoid in Lemma_3_2_7, the next ellipsoid-method construction derived from the same owner; - Matrix.toEuclideanLin, the ambient mathlib map turning a matrix into its Euclidean linear action. Best owner abstraction: - source-facing/core owner: affineEllipsoid; - bridge/view: mem_affineEllipsoid_iff. Primitive data: - a shape matrix H : Mat; - a center xBar : E. Derived API: - the ellipsoid affineEllipsoid H xBar; - the membership characterization mem_affineEllipsoid_iff. Source/core/bridge triage: - source-facing: affineEllipsoid; - core/canonical: the existing chapter owner from Lemma_3_2_7; - bridge/view: the companion membership equivalence. This file is therefore recall-only: the chapter already owns the ellipsoid and its defining membership theorem upstream, so no parallel local definition is kept here.
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
local notation "Mat" => Matrix (Fin n) (Fin n) ℝ
open Matrix
open scoped EllipsoidNotation
recall affineEllipsoid
(H : Mat) (xBar : E) :
Set EProposition 3.56
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.56: the terminal-step internal-iteration estimate is already owned by Lemma_3_3_9 under the canonical constrained-level method API.
recall ConstrainedLevelMethod.last_step_internal_iterations_le_uniform_internal_iteration_boundTheorem 3.56
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.56: let n ≥ 1 and consider the convex optimization problem min {f(y, x) | (y, x) ∈ Q} on ℝ × ℝ^(n - 1) with f(y, x) = max {|y|, ‖x‖²} and Q = {(y, x) | y² + ‖x‖² ≤ 1}. Then the optimal value is 0, attained at (0, 0). Moreover, the Kelley lower-bound execution is recorded here under the explicit nondegenerate side condition 2 ≤ n, which excludes the degenerate one-dimensional case n = 1.
open scoped ConstrainedArgmin
variable {n : ℕ}
theorem kelleyComplete_theorem_3_56
(hn : 1 ≤ n) :
KelleyCompleteTheoremConclusion nDefinition 3.57
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.57 lies in the chapter's Euclidean ellipsoid / cutting-plane domain. Sampled owner-style declarations: - centerCutEllipsoid in Lemma_3_2_7, the earlier Chapter 3 source-facing owner of the half-ellipsoid cut E₊; - mem_centerCutEllipsoid_iff in Lemma_3_2_7, the canonical companion membership theorem; - affineEllipsoid and mem_affineEllipsoid_iff in Lemma_3_2_7, the owner/view for the ambient ellipsoid E(H, x̄); - Matrix.toEuclideanLin in mathlib, the owner of the matrix action on EuclideanSpace ℝ (Fin n). Best owner abstraction: - source-facing/core owner: centerCutEllipsoid; - bridge/view: the textbook sign-reversed inequality 0 ≤ inner ℝ g (xBar - x). Primitive data: - H : Mat; - xBar : E; - g : E. Derived API: - the half-ellipsoid centerCutEllipsoid H xBar g; - its canonical membership theorem mem_centerCutEllipsoid_iff; - the textbook reformulation below, obtained by expanding ellipsoid membership and rewriting inner ℝ g (x - xBar) ≤ 0 as 0 ≤ inner ℝ g (xBar - x). Source/core/bridge triage: - source-facing: the half-ellipsoid E₊; - core/canonical: the existing chapter owner from Lemma_3_2_7; - bridge/view: the expanded textbook membership formula. This file should therefore reuse the owner directly and keep only the thin bridge needed to match the textbook phrasing, rather than maintaining a parallel local definition halfEllipsoid.
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
local notation "Mat" => Matrix (Fin n) (Fin n) ℝ
open Matrix
open scoped EllipsoidNotation
recall centerCutEllipsoid
(H : Mat) (xBar g : E) :
Set EProposition 3.57
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.57 is the direct recall of the chapter owner theorem constrained_level_total_internal_iterations_le.
recall constrained_level_total_internal_iterations_leTheorem 3.57
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.57: for a bounded feasible set Q, a level-method scalar history history, and a reference value fStar, if the history lower bounds stay below fStar and the textbook block-length estimate holds on every positive-terminal-gap interval where the gap δ_k does not shrink by more than the factor 1 - α, then the method satisfies the stopping criterion after at most ⌊M_f^2 D^2 / (ε^2 α (1 - α)^2 (2 - α))⌋ + 1 iterations; moreover any stopping index k satisfies f_k^* - fStar ≤ ε, with the source-faithful specialization D = Metric.diam Q.
universe u
variable {X : Type u} [PseudoMetricSpace X]
open scoped LevelMethodNotation
theorem levelMethod_stops_and_optimalValue_sub_fStar_le_epsilon_of_diam
{Q : Set X} (hQ_bounded : Bornology.IsBounded Q)
{history : LevelMethodHistory} {fStar : ℝ}
{M_f : NNReal} {ε : Set.Ioi (0 : ℝ)} {α : Set.Ioo (0 : ℝ) 1}
(hvalidLower : ∀ k : ℕ, history.approximateOptimalValue k ≤ fStar)
(hblock :
∀ {k p : ℕ}, k ≤ p →
δ[history](p) ≥ (1 - α) * δ[history](k) →
0 < δ[history](p) →
((p + 1 - k : ℕ) : ℝ) ≤
(M_f ^ (2 : ℕ) * (Metric.diam Q) ^ (2 : ℕ)) /
((1 - α) ^ (2 : ℕ) * δ[history](p) ^ (2 : ℕ))) :
(∃ k ≤ levelMethodIterationCap M_f (Metric.diam Q) ε α,
history.shouldStop ε k) ∧
∀ k : ℕ,
history.shouldStop ε k →
history.optimalValue k - fStar ≤ εDefinition 3.58
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.58: for a feasible sequence X, the best feasible objective value up to step k is the direct owner specialization below.
universe u
variable {α : Type u}
variable {X : ℕ → α}
variable (values : α → ℝ) (k : ℕ)
local notation "sampledValues" => values ∘ X
#check (bestFunctionValueUpTo sampledValues k : ℝ)Definition 3.59
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.59: a set Q satisfies the interior ball condition with parameter ρ when ρ is positive and Q contains an open ball of radius ρ. In the textbook Euclidean setting, this is exactly the interior Euclidean ball condition on Q ⊆ ℝⁿ.
variable {E : Type*} [PseudoMetricSpace E]
open Set Metric
namespace Set
def SatisfiesInteriorBallCondition (Q : Set E) (ρ : ℝ) : Prop :=
0 < ρ ∧ ∃ xBar : E, ball xBar ρ ⊆ QDefinition 3.60
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.60: a polyhedral localization set is a bounded set that admits a finite inner-product half-space presentation Q = {x | ∀ j, ⟪a j, x⟫ ≤ b j}.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
namespace Set
def IsPolyhedralLocalizationSet (Q : Set E) : Prop :=
Bornology.IsBounded Q ∧
∃ (m : ℕ) (a : Fin m → E) (b : Fin m → ℝ), Q = innerLePolyhedron a bDefinition 3.61
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.61: a point y is an inscribed-ellipsoid center of Ek when it is the center of some maximal-volume ellipsoid contained in Ek.
open MeasureTheory
open scoped EllipsoidNotation
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
local notation "Mat" => Matrix (Fin n) (Fin n) ℝ
def IsInscribedEllipsoidCenter (Ek : Set E) (y : E) : Prop :=
∃ H : Mat, IsMaximalVolumeInscribedEllipsoid Ek H yDefinition 3.62
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.62: a point y of the strict positive-slack locus is an analytic center exactly when it globally minimizes the analytic barrier x ↦ -∑ⱼ log (bⱼ - ⟪aⱼ, x⟫) on that strict domain. If one additionally knows that the chosen inequality presentation recovers interior (innerLePolyhedron a b), the same minimizer condition may be transported to that intrinsic interior by the bridge theorem above. The canonical owner is IsMinOn.
open AffineMap
open scoped BigOperators
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable {m : ℕ}
variable (a : Fin m → E) (b : Fin m → ℝ) (y : AnalyticBarrierPoint a b)
recall IsMinOnDefinition 3.63
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.63: for a bounded polyhedron innerLePolyhedron a b with nonempty interior, a volumetric center is a strict-slack point that globally minimizes the volumetric barrier x ↦ log det ∇²F(x) on the strict barrier domain. The canonical owner is IsMinOn, so this file reuses IsMinOn directly on volumetricBarrier a b rather than adding a parallel IsVolumetricCenter wrapper. If one additionally knows that the chosen presentation identifies analyticBarrierDomain a b with interior (innerLePolyhedron a b), the same minimizer condition transports to the intrinsic interior by isMinOn_volumetricBarrier_iff_interior. The boundedness hypothesis belongs only to volumetricBarrier_hessian_det_pos; it is not primitive data of the center predicate itself.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable {m : ℕ}
variable (a : Fin m → E) (b : Fin m → ℝ)
variable [FiniteDimensional ℝ E]
variable (y : AnalyticBarrierPoint a b)
recall IsMinOnDefinition 3.64
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.64: a function belongs to the source-facing class 𝓕_M^{0,0}(Q), written in Lean as 𝓕⁰⁰[M](Q), when it is convex on Q and M-Lipschitz on Q.
universe u
variable {E : Type u} [PseudoMetricSpace E] [AddCommMonoid E] [Module ℝ E]
variable {Q : Set E} {M : NNReal} {f : E → ℝ}
def ConvexLipschitzOn (Q : Set E) (M : NNReal) (f : E → ℝ) : Prop :=
ConvexOn ℝ Q f ∧ LipschitzOnWith M f QDefinition 3.65
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.65 lies in the chapter's finite max-affine model domain for nonsmooth convex optimization. The owner declaration already lives on an arbitrary real inner-product space, so this recall keeps that canonical ambient generality instead of re-specializing it to ℝⁿ. Sampled owner-style declarations: - nonsmoothModel in Lemma_3_3_2, the earlier chapter owner for the finite maximum of sampled affine minorants - nonsmoothModel_apply in Lemma_3_3_2, the pointwise evaluation bridge for that owner - maxTypeObjective in Chap02/Lemma_2_18, the ambient finite-maximum owner pattern specialized by nonsmoothModel - maxTypeObjective_apply in Chap02/Lemma_2_18, the corresponding owner evaluation bridge Best owner abstraction: - nonsmoothModel f xSeq g k Primitive data: - the objective f : E → ℝ - the sample sequence xSeq : ℕ → E - the chosen sampled slopes g : ℕ → E - the prefix length k : ℕ Derived API: - the source-facing notation f̂[xSeq; f; g](k) for the textbook model \hat f_k(X; ·) - the explicit finite-maximum evaluation formula from nonsmoothModel_apply - the later linear-program and quadratic-program reformulations in Definition_3_69 - the later pointwise domination lemmas in Definition_3_73 Source/core/bridge triage: - source-facing: the textbook nonsmooth model f̂[xSeq; f; g](k) = \hat f_k(X; ·) - core/canonical: nonsmoothModel f xSeq g k - bridge/view: nonsmoothModel_apply Definition 3.65 adds no new mathematical data beyond this existing owner declaration. This file is therefore recall-only and introduces no parallel public alias such as sampledAffineMaxModel.
universe u
open scoped NonsmoothModelNotation
recall nonsmoothModel {E : Type u} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
(f : E → ℝ) (xSeq g : ℕ → E) (k : ℕ) :
E → ℝDefinition 3.66
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.66: the complete-data optimization problem is the constrained minimization problem on ℝ × ℝⁿ with objective f(y, x) = max {|y|, ‖x‖²} and feasible set Q = {(y, x) | y² + ‖x‖² ≤ 1}.
open scoped ConstrainedArgmin
variable (n : ℕ)
local notation "Z" => ℝ × EuclideanSpace ℝ (Fin n)
abbrev kelleyCompleteProblem :
SetConstrainedMinimizationProblem ZDefinition 3.67
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.67: the model optimal set Z_k^* is the canonical constrained argmin argmin[Q] (hatf Z).
open scoped ConstrainedArgmin
universe u v
variable {X : Type u} {E : Type v}
variable (Q : Set E) (hatf : X → E → ℝ) (Z : X) (z : E)
local notation "modelProblem" => (SetConstrainedMinimizationProblem.mk Q (hatf Z) :
SetConstrainedMinimizationProblem E)
#check (argmin[Q] (hatf Z) : Set E)Definition 3.68
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.68: the exact model minimum \hat f_k^* is the canonical EReal owner (levelMethodApproximateProblem Q model k).optimalValue. The real history coordinate fhat(history, k) is supplied explicitly via hatf, together with the displayed exactness bridge to the canonical EReal owner.
universe u
open scoped LevelMethodNotation
open scoped LevelMethodNotation
variable
{X : Type u}
(Q : Set X)
(model : ℕ → X → ℝ)
(hatf : ℕ → ℝ)
(hhat :
∀ k : ℕ,
((hatf k : ℝ) : EReal) = (levelMethodApproximateProblem Q model k).optimalValue)
(f : X → ℝ)
(xSeq : ℕ → X)
variable (k : ℕ) (α : ℝ)
#check ((levelMethodApproximateProblem Q model k).optimalValue : EReal)Definition 3.69
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.69 (1): when the model is the finite maximum of the sampled affine minorants, and Q is presented as the polyhedron innerLePolyhedron a b, the minimal model value \hat f_k^* = history.approximateOptimalValue k is the optimal value of the linear program minimizing t subject to the polyhedron inequalities and the sampled affine inequalities.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
open scoped LevelMethodNotation
open scoped NonsmoothModelNotation
variable (Q : Set E) (f : E → ℝ) (xSeq g : ℕ → E) (hatf : ℕ → ℝ)
variable
(hhat :
∀ k : ℕ,
((hatf k : ℝ) : EReal) =
(levelMethodApproximateProblem Q (fun j ↦ f̂[xSeq; f; g](j)) k).optimalValue)
local notation "model" => fun j ↦ f̂[xSeq; f; g](j)
local notation "history" => (levelMethodHistoryFromApproximateValues hatf f xSeq : LevelMethodHistory)
variable {m : ℕ} (a : Fin m → E) (b : Fin m → ℝ) (hQ : Q = innerLePolyhedron a b)
theorem nonsmoothModelOptimalValue_eq_linearProgramValue
(hhat :
∀ k : ℕ,
((hatf k : ℝ) : EReal) =
(levelMethodApproximateProblem Q (fun j ↦ f̂[xSeq; f; g](j)) k).optimalValue)
(hQ : Q = innerLePolyhedron a b)
(k : ℕ) :
fhat(history, k) =
sInf
(Prod.snd ''
{xt : E × ℝ |
(∀ j : Fin m, inner ℝ (a j) xt.1 ≤ b j) ∧
∀ i : Fin (k + 1),
f (xSeq i) + inner ℝ (g i) (xt.1 - xSeq i) ≤ xt.2})Definition 3.70
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.70: the textbook constant M_f is the maximum of the norms ‖g‖ of all subgradients g ∈ ∂f(x) with x ∈ Q; in Lean this is the canonical maximum-attainment predicate on the direct set of attainable subgradient norms.
open scoped WithTopConvexAnalysis
universe u
variable {E : Type u} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
variable (Q : Set E) (f : E → ℝ) (M_f : ℝ)
#check
IsGreatest
({r | ∃ x ∈ Q, ∃ g ∈ ∂ (fun y ↦ (f y : WithTop ℝ))(x), ‖g‖ = r} : Set ℝ)
M_fDefinition 3.71
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.71: the level-parameter objective is the chapter's scalar iteration-cap factor α ↦ α (1 - α)^2 (2 - α), equivalently (1 - α)^2 (1 - (1 - α)^2), on [0, 1].
open Set Real
def levelParameterObjective (α : ℝ) : ℝ :=
α * (1 - α) ^ 2 * (2 - α)Definition 3.72
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.72: a finitely constrained Lipschitz minimization problem on a nonempty bounded closed convex set Q ⊆ ℝⁿ is exactly a general minimization problem whose basic feasible set Q is nonempty bounded closed and convex, whose scalar constraints are all of the form fⱼ(x) ≤ 0, and whose objective and constraint functions are globally Lipschitz on Q. The source spelling uses Fin m; the aggregate-constraint reformulation below does not need any extra nonemptiness hypothesis on that index type, because the owner theorem already handles the empty-index case. The canonical Chapter 1 owner predicate is recovered separately by isFunctionalConstraintProblem_of_lipschitz.
namespace GeneralMinimizationProblem
variable {n m : ℕ}
variable (problem : GeneralMinimizationProblem n m)
#check (
0 < m ∧
problem.basicFeasibleSet.Nonempty ∧
Bornology.IsBounded problem.basicFeasibleSet ∧
IsClosed problem.basicFeasibleSet ∧
Convex ℝ problem.basicFeasibleSet ∧
problem.HasLeConstraints ∧
(∃ L : NNReal, LipschitzWith L problem.objective) ∧
∀ j : Fin m, ∃ L : NNReal, LipschitzWith L (problem.constraints j)
)
Definition 3.73
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.73: both textbook sampled models \hat f_k(X; ·) and \check f_k(X; ·) are direct owner specializations of the same chapter owner nonsmoothModel, obtained by instantiating (φ, g) with (f, g) and (fBar, gBar) respectively.
universe u
variable {E : Type u} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
open scoped NonsmoothModelNotation
variable (φ : E → ℝ) (X : ℕ → E) (g : ℕ → E) (k : ℕ)
#check (f̂[X; φ; g](k) : E → ℝ)Definition 3.74
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.74: the textbook parametric function f(t; x) = max {f(x) - t, \bar f(x)} is the chapter owner setConstrainedParametricObjective f fBar, and its value function f^*(t) = \min_{x ∈ Q} f(t; x) is the owner parametricValueFunction Q f fBar.
recall setConstrainedParametricObjectiveDefinition 3.75
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.75: for the sampled history X = {x_k} and a fixed stage k, the corresponding model value function \hat f_k^*(X; ·) is the direct specialization of the chapter owner parametricValueFunction to the model pair (\hat f_k(X; ·), \check f_k(X; ·)).
universe u
variable {X : Type u}
variable (Q : Set X)
variable (hatModel checkModel : (ℕ → X) → ℕ → X → ℝ) (xSeq : ℕ → X) (k : ℕ)
#check (parametricValueFunction Q (hatModel xSeq k) (checkModel xSeq k) : ℝ → EReal)Definition 3.76
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.76: the textbook record value f_k^*(X; t) is the best sampled value of the fixed-stage parametric model from Definition 3.75.
universe u
variable {X : Type u}
variable (hatModel checkModel : (ℕ → X) → ℕ → X → ℝ) (xSeq : ℕ → X) (k : ℕ) (t : ℝ)
local notation "stageObjective" =>
setConstrainedParametricObjective (hatModel xSeq k) (checkModel xSeq k)
local notation "sampledValues" =>
stageObjective t ∘ xSeq
#check (bestFunctionValueUpTo sampledValues k : ℝ)Definition 3.77
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.77: the textbook constant M_f is the maximum of the norms ‖g‖ of all relative subgradients g ∈ ∂_Q f(x) ∪ ∂_Q \bar f(x) at feasible base points x; in Lean this is the canonical maximum-attainment predicate on the direct set of attainable relative-subgradient norms, with feasibility carried by subdifferentialWithin.
open scoped WithTopConvexAnalysis
universe u
variable {E : Type u} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
variable (Q : Set E) (f fBar : E → ℝ) (M_f : ℝ)
#check
IsGreatest
({r | ∃ x g, g ∈ ∂[Q] f(x) ∪ ∂[Q] fBar(x) ∧ ‖g‖ = r} : Set ℝ)
M_fDefinition 3.78
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.78: at fixed sampled affine models, the auxiliary slice x ↦ max (hatModel x - t) (checkModel x) is the direct owner specialization below.
universe u
variable {E : Type u} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
variable {k : ℕ}
variable (Q : Set E)
variable (f fBar : E → ℝ)
variable (xSample gSample gBarSample : Fin (k + 1) → E)
local notation "hatMinorant" =>
fun j : Fin (k + 1) ↦
sampledAffineMinorant (xSample j) (gSample j) (f (xSample j))
local notation "checkMinorant" =>
fun j : Fin (k + 1) ↦
sampledAffineMinorant (xSample j) (gBarSample j) (fBar (xSample j))
local notation "hatMinorantFn" =>
fun j : Fin (k + 1) ↦ (hatMinorant j : E → ℝ)
local notation "checkMinorantFn" =>
fun j : Fin (k + 1) ↦ (checkMinorant j : E → ℝ)
local notation "hatModel" =>
maxTypeObjective hatMinorantFn
local notation "checkModel" =>
maxTypeObjective checkMinorantFn
local notation "hatFn" =>
fun _ : Unit ↦ fun _ : Unit ↦ hatModel
local notation "checkFn" =>
fun _ : Unit ↦ fun _ : Unit ↦ checkModel
#check (setConstrainedParametricObjective hatModel checkModel : ℝ → E → ℝ)Definition 3.79
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.79: for a nonempty finite family f₁, …, f_m, sample points x₀, …, x_k, and chosen vectors g_i(x_j), the complete model is the maximum over all sampled affine minorants f_i(x_j) + ⟪g_i(x_j), x - x_j⟫. The associated pointwise maximum function \bar f(x) = max_i f_i(x) is already the owner maxTypeObjective f.
universe u
variable {m k : ℕ}
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable [NeZero m] (f : Fin m → E → ℝ) (xSample : Fin (k + 1) → E)
variable (subgradient : Fin (k + 1) → Fin m → E)
abbrev completeModel : E → ℝ :=
maxTypeObjective fun ji : Fin (k + 1) × Fin m ↦
fun x ↦
f ji.2 (xSample ji.1) + inner ℝ (subgradient ji.1 ji.2) (x - xSample ji.1)3.1 — General Convex Functions
Corollary 3.1.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Corollary 3.1.1 is a source-facing recall in the convex-analysis maximum-principle domain. Primary domain: - convex functions and the finite convex-hull maximum principle Sampled owner-style declarations: - ConvexOn.map_sum_le - ConvexOn.le_sup_of_mem_convexHull - ConvexOn.inf_le_of_mem_convexHull - convexOn_value_le_max_of_convex_combination Best owner abstraction: - ConvexOn.le_sup_of_mem_convexHull Primitive data: - a finite vertex set t : Finset E - a convex function hf : ConvexOn ℝ s f - the inclusion (t : Set E) ⊆ s - a point x with x ∈ convexHull ℝ (t : Set E) Derived API: - the canonical bound f x ≤ t.sup' ... f Source/core/bridge triage: - source-facing: the textbook corollary for finitely many points x₁, …, xₘ - core/canonical: ConvexOn.le_sup_of_mem_convexHull - bridge/view: the earlier chapter theorem convexOn_value_le_max_of_convex_combination, which packages the convex-combination input in the chapter owner is_convex_combination_of; the textbook coefficient display is kept only in the companion theorem convexOn_value_le_max_of_convex_combination_of_coefficients The coefficient-based chapter statement is bridge API, while this corollary is governed directly by the convex-hull owner theorem. This file therefore recalls the canonical owner declaration instead of keeping a specialized local shell.
recall ConvexOn.le_sup_of_mem_convexHullLemma 3.1.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.1 is a recall-only item in the finite Jensen / convex-combination domain. Primary mathematical domain: - Jensen's inequality for convex functions on finite weighted sums Relevant owner-style declarations sampled before refinement: - ConvexOn.map_centerMass_le - ConvexOn.map_sum_le - ConvexOn.map_add_sum_le - Convex.sum_mem Best owner abstraction: - ConvexOn.map_sum_le Primitive data: - a convex-on-set witness hf : ConvexOn ℝ s f - a finite family x : Fin m → E with x i ∈ s - a coefficient family α : Fin m → ℝ Derived API: - Jensen's inequality at the weighted sum ∑ i, α i • x i - the positivity-to-nonnegativity specialization 0 < α i → 0 ≤ α i, used only at call sites - the separate domain-membership consequence from Convex.sum_mem Source/core/bridge triage: - source-facing: the textbook finite Jensen inequality with positive coefficients - core/canonical: ConvexOn.map_sum_le - bridge/view: the redundant positivity specialization from the source wording to the owner theorem's nonnegative-weight interface The exact mathematical content here is already owned canonically by ConvexOn.map_sum_le, and the earlier chapter file Lemma_3_1.lean already centers that owner theorem. This file therefore does not keep a second local example proof with the stronger positivity hypotheses; those hypotheses are bridge-only and should be discharged at the use site by passing fun i ↦ (hαpos i).le into the owner theorem.
recall ConvexOn.map_sum_leTheorem 3.1.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.1: an ℝ ∪ {+∞}-valued function on ℝⁿ is convex on its effective domain if and only if its effective domain is convex and every forward affine extrapolation point y + β • (y - x) that remains in that domain satisfies the secant-line lower bound determined by x and y. The statement is generalized from the textbook ℝⁿ setting to an arbitrary real module.
universe u
open scoped WithTopConvexAnalysis
variable {X : Type u} [AddCommGroup X] [Module ℝ X]
theorem convexOn_effectiveDomain_iff_affine_ray_inequality
(f : X → WithTop ℝ) :
ConvexOn ℝ (dom f) (withTopRealPart f) ↔
Convex ℝ (dom f) ∧
∀ ⦃x y : X⦄, x ∈ dom f → y ∈ dom f →
∀ ⦃β : ℝ⦄, 0 ≤ β →
y + β • (y - x) ∈ dom f →
withTopRealPart f (y + β • (y - x)) ≥
withTopRealPart f y + β * (withTopRealPart f y - withTopRealPart f x)Corollary 3.1.1.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Corollary 3.1.1.1 lies in the convex-analysis domain of the finite convex-hull maximum principle. Sampled owner-style declarations: - ConvexOn.map_sum_le - ConvexOn.le_sup_of_mem_convexHull - ConvexOn.inf_le_of_mem_convexHull - convexOn_value_le_max_of_convex_combination Best owner abstraction: - ConvexOn.le_sup_of_mem_convexHull Primitive data: - a finite set t : Finset E - a convex function hf : ConvexOn ℝ s f - the inclusion (t : Set E) ⊆ s - a point x with x ∈ convexHull ℝ (t : Set E) Derived API: - the canonical bound f x ≤ t.sup' ... f Source/core/bridge triage: - source-facing: the corollary that the value at a convex combination is bounded by the maximum of the endpoint values - core/canonical: ConvexOn.le_sup_of_mem_convexHull - bridge/view: the chapter theorem convexOn_value_le_max_of_convex_combination, which uses the chapter owner is_convex_combination_of; the displayed equality x = ∑ i, α i • points i is kept only in the companion theorem convexOn_value_le_max_of_convex_combination_of_coefficients The coefficient-level theorem is bridge API, not the owner of this item. This file therefore recalls the canonical convex-hull maximum principle directly; when the source presentation names a point by an equality x = ∑ i, α i • points i, that equality should be used only as a bridge to x ∈ convexHull ℝ (Set.range points) at the call site or in a companion theorem.
recall ConvexOn.le_sup_of_mem_convexHullDefinition 3.1.1.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.1.1, generalized from the textbook ℝⁿ setting: a general convex minimization problem consists of minimizing an ambient extended-real-valued convex objective function over a closed convex set Q ⊆ X, subject to finitely many ambient extended-real-valued convex inequality constraints fᵢ(x) ≤ 0; no differentiability is assumed for the objective or constraint functions.
universe u
open scoped WithTopConvexAnalysis
structure GeneralConvexMinimizationProblem
(X : Type u) [TopologicalSpace X] [AddCommMonoid X] [Module ℝ X] (m : ℕ) where
/-- The ambient feasible set `Q ⊆ X`. -/
feasibleSet : Set X
/-- The ambient extended-real-valued objective function. -/
objective : X → WithTop ℝ
/-- The finitely many ambient extended-real-valued inequality constraints. -/
constraints : Fin m → X → WithTop ℝ
/-- The ambient feasible set `Q` is closed. -/
feasibleSet_closed : IsClosed feasibleSet
/-- The ambient feasible set `Q` is convex. -/
feasibleSet_convex : Convex ℝ feasibleSet
/-- The objective is convex in the Chapter 3 `WithTop` sense. -/
objective_convex : ConvexOn ℝ (dom objective) (withTopRealPart objective)
/-- Each inequality constraint is convex in the Chapter 3 `WithTop` sense. -/
constraints_convex (i : Fin m) :
ConvexOn ℝ (dom (constraints i)) (withTopRealPart (constraints i))Lemma 3.1.1.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.1.1 is a recall-only Euclidean specialization in the finite Jensen domain. Primary mathematical domain: - Jensen's inequality for convex functions on finite weighted sums Relevant owner-style declarations sampled before refinement: - Lemma_3_1_1, the earlier chapter recall of ConvexOn.map_sum_le - ConvexOn.map_centerMass_le - ConvexOn.map_sum_le - Convex.sum_mem Best owner abstraction: - ConvexOn.map_sum_le Primitive data: - a convex-on-set witness hf : ConvexOn ℝ s f - a finite family x : Fin m → EuclideanSpace ℝ (Fin n) with x i ∈ s - a coefficient family α : Fin m → ℝ summing to 1 Derived API: - the Euclidean specialization of Jensen's inequality - the positivity-to-nonnegativity bridge 0 < α i → 0 ≤ α i, which belongs only at use sites Source/core/bridge triage: - source-facing: the textbook Euclidean specialization - core/canonical: ConvexOn.map_sum_le - bridge/view: specializing the ambient module to EuclideanSpace ℝ (Fin n) This item adds no new mathematics beyond the owner theorem. The source's positivity wording is a use-site bridge to the owner's nonnegativity hypothesis, so the file keeps the owner declaration as the main public entry and lets the Euclidean specialization elaborate directly at call sites.
recall ConvexOn.map_sum_leProposition 3.1.1.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.1.1.1 lies in the chapter's extended-real effective-epigraph closedness domain. Primary domain: - closedness of effective epigraphs of EReal-valued functions on topological spaces, expressed through continuity of the finite real part on the effective domain. Sampled owner-style declarations: - chapter dom f and extendedRealRealPart f from Definition_3_1_1_3, the canonical finite-value owner/bridge vocabulary in this domain; - isClosed_effectiveEpigraph_of_continuousOn_of_isClosed_dom in Proposition_3_2, the source-facing chapter theorem on effectiveEpigraph f; - isClosed_effectiveEpigraph_of_lowerSemicontinuousOn_of_isClosed_dom in Proposition_3_2, the stronger companion obtained by weakening continuity to lower semicontinuity; - mathlib ContinuousOn.lowerSemicontinuousOn, the canonical bridge from the source-facing continuity statement to the companion strengthening; - mathlib LowerSemicontinuousOn.isClosed_re_epigraph, the canonical real-epigraph closedness owner used by the chapter theorem. Best owner abstraction: - the existing chapter theorem isClosed_effectiveEpigraph_of_continuousOn_of_isClosed_dom. Primitive data: - the EReal-valued function f; - continuity of extendedRealRealPart f on dom f; - closedness of dom f. Derived API: - closedness of the effective epigraph {p : X × ℝ | p.1 ∈ dom f ∧ f p.1 ≤ p.2}. - companion strengthening under lower semicontinuity on dom f. Source/core/bridge triage: - source-facing: this effective-epigraph closedness proposition; - core/canonical: mathlib lower-semicontinuity / isClosed_re_epigraph; - bridge/view: the chapter theorem isClosed_effectiveEpigraph_of_continuousOn_of_isClosed_dom. This file therefore remains recall-only: the proposition-level statement already exists upstream in the minimal chapter closure with the correct canonical dom f owner surface, so introducing a second local theorem here would only recreate the duplicate-wheel problem.
recall isClosed_effectiveEpigraph_of_continuousOn_of_isClosed_domTheorem 3.1.1.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.1.1: an extended-real-valued function on a real additive module is convex exactly when its effective domain is convex and every forward affine extrapolation point y + β • (y - x) in that domain satisfies the supporting-line inequality determined by the secant through x and y. Specializing X to EuclideanSpace ℝ (Fin n) recovers the textbook ℝⁿ statement.
universe u
open scoped ConvexAnalysis
variable {X : Type u} [AddCommGroup X] [Module ℝ X]
theorem isConvexExtendedRealFunction_iff_affine_ray_inequality
(f : X → EReal) :
ConvexOn ℝ (dom f) (extendedRealRealPart f) ↔
Convex ℝ (dom f) ∧
∀ ⦃x y : X⦄, x ∈ dom f → y ∈ dom f →
∀ ⦃β : ℝ⦄, 0 ≤ β →
y + β • (y - x) ∈ dom f →
extendedRealRealPart f (y + β • (y - x)) ≥
extendedRealRealPart f y +
β * (extendedRealRealPart f y - extendedRealRealPart f x)Corollary 3.1.1.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Corollary 3.1.1.2 lies in the convex-analysis domain of the finite convex-hull maximum principle. Sampled owner-style declarations: - ConvexOn.exists_ge_of_mem_convexHull - ConvexOn.le_sup_of_mem_convexHull - ConvexOn.inf_le_of_mem_convexHull Best owner abstraction: - ConvexOn.exists_ge_of_mem_convexHull Primitive data: - a vertex set t : Set E, or equivalently a finite vertex family viewed through Set.range - a set s and a convex function hf : ConvexOn ℝ s f - the inclusion t ⊆ s - a point x ∈ convexHull ℝ t Derived API: - a vertex y ∈ t with f x ≤ f y Source/core/bridge triage: - source-facing: the corollary that the maximum of a convex function on a finite convex hull is attained at a vertex - core/canonical: ConvexOn.exists_ge_of_mem_convexHull - bridge/view: the finite-sup theorem ConvexOn.le_sup_of_mem_convexHull and coefficient or index-level attainment formulations derived from the owner theorem The chapter theorem convexOn_value_le_max_of_convex_combination is a bridge/view reformulation; this file keeps only the canonical witness-level convex-hull maximum-principle recall.
recall ConvexOn.exists_ge_of_mem_convexHullDefinition 3.1.1.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.1.2: the chapter's effective-domain owner for an extended-real-valued function, namely the set of points where the value is a finite real. The textbook then treats the nonemptiness condition on this domain as a standing assumption.
universe u
abbrev extendedRealEffectiveDomain {X : Type u} (f : X → EReal) : Set X :=
{x | f x ≠ ⊤ ∧ f x ≠ ⊥}Lemma 3.1.1.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.1.2 is a source-facing recall in the chapter's closed-convex WithTop-valued convex-analysis domain. Primary domain: - restriction of a closed convex extended-real-valued function to a closed convex subset. Sampled owner-style declarations: - ClosedConvexOn - ClosedConvexOn.restrict - ClosedConvexOn.isClosed_constrainedEpigraph - mathlib ConvexOn.subset Best owner abstraction: - ClosedConvexOn Primitive data: - a witness hf : ClosedConvexOn Q f - the subset, closedness, and convexity data for Q₁ ⊆ Q Derived API: - the canonical owner theorem ClosedConvexOn.restrict Source/core/bridge triage: - source-facing: the restriction lemma for closed convex functions on a closed convex subset - core/canonical: ClosedConvexOn - bridge/view: constrainedEpigraph together with mathlib ConvexOn.subset The main entry now reuses the owner theorem at the same ambient abstraction level as ClosedConvexOn itself, instead of recalling a later ℝⁿ-specialized duplicate.
universe u
variable {X : Type u} [TopologicalSpace X] [AddCommMonoid X] [Module ℝ X]
recall ClosedConvexOn.restrict
{f : X → WithTop ℝ} {Q Q₁ : Set X}
(hf : ClosedConvexOn Q f)
(hQ₁_closed : IsClosed Q₁)
(hQ₁_convex : Convex ℝ Q₁)
(hQ₁Q : Q₁ ⊆ Q) :
ClosedConvexOn Q₁ fProposition 3.1.1.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.1.1.2 is the direct recall of the chapter's half-space description of the epigraph of the absolute value function.
recall abs_epigraph_eq_inter_halfspacesTheorem 3.1.1.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.1.2: the owner ConvexOn formulation for the finite real part of an extended-real-valued function is equivalent to convexity of its textbook effective epigraph in X × ℝ.
universe u
open scoped ConvexAnalysis
variable {X : Type u} [AddCommMonoid X] [Module ℝ X]
theorem isConvexExtendedRealFunction_iff_convex_epigraph
(f : X → EReal) :
ConvexOn ℝ (dom f) (extendedRealRealPart f) ↔
Convex ℝ (effectiveEpigraph f)Definition 3.1.1.3
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.1.3 (2): an EReal-valued function is strictly convex when its finite real part is strictly convex on the same effective domain.
universe u
open scoped ConvexAnalysis
variable {X : Type u} [AddCommMonoid X] [Module ℝ X]
variable (f : X → EReal)
#check StrictConvexOn ℝ (dom f) (extendedRealRealPart f)Proposition 3.1.1.3
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.1.1.3, generalized from the textbook ℝⁿ setting: a continuous convex real-valued function on a real topological module is a closed convex function; equivalently, after coercing f to WithTop ℝ, its epigraph is a closed convex subset of X × ℝ.
universe u
open scoped WithTopConvexAnalysis
variable {X : Type u} [TopologicalSpace X] [AddCommMonoid X] [Module ℝ X]
theorem closedConvexFunction_coe_of_convexOn_continuous
{f : X → ℝ}
(hf_convex : ConvexOn ℝ Set.univ f) (hf_cont : Continuous f) :
ClosedConvexFunction (fun x ↦ (f x : WithTop ℝ))Theorem 3.1.1.3
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.1.3: if f is convex, then each sublevel set {x ∈ dom f | extendedRealRealPart f x ≤ β} is convex. This is exactly the canonical owner theorem ConvexOn.convex_le, specialized to the chapter bridge extendedRealRealPart; the source-facing sublevel set {x ∈ dom f | f x ≤ β} is identified with this owner surface by extendedRealSublevelSet_dom_eq f.
universe u
open scoped ConvexAnalysis
variable {X : Type u} [AddCommMonoid X] [Module ℝ X]
variable (f : X → EReal)
#check
(ConvexOn.convex_le :
ConvexOn ℝ (dom f) (extendedRealRealPart f) →
∀ β : ℝ, Convex ℝ {x ∈ dom f | extendedRealRealPart f x ≤ β})Definition 3.1.1.4
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.1.4: a point is a convex combination of a finite family of points when it is the canonical convex combination associated to some simplex weight vector on that family.
open scoped BigOperators
open ConvexSpace
universe u v w
-- Declarations for this item will be appended below by the statement pipeline.
variable (R : Type u) [PartialOrder R] [Semiring R] [IsStrictOrderedRing R]
variable {ι : Type v}
variable {E : Type w} [ConvexSpace R E]
def is_convex_combination_of (points : ι → E) (x : E) : Prop :=
∃ w : StdSimplex R ι, x = convexCombination (w.map points)
Proposition 3.1.1.4
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.1.1.4 lies in the chapter's one-dimensional extended-real convex-analysis domain. Primary domain: - the reciprocal function on (0, ∞) and its WithTop ℝ epigraph/lower-semicontinuity package. Sampled owner-style declarations: - chapter positiveReciprocalExtension - chapter reciprocalEpigraphOnPositiveRay - chapter dom f and constrainedEpigraph Q f from Definition_3_3 - mathlib strictConvexOn_zpow - mathlib lowerSemicontinuous_iff_isClosed_epigraph Best owner abstraction: - the existing chapter owner declaration positiveReciprocalExtension from Proposition_3_5, together with the Chapter 2 owner set reciprocalEpigraphOnPositiveRay and the chapter canonical epigraph/domain owners constrainedEpigraph and dom from Definition_3_3. Primitive data: - the extended reciprocal function on ℝ Derived API: - the epigraph identification bridge to reciprocalEpigraphOnPositiveRay - convexity on Set.Ioi 0 - lower semicontinuity - closedness of the epigraph - identification and openness of the effective domain Source/core/bridge triage: - source-facing: the reciprocal-extension proposition package already owned by Proposition_3_5 - core/canonical: reciprocalEpigraphOnPositiveRay, dom, constrainedEpigraph, and the mathlib convexity/lower-semicontinuity owners - bridge/view: this numbering-local file, which should only recall the upstream chapter owner declarations and canonical expressions instead of reintroducing a parallel duplicate API This file is therefore recall-only. Keeping local names such as reciprocalWithTop would create a second public owner for the same mathematics with the same interface, which is exactly the duplicate-wheel pattern the chapter policy forbids.
recall positiveReciprocalExtensionTheorem 3.1.1.4
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.1.4: every real sublevel set of a closed convex function is closed and convex.
universe u
open scoped WithTopConvexAnalysis
variable {X : Type u} [TopologicalSpace X] [AddCommMonoid X] [Module ℝ X]
theorem ClosedConvexFunction.isClosed_convex_sublevelSet
{f : X → WithTop ℝ} (hf : ClosedConvexFunction f) (β : ℝ) :
IsClosed {x | f x ≤ β} ∧ Convex ℝ {x | f x ≤ β}Definition 3.1.1.5
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.1.5, generalized from the textbook ℝⁿ setting: an ℝ ∪ {+∞}-valued function is closed and convex on Q when Q ⊆ dom f and its constrained epigraph is a closed convex subset of X × ℝ.
universe u
open scoped WithTopConvexAnalysis
variable {X : Type u} [TopologicalSpace X] [AddCommMonoid X] [Module ℝ X]
def ClosedConvexOn (Q : Set X) (f : X → WithTop ℝ) : Prop :=
Q ⊆ dom f ∧
IsClosed (constrainedEpigraph Q f) ∧
Convex ℝ (constrainedEpigraph Q f)Proposition 3.1.1.5
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.1.1.5 lies in the chapter's closed-convex WithTop-valued convex-analysis domain. Primary domain: - closed convexity of seminorm functions on finite-dimensional real normed spaces, with the textbook ℝⁿ case as a specialization. Sampled owner-style declarations: - ClosedConvexOn and ClosedConvexFunction in Definition_3_1_1_5, the chapter owners for closed convex extended-real-valued functions; - Seminorm.closedConvexFunction in Proposition_3_6, the existing chapter theorem giving the canonical owner statement for x ↦ p x; - mathlib Seminorm.convexOn, the canonical convexity owner behind the theorem; - mathlib ConvexOn.continuousOn, the canonical continuity bridge on the open owner domain Set.univ. Best owner abstraction: - ClosedConvexFunction (fun x : E ↦ (p x : WithTop ℝ)). Primitive data: - a seminorm p : Seminorm ℝ E. Derived API: - the owner theorem Seminorm.closedConvexFunction; - the ClosedConvexOn Set.univ specialization obtained from the owner abbreviation because norm functions are finite everywhere. Source/core/bridge triage: - source-facing: the proposition that a norm function is closed and convex on all of ℝⁿ; - core/canonical: ClosedConvexFunction / ClosedConvexOn; - bridge/view: the identification of the norm function's effective domain with Set.univ. This file is therefore recall-only. Proposition_3_6 already introduced the correct owner-level theorem in the minimal chapter closure, so keeping another named theorem here would only duplicate that API under a second shell.
universe u
variable {E : Type u} [NormedAddCommGroup E] [NormedSpace ℝ E] [FiniteDimensional ℝ E]
recall Seminorm.closedConvexFunction
{E : Type u} [NormedAddCommGroup E] [NormedSpace ℝ E] [FiniteDimensional ℝ E]
(p : Seminorm ℝ E) :
ClosedConvexFunction (fun x : E ↦ (p x : WithTop ℝ))Definition 3.1.1.6
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.1.6 is a recall-only item in the seminorm-geometry domain. Layer targeted by this refinement: - source-facing recall of the core/canonical owner closed-ball API for seminorms Primary domain: - closed balls attached to seminorms. Sampled owner-style declarations: - Seminorm.closedBall - Seminorm.mem_closedBall - Seminorm.mem_closedBall_zero - Seminorm.closedBall_zero_eq Primitive data: - a seminorm p : Seminorm ℝ E - a center x₀ : E - a radius r : ℝ Derived API: - Seminorm.mem_closedBall gives the set-builder characterization x ∈ p.closedBall x₀ r ↔ p (x - x₀) ≤ r. Source/core/bridge triage: - source-facing: the textbook closed ball attached to a seminorm - core/canonical: Seminorm.closedBall - bridge/view: the defining membership lemma Seminorm.mem_closedBall This file therefore keeps only the general owner declaration and its defining bridge view. Downstream unit-ball files should use the same owner declaration directly and recall the zero-center specializations there.
recall Seminorm.closedBallProposition 3.1.1.6
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.1.1.6 lies in the chapter's finite-dimensional ℓ₁-geometry / convex-hull domain. Layer targeted by this refinement: - source-facing proposition stated at the canonical ℓ₁-closed-ball owner layer Sampled owner-style declarations: - EuclideanSpace.l1Seminorm - Seminorm.closedBall - l1_ball_eq_convexHull_signed_standard_basis_prop Best owner abstraction: - (EuclideanSpace.l1Seminorm n).closedBall x₀ r Primitive data: - the ambient dimension n : ℕ - the center x₀ : E - the radius r : ℝ Derived API: - the convex-hull description of that closed ball by the signed standard-basis vertices, already established in Proposition_3_7 Dimension side condition: - 0 < n, excluding the degenerate n = 0 case where the signed-vertex set is empty while the closed ball is nonempty for r ≥ 0 Source/core/bridge triage: - source-facing: the ℓ₁ closed ball / signed-vertex convex-hull equality - core/canonical: Seminorm.closedBall for EuclideanSpace.l1Seminorm n - bridge/view: this recall-only file Proposition_3_7 now already states the proposition on the canonical closed-ball owner surface. Keeping a second theorem here with the same interface would only duplicate that owner-level API under a different name, so this file is recall-only.
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
local notation "e[" i "]" => EuclideanSpace.single i (1 : ℝ)
recall l1_ball_eq_convexHull_signed_standard_basis_prop
(hn : 0 < n) (x₀ : E) (r : ℝ) (hr : 0 ≤ r) :
(EuclideanSpace.l1Seminorm n).closedBall x₀ r =
convexHull ℝ
(Set.range (fun i : Fin n ↦ x₀ + r • e[i]) ∪
Set.range (fun i : Fin n ↦ x₀ - r • e[i]))Definition 3.1.1.7
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.1.7: membership in the textbook ℓ_p closed ball centered at x₀ and of radius r is equivalent to the usual coordinate ℓ_p inequality.
open scoped BigOperators
open scoped EuclideanSpaceLp
theorem mem_lp_closedBall_coord_iff (n : ℕ) (p : EuclideanSpace.LpExponent)
(x₀ x : EuclideanSpace ℝ (Fin n)) (r : ℝ) :
x ∈ (EuclideanSpace.lpSeminorm n p).closedBall x₀ r ↔
(∑ i, |(x - x₀) i| ^ p.toReal) ^ (1 / p.toReal : ℝ) ≤ rProposition 3.1.1.7
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.1.1.7 lies in the chapter's source-facing unit-disk boundary-extension domain. Primary domain: - lower-semicontinuity of a WithTop ℝ-valued boundary extension on the Euclidean unit disk. Sampled owner-style declarations: - unitDiskBoundaryExtension in Proposition_3_8, the existing source-facing owner for the textbook construction; - unitDiskBoundaryExtension_convex_and_effectiveDomain in Proposition_3_8, the companion theorem supplying the owner-level convexity and effective-domain data of the same owner; - unitDiskBoundaryExtension_lowerSemicontinuous_iff_eq_zero in Proposition_3_8, which already has the exact textbook interface of this item; - mathlib LowerSemicontinuous together with the chapter's closed-convex recall surface in Theorem_3_1_4. Best owner abstraction: - the imported source-facing owner unitDiskBoundaryExtension. Primitive data: - the boundary datum φ : Metric.sphere (0 : EuclideanSpace ℝ (Fin 2)) 1 → ℝ; - the owner construction unitDiskBoundaryExtension φ. Derived API: - the open-disk value theorem already proved in Proposition_3_8; - the convexity/effective-domain companion theorem unitDiskBoundaryExtension_convex_and_effectiveDomain; - the lower-semicontinuity criterion unitDiskBoundaryExtension_lowerSemicontinuous_iff_eq_zero. Source/core/bridge triage: - source-facing: the unit-disk boundary extension and its vanishing criterion on the unit circle; - core/canonical: LowerSemicontinuous for WithTop ℝ-valued functions; - bridge/view: the owner-level convexity and effective-domain companion theorem from Proposition_3_8. This file previously rebuilt the same owner and theorem family under duplicate local declarations. Since Proposition_3_8 already provides the correct source-facing owner in the minimal chapter closure, this item is recall-only and reuses that canonical chapter declaration directly.
recall unitDiskBoundaryExtension_lowerSemicontinuous_iff_eq_zeroCorollary 3.1.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Corollary 3.1.2 lies in the convex-analysis domain of finite convex-hull maximum principles. Sampled owner-style declarations: - ConvexOn.exists_ge_of_mem_convexHull - ConvexOn.le_sup_of_mem_convexHull - ConvexOn.inf_le_of_mem_convexHull Best owner abstraction: - ConvexOn.exists_ge_of_mem_convexHull Primitive data: - a vertex set t : Set E, or equivalently a finite vertex family viewed through Set.range - a set s and a convex function hf : ConvexOn ℝ s f - the inclusion t ⊆ s - a point x ∈ convexHull ℝ t Derived API: - a vertex y ∈ t with f x ≤ f y Source/core/bridge triage: - source-facing: the corollary that the maximum of a convex function on a simplex is attained at a vertex once the vertices lie in the convex domain - core/canonical: ConvexOn.exists_ge_of_mem_convexHull - bridge/view: the finite-sup theorem ConvexOn.le_sup_of_mem_convexHull and coefficient or index-level attainment formulations derived from the owner theorem The chapter theorem convexOn_value_le_max_of_convex_combination is a bridge/view reformulation. This file keeps only the canonical witness-level convex-hull maximum-principle recall.
recall ConvexOn.exists_ge_of_mem_convexHullLemma 3.1.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.2 is a source-facing recall in the chapter's WithTop-valued convex-analysis domain. Primary domain: - restriction of a closed convex extended-real-valued function to a closed convex subset. Sampled owner-style declarations in this domain: - ClosedConvexOn from Definition_3_1_1_5 - ClosedConvexOn.restrict from Definition_3_1_1_5 - ClosedConvexOn.subset_withTopEffectiveDomain - ClosedConvexOn.isClosed_constrainedEpigraph - ClosedConvexOn.convex_constrainedEpigraph Best owner abstraction: - ClosedConvexOn Primitive data: - the owner witness hf : ClosedConvexOn Q f - the subset, closedness, and convexity data for Q₁ ⊆ Q Derived API: - the canonical owner theorem ClosedConvexOn.restrict Source/core/bridge triage: - source-facing: ClosedConvexOn.restrict - core/canonical: ClosedConvexOn - bridge/view: constrainedEpigraph, together with the closed/convex cylinder intersection proof internalized in the owner theorem This file is recall-only: the owner theorem now lives where ClosedConvexOn itself is defined, so the later chapter item reuses that exact owner theorem instead of introducing a duplicate specialized copy.
universe u
variable {X : Type u} [TopologicalSpace X] [AddCommMonoid X] [Module ℝ X]
recall ClosedConvexOn.restrict
{f : X → WithTop ℝ} {Q Q₁ : Set X}
(hf : ClosedConvexOn Q f)
(hQ₁_closed : IsClosed Q₁)
(hQ₁_convex : Convex ℝ Q₁)
(hQ₁Q : Q₁ ⊆ Q) :
ClosedConvexOn Q₁ fTheorem 3.1.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.2 is recall-only in the real convex-analysis / epigraph domain. Primary domain: - convexity of the epigraph of a real-valued function on a convexity domain, with the textbook specialization s ⊆ ℝⁿ. Sampled owner-style declarations: - mathlib ConvexOn; - mathlib ConvexOn.convex_epigraph; - mathlib convexOn_iff_convex_epigraph; - the chapter bridge convexOn_iff_convex_effective_epigraph in Theorem_3_3, which handles the WithTop ℝ-valued effective-epigraph variant by reducing to the same owner theorem. Best owner abstraction: - convexOn_iff_convex_epigraph. Primitive data: - a scalar type 𝕜, ambient module E, and codomain ordered additive module β in the owner theorem; - in the textbook specialization, 𝕜 = ℝ, E = EuclideanSpace ℝ (Fin n), and β = ℝ. Derived API: - convexity of the epigraph {p : E × ℝ | p.1 ∈ s ∧ f p.1 ≤ p.2}; - the one-direction owner theorem ConvexOn.convex_epigraph. Source/core/bridge triage: - source-facing: the Euclidean textbook epigraph characterization of convexity; - core/canonical: mathlib convexOn_iff_convex_epigraph; - bridge/view: the Euclidean specialization recorded below. The previous declaration convexOn_iff_convex_epigraph_euclidean was an exact-interface duplicate of the mathlib owner theorem. This file therefore recalls the canonical owner directly instead of keeping a second public theorem name for the same mathematics; the textbook Euclidean statement is its immediate specialization.
recall convexOn_iff_convex_epigraphCorollary 3.1.2.1
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Corollary 3.1.2.1: if f₁, …, f_m are closed convex real-valued functions on Q and every constrained level set of their finite pointwise maximum is bounded, then there exists a coefficient vector in the standard simplex StdSimplex ℝ ι whose canonical convex combination of the family values has the same constrained minimum value as that finite maximum; in Lean the boundedness hypothesis is expressed through the chapter owner constrainedSublevelSet Q (fun x ↦ ((maxTypeObjective fs x : ℝ) : WithTop ℝ)) α, and the conclusion is recorded as an equality of EReal infima over the subtype Q between maxTypeObjective fs and the textbook simplex-weighted objective fun x ↦ ∑ i, coeffs.weights i * fs i x. This weighted formula is the chapter bridge view of the canonical simplex combination supplied by StdSimplex.convexCombination_map_eq_sum, and the textbook ℝⁿ / m-indexed statement is recovered by specializing to E = EuclideanSpace ℝ (Fin n) and ι = Fin m, using the canonical proper-space instance on that finite-dimensional model.
universe u v
open scoped BigOperators
variable {ι : Type v} [Fintype ι] [Nonempty ι]
variable {E : Type u} [NormedAddCommGroup E] [NormedSpace ℝ E] [ProperSpace E]
theorem exists_stdSimplex_minimax_linearization_of_bounded_familyMaximumSublevelSets
{Q : Set E} {fs : ι → E → ℝ}
(hfs : ∀ i, ClosedConvexOn Q (fun x ↦ (fs i x : WithTop ℝ)))
(hbounded :
∀ α : ℝ,
Bornology.IsBounded
(constrainedSublevelSet Q (fun x ↦ ((maxTypeObjective fs x : ℝ) : WithTop ℝ)) α)) :
∃ coeffs : StdSimplex ℝ ι,
sInf (Set.range fun x : Q ↦ (maxTypeObjective fs x : EReal)) =
sInf (Set.range fun x : Q ↦
((∑ i, coeffs.weights i * fs i x : ℝ) : EReal))
Definition 3.1.2.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.2.1: the Fenchel dual of an ℝ ∪ {+∞}-valued function on a real inner-product space, obtained by evaluating fenchelConjugate along the Riesz map innerₗ.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
abbrev fenchelDual (f : E → WithTop ℝ) : E → EReal :=
fenchelConjugate (withTopToEReal ∘ f) ∘ innerₗ ELemma 3.1.2.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.2.1 lies in the chapter's support-function domain. Primary domain: - support functions of subsets of a real inner-product space and their behavior under convex hulls of two-set unions. Sampled owner-style declarations: - supportFunction from Definition_3_9 - supportFunction_apply - supportFunction_convexHull_union_eq_max from Lemma_3_3 - the duplicate local theorem shell in Lemma_3_1_3 Best owner abstraction: - the exact upstream theorem supportFunction_convexHull_union_eq_max from Lemma_3_3, stated at the same ambient owner level as supportFunction Primitive data: - two sets Q₁ Q₂ : Set E - a direction x : E Derived API: - the support function of convexHull ℝ (Q₁ ∪ Q₂) - its identification with the pointwise maximum of the two support functions Source/core/bridge triage: - source-facing: this two-set convex-hull support-function identity - core/canonical: the exact upstream theorem supportFunction_convexHull_union_eq_max - bridge/view: none needed; the target interface already exists upstream This file previously repeated the exact upstream theorem already present in Lemma_3_3 and again in Lemma_3_1_3. Since the project already has the precise owner interface, this numbered item is now a direct recall rather than a third parallel theorem shell. The textbook ℝⁿ statement is a specialization of that generalized owner theorem.
recall supportFunction_convexHull_union_eq_maxProposition 3.1.2.1
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.1.2.1: for a nonempty family of coordinatewise nonnegative finite weights, the weighted pointwise supremum built from the finite sums x ↦ ∑ i, weights i * f i x of closed convex functions is again a closed convex function.
open scoped BigOperators WithTopConvexAnalysis
universe u
variable {X : Type u} {ι : Type*} [Fintype ι]
[TopologicalSpace X] [AddCommMonoid X] [Module ℝ X]
theorem closedConvexFunction_pointwiseSupremumOn_nonneg_weighted
{Δ : Set (ι → ℝ)} {f : ι → X → WithTop ℝ}
(hΔ_nonempty : Δ.Nonempty)
(hΔ_nonneg : ∀ ⦃weights⦄, weights ∈ Δ → ∀ i, 0 ≤ weights i)
(hf : ∀ i, ClosedConvexFunction (f i)) :
ClosedConvexFunction
(pointwiseSupremumOn Δ (fun x weights ↦ ∑ i, (weights i : WithTop ℝ) * f i x))
Remark 3.1.2.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Remark 3.1.2.1: if Q is convex and contains 0 in its interior, then the everywhere-finite real representative withTopRealPart (ψ[Q]) of its Minkowski functional is positively homogeneous of degree 1 and convex on all of the ambient real topological vector space. In the textbook Euclidean bounded closed case, this is exactly the same Minkowski-functional conclusion.
open scoped Pointwise Topology ENNReal NNReal WithTopConvexAnalysis
universe u
variable {E : Type u} [TopologicalSpace E] [AddCommGroup E] [Module ℝ E]
[ContinuousSMul ℝ E]
open scoped MinkowskiFunctional
theorem minkowskiFunctional_posHom_and_convexOn_univ_of_convex_zero_mem_interior
{Q : Set E} (hQ_convex : Convex ℝ Q) (hQ_zero : (0 : E) ∈ interior Q) :
IsPositivelyHomogeneousOn 1 Set.univ (withTopRealPart (ψ[Q])) ∧
ConvexOn ℝ Set.univ (withTopRealPart (ψ[Q]))Theorem 3.1.2.1
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.2.1 is recall-only in the chapter's ClosedConvexOn owner API. Primary domain: - closure properties of closed convex WithTop ℝ-valued functions on a feasible set in a real topological module. Sampled owner-style declarations in this domain: - ClosedConvexOn from Definition_3_1_1_5 - ClosedConvexOn.nonneg_smul from Theorem_3_1_5 - ClosedConvexOn.add_inter from Theorem_3_1_5 - ClosedConvexOn.max_inter from Theorem_3_1_5 Best owner abstraction: - ClosedConvexOn Primitive data: - the owner witnesses hf, hf₁, hf₂ - the nonnegative scalar β Derived API: - ClosedConvexOn.nonneg_smul - ClosedConvexOn.add_inter - ClosedConvexOn.max_inter Source/core/bridge triage: - source-facing: the three closure properties recorded under Theorem 3.1.2.1 - core/canonical: the owner namespace ClosedConvexOn - bridge/view: this later numbered file, which now directly recalls the owner theorems instead of introducing parallel aliases smul_nonneg, add, and max The earlier file Theorem_3_1_5 already owns these exact closure operations with the chapter's canonical names. This file therefore reuses those owner entries directly rather than keeping a second public vocabulary for the same mathematics.
universe u
variable {X : Type u} [TopologicalSpace X] [AddCommMonoid X] [Module ℝ X]
recall ClosedConvexOn.nonneg_smul
{Q : Set X} {f : X → WithTop ℝ} {β : ℝ}
(hf : ClosedConvexOn Q f) (hβ : 0 ≤ β) :
ClosedConvexOn Q ((β : WithTop ℝ) • f)
Definition 3.1.2.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.2.2 is a recall-only item in the chapter's support-function domain. Primary domain: - support functions of subsets of a real inner-product space. Sampled owner-style declarations: - supportFunction from Nesterov.Chap03.Definition_3_9 - supportFunction_apply - supportFunction_convexHull_eq Best owner abstraction: - the chapter source-facing owner declaration supportFunction. Primitive data: - none; this is a recall-only item. Derived API: - the owner declaration supportFunction - the defining bridge supportFunction_apply Source/core/bridge triage: - source-facing: the textbook support function of a set - core/canonical: the owner declaration supportFunction - bridge/view: the defining evaluation lemma supportFunction_apply No exact mathlib owner for this EReal-valued support function was found in the sampled domain, so this file recalls the chapter owner declarations directly instead of keeping a parallel local copy of the same supremum construction. The textbook ℝⁿ statement is the specialization E = EuclideanSpace ℝ (Fin n).
universe u
open scoped SupportFunction
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
recall supportFunctionLemma 3.1.2.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.2.2 is a source-facing recall in the chapter's univariate closed-convex continuity domain. Primary domain: - relative continuity of univariate closed convex WithTop ℝ-valued functions on their effective domain. Sampled owner-style declarations: - ClosedConvexFunction from Definition_3_1_1_5 - ClosedConvexFunction.continuousOn_effectiveDomain_one_dimensional - mathlib ConvexOn.continuousOn - mathlib ConvexOn.continuousOn_interior Best owner abstraction: - the chapter owner ClosedConvexFunction, together with its exact univariate continuity theorem ClosedConvexFunction.continuousOn_effectiveDomain_one_dimensional Primitive data: - the effective domain dom f - the constrained-epigraph data packaged by ClosedConvexFunction f Derived API: - ClosedConvexFunction.continuousOn_effectiveDomain_one_dimensional Source/core/bridge triage: - source-facing: the one-dimensional continuity consequence stated in this lemma - core/canonical: ClosedConvexFunction - bridge/view: dom f, withTopRealPart f, and the ambient ConvexOn continuity lemmas sampled for domain style The previous file kept a second public theorem ClosedConvexFunction.continuousOn_effectiveDomain_univariate with the exact same interface as the upstream owner theorem. That shell carried no new mathematics, so this file now recalls the canonical chapter theorem directly instead of maintaining a parallel local copy.
recall ClosedConvexFunction.continuousOn_effectiveDomain_one_dimensionalProposition 3.1.2.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.1.2.2 recalls the chapter theorem supportFunction_smul.
universe u
variable {E : Type u} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
open scoped SupportFunction
recall supportFunction_smul
(Q : Set E) (hQ : Q.Nonempty) (x : E) (τ : NNReal) :
ξ[Q] (τ • x) = (τ : EReal) * ξ[Q] xRemark 3.1.2.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Remark 3.1.2.2: the support function of the lifted set quadraticSupportLift Q sends (g, γ) to the supremum over y ∈ Q of ⟪g, y⟫ - (γ / 2) ‖y‖², viewed in EReal.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
local notation "Z" => WithLp 2 (E × ℝ)
open WithLp
open scoped ConvexAnalysis
open scoped SupportFunction
theorem supportFunction_quadraticSupportLift_apply (Q : Set E) (g : E) (γ : ℝ) :
ξ[quadraticSupportLiftL2 Q] (toLp 2 (g, γ)) =
sSup
((fun y : E ↦ (((inner ℝ g y) - (γ / 2) * ‖y‖ ^ 2 : ℝ) : EReal)) '' Q)Theorem 3.1.2.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.2.2: if φ is closed and convex on S ⊆ ℝᵐ and g : ℝⁿ → ℝᵐ is affine, then the pullback x ↦ φ (g x) is closed and convex on the affine preimage {x | g x ∈ S}. This is the finite-dimensional specialization of ClosedConvexOn.comp_continuousAffineMap.
variable {m n : ℕ}
local notation "Eₙ" => EuclideanSpace ℝ (Fin n)
local notation "Eₘ" => EuclideanSpace ℝ (Fin m)
namespace ClosedConvexOn
theorem comp_affineMap
{m n : ℕ}
{S : Set (EuclideanSpace ℝ (Fin m))}
{φ : EuclideanSpace ℝ (Fin m) → WithTop ℝ}
(hφ : ClosedConvexOn S φ)
(g : EuclideanSpace ℝ (Fin n) →ᵃ[ℝ] EuclideanSpace ℝ (Fin m)) :
ClosedConvexOn (g ⁻¹' S) (φ ∘ g)Definition 3.1.2.3
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.2.3: a minimax linearization parameter for f₁, f₂ : Q → ℝ is a scalar lam ∈ [0, 1] such that the extended-real infimum of the pointwise maximum x ↦ max (f₁ x) (f₂ x) equals the extended-real infimum of the affine combination x ↦ lam * f₁ x + (1 - lam) * f₂ x.
universe u
variable {Q : Type u}
def IsMinimaxLinearizationParameter
(f₁ f₂ : Q → ℝ) (lam : unitInterval) : Prop :=
sInf (Set.range fun x ↦ ((max (f₁ x) (f₂ x) : ℝ) : EReal)) =
sInf (Set.range fun x ↦ ((AffineMap.lineMap (f₂ x) (f₁ x) (lam : ℝ) : ℝ) : EReal))Proposition 3.1.2.3
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.1.2.3 lies in the chapter's support-function / effective-domain domain. Sampled owner-style declarations: - extendedRealEffectiveDomain in Definition_3_1_1_2 - supportFunction in Definition_3_9 - supportFunction_mem_extendedRealEffectiveDomain_of_nonempty_bounded in Proposition_3_11 - supportFunction_dom_eq_univ_of_nonempty_bounded in Proposition_3_11 Best owner abstraction: - the existing chapter theorem supportFunction_dom_eq_univ_of_nonempty_bounded, together with its pointwise companion supportFunction_mem_extendedRealEffectiveDomain_of_nonempty_bounded Primitive data: - a set Q : Set E in a real inner-product space E - hypotheses Q.Nonempty and Bornology.IsBounded Q Derived API: - pointwise finiteness of supportFunction Q - the domain identity dom (supportFunction Q) = Set.univ Source/core/bridge triage: - source-facing: the bounded-set support-function finiteness statement - core/canonical: supportFunction_dom_eq_univ_of_nonempty_bounded - bridge/view: the pointwise companion theorem This file previously duplicated the exact pointwise companion theorem already formalized in Proposition_3_11 and also kept a renamed shell for the global domain theorem. Since the chapter already has the owner theorem family, and that family has now been generalized to arbitrary real inner-product spaces, this numbered item is recall-only and keeps no parallel local theorem names. The textbook ℝⁿ statement remains an immediate specialization.
recall supportFunction_mem_extendedRealEffectiveDomain_of_nonempty_boundedRemark 3.1.2.3
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Remark 3.1.2.3: if f is convex on E, then its perspective transform (τ, x) ↦ τ f (τ⁻¹ • x) is positively homogeneous on the cone τ > 0 together with the origin, and it is convex on that same domain. For the source statement, take E = ℝⁿ.
universe u
variable {E : Type u} [AddCommMonoid E] [Module ℝ E]
local notation "P" => ℝ × E
theorem perspectiveTransform_posHomogeneous_and_convexOn
{f : E → ℝ} (hf : ConvexOn ℝ Set.univ f) :
IsPositivelyHomogeneousOn 1 (perspectiveCone E) (perspectiveTransform f) ∧
ConvexOn ℝ (perspectiveCone E) (perspectiveTransform f)Theorem 3.1.2.3
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.2.3: if Q ⊆ X × Y is convex and φ : X × Y → EReal is convex on its effective domain with Q ⊆ dom φ, and every feasible fiber of the constrained partial infimum has a finite infimum, then the constrained partial infimum is convex in the chapter's EReal sense: its finite real part is convex on its finite-value domain. The hypothesis Q ⊆ dom φ restores the source-side input surface φ : ℝⁿ × ℝᵐ → ℝ ∪ {+∞}, while the extra finite-fiber hypothesis is the Lean rendering of the source conclusion f : \hat Q → ℝ ∪ {+∞}, which rules out -∞ values on feasible fibers.
universe u v
open scoped ConvexAnalysis
variable {X : Type u} {Y : Type v}
variable [AddCommMonoid X] [Module ℝ X]
variable [AddCommMonoid Y] [Module ℝ Y]
theorem partialInfProjection_convexOn
{Q : Set (X × Y)} {φ : X × Y → EReal}
(hQ : Convex ℝ Q)
(hφ : ConvexOn ℝ (dom φ) (extendedRealRealPart φ))
(hQdom : Q ⊆ dom φ)
(hfinite :
∀ ⦃x : X⦄, (∃ y : Y, (x, y) ∈ Q) →
x ∈ dom (partialInfProjection Q φ)) :
ConvexOn ℝ (dom (partialInfProjection Q φ))
(extendedRealRealPart (partialInfProjection Q φ))Theorem 3.1.2.4
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.2.4: for a nonempty index type, if each slice x ↦ φ x y is closed and convex on Q, then the pointwise supremum over all indices is closed and convex on its canonical effective domain.
universe u v
open scoped WithTopConvexAnalysis
variable {ι : Type u} {X : Type v}
variable [TopologicalSpace X] [AddCommMonoid X] [Module ℝ X] [Nonempty ι]
variable {Q : Set X} {φ : X → ι → WithTop ℝ}
#check
(show (∀ y : ι, ClosedConvexOn Q (fun x ↦ φ x y)) →
ClosedConvexOn (pointwiseSupremumOnEffectiveDomain Q (Set.univ : Set ι) φ)
(pointwiseSupremumOn (Set.univ : Set ι) φ) from
fun hφ ↦ by
let ⟨i⟩ := ‹Nonempty ι›
exact ClosedConvexOn.pointwise_sSup ⟨i, by simp⟩ (fun y _ ↦ hφ y))Theorem 3.1.2.5
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.2.5, at the generic owner level: if ψ is convex on domψ, φ is convex and nondecreasing on a convex set I containing ψ(domψ), then φ ∘ ψ is convex on domψ. The textbook real-interval statement is the specialization 𝕜 = ℝ, Y = Z = ℝ.
universe u v w x
open Set
namespace ConvexOn
variable {𝕜 : Type u} [Semiring 𝕜] [PartialOrder 𝕜]
variable {X : Type v} [AddCommMonoid X] [SMul 𝕜 X]
variable {Y : Type w} [AddCommMonoid Y] [PartialOrder Y] [SMul 𝕜 Y]
variable {Z : Type x} [AddCommMonoid Z] [PartialOrder Z] [SMul 𝕜 Z]
variable {domψ : Set X} {I : Set Y} {ψ : X → Y} {φ : Y → Z}
theorem comp_of_monotoneOn
(hφ : ConvexOn 𝕜 I φ) (hψ : ConvexOn 𝕜 domψ ψ)
(hφ_mono : MonotoneOn φ I) (hψ_maps : MapsTo ψ domψ I) :
ConvexOn 𝕜 domψ (φ ∘ ψ)Theorem 3.1.2.6
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.2.6: if f₁ and f₂ are real-valued functions whose WithTop lifts are closed and convex on a feasible set Q, and every constrained sublevel set of the pointwise maximum x ↦ max (f₁ x) (f₂ x) on Q is bounded, then there exists some λ* ∈ [0, 1] for which the minimum value of x ↦ max (f₁ x) (f₂ x) on Q equals the minimum value of the convex combination x ↦ λ* f₁ x + (1 - λ*) f₂ x; in Lean this minimum equality is recorded by the owner predicate IsMinimaxLinearizationParameter on the subtype Q. The textbook finite-dimensional ℝⁿ statement is recovered by equipping E = EuclideanSpace ℝ (Fin n) with the canonical ProperSpace instance.
universe u
open scoped WithTopConvexAnalysis
variable {E : Type u} [NormedAddCommGroup E] [NormedSpace ℝ E] [ProperSpace E]
theorem exists_minimax_parameter_of_bounded_constrainedSublevelSets
{Q : Set E} {f₁ f₂ : E → ℝ}
(hf₁ : ClosedConvexOn Q (fun x ↦ (f₁ x : WithTop ℝ)))
(hf₂ : ClosedConvexOn Q (fun x ↦ (f₂ x : WithTop ℝ)))
(hbounded :
∀ β : ℝ,
Bornology.IsBounded
(constrainedSublevelSet Q
(fun x ↦ ((max (f₁ x) (f₂ x) : ℝ) : WithTop ℝ)) β)) :
∃ lam : unitInterval,
IsMinimaxLinearizationParameter (fun x : Q ↦ f₁ x) (fun x : Q ↦ f₂ x) lamCorollary 3.1.3
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Corollary 3.1.3 is a recall-only item in the chapter's convex-analysis/minimax-linearization domain. Primary domain: - finite pointwise maxima of closed convex functions and bounded constrained sublevel sets. Sampled owner-style declarations: - ClosedConvexOn - maxTypeObjective - constrainedSublevelSet - exists_stdSimplex_minimax_linearization_of_bounded_familyMaximumSublevelSets Best owner abstraction: - source-facing: the bounded-level-set minimax-linearization corollary - core/canonical: maxTypeObjective fs, the bounded feasible sublevel owner constrainedSublevelSet Q (fun x ↦ ((maxTypeObjective fs x : ℝ) : WithTop ℝ)) α, and the simplex coefficient owner StdSimplex ℝ ι - bridge/view: the weighted-sum expression ∑ i, coeffs.weights i * fs i x, derived from the canonical simplex data Primitive data: - a nonempty finite family fs : ι → E → ℝ - closed convexity of each component on Q - boundedness of the constrained sublevel sets of maxTypeObjective fs Derived API: - the simplex coefficient vector coeffs : StdSimplex ℝ ι - the equality of constrained EReal infima between the finite maximum and the weighted sum The upstream max-sublevel-set theorem already has the exact interface needed here. This file therefore imports that owner theorem directly and recalls its canonical name rather than routing through a later recall surface or keeping a duplicate local wrapper. The textbook Fin m specialization is recovered by instantiating the recalled theorem at ι = Fin m; it does not need a separate local declaration here.
recall exists_stdSimplex_minimax_linearization_of_bounded_familyMaximumSublevelSets
Definition 3.1.3
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.3 belongs to the chapter's canonical one-sided directional-derivative API in Nesterov.Chap03.Definition_3_1_3_1. Primary domain: - one-sided directional derivatives of extended-real-valued functions on real modules. Relevant owner-style declarations sampled before refinement: - extendedRealEffectiveDomain - HasDerivWithinAt - HasDirectionalDerivAt - DirectionallyDifferentiableAt Owner abstraction: - HasDirectionalDerivAt Primitive data: - extendedRealEffectiveDomain f, the finite-value condition at the base point and along the ray. - the scalar slice fun α ↦ extendedRealRealPart f (x + α • p), organized canonically by the one-sided derivative owner HasDerivWithinAt ... (Set.Ici 0) 0. Derived API: - HasDirectionalDerivAt is the owner predicate for the one-sided directional derivative in every direction, including p = 0 via the constant ray. - DirectionallyDifferentiableAt is the existential wrapper around HasDirectionalDerivAt. - directionallyDifferentiableAt_iff_exists_hasDirectionalDerivAt is the companion specification theorem. Source/core/bridge triage: - source-facing: HasDirectionalDerivAt, DirectionallyDifferentiableAt - core/canonical: HasDerivWithinAt on the scalar slice fun α ↦ extendedRealRealPart f (x + α • p) - bridge/view: the secant-slope presentation of that scalar slice The support vocabulary is already owned upstream, so this file recalls only the source-facing owner surface and its thin specification theorem. As in the owner file, the source prose is Euclidean, but the declaration surface itself needs only the real-module structure used to form the ray x + α • p.
recall HasDirectionalDerivAtLemma 3.1.3
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.3 lies in the chapter's support-function domain. Primary domain: - support functions of subsets of a real inner-product space and their behavior under convex hulls of two-set unions. Sampled owner-style declarations: - supportFunction from Definition_3_9 - supportFunction_apply - supportFunction_convexHull_union_eq_max from Lemma_3_3 - the recall-only downstream use in Lemma_3_1_2_1 Best owner abstraction: - the exact upstream theorem supportFunction_convexHull_union_eq_max from Lemma_3_3, stated at the same ambient owner level as supportFunction Primitive data: - two sets Q₁ Q₂ : Set E - a direction x : E Derived API: - the support function of convexHull ℝ (Q₁ ∪ Q₂) - its identification with the pointwise maximum of the two support functions Source/core/bridge triage: - source-facing: this two-set convex-hull support-function identity - core/canonical: the exact upstream theorem supportFunction_convexHull_union_eq_max - bridge/view: none needed; the target interface already exists upstream This file previously repeated the exact theorem already owned by Lemma_3_3. Since the chapter already has the precise owner interface, this numbered item is refined to direct canonical recall/use instead of keeping a parallel local theorem shell. The textbook ℝⁿ statement is a specialization of that generalized owner theorem.
recall supportFunction_convexHull_union_eq_maxTheorem 3.1.3
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.3 is a recall-only item in the chapter's extended-real convex-analysis domain. Primary domain: - convex sublevel sets for EReal-valued convex functions on an ℝ-module. Relevant owner-style declarations sampled before refinement: - mathlib ConvexOn.convex_le - project mem_levelSet_iff - chapter extendedRealRealPart Best owner abstraction: - the owner theorem ConvexOn.convex_le Primitive data: - dom f - extendedRealRealPart f - the owner sublevel set {x ∈ dom f | extendedRealRealPart f x ≤ β} Derived API: - the convexity conclusion for that owner sublevel set The previous file recalled a chapter-local wrapper around the owner theorem. Since Theorem_3_1_1_3 now keeps only direct owner reuse and no parallel sublevel-set alias, this file checks the specialized owner theorem itself at the canonical owner surface.
universe u
open scoped ConvexAnalysis
variable {X : Type u} [AddCommMonoid X] [Module ℝ X]
variable (f : X → EReal)
#check
(show ConvexOn ℝ (dom f) (extendedRealRealPart f) →
∀ β : ℝ, Convex ℝ {x | x ∈ dom f ∧ extendedRealRealPart f x ≤ β} from
ConvexOn.convex_le)Definition 3.1.3.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.3.1: an extended-real-valued function has directional derivative d at x in direction p when x has finite value, the nearby values along the ray x + α • p are finite for α ↓ 0, and the directional slice α ↦ extendedRealRealPart f (x + α • p) has right derivative d at 0. This includes the zero direction, whose ray is constant and whose derivative is therefore 0.
universe u
open Filter Set
open scoped ConvexAnalysis Topology
variable {E : Type u} [AddCommMonoid E] [Module ℝ E]
def HasDirectionalDerivAt (f : E → EReal) (x p : E) (d : ℝ) : Prop :=
x ∈ dom f ∧
(∀ᶠ α : ℝ in 𝓝[>] (0 : ℝ), x + α • p ∈ dom f) ∧
HasDerivWithinAt (fun α ↦ extendedRealRealPart f (x + α • p)) d (Ici (0 : ℝ)) 0Lemma 3.1.3.1
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.3.1 (1): for a convex ℝ ∪ {+∞}-valued function and an interior point x of its effective domain, the finite directional derivative given by the explicit toReal view of convexDirectionalDerivative is convex on all directions.
open scoped ConvexAnalysis Topology WithTopConvexAnalysis
universe u
variable {E : Type u} [TopologicalSpace E] [AddCommGroup E] [Module ℝ E]
[IsTopologicalAddGroup E] [ContinuousSMul ℝ E]
variable {f : E → WithTop ℝ}
recall convexDirectionalDerivativeReal_convexOn_univ_of_mem_interior
Theorem 3.1.3.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.3.1: if a convex ℝ ∪ {+∞}-valued function has x₀ in the interior of its effective domain, then some ℓ₁-ball about x₀ stays inside the effective domain and the finite-value part of the function satisfies a local ℓ₁-Lipschitz estimate there. The bounded image consequence is recorded separately downstream.
open scoped Topology
open scoped WithTopConvexAnalysis
open EuclideanSpace
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
namespace ConvexOn
theorem exists_l1_ball_subset_effectiveDomain_and_abs_sub_le_of_mem_interior
{f : E → WithTop ℝ}
(hf : ConvexOn ℝ (dom f) (withTopRealPart f))
{x0 : E} (hx0 : x0 ∈ interior (dom f)) :
∃ ε > 0, ∃ L > 0,
(l1Seminorm n).ball x0 ε ⊆ dom f ∧
∀ ⦃y : E⦄, y ∈ (l1Seminorm n).ball x0 ε →Theorem 3.1.3.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.3.2: a convex ℝ ∪ {+∞}-valued function is directionally differentiable in every direction at every interior point of its effective domain.
universe u
open scoped ConvexAnalysis Topology WithTopConvexAnalysis
variable {E : Type u} [TopologicalSpace E] [AddCommGroup E] [Module ℝ E]
[IsTopologicalAddGroup E] [ContinuousSMul ℝ E]
namespace ConvexOn
theorem directionallyDifferentiableAt_of_mem_interior_dom
{f : E → WithTop ℝ}
(hf : ConvexOn ℝ (dom f) (withTopRealPart f))
{x p : E} (hx : x ∈ interior (dom f)) :
DirectionallyDifferentiableAt (withTopToEReal ∘ f) x pCorollary 3.1.4
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Corollary 3.1.4: if Q is a closed convex set in a nontrivial real inner-product space and x ∉ Q, then there exist a nonzero normal vector g and an offset γ such that Q lies in the lower closed half-space ⟪g, y⟫ ≤ γ and x lies strictly above it: γ < ⟪g, x⟫. The lower-side half-space inclusion is packaged by SeparatesPointFromWith Q x g γ. The textbook ℝⁿ statement is the specialization E = EuclideanSpace ℝ (Fin n) with n > 0.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E] [CompleteSpace E]
[Nontrivial E]
theorem exists_strictlySeparating_hyperplane_of_nonmem_closed_convex
(Q : Set E) (hQ_closed : IsClosed Q) (hQ_convex : Convex ℝ Q)
{x : E} (hx : x ∉ Q) :
∃ g : E, ∃ γ : ℝ, SeparatesPointFromWith Q x g γ ∧ γ < inner ℝ g xDefinition 3.1.4
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.4 is a source-facing recall in the chapter's affine-nesterovHyperplane domain. Primary domain: - affine hyperplanes, supporting hyperplanes, and separation of a point from a set in a real inner-product space. Specializing E = EuclideanSpace ℝ (Fin n) recovers the textbook ℝⁿ setting. Sampled owner-style declarations: - AffineHyperplane in Definition_3_1_4_1, the owner of a nonzero normal vector and an offset; - AffineHyperplane.IsSupporting, the owner-level support predicate; - nesterovHyperplane, the coordinate carrier view of an affine nesterovHyperplane; - StrictlySeparatesPointFromWith, the coordinate strict point-versus-set bridge; - mathlib Set.IsExposed, a nearby exposed-face owner that is stronger than this source-facing point-versus-set nesterovHyperplane recall. Best owner abstraction: - AffineHyperplane Source/core/bridge triage: - source-facing: the textbook nesterovHyperplane, supporting-nesterovHyperplane, and point-versus-set separation notions; - core/canonical: AffineHyperplane, whose primitive data are a nonzero normal vector and an offset; - bridge/view: the coordinate carrier nesterovHyperplane g γ, together with IsSupportingHyperplane, SeparatesPointFromWith, and StrictlySeparatesPointFromWith. Primitive data: - the nonzero normal vector and offset packaged by AffineHyperplane. Derived API: - the owner-level support and point-separation predicates on AffineHyperplane; - the coordinate bridge declarations nesterovHyperplane, IsSupportingHyperplane, SeparatesPointFromWith, and StrictlySeparatesPointFromWith. The later two-set separation layer already has its own numbered recall in Definition_3_12.lean. This file therefore recalls only the source-facing owner and bridge declarations for Definition 3.1.4 instead of keeping a broader inventory that duplicated later chapter material.
recall AffineHyperplaneLemma 3.1.4
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.4 is a source-facing recall in the chapter's univariate closed-convex continuity domain. Primary domain: - relative continuity of univariate closed convex WithTop ℝ-valued functions on their effective domain. Sampled owner-style declarations: - dom f, withTopRealPart from Definition_3_3 - ClosedConvexFunction from Definition_3_1_1_5 - ClosedConvexFunction.continuousOn_effectiveDomain_one_dimensional from Lemma_3_4 - mathlib continuousWithinAt_iff_continuousAt_restrict Best owner abstraction: - ClosedConvexFunction f, with dom f and withTopRealPart f as the canonical derived domain/view data. Primitive data: - the effective domain dom f - the finite-value representative withTopRealPart f - the owner hypothesis ClosedConvexFunction f Derived API: - ClosedConvexFunction.continuousOn_effectiveDomain_one_dimensional Source/core/bridge triage: - source-facing: continuity of the finite-value representative on the effective domain - core/canonical: ClosedConvexFunction - bridge/view: the restriction-based continuity formalization on the effective-domain subtype This file therefore recalls the upstream owner theorem directly instead of maintaining a second proof of the same continuity statement.
open scoped WithTopConvexAnalysis
recall ClosedConvexFunction.continuousOn_effectiveDomain_one_dimensional
{f : ℝ → WithTop ℝ} (hf : ClosedConvexFunction f) :
ContinuousOn (withTopRealPart f) (dom f)Theorem 3.1.4
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.4 is a source-facing recall in the chapter's closed-convex-function domain. Primary domain: - closed convex WithTop ℝ-valued functions and their real sublevel sets. Sampled owner-style declarations: - ClosedConvexFunction - ClosedConvexFunction.isClosed_convex_sublevelSet Source/core/bridge triage: - source-facing: the sublevel-set consequence recorded as Theorem 3.1.4 - core/canonical: ClosedConvexFunction - bridge/view: the epigraph-based owner API behind ClosedConvexFunction Primitive data: - the owner hypothesis ClosedConvexFunction f Derived API: - the recalled owner theorem ClosedConvexFunction.isClosed_convex_sublevelSet This file therefore reuses the canonical owner theorem directly, without rebuilding the broader five-result package that previously overreached the source theorem.
recall ClosedConvexFunction.isClosed_convex_sublevelSetCorollary 3.1.4.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Corollary 3.1.4.1 is the direct chapter reuse of the canonical owner theorem for separating a closed convex set from an exterior point. The coordinate (g, γ) form is recovered separately by exists_strictlySeparating_hyperplane_of_nonmem_closed_convex.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
recall areStronglySeparable_singleton_of_nonmem_closed_convexDefinition 3.1.4.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.4.1: an affine nesterovHyperplane in a real inner-product space is determined by a nonzero normal vector and an offset. Specializing E to EuclideanSpace ℝ (Fin n) recovers the textbook ℝⁿ presentation. The carrier set and the support/separation predicates below are derived from this owner object.
universe u
variable {E : Type u} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
structure AffineHyperplane (E : Type u) [SeminormedAddCommGroup E] [InnerProductSpace ℝ E] where
normal : E
normal_ne_zero : normal ≠ 0
offset : ℝTheorem 3.1.4.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.4.1 is the direct chapter recall of the existing strong-separation theorem with the same interface.
recall areStronglySeparable_of_disjoint_closed_convex_of_one_boundedCorollary 3.1.4.2
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Corollary 3.1.4.2 belongs to the chapter's support-function comparison domain. Primary domain: - support functions of subsets of a real Hilbert space and the extended-real effective domain. Sampled owner declarations: - extendedRealEffectiveDomain / dom from Definition_3_1_1_2 - supportFunction from Definition_3_9 - subset_of_supportFunction_le_on_domain from Theorem_3_17 - supportFunction_eq_on_common_domain_implies_eq from Theorem_3_17 Source-facing layer: - equality of closed convex sets from agreement of their support functions on the common finite-value domain, with empty/nonempty status already encoded by the shared effective domain. Core/canonical layer: - the owner constructions extendedRealEffectiveDomain and supportFunction, together with the exact comparison theorem supportFunction_eq_on_common_domain_implies_eq. Bridge/view: - the companion inclusion theorem subset_of_supportFunction_le_on_domain. Primitive data: - the sets Q₁, Q₂ and the canonical support-function/effective-domain constructions. Derived API: - the inclusion and equality comparison theorems for those owner constructions. This file therefore reuses the exact owner theorem from Theorem_3_17 instead of keeping a parallel local copy of supportFunction, extendedRealEffectiveDomain, and the same comparison statement under a second name.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E] [CompleteSpace E]
open scoped ConvexAnalysis SupportFunction
recall supportFunction_eq_on_common_domain_implies_eq
(Q₁ Q₂ : Set E) (hQ₁_closed : IsClosed Q₁) (hQ₂_closed : IsClosed Q₂)
(hQ₁_convex : Convex ℝ Q₁) (hQ₂_convex : Convex ℝ Q₂)
(hdom : dom ξ[Q₁] = dom ξ[Q₂])
(hξ : Set.EqOn ξ[Q₁] ξ[Q₂] (dom ξ[Q₁])) :
Q₁ = Q₂
Theorem 3.1.4.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.4.2 on the owner surface: if Q is a closed convex set in a finite-dimensional real inner-product space and x₀ is a boundary point of Q, then there exists an affine nesterovHyperplane H such that x₀ ∈ H and H supports Q. The textbook ℝⁿ statement is the specialization E = EuclideanSpace ℝ (Fin n).
universe u
open Filter
open scoped Topology
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
theorem exists_supporting_affineHyperplane_at_boundary_point_of_closed_convex
[FiniteDimensional ℝ E] (Q : Set E) (hQ_closed : IsClosed Q) (hQ_convex : Convex ℝ Q)
{x₀ : E} (hx₀ : x₀ ∈ frontier Q) :
∃ H : AffineHyperplane E, x₀ ∈ H ∧ H.IsSupporting QCorollary 3.1.5
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Corollary 3.1.5 (1) recalls the canonical support-function comparison theorem from Theorem_3_17.
universe u
open scoped ConvexAnalysis SupportFunction
recall subset_of_supportFunction_le_on_domain
{E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E] [CompleteSpace E]
(Q₁ Q₂ : Set E) (hQ₂_nonempty : Q₂.Nonempty)
(hQ₂_closed : IsClosed Q₂) (hQ₂_convex : Convex ℝ Q₂)
(hξ : ∀ g ∈ dom ξ[Q₂], ξ[Q₁] g ≤ ξ[Q₂] g) :
Q₁ ⊆ Q₂Definition 3.1.5
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.5, generalized from the textbook Euclidean setting: a vector g is a subgradient of an ℝ ∪ {+∞}-valued function f at x0 if x0 lies in the effective domain of f and the affine function y ↦ f x0 + ⟪g, y - x0⟫ supports f from below on that domain.
open scoped WithTopConvexAnalysis
universe u v
variable {V : Type u} [SeminormedAddCommGroup V] [InnerProductSpace ℝ V]
def IsSubgradientAt (f : V → WithTop ℝ) (x0 g : V) : Prop :=
x0 ∈ dom f ∧
∀ ⦃y : V⦄, y ∈ dom f →
f y ≥ f x0 + (inner ℝ g (y - x0) : WithTop ℝ)Lemma 3.1.5
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.5 is now a recall-only surface in the chapter's finite directional-derivative domain. Sampled owner-style declarations: - convexDirectionalDerivative in Theorem_3_21, the owner extended-valued directional derivative; - convexDirectionalDerivativeReal_convexOn_univ_of_mem_interior in Theorem_3_21, the owner convexity theorem for the theorem-level finite directional-derivative view; - convexDirectionalDerivativeReal_posHomOn_univ_of_mem_interior, whose canonical pointwise scaling surface is the owner projection map_smul, and convexDirectionalDerivativeReal_affine_support_of_mem_interior in Lemma_3_1_3_1, the source-facing real directional-derivative consequences. Best owner abstraction: - convexDirectionalDerivative, together with its theorem-level finite directional-derivative view. Primitive data: - none in this file; all directional-derivative data is already owned upstream. Derived API: - the recalled real-valued consequences used at this point in the chapter. Source/core/bridge triage: - source-facing: the real-valued convexity, positive-homogeneity, and affine-support statements; - core/canonical: convexDirectionalDerivative; - bridge/view: this recall surface. The previous version duplicated the same real-valued theorem surface with local proofs. This file now recalls the canonical chapter declarations directly instead of keeping a second parallel copy.
recall convexDirectionalDerivativeReal_convexOn_univ_of_mem_interior
Theorem 3.1.5
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.5 (2): the sum of two closed convex functions is closed and convex on the intersection of their feasible sets.
universe u
open scoped WithTopConvexAnalysis
variable {X : Type u} [TopologicalSpace X] [AddCommMonoid X] [Module ℝ X]
namespace ClosedConvexOn
-- Semantic recall: mathlib provides `ConvexOn.smul`, `ConvexOn.add`, and `ConvexOn.sup` for the
-- convexity half of these closure rules, while the chapter owner `ClosedConvexOn` packages the
-- feasible-domain and constrained-epigraph closedness data.
theorem add_inter
{Q₁ Q₂ : Set X} {f₁ f₂ : X → WithTop ℝ}
(hf₁ : ClosedConvexOn Q₁ f₁) (hf₂ : ClosedConvexOn Q₂ f₂) :
ClosedConvexOn (Q₁ ∩ Q₂) (f₁ + f₂)Corollary 3.1.5.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Corollary 3.1.5.1 lies in the chapter's extended-valued convex-analysis / subgradient domain. Primary domain: - subgradients and subdifferentials for ℝ ∪ {+∞}-valued functions on real inner-product spaces. Sampled owner-style declarations: - IsSubgradientAt - subdifferential - subgradient_nonneg_on_sublevelSet_of_mem_subdifferential - subgradient_inner_sub_nonneg_of_isMinOn Best owner abstraction: - the subdifferential owner API together with the minimizer-pairing theorem subgradient_inner_sub_nonneg_of_isMinOn Primitive data: - a feasible set Q, an extended-real objective f, points x0, xStar, g - the feasibility hypothesis x0 ∈ Q - the minimizing hypothesis IsMinOn f Q xStar Derived API: - the subgradient-membership view g ∈ subdifferential f x0 - the pairing inequality 0 ≤ inner ℝ g (x0 - xStar) Source/core/bridge triage: - source-facing: the corollary that a subgradient at a feasible point has nonnegative pairing with the displacement to a minimizer - core/canonical: subdifferential and the owner theorem subgradient_nonneg_on_sublevelSet_of_mem_subdifferential - bridge/view: the chapter theorem subgradient_inner_sub_nonneg_of_isMinOn The previous file introduced a second public theorem whose only extra step was rewriting the primitive predicate IsSubgradientAt as membership in the owner set subdifferential via mem_subdifferential_iff, and it had no downstream users. This file therefore recalls the canonical chapter theorem directly instead of keeping a parallel wrapper around the owner abstraction.
recall subgradient_inner_sub_nonneg_of_isMinOnDefinition 3.1.5.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.5.1: the constrained subdifferential of an extended-real-valued function f at a point x₀ relative to a set Q is the canonical set constrainedSubdifferential Q f x₀, written ∂[Q] f(x₀), consisting of all vectors whose affine support inequality holds for every y ∈ Q, with x₀ itself lying in Q ∩ dom f.
open scoped WithTopConvexAnalysis
universe u
recall constrainedSubdifferential
{V : Type u} [SeminormedAddCommGroup V] [InnerProductSpace ℝ V]
(Q : Set V) (f : V → WithTop ℝ) (x0 : V) : Set VLemma 3.1.5.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.5.1: if every point of Q has a nonempty constrained subdifferential, then f is finite on Q. The convexity conclusion, which does require convexity of Q, and the companion lower-semicontinuity conclusion are already the exact owner theorems recalled just below.
open scoped WithTopConvexAnalysis
universe u
variable {V : Type u} [SeminormedAddCommGroup V] [InnerProductSpace ℝ V]
variable {Q : Set V} {f : V → WithTop ℝ}
theorem subset_withTopEffectiveDomain_of_constrainedSubdifferential_nonempty
(hsubgrad : ∀ x ∈ Q, (∂[Q] f(x)).Nonempty) :
Q ⊆ dom fProposition 3.1.5.1
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Proposition 3.1.5.1: for the real positive-part function x ↦ max x 0, the set of real numbers g satisfying the global affine lower-support inequality max x 0 ≥ max 0 0 + g * (x - 0) for every x is exactly the interval [0, 1].
recall real_posPart_subdifferential_at_zero_eq_Icc
Theorem 3.1.5.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.5.1 is a recall-only surface in the chapter's extended-valued convex- subdifferential domain. Primary domain: - convex analysis of ℝ ∪ {+∞}-valued functions on Euclidean space. Sampled owner-style declarations: - dom and withTopRealPart in Definition_3_3, the chapter owners for the effective domain and finite-value representative; - subdifferential in Definition_3_1_5, the chapter owner for extended-valued subgradient sets; - subdifferential_nonempty_and_isBounded_of_convexOn_effectiveDomain_of_mem_interior in Theorem_3_1_15, the canonical chapter theorem on this owner surface. Best owner abstraction: - subdifferential_nonempty_and_isBounded_of_convexOn_effectiveDomain_of_mem_interior. Primitive data: - none in this file; the theorem and its owner-level hypotheses already live upstream. Derived API: - this recall-only source-facing entry point. Source/core/bridge triage: - source-facing: Theorem 3.1.5.1's nonempty-and-bounded subdifferential statement; - core/canonical: the upstream theorem in Theorem_3_1_15 on the chapter owners dom, withTopRealPart, and subdifferential; - bridge/view: this recall surface. The previous file kept a second theorem stub with the same statement as the canonical upstream theorem. This file now reuses that theorem directly instead of maintaining a parallel copy.
recall subdifferential_nonempty_and_isBounded_of_convexOn_effectiveDomain_of_mem_interiorDefinition 3.1.5.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.5.2 is the textbook positive-part map x ↦ (x)_+. In mathlib this notion is owned by the notation class PosPart, whose canonical map is posPart, written x⁺. Primary domain: - positive-part operations in ordered additive algebra. Relevant owner-style declarations sampled before refinement: - posPart - posPart_def - posPart_eq_ite Best owner abstraction: - posPart Primitive data: - none in this file; the positive-part operation is supplied upstream by mathlib instances. Derived API: - posPart_def - posPart_eq_ite Source/core/bridge triage: - source-facing: the textbook positive-part map x ↦ (x)_+ - core/canonical: posPart - bridge/view: posPart_def and posPart_eq_ite, giving the lattice and linear-order formulas This file therefore recalls the owner declaration directly together with its canonical lattice and linear-order formulas. Downstream specializations should use these upstream bridge lemmas directly instead of introducing parallel public shell theorems.
recall posPart {α : Type*} [PosPart α] : α → αTheorem 3.1.5.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.5.2 (1): for every x ∈ dom f, the Fenchel bidual is bounded above by the original value f x.
open scoped ConvexAnalysis WithTopConvexAnalysis
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable {f : E → WithTop ℝ}
variable {x : E}
-- Proof sketch: use `fenchelBidual_apply_eq_sSup_image_dom_of_dom_nonempty ⟨x, hx⟩` to rewrite
-- `(f⋆⋆) x` as the supremum over `dom (f⋆)` of the affine terms `⟪s, x⟫ - f⋆(s)`, then expand
-- `f⋆(s)` as a supremum over `dom f` and test the defining supremum at `y = x`.
theorem fenchelBidual_le_of_mem_dom
(hx : x ∈ dom f) :
(f⋆⋆) x ≤ withTopToEReal (f x)Definition 3.1.5.3
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.5.3: membership of -g in the normal cone to the sublevel set {x ∈ dom f | f x ≤ f x₀} at x₀ is equivalent to the textbook inequality ⟪g, x₀ - x⟫ ≥ 0 for every x with f x ≤ f x₀.
open scoped ConvexAnalysis
open scoped NormalCone
universe u
variable {V : Type u} [NormedAddCommGroup V] [InnerProductSpace ℝ V] [CompleteSpace V]
theorem level_set_inequality_at_iff
{f : V → EReal} {x0 g : V} :
(-g) ∈ N[{x | x ∈ dom f ∧ f x ≤ f x0}] x0 ↔
∀ ⦃x : V⦄, x ∈ dom f →
f x ≤ f x0 →
inner ℝ g (x0 - x) ≥ 0Theorem 3.1.5.3
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.5.3: at an interior point of the effective domain of a convex ℝ ∪ {+∞}-valued function, the finite directional derivative in direction p is the maximum of the pairings ⟪g, p⟫ over all subgradients g ∈ ∂f(x₀).
recall convexDirectionalDerivativeReal_isGreatest_subgradientPairing_of_mem_interiorDefinition 3.1.5.4
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.5.4: the common regular subdifferential of f on X is the intersection of the pointwise subdifferentials ∂f(x) over all x ∈ X.
universe u
open scoped WithTopConvexAnalysis
variable {V : Type u} [SeminormedAddCommGroup V] [InnerProductSpace ℝ V]
def commonRegularSubdifferential (f : V → WithTop ℝ) (X : Set V) : Set V :=
⋂ x ∈ X, ∂ f(x)Theorem 3.1.5.4
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.5.4 (1): if f is convex on its effective domain and x₀ lies in the interior of that domain, then the directional derivative p ↦ f′(x₀; p) is a finite convex function on all directions. The finiteness is built into the recalled owner surface f′[hx0] : E → ℝ.
open scoped WithTopConvexAnalysis
universe u
variable {E : Type u} [NormedAddCommGroup E] [NormedSpace ℝ E]
recall convexDirectionalDerivativeReal_convexOn_univ_of_mem_interiorDefinition 3.1.5.5
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.5.5: for a set-valued map S and a set X, the common-value set \hat{\mathcal S}(X) is the bounded intersection ⋂ x ∈ X, S x.
universe u v
open Set
variable {α : Type u} {β : Type v}
def commonValueSet (S : α → Set β) (X : Set α) : Set β :=
⋂ x ∈ X, S xTheorem 3.1.5.5
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.5.5: every subgradient g ∈ ∂f(x₀) is a supporting vector to the sublevel set {x | f x ≤ f x₀}, equivalently ⟪g, x - x₀⟫ ≤ 0 for every point of that set.
recall subgradient_nonpos_on_sublevelSet_of_mem_subdifferentialTheorem 3.1.5.6
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.5.6: if xStar minimizes f on Q and x0 ∈ Q, then every subgradient g ∈ ∂f(x0) has nonnegative pairing with the displacement x0 - xStar.
open scoped WithTopConvexAnalysis
universe u
variable {V : Type u} [SeminormedAddCommGroup V] [InnerProductSpace ℝ V]
theorem subgradient_inner_sub_nonneg_of_isMinOn
{Q : Set V} {f : V → WithTop ℝ} {x0 xStar g : V}
(hx0 : x0 ∈ Q) (hxStar : IsMinOn f Q xStar) (hg : g ∈ ∂ f(x0)) :
0 ≤ inner ℝ g (x0 - xStar)Theorem 3.1.5.7
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.5.7: every common subgradient g ∈ ∂̂ f(X) gives the affine increment formula f x₁ = f x₀ + ⟪g, x₁ - x₀⟫ on X.
recall eq_add_inner_of_mem_commonRegularSubdifferentialTheorem 3.1.5.8
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.5.8: if X^* := argmin[dom f] f, then XStar ⊆ X^* if and only if the zero vector lies in the common regular subdifferential ∂̂ f(XStar). The source-side assumptions that XStar is nonempty, closed, convex, and contained in dom f are redundant for this equivalence, so the canonical recall surface states the exact owner theorem without them.
open scoped ConstrainedArgmin WithTopConvexAnalysis
universe u
variable {V : Type u}
variable [SeminormedAddCommGroup V] [InnerProductSpace ℝ V]
recall subset_constrainedArgmin_effectiveDomain_iff_zero_mem_commonRegularSubdifferential
{f : V → WithTop ℝ} {XStar : Set V} :
XStar ⊆ argmin[dom f] f ↔ (0 : V) ∈ ∂̂ f(XStar)Corollary 3.1.6
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Corollary 3.1.6 is a recall-only item in the chapter's convex-analysis/subdifferential domain. Sampled owner-style declarations: - subdifferential - mem_subdifferential_iff - subgradient_nonneg_on_sublevelSet_of_mem_subdifferential - subgradient_inner_sub_nonneg_of_isMinOn Best owner abstraction: - the canonical subdifferential owner API together with the owner theorem subgradient_nonneg_on_sublevelSet_of_mem_subdifferential Primitive data: - a set Q, a function f, points x0, xStar, g - the membership hypothesis x0 ∈ Q - the minimizer hypothesis IsMinOn f Q xStar - the subdifferential-membership hypothesis g ∈ subdifferential f x0 Derived API: - the pairing inequality 0 ≤ inner ℝ g (x0 - xStar) Source/core/bridge triage: - source-facing: the corollary that every subgradient at a feasible point has nonnegative pairing with the displacement to a minimizer - core/canonical: the subdifferential owner API and the owner theorem subgradient_nonneg_on_sublevelSet_of_mem_subdifferential - bridge/view: the existing chapter theorem subgradient_inner_sub_nonneg_of_isMinOn, which already has the exact corollary interface This item therefore keeps only the direct recall of that source-facing bridge theorem instead of introducing a second local wrapper around the owner abstraction.
recall subgradient_inner_sub_nonneg_of_isMinOnDefinition 3.1.6
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.6: for a closed convex set X ⊆ dom f, the epigraph facet of X with respect to f is the common regular subdifferential ∂̂ f(X) = ⋂ x ∈ X, ∂ f(x).
open scoped WithTopConvexAnalysis
universe u
recall commonRegularSubdifferential
{V : Type u} [SeminormedAddCommGroup V] [InnerProductSpace ℝ V]
(f : V → WithTop ℝ) (X : Set V) : Set VTheorem 3.1.6
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.6: if φ is closed and convex on S ⊆ ℝᵐ, then the composition x ↦ φ (A x + b) is closed and convex on the inverse image {x ∈ ℝⁿ | A x + b ∈ S}.
variable {m n : ℕ}
local notation "Eₙ" => EuclideanSpace ℝ (Fin n)
local notation "Eₘ" => EuclideanSpace ℝ (Fin m)
namespace ClosedConvexOn
theorem comp_linearMap_add
{S : Set Eₘ} {φ : Eₘ → WithTop ℝ}
(hφ : ClosedConvexOn S φ)
(A : Eₙ →ₗ[ℝ] Eₘ) (b : Eₘ) :
ClosedConvexOn {x : Eₙ | A x + b ∈ S} (fun x ↦ φ (A x + b))Corollary 3.1.7
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Corollary 3.1.7 is the whole-space specialization of the chapter owner theorem convexOn_log_sum_exp_of_convexOn.
open scoped BigOperators
universe u v
variable {ι : Type u} {E : Type v} [AddCommMonoid E] [Module ℝ E]
#check
(show ∀ {t : Finset ι} {f : ι → E → ℝ},
t.Nonempty →
(∀ i ∈ t, ConvexOn ℝ Set.univ (f i)) →
ConvexOn ℝ Set.univ (fun x ↦ Real.log (∑ i ∈ t, Real.exp (f i x))) from
convexOn_log_sum_exp_of_convexOn Set.univ)Definition 3.1.7
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.1.7: a function on a cone is positively homogeneous of degree p when its domain is closed under nonnegative scalar multiplication and scaling the input by a bundled nonnegative scalar τ : NNReal scales the value by τ ^ p.
universe u v
variable {E : Type u} {F : Type v} [SMul NNReal E] [SMul ℝ F]
class IsPositivelyHomogeneousOn (p : ℝ) (s : Set E) (f : E → F) : Prop where
/-- A positively homogeneous function has a domain closed under nonnegative scalar
multiplication. -/
smul_mem {x : E} (hx : x ∈ s) (τ : NNReal) : τ • x ∈ s
/-- A positively homogeneous function satisfies the prescribed nonnegative scaling identity on
its domain. -/
map_smul {x : E} (hx : x ∈ s) (τ : NNReal) :
f (τ • x) = Real.rpow (τ : ℝ) p • f xLemma 3.1.7
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.7, derived owner form: if a convex extended-real-valued function on a complete real inner-product space is differentiable at an interior point of its effective domain, then its subdifferential there is the singleton containing the gradient of its finite real part.
open scoped Gradient Topology
open Filter
universe u
open scoped WithTopConvexAnalysis
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E] [CompleteSpace E]
theorem subdifferential_eq_singleton_gradient
{f : E → WithTop ℝ} (hf : ConvexOn ℝ (dom f) (withTopRealPart f)) {x : E}
(hx : x ∈ interior (dom f))
(hfd : DifferentiableAt ℝ (withTopRealPart f) x) :
∂ f(x) = {∇ (withTopRealPart f) x}Theorem 3.1.7
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.7 is a recall-only surface in the chapter's constrained partial-infimum / extended-real convex-analysis domain. Primary mathematical domain: - constrained fiberwise infima of real-valued convex functions, viewed through the chapter's EReal finite-value bridge. Relevant owner-style declarations sampled before refinement: - partialInfProjection in Theorem_3_1_2_3, the source-facing owner for constrained fiberwise infima; - extendedRealRealPart in Definition_3_1_1_3, the chapter bridge from EReal values to their real part on the finite-value domain; - partialInfProjection_convexOn in Theorem_3_1_2_3, the canonical convexity theorem for that owner surface; - partialInfProjection_convexOn_of_convexWithTop in Theorem_3_8, the later WithTop analogue that reuses the same owner abstraction. Best owner abstraction: - partialInfProjection_convexOn. Primitive data: - none in this file; the owner object and its convexity theorem already live upstream. Derived API: - this recall-only numbered entry point. Source/core/bridge triage: - source-facing: Theorem 3.1.7's statement that the constrained partial infimum of a real-valued convex function is convex; - core/canonical: partialInfProjection_convexOn on the chapter owner partialInfProjection; - bridge/view: the chapter EReal convexity surface ConvexOn ℝ (dom ψ) (extendedRealRealPart ψ). The previous file introduced a second theorem name with exactly the same statement as partialInfProjection_convexOn. Since the upstream theorem already lives on the correct owner abstraction and is reused elsewhere in the chapter, this file now recalls that canonical theorem directly instead of maintaining a duplicate wrapper.
recall partialInfProjection_convexOnCorollary 3.1.8
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Corollary 3.1.8 is a recall-only item in the chapter's extended-valued constrained-subdifferential domain. Primary domain: - convexity and lower-semicontinuity consequences of nonempty constrained subdifferentials for WithTop ℝ-valued functions on Euclidean space. Sampled owner-style declarations: - withTopRealPart - constrainedSubdifferential - convexOn_of_constrainedSubdifferential_nonempty - lowerSemicontinuousOn_of_constrainedSubdifferential_nonempty Best owner abstraction: - the owner theorem pair convexOn_of_constrainedSubdifferential_nonempty and lowerSemicontinuousOn_of_constrainedSubdifferential_nonempty, built from the chapter owner objects withTopRealPart and constrainedSubdifferential Primitive data: - a convex feasible set Q - a WithTop ℝ-valued objective f - pointwise nonemptiness of constrainedSubdifferential Q f x on Q Derived API: - convexity of withTopRealPart f on Q - lower semicontinuity of withTopRealPart f on Q Source/core/bridge triage: - source-facing: the textbook corollary collecting the convexity and lower-semicontinuity consequences - core/canonical: withTopRealPart and constrainedSubdifferential - bridge/view: none; both consequences already exist upstream with the exact public interfaces needed here The previous file-level conjunction theorem was a redundant wrapper around those two owner consequences, and it had no downstream users. This file therefore keeps only direct recalls of the upstream owner theorems instead of a parallel local packaging layer.
recall convexOn_of_constrainedSubdifferential_nonemptyLemma 3.1.8
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.8 is a recall-only bridge in the chapter's extended-valued convex-composition/subdifferential-calculus domain. Sampled owner-style declarations: - withTopEffectiveDomain in Definition_3_3, the chapter owner for the finite-value domain - withTopRealPart in Definition_3_3, the owner finite-value representative - subdifferential in Definition_3_1_5, the owner subgradient-set API - monotoneConvexComp_convexOn and subdifferential_monotoneConvexComp_eq_convexHull in Lemma_3_8, the source owner clauses for this monotone composition lemma on real inner-product spaces Best owner abstraction: - the source-facing composition object monotoneConvexComp together with its convexity and subdifferential owner clauses in Lemma_3_8 - the ambient effective-domain / finite-part / subdifferential notions remain owned upstream by Definition_3_3 and Definition_3_1_5 Primitive data: - none in this recall file Derived API: - monotoneConvexComp - monotoneConvexComp_apply_of_mem_effectiveDomain - monotoneConvexComp_convexOn - subdifferential_monotoneConvexComp_eq_convexHull Source/core/bridge triage: - source-facing: the monotone composition object and its two textbook clauses recalled below - core/canonical: withTopEffectiveDomain, withTopRealPart, ConvexOn ℝ (dom f) (withTopRealPart f), subdifferential - bridge/view: this recall-only file The previous version redefined the ambient effective-domain, finite-real-part, convexity, and subdifferential API locally, then split the owner theorem into two wrapper consequences. Those parallel wrappers had no downstream users. This file now recalls the owner composition surface directly instead of maintaining a second root API. The owner surface in Lemma_3_8 is now itself split into the two atomic textbook clauses, so this recall file follows that same canonical surface and drops the redundant properness hypotheses that were only artifacts of the earlier coordinate-specialized statement.
universe u
open scoped WithTopConvexAnalysis
recall monotoneConvexComp {V : Type u} (φ : ℝ → WithTop ℝ) (ψ : V → WithTop ℝ) :
V → WithTop ℝTheorem 3.1.8
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.8: if every slice x ↦ φ(x, y) with y ∈ Δ is closed and convex on Q, then the pointwise supremum x ↦ sup_{y ∈ Δ} φ(x, y) is closed and convex on the finite-value domain {x ∈ Q | sup_{y ∈ Δ} φ(x, y) < +∞}.
universe u v
open scoped WithTopConvexAnalysis
variable {ι : Type u} {X : Type v}
variable [TopologicalSpace X] [AddCommMonoid X] [Module ℝ X]
namespace ClosedConvexOn
theorem pointwise_sSup
{Q : Set X} {Δ : Set ι} {φ : X → ι → WithTop ℝ}
(hΔ : Δ.Nonempty)
(hφ : ∀ y ∈ Δ, ClosedConvexOn Q (fun x ↦ φ x y)) :
ClosedConvexOn (pointwiseSupremumOnEffectiveDomain Q Δ φ) (pointwiseSupremumOn Δ φ)Lemma 3.1.9
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.9 is the upstream neighborhood-form theorem subdifferential_eq_image_partialGradientFst_partialSubdifferentialSnd_of_nhds from Lemma_3_9.
recall subdifferential_eq_image_partialGradientFst_partialSubdifferentialSnd_of_nhdsTheorem 3.1.9
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.9: if ψ is convex on its domain domψ ⊆ X, and φ : Y → Z is convex and nondecreasing on all of Y, then the composition x ↦ φ (ψ x) is convex on domψ.
universe u v w
namespace ConvexOn
variable {𝕜 : Type u} [Semiring 𝕜] [PartialOrder 𝕜]
variable {X : Type v} [AddCommMonoid X] [SMul 𝕜 X]
variable {Y : Type w} [AddCommMonoid Y] [PartialOrder Y] [SMul 𝕜 Y]
variable {Z : Type*} [AddCommMonoid Z] [PartialOrder Z] [SMul 𝕜 Z]
variable {domψ : Set X} {ψ : X → Y} {φ : Y → Z}
theorem comp_of_monotone (hφ : ConvexOn 𝕜 Set.univ φ) (hψ : ConvexOn 𝕜 domψ ψ)
(hφ_mono : Monotone φ) :
ConvexOn 𝕜 domψ (φ ∘ ψ)Lemma 3.1.10
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.10 is a recall-only bridge in the chapter's extended-valued convex-analysis / continuous-subgradient-selection domain. Primary domain: - convex analysis of WithTop ℝ-valued functions on real inner-product spaces, with a continuous local selection of pointwise subgradients. Relevant sampled declarations in this domain: - subdifferential from Definition_3_1_5, the chapter owner for extended-valued subgradients; - withTopRealPart from Definition_3_3, the canonical real-valued representative on dom f; - HasGradientAt, the canonical gradient owner for differentiability with a specified gradient; - hasGradientAt_withTopRealPart_of_continuous_subgradient_selection from Lemma_3_10, the exact upstream owner theorem for this source item. Best owner abstraction: - the exact upstream owner theorem hasGradientAt_withTopRealPart_of_continuous_subgradient_selection from Lemma_3_10. Primitive data: - none in this recall file; the actual source-facing assumptions already live in Lemma_3_10. Derived API recalled here: - the owner HasGradientAt conclusion for the selected subgradient; - the derived differentiability consequence; - the derived gradient-identification consequence. Source/core/bridge triage: - source-facing: Lemma 3.1.10's differentiability statement from a continuous local subgradient selection; - core/canonical: subdifferential, withTopRealPart, and HasGradientAt; - bridge/view: the derived DifferentiableAt and gradient-equality consequences. The previous version drifted into the unrelated finite pointwise-supremum / active-set API from Lemma_3_1_13. This file instead reuses the actual Lemma_3_10 owner family directly and keeps the later numbered item aligned with its true source mathematics.
recall hasGradientAt_withTopRealPart_of_continuous_subgradient_selectionTheorem 3.1.10
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.10: if f₁ and f₂ are real-valued functions whose canonical WithTop lifts are closed and convex on a feasible set Q, and every constrained sublevel set of the pointwise maximum x ↦ max (f₁ x) (f₂ x) on Q is bounded, then there exists a parameter λ* ∈ [0, 1] for which the constrained minimum value of x ↦ max (f₁ x) (f₂ x) on Q equals that of the convex combination x ↦ λ* f₁ x + (1 - λ*) f₂ x; Lean records this conclusion by the owner predicate IsMinimaxLinearizationParameter on the subtype Q.
recall exists_minimax_parameter_of_bounded_constrainedSublevelSetsTheorem 3.1.11
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.11's main owner is mathlib's ConvexOn.locallyLipschitzOn_interior; in the chapter, it is applied to C = dom f and f = withTopRealPart f.
open scoped Topology
open scoped WithTopConvexAnalysis
universe u
recall ConvexOn.locallyLipschitzOn_interiorLemma 3.1.12
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.12: if f₁ and f₂ are closed convex and α₁, α₂ ≥ 0, then for the book-style restricted-domain weighted sum f = weightedAdd α₁ α₂ f₁ f₂ one has interior (dom f) = interior (dom f₁) ∩ interior (dom f₂), and at every x ∈ interior (dom f) the subdifferential of f equals the Minkowski sum of the scaled pointwise subdifferentials.
open scoped Pointwise Topology WithTopConvexAnalysis
universe u
variable {X : Type u}
variable {V : Type u} [NormedAddCommGroup V] [InnerProductSpace ℝ V] [CompleteSpace V]
variable {f₁ f₂ : V → WithTop ℝ} {α₁ α₂ : ℝ}
theorem subdifferential_nonneg_weighted_add_eq
(hf₁ : ClosedConvexFunction f₁)
(hf₂ : ClosedConvexFunction f₂)
(hα₁ : 0 ≤ α₁)
(hα₂ : 0 ≤ α₂)
{x : V}
(hx : x ∈ interior (dom (weightedAdd α₁ α₂ f₁ f₂))) :
∂ (weightedAdd α₁ α₂ f₁ f₂)(x) =
α₁ • ∂ f₁(x) + α₂ • ∂ f₂(x)Theorem 3.1.12
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3 1 12: a convex ℝ ∪ {+∞}-valued function has a finite one-sided directional derivative in every direction at every interior point of its effective domain.
universe u
open Filter Set
open scoped Topology WithTopConvexAnalysis
variable {E : Type u} [TopologicalSpace E] [AddCommGroup E] [Module ℝ E]
[IsTopologicalAddGroup E] [ContinuousSMul ℝ E]
theorem exists_tendsto_right_directionalSlope_of_convexOn_of_mem_interior_effectiveDomain
{f : E → WithTop ℝ}
(hf : ConvexOn ℝ (dom f) (withTopRealPart f))
{x p : E} (hx : x ∈ interior (dom f)) :
∃ d : ℝ,
(∀ᶠ α : ℝ in 𝓝[>] (0 : ℝ), x + α • p ∈ dom f) ∧
Tendsto
(fun α : ℝ ↦
(withTopRealPart f (x + α • p) - withTopRealPart f x) / α)
(𝓝[>] (0 : ℝ)) (𝓝 d)Lemma 3.1.13
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.13 is a recall-only bridge in the chapter's finite-family pointwise-supremum / subdifferential calculus domain for WithTop ℝ-valued convex functions. Primary domain: - finite nonempty pointwise suprema of closed convex extended-real-valued functions on the chapter's intrinsic ambient spaces. Relevant sampled declarations in this domain: - pointwiseSupremumOn - activePointwiseSupremumOnIndices - ClosedConvexFunction - subdifferential Best owner abstraction: - the chapter owner pointwiseSupremumOn, specialized in Lemma_3_13 to the finite set Set.univ. Primitive data: - none in this recall file; the imported owner surface already carries the only primitive mathematical inputs, namely a nonempty finite index type together with a family φ : X → ι → WithTop ℝ. Derived API recalled here: - the finite-supremum bridge pointwiseSupremumOn_univ_eq_sup' - the finite active-set bridge - the closed-convex, interior-domain, and active-subdifferential theorems for the Set.univ specialization Source/core/bridge triage: - source-facing: the finite-family specialization of Lemma 3.13 - core/canonical: pointwiseSupremumOn, activePointwiseSupremumOnIndices, ClosedConvexFunction, and subdifferential - bridge/view: pointwiseSupremumOn_univ_eq_sup', mem_activePointwiseSupremumOnIndices_univ_iff, and this numbered recall surface This file now recalls the owner-centered Set.univ specialization instead of maintaining a parallel finite-maximum wrapper layer.
universe u v
open scoped WithTopConvexAnalysis
variable {ι : Type u} [Fintype ι] [Nonempty ι]
recall pointwiseSupremumOn_univ_eq_sup'
Theorem 3.1.13
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3 1 13: if Q₁, Q₂ are closed convex subsets of a nontrivial proper real inner-product space with empty intersection, and one nonempty side is bounded, then they admit a strongly separating nesterovHyperplane. Specializing to finite-dimensional Euclidean spaces recovers the textbook ℝⁿ statement.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable [ProperSpace E] [Nontrivial E]
theorem areStronglySeparable_of_disjoint_closed_convex_of_bounded_one_side
(Q₁ Q₂ : Set E)
(hQ₁_closed : IsClosed Q₁) (hQ₂_closed : IsClosed Q₂)
(hQ₁_convex : Convex ℝ Q₁) (hQ₂_convex : Convex ℝ Q₂)
(hdisj : Disjoint Q₁ Q₂)
(hbounded : (Q₁.Nonempty ∧ Bornology.IsBounded Q₁) ∨
(Q₂.Nonempty ∧ Bornology.IsBounded Q₂)) :
AreStronglySeparable Q₁ Q₂Lemma 3.1.14
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.14, active-slice inclusion part: at every x ∈ pointwiseSupremumOnEffectiveDomain Q Δ φ, the constrained subdifferential of pointwiseSupremumOn Δ φ over pointwiseSupremumOnEffectiveDomain Q Δ φ contains the convex hull of the constrained subdifferentials of the active slices y ∈ I(x). The closed-convex part of Lemma 3.1.14 is the separate owner theorem ClosedConvexOn.pointwise_sSup.
open Set
open scoped WithTopConvexAnalysis
universe u v
variable {ι : Type u} {X : Type v}
variable {E : Type v} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
theorem convexHull_activePointwiseSupremumOnSubdifferentials_subset
{Q : Set E} {Δ : Set ι} {φ : E → ι → WithTop ℝ} {x : E}
(hx : x ∈ pointwiseSupremumOnEffectiveDomain Q Δ φ) :
convexHull ℝ
(⋃ y ∈ activePointwiseSupremumOnIndices Δ φ x,
∂[Q] (fun z ↦ φ z y)(x)) ⊆
∂[pointwiseSupremumOnEffectiveDomain Q Δ φ] (pointwiseSupremumOn Δ φ) (x)
Theorem 3.1.14
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.14 is a recall-only file in the chapter's affine-nesterovHyperplane support domain. Relevant sampled declarations: - AffineHyperplane - AffineHyperplane.IsSupporting - IsSupportingHyperplane - exists_supporting_hyperplane_at_boundary_point_of_closed_convex Best owner abstraction: - the earlier source-facing theorem exists_supporting_hyperplane_at_boundary_point_of_closed_convex, built on the chapter's AffineHyperplane owner API. Source/core/bridge triage: - source-facing: this numbered theorem item; - core/canonical: the earlier chapter theorem with the same mathematical content; - bridge/view: this later theorem name, which is only a textual restatement of the same result. Primitive data: - the closed convex set Q and the boundary point x₀. Derived API: - the coordinate witness (g, γ) together with membership in nesterovHyperplane g γ and the predicate IsSupportingHyperplane Q g γ. This file previously duplicated the exact theorem surface already provided by Theorem_3_1_4_2.lean. It now reuses that earlier declaration directly instead of keeping a second parallel theorem with the same statement under a different name.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
recall exists_supporting_hyperplane_at_boundary_point_of_closed_convex
[FiniteDimensional ℝ E] (Q : Set E) (hQ_closed : IsClosed Q) (hQ_convex : Convex ℝ Q)
{x₀ : E}
(hx₀ : x₀ ∈ frontier Q) :
∃ g : E, ∃ γ : ℝ, x₀ ∈ nesterovHyperplane g γ ∧ IsSupportingHyperplane Q g γLemma 3.1.15
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.15 lies in the chapter's real-valued subdifferential / positive-homogeneity domain. Primary domain: - subdifferentials of positively homogeneous real-valued functions on real inner-product spaces. Sampled owner-style declarations: - IsSubgradientAt in Definition_3_1_5, the chapter owner predicate for extended-valued subgradients; - subdifferential in Definition_3_1_5, the derived owner set-valued API; - IsPositivelyHomogeneousOn in Definition_3_1_7, the chapter owner predicate for positive homogeneity; - subdifferential_eq_subdifferential_zero_of_posHomogeneous in Lemma_3_15, the existing chapter theorem for this source fact. Best owner abstraction: - subdifferential_eq_subdifferential_zero_of_posHomogeneous, organized around subdifferential (fun y ↦ (f y : WithTop ℝ)) and IsPositivelyHomogeneousOn 1 Set.univ f. Primitive data: - a real inner-product space E; - a real-valued function f : E → ℝ; - the positive-homogeneity owner hypothesis IsPositivelyHomogeneousOn 1 Set.univ f. Derived API: - the subdifferential identity at x in terms of the origin subdifferential and the touching condition inner ℝ g x = f x. Source/core/bridge triage: - source-facing: Lemma 3.1.15's description of the subdifferential of a positively 1-homogeneous function; - core/canonical: IsSubgradientAt, subdifferential, and IsPositivelyHomogeneousOn; - bridge/view: none beyond the coercion fun y ↦ (f y : WithTop ℝ) already absorbed by the owner theorem in Lemma_3_15. The previous version duplicated a real-valued subgradient predicate, a real-valued subdifferential, and a parallel theorem with an extra convexity hypothesis. That convexity hypothesis is mathematically redundant here, and the owner theorem in Lemma_3_15 already states the source fact canonically. This file therefore recalls that theorem directly instead of keeping a second local wrapper API.
recall subdifferential_eq_subdifferential_zero_of_posHomogeneousTheorem 3.1.15
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3 1 15: generalized from the textbook Euclidean setting, if an ℝ ∪ {+∞}-valued convex function on a finite-dimensional real inner-product space is finite at an interior point x₀ of its effective domain, then the subdifferential ∂f(x₀) is nonempty and bounded. The textbook statement is the specialization E = EuclideanSpace ℝ (Fin n).
open scoped Topology
open scoped WithTopConvexAnalysis
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
local notation "Z" => WithLp 2 (E × ℝ)
local notation "eZ" => WithLp.prodContinuousLinearEquiv 2 ℝ E ℝ
theorem subdifferential_nonempty_and_isBounded_of_convexOn_effectiveDomain_of_mem_interior
[FiniteDimensional ℝ E] {f : E → WithTop ℝ}
(hf : ConvexOn ℝ (dom f) (withTopRealPart f))
{x0 : E} (hx0 : x0 ∈ interior (dom f)) :
(∂ f(x0)).Nonempty ∧ Bornology.IsBounded (∂ f(x0))Lemma 3.1.16
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.16 (1): if Λ ⊆ ℝ_+^m and each component f_i is convex, then the finite real part of the support-function composition x ↦ ξ[Λ] (vectorMap fs x) is convex on its effective domain.
open Set
open scoped BigOperators Gradient SupportFunction WithTopConvexAnalysis EuclideanOrthant
universe u
variable {m : ℕ} {E : Type u}
local notation "Eₘ" => EuclideanSpace ℝ (Fin m)
variable [AddCommMonoid E] [Module ℝ E]
theorem convexOn_supportFunction_comp_vectorMap
{Λ : Set Eₘ} {fs : Fin m → E → ℝ}
(hΛ_nonneg : Λ ⊆ ℝ₊^m)
(hfs_convex : ∀ i, ConvexOn ℝ Set.univ (fs i)) :
ConvexOn ℝ (extendedRealEffectiveDomain (ξ[Λ] ∘ vectorMap fs))
(extendedRealRealPart (ξ[Λ] ∘ vectorMap fs))
Theorem 3.1.16
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.16 (1): in the Euclidean specialization E = EuclideanSpace ℝ (Fin n), the Fenchel bidual is bounded above by the original value at every point of dom f.
open scoped ConvexAnalysis WithTopConvexAnalysis
variable {n : ℕ}
variable {f : EuclideanSpace ℝ (Fin n) → WithTop ℝ}
variable {x : EuclideanSpace ℝ (Fin n)}
#check
(fenchelBidual_le_of_mem_dom :
x ∈ dom f → (f⋆⋆) x ≤ withTopToEReal (f x))Lemma 3.1.17
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.17 (subdifferential part): in a finite-dimensional real inner-product space E, if F is convex and coordinatewise monotone on ℝ^m, and each component function f i is convex on the convex set Q, then at every interior feasible point x ∈ interior Q and for every gradient witness g = ∇F (f₁(x), ..., fₘ(x)), the constrained subdifferential ∂[Q] (fun y ↦ ((F ∘ vectorMap f) y : WithTop ℝ))(x) equals the weighted sum ∑ᵢ gᵢ • ∂[Q] (fun y ↦ (f i y : WithTop ℝ))(x). This keeps the owner notation ∂[Q] on the public theorem surface instead of unpacking it back to the raw set builder.
open scoped BigOperators Gradient Pointwise SupportFunction WithTopConvexAnalysis
universe u
variable {m : ℕ}
local notation "Y" => EuclideanSpace ℝ (Fin m)
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
theorem constrainedSubdifferential_comp_coordinatewiseMonotone_eq_weighted_sum
{Q : Set E} (hQ_convex : Convex ℝ Q)
{F : Y → ℝ} (hF_conv : ConvexOn ℝ Set.univ F)
{f : Fin m → E → ℝ} (hf_conv : ∀ i, ConvexOn ℝ Q (f i))
(hF_mono : Monotone (F ∘ (EuclideanSpace.equiv (Fin m) ℝ).symm))
[FiniteDimensional ℝ E]
{x : E} (hx : x ∈ interior Q) {g : Y}
(hF_grad : HasGradientAt F g (vectorMap f x)) :
∂[Q] (fun y ↦ ((F ∘ vectorMap f) y : WithTop ℝ))(x) =
∑ i : Fin m,
(g i) • ∂[Q] (fun y ↦ (f i y : WithTop ℝ))(x)
Theorem 3.1.17
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.17: for a convex ℝ ∪ {+∞}-valued function on ℝⁿ, the subdifferential with respect to the direction variable of p ↦ f'(x₀; p) at 0 coincides with the subdifferential of f at x₀. The companion directional-derivative max formula is recalled immediately below in the canonical IsGreatest owner form.
open scoped WithTopConvexAnalysis
open scoped Topology
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
variable {f : E → WithTop ℝ} {x0 : E}
theorem convexDirectionalDerivativeReal_subdifferential_eq_at_zero
(hf : ConvexOn ℝ (dom f) (withTopRealPart f))
(hx0 : x0 ∈ interior (dom f)) :
∂[Set.univ] f′[hx0](0) = ∂ f(x0)Lemma 3.1.18
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.18, owner-level equality form: for a convex set Q in a real normed space and a point xBar ∈ Q, the positive tangent cone at xBar is the closure of the canonical pointed cone hull of the feasible displacements Q - xBar. The textbook ℝⁿ closed boundary-point case is the specialization IsClosed Q and xBar ∈ frontier Q.
open scoped Pointwise
universe u
variable {E : Type u} [NormedAddCommGroup E] [NormedSpace ℝ E]
theorem posTangentConeAt_eq_closure_pointedConeHull_vsub_singleton
(Q : Set E) (hQ_convex : Convex ℝ Q) (xBar : E) (hxBar : xBar ∈ Q) :
posTangentConeAt Q xBar =
closure (PointedCone.hull ℝ (Q -ᵥ ({xBar} : Set E)))Theorem 3.1.18
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.18: every subgradient g ∈ ∂f(x₀) supports the lower level set 𝓛_f(f(x₀)) = {x ∈ dom f | f x ≤ f x₀} in the sense that ⟪g, x₀ - x⟫ ≥ 0 for every point of that set.
open scoped WithTopConvexAnalysis
universe u
variable {V : Type u} [SeminormedAddCommGroup V] [InnerProductSpace ℝ V]
theorem subgradient_nonneg_on_sublevelSet_of_mem_subdifferential
{f : V → WithTop ℝ} {x0 g : V}
(hg : g ∈ ∂ f(x0)) {x : V} (hx : f x ≤ f x0) :
0 ≤ inner ℝ g (x0 - x)Lemma 3.1.19
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.19: if xStar is a local minimizer of f on Q, then the intrinsic constrained directional-derivative map lineDerivWithin ℝ f Q xStar is nonnegative on the tangent cone posTangentConeAt Q xStar. The textbook ℝⁿ statement is the specialization to E = EuclideanSpace ℝ (Fin n).
open scoped Pointwise Topology
universe u
variable {E : Type u} [NormedAddCommGroup E] [NormedSpace ℝ E]
theorem directionalDerivative_nonneg_on_posTangentCone_of_isLocalMinOn
{Q : Set E} {f : E → ℝ} {xStar : E}
(hxStar : xStar ∈ Q) (hQ_convex : Convex ℝ Q) (hmin : IsLocalMinOn f Q xStar)
(hline_cont : Continuous (lineDerivWithin ℝ f Q xStar))
(hline_hom : IsPositivelyHomogeneousOn 1 Set.univ (lineDerivWithin ℝ f Q xStar))
(hline :
∀ (d : E) (_hd : ∀ᶠ β : ℝ in 𝓝[>] (0 : ℝ), xStar + β • d ∈ Q),
LineDifferentiableWithinAt ℝ f Q xStar d)
(p : E) (hp : p ∈ posTangentConeAt Q xStar) :
0 ≤ lineDerivWithin ℝ f Q xStar p
Theorem 3.1.19
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.19 (2): every common subgradient g ∈ ∂̂f(X) gives the affine increment formula f x₁ = f x₀ + ⟪g, x₁ - x₀⟫ for all x₀, x₁ ∈ X.
open scoped WithTopConvexAnalysis
universe u
variable {V : Type u} [SeminormedAddCommGroup V] [InnerProductSpace ℝ V]
theorem eq_add_inner_of_mem_commonRegularSubdifferential
{X : Set V} {f : V → ℝ}
{g x0 x1 : V} (hg : g ∈ ∂̂ f(X))
(hx0 : x0 ∈ X) (hx1 : x1 ∈ X) :
f x1 = f x0 + inner ℝ g (x1 - x0)Lemma 3.1.20
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.20 lies in the chapter's centered-ellipsoid / support-value domain. Sampled owner-style declarations: - affineEllipsoid in Lemma_3_2_7, the chapter owner for the textbook ellipsoid E(H, x̄); - isGreatest_inner_image_spdEllipsoid in Lemma_3_20, the exact maximum-attainment theorem on that owner ellipsoid; - IsGreatest.csSup_eq, the generic order-theoretic bridge from an explicit maximum statement to the corresponding supremum equality. Best owner abstraction: - source-facing/core owner: isGreatest_inner_image_spdEllipsoid; - bridge/view: hmax.csSup_eq. Primitive data: - A : Mat with hA : A.PosDef; - c : E. Derived API: - the centered ellipsoid affineEllipsoid A⁻¹ 0; - the image set ((fun x : E ↦ inner ℝ c x) '' affineEllipsoid A⁻¹ 0); - the companion supremum equality supplied canonically by hmax.csSup_eq. Source/core/bridge triage: - source-facing: the textbook maximum of ⟪c, x⟫ over the centered ellipsoid; - core/canonical: the earlier owner theorem isGreatest_inner_image_spdEllipsoid; - bridge/view: the generic companion IsGreatest.csSup_eq. This file is recall-only: once Lemma_3_20 is stated directly on the chapter ellipsoid owner, there is no reason to keep a second theorem name with the exact same interface.
open scoped InnerProduct
open scoped EllipsoidNotation
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
local notation "Mat" => Matrix (Fin n) (Fin n) ℝ
recall isGreatest_inner_image_spdEllipsoid
(A : Mat) (hA : A.PosDef) (c : E) :
IsGreatest ((fun x : E ↦ inner ℝ c x) '' E(A⁻¹, (0 : E)))
(Real.sqrt (inner ℝ c ((A⁻¹).toEuclideanLin c)))Theorem 3.1.20
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.20: a set X_* is contained in arg min_{x ∈ dom f} f(x) if and only if the zero vector lies in the common subdifferential ∂̂ f(X_*).
open scoped ConstrainedArgmin WithTopConvexAnalysis
universe u
variable {V : Type u} [SeminormedAddCommGroup V] [InnerProductSpace ℝ V]
theorem subset_constrainedArgmin_effectiveDomain_iff_zero_mem_commonRegularSubdifferential
{f : V → WithTop ℝ} {XStar : Set V} :
XStar ⊆ argmin[dom f] f ↔ (0 : V) ∈ ∂̂ f(XStar)Lemma 3.1.21
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.21 lies in the chapter's finite-dimensional Lagrangian-duality domain. Primary domain: - inequality-constrained Lagrangian duality with complementary slackness Sampled owner-style declarations: - LagrangianProblem.lagrangian and LagrangianProblem.lagrangianMinimizers in Chap01/Definition_1_10_2 - LagrangianProblem.dualOptimalValue_le_primalOptimalValue in Chap01/Proposition_1_10_8 - objective_gap_ge_weighted_constraint_violation_of_lagrangian_minimizer in Chap03/Lemma_3_21 Best owner abstraction: - the existing Chapter 3 theorem objective_gap_ge_weighted_constraint_violation_of_lagrangian_minimizer, stated using the owner problem : LagrangianProblem Q m together with the derived API problem.lagrangianMinimizers Primitive data: - the Lagrangian owner problem : LagrangianProblem Q m - the points xStar, xBar - the multiplier vector lambdaStar Derived API: - the owner Lagrangian-minimizer fiber problem.lagrangianMinimizers lambdaStar - complementary slackness at xStar Source/core/bridge triage: - source-facing: the textbook objective-gap inequality derived from a Lagrangian minimizer and complementary slackness - core/canonical: the Chapter 1 owner LagrangianProblem and its IsMinOn Lagrangian-minimizer interface - bridge/view: the previous local theorem name, which duplicated Lemma_3_21 without adding new mathematics The former file introduced a second public theorem name with exactly the same interface as the existing chapter theorem in Lemma_3_21. This refinement keeps Lemma 3.1.21 as direct canonical recall/use instead of a parallel wrapper theorem.
recall objective_gap_ge_weighted_constraint_violation_of_lagrangian_minimizerTheorem 3.1.21
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.21: if an ℝ ∪ {+∞}-valued function is homogeneous of degree p on its effective domain, then every subgradient g ∈ ∂ f(x) satisfies ⟪g, x⟫ = p f(x) on the finite real part of f.
open scoped WithTopConvexAnalysis
universe u
variable {E : Type u} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
theorem euler_homogeneous_function_theorem
{f : E → WithTop ℝ} {p : ℝ}
(hhom : IsPositivelyHomogeneousOn p (dom f) (withTopRealPart f))
{x g : E} (hg : g ∈ ∂ f(x)) :
inner ℝ g x = p * withTopRealPart f xLemma 3.1.22
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.22 lies in the chapter's parametric minimax / convex-analysis domain. Sampled owner-style declarations: - the slice-infimum value function u ↦ sInf ((fun x ↦ Ψ x u) '' P) - ClosedConvexOn.max_inter - exists_minimax_parameter_of_bounded_constrainedSublevelSets - exists_isMinOn_parametricMaximumObjective_eq_valueFunction_of_valueFunction_maximizer Best owner abstraction: - source-facing: the attained minimum of the two-slice maximum together with its value-function identification at a maximizing parameter; - core/canonical: the Chapter 3 owner theorem exists_isMinOn_parametricMaximumObjective_eq_valueFunction_of_valueFunction_maximizer; - bridge/view: the internal WithTop lift used only in the theorem's slice-geometry hypotheses. Primitive data: - the feasible set P - the parameter set S - the real-valued kernel Ψ - nonemptiness of P, which lets the dual closed-concavity owner recover Convex ℝ S - closed-convexity and bounded constrained sublevel sets of the primal slices - closed-concavity of the dual slices, encoded canonically by closed convexity of u ↦ -Ψ(x, u) - the canonical maximizing-parameter datum uStar ∈ S ∧ IsMaxOn (fun u ↦ sInf ((fun x ↦ Ψ x u) '' P)) S uStar Derived API: - the attained minimizer of x ↦ max (Ψ x u) (Ψ x uStar) - the equality of that minimum value with sInf ((fun x ↦ Ψ x uStar) '' P) Source/core/bridge triage: - source-facing: the textbook minimum-attainment statement for the maximum of two parameter slices of a convex-concave kernel; - core/canonical: the earlier Chapter 3 theorem above; - bridge/view: the internal WithTop lift used by the closed-convex owner API. The previous file introduced a second public theorem stated through the later wrapper maximinLowerValue, weakened the conclusion to a bare sInf equality, and kept redundant convexity data in the local API. This refinement removes that duplicate wrapper layer and makes the numbered item a direct recall of the canonical owner theorem from Lemma_3_22.
recall exists_isMinOn_parametricMaximumObjective_eq_valueFunction_of_valueFunction_maximizerTheorem 3.1.22
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.22: for bounded convex sets Q₁ and Q₂ with nonempty interior intersection in a finite-dimensional real inner-product space, the support function of Q₁ ∩ Q₂ at x is the minimum over all y of ξ_{Q₁}(x + y) + ξ_{Q₂}(-y), expressed here as the least element of the translated-sum value set.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
open scoped SupportFunction
theorem supportFunction_inter_isLeast_add_supportFunction
[FiniteDimensional ℝ E]
{Q₁ Q₂ : Set E}
(hQ₁_bounded : Bornology.IsBounded Q₁) (hQ₁_convex : Convex ℝ Q₁)
(hQ₂_bounded : Bornology.IsBounded Q₂) (hQ₂_convex : Convex ℝ Q₂)
(hQ_int : (interior (Q₁ ∩ Q₂)).Nonempty) (x : E) :
IsLeast
(Set.range fun y : E ↦ ξ[Q₁] (x + y) + ξ[Q₂] (-y))
(ξ[Q₁ ∩ Q₂] x)Lemma 3.1.23
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.23: if x_N satisfies the residual estimate (3.1.85), written here as f(x_N) - φ(\hat u_N) ≤ r_N, then the decomposition (f(x_N) - f^*) + (φ^* - φ(\hat u_N)) lies between 0 and the primal-dual gap f(x_N) - φ(\hat u_N), and that gap is itself bounded above by r_N.
universe u v
variable {E : Type u} {U : Type v} {α : Type*} [AddCommGroup α] [Preorder α]
[AddLeftMono α] [AddRightMono α]
theorem primal_dual_decomposition_mem_Icc_of_gap_le
{f : E → α} {φ : U → α} {xN : E} {uHat : U} {fStar φStar rN : α}
(h_primal : fStar ≤ f xN) (h_dual : φ uHat ≤ φStar)
(h_weak_duality : φStar ≤ fStar)
(h_gap : f xN - φ uHat ≤ rN) :
(f xN - fStar) + (φStar - φ uHat) ∈ Set.Icc 0 (f xN - φ uHat) ∧
f xN - φ uHat ≤ rNTheorem 3.1.23
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.23: a point xStar ∈ Q solves the convex composite minimization problem min_{x ∈ Q} (f x + Ψ x) if and only if it satisfies the variational inequality ⟪∇ f xStar, x - xStar⟫ + Ψ x ≥ Ψ xStar for every feasible point x ∈ Q.
open scoped Gradient WithTopConvexAnalysis
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E] [CompleteSpace E]
theorem isMinOn_add_convex_iff_forall_inner_gradient_add_ge
{Q : Set E} {f Ψ : E → ℝ}
(hf_conv : ConvexOn ℝ Q f)
(hΨ_conv : ConvexOn ℝ Q Ψ)
{xStar : E} (hxStar : xStar ∈ Q) (hf_diff : DifferentiableAt ℝ f xStar) :
IsMinOn (f + Ψ) Q xStar ↔
∀ x ∈ Q, inner ℝ (∇ f xStar) (x - xStar) + Ψ x ≥ Ψ xStarLemma 3.1.24
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.1.24: if the stagewise gap is bounded by an attained certificate maximum δMax N, and these maxima satisfy δMax N ≤ r_N, then for every stage N the decomposition (\hat f_N - f^*) + (\phi^* - φ(\hat u_N)) lies in the interval [0, \hat f_N - φ(\hat u_N)] and the gap itself is bounded by r_N.
open Filter
universe u v
variable {E : Type u} {U : Type v}
variable {P : Set E}
variable (δ : ℕ → E → ℝ) (hatf : ℕ → ℝ) (φ : U → ℝ) (uHat : ℕ → U)
variable (fStar φStar : ℝ) (r : ℕ → ℝ)
variable
(h_primal_lower : ∀ N, fStar ≤ hatf N)
(h_dual_upper_uHat : ∀ N, φ (uHat N) ≤ φStar)
(h_weak_duality : φStar ≤ fStar)
include h_primal_lower h_dual_upper_uHat h_weak_duality
local notation "gap" =>
fun N ↦ hatf N - φ (uHat N)
local notation "decomposition" =>
fun N ↦ (hatf N - fStar) + (φStar - φ (uHat N))
theorem primal_dual_decomposition_mem_Icc_and_gap_le_of_gap_le_max_delta
(δMax : ℕ → ℝ)
(hδMax : ∀ N, IsGreatest (δ N '' P) (δMax N))
(h_gap_le_δMax : ∀ N, hatf N - φ (uHat N) ≤ δMax N)
(h_δMax_le : ∀ N, δMax N ≤ r N)
(N : ℕ) :
decomposition N ∈ Set.Icc 0 (gap N) ∧ gap N ≤ r NTheorem 3.1.24
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3 1 24: a feasible point xStar ∈ Q belongs to the optimal set X* = argmin_{x ∈ Q} f(x) if and only if some subgradient at xStar lies in the normal cone of Q at xStar.
open scoped ConstrainedArgmin
open scoped TangentCone
open scoped WithTopConvexAnalysis
open scoped NormalCone
open scoped Topology
universe u
variable {V : Type u} [NormedAddCommGroup V] [InnerProductSpace ℝ V]
theorem mem_constrainedArgmin_iff_exists_subgradient_mem_normalCone
[FiniteDimensional ℝ V]
{Q : Set V} (hQ_convex : Convex ℝ Q)
{f : V → ℝ} (hf_conv : ConvexOn ℝ Set.univ f)
{xStar : V} (hxStar : xStar ∈ Q) :
xStar ∈ argmin[Q] f ↔
∃ gStar : V,
gStar ∈ ∂ (fun x : V ↦ (f x : WithTop ℝ))(xStar) ∧
gStar ∈ N[Q] xStarTheorem 3.1.25
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.25 (1): if φ : X × Y → (-∞, +∞] is convex on its effective domain and the canonical infimal projection over Q₂ is finite on Q₁, then its finite real part is convex on Q₁.
open scoped ConstrainedArgmin ConvexAnalysis WithTopConvexAnalysis
universe u v
variable {X : Type u} {Y : Type v}
variable [AddCommMonoid X] [Module ℝ X]
variable [AddCommMonoid Y] [Module ℝ Y]
variable {Q₁ : Set X} {Q₂ : Set Y} {φ : X × Y → WithTop ℝ}
theorem partialInfProjection_realPart_convexOn_of_convexWithTop
(hφ : ConvexOn ℝ (dom φ) (withTopRealPart φ))
(hQ₁_convex : Convex ℝ Q₁) (hQ₂_convex : Convex ℝ Q₂)
(hfinite :
Q₁ ⊆ dom (partialInfProjection (Set.univ ×ˢ Q₂) (withTopToEReal ∘ φ))) :
ConvexOn ℝ Q₁
(extendedRealRealPart (partialInfProjection (Set.univ ×ˢ Q₂) (withTopToEReal ∘ φ)))
Theorem 3.1.26
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.26: (Karush--Kuhn--Tucker) under convexity, ambient differentiability at every point of Q, and a strict-feasibility point for the constraints, a point xStar is an optimal solution of the inequality-constrained problem min {f₀(x) | x ∈ Q, fᵢ(x) ≤ 0 for all i} if and only if there is a nonnegative multiplier vector such that xStar is primal feasible, the canonical gradients ∇ f0 xStar and ∇ (fi i) xStar yield the Lagrangian variational inequality on Q, and the coordinates obey complementary slackness at xStar.
open scoped BigOperators Gradient EuclideanOrthant
universe u
variable {E : Type u}
variable {m : ℕ}
variable [NormedAddCommGroup E] [InnerProductSpace ℝ E] [CompleteSpace E]
local notation "Λ" => EuclideanSpace ℝ (Fin m)
local notation "V" => Λ × ℝ
theorem isMinOn_iff_exists_karush_kuhn_tucker_multiplier
{Q : Set E} {f0 : E → ℝ} {fi : Fin m → E → ℝ}
(hf0_conv : ConvexOn ℝ Q f0)
(hfi_conv : ∀ i : Fin m, ConvexOn ℝ Q (fi i))
{xStar : E}
(hf0_diff : ∀ x ∈ Q, DifferentiableAt ℝ f0 x)
(hfi_diff : ∀ i : Fin m, ∀ x ∈ Q, DifferentiableAt ℝ (fi i) x)
(hSlater : ∃ x ∈ Q, ∀ i : Fin m, fi i x < 0) :
(xStar ∈ Q ∧
(∀ i : Fin m, fi i xStar ≤ 0) ∧
IsMinOn f0 (inequalityConstrainedFeasibleSet Q fi) xStar) ↔
∃ lam : Λ,
lam ∈ ℝ₊^m ∧
xStar ∈ Q ∧
(∀ i : Fin m, fi i xStar ≤ 0) ∧
(∀ x : E, x ∈ Q →
0 ≤
inner ℝ
(∇ f0 xStar + ∑ i : Fin m, (lam i) • ∇ (fi i) xStar)
(x - xStar)) ∧
∀ i : Fin m, lam i * fi i xStar = 0Theorem 3.1.27
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.27. A point xStar ∈ Q is optimal for the equality-constrained problem exactly when it satisfies the equality constraint and admits a subgradient-multiplier certificate whose adjoint norm obeys the Slater-radius bound.
open scoped ConvexAnalysis
open scoped TangentCone
open scoped Topology
open scoped NormalCone
open scoped WithTopConvexAnalysis
universe u v
variable {E : Type u} {Λ : Type v}
variable [NormedAddCommGroup E] [InnerProductSpace ℝ E] [FiniteDimensional ℝ E]
variable [NormedAddCommGroup Λ] [InnerProductSpace ℝ Λ] [FiniteDimensional ℝ Λ]
theorem isMinOn_linearEqualityFeasibleSet_iff_exists_subgradient_multiplier_with_bound
{Q : Set E} {f : E → WithTop ℝ}
(hfQ : ConvexOn ℝ Q (withTopRealPart f))
(hf : ConvexOn ℝ (dom f) (withTopRealPart f))
(hQ_subset_interior : Q ⊆ interior (dom f))
(hlevel_bounded : ∀ α : ℝ, Bornology.IsBounded (constrainedSublevelSet Q f α))
{A : E →ₗ[ℝ] Λ} {b : Λ} {xBar : E} {ε : ℝ}
(hbar : A xBar = b) (hε : 0 < ε) (hball : Metric.ball xBar ε ⊆ Q) {xStar : E}
(hxStarQ : xStar ∈ Q) :
EqualityConstrainedOptimalPoint Q f A b xStar ↔
HasSubgradientMultiplierCertificateWithBound Q f A b xBar ε xStar
Theorem 3.1.28
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.28 (1): the affine-fiber projected value function is convex on its finite-value domain under the standard convexity hypotheses on f and Q.
recall affinePartialInfProjection_realPart_convexOn
Theorem 3.1.29
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.29: if every primal slice x ↦ Ψ x u with u ∈ S attains the unique minimizer x u on P, and if the lower-value function u ↦ sInf ((fun x ↦ Ψ x u) '' P) attains its maximum on S, then the primal minimum of f on P equals the dual maximum of that lower-value function on S.
open scoped WithTopConvexAnalysis
universe u v
variable {X : Type u} {U : Type v}
variable [NormedAddCommGroup X] [NormedSpace ℝ X] [FiniteDimensional ℝ X]
variable [AddCommMonoid U] [Module ℝ U]
variable {P : Set X} {S : Set U} {Ψ : X → U → ℝ} {f : X → ℝ}
variable (x : U → X)
variable {uStar : U}
theorem minimax_eq_of_unique_slice_argmin_and_attained_dual_max
(hΨ_closedConvex :
∀ ⦃u : U⦄, u ∈ S → ClosedConvexOn P (fun x ↦ (Ψ x u : WithTop ℝ)))
(hΨ_concave : ∀ ⦃x : X⦄, x ∈ P → ConcaveOn ℝ S (fun v ↦ Ψ x v))
(hf_eq :
∀ ⦃x : X⦄, x ∈ P →
(f x : WithTop ℝ) = pointwiseSupremumOn S (fun x' u ↦ (Ψ x' u : WithTop ℝ)) x)
(hx_mem : ∀ ⦃u : U⦄, u ∈ S → x u ∈ P)
(hx_min : ∀ ⦃u : U⦄, u ∈ S → IsMinOn (fun p ↦ Ψ p u) P (x u))
(hx_unique : ∀ ⦃u : U⦄, u ∈ S → ∀ p : X, IsMinOn (fun q ↦ Ψ q u) P p → p = x u)
(huStar : uStar ∈ S)
(huStar_max : IsMaxOn (fun u ↦ sInf ((fun x ↦ Ψ x u) '' P)) S uStar) :
sInf (f '' P) = sSup ((fun u ↦ sInf ((fun x ↦ Ψ x u) '' P)) '' S)Theorem 3.1.30
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3 1 30: for a bounded max-representation primal-dual problem on proper real normed spaces, if the primal feasible set is nonempty and both the primal feasible set and dual set are bounded, then the primal minimum equals the dual maximum: min_{x ∈ P} f(x) = max_{u ∈ S} inf_{x ∈ P} Ψ(x, u). Here the data P, S, Ψ, and f are carried canonically by problem : MaxRepresentationPrimalDualProblem E U, rather than repeated as separate tuple arguments.
universe u v
open scoped BigOperators WithTopConvexAnalysis
variable {E : Type u} {U : Type v}
variable [NormedAddCommGroup E] [NormedSpace ℝ E] [ProperSpace E]
variable [NormedAddCommGroup U] [NormedSpace ℝ U] [ProperSpace U]
theorem minimax_eq_of_bounded_maxRepresentationPrimalDualProblem
(problem : MaxRepresentationPrimalDualProblem E U)
(hfeasible_nonempty : problem.feasibleSet.Nonempty)
(hfeasible_bounded : Bornology.IsBounded problem.feasibleSet)
(hdual_bounded : Bornology.IsBounded problem.dualSet) :
sInf (problem '' problem.feasibleSet) =
sSup ((fun u ↦ sInf ((fun x ↦ problem.kernel x u) '' problem.feasibleSet)) ''
problem.dualSet)
Theorem 3.1.31
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.1.31: let f : X → ℝ be a real-valued objective on P whose WithTop ℝ lift agrees with the faithful upper-envelope owner pointwiseSupremumOn S (fun x u ↦ (Ψ x u : WithTop ℝ)) on P. If xStar minimizes f on P, if some relative subgradient gStar ∈ ∂_P f(xStar) satisfying the first-order optimality relation ⟪gStar, x - xStar⟫ = 0 on P admits a simplex representation by active section subgradients g i ∈ ∂_P (Ψ(·, u_i))(xStar), and if the real lower slices x ↦ Ψ(x, u) are exposed only under bounded-below hypotheses on P, then the minimax relation (3.1.78) holds: min_{x ∈ P} f(x) = max_{u ∈ S} inf_{x ∈ P} Ψ(x, u).
open scoped BigOperators
local notation:max "∂[" Q "] " f:arg "(" x:arg ")" => subdifferentialWithin Q f x
universe u v w
variable {X : Type u} {U : Type v} {ι : Type w}
variable [Fintype ι]
variable [SeminormedAddCommGroup X] [InnerProductSpace ℝ X]
variable [AddCommGroup U] [Module ℝ U]
-- Proof sketch: use the constrained subgradient inequality for `gStar` together with the
-- optimality relation `⟪gStar, x - xStar⟫ = 0` on `P`. Rewrite `gStar` as the simplex-weighted sum
-- of the active section subgradients `g i`, combine their section inequalities, and use the
-- barycenter inequality
-- `∑ i, weights.weights i * Ψ(x, u i) ≤ Ψ(x, (Finset.univ).centerMass weights.weights u)` to show
-- that `f xStar ≤ sInf ((fun x ↦ Ψ x uBar) '' P)` for the barycenter parameter
-- `uBar = (Finset.univ).centerMass weights.weights u`. The faithful upper-owner bridge and weak
-- duality then yield the minimax equality `(3.1.78)`.
theorem minimax_eq_of_activeSubgradientRepresentation_at_minimizer
{P : Set X} {S : Set U} {Ψ : X → U → ℝ} {f : X → ℝ}
{xStar gStar : X}
(hf_eq :
∀ ⦃x : X⦄, x ∈ P →
(f x : WithTop ℝ) = pointwiseSupremumOn S (fun x' u ↦ (Ψ x' u : WithTop ℝ)) x)
(hxStar_min : IsMinOn f P xStar)
(hgStar_mem : gStar ∈ ∂[P] f(xStar))
(horth : ∀ ⦃x : X⦄, x ∈ P → inner ℝ gStar (x - xStar) = 0)
(weights : StdSimplex ℝ ι)
(u : ι → U) (g : ι → X)
(hu_active :
∀ i : ι,
u i ∈ activePointwiseSupremumOnIndices S
(fun x u ↦ (Ψ x u : WithTop ℝ)) xStar)
(hg_mem :
∀ i : ι, g i ∈ ∂[P] (fun x ↦ Ψ x (u i)) (xStar))
(hgStar_repr : gStar = ∑ i, weights.weights i • g i)
(hΨ_bddBelow :
∀ ⦃u : U⦄, u ∈ S → BddBelow ((fun x ↦ Ψ x u) '' P))
(hu_bar_mem :
(Finset.univ).centerMass weights.weights u ∈ S)
(hbar_domination :
∀ ⦃x : X⦄, x ∈ P →
(∑ i, weights.weights i * Ψ x (u i)) ≤
Ψ x ((Finset.univ).centerMass weights.weights u)) :
sInf (f '' P) =
sSup ((fun u ↦ sInf ((fun x ↦ Ψ x u) '' P)) '' S)
3.2 — Methods of Nonsmooth Minimization
Corollary 3.2.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Corollary 3.2.1 lies in the constrained strong-convexity / relative-subdifferential domain. Sampled owner-style declarations: - mathlib StrongConvexOn - subdifferentialWithin in Theorem_3_44 - mem_subdifferentialWithin_iff in Theorem_3_44 - StrongConvexOn.lower_bound_of_mem_subdifferentialWithin in Theorem_3_44 Best owner abstraction: - core/canonical: the generic lower-bound theorem StrongConvexOn.lower_bound_of_mem_subdifferentialWithin - supporting owner notions: StrongConvexOn Q μ f and g ∈ ∂[Q] f(x) Source/core/bridge triage: - source-facing: Corollary 3.2.1, the quadratic affine lower bound produced by a feasible subgradient of a strongly convex function - core/canonical: StrongConvexOn.lower_bound_of_mem_subdifferentialWithin - bridge/view: none; this file is a direct recall of the owner theorem Primitive data: - an inner-product space, feasible set Q, objective f, modulus μ, base point x, comparison point y, and vector g - the owner hypotheses StrongConvexOn Q μ f, g ∈ ∂[Q] f(x), and y ∈ Q Derived API: - the quadratic affine lower bound at y The earlier chapter duplicates of subdifferentialWithin are Euclidean-specialized, while Theorem_3_44 is the generic inner-product-space owner needed here. This corollary therefore keeps the owner theorem itself as the public center and adds no parallel local wrapper.
recall StrongConvexOn.lower_bound_of_mem_subdifferentialWithinDefinition 3.2.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.2.1 lies in the cutting-plane localization-set / localization-radius domain. Primary mathematical domain: - cutting-plane localization sets in real inner-product spaces together with their associated pointwise and best localization radii. Sampled owner-style declarations: - localizationSet - mem_localizationSet_iff - subgradientLocalizationMeasure - localization_radius Best owner abstraction for this file: - source-facing: the stage localization set localizationSet Q xSeq gSeq k, the associated pointwise radii subgradientLocalizationMeasure g xStar (xSeq i), and their minimum localization_radius xStar g xSeq k - core/canonical: the same project owners from Lemma_3_2_1 and Theorem_3_2_9 - bridge/view: localization_radius_le_measure and the closed-ball comparison theorems closedBall_subset_localizationSet_of_le_localization_radius and le_localization_radius_of_closedBall_subset_localizationSet Primitive data: - a feasible region Q - a reference point xStar - a query-point sequence xSeq - either a cut sequence gSeq for S_k, or the chosen subgradient selection g feeding the generic owner subgradientLocalizationMeasure behind v_i and v_k^* - a stage index k Derived API: - the defining expansion of S_k via mem_localizationSet_iff - the comparison v_k^* ≤ v_i via localization_radius_le_measure - the centered-ball characterization of v_k^* via the closed-ball inclusion and converse theorems This file stays at the source-facing layer: it recalls the canonical owners for S_k, the associated radii v_i and v_k^*, and the ball characterization of v_k^*, without introducing any local wrapper or parallel chapter API.
recall localizationSetLemma 3.2.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.2.1, generalized from the textbook Euclidean setting: if g x is a subgradient of f at x, then the increment f x - f xBar is bounded by the growth function ω_f(xBar; v_f(xBar; x)) built from the same chosen value g x, i.e. formula (3.2.11).
universe u
variable {V : Type u} [NormedAddCommGroup V] [InnerProductSpace ℝ V]
variable {f : V → ℝ} {g : V → V}
open scoped PointwiseGrowthFunction
open scoped SubgradientLocalizationMeasure
theorem sub_le_pointwiseGrowthFunction_of_localizationMeasure
(xBar x : V)
(hgx : IsSubgradientAt (fun y ↦ (f y : WithTop ℝ)) x (g x)) :
f x - f xBar ≤ ω[f; xBar] (v[g; xBar] x)Theorem 3.2.1
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.2.1: for every k with 0 ≤ k ≤ n - 1, there exist an objective f and a chosen minimizer x* in the class 𝒫(x₀, R, M) such that every first-order oracle whose replies satisfy the resisting-oracle prefix-support rule, together with every iterate sequence satisfying the span condition for that oracle, has objective gap at least MR / (2 (2 + √(k + 1))) at step k.
universe u
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
local notation "e[" i "]" => EuclideanSpace.single i (1 : ℝ)
open scoped LipschitzConvexProblemClass CoordinateSubspace
theorem exists_problem_with_nonsmooth_firstOrder_lower_bound
(n : ℕ) (x0 : EuclideanSpace ℝ (Fin n)) (R M : NNReal) {k : ℕ} (hk : k + 1 ≤ n) :
∃ f : EuclideanSpace ℝ (Fin n) → ℝ, ∃ xStar : EuclideanSpace ℝ (Fin n),
𝒫(x0, R, M) f xStar ∧
∀ oracle : FirstOrderOracle f,
HasCoordinateSupportGrowth x0 oracle.subgradient k →
∀ xSeq : ℕ → EuclideanSpace ℝ (Fin n),
SatisfiesLinearSpanCondition x0 oracle.subgradient xSeq k →
f (xSeq k) - f xStar ≥
((M : ℝ) * (R : ℝ)) / (2 * (2 + Real.sqrt (k + 1 : ℝ)))
Corollary 3.2.2
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Corollary 3.2.2 is a source-facing consequence in the chapter's constrained strong-convexity domain on proper real normed spaces. The textbook finite-dimensional real inner-product-space case is recovered by the canonical finite-dimensional properness instance. Sampled owner-style declarations: - mathlib StrongConvexOn - StrongConvexOn.isBounded_constrainedSublevelSet in Theorem_3_45 - StrongConvexOn.existsUnique_isMinOn_of_isClosed in Theorem_3_45 Best owner abstraction: - source-facing: existence of a feasible minimizer on a nonempty closed feasible set - core/canonical: StrongConvexOn Q μ f - bridge/view: projection from the unique-minimizer owner theorem to the existence-only statement Primitive data: - a feasible set Q, an objective f, and a strong-convexity modulus μ - the owner predicate StrongConvexOn Q μ f - continuity of f on the feasible set Derived API: - boundedness of constrained sublevel sets - under closed/nonempty feasible-set hypotheses, existence and uniqueness of a feasible minimizer - the source-facing existence-only consequence extracted from the unique-minimizer owner theorem Source/core/bridge triage: - source-facing: Corollary 3.2.2, which stops at boundedness of constrained level sets and the resulting existence of an optimal solution - core/canonical: StrongConvexOn Q μ f - bridge/view: projection from the unique-minimizer owner theorem to the existence-only statement The stronger uniqueness consequence is deferred to Corollary_3_2_3, so this file keeps the bounded-sublevel owner theorem as the public center and exposes only the existence bridge needed for the source text.
universe u recall StrongConvexOn.isBounded_constrainedSublevelSet
Definition 3.2.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Definition 3.2.2: a real-valued function is strongly convex on the feasible set Q exactly when there exists a positive modulus μ such that StrongConvexOn Q μ f. The textbook segment inequality is recorded below as the canonical bridge for this positive-existential owner.
universe u
variable {E : Type u} [NormedAddCommGroup E] [NormedSpace ℝ E]
variable {Q : Set E} {μ : ℝ} {f : E → ℝ}
#check (∃ μ > 0, StrongConvexOn Q μ f)Lemma 3.2.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.2.2: if f_k^* = min_{0 ≤ i ≤ k} f(x_i) and v_k^* = min_{0 ≤ i ≤ k} v_i, then the best function-value gap up to step k is bounded by the modulus value at the best radius. This specializes to ω_f(x^*; v_k^*) when ω = ω_f(x^*; ·), with ω allowed to take the value ⊤ when the growth profile is unbounded.
universe u
variable {X : Type u}
theorem bestFunctionValueGapUpTo_le_modulusAtBestRadius
(f : X → ℝ) (ω : ℝ → WithTop ℝ) (hω_mono : Monotone ω)
(xSeq : ℕ → X) (xStar : X) (radii : ℕ → ℝ) (k : ℕ)
(hbound : ∀ i : Fin (k + 1), f (xSeq i) - f xStar ≤ ω (radii i)) :
bestFunctionValueUpTo (fun i ↦ f (xSeq i)) k - f xStar ≤
ω (bestRadiusUpTo radii k)Theorem 3.2.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.2.2: if xStar is a global minimizer of f, f is M-Lipschitz on the closed Euclidean ball B₂(xStar, R), ‖x₀ - xStar‖ ≤ R, and the minimum localization measure v_k^* = bestRadiusUpTo (fun i ↦ v_f(xStar; xᵢ)) k satisfies the standard projected-subgradient estimate v_k^* ≤ (R^2 + ∑_{i=0}^k h_i^2) / (2 ∑_{i=0}^k h_i), then the best objective gap up to step k satisfies f_k^* - f^* ≤ M * (R^2 + ∑_{i=0}^k h_i^2) / (2 ∑_{i=0}^k h_i), where f_k^* = bestFunctionValueUpTo (fun i ↦ f (xSeq i)) k.
open scoped PointwiseGrowthFunction SubgradientLocalizationMeasure
universe u
variable {V : Type u} [NormedAddCommGroup V] [InnerProductSpace ℝ V]
theorem bestFunctionValueGapUpTo_le_lipschitz_mul_stepsize_ratio
{f : V → ℝ} {g : V → V}
(hg : ∀ x : V, IsSubgradientAt (fun y ↦ (f y : WithTop ℝ)) x (g x))
(xSeq : ℕ → V) {xStar : V} (hxStar : IsMinOn f Set.univ xStar)
(h : ℕ → ℝ) (k : ℕ) {R : ℝ} {M : NNReal}
(hLip : LipschitzOnWith M f (Metric.closedBall xStar R))
(hR : ‖xSeq 0 - xStar‖ ≤ R)
(hvk :
bestRadiusUpTo
(fun i ↦ v[g; xStar] (xSeq i)) k ≤
(R ^ (2 : ℕ) + ∑ i : Fin (k + 1), (h i) ^ (2 : ℕ)) /
(2 * ∑ i : Fin (k + 1), h i)) :
bestFunctionValueUpTo (fun i ↦ f (xSeq i)) k - f xStar ≤
(M : ℝ) *
((R ^ (2 : ℕ) + ∑ i : Fin (k + 1), (h i) ^ (2 : ℕ)) /
(2 * ∑ i : Fin (k + 1), h i))Corollary 3.2.3
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Corollary 3.2.3 (quadratic-growth part): this is the direct constrained owner theorem already exposed in Chapter 2.
universe u
namespace StrongConvexOn
variable {E : Type u} [NormedAddCommGroup E]
variable [NormedSpace ℝ E]
recall StrongConvexOn.quadratic_growth_of_isMinOn_of_memLemma 3.2.3
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.2.3: if f is μ-strongly convex on Q, then every interior point x of Q supports the quadratic lower bound f y ≥ f x + f'(x; y - x) + (μ / 2) * ‖x - y‖^2 at every y ∈ Q, where f'(x; y - x) is the one-sided directional derivative of the line restriction t ↦ f (x + t • (y - x)) at t = 0.
universe u
open Set AffineMap
open scoped Topology
variable {E : Type u} [NormedAddCommGroup E] [NormedSpace ℝ E]
namespace StrongConvexOn
theorem lower_tangent_derivWithin_of_mem_interior
{Q : Set E} {μ : ℝ} {f : E → ℝ} (hf : StrongConvexOn Q μ f)
{x y : E} (hx : x ∈ interior Q) (hy : y ∈ Q) :
f y ≥
f x +
derivWithin (fun t : ℝ ↦ f (x + t • (y - x))) (Set.Ici (0 : ℝ)) 0 +
(μ / 2) * ‖x - y‖ ^ (2 : ℕ)Theorem 3.2.3
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.2.3: let x_k = problem.iterates ε x₀ selectedConstraintAt k be the recursively generated run of method (3.2.24). If the branch choices are valid, if the objective f and the constraint functions f_j, j = 1, ..., m, are M-Lipschitz on the ball B₂(x*, ‖x₀ - x*‖), if x* belongs to the canonical constrained argmin set argmin[{x | ∀ j, f_j(x) ≤ 0}] f, and if the number of steps satisfies N ≥ (M² / ε²) ‖x₀ - x*‖², then the admissible iterate set 𝓕_A(N) is nonempty and the sampled minimum on 𝓕_A(N) satisfies f_N^* ≤ f(x*) + ε. The index-form consequence (𝒜(N)).Nonempty is a separate corollary via admissibleIterateSet_nonempty_iff.
universe u
open scoped ConstrainedArgmin WithTopConvexAnalysis
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable {m : ℕ}
namespace MultipleConstraintFirstOrderProblem
variable {problem : MultipleConstraintFirstOrderProblem E m}
variable {ε : ℝ} {x0 : E} {selectedConstraintAt : ℕ → Option (Fin m)}
open scoped MultipleConstraintFirstOrderProblemNotation
theorem admissibleIterateSet_nonempty_and_sampledMinimum_le_optimal_add_eps
(hvalid : problem.IsValidSelection ε x0 selectedConstraintAt) (xStar : E) (M : NNReal)
(hε : 0 < ε)
(hf_lipschitz :
LipschitzOnWith M problem.objective
(Metric.closedBall xStar ‖x0 - xStar‖))
(hconstraints_lipschitz :
∀ j, LipschitzOnWith M (problem.constraints j)
(Metric.closedBall xStar ‖x0 - xStar‖))
(hxStar :
xStar ∈ argmin[{x | ∀ j, problem.constraints j x ≤ 0}] problem.objective)
(N : ℕ)
(hN :
(((M : ℝ) ^ (2 : ℕ)) / (ε ^ (2 : ℕ))) * ‖x0 - xStar‖ ^ (2 : ℕ) ≤
(N : ℝ)) :
𝓕_A[problem, ε, x0, selectedConstraintAt](N) ≠ ∅ ∧
problem.sampledOptimalValue ε x0 selectedConstraintAt N ≤
problem.objective xStar + ε
Corollary 3.2.4
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Corollary 3.2.4 lives in the localization-radius / ellipsoid-volume domain. Sampled owner declarations: - localization_radius and localization_radius_le_outer_radius_mul_volume_ratio_rpow from Theorem_3_2_9 - inner_ball_radius_le_outer_radius_mul_volume_ratio_rpow_of_convex from Theorem_3_51 - selected_radius_bound_of_positive_index and selected_index_pos_of_volume_drop from Theorem_3_52 Owner abstraction: - the chapter localization-radius API, with Theorem_3_52 as the selected-index bridge. Primitive data: - Q, the raw query sequence querySeq, the localization map g, the comparison sequence Ell, and the owner selected-index data Nat.count / Nat.nth. Derived API: - the corollary's two source-facing conclusions are already exactly the bridge theorems from Theorem_3_52, so adding local copies here would only duplicate API. Triage: - source-facing: the selected-stage radius bound and positivity of Nat.count (fun j ↦ querySeq j ∈ Q) k - core/canonical: the localization-radius bound from Theorem_3_2_9 - bridge/view: the selected-index specialization from Theorem_3_52 This file therefore keeps Corollary 3.2.4 as direct canonical recalls and adds no parallel wrapper theorems.
recall selected_radius_bound_of_positive_indexLemma 3.2.4
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.2.4 [Chapter3_4.json:81]: every realized midpoint-bisection box in the chapter's resisting-oracle construction contains the Euclidean closed ball centered at its midpoint with textbook radius r_k = (R / 2) * (1 / 2)^(k / n), where k = state.depth; this is the geometric containment lemma that immediately yields the subsequent complexity result.
open FeasibilityResistingOracleState
open Lean Elab Tactic Meta
variable {n : ℕ}
theorem FeasibilityResistingOracleState.closedBall_subset_currentBox
(R : ℝ) (hn : 0 < n) (state : FeasibilityResistingOracleState n) :
Metric.closedBall (state.currentCenter R hn)
((R / 2) * Real.rpow (1 / 2 : ℝ) ((state.depth : ℝ) / n)) ⊆
state.currentBox R hn
Theorem 3.2.4
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.2.4: under the large-iteration hypothesis, the Definition 3.46 primal average is pointwise bounded below by the fixed-λ_t Lagrangian up to the error M h.
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable {m : ℕ}
namespace ApproximateLagrangeMultiplierSwitchingMethod
variable {problem : ProjectedMultipleConstraintFirstOrderProblem E m}
open scoped ApproximateLagrangeMultiplierSwitchingNotation
local notation "lagrangianProblem" => problem.toLagrangianProblem
local notation "lagrangianFeasibleSet" => problem.toLagrangianProblem.feasibleSet
open scoped ApproximateLagrangeMultiplierSwitchingNotation
lemma pointwiseLagrangianLowerBound_of_large_iteration_count
[CompleteSpace E]
(method : ApproximateLagrangeMultiplierSwitchingMethod problem) (R : NNReal) (t : ℕ)
(hdenom : method.HasApproximateDualMultiplierDenominators t)
(hh : 0 < method.h)
(hQ_bounded : ∀ x ∈ problem.feasibleSet, ‖x - method.x0‖ ≤ R)
(ht : ((R : ℝ) ^ (2 : ℕ)) / (method.h ^ (2 : ℕ)) < (t : ℝ))
(x : problem.feasibleSet) :
(((A₀[method](t)).centerMass
(fun k ↦ ‖method.objectiveSubgradient k‖⁻¹)
(fun k ↦ problem.objective (method k)) : ℝ) - M[method](t) * method.h) ≤
problem.toLagrangianProblem.lagrangian x (λ[method](t; hdenom))
Lemma 3.2.5
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.2.5 lies in the chapter's ellipsoid-method localization-containment domain. Primary domain: - ellipsoid-method cutting-plane recursions and the containment of the retained localization sets inside the ambient ellipsoid sequence. Sampled owner-style declarations: - localizationSets in Definition_3_52, the source-facing recursive retained-region family; - GeneralCuttingPlaneScheme.localizationSets_subset_localizer in Algorithm_3_6, the core containment theorem for any cutting-plane localizer; - EllipsoidMethod.toGeneralCuttingPlaneScheme in Algorithm_3_8, the bridge from the ellipsoid recursion to the cutting-plane owner abstraction; - EllipsoidMethod.localizationSets_subset_associatedEllipsoid in Algorithm_3_8, the exact ellipsoid-method containment theorem. Best owner abstraction: - source-facing: the ellipsoid-method localization sets and associated ellipsoid sequence; - core/canonical: GeneralCuttingPlaneScheme.localizationSets_subset_localizer; - bridge/view: EllipsoidMethod.localizationSets_subset_associatedEllipsoid. Primitive data: - the ambient convex minimization problem with separation oracle; - the initial center and radius; - the definedness, initial-cover, and positive-definiteness hypotheses for the ellipsoid recursion. Derived API: - the canonical query/cut localization sets of the ellipsoid recursion; - the associated ellipsoid sequence; - the source-facing containment theorem recalled below. The previous file kept an arbitrary induction shell on unrelated sequences S, Ell, and i. That shell was not the chapter owner abstraction, and the exact source-facing ellipsoid-method statement already exists upstream. This item is therefore recall-only and keeps the canonical bridge theorem as its public center.
recall EllipsoidMethod.localizationSets_subset_associatedEllipsoidTheorem 3.2.5
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.2.5 is the earlier chapter theorem exists_stronglyConvexLipschitzProblem_with_span_method_lower_bound.
recall exists_stronglyConvexLipschitzProblem_with_span_method_lower_bound
Lemma 3.2.6
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.2.6: let g be a direction in a finite-dimensional real inner-product space, and define S₊ = S ∩ cuttingHalfspace (⨍ x in S, x) g = {x ∈ S | ⟪g, (⨍ x in S, x) - x⟫_ℝ ≥ 0}. Then vol S₊ / vol S ≤ 1 - 1 / e for convex S of finite positive volume.
open MeasureTheory
universe u v
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable [FiniteDimensional ℝ E]
theorem centerOfGravityCut_volumeRatio_le_one_sub_inv_e
(S : Set E) (g : E) (hS_convex : Convex ℝ S) (hS_finite : volume S ≠ ⊤)
(hS_pos : volume S ≠ 0) (hg : g ≠ 0) :
(volume (S ∩ cuttingHalfspace (⨍ x in S, x) g)).toReal / (volume S).toReal ≤
1 - 1 / Real.exp 1
Theorem 3.2.6
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.2.6: for the projected subgradient iteration x_{k+1} = π_Q (x_k - (2 ε / ‖g_k‖²) • g_k) on a μ-strongly convex objective over the convex feasible set Q in a real inner-product space, if every chosen relative subgradient g_k ∈ ∂[Q] f(x_k) has norm at most M and the iteration budget N satisfies N ≥ (M² / (μ ε)) log (M ‖xSeq 0 - x*‖ / ε), then the sampled best value bestFunctionValueUpTo (fun i ↦ f (xSeq i)) N is at most ε above the optimal value f xStar. The minimizer hypothesis is paired with the explicit feasibility assumption xStar ∈ Q; g_k ∈ ∂[Q] f(x_k) already forces x_k ∈ Q, so no separate feasibility assumption on x₀ is needed in the public API. The logarithmic budget is stated directly in terms of the canonically determined initial iterate xSeq 0.
open scoped WithTopConvexAnalysis
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
theorem bestFunctionValueUpTo_le_optimalValue_add_eps_of_projected_subgradient_log_budget_strongConvexOn
{Q : Set E} (projQ : E → E)
(hproj : ∀ x : E, IsProjectionPointOn Q x (projQ x))
{f : E → ℝ} {μ M ε : ℝ} (hμ : 0 < μ) (hε : 0 < ε)
(hf : StrongConvexOn Q μ f)
{xStar : E} (hxStar : IsMinOn f Q xStar)
(hxStar_mem : xStar ∈ Q)
(xSeq g : ℕ → E)
(hsubgrad : ∀ k : ℕ, g k ∈ ∂[Q] f((xSeq k)))
(hsubgrad_norm : ∀ k : ℕ, ‖g k‖ ≤ M)
(hxSeq_succ :
∀ k : ℕ,
xSeq (k + 1) =
projQ (xSeq k - ((2 * ε) / ‖g k‖ ^ (2 : ℕ)) • g k))
{N : ℕ}
(hN :
(N : ℝ) ≥ (M ^ (2 : ℕ)) / (μ * ε) * Real.log (M * ‖xSeq 0 - xStar‖ / ε)) :
bestFunctionValueUpTo (fun i ↦ f (xSeq i)) N ≤ f xStar + εLemma 3.2.7
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.2.7: for a positive-definite ellipsoid E(H, x̄) and a nonzero cutting direction g, the center cut E₊ is contained in the updated ellipsoid E(H₊, x̄₊), and the Lebesgue volume of the updated ellipsoid is at most (1 - 1 / (n + 1)^2)^(n / 2) times the original volume.
open MeasureTheory Matrix
open scoped RealInnerProductSpace MatrixOrder MatrixPosDef
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
local notation "Mat" => Matrix (Fin n) (Fin n) ℝ
open scoped EllipsoidNotation
theorem centerCutEllipsoid_subset_updatedEllipsoid_and_volume_le
(H : Mat) (hH : H.PosDef) (xBar g : E) (hg : g ≠ 0) (hn : 1 < n) :
E₊(H, xBar, g) ⊆ E(H₊(H, g), x̄₊(H, xBar, g)) ∧
((volume (E(H₊(H, g), x̄₊(H, xBar, g)))).toReal ≤
Real.rpow (1 - 1 / (((n : ℝ) + 1) ^ (2 : ℕ))) ((n : ℝ) / 2) *
(volume (E(H, xBar))).toReal)Theorem 3.2.7
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.2.7 is the chapter owner theorem exists_hard_feasibility_problem_for_short_separation_oracle_algorithms.
recall exists_hard_feasibility_problem_for_short_separation_oracle_algorithms
Theorem 3.2.8
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.2.8 is the chapter owner theorem value_oracle_query_lower_bound_of_uniform_epsilon_guarantee.
recall value_oracle_query_lower_bound_of_uniform_epsilon_guarantee
Theorem 3.2.9
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.2.9 at the intrinsic owner level: if 0 < Module.finrank ℝ E, if the stage set localizationSet Q xSeq (g ∘ xSeq) k is measurable, and if the best localization radius is already known to stay inside the outer ball of radius D, then that radius is bounded by D times the Module.finrank ℝ E-th root of the Haar-measure ratio μ(S_k(X)) / μ(Q) for μ = Measure.addHaar. Specializing to E = EuclideanSpace ℝ (Fin n) with 0 < n recovers the textbook ℝⁿ statement, since the normalization factor cancels in the ratio.
open MeasureTheory
open scoped SubgradientLocalizationMeasure
universe u
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable [FiniteDimensional ℝ E] [MeasurableSpace E] [BorelSpace E]
local notation "μ" => (Measure.addHaar : Measure E)
theorem localization_radius_le_outer_radius_mul_volume_ratio_rpow
(hdim : 0 < Module.finrank ℝ E)
{Q : Set E} {xStar : E} {g : E → E} {xSeq : ℕ → E} {k : ℕ} {D : ℝ}
(hQ_convex : Convex ℝ Q) (hxStar : xStar ∈ Q)
(hQ_pos : 0 < Measure.real μ Q)
(hQ_subset : Q ⊆ Metric.closedBall xStar D)
(hSk_meas : MeasurableSet (localizationSet Q xSeq (g ∘ xSeq) k))
(hloc_le_D : localization_radius xStar g xSeq k ≤ D) :
localization_radius xStar g xSeq k ≤
D *
Real.rpow
(Measure.real μ (localizationSet Q xSeq (g ∘ xSeq) k) / Measure.real μ Q)
(1 / (Module.finrank ℝ E : ℝ))Theorem 3.2.10
— source only (outside the published documentation root) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.2.10: at the chapter owner level, for the center-of-gravity cutting-plane method on a bounded convex minimization problem, if xStar minimizes the objective on the feasible set Q = problem.feasibleSet, if xStar ∈ Q, if Q ⊆ B₂(xStar, D), and if the objective is M-Lipschitz on that ball, then the best sampled value f_k^* = bestFunctionValueUpTo (fun i ↦ problem (iterate problem i)) k along the centroid iterates satisfies f_k^* - problem xStar ≤ M D (1 - 1 / e)^(k / dim), where dim = Module.finrank ℝ E. The factor (1 - 1 / e)^(k / dim) is derived from the centroid-cut volume contraction and the localization-radius/volume comparison, not assumed as primitive data.
universe u
open MeasureTheory
open scoped SubgradientLocalizationMeasure
open CenterOfGravityMethod
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
variable [FiniteDimensional ℝ E]
local notation "dim" => Module.finrank ℝ E
theorem CenterOfGravityMethod.bestFunctionValueUpTo_sub_optimalValue_le_lipschitz_geometricDecay
(problem : ConvexMinimizationWithSeparationOracle E)
{xStar : E} {D : ℝ} {M : NNReal}
(hxStar : IsMinOn problem problem.feasibleSet xStar)
(hxStar_mem : xStar ∈ problem.feasibleSet)
(hQ_subset : problem.feasibleSet ⊆ Metric.closedBall xStar D)
(hLip : LipschitzOnWith M problem (Metric.closedBall xStar D))
(k : ℕ) :
bestFunctionValueUpTo (fun i ↦ problem (iterate problem i)) k - problem xStar ≤
(M : ℝ) * D * Real.rpow (1 - 1 / Real.exp 1) ((k : ℝ) / dim)
Theorem 3.2.11
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.2.11: if xSeq is the feasible subsequence selected by the cutting-plane method, f is M-Lipschitz on B₂(xStar, R), xStar realizes the optimal value f^* = f(xStar) on that ball, and the selected feasible iterate xSeq (i k) satisfies the standard geometric distance estimate, then the best feasible objective value bestFunctionValueUpTo (fun j ↦ f (xSeq j)) (i k) among the first i(k) + 1 selected points satisfies the same decay estimate relative to f^*.
open MeasureTheory
variable {n : ℕ}
local notation "E" => EuclideanSpace ℝ (Fin n)
theorem selected_feasible_bestValueGap_le_lipschitz_radius_mul_geometricDecay_volumeRatio
{f : E → ℝ} {xStar x0 : E} {R : ℝ} {M : NNReal} {Q : Set E}
{xSeq : ℕ → E} {i : ℕ → ℕ} (k : ℕ)
(hxStar_opt : IsMinOn f (Metric.closedBall xStar R) xStar)
(hf_lipschitz : LipschitzOnWith M f (Metric.closedBall xStar R))
(hselected_mem : xSeq (i k) ∈ Metric.closedBall xStar R)
(hselected_dist :
‖xSeq (i k) - xStar‖ ≤
R *
Real.rpow (1 - 1 / (((n : ℝ) + 1) ^ (2 : ℕ))) ((k : ℝ) / 2) *
Real.rpow
((volume (Metric.closedBall x0 R)).toReal / (volume Q).toReal)
(1 / (n : ℝ))) :
bestFunctionValueUpTo (fun j ↦ f (xSeq j)) (i k) - f xStar ≤
(M : ℝ) * R *
Real.rpow (1 - 1 / (((n : ℝ) + 1) ^ (2 : ℕ))) ((k : ℝ) / 2) *
Real.rpow
((volume (Metric.closedBall x0 R)).toReal / (volume Q).toReal)
(1 / (n : ℝ))3.3 — Methods with Complete Data
Lemma 3.3.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.3.1: if the record values satisfy f_{j+1}^* ≤ f_j^*, the model gaps satisfy δ_{j+1} ≤ δ_j, and the gap still obeys δ_p ≥ (1 - α) δ_k, then every intermediate level value ℓ_i(α) with k ≤ i ≤ p is at least \hat f_p^*.
universe u
open scoped LevelMethodNotation
namespace LevelMethodHistory
theorem levelValue_ge_approximateOptimalValue_of_gap_large_enough
(history : LevelMethodHistory) {α : ℝ} {k p : ℕ}
(hα : α ≤ 1)
(hoptimal_succ : ∀ j : ℕ, history.optimalValue (j + 1) ≤ history.optimalValue j)
(hgap_succ : ∀ j : ℕ, history.gap (j + 1) ≤ history.gap j)
(hgap_large : δ[history](p) ≥ (1 - α) * δ[history](k)) :
∀ {i : ℕ}, k ≤ i → i ≤ p → ℓ[history](α, i) ≥ fhat(history, p)Theorem 3.3.1
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.3.1 (iteration-cap part): if each level-method block [k, p] that is still in the positive-gap regime and satisfies δ_p ≥ (1 - α) δ_k obeys the interval bound from Lemma 3.3.3, then the Level Method reaches the canonical stopping predicate δ_k ≤ ε after at most ⌊M_f^2 D^2 / (ε^2 α (1 - α)^2 (2 - α))⌋ + 1 iterations.
open scoped LevelMethodNotation
theorem exists_stopping_index_le_levelMethodIterationCap
(history : LevelMethodHistory) {M_f D ε α : ℝ}
(hε : 0 < ε)
(hα : α ∈ Set.Ioo (0 : ℝ) 1)
(hblock :
∀ {k p : ℕ}, k ≤ p →
δ[history](p) ≥ (1 - α) * δ[history](k) →
0 < δ[history](p) →
((p + 1 - k : ℕ) : ℝ) ≤
(M_f ^ (2 : ℕ) * D ^ (2 : ℕ)) /
((1 - α) ^ (2 : ℕ) * δ[history](p) ^ (2 : ℕ))) :
∃ k ≤ levelMethodIterationCap M_f D ε α,
history.shouldStop ε kLemma 3.3.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.3.2: if the level value at iteration k dominates the sampled model at x_{k+1}, the record value satisfies f_k^* ≤ f(x_k), and the selected subgradient norm is bounded by M_f, then the step length satisfies ‖x_{k+1} - x_k‖ ≥ ((1 - α) δ_k) / M_f.
universe u
variable {E : Type u} [SeminormedAddCommGroup E] [InnerProductSpace ℝ E]
open scoped NonsmoothModelNotation
lemma step_norm_lower_bound_of_level_method_assumptions
{f : E → ℝ} {X g : ℕ → E} {history : LevelMethodHistory}
(α Mf : ℝ) {k : ℕ}
(hrecord_le_current : history.optimalValue k ≤ f (X k))
(hlevel_ge_model :
history.levelValue α k ≥
f̂[X; f; g](k) (X (k + 1)))
(hsubgradient_bound : ‖g k‖ ≤ Mf)
(hMf_pos : 0 < Mf) :
‖X (k + 1) - X k‖ ≥
((1 - α) * history.gap k) / MfTheorem 3.3.2
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.3.2: in view of Lemma 3.3.7, the master process reaches the global-stop threshold, and hence the estimate (3.3.9), after at most log ((t₀ - t^*) / ((1 - ε) ε)) / log (2 (1 - ε)) full iterations.
open HasGeometricRateOfConvergence
universe u v
variable {χ : Type u} {ι : Type v}
theorem selected_exactValue_le_epsilon_at_natCeil_masterIterationCountBound
{ε tStar : ℝ} {X : χ} {t : ℕ → ℝ} {j : ℕ → ι}
{exactValue : ι → χ → ℝ → ℝ}
(hε : 0 < ε)
(hε_contract : 1 < 2 * (1 - ε))
(hgeom :
∀ k : ℕ,
exactValue (j k) X (t k) ≤
((t 0 - tStar) / (1 - ε)) * ((1 / (2 * (1 - ε))) ^ k)) :
exactValue
(j ⌈Real.log ((t 0 - tStar) / ((1 - ε) * ε)) / Real.log (2 * (1 - ε))⌉₊)
X
(t ⌈Real.log ((t 0 - tStar) / ((1 - ε) * ε)) / Real.log (2 * (1 - ε))⌉₊) ≤
εLemma 3.3.3
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.3.3: let Q be the bounded feasible set from problem (3.3.1) with diam Q ≤ D. If the gap satisfies δ_p ≥ (1 - α) δ_k, and if the iterates are generated by projections onto the level sets 𝓛_i(α) = {x ∈ Q | \hat f_i(X; x) ≤ (1 - α)\hat f_i^* + α f_i^*} with the step-size lower bound coming from Lemma 3.3.2, then p + 1 - k ≤ M_f^2 D^2 / ((1 - α)^2 δ_p^2).
universe u
open scoped LevelMethodNotation
variable {E : Type u} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
theorem iteration_count_bound_of_bounded_feasible_set
(history : LevelMethodHistory) (Q : Set E) (model : ℕ → E → ℝ)
(x : ℕ → E) {α Mf D : ℝ} {k p : ℕ} {xStar : E}
(hα : α < 1)
(hoptimal_mono :
∀ ⦃i : ℕ⦄, k ≤ i → i ≤ p → fstar(history, p) ≤ fstar(history, i))
(hgap_antitone :
∀ ⦃i j : ℕ⦄, k ≤ i → i ≤ j → j ≤ p → δ[history](j) ≤ δ[history](i))
(hgap_large : δ[history](p) ≥ (1 - α) * δ[history](k))
(hgap_pos : 0 < δ[history](p))
(hMf_pos : 0 < Mf)
(hQ_bounded : Bornology.IsBounded Q)
(hdiam : Metric.diam Q ≤ D)
(hxk_mem : x k ∈ Q)
(hxStar_mem : xStar ∈ Q)
(hxStar_terminal : model p xStar = fhat(history, p))
(hmodel_le_terminal :
∀ ⦃i : ℕ⦄, k ≤ i → i ≤ p → model i xStar ≤ model p xStar)
(hconvex :
∀ ⦃i : ℕ⦄, k ≤ i → i ≤ p →
Convex ℝ (𝓛[Q, model, history](α, i)))
(hprojection :
∀ ⦃i : ℕ⦄, k ≤ i → i ≤ p →
IsProjectionPointOn (𝓛[Q, model, history](α, i)) (x i) (x (i + 1)))
(hstep_lower :
∀ ⦃i : ℕ⦄, k ≤ i → i ≤ p →
‖x (i + 1) - x i‖ ≥
((1 - α) * δ[history](i)) / Mf) :
((p + 1 - k : ℕ) : ℝ) ≤
(Mf ^ (2 : ℕ) * D ^ (2 : ℕ)) /
((1 - α) ^ (2 : ℕ) * δ[history](p) ^ (2 : ℕ))Theorem 3.3.3
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Theorem 3.3.3: if a constrained level method has a globally stopping master step, if each preceding full master step before the first globally stopping one is bounded by the canonical one-run cap levelMethodIterationCap M_f D (χ ε) α, if the terminal globally stopping master step satisfies the raw bound constrainedLevelMethodInternalIterationBound M_f D χ ε α, and if the number of preceding full master steps is bounded by the natural ceiling of the logarithmic threshold from Theorem 3.3.2, then the total number of internal iterations executed up to that first globally stopping step is bounded by (N(ε) + 1) copies of the canonical cap levelMethodIterationCap M_f D (χ ε) α.
universe u
variable {E : Type u} [PseudoMetricSpace E]
open ConstrainedLevelMethod
open scoped BigOperators
theorem constrained_level_total_internal_iterations_le
(method : ConstrainedLevelMethodInput E)
(hrelative : method.RelativeStoppingExists)
(hfinite : method.SelectedThresholdFinite hrelative)
{k : ℕ}
(hfirst : IsLeast {i : ℕ | globallyStopsAt method hrelative hfinite i} k)
{M_f D tStar : ℝ}
(houter :
k ≤
⌈Real.log
((method.initialParameter - tStar) /
((1 - method.chi) * method.epsilon)) /
Real.log (2 * (1 - method.chi))⌉₊)
(hfull_internal :
∀ i < k,
stoppingIndex method hrelative hfinite i ≤
levelMethodIterationCap
M_f D (method.chi * method.epsilon) method.levelCoefficient)
(hterminal_internal :
(globalStopIndex method hrelative hfinite k hfirst.1 : ℝ) ≤
constrainedLevelMethodInternalIterationBound
M_f D method.chi method.epsilon method.levelCoefficient) :
(∑ i ∈ Finset.range k, (stoppingIndex method hrelative hfinite i : ℝ)) +
(globalStopIndex method hrelative hfinite k hfirst.1 : ℝ) ≤
(((⌈Real.log
((method.initialParameter - tStar) /
((1 - method.chi) * method.epsilon)) /
Real.log (2 * (1 - method.chi))⌉₊ : ℕ) : ℝ) + 1) *
levelMethodIterationCap
M_f D (method.chi * method.epsilon) method.levelCoefficientLemma 3.3.4
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.3.4: if xStar attains the minimum of hatFn k X on the feasible set Q ∩ {x | checkFn k X x ≤ 0}, then t_k^*(X) equals that constrained minimum value.
universe u v w
open scoped ConstrainedArgmin
variable {Index : Type u} {Param : Type v} {Decision : Type w}
open scoped ConstrainedThreshold
theorem constrainedThreshold_eq_minimum_of_feasible_minimizer
(Q : Set Decision) (hatFn checkFn : Index → Param → Decision → ℝ) (k : Index)
(X : Param) {xStar : Decision}
(hxStar : xStar ∈ argmin[Q ∩ {x | checkFn k X x ≤ 0}] (hatFn k X)) :
t*[Q; hatFn; checkFn](k, X) = (hatFn k X xStar : EReal)Lemma 3.3.5
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.3.5: if the finite real-part view of the complete-data owner value parametricValueFunction Q (hatFn k X) (checkFn k X) t₁ is positive for some t₀ < t₁ ≤ τ, the owner value at τ is finite and nonpositive, and the scalar slice is convex on (-∞, τ], then τ lies strictly to the right of t₁ and the displayed secant lower bound holds.
open scoped ConvexAnalysis
universe u v w
variable {Index : Type u} {Param : Type v} {Decision : Type w}
variable (Q : Set Decision) (hatFn checkFn : Index → Param → Decision → ℝ)
variable (k : Index) (X : Param)
local notation "valueFn" =>
parametricValueFunction Q (hatFn k X) (checkFn k X)
local notation "modelValue" =>
extendedRealRealPart valueFn
theorem parametricValueFunction_strict_lt_right_and_secant_lower_bound
{t0 t1 τ : ℝ}
(ht0_lt_t1 : t0 < t1)
(ht1_le_right : t1 ≤ τ)
(hpos : 0 < modelValue t1)
(hτ_dom : τ ∈ dom valueFn)
(hright_nonpos : valueFn τ ≤ (0 : EReal))
(hconvex : ConvexOn ℝ (Set.Iic τ) modelValue) :
t1 < τ ∧
modelValue t0 ≥
modelValue t1 + ((t1 - t0) / (τ - t1)) * modelValue t1Lemma 3.3.6
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.3.6: for any model f and any Δ ≥ 0, shifting the parameter from t to t + Δ decreases the chapter parametric value by at most Δ. The textbook exact value f^*(t) and the approximate value \hat f_k^*(X; t) are the corresponding specializations of this owner theorem.
universe u
variable {α : Type u}
theorem parametricValueFunction_sub_le_shift
(Q : Set α) (f barf : α → ℝ) (Δ t : ℝ) (hΔ : 0 ≤ Δ) :
parametricValueFunction Q f barf t - Δ ≤
parametricValueFunction Q f barf (t + Δ)Lemma 3.3.7
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.3.7 in canonical owner form: under the complete-data level-method hypotheses with ε < 1 and the three comparison inequalities between f_{j(k)}^*(X; t_k) and \hat f_{j(k)}^*(X; t_k), if the initial approximate value is bounded by the initial gap, then the selected exact-value sequence has geometric rate with parameter 1 - (2 * (1 - ε))⁻¹ and initial constant (t 0 - tStar) / (1 - ε). The explicit threshold recursion and step-size bounds from the source are redundant for this scalar consequence, and the stronger contractivity side condition belongs only to later iteration-threshold consequences.
open HasGeometricRateOfConvergence
universe u v
variable {χ : Type u} {ι : Type v}
theorem selectedExactValue_hasGeometricRateOfConvergence
{ε tStar : ℝ} {X : χ} {t : ℕ → ℝ} {j : ℕ → ι}
{exactValue estimatedValue : ι → χ → ℝ → ℝ}
(hε : ε < 1)
(hestimated_ge :
∀ k : ℕ,
estimatedValue (j k) X (t k) ≥ (1 - ε) * exactValue (j k) X (t k))
(hestimated_prev_ge_two_curr :
∀ k : ℕ,
estimatedValue (j (k + 1)) X (t k) ≥
2 * estimatedValue (j (k + 1)) X (t (k + 1)))
(hexact_prev_ge_estimated_prev :
∀ k : ℕ,
exactValue (j k) X (t k) ≥ estimatedValue (j (k + 1)) X (t k))
(hestimated0_le_gap :
estimatedValue (j 0) X (t 0) ≤ t 0 - tStar) :
HasGeometricRateOfConvergence
(fun k ↦ exactValue (j k) X (t k))
(1 - (2 * (1 - ε))⁻¹)
((t 0 - tStar) / (1 - ε))Lemma 3.3.8
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.3.8: once the global-stop condition holds at master step k, some produced internal iterate attains the stopping record value and therefore satisfies the displayed objective and constraint bounds.
universe u
variable {E : Type u} [PseudoMetricSpace E]
namespace ConstrainedLevelMethod
theorem global_stop_exists_internal_witness_and_component_bounds
(method : ConstrainedLevelMethodInput E)
(hrelative : method.RelativeStoppingExists)
(hfinite : method.SelectedThresholdFinite hrelative)
(k : ℕ)
(hstop : globallyStopsAt method hrelative hfinite k)
(tStar : ℝ)
(htk : parameter method hrelative hfinite k ≤ tStar) :
∃ jStar : ℕ,
jStar ≤ globalStopIndex method hrelative hfinite k hstop ∧
method.exactObjectiveAt (parameter method hrelative hfinite k)
(internalIterate method hrelative hfinite k jStar) =
(history method hrelative hfinite k).optimalValue
(globalStopIndex method hrelative hfinite k hstop) ∧
method.exactObjectiveAt (parameter method hrelative hfinite k)
(internalIterate method hrelative hfinite k jStar) ≤
method.epsilon ∧
method.objective (internalIterate method hrelative hfinite k jStar) ≤
tStar + method.epsilon ∧
method.constraint (internalIterate method hrelative hfinite k jStar) ≤
method.epsilonLemma 3.3.9
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
Lemma 3.3.9 (1): if master step k is a full step in the sense that the selected exact record value is still at least ε, and if the chapter owner hypotheses of Theorem 3.3.1 hold for the inner history method.history k, then the canonical full-step count j(k) - j(k - 1) represented by method.stoppingIndex k is bounded by the chapter owner levelMethodIterationCap evaluated at the uniform tolerance χ ε.
universe u
variable {E : Type u} [PseudoMetricSpace E]
namespace ConstrainedLevelMethod
theorem full_step_increment_le_uniform_internal_iteration_bound
(method : ConstrainedLevelMethodInput E)
(hrelative : method.RelativeStoppingExists)
(hfinite : method.SelectedThresholdFinite hrelative)
(k : ℕ) (M_f D : ℝ)
(hχ : 0 < method.chi)
(hε : 0 < method.epsilon)
(hα : method.levelCoefficient ∈ Set.Ioo (0 : ℝ) 1)
(hblock :
∀ {i p : ℕ}, i ≤ p →
(history method hrelative hfinite k).gap p ≥
(1 - method.levelCoefficient) * (history method hrelative hfinite k).gap i →
0 < (history method hrelative hfinite k).gap p →
((p + 1 - i : ℕ) : ℝ) ≤
M_f ^ (2 : ℕ) * D ^ (2 : ℕ) /
((1 - method.levelCoefficient) ^ (2 : ℕ) *
(history method hrelative hfinite k).gap p ^ (2 : ℕ)))
(hfull_step :
method.epsilon ≤
(history method hrelative hfinite k).optimalValue
(stoppingIndex method hrelative hfinite k)) :
stoppingIndex method hrelative hfinite k ≤
levelMethodIterationCap
M_f D (method.chi * method.epsilon) method.levelCoefficientSupporting modules
These source modules support the indexed items and are listed separately so the chapter inventory remains complete.
-
Lemma_3_2_6/Coordinate.lean(Documentation unavailable) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
-
Lemma_3_2_6/Core.lean(Documentation unavailable) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
-
Lemma_3_2_6/Profile.lean(Documentation unavailable) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
-
Lemma_3_2_6/RealLine.lean(Documentation unavailable) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
-
Lemma_3_2_6/Slice.lean(Documentation unavailable) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
-
LinearEqualityFeasibleSet.lean(Documentation) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.
-
PointwiseSupremumOn.lean(Documentation) (Lean source)
Lean contributors (project-level attribution): Chenyi Li, Siyuan Shao, Yijie Wang, Feiming Wang, Weiran Shi, Yuhao Jiang, Zebo Liu, Wentao Long.