openAI-math-dataset / lean /OAI /Analysis /Brenier /InterceptContinuity.lean
roseking's picture
Add lean source and binary files (part 12)
9655094 verified
Raw History Blame Contribute Delete
12.7 kB
import OAI.Analysis.Brenier.FixedMass
namespace OAI
/-! Compactness and uniqueness imply continuous dependence of minimizers.
No differentiability or inverse-function theorem is used. -/
noncomputable section
open Set Filter
open scoped Topology
namespace Problem358.CompactArgmin
variable {P C I : Type*} [TopologicalSpace P] [TopologicalSpace C]
/-- A compactly constrained family of minimizers converges to the unique minimizer
of the limiting continuous objective. The feasible set may be noncompact. -/
theorem tendsto_of_unique_minimizer {l : Filter I} {p : I → P} {p₀ : P}
{c : I → C} {c₀ : C} {S K : Set C} {F : P × C → ℝ}
(hK : IsCompact K) (hS : IsClosed S) (hF : Continuous F)
(hp : Tendsto p l (𝓝 p₀)) (hbounded : ∀ᶠ i in l, c i ∈ K)
(hfeasible : ∀ᶠ i in l, c i ∈ S)
(hmin : ∀ᶠ i in l, ∀ z ∈ S, F (p i, c i) ≤ F (p i, z))
(hunique : ∀ z ∈ S, IsMinOn (fun w => F (p₀, w)) S z → z = c₀) :
Tendsto c l (𝓝 c₀) := by
apply hK.tendsto_nhds_of_unique_mapClusterPt hbounded
intro z _ hz
obtain ⟨U, hUl, hUz⟩ := mapClusterPt_iff_ultrafilter.mp hz
have hzS : z ∈ S := hS.mem_of_tendsto hUz (hUl hfeasible)
apply hunique z hzS
intro w hw
have hleft : Tendsto (fun i => F (p i, c i)) (U : Filter I) (𝓝 (F (p₀, z))) :=
hF.continuousAt.tendsto.comp ((hp.mono_left hUl).prodMk_nhds hUz)
have hright : Tendsto (fun i => F (p i, w)) (U : Filter I) (𝓝 (F (p₀, w))) :=
hF.continuousAt.tendsto.comp ((hp.mono_left hUl).prodMk_nhds tendsto_const_nhds)
apply le_of_tendsto_of_tendsto hleft hright
filter_upwards [hUl hmin] with i hi
exact hi w hw
/-- Local compact bounds and uniqueness give continuity of an arbitrary selected minimizer. -/
theorem continuousAt_unique_minimizer {c : P → C} {p : P}
{S K : Set C} {F : P × C → ℝ}
(hK : IsCompact K) (hS : IsClosed S) (hF : Continuous F)
(hbounded : ∀ᶠ q in 𝓝 p, c q ∈ K)
(hfeasible : ∀ᶠ q in 𝓝 p, c q ∈ S)
(hmin : ∀ᶠ q in 𝓝 p, ∀ z ∈ S, F (q, c q) ≤ F (q, z))
(hunique : ∀ z ∈ S, IsMinOn (fun w => F (p, w)) S z → z = c p) :
ContinuousAt c p :=
tendsto_of_unique_minimizer hK hS hF tendsto_id hbounded hfeasible hmin hunique
end Problem358.CompactArgmin
end
noncomputable section
open Set MeasureTheory
open scoped RealInnerProductSpace
namespace Problem358.MaxPotential
variable {ι : Type*} [Fintype ι] [Nonempty ι]
/-- Finite affine maxima depend jointly continuously on sites, intercepts, and source points. -/
theorem continuous_joint_potential {d : ℕ} :
Continuous (fun p : ((ι → E d) × (ι → ℝ)) × E d =>
potential p.1.1 p.1.2 p.2) := by
have h : Continuous (Finset.univ.sup' Finset.univ_nonempty
(fun i (p : ((ι → E d) × (ι → ℝ)) × E d) => ⟪p.2, p.1.1 i⟫ + p.1.2 i)) := by
apply Continuous.finset_sup'
intro i _
fun_prop
convert h using 1
funext p
simp only [potential, Finset.sup'_apply]
/-- The integrated finite affine objective is jointly continuous in all sites and intercepts. -/
theorem continuous_uniform_integral_potential {d : ℕ} {K : Set (E d)}
(hK : IsCompact K) (hne : (interior K).Nonempty) :
Continuous (fun p : (ι → E d) × (ι → ℝ) =>
∫ x, potential p.1 p.2 x ∂uniformMeasure K) := by
let := UniformSource.probability hK hne
have h := continuous_parametric_integral_of_continuous
(μ := uniformMeasure K)
(f := fun (p : (ι → E d) × (ι → ℝ)) (x : E d) => potential p.1 p.2 x)
(continuous_joint_potential (ι := ι) (d := d)) hK
rwa [Measure.restrict_eq_self_of_ae_mem (UniformSource.ae_mem hK.measurableSet)] at h
end Problem358.MaxPotential
end
noncomputable section
open Set MeasureTheory Filter
open scoped RealInnerProductSpace BigOperators Topology
namespace Problem358.MaxPotential
variable {ι : Type*} [Fintype ι] [Nonempty ι]
/-- Any normalized minimizing intercept selection depends continuously on the sites
at every injective configuration. This needs neither moving-cell derivatives nor IFT. -/
theorem continuousAt_uniform_intercept_selection {d : ℕ} {K : Set (E d)}
(hK : IsCompact K) (hconv : Convex ℝ K) (hne : (interior K).Nonempty)
(m : ι → ℝ) (hm : ∀ i, 0 < m i) (hsum : ∑ i, m i = 1)
(c : (ι → E d) → ι → ℝ)
(hc0 : ∀ y, ∑ i, m i * c y i = 0)
(hcmin : ∀ y (a : ι → ℝ), (∑ i, m i * a i) = 0 →
(∫ x, potential y (c y) x ∂uniformMeasure K) ≤ ∫ x, potential y a x ∂uniformMeasure K)
(y : ι → E d) (hy : Function.Injective y) : ContinuousAt c y := by
let := UniformSource.probability hK hne
let L : NNReal := ⟨‖y‖ + 1, by positivity⟩
let B : ℝ := 2 * (L : ℝ) * ∫ x, ‖x‖ ∂uniformMeasure K
let S : Set (ι → ℝ) := {a | ∑ i, m i * a i = 0}
let J : Set (ι → ℝ) := Set.Icc (fun i => -B / m i) (fun _ => B)
have hS : IsClosed S := isClosed_eq (by fun_prop) continuous_const
have hJ : IsCompact J := isCompact_Icc
have hnorm : ∀ᶠ z : ι → E d in 𝓝 y, ‖z‖ < (L : ℝ) := by
exact continuous_norm.continuousAt.eventually_lt continuousAt_const (by change ‖y‖ < ‖y‖ + 1; linarith)
have hbounded : ∀ᶠ z in 𝓝 y, c z ∈ J := by
filter_upwards [hnorm] with z hz
apply minimizer_mem_box (uniformMeasure K) z (c z) m L
(fun i => (norm_le_pi_norm z i).trans hz.le) hm hsum (hc0 z)
(UniformSource.integrable_norm hK)
exact hcmin z 0 (by simp)
apply CompactArgmin.continuousAt_unique_minimizer hJ hS
(continuous_uniform_integral_potential (ι := ι) hK hne) hbounded
(Filter.Eventually.of_forall hc0) (Filter.Eventually.of_forall hcmin)
intro a ha hamin
exact uniform_normalized_minimizer_unique hK hconv hne y hy m hm hsum
a (c y) ha (hc0 y) hamin (hcmin y)
/-- A globally defined normalized minimizer selection exists, and is continuous
on the open set of pairwise distinct sites. -/
theorem exists_continuous_uniform_intercept_selection {d : ℕ} {K : Set (E d)}
(hK : IsCompact K) (hconv : Convex ℝ K) (hne : (interior K).Nonempty)
(m : ι → ℝ) (hm : ∀ i, 0 < m i) (hsum : ∑ i, m i = 1) :
∃ c : (ι → E d) → ι → ℝ,
(∀ y, ∑ i, m i * c y i = 0) ∧
(∀ y (a : ι → ℝ), (∑ i, m i * a i) = 0 →
(∫ x, potential y (c y) x ∂uniformMeasure K) ≤ ∫ x, potential y a x ∂uniformMeasure K) ∧
ContinuousOn c {y | Function.Injective y} := by
classical
let := UniformSource.probability hK hne
have hex : ∀ y : ι → E d, ∃ c : ι → ℝ,
(∑ i, m i * c i) = 0 ∧
∀ a : ι → ℝ, (∑ i, m i * a i) = 0 →
(∫ x, potential y c x ∂uniformMeasure K) ≤ ∫ x, potential y a x ∂uniformMeasure K := by
intro y
obtain ⟨c, hc0, hcmin⟩ := exists_minimum_objective (uniformMeasure K) y m hm hsum
‖y‖₊ (fun i => norm_le_pi_norm y i) (UniformSource.integrable_norm hK)
refine ⟨c, hc0, ?_⟩
intro a ha
simpa only [objective, hc0, ha, sub_zero] using hcmin a ha
choose c hc0 hcmin using hex
refine ⟨c, hc0, hcmin, ?_⟩
intro y hy
exact (continuousAt_uniform_intercept_selection hK hconv hne m hm hsum c hc0 hcmin y hy).continuousWithinAt
end Problem358.MaxPotential
end
noncomputable section
open Set MeasureTheory Filter
open scoped BigOperators RealInnerProductSpace Topology
namespace Problem358.FixedMassIntercepts
variable {d : ℕ} {ι : Type*} [Fintype ι]
/-- Distinct finite site configurations form an open subset of the ambient coordinate space. -/
theorem isOpen_injective_sites : IsOpen {y : ι → E d | Function.Injective y} := by
have heq : {y : ι → E d | Function.Injective y} =
⋂ i : ι, ⋂ j : ι, {y | i ≠ j → y i ≠ y j} := by
ext y
simp only [mem_ofPred_eq, mem_iInter]
exact Function.injective_iff_pairwise_ne
rw [heq]
apply isOpen_iInter_of_finite
intro i
apply isOpen_iInter_of_finite
intro j
by_cases hij : i = j
· simp [hij]
· have hset : {y : ι → E d | i ≠ j → y i ≠ y j} =
{y : ι → E d | y i = y j}ᶜ := by
ext y
simp [hij]
rw [hset]
exact (isClosed_eq (show Continuous (fun y : ι → E d => y i) from continuous_apply i)
(show Continuous (fun y : ι → E d => y j) from continuous_apply j)).isOpen_compl
variable [Nonempty ι] [MeasurableSpace ι] [MeasurableSingletonClass ι]
/-- The actual prescribed intercepts minimize the integrated affine potential. -/
theorem intercepts_minimize {K : Set (E d)}
(hK : IsCompact K) (hconv : Convex ℝ K) (hne : (interior K).Nonempty)
(m : ι → ℝ) (hm : ∀ i, 0 < m i) (hsum : ∑ i, m i = 1)
(y : ι → E d) (hy : Function.Injective y) (a : ι → ℝ)
(ha : ∑ i, m i * a i = 0) :
(∫ x, MaxPotential.potential y (intercepts hK hconv hne m hm hsum y) x
∂uniformMeasure K) ≤ ∫ x, MaxPotential.potential y a x ∂uniformMeasure K := by
let := UniformSource.probability hK hne
let c := intercepts hK hconv hne m hm hsum y
obtain ⟨k, hk, _, hmax⟩ := exists_measurable_argmax
(fun i (x : E d) => ⟪x, y i⟫ + (c : ι → ℝ) i)
(fun i => ((continuous_id.inner continuous_const).add continuous_const).measurable)
apply MaxPotential.isMin_of_cell_masses (uniformMeasure K) y c m k c.property
(fun i => hk (measurableSet_singleton i)) (Eventually.of_forall hmax)
_ _ a ha
· intro i
exact (masses_eq_selector hK.measurableSet y hy c k hmax i).symm.trans
(congrFun (intercepts_spec hK hconv hne m hm hsum y hy) i)
· intro b
exact UniformSource.integrable_continuousOn hK
(MaxPotential.continuous_potential y b).continuousOn
/-- Canonical closed-cell intercepts are continuous at every distinct-site configuration.
This compactness argument is independent of the moving-boundary C1 theorem. -/
theorem continuousAt_intercepts_coe {K : Set (E d)}
(hK : IsCompact K) (hconv : Convex ℝ K) (hne : (interior K).Nonempty)
(m : ι → ℝ) (hm : ∀ i, 0 < m i) (hsum : ∑ i, m i = 1)
(y : ι → E d) (hy : Function.Injective y) :
ContinuousAt (fun z => (intercepts hK hconv hne m hm hsum z : ι → ℝ)) y := by
obtain ⟨b, hb0, hbmin, hbcont⟩ :=
MaxPotential.exists_continuous_uniform_intercept_selection hK hconv hne m hm hsum
have heq : ∀ z : ι → E d, Function.Injective z →
(intercepts hK hconv hne m hm hsum z : ι → ℝ) = b z := by
intro z hz
apply MaxPotential.uniform_normalized_minimizer_unique hK hconv hne z hz m hm hsum
_ _ (intercepts hK hconv hne m hm hsum z).property (hb0 z)
· exact intercepts_minimize hK hconv hne m hm hsum z hz
· exact hbmin z
have h_nhds : {z : ι → E d | Function.Injective z} ∈ 𝓝 y :=
isOpen_injective_sites.mem_nhds hy
apply (hbcont.continuousAt h_nhds).congr_of_eventuallyEq
filter_upwards [h_nhds] with z hz
exact heq z hz
/-- On repeated-site configurations the chosen extension is exactly zero. -/
theorem intercepts_eq_zero_of_not_injective {K : Set (E d)}
(hK : IsCompact K) (hconv : Convex ℝ K) (hne : (interior K).Nonempty)
(m : ι → ℝ) (hm : ∀ i, 0 < m i) (hsum : ∑ i, m i = 1)
(y : ι → E d) (hy : ¬Function.Injective y) :
intercepts hK hconv hne m hm hsum y = 0 := by
simp only [intercepts, dite_eq_right hy]
/-- The actual normalized intercept choice is globally Borel measurable, including
at repeated-site configurations, where its value was explicitly set to zero. -/
theorem measurable_intercepts_coe {K : Set (E d)}
(hK : IsCompact K) (hconv : Convex ℝ K) (hne : (interior K).Nonempty)
(m : ι → ℝ) (hm : ∀ i, 0 < m i) (hsum : ∑ i, m i = 1) :
Measurable (fun y => (intercepts hK hconv hne m hm hsum y : ι → ℝ)) := by
classical
let c : (ι → E d) → ι → ℝ := fun y => intercepts hK hconv hne m hm hsum y
have hc : ContinuousOn c {y | Function.Injective y} := fun y hy =>
(continuousAt_intercepts_coe hK hconv hne m hm hsum y hy).continuousWithinAt
have hmeas := hc.measurable_piecewise (continuous_const.continuousOn :
ContinuousOn (fun _ : ι → E d => (0 : ι → ℝ)) {y | Function.Injective y}ᶜ)
isOpen_injective_sites.measurableSet
have heq : ({y : ι → E d | Function.Injective y}.piecewise c (fun _ => 0)) = c := by
funext y
by_cases hy : Function.Injective y
· simp only [piecewise, mem_ofPred_eq, ite_eq_left hy]
· simp only [piecewise, mem_ofPred_eq, ite_eq_right hy]
symm
exact congrArg Subtype.val (intercepts_eq_zero_of_not_injective hK hconv hne m hm hsum y hy)
rwa [heq] at hmeas
end Problem358.FixedMassIntercepts
end
end OAI