math / data /lean /OAI /Dynamics /MultipleMixing /Entropy.lean
introvoyz045's picture
Migrated from GitHub
4a56cd6 verified
Raw History Blame Contribute Delete
75.3 kB
import OAI.Dynamics.MultipleMixing.Foundations namespace OAI noncomputable section namespace Rokhlin section open MeasureTheory Filter open scoped BigOperators Topology variable {A B C : Type*} [Fintype A] [Fintype B] [Fintype C] lemma shannonEntropy_eq_negMulLog (p : A → ℝ) : shannonEntropy p = ∑ a, Real.negMulLog (p a) := rfl def rowEntropy (p : A → ℝ) : ℝ := shannonEntropy p - Real.negMulLog (∑ a, p a) lemma rowEntropy_zero : rowEntropy (fun _ : A => 0) = 0 := by simp [rowEntropy,shannonEntropy,Real.negMulLog] lemma zero_row {p : A → ℝ} (hp : ∀ a, 0 ≤ p a) (h : ∑ a, p a = 0) (a : A) : p a = 0 := by exact le_antisymm (h ▸ Finset.single_le_sum (fun a _ => hp a) (Finset.mem_univ a)) (hp a) lemma rowEntropy_normalize (p : A → ℝ) (hp : ∀ a, 0 ≤ p a) : rowEntropy p = (∑ a, p a) * shannonEntropy (fun a => p a / ∑ a, p a) := by classical let w := ∑ a, p a have hw : 0 ≤ w := Finset.sum_nonneg (fun a _ => hp a) by_cases hz : w = 0 · have h0 : p = fun _ => 0 := funext (zero_row hp hz) subst p simp [rowEntropy_zero] · have hs : ∑ a, p a / w = 1 := by rw [←Finset.sum_div,div_self hz] have he (a : A) : Real.negMulLog (p a) = p a / w * Real.negMulLog w + w * Real.negMulLog (p a / w) := by conv_lhs => rw [show p a = w * (p a / w) by field_simp] exact Real.negMulLog_mul _ _ simp only [rowEntropy,shannonEntropy_eq_negMulLog] change (∑ a, Real.negMulLog (p a)) - Real.negMulLog w = _ simp_rw [he] rw [Finset.sum_add_distrib,←Finset.sum_mul,hs,one_mul,←Finset.mul_sum] ring lemma rowEntropy_sum_le (p : B → A → ℝ) (hp : ∀ b a, 0 ≤ p b a) : (∑ b, rowEntropy (p b)) ≤ rowEntropy (fun a => ∑ b, p b a) := by classical let w (b : B) := ∑ a, p b a let s := ∑ b, w b have hw (b : B) : 0 ≤ w b := Finset.sum_nonneg (fun a _ => hp b a) have hs : 0 ≤ s := Finset.sum_nonneg (fun b _ => hw b) by_cases hz : s = 0 · have hw0 b : w b = 0 := zero_row hw hz b have hp0 b a : p b a = 0 := zero_row (hp b) (hw0 b) a have hpz : p = fun _ _ => 0 := funext fun b => funext (hp0 b) rw [hpz] simp only [Finset.sum_const_zero,rowEntropy_zero,le_refl] · have hspos : 0 < s := lt_of_le_of_ne hs (Ne.symm hz) have hweights : ∑ b, w b / s = 1 := by rw [←Finset.sum_div,div_self hz] have he (a : A) : ∑ b, (w b / s) * (p b a / w b) = (∑ b, p b a) / s := by rw [Finset.sum_div] apply Finset.sum_congr rfl intro b _ by_cases hwb : w b = 0 · rw [zero_row (hp b) hwb a,hwb]; simp · field_simp have hj (a : A) : ∑ b, (w b / s) * Real.negMulLog (p b a / w b) ≤ Real.negMulLog ((∑ b, p b a) / s) := by have h := Real.concaveOn_negMulLog.le_map_sum (t := Finset.univ) (w := fun b => w b / s) (p := fun b => p b a / w b) (fun b _ => div_nonneg (hw b) hs) hweights (fun b _ => div_nonneg (hp b a) (hw b)) simpa only [smul_eq_mul,he] using h rw [rowEntropy_normalize _ (fun a => Finset.sum_nonneg (fun b _ => hp b a))] have hsum : (∑ a, ∑ b, p b a) = s := Finset.sum_comm rw [hsum] calc ∑ b, rowEntropy (p b) = s * ∑ a, ∑ b, (w b / s) * Real.negMulLog (p b a / w b) := by simp_rw [rowEntropy_normalize _ (hp _),shannonEntropy_eq_negMulLog] rw [Finset.sum_comm,Finset.mul_sum] apply Finset.sum_congr rfl intro b _ rw [←Finset.mul_sum,←mul_assoc,mul_div_cancel₀ _ hz] _ ≤ s * ∑ a, Real.negMulLog ((∑ b, p b a) / s) := mul_le_mul_of_nonneg_left (Finset.sum_le_sum (fun a _ => hj a)) hs _ = _ := rfl def conditionalShannon (p : B → A → ℝ) : ℝ := ∑ b, rowEntropy (p b) lemma shannon_chain_rule (p : B → A → ℝ) : shannonEntropy (fun z : B × A => p z.1 z.2) = shannonEntropy (fun b => ∑ a, p b a) + conditionalShannon p := by simp only [conditionalShannon,rowEntropy,shannonEntropy_eq_negMulLog,Fintype.sum_prod_type, Finset.sum_sub_distrib] ring lemma conditionalShannon_refine (p : C → B → A → ℝ) (hp : ∀ c b a, 0 ≤ p c b a) : conditionalShannon (fun z : C × B => p z.1 z.2) ≤ conditionalShannon (fun b a => ∑ c, p c b a) := by simp only [conditionalShannon,Fintype.sum_prod_type] rw [Finset.sum_comm] exact Finset.sum_le_sum fun b _ => rowEntropy_sum_le (fun c a => p c b a) (fun c a => hp c b a) end section open MeasureTheory Filter open scoped BigOperators Topology variable {Ω A B C : Type*} [MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ] [Fintype A] [Fintype B] [Fintype C] [MeasurableSpace A] [MeasurableSingletonClass A] [MeasurableSpace B] [MeasurableSingletonClass B] [MeasurableSpace C] [MeasurableSingletonClass C] def randomEntropy (F : Ω → A) : ℝ := shannonEntropy (fun a => μ.real (F ⁻¹' {a})) def randomConditionalEntropy (F : Ω → A) (G : Ω → B) : ℝ := conditionalShannon (fun b a => μ.real {x | G x = b ∧ F x = a}) omit [Fintype B] in lemma sum_joint_mass (F : Ω → A) (G : Ω → B) (hF : Measurable F) (hG : Measurable G) (b : B) : (∑ a, μ.real {x | G x = b ∧ F x = a}) = μ.real (G ⁻¹' {b}) := by rw [←measureReal_iUnion_fintype] · congr 1 ext x simp · intro a a' haa apply Set.disjoint_left.mpr intro x hx hx' exact haa (hx.2.symm.trans hx'.2) · intro a exact (hG (measurableSet_singleton b)).inter (hF (measurableSet_singleton a)) lemma randomEntropy_pair (F : Ω → A) (G : Ω → B) (hF : Measurable F) (hG : Measurable G) : randomEntropy μ (fun x => (G x,F x)) = randomEntropy μ G + randomConditionalEntropy μ F G := by have h := shannon_chain_rule (fun b a => μ.real {x | G x = b ∧ F x = a}) simp_rw [sum_joint_mass μ F G hF hG] at h simpa only [randomEntropy,randomConditionalEntropy,Set.preimage,Set.mem_singleton_iff,Prod.ext_iff] using h lemma randomEntropy_nonneg (F : Ω → A) (hF : Measurable F) : 0 ≤ randomEntropy μ F := by apply shannonEntropy_nonneg _ (fun _ => measureReal_nonneg) simpa using (sum_measureReal_preimage_singleton (μ := μ) Finset.univ (fun a _ => hF (measurableSet_singleton a))) lemma randomConditionalEntropy_nonneg (F : Ω → A) (G : Ω → B) (hF : Measurable F) (hG : Measurable G) : 0 ≤ randomConditionalEntropy μ F G := by have h := randomEntropy_comp_le (η := μ) (fun x => (G x,F x)) (hG.prodMk hF) Prod.fst change randomEntropy μ G ≤ randomEntropy μ (fun x => (G x,F x)) at h rw [randomEntropy_pair μ F G hF hG] at h linarith lemma randomConditionalEntropy_le (F : Ω → A) (G : Ω → B) (hF : Measurable F) (hG : Measurable G) : randomConditionalEntropy μ F G ≤ randomEntropy μ F := by have h := rowEntropy_sum_le (fun b a => μ.real {x | G x = b ∧ F x = a}) (fun _ _ => measureReal_nonneg) have hsum (a : A) : (∑ b, μ.real {x | G x = b ∧ F x = a}) = μ.real (F ⁻¹' {a}) := by have hh := sum_joint_mass μ G F hG hF a convert hh using 1 apply Finset.sum_congr rfl intro b _ congr 1 ext x exact and_comm simp_rw [hsum] at h have hmass : (∑ a, μ.real (F ⁻¹' {a})) = 1 := by simpa using (sum_measureReal_preimage_singleton (μ := μ) Finset.univ (fun a _ => hF (measurableSet_singleton a))) simpa only [randomConditionalEntropy,conditionalShannon, rowEntropy,hmass,Real.negMulLog_one, sub_zero,randomEntropy] using h lemma randomConditionalEntropy_refine (F : Ω → A) (G : Ω → B) (K : Ω → C) (hF : Measurable F) (hG : Measurable G) (hK : Measurable K) : randomConditionalEntropy μ F (fun x => (K x,G x)) ≤ randomConditionalEntropy μ F G := by have hp := conditionalShannon_refine (fun c b a => μ.real {x | K x = c ∧ G x = b ∧ F x = a}) (fun _ _ _ => measureReal_nonneg) have he (b : B) (a : A) : (∑ c, μ.real {x | K x = c ∧ G x = b ∧ F x = a}) = μ.real {x | G x = b ∧ F x = a} := by have hh := sum_joint_mass μ K (fun x => (G x,F x)) hK (hG.prodMk hF) (b,a) simpa only [Set.preimage,Set.mem_singleton_iff,Prod.mk.injEq,and_comm,and_left_comm] using hh simp_rw [he] at hp simpa only [randomConditionalEntropy,Prod.ext_iff,and_assoc] using hp lemma randomEntropy_prod_le (F : Ω → A) (G : Ω → B) (hF : Measurable F) (hG : Measurable G) : randomEntropy μ (fun x => (G x,F x)) ≤ randomEntropy μ G + randomEntropy μ F := by rw [randomEntropy_pair μ F G hF hG] exact add_le_add_right (randomConditionalEntropy_le μ F G hF hG) _ lemma randomEntropy_recode (F : Ω → A) (hF : Measurable F) (e : A ≃ B) : randomEntropy μ (e ∘ F) = randomEntropy μ F := by apply le_antisymm (randomEntropy_comp_le F hF e) have h := randomEntropy_comp_le (η := μ) (e ∘ F) ((measurable_of_finite e).comp hF) e.symm simpa only [Function.comp_def,Equiv.symm_apply_apply] using h end section open MeasureTheory Filter open scoped Topology BigOperators symmDiff variable {Ω A B : Type*} [MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ] [Fintype A] [Fintype B] [MeasurableSpace A] [MeasurableSingletonClass A] [MeasurableSpace B] [MeasurableSingletonClass B] omit [MeasurableSpace A] [MeasurableSingletonClass A] in lemma continuous_shannonEntropy : Continuous (shannonEntropy (A := A)) := by unfold shannonEntropy exact continuous_finsetSum _ fun a _ => Real.continuous_negMulLog.comp (continuous_apply a) omit [IsProbabilityMeasure μ] [MeasurableSpace A] [MeasurableSingletonClass A] in lemma randomEntropy_congr_ae {F G : Ω → A} (h : F =ᵐ[μ] G) : randomEntropy μ F = randomEntropy μ G := by unfold randomEntropy congr 1 funext a apply measureReal_congr filter_upwards [h] with x hx apply propext change F x = a ↔ G x = a rw [hx] omit [IsProbabilityMeasure μ] in lemma randomEntropy_measurePreserving {Υ : Type*} [MeasurableSpace Υ] (ν : Measure Υ) [IsProbabilityMeasure ν] {T : Ω → Υ} (hT : MeasurePreserving T μ ν) (F : Υ → A) (hF : Measurable F) : randomEntropy μ (F ∘ T) = randomEntropy ν F := by unfold randomEntropy congr 1 funext a exact hT.measureReal_preimage (hF (measurableSet_singleton a)).nullMeasurableSet omit [Fintype A] in lemma atom_mass_disagreement {F G : Ω → A} (hF : Measurable F) (hG : Measurable G) (a : A) : |μ.real (F ⁻¹' {a}) - μ.real (G ⁻¹' {a})| ≤ μ.real {x | F x ≠ G x} := by apply (abs_measureReal_sub_le_measureReal_symmDiff (hF (measurableSet_singleton a)).nullMeasurableSet (hG (measurableSet_singleton a)).nullMeasurableSet).trans refine measureReal_mono ?_ (by finiteness) intro x hx simp only [Set.mem_symmDiff,Set.mem_preimage,Set.mem_singleton_iff,Set.mem_ofPred_eq] at hx ⊢ rcases hx with hx | hx <;> aesop lemma randomEntropy_tendsto_of_disagreement {ι : Type*} {l : Filter ι} {Fn : ι → Ω → A} {F : Ω → A} (hFn : ∀ n, Measurable (Fn n)) (hF : Measurable F) (h : Tendsto (fun n => μ.real {x | Fn n x ≠ F x}) l (𝓝 0)) : Tendsto (fun n => randomEntropy μ (Fn n)) l (𝓝 (randomEntropy μ F)) := by apply continuous_shannonEntropy.tendsto _ |>.comp apply tendsto_pi_nhds.mpr intro a apply (tendsto_iff_norm_sub_tendsto_zero).mpr exact squeeze_zero_norm (fun n => by simpa only [Real.norm_eq_abs,abs_abs] using atom_mass_disagreement μ (hFn n) hF a) h lemma randomEntropy_self_pair (F : Ω → A) (hF : Measurable F) : randomEntropy μ (fun x => (F x,F x)) = randomEntropy μ F := by apply le_antisymm (randomEntropy_comp_le F hF (fun a => (a,a))) have h := randomEntropy_comp_le (η := μ) (fun x => (F x,F x)) (hF.prodMk hF) Prod.fst exact h lemma randomConditionalEntropy_self (F : Ω → A) (hF : Measurable F) : randomConditionalEntropy μ F F = 0 := by have h := randomEntropy_pair μ F F hF hF rw [randomEntropy_self_pair μ F hF] at h linarith lemma randomConditionalEntropy_tendsto_self {ι : Type*} {l : Filter ι} {Fn : ι → Ω → A} {F : Ω → A} (hFn : ∀ n, Measurable (Fn n)) (hF : Measurable F) (h : Tendsto (fun n => μ.real {x | Fn n x ≠ F x}) l (𝓝 0)) : Tendsto (fun n => randomConditionalEntropy μ F (Fn n)) l (𝓝 0) := by have hp : Tendsto (fun n => μ.real {x | (Fn n x,F x) ≠ (F x,F x)}) l (𝓝 0) := by have he : (fun n => μ.real {x | (Fn n x,F x) ≠ (F x,F x)}) = (fun n => μ.real {x | Fn n x ≠ F x}) := by funext n congr 1 ext x constructor · intro hn he exact hn (congrArg (fun a => (a,F x)) he) · intro hn he exact hn (congrArg Prod.fst he) rwa [he] have ht := (randomEntropy_tendsto_of_disagreement μ (fun n => (hFn n).prodMk hF) (hF.prodMk hF) hp).sub (randomEntropy_tendsto_of_disagreement μ hFn hF h) have he n : randomEntropy μ (fun x => (Fn n x,F x)) - randomEntropy μ (Fn n) = randomConditionalEntropy μ F (Fn n) := by rw [randomEntropy_pair μ F (Fn n) hF (hFn n)]; ring simpa only [he,randomEntropy_self_pair μ F hF,sub_self] using ht end section open MeasureTheory Filter open scoped Topology BigOperators variable {Ω A B C D : Type*} [MeasurableSpace Ω] (μ : Measure Ω) [retained0 : IsProbabilityMeasure μ] [Fintype A] [Fintype B] [Fintype C] [Fintype D] [retained1 : MeasurableSpace A] [retained2 : MeasurableSingletonClass A] [retained3 : MeasurableSpace B] [retained4 : MeasurableSingletonClass B] [MeasurableSpace C] [MeasurableSingletonClass C] [MeasurableSpace D] [MeasurableSingletonClass D] lemma randomEntropy_mutual_factor (F : Ω → A) (G : Ω → B) (hF : Measurable F) (hG : Measurable G) (e : A → B) (r : B → A) (he : G = e ∘ F) (hr : F = r ∘ G) : randomEntropy μ F = randomEntropy μ G := by apply le_antisymm · rw [hr] exact randomEntropy_comp_le G hG r · rw [he] exact randomEntropy_comp_le F hF e lemma randomEntropy_pair_factor (F : Ω → A) (hF : Measurable F) (e : A → B) : randomEntropy μ (fun x => (F x,e (F x))) = randomEntropy μ F := by apply randomEntropy_mutual_factor μ _ F (hF.prodMk ((measurable_of_finite e).comp hF)) hF Prod.fst (fun a => (a,e a)) rfl rfl lemma randomConditionalEntropy_condition_redundant (F : Ω → A) (G : Ω → B) (hF : Measurable F) (hG : Measurable G) (e : B → C) : randomConditionalEntropy μ F (fun x => (G x,e (G x))) = randomConditionalEntropy μ F G := by have hp := randomEntropy_pair μ F (fun x => (G x,e (G x))) hF (hG.prodMk ((measurable_of_finite e).comp hG)) have hq := randomEntropy_pair μ F G hF hG have hj : randomEntropy μ (fun x => ((G x,e (G x)),F x)) = randomEntropy μ (fun x => (G x,F x)) := by exact randomEntropy_mutual_factor μ _ _ ((hG.prodMk ((measurable_of_finite e).comp hG)).prodMk hF) (hG.prodMk hF) (fun z => (z.1.1,z.2)) (fun z => ((z.1,e z.1),z.2)) rfl rfl rw [hj,randomEntropy_pair_factor μ G hG e] at hp linarith lemma randomConditionalEntropy_condition_comp (F : Ω → A) (G : Ω → B) (hF : Measurable F) (hG : Measurable G) (e : B → C) : randomConditionalEntropy μ F G ≤ randomConditionalEntropy μ F (e ∘ G) := by have h := randomConditionalEntropy_refine μ F (e ∘ G) G hF ((measurable_of_finite e).comp hG) hG change randomConditionalEntropy μ F (fun x => (G x,e (G x))) ≤ _ at h rwa [randomConditionalEntropy_condition_redundant μ F G hF hG e] at h lemma randomConditionalEntropy_measurable_factor (F : Ω → A) (G : Ω → B) (K : Ω → C) (hF : Measurable F) (hK : Measurable K) [Nonempty B] (hG : Measurable[MeasurableSpace.comap K inferInstance] G) : randomConditionalEntropy μ F K ≤ randomConditionalEntropy μ F G := by obtain ⟨e,he⟩ := (Function.factorsThrough_iff G).mp hG.factorsThrough rw [he] exact randomConditionalEntropy_condition_comp μ F K hF hK e lemma randomConditionalEntropy_measurePreserving {Υ : Type*} [MeasurableSpace Υ] (ν : Measure Υ) [IsProbabilityMeasure ν] {T : Ω → Υ} (hT : MeasurePreserving T μ ν) (F : Υ → A) (G : Υ → B) (hF : Measurable F) (hG : Measurable G) : randomConditionalEntropy μ (F ∘ T) (G ∘ T) = randomConditionalEntropy ν F G := by have h := randomEntropy_pair μ (F ∘ T) (G ∘ T) (hF.comp hT.measurable) (hG.comp hT.measurable) have h' := randomEntropy_pair ν F G hF hG have hp := randomEntropy_measurePreserving μ ν hT (fun x => (G x,F x)) (hG.prodMk hF) have hg := randomEntropy_measurePreserving μ ν hT G hG change randomEntropy μ (fun x => ((G ∘ T) x,(F ∘ T) x)) = _ at hp rw [hp,hg] at h linarith include retained0 retained1 retained2 retained3 retained4 in lemma randomConditionalEntropy_congr_ae {F F' : Ω → A} {G G' : Ω → B} (hF : F =ᵐ[μ] F') (hG : G =ᵐ[μ] G') : randomConditionalEntropy μ F G = randomConditionalEntropy μ F' G' := by let _ := retained0 let _ := retained1 let _ := retained2 let _ := retained3 let _ := retained4 unfold randomConditionalEntropy congr 1 funext b a apply measureReal_congr filter_upwards [hF,hG] with x hx hy apply propext rw [hx,hy] end section open MeasureTheory Filter open scoped Topology BigOperators variable {Ω A B C : Type*} [MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ] [Fintype A] [Fintype B] [Fintype C] [MeasurableSpace A] [MeasurableSingletonClass A] [MeasurableSpace B] [MeasurableSingletonClass B] [MeasurableSpace C] [MeasurableSingletonClass C] lemma randomConditionalEntropy_chain (F : Ω → A) (G : Ω → B) (K : Ω → C) (hF : Measurable F) (hG : Measurable G) (hK : Measurable K) : randomConditionalEntropy μ (fun x => (G x,F x)) K = randomConditionalEntropy μ G K + randomConditionalEntropy μ F (fun x => (K x,G x)) := by have h1 := randomEntropy_pair μ (fun x => (G x,F x)) K (hG.prodMk hF) hK have h2 := randomEntropy_pair μ F (fun x => (K x,G x)) hF (hK.prodMk hG) have h3 := randomEntropy_pair μ G K hG hK have he : randomEntropy μ (fun x => (K x,(G x,F x))) = randomEntropy μ (fun x => ((K x,G x),F x)) := by exact randomEntropy_mutual_factor μ _ _ (hK.prodMk (hG.prodMk hF)) ((hK.prodMk hG).prodMk hF) (fun z => ((z.1,z.2.1),z.2.2)) (fun z => (z.1.1,(z.1.2,z.2))) rfl rfl linarith lemma randomConditionalEntropy_pair_le (F : Ω → A) (G : Ω → B) (K : Ω → C) (hF : Measurable F) (hG : Measurable G) (hK : Measurable K) : randomConditionalEntropy μ (fun x => (G x,F x)) K ≤ randomConditionalEntropy μ G K + randomConditionalEntropy μ F K := by rw [randomConditionalEntropy_chain μ F G K hF hG hK] have h := randomConditionalEntropy_condition_comp μ F (fun x => (K x,G x)) hF (hK.prodMk hG) Prod.fst exact add_le_add_right h _ lemma randomConditionalEntropy_mutual_factor (F : Ω → A) (G : Ω → B) (K : Ω → C) (hF : Measurable F) (hG : Measurable G) (hK : Measurable K) (e : A → B) (r : B → A) (he : G = e ∘ F) (hr : F = r ∘ G) : randomConditionalEntropy μ F K = randomConditionalEntropy μ G K := by have h1 := randomEntropy_pair μ F K hF hK have h2 := randomEntropy_pair μ G K hG hK have he' : (fun x => (K x,G x)) = (fun z : C × A => (z.1,e z.2)) ∘ (fun x => (K x,F x)) := by rw [he]; rfl have hr' : (fun x => (K x,F x)) = (fun z : C × B => (z.1,r z.2)) ∘ (fun x => (K x,G x)) := by rw [hr]; rfl have h := randomEntropy_mutual_factor μ _ _ (hK.prodMk hF) (hK.prodMk hG) (fun z : C × A => (z.1,e z.2)) (fun z : C × B => (z.1,r z.2)) he' hr' linarith omit [MeasurableSpace A] [MeasurableSingletonClass A] in lemma randomEntropy_fintype_one [Unique A] (F : Ω → A) : randomEntropy μ F = 0 := by classical have he : F = fun _ => default := funext fun x => Unique.eq_default _ rw [he] simp [randomEntropy,shannonEntropy] lemma randomConditionalEntropy_fintype_one [Unique A] (F : Ω → A) (G : Ω → B) (hF : Measurable F) (hG : Measurable G) : randomConditionalEntropy μ F G = 0 := by apply le_antisymm _ (randomConditionalEntropy_nonneg μ F G hF hG) have h := randomConditionalEntropy_le μ F G hF hG rwa [randomEntropy_fintype_one μ F] at h def integerWord (F : ℤ → Ω → A) (t : ℤ) (n : ℕ) : Ω → Fin n → A := fun x i => F (t+i.val) x omit [Fintype A] [MeasurableSingletonClass A] in lemma measurable_integerWord (F : ℤ → Ω → A) (hF : ∀ t, Measurable (F t)) (t : ℤ) (n : ℕ) : Measurable (integerWord F t n) := Measurable.of_eval fun _ => hF _ lemma integerWord_pair (F : ℤ → Ω → A) (hF : ∀ t, Measurable (F t)) (t : ℤ) (m n : ℕ) : randomEntropy μ (integerWord F t (m+n)) = randomEntropy μ (fun x => (integerWord F t m x,integerWord F (t+m) n x)) := by apply randomEntropy_mutual_factor μ _ _ (measurable_integerWord F hF _ _) ((measurable_integerWord F hF _ _).prodMk (measurable_integerWord F hF _ _)) (fun z => (fun i => z (i.castAdd n),fun i => z (i.natAdd m))) (fun z => Fin.append z.1 z.2) · funext x apply Prod.ext · rfl · funext i simp only [integerWord,Function.comp_apply,Fin.val_natAdd,Nat.cast_add] congr 1 ring · funext x i induction i using Fin.addCases with | left i => simp [integerWord] | right i => simp only [integerWord,Function.comp_apply,Fin.append_right,Fin.val_natAdd,Nat.cast_add]; congr 1; ring lemma integerWord_add (F : ℤ → Ω → A) (hF : ∀ t, Measurable (F t)) (t : ℤ) (m n : ℕ) : randomEntropy μ (integerWord F t (m+n)) = randomEntropy μ (integerWord F t m) + randomConditionalEntropy μ (integerWord F (t+m) n) (integerWord F t m) := by rw [integerWord_pair μ F hF t m n,randomEntropy_pair μ _ _ (measurable_integerWord F hF _ _) (measurable_integerWord F hF _ _)] lemma integerWord_succ (F : ℤ → Ω → A) (hF : ∀ t, Measurable (F t)) (t : ℤ) (n : ℕ) : randomEntropy μ (integerWord F t (n+1)) = randomEntropy μ (integerWord F t n) + randomConditionalEntropy μ (F (t+n)) (integerWord F t n) := by rw [integerWord_add μ F hF t n 1] congr 1 apply randomConditionalEntropy_mutual_factor μ _ _ _ (measurable_integerWord F hF _ _) (hF _) (measurable_integerWord F hF _ _) (fun z => z 0) (fun a _ => a) · funext x simp [integerWord] · funext x i have hi : i = 0 := Subsingleton.elim _ _ subst i simp only [integerWord,Fin.val_zero,Nat.cast_zero,add_zero,Function.comp_apply] end section open Filter open scoped Topology BigOperators theorem entropy_rate_of_decreasing_increments (H : ℕ → ℝ) (h0 : H 0 = 0) (hmono : Antitone (fun n => H (n+1) - H n)) (hpos : ∀ n, 0 ≤ H (n+1) - H n) : ∃ h : ℝ, 0 ≤ h ∧ (∀ n, h ≤ H (n+1) - H n) ∧ Tendsto (fun n : ℕ => H n / n) atTop (𝓝 h) := by let a : ℕ → ℝ := fun n => H (n+1) - H n have hb : BddBelow (Set.range a) := ⟨0,by rintro _ ⟨n,rfl⟩; exact hpos n⟩ refine ⟨⨅ n, a n,?_,fun n => ciInf_le hb n,?_⟩ · exact le_ciInf (fun n => hpos n) · have ht := (tendsto_atTop_ciInf hmono hb).cesaro have he n : (∑ j ∈ Finset.range n, a j) = H n := by dsimp [a] rw [Finset.sum_range_sub,h0,sub_zero] simpa only [a,Finset.sum_range_sub,h0,sub_zero,div_eq_mul_inv,mul_comm] using ht lemma entropy_block_lower (H : ℕ → ℝ) (h : ℝ) (ha : ∀ n, h ≤ H (n+1) - H n) (m n : ℕ) : (n : ℝ) * h ≤ H (m+n) - H m := by induction n with | zero => simp | succ n ih => have hh := ha (m+n) simp only [Nat.cast_add,Nat.cast_one] rw [show m + (n+1) = (m+n)+1 by omega] nlinarith theorem entropy_zero_of_tail_bounds (H K : ℕ → ℝ) (h : ℝ) (hrate : Tendsto (fun n : ℕ => H n / n) atTop (𝓝 h)) (hK : ∀ n, 0 ≤ K n) (hb : ∀ ε : ℝ, 0 < ε → ∃ c : ℕ, ∀ n : ℕ, K n + n * h ≤ H (n+c) + n * ε) : Tendsto (fun n : ℕ => K n / n) atTop (𝓝 0) := by apply tendsto_order.mpr constructor · intro a ha exact Eventually.of_forall fun n => ha.trans_le (div_nonneg (hK n) (Nat.cast_nonneg n)) · intro ε hε obtain ⟨c,hc⟩ := hb (ε/2) (by positivity) have hd : Tendsto (fun n : ℕ => (n+c : ℝ) / n) atTop (𝓝 1) := by have ht : Tendsto (fun n : ℕ => (c:ℝ) / n) atTop (𝓝 0) := tendsto_const_nhds.div_atTop (tendsto_natCast_atTop_atTop (R := ℝ)) have hh : Tendsto (fun n : ℕ => 1 + c / (n : ℝ)) atTop (𝓝 (1+0)) := tendsto_const_nhds.add ht simp only [add_zero] at hh apply hh.congr' filter_upwards [eventually_ge_atTop 1] with n hn have hn' : (n : ℝ) ≠ 0 := by exact_mod_cast (by omega : n ≠ 0) field_simp have hsh : Tendsto (fun n : ℕ => H (n+c) / n) atTop (𝓝 h) := by have hh := (hrate.comp (tendsto_add_atTop_nat c)).mul hd rw [mul_one] at hh apply hh.congr' filter_upwards [eventually_ge_atTop 1] with n hn have hn' : (n+c : ℝ) ≠ 0 := by positivity simp only [Function.comp_def,Nat.cast_add] field_simp have hu := (hsh.sub_const h).add_const (ε/2) simp only [sub_self,zero_add] at hu filter_upwards [hu.eventually (gt_mem_nhds (show ε/2 < ε by linarith)), eventually_ge_atTop 1] with n hh hn have hn' : (0 : ℝ) < n := by exact_mod_cast (by omega : 0 < n) apply lt_of_le_of_lt _ hh apply (div_le_iff₀ hn').mpr have hf := hc n have he : (H (n+c) / n - h + ε/2) * n = H (n+c) - n*h + n*(ε/2) := by field_simp rw [he] linarith end section open MeasureTheory Filter open scoped Topology BigOperators variable {Ω A : Type*} [retained5 : MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ] [Fintype A] [MeasurableSpace A] [MeasurableSingletonClass A] omit [MeasurableSpace Ω] [Fintype A] [MeasurableSpace A] [MeasurableSingletonClass A] in lemma integerWord_subword (F : ℤ → Ω → A) (s : ℤ) (l n m : ℕ) (h : l+n ≤ m) : integerWord F (s+l) n = (fun z : Fin m → A => fun i : Fin n => z ⟨l+i.val,by omega⟩) ∘ integerWord F s m := by funext x i simp only [integerWord,Function.comp_apply,Nat.cast_add] congr 1 ring omit [MeasurableSpace Ω] in lemma integerWord_subword_measurable (F : ℤ → Ω → A) (s : ℤ) (l n m : ℕ) (h : l+n ≤ m) : Measurable[MeasurableSpace.comap (integerWord F s m) inferInstance] (integerWord F (s+l) n) := by rw [integerWord_subword F s l n m h] exact (measurable_of_finite _).comp (comap_measurable _) include retained5 in lemma pastWord_comap_mono (F : ℤ → Ω → A) (t : ℤ) : Monotone (fun n : ℕ => MeasurableSpace.comap (integerWord F (t-n) n) inferInstance) := by let _ := retained5 intro m n hmn have hh := integerWord_subword_measurable F (t-n) (n-m) m n (by omega) have he : t - (n:ℤ) + (n-m:ℕ) = t-m := by rw [Nat.cast_sub hmn]; ring rw [he] at hh exact hh.comap_le include retained5 in lemma windowWord_comap_mono (F : ℤ → Ω → A) : Monotone (fun n : ℕ => MeasurableSpace.comap (integerWord F (-(n:ℤ)) (2*n+1)) inferInstance) := by let _ := retained5 intro m n hmn have hh := integerWord_subword_measurable F (-(n:ℤ)) (n-m) (2*m+1) (2*n+1) (by omega) have he : -(n:ℤ) + (n-m:ℕ) = -(m:ℤ) := by rw [Nat.cast_sub hmn]; ring rw [he] at hh exact hh.comap_le variable (U : ℤ → Ω → Ω) (hU : ∀ s, MeasurePreserving (U s) μ μ) (F : ℤ → Ω → A) (hF : ∀ t, Measurable (F t)) (htrans : ∀ t s x, F (t+s) x = F t (U s x)) omit [MeasurableSpace Ω] [Fintype A] [MeasurableSpace A] [MeasurableSingletonClass A] in include htrans in lemma integerWord_translate (t s : ℤ) (n : ℕ) : integerWord F (t+s) n = integerWord F t n ∘ U s := by funext x i change F ((t+s)+i.val) x = F (t+i.val) (U s x) rw [show (t+s)+(i.val:ℤ) = (t+i.val)+s by ring,htrans] include hU hF htrans lemma integerWord_entropy_translate (t : ℤ) (n : ℕ) : randomEntropy μ (integerWord F t n) = randomEntropy μ (integerWord F 0 n) := by rw [show t = 0+t by simp,integerWord_translate U F htrans] exact randomEntropy_measurePreserving μ μ (hU t) _ (measurable_integerWord F hF _ _) lemma conditionalEntropy_past_translate (t : ℤ) (n : ℕ) : randomConditionalEntropy μ (F t) (integerWord F (t-n) n) = randomConditionalEntropy μ (F 0) (integerWord F (-(n:ℤ)) n) := by have he : integerWord F (t-n) n = integerWord F (-(n:ℤ)) n ∘ U t := by rw [show t - (n:ℤ) = -(n:ℤ)+t by ring,integerWord_translate U F htrans] have hf : F t = F 0 ∘ U t := by funext x; simpa using htrans 0 t x rw [he,hf] exact randomConditionalEntropy_measurePreserving μ μ (hU t) _ _ (hF _) (measurable_integerWord F hF _ _) lemma entropy_increment_past (n : ℕ) : randomEntropy μ (integerWord F 0 (n+1)) - randomEntropy μ (integerWord F 0 n) = randomConditionalEntropy μ (F 0) (integerWord F (-(n:ℤ)) n) := by have hh := integerWord_succ μ F hF (-(n:ℤ)) n rw [integerWord_entropy_translate μ U hU F hF htrans (-(n:ℤ)) (n+1), integerWord_entropy_translate μ U hU F hF htrans (-(n:ℤ)) n] at hh simp only [neg_add_cancel] at hh linarith theorem stationary_entropy_rate [Nonempty A] : ∃ h : ℝ, 0 ≤ h ∧ (∀ (t : ℤ) (n : ℕ), h ≤ randomConditionalEntropy μ (F t) (integerWord F (t-n) n)) ∧ Tendsto (fun n : ℕ => randomEntropy μ (integerWord F 0 n) / n) atTop (𝓝 h) := by let H : ℕ → ℝ := fun n => randomEntropy μ (integerWord F 0 n) have h0 : H 0 = 0 := randomEntropy_fintype_one μ _ have he n : H (n+1) - H n = randomConditionalEntropy μ (F 0) (integerWord F (-(n:ℤ)) n) := entropy_increment_past μ U hU F hF htrans n have hmono : Antitone (fun n => H (n+1) - H n) := by intro m n hmn change H (n+1) - H n ≤ H (m+1) - H m rw [he,he] apply randomConditionalEntropy_measurable_factor μ _ _ _ (hF 0) (measurable_integerWord F hF _ _) have hh := (pastWord_comap_mono F 0 hmn) simp only [zero_sub] at hh exact (comap_measurable _).mono hh le_rfl have hpos n : 0 ≤ H (n+1)-H n := by rw [he] exact randomConditionalEntropy_nonneg μ _ _ (hF _) (measurable_integerWord F hF _ _) obtain ⟨h,hpos,hle,ht⟩ := entropy_rate_of_decreasing_increments H h0 hmono hpos refine ⟨h,hpos,?_,ht⟩ intro t n rw [conditionalEntropy_past_translate μ U hU F hF htrans,←he] exact hle n end section open MeasureTheory Filter Set variable {Ω : Type*} [m₀ : MeasurableSpace Ω] (μ : Measure Ω) @[instance_reducible] def completedFactor (m : MeasurableSpace Ω) : MeasurableSpace Ω where MeasurableSet' s := MeasurableSet[m₀] s ∧ ∃ t : Set Ω, MeasurableSet[m] t ∧ ∀ᵐ x ∂μ, x ∈ s ↔ x ∈ t measurableSet_empty := ⟨@MeasurableSet.empty Ω m₀,∅,@MeasurableSet.empty Ω m,Eventually.of_forall fun _ => Iff.rfl⟩ measurableSet_compl := by rintro s ⟨hs,t,ht,he⟩ exact ⟨hs.compl,tᶜ,ht.compl,he.mono fun x hx => not_congr hx⟩ measurableSet_iUnion := by intro s hs choose t ht he using fun n => (hs n).2 refine ⟨MeasurableSet.iUnion fun n => (hs n).1,⋃ n, t n,MeasurableSet.iUnion ht,?_⟩ filter_upwards [ae_all_iff.mpr he] with x hx simp only [mem_iUnion] exact exists_congr fun n => hx n lemma completedFactor_le (m : MeasurableSpace Ω) : completedFactor (m₀ := m₀) μ m ≤ m₀ := fun _ hs => hs.1 lemma le_completedFactor {m : MeasurableSpace Ω} (hm : m ≤ m₀) : m ≤ completedFactor (m₀ := m₀) μ m := fun s hs => ⟨hm s hs,s,hs,Eventually.of_forall fun _ => Iff.rfl⟩ lemma completedFactor_mono : Monotone (completedFactor (m₀ := m₀) μ) := by intro m n h s hs obtain ⟨hs,t,ht,he⟩ := hs exact ⟨hs,t,h t ht,he⟩ lemma completedFactor_congr_set (m : MeasurableSpace Ω) (s t : Set Ω) (hs : MeasurableSet[m₀] s) (ht : MeasurableSet[completedFactor (m₀ := m₀) μ m] t) (he : ∀ᵐ x ∂μ, x ∈ s ↔ x ∈ t) : MeasurableSet[completedFactor (m₀ := m₀) μ m] s := by obtain ⟨_,r,hr,her⟩ := ht exact ⟨hs,r,hr,he.and her |>.mono fun x hx => hx.1.trans hx.2⟩ lemma measurable_completedFactor_of_ae {Y : Type*} [TopologicalSpace Y] [TopologicalSpace.PseudoMetrizableSpace Y] [MeasurableSpace Y] [BorelSpace Y] (m : MeasurableSpace Ω) {f : Ω → Y} (hf : Measurable[m₀] f) (h : AEStronglyMeasurable[m] f μ) : Measurable[completedFactor (m₀ := m₀) μ m] f := by intro s hs exact ⟨hf hs,(h.mk f) ⁻¹' s,h.stronglyMeasurable_mk.measurable hs, h.ae_eq_mk.mono fun x hx => by simp only [mem_preimage,hx]⟩ lemma completedFactor_iInf_measurable {Y : Type*} [TopologicalSpace Y] [TopologicalSpace.PseudoMetrizableSpace Y] [MeasurableSpace Y] [BorelSpace Y] (m : ℕ → MeasurableSpace Ω) {f : Ω → Y} (hf : Measurable[m₀] f) (h : ∀ n, AEStronglyMeasurable[m n] f μ) : Measurable[⨅ n, completedFactor (m₀ := m₀) μ (m n)] f := by intro s hs exact MeasurableSpace.measurableSet_iInf.mpr fun n => measurable_completedFactor_of_ae μ (m n) hf (h n) hs end section open MeasureTheory Filter Set open scoped Topology BigOperators variable {Ω : Type*} [m₀ : MeasurableSpace Ω] (μ : Measure Ω) lemma aeStrong_limit_factor {E : Type*} [TopologicalSpace E] [Nonempty E] [TopologicalSpace.IsCompletelyMetrizableSpace E] (m : MeasurableSpace Ω) {f : ℕ → Ω → E} {g : Ω → E} (hf : ∀ n, AEStronglyMeasurable[m] (f n) μ) (ht : ∀ᵐ x ∂μ, Tendsto (fun n => f n x) atTop (𝓝 (g x))) : AEStronglyMeasurable[m] g μ := by let F : ℕ → Ω → E := fun n => (hf n).mk (f n) have hF : ∀ n, StronglyMeasurable[m] (F n) := fun n => (hf n).stronglyMeasurable_mk refine ⟨fun x => limUnder atTop (fun n => F n x),StronglyMeasurable.limUnder hF,?_⟩ filter_upwards [ht,ae_all_iff.mpr (fun n => (hf n).ae_eq_mk)] with x hx he have h : Tendsto (fun n => F n x) atTop (𝓝 (g x)) := by convert hx using 1 funext n exact (he n).symm exact h.limUnder_eq.symm lemma simpleFunc_aeStrong_of_completed (m : MeasurableSpace Ω) (g : @SimpleFunc Ω (completedFactor (m₀ := m₀) μ m) ℝ) : AEStronglyMeasurable[m] g μ := by classical have hsets : ∀ c : ℝ, ∃ t : Set Ω, MeasurableSet[m] t ∧ ∀ᵐ x ∂μ, g x = c ↔ x ∈ t := fun c => (@SimpleFunc.measurableSet_fiber Ω ℝ (completedFactor (m₀ := m₀) μ m) g c).2 choose t ht he using hsets let G : Ω → ℝ := fun x => ∑ c ∈ (@SimpleFunc.range Ω ℝ (completedFactor (m₀ := m₀) μ m) g), (t c).indicator (fun _ => c) x have hm : StronglyMeasurable[m] G := Finset.stronglyMeasurable_fun_sum _ fun c _ => stronglyMeasurable_const.indicator (ht c) refine ⟨G,hm,?_⟩ filter_upwards [ae_all_iff.mpr (fun c : (@SimpleFunc.range Ω ℝ (completedFactor (m₀ := m₀) μ m) g) => he c.val)] with x hx have hi (c : ℝ) (hc : c ∈ (@SimpleFunc.range Ω ℝ (completedFactor (m₀ := m₀) μ m) g)) : (t c).indicator (fun _ => c) x = if g x = c then c else 0 := by simp only [Set.indicator_apply] rw [←hx ⟨c,hc⟩] dsimp [G] simp_rw [Finset.sum_congr rfl hi] symm rw [Finset.sum_eq_single (g x)] · simp only [ite_true] · intro c hc hne exact ite_eq_right (Ne.symm hne) · intro hnot exact False.elim (hnot (@SimpleFunc.mem_range_self Ω ℝ (completedFactor (m₀ := m₀) μ m) g x)) lemma aeStrong_of_measurable_completed (m : MeasurableSpace Ω) {f : Ω → ℝ} (hf : Measurable[completedFactor (m₀ := m₀) μ m] f) : AEStronglyMeasurable[m] f μ := by have hs := hf.stronglyMeasurable exact aeStrong_limit_factor (m₀ := m₀) μ m (fun n => simpleFunc_aeStrong_of_completed (m₀ := m₀) μ m (hs.approx n)) (Eventually.of_forall hs.tendsto_approx) end section open MeasureTheory Filter Set open scoped Topology BigOperators symmDiff variable {Ω : Type*} [m₀ : MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ] lemma set_approx_iSup_completed (m : ℕ → MeasurableSpace Ω) (hm : ∀ n, m n ≤ m₀) (hmono : Monotone m) {s : Set Ω} (hs : MeasurableSet[completedFactor (m₀ := m₀) μ (⨆ n, m n)] s) (ε : ℝ) (hε : 0 < ε) : ∃ n t, MeasurableSet[m n] t ∧ μ.real (s ∆ t) < ε := by have hle : (⨆ n, m n) ≤ m₀ := iSup_le hm let C : Set (Set Ω) := {t | ∃ n, MeasurableSet[m n] t} have hCalg : IsSetAlgebra C := by refine ⟨⟨0,@MeasurableSet.empty Ω (m 0)⟩,?_,?_⟩ · rintro t ⟨n,hn⟩ exact ⟨n,hn.compl⟩ · rintro t t' ⟨n,hn⟩ ⟨k,hk⟩ exact ⟨max n k,((hmono (le_max_left n k)) _ hn).union ((hmono (le_max_right n k)) _ hk)⟩ have hCgen : (⨆ n, m n) = MeasurableSpace.generateFrom C := by apply le_antisymm · apply iSup_le intro n t ht exact MeasurableSpace.measurableSet_generateFrom ⟨n,ht⟩ · apply MeasurableSpace.generateFrom_le rintro t ⟨n,hn⟩ exact (le_iSup m n) _ hn have hdense := @Measure.MeasureDense.of_generateFrom_isSetAlgebra_finite Ω (⨆ n, m n) (μ.trim hle) C inferInstance hCalg hCgen obtain ⟨t,ht,he⟩ := hs.2 obtain ⟨r,⟨n,hn⟩,hr⟩ := @Measure.MeasureDense.approx Ω (⨆ n, m n) (μ.trim hle) C hdense t ht (by finiteness) ε hε have hrAll : MeasurableSet[(⨆ n, m n)] r := (le_iSup m n) _ hn rw [trim_measurableSet_eq hle (ht.symmDiff hrAll)] at hr have he' : s ∆ r =ᵐ[μ] t ∆ r := by filter_upwards [he] with x hx apply propext simp only [Set.mem_symmDiff,hx] refine ⟨n,r,hn,?_⟩ have her : μ.real (s ∆ r) = μ.real (t ∆ r) := measureReal_congr he' apply lt_of_eq_of_lt her have hh := ENNReal.toReal_lt_toReal (measure_ne_top μ (t ∆ r)) ENNReal.ofReal_ne_top |>.mpr hr simpa only [Measure.real,ENNReal.toReal_ofReal hε.le] using hh variable {B : Type*} [Fintype B] [Nonempty B] [MeasurableSpace B] [MeasurableSingletonClass B] def pickFiniteBit (v : B → Bool) : B := if h : ∃ b, v b = true then Classical.choose h else Classical.choice ‹Nonempty B› omit [MeasurableSpace B] [MeasurableSingletonClass B] in lemma pickFiniteBit_eq (b : B) (v : B → Bool) (hv : ∀ a, v a = true ↔ a = b) : pickFiniteBit v = b := by have he : ∃ a, v a = true := ⟨b,(hv b).mpr rfl⟩ simp only [pickFiniteBit,dite_eq_left he] exact (hv _).mp (Classical.choose_spec he) lemma finite_approx_iSup_completed (m : ℕ → MeasurableSpace Ω) (hm : ∀ n, m n ≤ m₀) (hmono : Monotone m) (F : Ω → B) (hF : Measurable[completedFactor (m₀ := m₀) μ (⨆ n, m n)] F) (ε : ℝ) (hε : 0 < ε) : ∃ n G, Measurable[m n] G ∧ μ.real {x | G x ≠ F x} < ε := by classical let δ := ε / (Fintype.card B + 1 : ℝ) have hδ : 0 < δ := div_pos hε (by positivity) have hsets (b : B) := set_approx_iSup_completed μ m hm hmono (hF (measurableSet_singleton b)) δ hδ choose n r hr hεr using hsets let N := Finset.univ.sup n have hrN (b : B) : MeasurableSet[m N] (r b) := (hmono (Finset.le_sup (f := n) (Finset.mem_univ b))) _ (hr b) let code : Ω → B → Bool := fun x b => if x ∈ r b then true else false have hcode : Measurable[m N] code := by refine @Measurable.of_eval Ω B (fun _ => Bool) (m N) _ code ?_ intro b exact Measurable.ite (hrN b) measurable_const measurable_const let G : Ω → B := pickFiniteBit ∘ code have hG : Measurable[m N] G := (measurable_of_finite pickFiniteBit).comp hcode have hsub : {x | G x ≠ F x} ⊆ ⋃ b, ((F ⁻¹' {b}) ∆ r b) := by intro x hx by_contra hn have hbits (b : B) : code x b = true ↔ b = F x := by have hh : x ∉ (F ⁻¹' {b}) ∆ r b := fun hh => hn (Set.mem_iUnion.mpr ⟨b,hh⟩) simp only [Set.mem_symmDiff,Set.mem_preimage,Set.mem_singleton_iff,not_or] at hh simp only [code,ite_eq_left_iff,Bool.false_eq_true,imp_false] aesop exact hx (pickFiniteBit_eq (F x) (code x) hbits) refine ⟨N,G,hG,lt_of_le_of_lt ((measureReal_mono hsub).trans (measureReal_iUnion_fintype_le _)) ?_⟩ calc (∑ b, μ.real ((F ⁻¹' {b}) ∆ r b)) ≤ ∑ _ : B, δ := Finset.sum_le_sum fun b _ => (hεr b).le _ = (Fintype.card B : ℝ) * δ := by simp _ < ε := by have hd : (0 : ℝ) < Fintype.card B + 1 := by positivity dsimp [δ] rw [← mul_div_assoc] exact (div_lt_iff₀ hd).mpr (by nlinarith) end section open MeasureTheory Filter open scoped Topology BigOperators variable {Ω A B : Type*} [m₀ : MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ] [Fintype A] [Fintype B] [MeasurableSpace A] [MeasurableSingletonClass A] [MeasurableSpace B] [MeasurableSingletonClass B] lemma randomConditionalEntropy_tendsto_condition {ι : Type*} {l : Filter ι} (F : Ω → A) (hF : Measurable F) {Gn : ι → Ω → B} {G : Ω → B} (hGn : ∀ n, Measurable (Gn n)) (hG : Measurable G) (h : Tendsto (fun n => μ.real {x | Gn n x ≠ G x}) l (𝓝 0)) : Tendsto (fun n => randomConditionalEntropy μ F (Gn n)) l (𝓝 (randomConditionalEntropy μ F G)) := by have hp : Tendsto (fun n => μ.real {x | (Gn n x,F x) ≠ (G x,F x)}) l (𝓝 0) := by have he : (fun n => μ.real {x | (Gn n x,F x) ≠ (G x,F x)}) = (fun n => μ.real {x | Gn n x ≠ G x}) := by funext n congr 1 ext x constructor · intro hn he exact hn (congrArg (fun a => (a,F x)) he) · intro hn he exact hn (congrArg Prod.fst he) rwa [he] have ht := (randomEntropy_tendsto_of_disagreement μ (fun n => (hGn n).prodMk hF) (hG.prodMk hF) hp).sub (randomEntropy_tendsto_of_disagreement μ hGn hG h) have he (H : Ω → B) (hH : Measurable H) : randomEntropy μ (fun x => (H x,F x)) - randomEntropy μ H = randomConditionalEntropy μ F H := by rw [randomEntropy_pair μ F H hF hH]; ring simpa only [he G hG,he _ (hGn _)] using ht lemma finite_approx_completed_sequence [Nonempty B] (m : ℕ → MeasurableSpace Ω) (hm : ∀ n, m n ≤ m₀) (hmono : Monotone m) (G : Ω → B) (hG : Measurable[completedFactor (m₀ := m₀) μ (⨆ n, m n)] G) : ∃ n : ℕ → ℕ, ∃ Gn : ℕ → Ω → B, (∀ l, Measurable[m (n l)] (Gn l)) ∧ Tendsto (fun l => μ.real {x | Gn l x ≠ G x}) atTop (𝓝 0) := by have hs (l : ℕ) := finite_approx_iSup_completed μ m hm hmono G hG (1 / (l+1 : ℝ)) (by positivity) choose n Gn hGn he using hs refine ⟨n,Gn,hGn,?_⟩ have ht : Tendsto (fun l : ℕ => 1 / (l+1 : ℝ)) atTop (𝓝 0) := by exact tendsto_const_nhds.div_atTop ((tendsto_natCast_atTop_atTop (R := ℝ)).atTop_add tendsto_const_nhds) exact squeeze_zero (fun _ => measureReal_nonneg) (fun l => (he l).le) ht lemma exists_conditional_entropy_lt_of_completed [Nonempty A] {C : ℕ → Type*} [∀ n, Fintype (C n)] [∀ n, MeasurableSpace (C n)] [∀ n, MeasurableSingletonClass (C n)] (K : ∀ n, Ω → C n) (hK : ∀ n, Measurable (K n)) (hmono : Monotone (fun n => MeasurableSpace.comap (K n) inferInstance)) (F : Ω → A) (hF : Measurable[completedFactor (m₀ := m₀) μ (⨆ n, MeasurableSpace.comap (K n) inferInstance)] F) (ε : ℝ) (hε : 0 < ε) : ∃ n, randomConditionalEntropy μ F (K n) < ε := by have hm n : MeasurableSpace.comap (K n) inferInstance ≤ m₀ := (hK n).comap_le obtain ⟨n,G,hG,ht⟩ := finite_approx_completed_sequence μ _ hm hmono F hF have hF0 : Measurable F := hF.mono (completedFactor_le μ _) le_rfl have hG0 l : Measurable (G l) := (hG l).mono (hm _) le_rfl have hcond := randomConditionalEntropy_tendsto_self μ hG0 hF0 ht obtain ⟨l,hl⟩ := (hcond.eventually (gt_mem_nhds hε)).exists refine ⟨n l,lt_of_le_of_lt ?_ hl⟩ exact randomConditionalEntropy_measurable_factor μ F (G l) (K (n l)) hF0 (hK _) (hG l) lemma conditional_entropy_lower_of_completed [Nonempty B] {C : ℕ → Type*} [∀ n, Fintype (C n)] [∀ n, MeasurableSpace (C n)] [∀ n, MeasurableSingletonClass (C n)] (K : ∀ n, Ω → C n) (hK : ∀ n, Measurable (K n)) (hmono : Monotone (fun n => MeasurableSpace.comap (K n) inferInstance)) (F : Ω → A) (hF : Measurable F) (G : Ω → B) (hG : Measurable[completedFactor (m₀ := m₀) μ (⨆ n, MeasurableSpace.comap (K n) inferInstance)] G) (r : ℝ) (hr : ∀ n, r ≤ randomConditionalEntropy μ F (K n)) : r ≤ randomConditionalEntropy μ F G := by have hm n : MeasurableSpace.comap (K n) inferInstance ≤ m₀ := (hK n).comap_le obtain ⟨n,Gn,hGn,ht⟩ := finite_approx_completed_sequence μ _ hm hmono G hG have hG0 : Measurable G := hG.mono (completedFactor_le μ _) le_rfl have hGn0 l : Measurable (Gn l) := (hGn l).mono (hm _) le_rfl have hcond := randomConditionalEntropy_tendsto_condition μ F hF hGn0 hG0 ht apply ge_of_tendsto hcond exact Eventually.of_forall fun l => (hr (n l)).trans (randomConditionalEntropy_measurable_factor μ F (Gn l) (K (n l)) hF (hK _) (hGn l)) end section open MeasureTheory Filter open scoped Topology BigOperators variable {Ω A B : Type*} [MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ] [Fintype A] [Fintype B] [MeasurableSpace A] [MeasurableSingletonClass A] [MeasurableSpace B] [MeasurableSingletonClass B] lemma integerWord_conditional_succ (F : ℤ → Ω → A) (hF : ∀ t, Measurable (F t)) (G : Ω → B) (hG : Measurable G) (t : ℤ) (n : ℕ) : randomConditionalEntropy μ (integerWord F t (n+1)) G = randomConditionalEntropy μ (integerWord F t n) G + randomConditionalEntropy μ (F (t+n)) (fun x => (G x,integerWord F t n x)) := by have he : randomConditionalEntropy μ (integerWord F t (n+1)) G = randomConditionalEntropy μ (fun x => (integerWord F t n x,F (t+n) x)) G := by apply randomConditionalEntropy_mutual_factor μ _ _ G (measurable_integerWord F hF _ _) ((measurable_integerWord F hF _ _).prodMk (hF _)) hG (fun z => (fun i : Fin n => z i.castSucc,z (Fin.last n))) (fun z => Fin.snoc z.1 z.2) · rfl · funext x i refine Fin.lastCases ?_ (fun j => ?_) i · simp only [Function.comp_apply,Fin.snoc_last,integerWord,Fin.val_last] · simp only [Function.comp_apply,Fin.snoc_castSucc,integerWord,Fin.val_castSucc] rw [he,randomConditionalEntropy_chain μ _ _ _ (hF _) (measurable_integerWord F hF _ _) hG] lemma integerWord_conditional_le_sum (F : ℤ → Ω → A) (hF : ∀ t, Measurable (F t)) (G : Ω → B) (hG : Measurable G) (t : ℤ) (n : ℕ) : randomConditionalEntropy μ (integerWord F t n) G ≤ ∑ j ∈ Finset.range n, randomConditionalEntropy μ (F (t+j)) G := by induction n with | zero => simp [randomConditionalEntropy_fintype_one μ _ G (measurable_integerWord F hF _ _) hG] | succ n ih => rw [integerWord_conditional_succ μ F hF G hG t n,Finset.sum_range_succ] exact add_le_add ih (randomConditionalEntropy_condition_comp μ (F (t+n)) (fun x => (G x,integerWord F t n x)) (hF _) (hG.prodMk (measurable_integerWord F hF _ _)) Prod.fst) lemma integerWord_conditional_lower (F : ℤ → Ω → A) (hF : ∀ t, Measurable (F t)) (G : Ω → B) (hG : Measurable G) (t : ℤ) (n : ℕ) (r : ℝ) (hr : ∀ j < n, r ≤ randomConditionalEntropy μ (F (t+j)) (fun x => (G x,integerWord F t j x))) : (n : ℝ) * r ≤ randomConditionalEntropy μ (integerWord F t n) G := by induction n with | zero => simpa only [Nat.cast_zero,zero_mul] using randomConditionalEntropy_nonneg μ _ _ (measurable_integerWord F hF t 0) hG | succ n ih => rw [integerWord_conditional_succ μ F hF G hG t n,Nat.cast_add,Nat.cast_one,add_mul,one_mul] exact add_le_add (ih (fun j hj => hr j (by omega))) (hr n (by omega)) end section open MeasureTheory Filter open scoped Topology BigOperators variable {Ω A B : Type*} [m₀ : MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ] [Fintype A] [Fintype B] [Nonempty A] [Nonempty B] [MeasurableSpace A] [MeasurableSingletonClass A] [MeasurableSpace B] [MeasurableSingletonClass B] abbrev processPast (F : ℤ → Ω → A) (t : ℤ) : MeasurableSpace Ω := ⨆ n : ℕ, MeasurableSpace.comap (integerWord F (t-n) n) inferInstance abbrev processFull (F : ℤ → Ω → A) : MeasurableSpace Ω := ⨆ n : ℕ, MeasurableSpace.comap (integerWord F (-(n:ℤ)) (2*n+1)) inferInstance omit [Fintype A] [Nonempty A] [MeasurableSingletonClass A] in lemma processPast_le (F : ℤ → Ω → A) (hF : ∀ t, Measurable (F t)) (t : ℤ) : processPast F t ≤ m₀ := iSup_le fun _n => (measurable_integerWord F hF _ _).comap_le omit [Fintype A] [Nonempty A] [MeasurableSingletonClass A] in lemma processFull_le (F : ℤ → Ω → A) (hF : ∀ t, Measurable (F t)) : processFull F ≤ m₀ := iSup_le fun _n => (measurable_integerWord F hF _ _).comap_le omit [Nonempty A] in include m₀ in lemma processPast_zero_le_full (F : ℤ → Ω → A) : processPast F 0 ≤ processFull F := by let _ := m₀ apply iSup_le intro n have hh := integerWord_subword_measurable F (-(n:ℤ)) 0 n (2*n+1) (by omega) simp only [Nat.cast_zero,add_zero] at hh have hh' : MeasurableSpace.comap (integerWord F (0-(n:ℤ)) n) inferInstance ≤ MeasurableSpace.comap (integerWord F (-(n:ℤ)) (2*n+1)) inferInstance := by simpa only [zero_sub] using hh.comap_le change _ ≤ ⨆ m : ℕ, MeasurableSpace.comap (integerWord F (-(m:ℤ)) (2*m+1)) inferInstance exact le_iSup_of_le n hh' variable (U : ℤ → Ω → Ω) (hU : ∀ s, MeasurePreserving (U s) μ μ) (F : ℤ → Ω → A) (hF : ∀ t, Measurable (F t)) (htrans : ∀ t s x, F (t+s) x = F t (U s x)) (G : ℤ → Ω → B) (hG : ∀ t, Measurable (G t)) (hGtrans : ∀ t s x, G (t+s) x = G t (U s x)) (htail : ∀ t s, Measurable[completedFactor (m₀ := m₀) μ (processPast F t)] (G s)) include htail hF in lemma tail_word_conditional_lower (n : ℕ) (r : ℝ) (hr : ∀ (t : ℤ) (m : ℕ), r ≤ randomConditionalEntropy μ (F t) (integerWord F (t-m) m)) : (n : ℝ) * r ≤ randomConditionalEntropy μ (integerWord F 0 n) (integerWord G 0 n) := by have hg s : Measurable (G s) := (htail 0 s).mono (completedFactor_le μ _) le_rfl apply integerWord_conditional_lower μ F hF (integerWord G 0 n) (measurable_integerWord G hg _ _) 0 n r intro j hj apply conditional_entropy_lower_of_completed μ (fun m => integerWord F ((0+(j:ℤ))-m) m) (fun m => measurable_integerWord F hF _ _) (pastWord_comap_mono F _) (F (0+(j:ℤ))) (hF _) _ ?_ r (hr _) apply Measurable.prodMk · exact @Measurable.of_eval Ω (Fin n) (fun _ => B) (completedFactor (m₀ := m₀) μ (processPast F (0+(j:ℤ)))) _ (integerWord G 0 n) (fun i => htail _ _) · have hh : MeasurableSpace.comap (integerWord F 0 j) inferInstance ≤ processPast F (0+(j:ℤ)) := by have hh := le_iSup (fun m : ℕ => MeasurableSpace.comap (integerWord F ((0+(j:ℤ))-m) m) inferInstance) j simpa only [processPast,zero_add,sub_self] using hh exact (comap_measurable _).mono (hh.trans (le_completedFactor μ (processPast_le F hF _))) le_rfl omit [Nonempty B] in include hU hF htrans hG hGtrans in lemma tail_word_conditional_upper (M n : ℕ) (ε : ℝ) (he : randomConditionalEntropy μ (G 0) (integerWord F (-(M:ℤ)) (2*M+1)) ≤ ε) : randomConditionalEntropy μ (integerWord G 0 n) (integerWord F (-(M:ℤ)) (n+2*M)) ≤ n*ε := by apply (integerWord_conditional_le_sum μ G hG _ (measurable_integerWord F hF _ _) 0 n).trans calc _ ≤ ∑ j ∈ Finset.range n, ε := by apply Finset.sum_le_sum intro j hj have hj' := Finset.mem_range.mp hj have hs := integerWord_subword_measurable F (-(M:ℤ)) j (2*M+1) (n+2*M) (by omega) have hc := randomConditionalEntropy_measurable_factor μ (G (0+(j:ℤ))) (integerWord F (-(M:ℤ)+(j:ℤ)) (2*M+1)) (integerWord F (-(M:ℤ)) (n+2*M)) (hG _) (measurable_integerWord F hF _ _) hs have hf : G (0+(j:ℤ)) = G 0 ∘ U j := funext fun x => hGtrans 0 j x rw [integerWord_translate U F htrans,hf, randomConditionalEntropy_measurePreserving μ μ (hU j) _ _ (hG _) (measurable_integerWord F hF _ _)] at hc simpa only [hf] using hc.trans he _ = n*ε := by simp include hU hF htrans hG hGtrans htail in theorem tail_process_entropy_zero : Tendsto (fun n : ℕ => randomEntropy μ (integerWord G 0 n) / n) atTop (𝓝 0) := by obtain ⟨h,hpos,hlower,hrate⟩ := stationary_entropy_rate μ U hU F hF htrans apply entropy_zero_of_tail_bounds (fun n => randomEntropy μ (integerWord F 0 n)) (fun n => randomEntropy μ (integerWord G 0 n)) h hrate (fun n => randomEntropy_nonneg μ _ (measurable_integerWord G hG _ _)) intro ε hε have hfull : Measurable[completedFactor (m₀ := m₀) μ (processFull F)] (G 0) := (htail 0 0).mono (completedFactor_mono μ (processPast_zero_le_full F)) le_rfl obtain ⟨M,hM⟩ := exists_conditional_entropy_lt_of_completed μ (fun m => integerWord F (-(m:ℤ)) (2*m+1)) (fun m => measurable_integerWord F hF _ _) (windowWord_comap_mono F) (G 0) hfull ε hε refine ⟨2*M,fun n => ?_⟩ have hlo := tail_word_conditional_lower μ F hF G htail n h hlower have hup := tail_word_conditional_upper μ U hU F hF htrans G hG hGtrans M n ε hM.le have he : randomEntropy μ (fun x => (integerWord G 0 n x,integerWord F 0 n x)) ≤ randomEntropy μ (fun x => (integerWord F (-(M:ℤ)) (n+2*M) x,integerWord G 0 n x)) := by let e : (Fin (n+2*M) → A) × (Fin n → B) → (Fin n → B) × (Fin n → A) := fun z => (z.2,fun i => z.1 ⟨M+i.val,by omega⟩) have heq : (fun x => (integerWord G 0 n x,integerWord F 0 n x)) = e ∘ (fun x => (integerWord F (-(M:ℤ)) (n+2*M) x,integerWord G 0 n x)) := by funext x apply Prod.ext · rfl · funext i dsimp [e,integerWord] congr 1 ring rw [heq] exact randomEntropy_comp_le _ ((measurable_integerWord F hF _ _).prodMk (measurable_integerWord G hG _ _)) e rw [randomEntropy_pair μ _ _ (measurable_integerWord F hF _ _) (measurable_integerWord G hG _ _), randomEntropy_pair μ _ _ (measurable_integerWord G hG _ _) (measurable_integerWord F hF _ _), integerWord_entropy_translate μ U hU F hF htrans] at he linarith end section open Filter open scoped Topology variable {𝕜 H : Type*} [RCLike 𝕜] [NormedAddCommGroup H] [InnerProductSpace 𝕜 H] [CompleteSpace H] local instance complete_iInf (K : ℕ → Submodule 𝕜 H) [∀ n, CompleteSpace (K n)] : CompleteSpace ↥(⨅ n, K n) := by have h : IsClosed (↑(⨅ n, K n) : Set H) := by simp only [Submodule.coe_iInf] exact isClosed_iInter fun n => (completeSpace_coe_iff_isComplete.mp (inferInstance : CompleteSpace (K n))).isClosed exact h.completeSpace_coe lemma antitone_projection_tendsto (K : ℕ → Submodule 𝕜 H) [∀ n, CompleteSpace (K n)] (hK : Antitone K) (x : H) : Tendsto (fun n => (K n).starProjection x) atTop (𝓝 ((⨅ n, K n).starProjection x)) := by have he : (⨆ n, (K n)ᗮ)ᗮ = ⨅ n, K n := by apply le_antisymm · apply le_iInf intro n simpa only [Submodule.orthogonal_orthogonal] using (Submodule.orthogonal_le (le_iSup (fun n => (K n)ᗮ) n)) · intro x hx y hy have hy' : y ∈ (⨅ n, K n)ᗮ := (iSup_le (fun n => Submodule.orthogonal_le (iInf_le K n))) hy exact inner_eq_zero_symm.mp (hy' x hx) have he' : (⨆ n, (K n)ᗮ).topologicalClosure = (⨅ n, K n)ᗮ := by rw [←Submodule.orthogonal_orthogonal_eq_closure,he] have ht := Submodule.starProjection_tendsto_closure_iSup (fun n => (K n)ᗮ) (fun a b h => Submodule.orthogonal_le (hK h)) x simp only [he'] at ht have hh := (tendsto_const_nhds (x := x) (f := atTop)).sub ht simpa only [Submodule.starProjection_orthogonal_val,sub_sub_cancel] using hh end section open MeasureTheory Filter open scoped ENNReal Topology variable {Ω : Type*} [m₀ : MeasurableSpace Ω] (μ : Measure Ω) lemma lpMeas_completed_eq (m : MeasurableSpace Ω) (hm : m ≤ m₀) : @lpMeas Ω ℝ ℝ _ _ _ (completedFactor (m₀ := m₀) μ m) m₀ 2 μ = @lpMeas Ω ℝ ℝ _ _ _ m m₀ 2 μ := by ext f simp only [mem_lpMeas_iff_aestronglyMeasurable] constructor · intro h have hm' : Measurable[completedFactor (m₀ := m₀) μ m] (h.mk f) := h.stronglyMeasurable_mk.measurable exact (aeStrong_of_measurable_completed (m₀ := m₀) μ m hm').congr h.ae_eq_mk.symm · exact fun h => h.mono (le_completedFactor (m₀ := m₀) μ hm) lemma lpMeas_completed_iInf (m : ℕ → MeasurableSpace Ω) (hm : ∀ n, m n ≤ m₀) : @lpMeas Ω ℝ ℝ _ _ _ (⨅ n, completedFactor (m₀ := m₀) μ (m n)) m₀ 2 μ = ⨅ n, @lpMeas Ω ℝ ℝ _ _ _ (m n) m₀ 2 μ := by ext f simp only [Submodule.mem_iInf,mem_lpMeas_iff_aestronglyMeasurable] constructor · intro h n have h' := h.mono (iInf_le (fun n => completedFactor (m₀ := m₀) μ (m n)) n) have hh : f ∈ @lpMeas Ω ℝ ℝ _ _ _ (completedFactor (m₀ := m₀) μ (m n)) m₀ 2 μ := by exact mem_lpMeas_iff_aestronglyMeasurable.mpr h' rw [lpMeas_completed_eq μ (m n) (hm n)] at hh exact mem_lpMeas_iff_aestronglyMeasurable.mp hh · intro h exact (completedFactor_iInf_measurable μ m (Lp.stronglyMeasurable f).measurable h).stronglyMeasurable.aestronglyMeasurable lemma reverse_conditional_L2 (m : ℕ → MeasurableSpace Ω) (hm : ∀ n, m n ≤ m₀) (hanti : Antitone m) (f : Lp ℝ 2 μ) : Tendsto (fun n => (condExpL2 ℝ ℝ (hm n) f : Lp ℝ 2 μ)) atTop (𝓝 (condExpL2 ℝ ℝ (show (⨅ n, completedFactor (m₀ := m₀) μ (m n)) ≤ m₀ from (iInf_le _ 0).trans (completedFactor_le μ (m 0))) f : Lp ℝ 2 μ)) := by let K : ℕ → Submodule ℝ (Lp ℝ 2 μ) := fun n => @lpMeas Ω ℝ ℝ _ _ _ (m n) m₀ 2 μ let (n : ℕ) : Fact (m n ≤ m₀) := ⟨hm n⟩ have hK : Antitone K := by intro a b hab f hf exact mem_lpMeas_iff_aestronglyMeasurable.mpr ((mem_lpMeas_iff_aestronglyMeasurable.mp hf).mono (hanti hab)) have h := antitone_projection_tendsto K hK f have ht : Fact ((⨅ n, completedFactor (m₀ := m₀) μ (m n)) ≤ m₀) := ⟨(iInf_le _ 0).trans (completedFactor_le μ (m 0))⟩ let := ht simpa only [K, ←lpMeas_completed_iInf μ m hm, Submodule.starProjection_apply,condExpL2] using h end section open MeasureTheory Filter open scoped Topology RealInnerProductSpace variable {Ω : Type*} [m₀ : MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ] omit [IsProbabilityMeasure μ] in lemma reverse_tail_pairing (m : ℕ → MeasurableSpace Ω) (hm : ∀ n, m n ≤ m₀) (hanti : Antitone m) (f : Lp ℝ 2 μ) (g : ℕ → Lp ℝ 2 μ) (hg : ∀ n, AEStronglyMeasurable[m n] (g n) μ) (hb : ∀ n, ‖g n‖ ≤ 1) : Tendsto (fun n => ⟪f - (condExpL2 ℝ ℝ (show (⨅ n, completedFactor (m₀ := m₀) μ (m n)) ≤ m₀ from (iInf_le _ 0).trans (completedFactor_le μ (m 0))) f : Lp ℝ 2 μ), g n⟫) atTop (𝓝 0) := by let ft : Lp ℝ 2 μ := condExpL2 ℝ ℝ (show (⨅ n, completedFactor (m₀ := m₀) μ (m n)) ≤ m₀ from (iInf_le _ 0).trans (completedFactor_le μ (m 0))) f have he n : ⟪f-ft,g n⟫ = ⟪(condExpL2 ℝ ℝ (hm n) f : Lp ℝ 2 μ)-ft,g n⟫ := by simp only [inner_sub_left,inner_condExpL2_eq_inner_fun (hm n) f (g n) (hg n)] have hbound n : |⟪f-ft,g n⟫| ≤ ‖(condExpL2 ℝ ℝ (hm n) f : Lp ℝ 2 μ)-ft‖ := by rw [he] exact (abs_real_inner_le_norm _ _).trans (by nlinarith [norm_nonneg ((condExpL2 ℝ ℝ (hm n) f : Lp ℝ 2 μ)-ft),hb n]) have ht : Tendsto (fun n => ‖(condExpL2 ℝ ℝ (hm n) f : Lp ℝ 2 μ)-ft‖) atTop (𝓝 0) := by simpa only [ft,sub_self,norm_zero] using ((reverse_conditional_L2 μ m hm hanti f).sub (tendsto_const_nhds (x := ft))).norm exact squeeze_zero_norm (fun n => by simpa only [Real.norm_eq_abs] using hbound n) ht lemma reverse_tail_pairing_of_bound (m : ℕ → MeasurableSpace Ω) (hm : ∀ n, m n ≤ m₀) (hanti : Antitone m) (f : Lp ℝ 2 μ) (g : ℕ → Lp ℝ 2 μ) (hg : ∀ n, AEStronglyMeasurable[m n] (g n) μ) (hb : ∀ n, ∀ᵐ x ∂μ, |g n x| ≤ 1) : Tendsto (fun n => ⟪f - (condExpL2 ℝ ℝ (show (⨅ n, completedFactor (m₀ := m₀) μ (m n)) ≤ m₀ from (iInf_le _ 0).trans (completedFactor_le μ (m 0))) f : Lp ℝ 2 μ), g n⟫) atTop (𝓝 0) := by apply reverse_tail_pairing μ m hm hanti f g hg intro n simpa only [measureUnivNNReal,measure_univ,ENNReal.toNNReal_one,NNReal.coe_one,Real.one_rpow,one_mul] using Lp.norm_le_of_ae_bound (f := g n) (by norm_num : (0 : ℝ) ≤ 1) (hb n) end section open MeasureTheory Filter variable {Ω A : Type*} [m₀ : MeasurableSpace Ω] (μ : Measure Ω) [MeasurableSpace A] omit m₀ in lemma processPast_eq_generators (F : ℤ → Ω → A) (t : ℤ) : processPast F t = ⨆ q : {q : ℤ // q < t}, MeasurableSpace.comap (F q) inferInstance := by change (⨆ n : ℕ, MeasurableSpace.comap (fun x (i : Fin n) => F (t-(n:ℤ)+(i.val:ℤ)) x) inferInstance) = _ simp_rw [MeasurableSpace.comap_process_pi] apply le_antisymm · apply iSup_le intro n apply iSup_le intro i apply le_iSup_of_le ⟨t-(n:ℤ)+(i.val:ℤ),by omega⟩ exact le_rfl · apply iSup_le intro q let n := (t-q.val).toNat have hn : (n:ℤ) = t-q.val := by dsimp [n]; omega have hnpos : 0 < n := by have := q.property; omega apply le_iSup_of_le n apply le_iSup_of_le (⟨0,hnpos⟩ : Fin n) simp only [Nat.cast_zero,add_zero,hn,sub_sub_cancel] exact le_rfl omit m₀ in lemma processPast_mono (F : ℤ → Ω → A) : Monotone (processPast F) := by intro s t h rw [processPast_eq_generators,processPast_eq_generators] apply iSup_le intro q exact le_iSup_of_le ⟨q.val,q.property.trans_le h⟩ le_rfl lemma completedFactor_comap {f : Ω → Ω} (hf : MeasurePreserving f μ μ) {m n : MeasurableSpace Ω} (h : MeasurableSpace.comap f m ≤ n) : MeasurableSpace.comap f (completedFactor (m₀ := m₀) μ m) ≤ completedFactor (m₀ := m₀) μ n := by rintro s ⟨r,hr,rfl⟩ obtain ⟨hr,v,hv,he⟩ := hr change MeasurableSet[m₀] (f ⁻¹' r) ∧ ∃ v, MeasurableSet[n] v ∧ ∀ᵐ x ∂μ, x ∈ f ⁻¹' r ↔ x ∈ v let : MeasurableSpace Ω := m₀ have hfm : @Measurable Ω Ω m₀ m₀ f := hf.measurable have hea : ∀ᵐ x ∂μ, f x ∈ r ↔ f x ∈ v := hf.quasiMeasurePreserving.ae he exact ⟨hfm hr,f ⁻¹' v,h _ ⟨v,hv,rfl⟩,hea⟩ omit m₀ in lemma processPast_comap (U : ℤ → Ω → Ω) (F : ℤ → Ω → A) (htrans : ∀ t s x, F (t+s) x = F t (U s x)) (t s : ℤ) : MeasurableSpace.comap (U s) (processPast F t) ≤ processPast F (t+s) := by rw [processPast_eq_generators,MeasurableSpace.comap_iSup,processPast_eq_generators] apply iSup_le intro q rw [MeasurableSpace.comap_comp] have he : F (q.val+s) = F q ∘ U s := funext fun x => htrans q s x rw [←he] exact le_iSup_of_le ⟨q.val+s,by have := q.property; omega⟩ le_rfl abbrev processTail (F : ℤ → Ω → A) : MeasurableSpace Ω := ⨅ n : ℕ, completedFactor (m₀ := m₀) μ (processPast F (-(n:ℤ))) lemma processTail_le (F : ℤ → Ω → A) : processTail μ F ≤ m₀ := (iInf_le _ 0).trans (completedFactor_le μ _) lemma processTail_le_past (F : ℤ → Ω → A) (t : ℤ) : processTail μ F ≤ completedFactor (m₀ := m₀) μ (processPast F t) := by have h : -(((-t).toNat : ℕ) : ℤ) ≤ t := by omega exact (iInf_le _ ((-t).toNat)).trans (completedFactor_mono μ (processPast_mono F h)) lemma processTail_comap (U : ℤ → Ω → Ω) (hU : ∀ s, MeasurePreserving (U s) μ μ) (F : ℤ → Ω → A) (htrans : ∀ t s x, F (t+s) x = F t (U s x)) (s : ℤ) : MeasurableSpace.comap (U s) (processTail μ F) ≤ processTail μ F := by apply le_iInf intro n have h := MeasurableSpace.comap_mono (g := U s) (processTail_le_past μ F (-(n:ℤ)-s)) apply h.trans have ht := completedFactor_comap μ (hU s) (processPast_comap U F htrans (-(n:ℤ)-s) s) simpa only [sub_add_cancel] using ht lemma measurable_tail_translate {B : Type*} [MeasurableSpace B] (U : ℤ → Ω → Ω) (hU : ∀ s, MeasurePreserving (U s) μ μ) (F : ℤ → Ω → A) (htrans : ∀ t s x, F (t+s) x = F t (U s x)) (g : Ω → B) (hg : Measurable[processTail μ F] g) (s t : ℤ) : Measurable[completedFactor (m₀ := m₀) μ (processPast F t)] (g ∘ U s) := by have hu : @Measurable Ω Ω (processTail μ F) (processTail μ F) (U s) := measurable_iff_comap_le.mpr (processTail_comap μ U hU F htrans s) exact (hg.comp hu).mono (processTail_le_past μ F t) le_rfl end section open MeasureTheory Filter open scoped Topology ENNReal NNReal lemma antitone_starProjection_tendsto {𝕜 H : Type*} [RCLike 𝕜] [NormedAddCommGroup H] [InnerProductSpace 𝕜 H] [CompleteSpace H] (V : ℕ → Submodule 𝕜 H) [∀ n, (V n).HasOrthogonalProjection] [(⨅ n, V n).HasOrthogonalProjection] (hV : Antitone V) (x : H) : Tendsto (fun n => (V n).starProjection x) atTop (𝓝 ((⨅ n, V n).starProjection x)) := by have hmono : Monotone (fun n => (V n)ᗮ) := fun _ _ h => Submodule.orthogonal_le (hV h) have he : (⨆ n, (V n)ᗮ).topologicalClosure = (⨅ n, V n)ᗮ := by rw [← Submodule.orthogonal_orthogonal_eq_closure, ← Submodule.iInf_orthogonal] simp only [Submodule.orthogonal_orthogonal] have ht := Submodule.starProjection_tendsto_closure_iSup (fun n => (V n)ᗮ) hmono x simp only [he] at ht have hh := (tendsto_const_nhds (x := x) (f := atTop)).sub ht simpa only [Submodule.starProjection_orthogonal, sub_apply, ContinuousLinearMap.id_apply, sub_sub_cancel] using hh variable {Ω : Type*} {m₀ : MeasurableSpace Ω} {μ : Measure Ω} lemma lpMeas_iInf_antitone (m : ℕ → MeasurableSpace Ω) (_hm : ∀ n, m n ≤ m₀) (hanti : Antitone m) : lpMeas ℝ ℝ (⨅ n, m n) 2 μ = ⨅ n, lpMeas ℝ ℝ (m n) 2 μ := by apply le_antisymm · apply le_iInf intro n f hf exact mem_lpMeas_iff_aestronglyMeasurable.mpr ((mem_lpMeas_iff_aestronglyMeasurable.mp hf).mono (iInf_le _ n)) · intro f hf have hf' (n : ℕ) : AEStronglyMeasurable[m n] f μ := mem_lpMeas_iff_aestronglyMeasurable.mp (((Submodule.mem_iInf _).mp hf) n) let g : ℕ → Ω → ℝ := fun n => (hf' n).mk f let F : Ω → ℝ := fun ω => limUnder atTop (fun n => g n ω) have hg (n : ℕ) : StronglyMeasurable[m n] (g n) := (hf' n).stronglyMeasurable_mk have hF (j : ℕ) : StronglyMeasurable[m j] F := by have hg' (n : ℕ) : StronglyMeasurable[m j] (g (n+j)) := (hg (n+j)).mono (hanti (Nat.le_add_left j n)) let : MeasurableSpace Ω := m j have hh := StronglyMeasurable.limUnder (l := atTop) hg' convert hh using 1 funext ω dsimp [F] unfold limUnder rw [show (fun n => g (n+j) ω) = (fun n => g n ω) ∘ (fun n => n+j) from rfl, ← Filter.map_map, map_add_atTop_eq_nat] have hmeas : Measurable[⨅ n, m n] F := by intro S hS exact MeasurableSpace.measurableSet_iInf.mpr (fun j => (hF j).measurable hS) apply mem_lpMeas_iff_aestronglyMeasurable.mpr refine hmeas.stronglyMeasurable.aestronglyMeasurable.congr ?_ filter_upwards [ae_all_iff.mpr (fun n => (hf' n).ae_eq_mk)] with ω hω have he : (fun n => g n ω) = fun _ => f ω := funext (fun n => (hω n).symm) change limUnder atTop (fun n => g n ω) = f ω rw [he] exact tendsto_const_nhds.limUnder_eq lemma condExpL2_antitone_tendsto (m : ℕ → MeasurableSpace Ω) (hm : ∀ n, m n ≤ m₀) (hanti : Antitone m) (f : Lp ℝ 2 μ) : Tendsto (fun n => (condExpL2 ℝ ℝ (hm n) f : Lp ℝ 2 μ)) atTop (𝓝 (condExpL2 ℝ ℝ ((iInf_le m 0).trans (hm 0)) f : Lp ℝ 2 μ)) := by let V (n : ℕ) := lpMeas ℝ ℝ (m n) 2 μ have inst (n : ℕ) : (V n).HasOrthogonalProjection := by let : Fact (m n ≤ m₀) := ⟨hm n⟩ dsimp [V] infer_instance have hme : (⨅ n, V n) = lpMeas ℝ ℝ (⨅ n, m n) 2 μ := (lpMeas_iInf_antitone m hm hanti).symm have inst' : (⨅ n, V n).HasOrthogonalProjection := by rw [hme] let : Fact ((⨅ n, m n) ≤ m₀) := ⟨(iInf_le m 0).trans (hm 0)⟩ infer_instance let := inst let := inst' have hV : Antitone V := by intro i j hij f hf exact mem_lpMeas_iff_aestronglyMeasurable.mpr ((mem_lpMeas_iff_aestronglyMeasurable.mp hf).mono (hanti hij)) have h := antitone_starProjection_tendsto V hV f simp only [hme] at h exact h lemma integral_norm_L2_le [IsProbabilityMeasure μ] (f : Lp ℝ 2 μ) : (∫ ω, ‖f ω‖ ∂μ) ≤ ‖f‖ := by rw [← lpNorm_one_eq_integral_norm (Lp.aestronglyMeasurable f), lpNorm, Lp.norm_def] exact ENNReal.toReal_mono (Lp.eLpNorm_ne_top f) (eLpNorm_le_eLpNorm_of_exponent_le (by norm_num : (1 : ℝ≥0∞) ≤ 2)) lemma condExp_antitone_L1_tendsto [IsProbabilityMeasure μ] (m : ℕ → MeasurableSpace Ω) (hm : ∀ n, m n ≤ m₀) (hanti : Antitone m) {f : Ω → ℝ} (hf : MemLp f 2 μ) : Tendsto (fun n => ∫ ω, ‖μ[f | m n] ω - μ[f | ⨅ i, m i] ω‖ ∂μ) atTop (𝓝 0) := by let ht : (⨅ i, m i) ≤ m₀ := (iInf_le m 0).trans (hm 0) let c (n : ℕ) : Lp ℝ 2 μ := condExpL2 ℝ ℝ (hm n) hf.toLp let d : Lp ℝ 2 μ := condExpL2 ℝ ℝ ht hf.toLp have hc := condExpL2_antitone_tendsto m hm hanti hf.toLp change Tendsto c atTop (𝓝 d) at hc have hc' : Tendsto (fun n => ‖c n - d‖) atTop (𝓝 0) := by simpa using (hc.sub (tendsto_const_nhds (x := d) (f := atTop))).norm apply squeeze_zero (fun _ => integral_nonneg (fun _ => norm_nonneg _)) _ hc' intro n calc (∫ ω, ‖μ[f | m n] ω - μ[f | ⨅ i, m i] ω‖ ∂μ) = ∫ ω, ‖(c n - d) ω‖ ∂μ := by apply integral_congr_ae filter_upwards [hf.condExpL2_ae_eq_condExp (𝕜 := ℝ) (hm n), hf.condExpL2_ae_eq_condExp (𝕜 := ℝ) ht, Lp.coeFn_sub (c n) d] with ω hn hd he rw [he] change c n ω = μ[f | m n] ω at hn change d ω = μ[f | ⨅ i, m i] ω at hd simp only [Pi.sub_apply, hn, hd] _ ≤ ‖c n - d‖ := integral_norm_L2_le _ end open MeasureTheory Filter open scoped Topology variable {Ω : Type*} [m₀ : MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ] def clippedConditional (m : MeasurableSpace Ω) (f : Ω → ℝ) : Ω → ℝ := fun x => max 0 (min 1 (μ[f | m] x)) omit [IsProbabilityMeasure μ] in lemma clippedConditional_bounds (m : MeasurableSpace Ω) (f : Ω → ℝ) (x : Ω) : clippedConditional (m₀ := m₀) μ m f x ∈ Set.Icc (0:ℝ) 1 := by exact ⟨le_max_left _ _,max_le (by norm_num) (min_le_left _ _)⟩ omit [IsProbabilityMeasure μ] in lemma clippedConditional_measurable (m : MeasurableSpace Ω) (f : Ω → ℝ) : Measurable[m] (clippedConditional (m₀ := m₀) μ m f) := by exact measurable_const.max (measurable_const.min stronglyMeasurable_condExp.measurable) omit [IsProbabilityMeasure μ] in lemma clippedConditional_ae (m : MeasurableSpace Ω) (f : Ω → ℝ) (hf : ∀ᵐ x ∂μ, f x ∈ Set.Icc (0:ℝ) 1) : clippedConditional (m₀ := m₀) μ m f =ᵐ[μ] μ[f | m] := by let : MeasurableSpace Ω := m₀ have hn := condExp_nonneg (m := m) (hf.mono fun _ h => h.1) have hb := condExp_le_nonneg_const (m := m) (show (0:ℝ) ≤ 1 by norm_num) (hf.mono fun _ h => h.2) filter_upwards [hn,hb] with x hx hy change (0:ℝ) ≤ μ[f | m] x at hx simp only [clippedConditional,min_eq_right hy,max_eq_right hx] lemma clippedConditional_mean (m : MeasurableSpace Ω) (hm : m ≤ m₀) (f : Ω → ℝ) (hf : ∀ᵐ x ∂μ, f x ∈ Set.Icc (0:ℝ) 1) : (∫ x, clippedConditional (m₀ := m₀) μ m f x ∂μ) = ∫ x, f x ∂μ := by let : MeasurableSpace Ω := m₀ rw [integral_congr_ae (clippedConditional_ae (m₀ := m₀) μ m f hf),integral_condExp hm] lemma exists_simple_interval_approx (m : MeasurableSpace Ω) (hm : m ≤ m₀) (g : Ω → ℝ) (hg : Measurable[m] g) (hb : ∀ x, g x ∈ Set.Icc (0:ℝ) 1) (ε : ℝ) (hε : 0 < ε) : ∃ s : @SimpleFunc Ω m ℝ, (∀ x, s x ∈ Set.Icc (0:ℝ) 1) ∧ (∫ x, |s x-g x| ∂μ) < ε := by let s : ℕ → @SimpleFunc Ω m ℝ := fun n => SimpleFunc.approxOn g hg (Set.Icc 0 1) 0 (by norm_num) n have hs n x : s n x ∈ Set.Icc (0:ℝ) 1 := SimpleFunc.approxOn_mem hg (by norm_num) n x have hsm n : Measurable[m₀] (s n) := (s n).measurable.mono hm le_rfl have hgm : Measurable[m₀] g := hg.mono hm le_rfl have hlim x : Tendsto (fun n => |s n x-g x|) atTop (𝓝 (0:ℝ)) := by have ht : Tendsto (fun n => s n x) atTop (𝓝 (g x)) := SimpleFunc.tendsto_approxOn hg (by norm_num) (subset_closure (hb x)) simpa only [sub_self,abs_zero] using (ht.sub_const (g x)).abs have hbound n : ∀ᵐ x ∂μ, ‖|s n x-g x|‖ ≤ (1:ℝ) := by filter_upwards with x rw [Real.norm_eq_abs,abs_abs] exact abs_le.mpr ⟨by have h1 := (hs n x).1; have h2 := (hb x).2; linarith, by have h1 := (hs n x).2; have h2 := (hb x).1; linarith⟩ let : MeasurableSpace Ω := m₀ have ht : Tendsto (fun n => ∫ x, |s n x-g x| ∂μ) atTop (𝓝 (0:ℝ)) := by simpa only [integral_zero,Pi.sub_apply] using tendsto_integral_of_dominated_convergence (μ := μ) (fun _ => (1:ℝ)) (fun n => ((hsm n).sub hgm).abs.aestronglyMeasurable) (integrable_const _) hbound (Eventually.of_forall hlim) obtain ⟨n,hn⟩ := (ht.eventually (gt_mem_nhds hε)).exists exact ⟨s n,hs n,hn⟩ lemma reverse_condExp_completed_L1 (m : ℕ → MeasurableSpace Ω) (hm : ∀ n, m n ≤ m₀) (hanti : Antitone m) {f : Ω → ℝ} (hf : MemLp f 2 μ) : Tendsto (fun n => ∫ x, ‖μ[f | m n] x - μ[f | ⨅ j, completedFactor (m₀ := m₀) μ (m j)] x‖ ∂μ) atTop (𝓝 (0:ℝ)) := by let ht : (⨅ j, completedFactor (m₀ := m₀) μ (m j)) ≤ m₀ := (iInf_le _ 0).trans (completedFactor_le μ _) let c n : Lp ℝ 2 μ := condExpL2 ℝ ℝ (hm n) hf.toLp let d : Lp ℝ 2 μ := condExpL2 ℝ ℝ ht hf.toLp have hc := reverse_conditional_L2 μ m hm hanti hf.toLp change Tendsto c atTop (𝓝 d) at hc have hnorm : Tendsto (fun n => ‖c n-d‖) atTop (𝓝 (0:ℝ)) := by simpa only [sub_self,norm_zero] using (hc.sub (tendsto_const_nhds (x := d))).norm apply squeeze_zero (fun _ => integral_nonneg fun _ => norm_nonneg _) _ hnorm intro n have he : (∫ x, ‖μ[f | m n] x-μ[f | ⨅ j, completedFactor (m₀ := m₀) μ (m j)] x‖ ∂μ) = ∫ x, ‖(c n-d) x‖ ∂μ := by apply integral_congr_ae filter_upwards [hf.condExpL2_ae_eq_condExp (𝕜 := ℝ) (hm n), hf.condExpL2_ae_eq_condExp (𝕜 := ℝ) ht,Lp.coeFn_sub (c n) d] with x hx hy hz rw [hz] change c n x = _ at hx change d x = _ at hy simp only [Pi.sub_apply,hx,hy] rw [he] exact integral_norm_L2_le _ end Rokhlin end end OAI
import OAI.Dynamics.MultipleMixing.Foundations
namespace OAI
noncomputable section
namespace Rokhlin
section
open MeasureTheory Filter
open scoped BigOperators Topology
variable {A B C : Type*} [Fintype A] [Fintype B] [Fintype C]
lemma shannonEntropy_eq_negMulLog (p : A → ℝ) :
shannonEntropy p = ∑ a, Real.negMulLog (p a) := rfl
def rowEntropy (p : A → ℝ) : ℝ :=
shannonEntropy p - Real.negMulLog (∑ a, p a)
lemma rowEntropy_zero : rowEntropy (fun _ : A => 0) = 0 := by
simp [rowEntropy,shannonEntropy,Real.negMulLog]
lemma zero_row {p : A → ℝ} (hp : ∀ a, 0 ≤ p a) (h : ∑ a, p a = 0) (a : A) : p a = 0 := by
exact le_antisymm (h ▸ Finset.single_le_sum (fun a _ => hp a) (Finset.mem_univ a)) (hp a)
lemma rowEntropy_normalize (p : A → ℝ) (hp : ∀ a, 0 ≤ p a) :
rowEntropy p = (∑ a, p a) * shannonEntropy (fun a => p a / ∑ a, p a) := by
classical
let w := ∑ a, p a
have hw : 0 ≤ w := Finset.sum_nonneg (fun a _ => hp a)
by_cases hz : w = 0
· have h0 : p = fun _ => 0 := funext (zero_row hp hz)
subst p
simp [rowEntropy_zero]
· have hs : ∑ a, p a / w = 1 := by rw [←Finset.sum_div,div_self hz]
have he (a : A) : Real.negMulLog (p a) =
p a / w * Real.negMulLog w + w * Real.negMulLog (p a / w) := by
conv_lhs => rw [show p a = w * (p a / w) by field_simp]
exact Real.negMulLog_mul _ _
simp only [rowEntropy,shannonEntropy_eq_negMulLog]
change (∑ a, Real.negMulLog (p a)) - Real.negMulLog w = _
simp_rw [he]
rw [Finset.sum_add_distrib,←Finset.sum_mul,hs,one_mul,←Finset.mul_sum]
ring
lemma rowEntropy_sum_le (p : B → A → ℝ) (hp : ∀ b a, 0 ≤ p b a) :
(∑ b, rowEntropy (p b)) ≤ rowEntropy (fun a => ∑ b, p b a) := by
classical
let w (b : B) := ∑ a, p b a
let s := ∑ b, w b
have hw (b : B) : 0 ≤ w b := Finset.sum_nonneg (fun a _ => hp b a)
have hs : 0 ≤ s := Finset.sum_nonneg (fun b _ => hw b)
by_cases hz : s = 0
· have hw0 b : w b = 0 := zero_row hw hz b
have hp0 b a : p b a = 0 := zero_row (hp b) (hw0 b) a
have hpz : p = fun _ _ => 0 := funext fun b => funext (hp0 b)
rw [hpz]
simp only [Finset.sum_const_zero,rowEntropy_zero,le_refl]
· have hspos : 0 < s := lt_of_le_of_ne hs (Ne.symm hz)
have hweights : ∑ b, w b / s = 1 := by rw [←Finset.sum_div,div_self hz]
have he (a : A) : ∑ b, (w b / s) * (p b a / w b) = (∑ b, p b a) / s := by
rw [Finset.sum_div]
apply Finset.sum_congr rfl
intro b _
by_cases hwb : w b = 0
· rw [zero_row (hp b) hwb a,hwb]; simp
· field_simp
have hj (a : A) : ∑ b, (w b / s) * Real.negMulLog (p b a / w b) ≤
Real.negMulLog ((∑ b, p b a) / s) := by
have h := Real.concaveOn_negMulLog.le_map_sum
(t := Finset.univ) (w := fun b => w b / s) (p := fun b => p b a / w b)
(fun b _ => div_nonneg (hw b) hs) hweights
(fun b _ => div_nonneg (hp b a) (hw b))
simpa only [smul_eq_mul,he] using h
rw [rowEntropy_normalize _ (fun a => Finset.sum_nonneg (fun b _ => hp b a))]
have hsum : (∑ a, ∑ b, p b a) = s := Finset.sum_comm
rw [hsum]
calc
∑ b, rowEntropy (p b) = s * ∑ a, ∑ b, (w b / s) * Real.negMulLog (p b a / w b) := by
simp_rw [rowEntropy_normalize _ (hp _),shannonEntropy_eq_negMulLog]
rw [Finset.sum_comm,Finset.mul_sum]
apply Finset.sum_congr rfl
intro b _
rw [←Finset.mul_sum,←mul_assoc,mul_div_cancel₀ _ hz]
_ ≤ s * ∑ a, Real.negMulLog ((∑ b, p b a) / s) :=
mul_le_mul_of_nonneg_left (Finset.sum_le_sum (fun a _ => hj a)) hs
_ = _ := rfl
def conditionalShannon (p : B → A → ℝ) : ℝ := ∑ b, rowEntropy (p b)
lemma shannon_chain_rule (p : B → A → ℝ) :
shannonEntropy (fun z : B × A => p z.1 z.2) =
shannonEntropy (fun b => ∑ a, p b a) + conditionalShannon p := by
simp only [conditionalShannon,rowEntropy,shannonEntropy_eq_negMulLog,Fintype.sum_prod_type,
Finset.sum_sub_distrib]
ring
lemma conditionalShannon_refine (p : C → B → A → ℝ) (hp : ∀ c b a, 0 ≤ p c b a) :
conditionalShannon (fun z : C × B => p z.1 z.2) ≤
conditionalShannon (fun b a => ∑ c, p c b a) := by