Chapter 09 — Lower Semicontinuous Convex Functions

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

Theorem 9.1

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

Source header · Attribution source

Theorem 9.1: for a convex extended-real-valued function on a real Hilbert space, weak sequential lower semicontinuity, strong sequential lower semicontinuity, strong lower semicontinuity, and weak lower semicontinuity are equivalent. The weak conditions are expressed through toWeakSpace ℝ H and WeaklyLowerSemicontinuous.

open Filter open scoped InnerProductSpace Topology universe u namespace ERealFunction theorem convex_lowerSemicontinuity_tfae {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] {f : H → EReal} (hconv : Convex ℝ (epigraph f)) : List.TFAE [ (∀ ⦃xₙ : ℕ → H⦄ ⦃x : H⦄, Tendsto (fun n ↦ toWeakSpace ℝ H (xₙ n)) atTop (𝓝 (toWeakSpace ℝ H x)) → f x ≤ liminf (f ∘ xₙ) atTop), (∀ ⦃xₙ : ℕ → H⦄ ⦃x : H⦄, Tendsto xₙ atTop (𝓝 x) → f x ≤ liminf (f ∘ xₙ) atTop), LowerSemicontinuous f, WeaklyLowerSemicontinuous f ]

Definition 9.2

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

Source header · Attribution source

Definition 9.2: gamma H is the textbook class Γ(ℋ) of extended-real-valued functions on a real sequential topological module that are convex and lower semicontinuous.

open Filter universe u namespace ERealFunction variable {H : Type u} def gamma (H : Type u) [TopologicalSpace H] [SequentialSpace H] [AddCommMonoid H] [Module ℝ H] : Set (H → EReal) := {f | IsConvex f ∧ LowerSemicontinuous f}

Proposition 9.3

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

Source header · Attribution source

Proposition 9.3: the pointwise supremum of a family of functions in Γ(ℋ) again belongs to Γ(ℋ).

universe u v namespace ERealFunction variable {H : Type u} [TopologicalSpace H] [SequentialSpace H] [AddCommGroup H] [Module ℝ H] variable {I : Type v} theorem iSup_mem_gamma (f : I → H → EReal) (hf : ∀ i, f i ∈ gamma H) : (fun x ↦ ⨆ i, f i x) ∈ gamma H

Corollary 9.4

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

Source header · Attribution source

Corollary 9.4: if (fᵢ)ᵢ is a family in Γ(ℋ), then the source-defined pointwise sum familySum f again belongs to Γ(ℋ) whenever either (i) I is finite and no fᵢ attains -∞, or (ii) every fᵢ is pointwise nonnegative.

open scoped BigOperators universe u v namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] variable {I : Type v} theorem familySum_mem_gamma (f : I → H → EReal) (hf : ∀ i, f i ∈ gamma H) (hcases : (Finite I ∧ ∀ i x, f i x ≠ ⊥) ∨ ∀ i x, (0 : EReal) ≤ f i x) : familySum f ∈ gamma H

Proposition 9.5

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

Source header · Attribution source

Proposition 9.5: if f is proper, lower semicontinuous, and convex on K, and dom f meets the range of a continuous linear map L : H →L[ℝ] K, then f ∘ L is proper, lower semicontinuous, and convex on H.

universe u v namespace ERealFunction variable {H : Type u} {K : Type v} variable [NormedAddCommGroup H] [NormedSpace ℝ H] variable [NormedAddCommGroup K] [NormedSpace ℝ K] theorem isProper_and_mem_gamma_comp_continuousLinearMap_of_dom_inter_range_nonempty (f : K → EReal) (L : H →L[ℝ] K) (hf_proper : IsProper f) (hf_gamma : f ∈ gamma K) (hdom : (dom f ∩ Set.range L).Nonempty) : IsProper (f ∘ L) ∧ f ∘ L ∈ gamma H

Proposition 9.6

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

Source header · Attribution source

Proposition 9.6: if a function in Γ(ℋ) takes the value -∞ at some point, then it is nowhere real-valued: every value is either -∞ or +∞.

open Filter universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [NormedSpace ℝ H] theorem eq_bot_or_eq_top_of_mem_gamma_of_eq_bot {f : H → EReal} (hf : f ∈ gamma H) {x : H} (hx : f x = ⊥) (y : H) : f y = ⊥ ∨ f y = ⊤

Definition 9.7

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

Source header · Attribution source

Definition 9.7: the lower semicontinuous convex envelope of f is the supremum, in the function lattice, of all lower semicontinuous convex minorants of f.

universe u namespace ERealFunction variable {H : Type u} [TopologicalSpace H] [AddCommGroup H] [Module ℝ H] noncomputable def lowerSemicontinuousConvexEnvelope (f : H → EReal) : H → EReal := sSup (lowerSemicontinuousConvexMinorants f)

Proposition 9.8

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

Source header · Attribution source

Proposition 9.8 (1): part (i). The lower semicontinuous convex envelope is pointwise majorized by f.

open Set universe u namespace ERealFunction variable {H : Type u} [TopologicalSpace H] [AddCommGroup H] [Module ℝ H] [IsTopologicalAddGroup H] [ContinuousConstSMul ℝ H] theorem lowerSemicontinuousConvexEnvelope_le (f : H → EReal) : lowerSemicontinuousConvexEnvelope f ≤ f

Theorem 9.9

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

Source header · Attribution source

Theorem 9.9: the epigraph of the lower semicontinuous convex envelope of f is the closure of the convex hull of the epigraph of f.

open Set open scoped InnerProductSpace open WithLp universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] theorem epigraph_lowerSemicontinuousConvexEnvelope_eq_closure_convexHull_epigraph (f : H → EReal) : epigraph (lowerSemicontinuousConvexEnvelope f) = closure (convexHull ℝ (epigraph f))

Corollary 9.10

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

Source header · Attribution source

Corollary 9.10: for a convex extended-real-valued function, the lower semicontinuous convex envelope \bar f agrees with the lower semicontinuous envelope \check f.

open Set universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] theorem lowerSemicontinuousConvexEnvelope_eq_lowerSemicontinuousEnvelope_of_convex_epigraph (f : H → EReal) (hconv : Convex ℝ (epigraph f)) : lowerSemicontinuousConvexEnvelope f = lowerSemicontinuousEnvelope f

Corollary 9.11

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

Source header · Attribution source

Corollary 9.11 (1): the strict negative level set of f is contained in its nonpositive level set.

open Set universe u namespace ERealFunction variable {H : Type u} -- Proof sketch: this is the pointwise implication `f x < 0 → f x ≤ 0`, rewritten in terms of the -- strict and closed lower level sets at height `0`. theorem strictLowerLevelSet_zero_subset_lowerLevelSet_zero (f : H → EReal) : strictLowerLevelSet f 0 ⊆ lowerLevelSet f 0

Definition 9.12

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

Source header · Attribution source

Definition 9.12: Γ₀(H) is the set of lower semicontinuous convex ]-∞,+∞]-valued functions on H, with convexity imposed on the effective domain so that properness is built into the definition.

universe u namespace ERealFunction variable {H : Type u} [TopologicalSpace H] [AddCommGroup H] [Module ℝ H] def gammaZero (H : Type u) [TopologicalSpace H] [AddCommGroup H] [Module ℝ H] : Set (H → Set.Ioi (⊥ : EReal)) := {f | LowerSemicontinuous (fun x : H ↦ (f x : EReal)) ∧ ConvexOn f (effectiveDomain f)}

Example 9.13

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

Source header · Attribution source

Example 9.13: if each φ i belongs to Γ₀(ℝ) and satisfies φ i ≥ φ i(0) = 0, then the function obtained by summing the coordinate values φ i ⟪x, e i⟫_ℝ belongs to Γ₀(H).

open scoped BigOperators InnerProductSpace universe u v namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] variable {I : Type v} theorem innerProductSeriesFunction_mem_gammaZero (e : I → H) (φ : I → ℝ → Set.Ioi (⊥ : EReal)) (hφ : ∀ i, φ i ∈ Γ₀(ℝ)) (hφ_zero : ∀ i, (φ i 0 : EReal) = 0) (hφ_nonneg : ∀ i t, (φ i 0 : EReal) ≤ (φ i t : EReal)) : innerProductSeriesFunction e φ hφ_zero hφ_nonneg ∈ Γ₀(H)

Proposition 9.14

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

Source header · Attribution source

Proposition 9.14: if f is lower semicontinuous and has convex epigraph, and if x and y lie in the effective domain of f, then the values of f along the segment point x_α = (1 - α) • x + α • y = AffineMap.lineMap x y α converge to f x as α ↓ 0.

open scoped Topology universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [NormedSpace ℝ H] -- Proof sketch: combine the convexity inequality along the segment from `x` to `y`, which bounds -- the right limsup by `f x`, with lower semicontinuity of `f` and the fact that -- `AffineMap.lineMap x y α → x` as `α ↓ 0` to obtain the matching lower liminf inequality. theorem tendsto_apply_lineMap_right_zero_of_lowerSemicontinuous_epigraph_convex {f : H → Set.Ioi (⊥ : EReal)} (hf_lsc : LowerSemicontinuous (fun z : H ↦ (f z : EReal))) (hf_conv : Convex ℝ {p : H × ℝ | (f p.1 : EReal) ≤ (p.2 : EReal)}) {x y : H} (hx : (f x : EReal) < ⊤) (hy : (f y : EReal) < ⊤) : Filter.Tendsto (fun α : ℝ ↦ (f (AffineMap.lineMap x y α) : EReal)) (𝓝[>] (0 : ℝ)) (𝓝 (f x : EReal))

Corollary 9.15

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

Source header · Attribution source

Corollary 9.15: if f ∈ Γ₀(ℝ), then the real-valued representative of f is continuous on its effective domain; equivalently, the restriction f|_{dom f} : dom f → ℝ is continuous.

open Set open scoped Topology namespace ERealFunction theorem continuousOn_toReal_effectiveDomain_of_mem_gammaZero {f : ℝ → Set.Ioi (⊥ : EReal)} (hf : f ∈ Γ₀(ℝ)) : ContinuousOn (fun x : ℝ ↦ (f x : EReal).toReal) (effectiveDomain f)

Proposition 9.16

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

Source header · Attribution source

Proposition 9.16: if f and g belong to Γ₀(H × ℝ), their effective domains are contained in the closed upper half-space, they agree on the open upper half-space, and their common restriction there is proper, then f = g on all of H × ℝ.

open Set open scoped Topology universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [NormedSpace ℝ H] theorem eq_of_eqOn_openUpperHalfSpace_of_effectiveDomain_union_subset_closedUpperHalfSpace {f g : H × ℝ → Set.Ioi (⊥ : EReal)} (hf : f ∈ Γ₀(H × ℝ)) (hg : g ∈ Γ₀(H × ℝ)) (hdom : effectiveDomain f ∪ effectiveDomain g ⊆ Set.univ ×ˢ Set.Ici (0 : ℝ)) (heq : Set.EqOn f g (Set.univ ×ˢ Set.Ioi (0 : ℝ))) (hproper : (effectiveDomain f ∩ (Set.univ ×ˢ Set.Ioi (0 : ℝ))).Nonempty) : f = g

Fact 9.17

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

Source header · Attribution source

Fact 9.17: for f, g ∈ Γ₀(H) on a real Hilbert space, the interior of the difference of their effective domains coincides with the core of that difference.

open Set Topology open scoped Pointwise universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] theorem interior_sub_effectiveDomain_eq_core_sub_effectiveDomain {f g : H → Set.Ioi (⊥ : EReal)} (hf : f ∈ Γ₀(H)) (hg : g ∈ Γ₀(H)) : interior (effectiveDomain f - effectiveDomain g) = Set.core (effectiveDomain f - effectiveDomain g)

Proposition 9.18

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

Source header · Attribution source

Proposition 9.18: for f ∈ Γ₀(H), a pair (p, π) is the metric projection of (x, ξ) onto the real-height epigraph of f if and only if π majorizes both ξ and f p, and the resulting variational inequality holds against every point of effectiveDomain f.

open scoped InnerProductSpace open WithLp universe u v namespace ERealFunction variable {H : Type u} {K : Type v} variable [NormedAddCommGroup H] [InnerProductSpace ℝ H] variable [NormedAddCommGroup K] [InnerProductSpace ℝ K] variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] theorem eq_projectionPoint_epigraph_iff_max_le_and_variational_inequality_of_mem_gammaZero {f : H → Set.Ioi (⊥ : EReal)} (hf : f ∈ Γ₀(H)) {x p : H} {ξ π : ℝ} : (p, π) = projectionPoint (epigraph (fun y : H ↦ (f y : EReal))) (isChebyshev_epigraph_of_mem_gammaZero hf) (x, ξ) ↔ max (ξ : EReal) (f p : EReal) ≤ (π : EReal) ∧ ∀ y ∈ effectiveDomain f, ⟪y - p, x - p⟫_ℝ + ((f y : EReal).toReal - π) * (ξ - π) ≤ 0

Proposition 9.19

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

Source header · Attribution source

Proposition 9.19 (1): if (p, π) is the projection of (x, ξ) onto the real-height epigraph of f, then π equals f p and this common value lies strictly above ξ.

open scoped InnerProductSpace open WithLp universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] theorem strict_lt_value_and_value_eq_height_of_eq_projectionPoint_epigraph_of_mem_gammaZero {f : H → Set.Ioi (⊥ : EReal)} (hf : f ∈ Γ₀(H)) {x p : H} {ξ π : ℝ} (hx : x ∈ effectiveDomain f) (hξ : ξ < (f x : EReal).toReal) : (p, π) = projectionPoint (epigraph (fun y : H ↦ (f y : EReal))) (isChebyshev_epigraph_of_mem_gammaZero hf) (x, ξ) → ξ < (f p : EReal).toReal ∧ (f p : EReal).toReal = π

Theorem 9.20

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

Source header · Attribution source

Theorem 9.20: every function in Γ₀(H) on a real Hilbert space admits a continuous affine minorant; more precisely, there exist a point p ∈ dom f and a vector u such that the affine functional y ↦ ⟪y - p, u⟫_ℝ + f(p) is pointwise bounded above by f.

open scoped InnerProductSpace open WithLp universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] theorem exists_affine_minorant_of_mem_gammaZero {f : H → Set.Ioi (⊥ : EReal)} (hf : f ∈ Γ₀(H)) : ∃ p ∈ effectiveDomain f, ∃ u : H, ∀ y : H, ((⟪y - p, u⟫_ℝ + (f p : EReal).toReal : ℝ) : EReal) ≤ (f y : EReal)

Corollary 9.21

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

Source header · Attribution source

Corollary 9.21: every Γ₀(H) function on a real Hilbert space is bounded below on every bounded subset of H, hence in particular on every nonempty bounded subset.

open scoped InnerProductSpace open WithLp universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] theorem bddBelow_on_bounded_of_mem_gammaZero {f : H → Set.Ioi (⊥ : EReal)} (hf : f ∈ Γ₀(H)) {C : Set H} (hC_bounded : Bornology.IsBounded C) : BddBelow (Set.range fun x : C ↦ (f x : EReal))

Example 9.22

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

Source header · Attribution source

Example 9.22: a discontinuous real linear functional has no continuous affine minorant. This specializes the textbook Hilbert-space example, where infinite dimensionality is only used to guarantee the existence of such a discontinuous functional.

universe u variable {H : Type u} [NormedAddCommGroup H] [NormedSpace ℝ H] theorem no_continuous_affine_minorant_of_not_continuous (f : H →ₗ[ℝ] ℝ) (hf : ¬ Continuous f) : ¬ ∃ g : H →ᴬ[ℝ] ℝ, ∀ x : H, g x ≤ f x

Theorem 9.23

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

Source header · Attribution source

Theorem 9.23: if f ∈ Γ₀(H) and x lies in the interior of its effective domain, then f admits a supporting continuous affine minorant through (x, f x); equivalently, in coordinates, one may write this affine minorant as y ↦ ⟪y - x, u⟫_ℝ + (f x).toReal for some u : H.

open Set open scoped InnerProductSpace open WithLp universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] theorem exists_supporting_affine_minorant_of_mem_interior_effectiveDomain_of_mem_gammaZero {f : H → Set.Ioi (⊥ : EReal)} (hf : f ∈ Γ₀(H)) {x : H} (hx : x ∈ interior (effectiveDomain f)) : ∃ g : H →ᴬ[ℝ] ℝ, g x = (f x : EReal).toReal ∧ ∀ y : H, (g y : EReal) ≤ (f y : EReal)

Proposition 9.24

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

Source header · Attribution source

Proposition 9.24: Jensen's inequality for a Γ₀(ℝ) function over a finite nonzero measure space. If x is integrable, x takes values in effectiveDomain φ almost everywhere, EReal.toReal ∘ φ ∘ x is integrable, and the average value of x lies in interior (effectiveDomain φ), then the real-valued representative of φ at that average is bounded above by the average of the real-valued representative along x.

open MeasureTheory open Set universe u namespace ERealFunction variable {Ω : Type u} [MeasurableSpace Ω] {μ : Measure Ω} [IsFiniteMeasure μ] [NeZero μ] theorem jensen_inequality_of_mem_gammaZero (φ : ℝ → Set.Ioi (⊥ : EReal)) (hφ : φ ∈ Γ₀(ℝ)) (x : Ω → ℝ) (hx : Integrable x μ) (hφx_mem : ∀ᵐ ω ∂μ, x ω ∈ effectiveDomain φ) (hφx_toReal : Integrable (fun ω ↦ (φ (x ω) : EReal).toReal) μ) (h_average_mem : (⨍ ω, x ω ∂μ) ∈ interior (effectiveDomain φ)) : (φ (⨍ ω, x ω ∂μ) : EReal).toReal ≤ ⨍ ω, (φ (x ω) : EReal).toReal ∂μ

Example 9.25

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

Source header · Attribution source

Example 9.25 (1): on a finite measure space, every H-valued L^q function with 0 < p < q satisfies the standard comparison estimate between its L^p and L^q norms.

open MeasureTheory open Set universe u v variable {Ω : Type u} [MeasurableSpace Ω] {μ : Measure Ω} [IsFiniteMeasure μ] variable {H : Type v} [NormedAddCommGroup H] -- Proof sketch: rewrite the textbook left- and right-hand sides as the corresponding -- `MeasureTheory.eLpNorm` expressions for exponents `ENNReal.ofReal p` and `ENNReal.ofReal q`, -- then apply `MeasureTheory.eLpNorm_le_eLpNorm_mul_rpow_measure_univ` to `x`. The hypothesis -- `hpq : p < q` supplies the required exponent monotonicity. theorem lp_norm_le_measure_univ_rpow_mul_lq_norm {p q : ℝ} (hp : 0 < p) (hpq : p < q) {x : Ω → H} (hx : MemLp x (ENNReal.ofReal q) μ) : (∫ ω, ‖x ω‖ ^ p ∂μ) ^ (1 / p) ≤ ((μ Set.univ).toReal) ^ (1 / p - 1 / q) * (∫ ω, ‖x ω‖ ^ q ∂μ) ^ (1 / q)

Example 9.26

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

Source header · Attribution source

Example 9.26: on a probability space, the L^p norm of a real-valued random variable is bounded above by its L^q norm whenever 0 < p < q. This is the probability-space specialization of lp_norm_le_measure_univ_rpow_mul_lq_norm to H = ℝ.

open MeasureTheory universe u variable {Ω : Type u} [MeasurableSpace Ω] {P : Measure Ω} -- Proof sketch: this is the probability-measure specialization of the chapter's finite-measure -- owner theorem `lp_norm_le_measure_univ_rpow_mul_lq_norm`; the factor -- `((P Set.univ).toReal) ^ (1 / p - 1 / q)` simplifies to `1` because `P` is a probability -- measure. theorem lp_norm_le_lq_norm_of_isProbabilityMeasure [IsProbabilityMeasure P] {p q : ℝ} (hp : 0 < p) (hpq : p < q) {X : Ω → ℝ} (hX : MemLp X (ENNReal.ofReal q) P) : (∫ ω, |X ω| ^ p ∂P) ^ (1 / p) ≤ (∫ ω, |X ω| ^ q ∂P) ^ (1 / q)

Proposition 9.27

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

Source header · Attribution source

Proposition 9.27: for a proper convex ]-∞,+∞]-valued function, the directional difference quotient based at a point of the effective domain is increasing on ℝ_{++}.

universe u namespace ERealFunction variable {H : Type u} [AddCommGroup H] [Module ℝ H] theorem directionalDifferenceQuotient_monotone (f : H → Set.Ioi (⊥ : EReal)) (hconv : ConvexOn f (effectiveDomain f)) {x : H} (hx : x ∈ effectiveDomain f) (y : H) : Monotone (directionalDifferenceQuotient f x y)

Definition 9.28

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

Source header · Attribution source

Definition 9.28: the recession function sends y to the supremum of the translated increments f (x + y) - f x over the effective domain of f.

universe u namespace ERealFunction variable {H : Type u} [AddCommGroup H] noncomputable def recessionFunction (f : H → Set.Ioi (⊥ : EReal)) (hdom : (effectiveDomain f).Nonempty) : H → Set.Ioi (⊥ : EReal) := fun y ↦ ⟨sSup ((fun x : H ↦ (f (x + y) : EReal) - (f x : EReal)) '' effectiveDomain f), recessionFunction_sup_gt_bot f hdom y⟩

Proposition 9.29

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

Source header · Attribution source

Proposition 9.29: for a proper convex ]-∞,+∞]-valued function, the epigraph of its recession function is the recession cone of the epigraph.

open scoped Pointwise universe u namespace ERealFunction variable {H : Type u} [AddCommGroup H] [Module ℝ H] theorem epigraph_recessionFunction_eq_recessionCone_epigraph (f : H → Set.Ioi (⊥ : EReal)) (hconv : ConvexOn f (effectiveDomain f)) : epigraph (fun y : H ↦ (recessionFunction f hconv.nonempty y : EReal)) = Set.recessionCone (epigraph (fun y : H ↦ (f y : EReal)))

Proposition 9.30

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

Source header · Attribution source

Proposition 9.30 (1): clause (i). The recession function of a Γ₀(H) function again belongs to Γ₀(H).

open scoped Pointwise universe u v namespace ERealFunction variable {H : Type u} [TopologicalSpace H] [AddCommGroup H] [Module ℝ H] [SequentialSpace H] [IsTopologicalAddGroup H] [ContinuousSMul ℝ H] theorem recessionFunction_mem_gammaZero (f : H → Set.Ioi (⊥ : EReal)) (hf : f ∈ Γ₀(H)) : recessionFunction f hf.2.nonempty ∈ Γ₀(H)

Example 9.31

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

Source header · Attribution source

Example 9.31: a positively homogeneous function in Γ₀(H) that is finite and zero at the origin coincides with its recession function.

universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] theorem recessionFunction_eq_self_of_positivelyHomogeneous (f : H → Set.Ioi (⊥ : EReal)) (hf : f ∈ Γ₀(H)) (hzero : (f 0 : EReal) = 0) (hph : ∀ ⦃a : ℝ⦄, 0 < a → ∀ x : H, (f (a • x) : EReal) = (a : EReal) * (f x : EReal)) : recessionFunction f hf.2.nonempty = f

Example 9.32

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

Source header · Attribution source

Example 9.32: if f ∈ Γ₀(H) is supercoercive, meaning f x / ‖x‖ → +∞ as ‖x‖ → +∞, then its recession function is the extended-real indicator of the singleton {0}.

universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] theorem recessionFunction_eq_indicator_singleton_zero_of_supercoercive {f : H → Set.Ioi (⊥ : EReal)} (hf : f ∈ Γ₀(H)) (hsuper : Filter.Tendsto (fun x : H ↦ (f x : EReal) / ‖x‖) (Filter.comap (fun x : H ↦ ‖x‖) Filter.atTop) (nhds (⊤ : EReal))) : (fun y : H ↦ (recessionFunction f hf.2.nonempty y : EReal)) = Set.indicator ({(0 : H)}ᶜ) (fun _ : H ↦ (⊤ : EReal))

Proposition 9.33

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

Source header · Attribution source

Proposition 9.33: the piecewise boundary-liminf extension of a proper convex function with open effective domain and continuous real-valued restriction agrees with its lower semicontinuous convex envelope.

open Set Filter universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] theorem boundaryLiminfExtensionEReal_eq_lowerSemicontinuousConvexEnvelope (g : H → Set.Ioi (⊥ : EReal)) (hconv : ConvexOn g (effectiveDomain g)) (hopen : IsOpen (effectiveDomain g)) (hcont : ContinuousOn (fun x : H ↦ (g x : EReal).toReal) (effectiveDomain g)) : boundaryLiminfExtensionEReal g = lowerSemicontinuousConvexEnvelope (fun x : H ↦ (g x : EReal))

Proposition 9.34

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

Source header · Attribution source

Proposition 9.34: the one-sided-limit extension of a proper strictly convex function on ]\alpha,\beta[ is strictly convex.

open Set Filter open scoped Topology namespace ERealFunction theorem oneSidedLimitExtension_strictlyConvex (g : ℝ → Set.Ioi (⊥ : EReal)) (α β : EReal) (hαβ : α < β) (hdom : effectiveDomain g = erealOpenInterval α β) (hstrict : StrictlyConvex g) (hαlim : ∀ ⦃x : ℝ⦄, α = (x : EReal) → ⊥ < Filter.liminf (fun y : ℝ ↦ (g y : EReal)) (nhdsWithin x (Set.Ioi x))) (hβlim : ∀ ⦃x : ℝ⦄, β = (x : EReal) → ⊥ < Filter.liminf (fun y : ℝ ↦ (g y : EReal)) (nhdsWithin x (Set.Iio x))) : StrictlyConvex (oneSidedLimitExtension g α β hαlim hβlim)

Example 9.35

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

Source header · Attribution source

Example 9.35: the function equal to x \log x - x on (0,+∞), equal to 0 at 0, and equal to +∞ on (-∞,0) belongs to Γ₀(ℝ); equivalently, it is proper, lower semicontinuous, and strictly convex on its effective domain [0,+∞).

open Filter Set open scoped Topology namespace ERealFunction theorem boltzmannEntropy_mem_gammaZero : boltzmannEntropy ∈ Γ₀(ℝ)

Example 9.36

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

Source header · Attribution source

Example 9.36 (i): the everywhere-finite exponential function belongs to Γ₀(ℝ).

namespace ERealFunction theorem exp_mem_gammaZero : Real.exp.toEReal ∈ Γ₀(ℝ)

Remark 9.37

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

Source header · Attribution source

Remark 9.37: for a finite family of proper functions on real Hilbert spaces, the direct-sum function on the Hilbert direct sum is convex on its effective domain exactly when each coordinate function is convex on its own effective domain, and it is strictly convex exactly when each coordinate function is strictly convex.

open scoped BigOperators universe u v namespace ERealFunction variable {I : Type v} [Fintype I] variable {H : I → Type u} variable [∀ i, NormedAddCommGroup (H i)] variable [∀ i, InnerProductSpace ℝ (H i)] theorem directSumFunction_convexOn_effectiveDomain_and_strictlyConvex_iff (f : ∀ i, H i → Set.Ioi (⊥ : EReal)) (hproper : ∀ i, IsProper (fun x : H i ↦ (f i x : EReal))) : (ConvexOn (directSumFunction f) (effectiveDomain (directSumFunction f)) ↔ ∀ i, ConvexOn (f i) (effectiveDomain (f i))) ∧ (StrictlyConvex (directSumFunction f) ↔ ∀ i, StrictlyConvex (f i))

Proposition 9.38

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

Source header · Attribution source

Proposition 9.38: if p ∈ (1, +∞) and p* = p / (p - 1) = Real.conjExponent p, then for x ∈ L^p(μ) and y ∈ L^{p*}(μ) one has Hölder's inequality ∫ |x y| ≤ (∫ |x|^p)^(1/p) * (∫ |y|^p*)^(1/p*).

open MeasureTheory universe u variable {Ω : Type u} [MeasurableSpace Ω] {μ : Measure Ω} theorem holder_integral_abs_mul_le {p : ℝ} (hp : 1 < p) {x y : Ω → ℝ} (hx : MemLp x (ENNReal.ofReal p) μ) (hy : MemLp y (ENNReal.ofReal (Real.conjExponent p)) μ) : ∫ ω, |x ω * y ω| ∂μ ≤ (∫ ω, |x ω| ^ p ∂μ) ^ (1 / p) * (∫ ω, |y ω| ^ Real.conjExponent p ∂μ) ^ (1 / Real.conjExponent p)

Example 9.39

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

Source header · Attribution source

Example 9.39: for finite real families, the sum of the absolute values of the coordinate products is bounded by the product of the ℓ^p norm of x and the ℓ^{p*} norm of y, where p* = p.conjExponent.

open scoped BigOperators ENNReal universe u variable {ι : Type u} [Fintype ι] -- Proof sketch: specialize the finite-sum Hölder inequality from `Real.inner_le_Lp_mul_Lq` to -- `Finset.univ`, with exponents `p.toReal` and `p.conjExponent.toReal`, then rewrite the two -- right-hand factors using the canonical `WithLp` norm formulas on finite products. theorem holder_sum_abs_mul_le_withLpNorm_mul_conjExponent_withLpNorm (p : ENNReal) (hp : 1 ≤ p) (x y : ι → ℝ) : ∑ i, |x i * y i| ≤ ‖WithLp.toLp p x‖ * ‖WithLp.toLp p.conjExponent y‖

Proposition 9.40

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

Source header · Attribution source

Proposition 9.40: if φ ∈ Γ₀(H) and either (i) μ(Ω) < +∞ or (ii) φ ≥ φ(0) = 0, then the integral functional induced by φ belongs to Γ₀(L²((Ω,\mathcal F,\mu); H)).

open MeasureTheory open scoped ENNReal InnerProductSpace universe u v namespace ERealFunction variable {Ω : Type u} [MeasurableSpace Ω] {μ : Measure Ω} variable {H : Type v} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] [MeasurableSpace H] [BorelSpace H] [TopologicalSpace.SeparableSpace H] theorem integralFunctional_mem_gammaZero (μ : Measure Ω) (φ : H → Set.Ioi (⊥ : EReal)) (hφ : φ ∈ Γ₀(H)) (hfinite_or_nonneg : μ Set.univ < ∞ ∨ ((φ 0 : EReal) = 0 ∧ ∀ z : H, (φ 0 : EReal) ≤ (φ z : EReal))) : integralFunctional μ φ ∈ Γ₀(Ω →₂[μ] H)

Example 9.41

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

Source header · Attribution source

Example 9.41: on a finite measure space, the integral functional induced by the entropy integrand t ↦ t \log t - t on [0,+∞) belongs to Γ₀(L²((Ω,\mathcal F,\mu); \mathbb R)).

open MeasureTheory open Filter Set open scoped BigOperators open scoped Topology universe u namespace ERealFunction variable {Ω : Type u} [MeasurableSpace Ω] {μ : Measure Ω} theorem boltzmannEntropy_integralFunctional_mem_gammaZero (hμ : μ Set.univ < ⊤) : integralFunctional μ boltzmannEntropy ∈ Γ₀(Ω →₂[μ] ℝ)

Proposition 9.42

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

Source header · Attribution source

Proposition 9.42: for φ ∈ Γ₀(H), the lower semicontinuous envelope of the perspective of φ is the closed perspective obtained by inserting the recession function on the zero-height slice.

open Set open scoped Pointwise InnerProductSpace universe u namespace ERealFunction variable {H : Type u} variable [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] open WithLp theorem lowerSemicontinuousEnvelope_perspective_eq_closedPerspectiveEReal (φ : H → Set.Ioi (⊥ : EReal)) (hφ : φ ∈ Γ₀(H)) : lowerSemicontinuousEnvelope (perspective (fun x : H ↦ (φ x : EReal))) = closedPerspectiveEReal φ hφ.2.nonempty

Example 9.43

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

Source header · Attribution source

Example 9.43: for p > 1, the function g : ℝ × H → ]-∞,+∞] given by g(ξ, x) = ‖x‖^p / ξ^(p - 1) for ξ > 0, g(0, 0) = 0, and g(ξ, x) = +∞ otherwise belongs to Γ₀(ℝ × H).

universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [NormedSpace ℝ H] open scoped Topology theorem normPowerPerspectiveAtOrigin_mem_gammaZero (p : ℝ) (hp : 1 < p) : normPowerPerspectiveAtOrigin p ∈ Γ₀(ℝ × H)

Corollary 9.44

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

Source header · Attribution source

Corollary 9.44: if there exists z with ρ ≤ ⟪z, u⟫_ℝ and L z - r ∈ (⟪z, u⟫_ℝ - ρ) • effectiveDomain φ, then the function obtained from the closed perspective of φ by the substitution x ↦ (⟪x, u⟫_ℝ - ρ, L x - r) belongs to Γ₀(H).

open scoped InnerProductSpace Pointwise universe u v namespace ERealFunction variable {H : Type u} {K : Type v} variable [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] variable [NormedAddCommGroup K] [InnerProductSpace ℝ K] [CompleteSpace K] theorem closedPerspective_comp_mem_gammaZero (φ : K → Set.Ioi (⊥ : EReal)) (hφ : φ ∈ Γ₀(K)) (L : H →L[ℝ] K) (r : K) (u : H) (ρ : ℝ) (hz : ∃ z : H, ρ ≤ ⟪z, u⟫_ℝ ∧ L z - r ∈ (⟪z, u⟫_ℝ - ρ) • effectiveDomain φ) : closedPerspective φ hφ.2.nonempty ∘ (fun x : H ↦ (⟪x, u⟫_ℝ - ρ, L x - r)) ∈ Γ₀(H)

Example 9.45

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

Source header · Attribution source

Example 9.45: for p > 1, if there exists z with either ρ < ⟪z, u⟫_ℝ or ⟪z, u⟫_ℝ = ρ and L z = r, then the function x ↦ ‖L x - r‖^p / (⟪x, u⟫_ℝ - ρ)^(p - 1) on the half-space ρ < ⟪x, u⟫_ℝ, extended by 0 when ⟪x, u⟫_ℝ = ρ and L x = r, and by +∞ otherwise, belongs to Γ₀(H).

open scoped InnerProductSpace universe u v namespace ERealFunction variable {H : Type u} (K : Type v) variable [NormedAddCommGroup H] [InnerProductSpace ℝ H] variable [NormedAddCommGroup K] [NormedSpace ℝ K] theorem normPowerPerspective_mem_gammaZero (p : ℝ) (hp : 1 < p) (L : H →L[ℝ] K) (r : K) (u : H) (ρ : ℝ) (hz : ∃ z : H, ρ < ⟪z, u⟫_ℝ ∨ ⟪z, u⟫_ℝ = ρ ∧ L z = r) : normPowerPerspective K p L r u ρ ∈ Γ₀(H)

Example 9.46

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

Source header · Attribution source

Example 9.46: for φ ∈ Γ₀(ℝ), the finite-dimensional function d_φ obtained by summing the scalar closed perspective coordinatewise on the canonical Fin N → ℝ model of ℝ^N × ℝ^N belongs to Γ₀(ℝ^N × ℝ^N).

open scoped BigOperators namespace ERealFunction theorem coordinatePhiDivergence_mem_gammaZero (N : ℕ) (φ : ℝ → Set.Ioi (⊥ : EReal)) (hφ : φ ∈ Γ₀(ℝ)) : coordinatePhiDivergence N φ hφ.2.nonempty ∈ Γ₀((Fin N → ℝ) × (Fin N → ℝ))

Remark 9.47

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

Source header · Attribution source

Remark 9.47: if φ ∈ Γ₀(ℝ) is supercoercive, so (φ t : EReal) / ‖t‖ → +∞ as ‖t‖ → +∞, then the finite-dimensional φ-divergence from Example 9.46 reduces to the sum of the positive coordinate perspective terms when y has no negative coordinates and every zero coordinate of y also has vanishing x-coordinate, and equals +∞ otherwise.

open scoped BigOperators namespace ERealFunction theorem coordinatePhiDivergence_eq_reduced_formula_of_supercoercive (N : ℕ) (φ : ℝ → Set.Ioi (⊥ : EReal)) (hφ : φ ∈ Γ₀(ℝ)) (hsuper : Filter.Tendsto (fun t : ℝ ↦ (φ t : EReal) / ‖t‖) (Filter.comap (fun t : ℝ ↦ ‖t‖) Filter.atTop) (nhds (⊤ : EReal))) (x y : Fin N → ℝ) : (coordinatePhiDivergence N φ hφ.2.nonempty (x, y) : EReal) = if coordinatePhiDivergenceNegativeIndices y = ∅ ∧ ∀ i ∈ coordinatePhiDivergenceZeroIndices y, x i = 0 then ∑ i ∈ coordinatePhiDivergencePositiveIndices y, (y i : EReal) * (φ (x i / y i) : EReal) else ⊤

Example 9.48

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

Source header · Attribution source

Example 9.48: the explicit formula obtained by allowing zero coordinates in the second variable and summing only over the common positive coordinates is the lower semicontinuous hull of the Kullback-Leibler divergence of Example 8.27.

open Set open scoped BigOperators namespace ERealFunction theorem lowerSemicontinuousEnvelope_kullbackLeiblerDivergence_eq_relativeEntropyLowerSemicontinuousHull (N : ℕ) : lowerSemicontinuousEnvelope (kullbackLeiblerDivergence N) = relativeEntropyLowerSemicontinuousHull N