Chapter 13 — Conjugation

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

Definition 13.1

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

Source header · Attribution source

Definition 13.1: the conjugate of an extended-real-valued function on a real inner-product space is the function u ↦ sup_x (⟪x, u⟫ - f x).

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] noncomputable def conjugate (f : H → EReal) : H → EReal := fun u ↦ ⨆ x : H, ((⟪x, u⟫_ℝ : ℝ) : EReal) - f x

Example 13.2

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

Source header · Attribution source

Example 13.2 (1): clause (i). For p ∈ ]1,+∞[, the conjugate of x ↦ |x|^p / p is u ↦ |u|^{p*} / p*, where p* = Real.conjExponent p = p / (p - 1).

open scoped ERealFunction open scoped InnerProductSpace open Real namespace ERealFunction theorem conjugate_absRpowDivided (p : ℝ) (hp : 1 < p) (u : ℝ) : ((fun x : ℝ ↦ (|x| ^ p) / p).toEReal.asEReal)∗ u = ((fun x : ℝ ↦ (|x| ^ Real.conjExponent p) / Real.conjExponent p).toEReal.asEReal) u

Example 13.3

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

Source header · Attribution source

Example 13.3 (1): clause (i). The conjugate of the indicator ι[C] is the support function σ[C].

open scoped InnerProductSpace Set open Metric universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] -- Proof sketch: expand `conjugate` for the complement indicator of `C`. On `C` the indicator -- term vanishes, and outside `C` it contributes `⊤`, so only points of `C` survive in the -- defining supremum; what remains is exactly the support function `σ[C]`. theorem conjugate_indicator_eq_supportFunction (C : Set H) : ((ι[C]).asEReal)∗ = σ[C]

Example 13.4

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

Source header · Attribution source

Example 13.4: for φ : H → ]-∞,+∞] and γ ∈ ℝ_{++}, the Fenchel conjugate of the regularized function x ↦ φ x + ‖x‖² / (2γ) is u ↦ γ‖u‖² / 2 - {}^γφ(γu).

universe u open scoped InnerProductSpace namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] -- Proof sketch: expand the Fenchel-conjugate value of -- `x ↦ φ x + ‖x‖² / (2γ)` at `u`, complete the square in the quadratic term, and recognize the -- remaining infimum as the `γ`-Moreau envelope of `φ` evaluated at `γ • u`. theorem conjugate_regularized_eq_scaledQuadratic_sub_moreauEnvelope (φ : H → Set.Ioi (⊥ : EReal)) (γ : Set.Ioi (0 : ℝ)) : ((φ + moreauQuadraticKernel γ).asEReal)∗ = fun u : H ↦ ((((γ : ℝ) / 2) * ‖u‖ ^ 2 : ℝ) : EReal) - ({}^[γ] φ) ((γ : ℝ) • u)

Example 13.5

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

Source header · Attribution source

Example 13.5: if C is nonempty and f = ι_C + ‖·‖² / 2, then the Fenchel conjugate f* is u ↦ (‖u‖² - d(u, C)²) / 2.

open Set open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] -- Proof sketch: apply Example 13.4 to the indicator of `C` with parameter `γ = 1`, then rewrite -- the Moreau envelope term using the squared-distance formula for the indicator function. theorem fenchelConjugate_indicator_add_halfSquaredNorm_eq_sqNorm_sub_sqInfDist_div_two (C : Set H) (hC_nonempty : C.Nonempty) : ((ι[C] + halfSquaredNorm).asEReal)∗ = fun u : H ↦ ((((‖u‖ ^ 2 - Metric.infDist u C ^ 2) / 2 : ℝ) : EReal))

Example 13.6

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

Source header · Attribution source

Example 13.6: for f = (1 / 2)‖·‖², the Fenchel conjugate f* equals f.

open Set open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] -- Proof sketch: specialize Example 13.5 to `C = univ`, use that `univ` is nonempty and that -- `Metric.infDist u univ = 0`, then rewrite the resulting right-hand side as `‖u‖² / 2`. theorem fenchelConjugate_halfSquaredNorm : ((halfSquaredNorm : H → Set.Ioi (⊥ : EReal)).asEReal)∗ = (halfSquaredNorm : H → Set.Ioi (⊥ : EReal)).asEReal

Example 13.7

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

Source header · Attribution source

Example 13.7: the Fenchel conjugate of ι_{B(0;ρ)} + ‖·‖² / 2 is the radial Huber profile u ↦ huberFunction ρ ‖u‖, viewed in EReal.

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] theorem fenchelConjugate_indicator_add_halfSquaredNorm_closedBall_eq_huberFunction_comp_norm (ρ : Set.Ioi (0 : ℝ)) : ((ι[Metric.closedBall (0 : H) (ρ : ℝ)] + halfSquaredNorm).asEReal)∗ = (huberFunction ρ ∘ (norm : H → ℝ)).toEReal.asEReal

Example 13.8

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

Source header · Attribution source

Example 13.8: if φ : ℝ → ]-∞,+∞] is even, then the Fenchel conjugate of the radial function x ↦ φ ‖x‖ is the radial function u ↦ φ^*(‖u‖).

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] -- Proof sketch: expand the two Fenchel conjugates. The radial parametrization `x = ρ z` with -- `‖z‖ = 1` reduces the supremum over `H` to a supremum over `ρ ≥ 0`, the evenness of `φ` extends -- this to all `ρ : ℝ`, and the support-function identity of the unit ball gives the factor `‖u‖`. theorem conjugate_comp_norm_eq_comp_norm_conjugate_of_even (φ : ℝ → Set.Ioi (⊥ : EReal)) (hφ_even : Function.Even φ) : (φ ∘ (norm : H → ℝ)).asEReal∗ = φ.asEReal∗ ∘ (norm : H → ℝ)

Example 13.9

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

Source header · Attribution source

Example 13.9: if φ attains at least one finite value, then the Fenchel conjugate of the perspective is the textbook indicator ι[C] of the set C = {(μ, u) | μ + φ^*(u) ≤ 0}.

open Set open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] theorem conjugate_perspective_eq_indicator_perspectiveConjugateSet (φ : H → Set.Ioi (⊥ : EReal)) (hdom : (effectiveDomain φ).Nonempty) : (perspective φ.asEReal)∗ = (ι[perspectiveConjugateSet φ]).asEReal

Proposition 13.10

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

Source header · Attribution source

Proposition 13.10 (1): clause (i). The conjugate at the origin is the negative infimum of the values of f.

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] theorem conjugate_zero_eq_neg_iInf (f : H → EReal) : f∗ 0 = - (⨅ x : H, f x)

Proposition 13.11

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

Source header · Attribution source

Proposition 13.11: if dom f is nonempty, then the support function of the epigraph of f, evaluated at (u, -μ) in the Hilbert product space H × ℝ, equals μ f*(u / μ) for μ > 0, equals the support function of dom f for μ = 0, and equals +∞ for μ < 0.

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] theorem supportFunction_epigraph_eq {f : H → EReal} (hdom : (dom f).Nonempty) (u : H) (μ : ℝ) : σ[epigraph f] (u, -μ) = if 0 < μ then (μ : EReal) * f∗ (μ⁻¹ • u) else if μ = 0 then σ[dom f] u else ⊤

Proposition 13.12

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

Source header · Attribution source

Proposition 13.12 (1): clause (i). A point (u, μ) lies in the epigraph of the conjugate if and only if the continuous affine functional x ↦ ⟪x, u⟫ - μ is a minorant of f.

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] -- Proof sketch: expand epigraph membership into `conjugate f u ≤ μ`, unfold the supremum -- definition of `conjugate f u`, and rewrite the resulting pointwise inequalities into the affine -- minorant form `⟪x,u⟫ - μ ≤ f x`. theorem mem_epigraph_conjugate_iff (f : H → EReal) (u : H) (μ : ℝ) : (u, μ) ∈ epigraph f∗ ↔ ∀ x : H, (((⟪x, u⟫_ℝ - μ : ℝ) : EReal) ≤ f x)

Proposition 13.13

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

Source header · Attribution source

Proposition 13.13: the Fenchel conjugate of an extended-real-valued function on a real inner-product space belongs to the textbook class Γ(H) of convex lower semicontinuous functions.

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] theorem conjugate_mem_gamma (f : H → EReal) : f∗ ∈ gamma H

Example 13.14

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

Source header · Attribution source

Example 13.14 (1): for γ ∈ ℝ_{++}, the function u ↦ (γ / 2) ‖u‖² - {}^γφ(γ u) belongs to Γ(H).

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] -- Proof sketch: Example 13.4 identifies this function with the Fenchel conjugate of -- `x ↦ φ x + ‖x‖² / (2γ)`, and Proposition 13.13 places every conjugate in `Γ(H)`. theorem scaledQuadratic_sub_moreauEnvelope_comp_smul_mem_gamma (φ : H → Set.Ioi (⊥ : EReal)) (γ : Set.Ioi (0 : ℝ)) : (fun u : H ↦ ((((γ : ℝ) / 2) * ‖u‖ ^ 2 : ℝ) : EReal) - ({}^[γ] φ) ((γ : ℝ) • u)) ∈ gamma H

Proposition 13.15

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

Source header · Attribution source

Proposition 13.15: for a proper extended-real-valued function on a real inner-product space, the Fenchel--Young inequality states that f x + f*(u) dominates the pairing ⟪x, u⟫.

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] theorem fenchel_young_inequality {f : H → EReal} (hproper : IsProper f) (x u : H) : ((⟪x, u⟫_ℝ : ℝ) : EReal) ≤ f x + f∗ u

Proposition 13.16

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

Source header · Attribution source

Proposition 13.16 (1): clause (i). Every extended-real-valued function dominates its biconjugate.

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] -- Proof sketch: apply `fenchel_young_inequality` with the outer conjugation variable in place of -- `u`, then take the supremum over the primal variable to show each value of `f∗∗` is bounded -- above by the corresponding value of `f`. theorem biconjugate_le (f : H → EReal) : f∗∗ ≤ f

Example 13.17

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

Source header · Attribution source

Example 13.17: for f(x) = -|x|, the lower semicontinuous convex envelope \tilde f is identically -∞.

open scoped ERealFunction InnerProductSpace namespace ERealFunction theorem lowerSemicontinuousConvexEnvelope_negAbs_eq_bot : lowerSemicontinuousConvexEnvelope (fun x : ℝ ↦ (-|x| : EReal)) = fun _ : ℝ ↦ (⊥ : EReal)

Example 13.18

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

Source header · Attribution source

Example 13.18 (1): on the real vector space of self-adjoint operators, the Loewner partial order is the ambient operator order, characterized by ⟪A x, x⟫ ≤ ⟪B x, x⟫ for every x.

open ContinuousLinearMap open scoped InnerProductSpace universe u variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] notation "S(" H ")" => selfAdjoint (H →L[ℝ] H) infix:50 " ≺ " => ContinuousLinearMap.StrictlyLoewner -- Proof sketch: rewrite the ambient order by `ContinuousLinearMap.le_def`, express positivity via -- `ContinuousLinearMap.isPositive_iff'`, and use that the difference of two self-adjoint operators -- is self-adjoint to reduce to the quadratic-form inequality. theorem loewner_le_iff_forall_inner_le (A B : S(H)) : A ≤ B ↔ ∀ x : H, ⟪(A : H →L[ℝ] H) x, x⟫_ℝ ≤ ⟪(B : H →L[ℝ] H) x, x⟫_ℝ

Text 13.18.1

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

Source header · Attribution source

Text 13.18.1: clause (i). For self-adjoint operators, if β • Id ≤ B ≤ A, then taking inverses reverses the Loewner order: A⁻¹ ≤ B⁻¹.

open ContinuousLinearMap open scoped InnerProduct InnerProductSpace universe u variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] variable {α β : Set.Ioi (0 : ℝ)} {A B : S(H)} -- Proof sketch: from `β • 1 ≤ B ≤ A` in the self-adjoint Loewner order, both operators are -- positive and bounded below away from zero, hence invertible. Apply the standard order-reversing -- property of inversion on positive operators. theorem sandwiched_positive_operators_inverse_antitone (hBA : B ≤ A) (hβB : (β : ℝ) • 1 ≤ B) : (A : H →L[ℝ] H).inverse ≤ (B : H →L[ℝ] H).inverse

Proposition 13.19

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

Source header · Attribution source

Proposition 13.19: an extended-real-valued function on a real inner-product space is self-conjugate exactly when it is the canonical quadratic owner halfSquaredNorm.asEReal.

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] -- Proof sketch: one direction is Example 13.6, which identifies the conjugate of the quadratic -- `x ↦ ‖x‖² / 2` with itself. For the converse, apply the Fenchel--Young inequality to `f` at -- `(x, x)` to get the lower bound `‖x‖² / 2 ≤ f x`, then use the order-reversing property of -- conjugation from Proposition 13.16 together with `f = f∗` to force the reverse inequality. theorem self_conjugate_iff_eq_half_squared_norm (f : H → EReal) : f = f∗ ↔ f = halfSquaredNorm.asEReal

Remark 13.20

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

Source header · Attribution source

Remark 13.20 (1): clause (i). A cone in a nontrivial real inner-product space cannot be self-polar. In particular, a convex cone cannot be self-polar.

open scoped ERealFunction InnerProductSpace Pointwise open ERealFunction universe u namespace Set variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] variable [Nontrivial H] -- Proof sketch: if a cone were self-polar, then `0 ∈ K`, so the indicator of `K` would be -- self-conjugate by the standard indicator/polar-cone correspondence for nonempty cones, -- contradicting the uniqueness of self-conjugate functions on a nontrivial Hilbert space. theorem isCone_ne_polarCone_of_nontrivial {K : Set H} (hK_cone : IsCone K) : K ≠ Kᵒ⊖

Proposition 13.21

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

Source header · Attribution source

Proposition 13.21: the conjugate of an even extended-real-valued function on a real inner-product space is even.

universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] -- Proof sketch: unfold `conjugate`, rewrite `conjugate f (-u)` as the supremum of -- `x ↦ ⟪x, -u⟫ - f x`, replace `x` by `-x`, and use the evenness of `f`. theorem conjugate_even (f : H → EReal) (hf : Function.Even f) : Function.Even f∗

Proposition 13.22

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

Source header · Attribution source

Proposition 13.22: if f attains its minimum value 0 at the origin, then the conjugate is bounded below by its value at the origin: f*(0) ≤ f*(u) for every u.

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] -- Proof sketch: Proposition 13.10 identifies `f*(0)` with the negative infimum of `f`. The -- minimum assumptions imply `⨅ x, f x = 0`, so the value at the origin is `0`. theorem conjugate_zero_le_conjugate_of_minimum_at_zero (f : H → EReal) (h_min : ∀ x : H, f 0 ≤ f x) (h_zero : f 0 = 0) (u : H) : f∗ 0 ≤ f∗ u

Proposition 13.23

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

Source header · Attribution source

Proposition 13.23 (1): clause (i). Scaling an extended-real-valued function by a positive scalar sends its conjugate to the same scalar multiple of the conjugate evaluated at the inversely scaled dual variable.

open scoped InnerProductSpace translate universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] -- Proof sketch: expand `conjugate`, pull the positive scalar `α` out of the supremum, and rewrite -- `⟪x, u⟫ - α f x` as `α * (⟪x, α⁻¹ • u⟫ - f x)`. theorem conjugate_pos_smul (f : H → EReal) (α : Set.Ioi (0 : ℝ)) : (fun x : H ↦ ((α : ℝ) : EReal) * f x)∗ = fun u : H ↦ ((α : ℝ) : EReal) * f∗ (((α : ℝ)⁻¹) • u)

Proposition 13.24

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

Source header · Attribution source

Proposition 13.24 (1): the Fenchel conjugate of the infimal convolution f □ g is the pointwise sum f* + g*.

open scoped InnerProductSpace universe u v namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] -- Proof sketch: expand the conjugate of `f □ g`, rewrite the defining infimum as a translated -- supremum over decompositions `x = y + z`, and separate the two resulting suprema. theorem conjugate_infimalConvolution_eq (f g : H → Set.Ioi (⊥ : EReal)) : (f □ g)∗ = f.asEReal∗ + g.asEReal∗

Corollary 13.25

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

Source header · Attribution source

Corollary 13.25 (1): if the infimal postcomposition L ▷ f never attains -∞, then the conjugate of its infimal convolution with g is (f* ∘ L*) + g*.

open scoped InnerProductSpace universe u v namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] variable {K : Type v} [NormedAddCommGroup K] [InnerProductSpace ℝ K] variable [CompleteSpace H] [CompleteSpace K] -- Proof sketch: use the hypothesis `⊥ ∉ range (L ▷ f)` to regard `L ▷ f` as an -- `]-∞,+∞]`-valued function, apply Proposition 13.24(1) to its infimal convolution with `g`, -- and then rewrite `(L ▷ f)^*` by Proposition 13.24(4). theorem conjugate_infimalConvolution_infimalPostcomposition_eq (f : H → Set.Ioi (⊥ : EReal)) (g : K → Set.Ioi (⊥ : EReal)) (L : H →L[ℝ] K) (hL : (⊥ : EReal) ∉ Set.range (L ▷ f)) : (infimalConvolution (L ▷ f) g.asEReal)∗ = f.asEReal∗ ∘ L.adjoint + g.asEReal∗

Example 13.26

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

Source header · Attribution source

Example 13.26: if C is a nonempty closed convex subset of H and φ is increasing on [0, ∞) and even, then the Fenchel conjugate of x ↦ φ (d(x, C)) is σ[C] + (u ↦ φ^*(‖u‖)).

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] -- Proof sketch: rewrite `x ↦ φ (d(x, C))` as the infimal convolution of `ι[C]` with the radial -- kernel `x ↦ φ ‖x‖`, the monotone-radial companion of -- `distanceToSet_eq_indicator_infimalConvolution_norm` from Example 12.2. Then apply Proposition -- 13.24(i), Example 13.3(i), and Example 13.8 to identify the three conjugates. theorem fenchelConjugate_comp_infDist_eq_supportFunction_add_comp_norm (C : Set H) (hC_nonempty : C.Nonempty) (hC_closed : IsClosed C) (hC_convex : Convex ℝ C) (φ : ℝ → Set.Ioi (⊥ : EReal)) (hφ_mono : MonotoneOn φ (Set.Ici (0 : ℝ))) (hφ_even : Function.Even φ) : (φ ∘ fun x : H ↦ Metric.infDist x C).asEReal∗ = σ[C] + φ.asEReal∗ ∘ norm

Example 13.27

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

Source header · Attribution source

Example 13.27 (1): if C is a nonempty subset of H, then the Fenchel conjugate of the distance function d_C is the support function σ[C] plus the indicator of the closed unit ball B(0;1).

open scoped InnerProductSpace open Metric universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] -- Proof sketch: rewrite `Metric.infDist` through the source-facing distance decomposition from -- Example 12.2, then apply Proposition 13.24(i), Example 13.3(i), and Example 13.3(v). The -- resulting support function already depends only on the underlying set `C`. theorem fenchelConjugate_infDist_eq_supportFunction_add_indicator_closedUnitBall (C : Set H) (hC_nonempty : C.Nonempty) : (fun x : H ↦ Metric.infDist x C).toEReal.asEReal∗ = σ[C] + (ι[closedBall (0 : H) 1]).asEReal

Proposition 13.28

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

Source header · Attribution source

Proposition 13.28 (2): the Fenchel conjugate of a pointwise supremum is bounded above by the pointwise infimum of the Fenchel conjugates. In particular, this recovers textbook clause (ii) for a family of proper functions.

open scoped InnerProductSpace universe u v namespace ERealFunction variable {I : Type v} {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] -- Proof sketch: for each index `i`, the pointwise inequality `f i ≤ ⨆ j : I, f j` holds. Apply -- the order-reversing property of Fenchel conjugation from Proposition 13.16(ii), then take the -- infimum over `i` on the right-hand side. theorem conjugate_iSup_le_iInf_conjugate (f : I → H → EReal) : (⨆ i : I, f i)∗ ≤ ⨅ i : I, (f i)∗

Proposition 13.29

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

Source header · Attribution source

Proposition 13.29 (1): for an ]-∞,+∞]-valued function f and γ > 0, the Fenchel conjugate of γ f - q equals γ (γ q - f*)* - q, where q(x) = ‖x‖² / 2.

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] variable (f : H → Set.Ioi (⊥ : EReal)) (γ : Set.Ioi (0 : ℝ)) -- Proof sketch: apply Proposition 13.28 to the infimum presentation of `γ q - f*`, use the -- positive-scaling conjugation rule from Proposition 13.23, and simplify the resulting quadratic -- terms to obtain formula (13.22). theorem conjugate_smul_sub_halfSquaredNorm_eq : (fun x : H ↦ ((γ : ℝ) : EReal) * f.asEReal x - halfSquaredNorm.asEReal x)∗ = fun u : H ↦ ((γ : ℝ) : EReal) * ((fun v : H ↦ ((γ : ℝ) : EReal) * halfSquaredNorm.asEReal v - f.asEReal∗ v)∗ u) - halfSquaredNorm.asEReal u

Proposition 13.30

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

Source header · Attribution source

Proposition 13.30: after coercing a finite direct sum of ]-∞,+∞]-valued functions to EReal, its Fenchel conjugate is the finite sum of the coordinate Fenchel conjugates.

open scoped ERealFunction InnerProductSpace universe u v namespace ERealFunction variable {I : Type v} variable {H : I → Type u} variable [∀ i, NormedAddCommGroup (H i)] [∀ i, InnerProductSpace ℝ (H i)] variable [Fintype I] -- Proof sketch: rewrite the `Γ₀`-valued direct sum as the Chapter 8 owner `⨁ i, ...`, apply the -- owner-level conjugate identity above, then specialize the resulting finite Hilbert sum back to -- an ordinary coordinate sum. theorem conjugate_directSumFunction_eq_sum_conjugate (f : ∀ i, H i → Set.Ioi (⊥ : EReal)) : (directSumFunction f).asEReal∗ = fun u : lp H 2 ↦ ∑ i, (f i).asEReal∗ (u i)

Example 13.31

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

Source header · Attribution source

Example 13.31: on ℝ^N, for p ∈ ]1,+∞[, the Fenchel conjugate of x ↦ ‖x‖_p^p / p is u ↦ ‖u‖_{p*}^{p*} / p*, where p* = Real.conjExponent p = p / (p - 1).

open scoped ENNReal InnerProductSpace namespace ERealFunction theorem fenchelConjugate_lpNormPowerDivided_eq_lpNormPowerDivided_conjExponent (N : ℕ) (p : ℝ) (hp : 1 < p) : (fun x : EuclideanSpace ℝ (Fin N) ↦ ((‖x‖_[ENNReal.ofReal p] ^ p / p : ℝ) : EReal))∗ = fun u : EuclideanSpace ℝ (Fin N) ↦ ((‖u‖_[ENNReal.ofReal (Real.conjExponent p)] ^ Real.conjExponent p / Real.conjExponent p : ℝ) : EReal)

Example 13.32

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

Source header · Attribution source

Example 13.32: on ℝ^N, for p ∈ [1,+∞], the conjugate of the coordinate ℓ^p norm is the indicator of the conjugate-exponent unit ball B_{p*}.

open WithLp open scoped ENNReal namespace ERealFunction theorem conjugate_lpNorm_eq_indicator_lpClosedUnitBall_conjExponent (N : ℕ) (p : ℝ≥0∞) [Fact (1 ≤ p)] : (fun x : EuclideanSpace ℝ (Fin N) ↦ (‖x‖_[p] : EReal))∗ = (ι[Set.lpClosedUnitBall N p.conjExponent]).asEReal

Proposition 13.33

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

Source header · Attribution source

Proposition 13.33: if f(x) = inf_{y ∈ K} F(x, y), then the conjugate of f equals the canonical Hilbert-product conjugate of F along the slice u ↦ (u, 0).

open scoped InnerProductSpace universe u v namespace ERealFunction variable {H : Type u} {K : Type v} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [NormedAddCommGroup K] [InnerProductSpace ℝ K] theorem conjugate_infimalPostcomposition_fst_eq_conjugate_zeroSecond (F : H × K → EReal) : (Prod.fst ▷ F)∗ = fun u ↦ F∗ (u, (0 : K))

Definition 13.34

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

Source header · Attribution source

Definition 13.34: an extended-real-valued function on H × H is autoconjugate when its Fenchel conjugate for the canonical pairing on the Hilbert product equals its transpose.

open scoped InnerProductSpace open WithLp universe u v namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] def autoconjugate (F : H × H → EReal) : Prop := F∗ = Fᵀ

Proposition 13.35

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

Source header · Attribution source

Proposition 13.35: on H × H, Fenchel conjugation commutes with transpose ᵀ.

open scoped InnerProductSpace open WithLp universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] theorem transpose_conjugate (F : H × H → EReal) : (F∗)ᵀ = (Fᵀ)∗

Proposition 13.36

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

Source header · Attribution source

Proposition 13.36 (1): a proper autoconjugate function on H × H dominates the canonical pairing ⟪x, u⟫ pointwise.

open scoped InnerProductSpace open WithLp universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] theorem pairing_le_autoconjugate {F : H × H → EReal} (hproper : IsProper F) (hauto : autoconjugate F) (x u : H) : ((⟪x, u⟫_ℝ : ℝ) : EReal) ≤ F (x, u)

Theorem 13.37

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

Source header · Attribution source

Theorem 13.37: for a proper extended-real-valued function on a real Hilbert space, being lower semicontinuous and convex, equivalently belonging to Γ(H), is equivalent to coinciding with its Fenchel biconjugate.

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] -- Proof sketch: if `f** = f`, then `f` is a Fenchel conjugate by Definition 13.1, so -- Proposition 13.13 puts it in `Γ(H)`. Conversely, if `f ∈ Γ(H)`, apply the supporting -- hyperplane argument of the Fenchel--Moreau theorem to get `f ≤ f**`, then combine with -- Proposition 13.16(i), which always gives `f** ≤ f`. theorem mem_gamma_iff_eq_biconjugate_of_is_proper {f : H → EReal} (hproper : IsProper f) : f ∈ Γ(H) ↔ f∗∗ = f

Corollary 13.38

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

Source header · Attribution source

Corollary 13.38: if f ∈ Γ₀(H), then its canonical Γ₀(H)-valued Fenchel conjugate again belongs to Γ₀(H), and the Fenchel biconjugate of the EReal-valued coercion of f is f itself.

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] -- Proof sketch: Theorem 13.37 applies to the `EReal`-valued coercion of `f` because -- `isProper_of_mem_gammaZero hf` gives properness and `asEReal_mem_gamma_of_mem_gammaZero hf` -- gives the `gamma H` hypothesis. theorem conjugate_mem_gammaZero_and_biconjugate_eq {f : H → Set.Ioi (⊥ : EReal)} (hf : f ∈ Γ₀(H)) : f∗[hf] ∈ Γ₀(H) ∧ f.asEReal∗∗ = f.asEReal

Corollary 13.39

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

Source header · Attribution source

Corollary 13.39: for f ∈ Γ₀(H) and g : H → [-∞,+∞], one has f ≤ g if and only if the Fenchel conjugates satisfy f* ≥ g*.

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] -- Proof sketch: the forward implication is Proposition 13.16(ii) applied to the canonical -- `EReal`-valued owner `f.asEReal`. For the reverse implication, apply Proposition 13.16(ii) to -- the conjugate inequality to get `f.asEReal∗∗ ≤ g∗∗`, rewrite the left-hand side using -- Corollary 13.38, and finish with Proposition 13.16(i), which gives `g∗∗ ≤ g`. theorem le_iff_conjugate_ge_of_mem_gammaZero {f : H → Set.Ioi (⊥ : EReal)} (hf : f ∈ Γ₀(H)) {g : H → EReal} : f.asEReal ≤ g ↔ f.asEReal∗ ≥ g∗

Corollary 13.40

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

Source header · Attribution source

Corollary 13.40: for functions f and g in Γ₀(H), one has f = g* if and only if g = f*, stated using the canonical Γ₀(H)-valued conjugate owner gammaZeroConjugate.

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] -- Proof sketch: `gammaZeroConjugate` packages Fenchel conjugation inside `Γ₀(H)`, so applying it -- twice yields the biconjugate. Corollary 13.38 identifies that biconjugate with the original -- function for every member of `Γ₀(H)`. theorem eq_conjugate_iff_eq_conjugate_of_mem_gammaZero (f g : H → Set.Ioi (⊥ : EReal)) (hf : f ∈ Γ₀(H)) (hg : g ∈ Γ₀(H)) : f = gammaZeroConjugate g hg ↔ g = gammaZeroConjugate f hf

Example 13.41

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

Source header · Attribution source

Example 13.41: the support function of the coordinate ℓ^p unit ball in ℝ^N is the coordinate ℓ^{p*} norm, written on the theorem surface with the Chapter 7 notation u ↦ ‖u‖_[p*] and viewed in EReal.

open scoped ENNReal InnerProductSpace open Set namespace ERealFunction -- Proof sketch: apply the Chapter 13 owner -- `conjugate_lpNorm_eq_indicator_lpClosedUnitBall_conjExponent` to `q = p.conjExponent`, rewrite -- `((ι[lpClosedUnitBall N p]).asEReal)∗` as `σ[lpClosedUnitBall N p]` using -- `conjugate_indicator_eq_supportFunction`, and then use -- `eq_conjugate_iff_eq_conjugate_of_mem_gammaZero` to swap the conjugation relation back to the -- Chapter 7 owner `EuclideanSpace.lpNorm`, written pointwise as `‖u‖_[p.conjExponent]`. theorem lpDualNorm_eq_lpNorm_conjExponent (N : ℕ) (p : ℝ≥0∞) (hp : 1 ≤ p) : σ[lpClosedUnitBall N p] = fun u ↦ (‖u‖_[p.conjExponent] : EReal)

Corollary 13.42

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

Source header · Attribution source

Corollary 13.42 in canonical support-function form: if f ∈ Γ₀(ℋ), then f is the support function of the epigraph of its Fenchel conjugate, evaluated along (x, -1).

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] theorem eq_supportFunction_epigraph_conjugate_of_mem_gammaZero {f : H → Set.Ioi (⊥ : EReal)} (hf : f ∈ Γ₀(H)) : f.asEReal = fun x ↦ σ[epigraph (f.asEReal∗)] (x, -1)

Example 13.43

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

Source header · Attribution source

Example 13.43: the Fenchel biconjugate of the indicator of C is the indicator of the closed convex hull of C.

open scoped InnerProductSpace Pointwise Set universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] -- Proof sketch: if `C = ∅`, then both sides are the constant function `⊤`. Otherwise apply the -- Fenchel--Moreau theorem to the proper indicator of `C`, then use Example 13.3(i) and -- Proposition 7.13 to identify the conjugate support function with the indicator of -- `closedConvexHull ℝ C`. theorem biconjugate_indicator_eq_indicator_closedConvexHull (C : Set H) : ((ι[C]).asEReal)∗∗ = (ι[closedConvexHull ℝ C]).asEReal

Proposition 13.44

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

Source header · Attribution source

Proposition 13.44, clause (i): for a convex extended-real-valued function, lower semicontinuity at x is equivalent to the lower semicontinuous convex envelope agreeing with f at x.

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] -- Proof sketch: Lemma 1.32 rewrites lower semicontinuity at `x` as -- `lowerSemicontinuousEnvelope f x = f x`, while Corollary 9.10 identifies -- `lowerSemicontinuousConvexEnvelope f` with `lowerSemicontinuousEnvelope f` for convex `f`. theorem lscAt_iff_lowerSemicontinuousConvexEnvelope_eq_self_of_convex {f : H → EReal} (hconv : IsConvex f) {x : H} : LowerSemicontinuousAt f x ↔ lowerSemicontinuousConvexEnvelope f x = f x

Proposition 13.45

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

Source header · Attribution source

Proposition 13.45: if the domain of the Fenchel conjugate f* is nonempty, equivalently if f admits a continuous affine minorant, then f** = \breve f; otherwise f** is identically -∞.

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] -- Proof sketch: if `(dom f∗).Nonempty`, then Proposition 13.12 supplies a continuous affine -- minorant of `f`, which makes `lowerSemicontinuousConvexEnvelope f` proper because it lies -- between that affine minorant and `f`. Proposition 13.16(iv) gives -- `conjugate (lowerSemicontinuousConvexEnvelope f) = conjugate f`, so taking conjugates again and -- applying Theorem 13.37 to the proper function `lowerSemicontinuousConvexEnvelope f` yields -- `f∗∗ = lowerSemicontinuousConvexEnvelope f`. variable [CompleteSpace H] theorem biconjugate_eq_lowerSemicontinuousConvexEnvelope_or_bot (f : H → EReal) : f∗∗ = if (dom f∗).Nonempty then lowerSemicontinuousConvexEnvelope f else ⊥

Proposition 13.46

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

Source header · Attribution source

Proposition 13.46 (1), left inclusion: if the Fenchel conjugate of an extended-real-valued function on a real Hilbert space has nonempty domain, then the domain of f is contained in the domain of f∗∗.

open Set open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] theorem dom_subset_dom_biconjugate_of_dom_conjugate_nonempty {f : H → EReal} (hdom : (dom f∗).Nonempty) : dom f ⊆ dom f∗∗

Proposition 13.47

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

Source header · Attribution source

Proposition 13.47: for a family of functions in Γ₀(H) whose pointwise supremum is not identically +∞, the Fenchel conjugate of that supremum is the lower semicontinuous convex envelope of the pointwise infimum of the Fenchel conjugates.

open scoped InnerProductSpace universe u v namespace ERealFunction variable {I : Type v} variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] -- Proof sketch: use the chapter owner `f.asEReal` for the `EReal` coercion of each `Γ₀(H)` -- function and its canonical conjugate `(f i).asEReal∗`. Corollary 13.38 identifies every -- `f i.asEReal` with its biconjugate, Proposition 13.28(i) gives -- `conjugate (⨅ i, (f i).asEReal∗) = ⨆ i, f i.asEReal`, and -- Proposition 13.45 turns the resulting biconjugate into the lower semicontinuous convex envelope -- under the hypothesis that the supremum is not identically `+∞`. theorem conjugate_iSup_eq_lowerSemicontinuousConvexEnvelope_iInf_conjugate_of_mem_gammaZero (f : I → H → Set.Ioi (⊥ : EReal)) (hf : ∀ i, f i ∈ Γ₀(H)) (hfinite : (⨆ i : I, (f i).asEReal) ≠ (⊤ : H → EReal)) : (⨆ i : I, (f i).asEReal)∗ = lowerSemicontinuousConvexEnvelope (⨅ i : I, (f i).asEReal∗)

Proposition 13.48

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

Source header · Attribution source

Proposition 13.48: if range L meets effectiveDomain g, then the Fenchel conjugate of g ∘ L is the infimal postcomposition of the Fenchel conjugate of g along L.adjoint.

open Set open scoped InnerProductSpace universe u v namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] [CompleteSpace H] variable {K : Type v} [NormedAddCommGroup K] [InnerProductSpace ℝ K] [CompleteSpace K] variable (g : K → Set.Ioi (⊥ : EReal)) (hg : g ∈ Γ₀(K)) (L : H →L[ℝ] K) -- Proof sketch: Proposition 13.24(5) gives the universal inequality -- `(g.asEReal ∘ L)∗ ≤ L.adjoint ▷ g.asEReal∗`, and the range-domain hypothesis is the textbook -- regularity condition under which that infimal postcomposition has no duality gap. theorem conjugate_comp_eq_adjointInfimalPostcomposition (hdom : (range L ∩ effectiveDomain g).Nonempty) : (g.asEReal ∘ L)∗ = L.adjoint ▷ g.asEReal∗

Proposition 13.49

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

Source header · Attribution source

Proposition 13.49 (1): for f ∈ Γ₀(ℋ), the recession function of the Fenchel conjugate f* coincides with the support function of the effective domain of f.

open scoped InnerProductSpace universe u namespace ERealFunction variable {H : Type u} [NormedAddCommGroup H] [InnerProductSpace ℝ H] -- Proof sketch: apply Proposition 13.11 to the support function of the epigraph of the -- `EReal`-valued coercion of `f`, identify the zero-height slice with the recession function of -- `f*`, and rewrite the remaining support term as the support function of `effectiveDomain f`. theorem recessionFunction_gammaZeroConjugate_eq_supportFunction_effectiveDomain (f : H → Set.Ioi (⊥ : EReal)) (hf : f ∈ Γ₀(H)) : (recessionFunction (gammaZeroConjugate f hf) (gammaZeroConjugate_mem_gammaZero hf).2.nonempty).asEReal = σ[effectiveDomain f]

Proposition 13.50

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

Source header · Attribution source

Proposition 13.50 (1): clause (i) for g = integralFunctional (gammaZeroConjugate φ hφ). Under the same hypotheses, the integral functional induced by the Fenchel conjugate 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] variable (φ : H → Set.Ioi (⊥ : EReal)) (hφ : φ ∈ Γ₀(H)) theorem integralFunctional_gammaZeroConjugate_mem_gammaZero (hfinite_or_nonneg : μ Set.univ < ∞ ∨ ((φ 0 : EReal) = 0 ∧ ∀ z : H, (φ 0 : EReal) ≤ (φ z : EReal))) : integralFunctional μ (gammaZeroConjugate φ hφ) ∈ Γ₀(Ω →₂[μ] H)

Example 13.51

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

Source header · Attribution source

Example 13.51: on a complete sigma-finite measure space, if f(X) = ∫ ω, |X ω|^p / p ∂P on L²((Ω,\mathcal F,P);\mathbb R) with p ∈ ]1,+∞[, then f* (X) = ∫ ω, |X ω|^{p*} / p* ∂P, where p* = p.conjExponent = p / (p - 1).

open MeasureTheory open scoped ERealFunction universe u namespace ERealFunction variable {Ω : Type u} [MeasurableSpace Ω] {P : Measure Ω} variable [P.IsComplete] [SigmaFinite P] -- Proof sketch: rewrite the canonical packaged conjugate `gammaZeroConjugate` of the scalar -- integrand via Example 13.2(i). theorem fenchelConjugate_integralFunctional_absPowerDivided_eq_integralFunctional_conjExponent (p : ℝ) (hp : 1 < p) : ((integralFunctional P ((fun t : ℝ ↦ |t| ^ p / p).toEReal)).asEReal)∗ = (integralFunctional P ((fun t : ℝ ↦ |t| ^ p.conjExponent / p.conjExponent).toEReal)).asEReal