Datasets:

Modalities:
Text
Languages:
English
Libraries:
Datasets
License:
Zhangir Azerbayev
squashed?
4365a98
raw
history blame
21.2 kB
/-
Copyright (c) 2019 Sébastien Gouëzel. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Sébastien Gouëzel
-/
import analysis.specific_limits.basic
import topology.metric_space.hausdorff_distance
import topology.sets.compacts
/-!
# Closed subsets
This file defines the metric and emetric space structure on the types of closed subsets and nonempty
compact subsets of a metric or emetric space.
The Hausdorff distance induces an emetric space structure on the type of closed subsets
of an emetric space, called `closeds`. Its completeness, resp. compactness, resp.
second-countability, follow from the corresponding properties of the original space.
In a metric space, the type of nonempty compact subsets (called `nonempty_compacts`) also
inherits a metric space structure from the Hausdorff distance, as the Hausdorff edistance is
always finite in this context.
-/
noncomputable theory
open_locale classical topological_space ennreal
universe u
open classical set function topological_space filter
namespace emetric
section
variables {α : Type u} [emetric_space α] {s : set α}
/-- In emetric spaces, the Hausdorff edistance defines an emetric space structure
on the type of closed subsets -/
instance closeds.emetric_space : emetric_space (closeds α) :=
{ edist := λs t, Hausdorff_edist (s : set α) t,
edist_self := λs, Hausdorff_edist_self,
edist_comm := λs t, Hausdorff_edist_comm,
edist_triangle := λs t u, Hausdorff_edist_triangle,
eq_of_edist_eq_zero :=
λ s t h, closeds.ext $ (Hausdorff_edist_zero_iff_eq_of_closed s.closed t.closed).1 h }
/-- The edistance to a closed set depends continuously on the point and the set -/
lemma continuous_inf_edist_Hausdorff_edist :
continuous (λ p : α × (closeds α), inf_edist p.1 p.2) :=
begin
refine continuous_of_le_add_edist 2 (by simp) _,
rintros ⟨x, s⟩ ⟨y, t⟩,
calc inf_edist x s ≤ inf_edist x t + Hausdorff_edist (t : set α) s :
inf_edist_le_inf_edist_add_Hausdorff_edist
... ≤ inf_edist y t + edist x y + Hausdorff_edist (t : set α) s :
add_le_add_right inf_edist_le_inf_edist_add_edist _
... = inf_edist y t + (edist x y + Hausdorff_edist (s : set α) t)
: by rw [add_assoc, Hausdorff_edist_comm]
... ≤ inf_edist y t + (edist (x, s) (y, t) + edist (x, s) (y, t)) :
add_le_add_left (add_le_add (le_max_left _ _) (le_max_right _ _)) _
... = inf_edist y t + 2 * edist (x, s) (y, t) :
by rw [← mul_two, mul_comm]
end
/-- Subsets of a given closed subset form a closed set -/
lemma is_closed_subsets_of_is_closed (hs : is_closed s) :
is_closed {t : closeds α | (t : set α) ⊆ s} :=
begin
refine is_closed_of_closure_subset (λt ht x hx, _),
-- t : closeds α, ht : t ∈ closure {t : closeds α | t ⊆ s},
-- x : α, hx : x ∈ t
-- goal : x ∈ s
have : x ∈ closure s,
{ refine mem_closure_iff.2 (λε εpos, _),
rcases mem_closure_iff.1 ht ε εpos with ⟨u, hu, Dtu⟩,
-- u : closeds α, hu : u ∈ {t : closeds α | t ⊆ s}, hu' : edist t u < ε
rcases exists_edist_lt_of_Hausdorff_edist_lt hx Dtu with ⟨y, hy, Dxy⟩,
-- y : α, hy : y ∈ u, Dxy : edist x y < ε
exact ⟨y, hu hy, Dxy⟩ },
rwa hs.closure_eq at this,
end
/-- By definition, the edistance on `closeds α` is given by the Hausdorff edistance -/
lemma closeds.edist_eq {s t : closeds α} : edist s t = Hausdorff_edist (s : set α) t := rfl
/-- In a complete space, the type of closed subsets is complete for the
Hausdorff edistance. -/
instance closeds.complete_space [complete_space α] : complete_space (closeds α) :=
begin
/- We will show that, if a sequence of sets `s n` satisfies
`edist (s n) (s (n+1)) < 2^{-n}`, then it converges. This is enough to guarantee
completeness, by a standard completeness criterion.
We use the shorthand `B n = 2^{-n}` in ennreal. -/
let B : ℕ → ℝ≥0∞ := λ n, (2⁻¹)^n,
have B_pos : ∀ n, (0:ℝ≥0∞) < B n,
by simp [B, ennreal.pow_pos],
have B_ne_top : ∀ n, B n ≠ ⊤,
by simp [B, ennreal.pow_ne_top],
/- Consider a sequence of closed sets `s n` with `edist (s n) (s (n+1)) < B n`.
We will show that it converges. The limit set is t0 = ⋂n, closure (⋃m≥n, s m).
We will have to show that a point in `s n` is close to a point in `t0`, and a point
in `t0` is close to a point in `s n`. The completeness then follows from a
standard criterion. -/
refine complete_of_convergent_controlled_sequences B B_pos (λs hs, _),
let t0 := ⋂ n, closure (⋃ m ≥ n, s m : set α),
let t : closeds α := ⟨t0, is_closed_Inter (λ_, is_closed_closure)⟩,
use t,
-- The inequality is written this way to agree with `edist_le_of_edist_le_geometric_of_tendsto₀`
have I1 : ∀ n, ∀ x ∈ s n, ∃ y ∈ t0, edist x y ≤ 2 * B n,
{ /- This is the main difficulty of the proof. Starting from `x ∈ s n`, we want
to find a point in `t0` which is close to `x`. Define inductively a sequence of
points `z m` with `z n = x` and `z m ∈ s m` and `edist (z m) (z (m+1)) ≤ B m`. This is
possible since the Hausdorff distance between `s m` and `s (m+1)` is at most `B m`.
This sequence is a Cauchy sequence, therefore converging as the space is complete, to
a limit which satisfies the required properties. -/
assume n x hx,
obtain ⟨z, hz₀, hz⟩ : ∃ z : Π l, s (n + l), (z 0 : α) = x ∧
∀ k, edist (z k:α) (z (k+1):α) ≤ B n / 2^k,
{ -- We prove existence of the sequence by induction.
have : ∀ l (z : s (n + l)), ∃ z' : s (n + l + 1), edist (z : α) z' ≤ B n / 2^l,
{ assume l z,
obtain ⟨z', z'_mem, hz'⟩ : ∃ z' ∈ s (n + l + 1), edist (z : α) z' < B n / 2^l,
{ refine exists_edist_lt_of_Hausdorff_edist_lt _ _,
{ exact s (n + l) },
{ exact z.2 },
simp only [B, ennreal.inv_pow, div_eq_mul_inv],
rw [← pow_add],
apply hs; simp },
exact ⟨⟨z', z'_mem⟩, le_of_lt hz'⟩ },
use [λ k, nat.rec_on k ⟨x, hx⟩ (λl z, some (this l z)), rfl],
exact λ k, some_spec (this k _) },
-- it follows from the previous bound that `z` is a Cauchy sequence
have : cauchy_seq (λ k, ((z k):α)),
from cauchy_seq_of_edist_le_geometric_two (B n) (B_ne_top n) hz,
-- therefore, it converges
rcases cauchy_seq_tendsto_of_complete this with ⟨y, y_lim⟩,
use y,
-- the limit point `y` will be the desired point, in `t0` and close to our initial point `x`.
-- First, we check it belongs to `t0`.
have : y ∈ t0 := mem_Inter.2 (λk, mem_closure_of_tendsto y_lim
begin
simp only [exists_prop, set.mem_Union, filter.eventually_at_top, set.mem_preimage,
set.preimage_Union],
exact ⟨k, λ m hm, ⟨n+m, zero_add k ▸ add_le_add (zero_le n) hm, (z m).2⟩⟩
end),
use this,
-- Then, we check that `y` is close to `x = z n`. This follows from the fact that `y`
-- is the limit of `z k`, and the distance between `z n` and `z k` has already been estimated.
rw [← hz₀],
exact edist_le_of_edist_le_geometric_two_of_tendsto₀ (B n) hz y_lim },
have I2 : ∀ n, ∀ x ∈ t0, ∃ y ∈ s n, edist x y ≤ 2 * B n,
{ /- For the (much easier) reverse inequality, we start from a point `x ∈ t0` and we want
to find a point `y ∈ s n` which is close to `x`.
`x` belongs to `t0`, the intersection of the closures. In particular, it is well
approximated by a point `z` in `⋃m≥n, s m`, say in `s m`. Since `s m` and
`s n` are close, this point is itself well approximated by a point `y` in `s n`,
as required. -/
assume n x xt0,
have : x ∈ closure (⋃ m ≥ n, s m : set α), by apply mem_Inter.1 xt0 n,
rcases mem_closure_iff.1 this (B n) (B_pos n) with ⟨z, hz, Dxz⟩,
-- z : α, Dxz : edist x z < B n,
simp only [exists_prop, set.mem_Union] at hz,
rcases hz with ⟨m, ⟨m_ge_n, hm⟩⟩,
-- m : ℕ, m_ge_n : m ≥ n, hm : z ∈ s m
have : Hausdorff_edist (s m : set α) (s n) < B n := hs n m n m_ge_n (le_refl n),
rcases exists_edist_lt_of_Hausdorff_edist_lt hm this with ⟨y, hy, Dzy⟩,
-- y : α, hy : y ∈ s n, Dzy : edist z y < B n
exact ⟨y, hy, calc
edist x y ≤ edist x z + edist z y : edist_triangle _ _ _
... ≤ B n + B n : add_le_add (le_of_lt Dxz) (le_of_lt Dzy)
... = 2 * B n : (two_mul _).symm ⟩ },
-- Deduce from the above inequalities that the distance between `s n` and `t0` is at most `2 B n`.
have main : ∀n:ℕ, edist (s n) t ≤ 2 * B n := λn, Hausdorff_edist_le_of_mem_edist (I1 n) (I2 n),
-- from this, the convergence of `s n` to `t0` follows.
refine tendsto_at_top.2 (λε εpos, _),
have : tendsto (λn, 2 * B n) at_top (𝓝 (2 * 0)),
from ennreal.tendsto.const_mul
(ennreal.tendsto_pow_at_top_nhds_0_of_lt_1 $ by simp [ennreal.one_lt_two])
(or.inr $ by simp),
rw mul_zero at this,
obtain ⟨N, hN⟩ : ∃ N, ∀ b ≥ N, ε > 2 * B b,
from ((tendsto_order.1 this).2 ε εpos).exists_forall_of_at_top,
exact ⟨N, λn hn, lt_of_le_of_lt (main n) (hN n hn)⟩
end
/-- In a compact space, the type of closed subsets is compact. -/
instance closeds.compact_space [compact_space α] : compact_space (closeds α) :=
begin
/- by completeness, it suffices to show that it is totally bounded,
i.e., for all ε>0, there is a finite set which is ε-dense.
start from a set `s` which is ε-dense in α. Then the subsets of `s`
are finitely many, and ε-dense for the Hausdorff distance. -/
refine compact_of_totally_bounded_is_closed (emetric.totally_bounded_iff.2 (λε εpos, _))
is_closed_univ,
rcases exists_between εpos with ⟨δ, δpos, δlt⟩,
rcases emetric.totally_bounded_iff.1
(compact_iff_totally_bounded_complete.1 (@compact_univ α _ _)).1 δ δpos with ⟨s, fs, hs⟩,
-- s : set α, fs : s.finite, hs : univ ⊆ ⋃ (y : α) (H : y ∈ s), eball y δ
-- we first show that any set is well approximated by a subset of `s`.
have main : ∀ u : set α, ∃v ⊆ s, Hausdorff_edist u v ≤ δ,
{ assume u,
let v := {x : α | x ∈ s ∧ ∃y∈u, edist x y < δ},
existsi [v, ((λx hx, hx.1) : v ⊆ s)],
refine Hausdorff_edist_le_of_mem_edist _ _,
{ assume x hx,
have : x ∈ ⋃y ∈ s, ball y δ := hs (by simp),
rcases mem_Union₂.1 this with ⟨y, ys, dy⟩,
have : edist y x < δ := by simp at dy; rwa [edist_comm] at dy,
exact ⟨y, ⟨ys, ⟨x, hx, this⟩⟩, le_of_lt dy⟩ },
{ rintros x ⟨hx1, ⟨y, yu, hy⟩⟩,
exact ⟨y, yu, le_of_lt hy⟩ }},
-- introduce the set F of all subsets of `s` (seen as members of `closeds α`).
let F := {f : closeds α | (f : set α) ⊆ s},
refine ⟨F, _, λ u _, _⟩,
-- `F` is finite
{ apply @finite.of_finite_image _ _ F coe,
{ apply fs.finite_subsets.subset (λb, _),
simp only [and_imp, set.mem_image, set.mem_set_of_eq, exists_imp_distrib],
assume x hx hx',
rwa hx' at hx },
{ exact set_like.coe_injective.inj_on F } },
-- `F` is ε-dense
{ obtain ⟨t0, t0s, Dut0⟩ := main u,
have : is_closed t0 := (fs.subset t0s).is_compact.is_closed,
let t : closeds α := ⟨t0, this⟩,
have : t ∈ F := t0s,
have : edist u t < ε := lt_of_le_of_lt Dut0 δlt,
apply mem_Union₂.2,
exact ⟨t, ‹t ∈ F›, this⟩ }
end
/-- In an emetric space, the type of non-empty compact subsets is an emetric space,
where the edistance is the Hausdorff edistance -/
instance nonempty_compacts.emetric_space : emetric_space (nonempty_compacts α) :=
{ edist := λ s t, Hausdorff_edist (s : set α) t,
edist_self := λs, Hausdorff_edist_self,
edist_comm := λs t, Hausdorff_edist_comm,
edist_triangle := λs t u, Hausdorff_edist_triangle,
eq_of_edist_eq_zero := λ s t h, nonempty_compacts.ext $ begin
have : closure (s : set α) = closure t := Hausdorff_edist_zero_iff_closure_eq_closure.1 h,
rwa [s.compact.is_closed.closure_eq, t.compact.is_closed.closure_eq] at this,
end }
/-- `nonempty_compacts.to_closeds` is a uniform embedding (as it is an isometry) -/
lemma nonempty_compacts.to_closeds.uniform_embedding :
uniform_embedding (@nonempty_compacts.to_closeds α _ _) :=
isometry.uniform_embedding $ λx y, rfl
/-- The range of `nonempty_compacts.to_closeds` is closed in a complete space -/
lemma nonempty_compacts.is_closed_in_closeds [complete_space α] :
is_closed (range $ @nonempty_compacts.to_closeds α _ _) :=
begin
have : range nonempty_compacts.to_closeds =
{s : closeds α | (s : set α).nonempty ∧ is_compact (s : set α) },
{ ext s,
refine ⟨_, λ h, ⟨⟨⟨s, h.2⟩, h.1⟩, closeds.ext rfl⟩⟩,
rintro ⟨s, hs, rfl⟩,
exact ⟨s.nonempty, s.compact⟩ },
rw this,
refine is_closed_of_closure_subset (λs hs, ⟨_, _⟩),
{ -- take a set set t which is nonempty and at a finite distance of s
rcases mem_closure_iff.1 hs ⊤ ennreal.coe_lt_top with ⟨t, ht, Dst⟩,
rw edist_comm at Dst,
-- since `t` is nonempty, so is `s`
exact nonempty_of_Hausdorff_edist_ne_top ht.1 (ne_of_lt Dst) },
{ refine compact_iff_totally_bounded_complete.2 ⟨_, s.closed.is_complete⟩,
refine totally_bounded_iff.2 (λε (εpos : 0 < ε), _),
-- we have to show that s is covered by finitely many eballs of radius ε
-- pick a nonempty compact set t at distance at most ε/2 of s
rcases mem_closure_iff.1 hs (ε/2) (ennreal.half_pos εpos.ne') with ⟨t, ht, Dst⟩,
-- cover this space with finitely many balls of radius ε/2
rcases totally_bounded_iff.1 (compact_iff_totally_bounded_complete.1 ht.2).1 (ε/2)
(ennreal.half_pos εpos.ne') with ⟨u, fu, ut⟩,
refine ⟨u, ⟨fu, λx hx, _⟩⟩,
-- u : set α, fu : u.finite, ut : t ⊆ ⋃ (y : α) (H : y ∈ u), eball y (ε / 2)
-- then s is covered by the union of the balls centered at u of radius ε
rcases exists_edist_lt_of_Hausdorff_edist_lt hx Dst with ⟨z, hz, Dxz⟩,
rcases mem_Union₂.1 (ut hz) with ⟨y, hy, Dzy⟩,
have : edist x y < ε := calc
edist x y ≤ edist x z + edist z y : edist_triangle _ _ _
... < ε/2 + ε/2 : ennreal.add_lt_add Dxz Dzy
... = ε : ennreal.add_halves _,
exact mem_bUnion hy this },
end
/-- In a complete space, the type of nonempty compact subsets is complete. This follows
from the same statement for closed subsets -/
instance nonempty_compacts.complete_space [complete_space α] :
complete_space (nonempty_compacts α) :=
(complete_space_iff_is_complete_range
nonempty_compacts.to_closeds.uniform_embedding.to_uniform_inducing).2 $
nonempty_compacts.is_closed_in_closeds.is_complete
/-- In a compact space, the type of nonempty compact subsets is compact. This follows from
the same statement for closed subsets -/
instance nonempty_compacts.compact_space [compact_space α] : compact_space (nonempty_compacts α) :=
begin
rw nonempty_compacts.to_closeds.uniform_embedding.embedding.is_compact_iff_is_compact_image,
rw [image_univ],
exact nonempty_compacts.is_closed_in_closeds.is_compact
end
/-- In a second countable space, the type of nonempty compact subsets is second countable -/
instance nonempty_compacts.second_countable_topology [second_countable_topology α] :
second_countable_topology (nonempty_compacts α) :=
begin
haveI : separable_space (nonempty_compacts α) :=
begin
/- To obtain a countable dense subset of `nonempty_compacts α`, start from
a countable dense subset `s` of α, and then consider all its finite nonempty subsets.
This set is countable and made of nonempty compact sets. It turns out to be dense:
by total boundedness, any compact set `t` can be covered by finitely many small balls, and
approximations in `s` of the centers of these balls give the required finite approximation
of `t`. -/
rcases exists_countable_dense α with ⟨s, cs, s_dense⟩,
let v0 := {t : set α | t.finite ∧ t ⊆ s},
let v : set (nonempty_compacts α) := {t : nonempty_compacts α | (t : set α) ∈ v0},
refine ⟨⟨v, _, _⟩⟩,
{ have : v0.countable, from countable_set_of_finite_subset cs,
exact this.preimage set_like.coe_injective },
{ refine λt, mem_closure_iff.2 (λε εpos, _),
-- t is a compact nonempty set, that we have to approximate uniformly by a a set in `v`.
rcases exists_between εpos with ⟨δ, δpos, δlt⟩,
have δpos' : 0 < δ / 2, from ennreal.half_pos δpos.ne',
-- construct a map F associating to a point in α an approximating point in s, up to δ/2.
have Exy : ∀x, ∃y, y ∈ s ∧ edist x y < δ/2,
{ assume x,
rcases mem_closure_iff.1 (s_dense x) (δ/2) δpos' with ⟨y, ys, hy⟩,
exact ⟨y, ⟨ys, hy⟩⟩ },
let F := λx, some (Exy x),
have Fspec : ∀x, F x ∈ s ∧ edist x (F x) < δ/2 := λx, some_spec (Exy x),
-- cover `t` with finitely many balls. Their centers form a set `a`
have : totally_bounded (t : set α) := t.compact.totally_bounded,
rcases totally_bounded_iff.1 this (δ/2) δpos' with ⟨a, af, ta⟩,
-- a : set α, af : a.finite, ta : t ⊆ ⋃ (y : α) (H : y ∈ a), eball y (δ / 2)
-- replace each center by a nearby approximation in `s`, giving a new set `b`
let b := F '' a,
have : b.finite := af.image _,
have tb : ∀ x ∈ t, ∃ y ∈ b, edist x y < δ,
{ assume x hx,
rcases mem_Union₂.1 (ta hx) with ⟨z, za, Dxz⟩,
existsi [F z, mem_image_of_mem _ za],
calc edist x (F z) ≤ edist x z + edist z (F z) : edist_triangle _ _ _
... < δ/2 + δ/2 : ennreal.add_lt_add Dxz (Fspec z).2
... = δ : ennreal.add_halves _ },
-- keep only the points in `b` that are close to point in `t`, yielding a new set `c`
let c := {y ∈ b | ∃ x ∈ t, edist x y < δ},
have : c.finite := ‹b.finite›.subset (λx hx, hx.1),
-- points in `t` are well approximated by points in `c`
have tc : ∀ x ∈ t, ∃ y ∈ c, edist x y ≤ δ,
{ assume x hx,
rcases tb x hx with ⟨y, yv, Dxy⟩,
have : y ∈ c := by simp [c, -mem_image]; exact ⟨yv, ⟨x, hx, Dxy⟩⟩,
exact ⟨y, this, le_of_lt Dxy⟩ },
-- points in `c` are well approximated by points in `t`
have ct : ∀ y ∈ c, ∃ x ∈ t, edist y x ≤ δ,
{ rintro y ⟨hy1, x, xt, Dyx⟩,
have : edist y x ≤ δ := calc
edist y x = edist x y : edist_comm _ _
... ≤ δ : le_of_lt Dyx,
exact ⟨x, xt, this⟩ },
-- it follows that their Hausdorff distance is small
have : Hausdorff_edist (t :set α) c ≤ δ :=
Hausdorff_edist_le_of_mem_edist tc ct,
have Dtc : Hausdorff_edist (t : set α) c < ε := this.trans_lt δlt,
-- the set `c` is not empty, as it is well approximated by a nonempty set
have hc : c.nonempty,
from nonempty_of_Hausdorff_edist_ne_top t.nonempty (ne_top_of_lt Dtc),
-- let `d` be the version of `c` in the type `nonempty_compacts α`
let d : nonempty_compacts α := ⟨⟨c, ‹c.finite›.is_compact⟩, hc⟩,
have : c ⊆ s,
{ assume x hx,
rcases (mem_image _ _ _).1 hx.1 with ⟨y, ⟨ya, yx⟩⟩,
rw ← yx,
exact (Fspec y).1 },
have : d ∈ v := ⟨‹c.finite›, this⟩,
-- we have proved that `d` is a good approximation of `t` as requested
exact ⟨d, ‹d ∈ v›, Dtc⟩ },
end,
apply uniform_space.second_countable_of_separable,
end
end --section
end emetric --namespace
namespace metric
section
variables {α : Type u} [metric_space α]
/-- `nonempty_compacts α` inherits a metric space structure, as the Hausdorff
edistance between two such sets is finite. -/
instance nonempty_compacts.metric_space : metric_space (nonempty_compacts α) :=
emetric_space.to_metric_space $ λ x y, Hausdorff_edist_ne_top_of_nonempty_of_bounded
x.nonempty y.nonempty x.compact.bounded y.compact.bounded
/-- The distance on `nonempty_compacts α` is the Hausdorff distance, by construction -/
lemma nonempty_compacts.dist_eq {x y : nonempty_compacts α} :
dist x y = Hausdorff_dist (x : set α) y := rfl
lemma lipschitz_inf_dist_set (x : α) : lipschitz_with 1 (λ s : nonempty_compacts α, inf_dist x s) :=
lipschitz_with.of_le_add $ assume s t,
by { rw dist_comm,
exact inf_dist_le_inf_dist_add_Hausdorff_dist (edist_ne_top t s) }
lemma lipschitz_inf_dist : lipschitz_with 2 (λ p : α × (nonempty_compacts α), inf_dist p.1 p.2) :=
@lipschitz_with.uncurry _ _ _ _ _ _ (λ (x : α) (s : nonempty_compacts α), inf_dist x s) 1 1
(λ s, lipschitz_inf_dist_pt s) lipschitz_inf_dist_set
lemma uniform_continuous_inf_dist_Hausdorff_dist :
uniform_continuous (λ p : α × (nonempty_compacts α), inf_dist p.1 p.2) :=
lipschitz_inf_dist.uniform_continuous
end --section
end metric --namespace