math / data /lean /OAI /Dynamics /SmoothObstruction /WordEntropy.lean
introvoyz045's picture
Migrated from GitHub
4a56cd6 verified
Raw History Blame Contribute Delete
49.8 kB
import OAI.Dynamics.SmoothObstruction.WordHierarchy namespace OAI section namespace SmoothObstruction.WordFlatten open Finset FiniteLaw WordLaw SequentialLaw variable {A : Type*} [Fintype A] [DecidableEq A] {q H : ℕ} omit [Fintype A] in lemma test_flatten (m : Fin (q*H) → Option A) (w : Fin q → Fin H → A) : test m (flattenEquiv w) = ∏ b, test (fun r => m (finProdFinEquiv (b,r))) (w b) := by unfold test calc _ = ∏ br : Fin q × Fin H, mark (m (finProdFinEquiv br)) (w br.1 br.2) := by simpa only [flattenEquiv, Equiv.coe_fn_mk,Equiv.symm_apply_apply] using (finProdFinEquiv.prod_comp (fun j => mark (m j) (flattenEquiv w j))).symm _ = _ := Fintype.prod_prod_type _ omit [Fintype A] [DecidableEq A] in lemma count_grouped (m : Fin (q*H) → Option A) : (∑ b : Fin q, (univ.filter (fun r : Fin H => (m (finProdFinEquiv (b,r))).isSome)).card) = (univ.filter (fun r => (m r).isSome)).card := by simp only [card_eq_sum_ones, sum_filter] calc _ = ∑ br : Fin q × Fin H, if (m (finProdFinEquiv br)).isSome then 1 else 0 := (Fintype.sum_prod_type (fun br : Fin q × Fin H => if (m (finProdFinEquiv br)).isSome then (1:ℕ) else 0)).symm _ = _ := finProdFinEquiv.sum_comp (fun r => if (m r).isSome then 1 else 0) variable {V : Type*} [AddCommGroup V] [Module (ZMod 2) V] [FiniteDimensional (ZMod 2) V] [Fintype V] [DecidableEq V] (B : LinearMap.BilinForm (ZMod 2) V) {K : ℕ} omit [FiniteDimensional (ZMod 2) V] in /-- W3 itself: the actual single conditioned law controls every arbitrary partial list, with a single factor two, including lists crossing groups. -/ theorem flat_sublist (hd : K+1 ≤ Module.finrank (ZMod 2) V) (hK : 1 ≤ K) (E : Finset (Fin q → Fin H → {v : V // v ≠ 0})) (hE : 1/2 ≤ prob (rawLaw B hd H q) E) (m : Fin (q*H) → Option {v : V // v ≠ 0}) : expect (flatLaw B hd E hE) (test m) ≤ 2 * ((2 : ℝ)^(2*K) / Fintype.card {v : V // v ≠ 0}) ^ (univ.filter (fun r => (m r).isSome)).card := by unfold flatLaw rw [expect_transport] have ht : ((test m) ∘ flattenEquiv) = (fun w => ∏ b, test (fun r => m (finProdFinEquiv (b,r))) (w b)) := by funext w exact test_flatten m w rw [ht] have hh := conditioned_sublist B hd hK E hE (fun b r => m (finProdFinEquiv (b,r))) simpa only [count_grouped] using hh end SmoothObstruction.WordFlatten namespace SmoothObstruction.FiniteLaw open Finset FiniteEntropy variable {α β γ : Type*} [Fintype α] [Fintype β] [Fintype γ] noncomputable def uniform [Nonempty α] : Law α where mass := fun _ => (Fintype.card α : ℝ)⁻¹ nonneg := fun _ => by positivity total := by simp [Fintype.card_ne_zero] noncomputable def mixture (P : Law α) (Q : α → Law β) : Law β where mass b := ∑ a,P.mass a*(Q a).mass b nonneg b := sum_nonneg (fun a _ => mul_nonneg (P.nonneg a) ((Q a).nonneg b)) total := by rw [sum_comm] simp only [←mul_sum,(Q _).total,mul_one,P.total] noncomputable def TV (P Q : Law α) : ℝ := (1/2:ℝ)*∑ a,|P.mass a-Q.mass a| noncomputable def entropy (P : Law α) : ℝ := natEntropy P.mass lemma TV_nonneg (P Q : Law α) : 0 ≤ TV P Q := by unfold TV; positivity lemma TV_symm (P Q : Law α) : TV P Q=TV Q P := by simp only [TV,abs_sub_comm] lemma TV_triangle (P Q R : Law α) : TV P R ≤ TV P Q+TV Q R := by unfold TV rw [←mul_add,←sum_add_distrib] exact mul_le_mul_of_nonneg_left (sum_le_sum (fun a _ => abs_sub_le (P.mass a) (Q.mass a) (R.mass a))) (by norm_num) lemma push_sub [DecidableEq β] (P Q : Law α) (f : α → β) (b : β) : (push P f).mass b-(push Q f).mass b = ∑ a,if f a=b then P.mass a-Q.mass a else 0 := by classical simp only [push,←sum_sub_distrib] apply sum_congr rfl intro a _ split_ifs <;> ring lemma TV_push [DecidableEq β] (P Q : Law α) (f : α → β) : TV (push P f) (push Q f) ≤ TV P Q := by classical unfold TV apply mul_le_mul_of_nonneg_left _ (by norm_num) simp_rw [push_sub] calc _ ≤ ∑ b,∑ a,|if f a=b then P.mass a-Q.mass a else 0| := sum_le_sum (fun b _ => abs_sum_le_sum_abs _ _) _ = ∑ a,|P.mass a-Q.mass a| := by rw [sum_comm] simp only [apply_ite abs,abs_zero,Finset.sum_ite_eq,Finset.mem_univ,ite_true] lemma entropy_nonneg (P : Law α) : 0 ≤ entropy P := FiniteEntropy.entropy_nonneg P.nonneg P.total lemma entropy_mixture (P : Law α) (Q : α → Law β) : ∑ a,P.mass a*entropy (Q a) ≤ entropy (mixture P Q) := FiniteEntropy.entropy_mixture P.mass (fun a => (Q a).mass) P.nonneg P.total (fun a => (Q a).nonneg) lemma entropy_atom_lower (P : Law α) {M : ℝ} (_hM : 0 < M) (h : ∀ a,P.mass a ≤ M) : -Real.log M ≤ entropy P := by have hb (a : α) : P.mass a*(-Real.log M) ≤ Real.negMulLog (P.mass a) := by rcases (P.nonneg a).eq_or_lt with hz|hp · rw [←hz,zero_mul,Real.negMulLog_zero] · have hh := Real.log_le_log hp (h a) rw [Real.negMulLog_def] nlinarith [mul_le_mul_of_nonneg_left hh hp.le] have hs := sum_le_sum (fun a (_ha : a ∈ (univ : Finset α)) => hb a) simpa only [←sum_mul,P.total,one_mul,entropy,natEntropy] using hs /-- Exactly the one-bit TV cost used in source equation (tv-entropy), expressed first in natural units. -/ lemma entropy_tv_lower (P Q : Law α) {s : ℝ} (hs : TV P Q ≤ s) : entropy P-Real.log 2-s*Real.log (Fintype.card α) ≤ entropy Q := by have hcard : 1 ≤ Fintype.card α := by by_contra h have he : IsEmpty α := Fintype.card_eq_zero_iff.mp (by omega) have hh := P.total simp only [univ_eq_empty,sum_empty] at hh norm_num at hh have hc : 0 ≤ Real.log (Fintype.card α) := Real.log_nonneg (by exact_mod_cast hcard) have hh := FiniteEntropy.entropy_tv Q.nonneg P.nonneg Q.total P.total rw [←TV,TV_symm] at hh change entropy P ≤ entropy Q+Real.log 2+TV P Q*Real.log (Fintype.card α) at hh nlinarith [mul_le_mul_of_nonneg_right hs hc] lemma entropy_close_atom_lower (P Q : Law α) {M s : ℝ} (hM : 0 < M) (hP : ∀ a,P.mass a ≤ M) (hs : TV P Q ≤ s) : -Real.log M-Real.log 2-s*Real.log (Fintype.card α) ≤ entropy Q := by have ha := entropy_atom_lower P hM hP have hb := entropy_tv_lower P Q hs linarith /-- Mixtures over arbitrary parent laws retain each admissible-slot bound; only the total exceptional-slot mass is lost. -/ theorem entropy_good_mixture (P : Law α) (Q : α → Law β) (E : Finset α) {L : ℝ} (h : ∀ a ∈ E,L ≤ entropy (Q a)) : (∑ a ∈ E,P.mass a)*L ≤ entropy (mixture P Q) := by classical calc _ = ∑ a ∈ E,P.mass a*L := sum_mul _ _ _ _ ≤ ∑ a ∈ E,P.mass a*entropy (Q a) := sum_le_sum (fun a ha => mul_le_mul_of_nonneg_left (h a ha) (P.nonneg a)) _ ≤ ∑ a,P.mass a*entropy (Q a) := sum_le_sum_of_subset_of_nonneg (subset_univ E) (fun a _ _ => mul_nonneg (P.nonneg a) (entropy_nonneg (Q a))) _ ≤ _ := entropy_mixture P Q end SmoothObstruction.FiniteLaw namespace SmoothObstruction.SequentialLaw open Finset variable {A : Type*} [Fintype A] [DecidableEq A] {q N : ℕ} noncomputable def embeddedMarks (e : Fin q ↪ Fin N) (v : Fin q → A) : Fin N → Option A := fun i => if h : ∃ j, e j=i then some (v h.choose) else none omit [Fintype A] [DecidableEq A] in lemma embeddedMarks_at (e : Fin q ↪ Fin N) (v : Fin q → A) (j : Fin q) : embeddedMarks e v (e j)=some (v j) := by classical have h : ∃ k, e k=e j := ⟨j,rfl⟩ have hc := e.injective h.choose_spec simp only [embeddedMarks,dite_eq_left h,hc] omit [Fintype A] [DecidableEq A] in lemma embeddedMarks_some (e : Fin q ↪ Fin N) (v : Fin q → A) (i : Fin N) (a : A) : embeddedMarks e v i=some a ↔ ∃ j,e j=i ∧ v j=a := by classical constructor · intro h unfold embeddedMarks at h split_ifs at h with hi · exact ⟨hi.choose,hi.choose_spec,Option.some.inj h⟩ · rintro ⟨j,rfl,rfl⟩ exact embeddedMarks_at e v j omit [Fintype A] [DecidableEq A] in lemma embeddedMarks_count (e : Fin q ↪ Fin N) (v : Fin q → A) : (univ.filter (fun i => (embeddedMarks e v i).isSome)).card=q := by classical have he : univ.filter (fun i => (embeddedMarks e v i).isSome)=univ.map e := by ext i simp only [mem_filter,mem_univ,true_and,mem_map] constructor · intro hi cases hm : embeddedMarks e v i with | none => simp [hm] at hi | some a => obtain ⟨j,hj,_⟩ := (embeddedMarks_some e v i a).mp hm exact ⟨j,hj⟩ · rintro ⟨j,rfl⟩ simp only [embeddedMarks_at,Option.isSome_some] rw [he,card_map,card_univ,Fintype.card_fin] lemma embeddedMarks_test (e : Fin q ↪ Fin N) (v : Fin q → A) (w : Fin N → A) : test (embeddedMarks e v) w=if w ∘ e=v then 1 else 0 := by classical rw [test_eq_indicator] congr 1 apply propext constructor · intro h funext j exact h (e j) (v j) (embeddedMarks_at e v j) · intro h i a hm obtain ⟨j,rfl,hj⟩ := (embeddedMarks_some e v i a).mp hm exact (congrFun h j).trans hj end SmoothObstruction.SequentialLaw namespace SmoothObstruction.WordFlatten open Finset FiniteLaw WordLaw SequentialLaw variable {V : Type*} [AddCommGroup V] [Module (ZMod 2) V] [FiniteDimensional (ZMod 2) V] [Fintype V] [DecidableEq V] (B : LinearMap.BilinForm (ZMod 2) V) {K b H q : ℕ} omit [FiniteDimensional (ZMod 2) V] in /-- The pushed law of any distinct coordinate list has the exact W3 atom bound; the law is the recursively used single conditioned sampler. -/ theorem flat_window_atom (hd : K+1 ≤ Module.finrank (ZMod 2) V) (hK : 1 ≤ K) (E : Finset (Fin b → Fin H → {v : V // v ≠ 0})) (hE : 1/2 ≤ prob (rawLaw B hd H b) E) (e : Fin q ↪ Fin (b*H)) (v : Fin q → {v : V // v ≠ 0}) : (push (flatLaw B hd E hE) (fun w => w ∘ e)).mass v ≤ 2*((2:ℝ)^(2*K)/Fintype.card {v : V // v ≠ 0})^q := by have hh := flat_sublist B hd hK E hE (embeddedMarks e v) rw [embeddedMarks_count] at hh simp only [FiniteLaw.expect,embeddedMarks_test] at hh simpa only [push,FiniteLaw.expect,mul_ite,mul_one,mul_zero] using hh end SmoothObstruction.WordFlatten namespace SmoothObstruction.FiniteLaw open Finset variable {A B C : Type*} [Fintype A] [Fintype B] [Fintype C] lemma mass_ext (P Q : Law A) (h : P.mass=Q.mass) : P=Q := by cases P; cases Q; cases h; rfl lemma push_comp [DecidableEq B] [DecidableEq C] (P : Law A) (f : A → B) (g : B → C) : push (push P f) g=push P (g ∘ f) := by apply mass_ext funext c simp only [push] calc _ = ∑ b,∑ a,if g b=c then (if f a=b then P.mass a else 0) else 0 := by apply sum_congr rfl intro b _ split_ifs <;> simp _ = _ := by rw [sum_comm] apply sum_congr rfl intro a _ have he (b : B) : (if g b=c then (if f a=b then P.mass a else 0) else 0) = if f a=b then (if g b=c then P.mass a else 0) else 0 := by split_ifs <;> rfl simp only [he,Finset.sum_ite_eq,mem_univ,ite_true,Function.comp_apply] lemma push_uniform_mass [Nonempty A] [DecidableEq B] (f : A → B) (b : B) : (push uniform f).mass b=(∑ a,if f a=b then (1:ℝ) else 0)/Fintype.card A := by simp only [push,uniform,div_eq_mul_inv,sum_mul,ite_mul,one_mul,zero_mul] lemma uniform_reindex [Nonempty A] [Nonempty B] [DecidableEq C] (e : A ≃ B) (f : B → C) : push (uniform (α:=A)) (f ∘ e)=push (uniform (α:=B)) f := by apply mass_ext funext c rw [push_uniform_mass,push_uniform_mass,Fintype.card_congr e] congr 1 exact e.sum_comp (fun b => if f b=c then (1:ℝ) else 0) lemma push_observable {N : ℕ} [DecidableEq C] (P : Law (Fin N → A)) (f : (Fin N → A) → C) (c : C) : (push P f).mass c=WordFamily.observableLaw P f c := by simp only [push,WordFamily.observableLaw,expect,mul_ite,mul_one,mul_zero] lemma TV_empirical {M N : ℕ} [NeZero M] [DecidableEq C] (P : Law (Fin N → A)) (f : (Fin N → A) → C) (w : Fin M → Fin N → A) : TV (push uniform (f ∘ w)) (push P f)=WordFamily.empiricalTV P f w := by simp only [TV,push_uniform_mass,Fintype.card_fin,push_observable, WordFamily.empiricalTV,Function.comp_apply] lemma TV_grouped {q H : ℕ} [NeZero q] [NeZero H] [Nonempty A] [DecidableEq C] (f : A → C) (w : Fin q → Fin H → A) : TV (push uniform (f ∘ WordFlatten.flattenEquiv w)) (push uniform f) = (1/2:ℝ)*∑ c, |(∑ b,(∑ r,if f (w b r)=c then (1:ℝ) else 0)/(H:ℝ))/(q:ℝ)- (∑ a,if f a=c then (1:ℝ) else 0)/Fintype.card A| := by unfold TV congr 1 apply sum_congr rfl intro c _ rw [push_uniform_mass,push_uniform_mass,Fintype.card_fin] congr 2 have he : (∑ j : Fin (q*H),if (f ∘ WordFlatten.flattenEquiv w) j=c then (1:ℝ) else 0) = ∑ b,∑ r,if f (w b r)=c then (1:ℝ) else 0 := by rw [← finProdFinEquiv.sum_comp] simpa only [Function.comp_apply,WordFlatten.flattenEquiv,Equiv.coe_fn_mk, Equiv.symm_apply_apply] using Fintype.sum_prod_type (fun br : Fin q × Fin H => if f (w br.1 br.2)=c then (1:ℝ) else 0) rw [he,←sum_div,Nat.cast_mul] simp only [div_div] rw [mul_comm] lemma atom_entropy_formula {q K : ℕ} [Nonempty A] (P : Law (Fin q → A)) (hP : ∀ v,P.mass v ≤ 2*((2:ℝ)^(2*K)/Fintype.card A)^q) : (q:ℝ)*(Real.log (Fintype.card A)-(2*K:ℕ)*Real.log 2)-Real.log 2 ≤ entropy P := by have hc : (0:ℝ)<Fintype.card A := by exact_mod_cast Fintype.card_pos have hM : 0<2*((2:ℝ)^(2*K)/Fintype.card A)^q := by positivity have hh := entropy_atom_lower P hM hP rw [Real.log_mul (by norm_num) (by positivity),Real.log_pow, Real.log_div (by positivity) hc.ne',Real.log_pow] at hh convert hh using 1; ring lemma window_entropy_close {q K : ℕ} [Nonempty A] (P Q : Law (Fin q → A)) (hP : ∀ v,P.mass v ≤ 2*((2:ℝ)^(2*K)/Fintype.card A)^q) {s : ℝ} (hPQ : TV P Q ≤ s) : (q:ℝ)*((1-s)*Real.log (Fintype.card A)-(2*K:ℕ)*Real.log 2)-2*Real.log 2 ≤ entropy Q := by have ha := atom_entropy_formula P hP have ht := entropy_tv_lower P Q hPQ rw [Fintype.card_fun,Fintype.card_fin,Nat.cast_pow,Real.log_pow] at ht nlinarith end SmoothObstruction.FiniteLaw namespace SmoothObstruction.WordHierarchy open Finset WordParameters WordStep WordAlphabet FiniteLaw noncomputable instance levelNonempty (i : ℕ) : Nonempty (A i) := alphabet_nonempty i noncomputable instance groupNeZero (i : ℕ) : NeZero (groupLength i) := ⟨by rw [groupLength,q_mul_H]; exact (N_pos _).ne'⟩ noncomputable instance familyNeZero (i : ℕ) : NeZero (size i) := ⟨Fintype.card_ne_zero⟩ noncomputable def conditionedLabels (i : ℕ) : Law (Labels i) := WordFlatten.flatLaw (form (S i)) (goodDim i) (stage i).E (stage i).massE noncomputable def conditionedWindow (i : ℕ) (t : Offsets (S i)) : Law (Fin (D (S i)) → A i) := push (conditionedLabels i) (window (S i) t) noncomputable def uniformWindow (i : ℕ) (t : Offsets (S i)) : Law (Fin (D (S i)) → A i) := push uniform (fun a : A (i+1) => window (S i) t ((stage i).word a)) noncomputable def parentWindow (i : ℕ) (a : A (i+2)) (t : Offsets (S i)) : Law (Fin (D (S i)) → A i) := push uniform (fun j : Fin (groupLength (i+1)) => window (S i) t ((stage i).word ((stage (i+1)).word a j))) /-- W3 for the actual uniform selected family, in the finite-law API. -/ lemma uniformWindow_close (i : ℕ) (t : Offsets (S i)) : TV (uniformWindow i t) (conditionedWindow i t) ≤ 1/10000 := by have h := (stage i).windowTV t rw [←TV_empirical] at h change TV (push uniform ((window (S i) t) ∘ (stage i).w)) (conditionedWindow i t) ≤ _ at h have he : uniformWindow i t=push uniform ((window (S i) t) ∘ (stage i).w) := by exact uniform_reindex (Fintype.equivFin (A (i+1))) ((window (S i) t) ∘ (stage i).w) rwa [he] /-- W4 for each actual parent: uniformity is only over its slots, never an assumption on the parent-word distribution. -/ lemma parentWindow_close (i : ℕ) (a : A (i+2)) (t : Offsets (S i)) : TV (parentWindow i a t) (uniformWindow i t) ≤ 1/10000 := by let : NeZero (q (S (i+1))) := ⟨(q_pos _).ne'⟩ let : NeZero (H (S (i+1))) := ⟨(H_pos _).ne'⟩ have hh := TV_grouped (fun b => window (S i) t ((stage i).word b)) (WordFlatten.flattenEquiv.symm ((stage (i+1)).word a)) simp only [Equiv.apply_symm_apply] at hh change TV (parentWindow i a t) (uniformWindow i t)= WordLaw.tvFromUniformObservable (fun b => window (S i) t ((stage i).word b)) (WordFlatten.flattenEquiv.symm ((stage (i+1)).word a)) at hh rw [hh] exact stage_approximates i a t /-- The two and only two approximation costs used in the manuscript. -/ lemma parent_conditioned_close (i : ℕ) (a : A (i+2)) (t : Offsets (S i)) : TV (conditionedWindow i t) (parentWindow i a t) ≤ 2/10000 := by rw [TV_symm] have ht := TV_triangle (parentWindow i a t) (uniformWindow i t) (conditionedWindow i t) have h1 := parentWindow_close i a t have h2 := uniformWindow_close i t linarith def windowEmbedding (i : ℕ) (t : Offsets (S i)) {r : ℕ} (hr : r ≤ D (S i)) : Fin r ↪ Fin (groupLength i) where toFun k := ⟨t.val.val+k.val,by change t.val.val+k.val < q (S i)*H (S i); have h := t.property; have hk := k.isLt; omega⟩ inj' := by intro j k h; apply Fin.ext; have he := congrArg Fin.val h; dsimp at he; omega def takePrefix {level r d : ℕ} (hr : r ≤ d) (w : Fin d → A level) : Fin r → A level := fun k => w ⟨k.val,lt_of_lt_of_le k.isLt hr⟩ lemma conditioned_prefix_atom (i : ℕ) (t : Offsets (S i)) {r : ℕ} (hr : r ≤ D (S i)) (v : Fin r → A i) : (push (conditionedWindow i t) (takePrefix hr)).mass v ≤ 2*((2:ℝ)^(2*K (S i))/Fintype.card (A i))^r := by rw [conditionedWindow,push_comp] exact WordFlatten.flat_window_atom (form (S i)) (goodDim i) (by change 1 ≤ K (sequence 40000 i); have := (step_conditions initial_threshold i).1; omega) (stage i).E (stage i).massE (windowEmbedding i t hr) v /-- A real entropy lower bound for EVERY selected parent and admissible internal offset, with fully constructed comparison law and all atom bounds. -/ theorem parent_prefix_entropy (i : ℕ) (a : A (i+2)) (t : Offsets (S i)) {r : ℕ} (hr : r ≤ D (S i)) : (r:ℝ)*((1-2/10000)*Real.log (Fintype.card (A i))-(2*K (S i):ℕ)*Real.log 2)-2*Real.log 2 ≤ entropy (push (parentWindow i a t) (takePrefix hr)) := by exact window_entropy_close (push (conditionedWindow i t) (takePrefix hr)) _ (conditioned_prefix_atom i t hr) ((TV_push _ _ (takePrefix hr)).trans (parent_conditioned_close i a t)) end SmoothObstruction.WordHierarchy namespace SmoothObstruction.FiniteLaw open MeasureTheory Set Finset variable {X α β : Type*} [MeasurableSpace X] [Fintype α] [Fintype β] [MeasurableSpace α] [MeasurableSingletonClass α] [MeasurableSpace β] [MeasurableSingletonClass β] omit [MeasurableSpace α] [MeasurableSingletonClass α] in lemma observed_ext {P Q : Law α} (h : P.mass=Q.mass) : P=Q := by cases P; cases Q; cases h; rfl noncomputable def observed (μ : Measure X) [IsProbabilityMeasure μ] (f : X → α) (hf : Measurable f) : Law α where mass a := μ.real {x | f x=a} nonneg _ := measureReal_nonneg total := by have h := Alignment.sum_fibers μ univ hf simpa only [Set.mem_univ,true_and,probReal_univ] using h lemma observed_push [DecidableEq β] (μ : Measure X) [IsProbabilityMeasure μ] (f : X → α) (hf : Measurable f) (g : α → β) : push (observed μ f hf) g = observed μ (g ∘ f) ((measurable_of_countable g).comp hf) := by classical apply observed_ext funext b dsimp only [push,FiniteEntropy.pushMass,observed,Function.comp_apply] rw [←Alignment.sum_fibers μ {x | g (f x)=b} hf] apply sum_congr rfl intro a _ by_cases ha : g a=b · rw [ite_eq_left ha] congr 1 ext x simp only [Set.mem_ofPred_eq] exact ⟨fun hx => ⟨by rw [hx,ha],hx⟩,And.right⟩ · rw [ite_eq_right ha] have he : {x | x∈{x | g (f x)=b} ∧ f x=a}=∅ := by ext x simp only [Set.mem_ofPred_eq,Set.mem_empty_iff_false,iff_false] rintro ⟨hb,hf⟩ exact ha (hf ▸ hb) rw [he,measureReal_empty] noncomputable def conditionalObserved (μ : Measure X) [IsProbabilityMeasure μ] (f : X → α) (hf : Measurable f) (k : X → β) (_hk : Measurable k) [Nonempty α] (b : β) : Law α := by let p := μ.real {x | k x=b} by_cases hp : p=0 · exact uniform · refine ⟨(fun a => μ.real {x | k x=b ∧ f x=a}/p),fun _ => div_nonneg measureReal_nonneg measureReal_nonneg,?_⟩ rw [←sum_div] have hs := Alignment.sum_fibers μ {x | k x=b} hf simp only [Set.mem_ofPred_eq] at hs rw [hs] exact div_self hp lemma conditionalObserved_mass (μ : Measure X) [IsProbabilityMeasure μ] (f : X → α) (hf : Measurable f) (k : X → β) (hk : Measurable k) [Nonempty α] (b : β) (a : α) : (observed μ k hk).mass b*(conditionalObserved μ f hf k hk b).mass a = μ.real {x | k x=b ∧ f x=a} := by change μ.real {x | k x=b}*_= _ unfold conditionalObserved dsimp only split_ifs with hp · rw [hp,zero_mul] apply (le_antisymm ?_ measureReal_nonneg).symm calc _ ≤ μ.real {x | k x=b} := measureReal_mono (by intro x hx; exact hx.1) _ = 0 := hp · exact mul_div_cancel₀ _ hp /-- Disintegration over a finite actual observation. Zero conditional masses are represented by a harmless uniform law and contribute exactly zero. -/ theorem observed_mixture (μ : Measure X) [IsProbabilityMeasure μ] (f : X → α) (hf : Measurable f) (k : X → β) (hk : Measurable k) [Nonempty α] : observed μ f hf=mixture (observed μ k hk) (conditionalObserved μ f hf k hk) := by apply observed_ext funext a change μ.real {x | f x=a}=∑ b,_ simp only [conditionalObserved_mass] have h := Alignment.sum_fibers μ {x | f x=a} hk simpa only [Set.mem_ofPred_eq,and_comm] using h.symm end SmoothObstruction.FiniteLaw namespace SmoothObstruction.WordHierarchy open WordParameters WordStep WordConcatenation MeasureTheory Set Finset FiniteLaw open scoped ENNReal noncomputable instance stageLengthNeZero (i : ℕ) : NeZero (N (S i)) := ⟨(N_pos _).ne'⟩ lemma measurable_slot (i : ℕ) : Measurable (slot i) := by have hh : slot i = (fun a : ZMod (length 40000 (i+1)) => (⟨a.val / length 40000 i,by apply (Nat.div_lt_iff_lt_mul (length_pos 40000 i)).mpr simpa only [length_step,Nat.mul_comm] using ZMod.val_lt a⟩ : Fin (N (S i)))) ∘ phase (i+1) := rfl rw [hh] exact (measurable_of_countable _).comp (measurable_phase (i+1)) lemma pack_slots (i : ℕ) (x : Extension) : pack (slot (i+1) x,slot i x)=parentSlot i x := by apply Fin.ext rw [pack_val,slot_pair] lemma slots_iff (i : ℕ) (x : Extension) (u : Fin (N (S i))) (h : Fin (N (S (i+1)))) : slot i x=u ∧ slot (i+1) x=h ↔ parentSlot i x=pack (h,u) := by rw [←pack_slots,(Equiv.injective pack).eq_iff,Prod.mk.injEq,and_comm] lemma boundary_phase_zero (μ : Measure Extension) (i : ℕ) : ∀ᵐ x ∂boundary μ i,phase i x=0 := by exact Measure.ae_smul_measure (ae_restrict_mem (measurableSet_eq_fun (measurable_phase i) measurable_const)) _ lemma boundary_slots_mass (μ : Measure Extension) [IsFiniteMeasure μ] (hT : MeasurePreserving T μ μ) (i : ℕ) (v : A (i+2)) (u : Fin (N (S i))) (h : Fin (N (S (i+1)))) : (boundary μ i).real {x | label (i+2) x=v ∧ slot i x=u ∧ slot (i+1) x=h} = μ.real {x | label (i+2) x=v}/((N (S i):ℝ)*N (S (i+1))) := by have he : {x | label (i+2) x=v ∧ slot i x=u ∧ slot (i+1) x=h} = {x | parentSlot i x=pack (h,u) ∧ label (i+2) x=v} := by ext x simp only [Set.mem_ofPred_eq,←slots_iff,and_comm] rw [he] apply (eq_div_iff (mul_ne_zero (Nat.cast_ne_zero.mpr (N_pos _).ne') (Nat.cast_ne_zero.mpr (N_pos _).ne') : (N (S i):ℝ)*N (S (i+1))≠0)).mpr have hh := boundary_parentSlot_mass μ hT i v (pack (h,u)) simpa only [Nat.cast_mul,mul_comm] using hh noncomputable def windowKey (i : ℕ) (x : Extension) : A (i+2) × Fin (N (S i)) := (label (i+2) x,slot i x) lemma measurable_windowKey (i : ℕ) : Measurable (windowKey i) := (measurable_label _).prodMk (measurable_slot _) lemma boundary_key_mass (μ : Measure Extension) [IsProbabilityMeasure μ] (hT : MeasurePreserving T μ μ) (i : ℕ) (v : A (i+2)) (u : Fin (N (S i))) : (boundary μ i).real {x | windowKey i x=(v,u)} = μ.real {x | label (i+2) x=v}/N (S i) := by let : IsProbabilityMeasure (boundary μ i) := boundary_isProbability μ hT i rw [←Alignment.sum_fibers (boundary μ i) {x | windowKey i x=(v,u)} (measurable_slot (i+1))] have hs (h : Fin (N (S (i+1)))) : (boundary μ i).real {x | x∈{x | windowKey i x=(v,u)} ∧ slot (i+1) x=h} = μ.real {x | label (i+2) x=v}/((N (S i):ℝ)*N (S (i+1))) := by convert boundary_slots_mass μ hT i v u h using 2 ext x simp only [Set.mem_ofPred_eq,windowKey,Prod.mk.injEq,and_assoc] simp only [hs,sum_const,card_univ,Fintype.card_fin,nsmul_eq_mul] field_simp [Nat.cast_ne_zero.mpr (N_pos (S (i+1))).ne'] end SmoothObstruction.WordHierarchy namespace SmoothObstruction.WordHierarchy open WordParameters WordStep WordConcatenation MeasureTheory Set Finset FiniteLaw /-- The actual r-child window at an internal offset of a specified parent. -/ noncomputable def sampledWindow (i r : ℕ) (u : Fin (N (S i))) (hr : u.val+r ≤ N (S i)) (v : A (i+2)) (h : Fin (N (S (i+1)))) : Fin r → A i := fun j => children i (children (i+1) v h) ⟨u.val+j.val,by have := j.isLt; omega⟩ noncomputable def sampledLaw (i r : ℕ) (u : Fin (N (S i))) (hr : u.val+r ≤ N (S i)) (v : A (i+2)) : Law (Fin r → A i) := push uniform (sampledWindow i r u hr v) lemma labelList_key (i r : ℕ) (u : Fin (N (S i))) (hr : u.val+r ≤ N (S i)) (v : A (i+2)) (h : Fin (N (S (i+1)))) (x : Extension) (hx : phase i x=0) (hk : windowKey i x=(v,u)) (hh : slot (i+1) x=h) : labelList i r x=sampledWindow i r u hr v h := by have hv : label (i+2) x=v := congrArg Prod.fst hk have hu : slot i x=u := congrArg Prod.snd hk rw [labelList_inside i r x hx (by simpa only [hu] using hr)] simp only [hv,hu,hh] rfl /-- Exact disintegration at an admissible offset. Stationarity gives slot uniformity even when the parent marginal has zero or varying masses. -/ theorem boundary_window_joint (μ : Measure Extension) [IsProbabilityMeasure μ] (hT : MeasurePreserving T μ μ) (i r : ℕ) (u : Fin (N (S i))) (hr : u.val+r ≤ N (S i)) (v : A (i+2)) (w : Fin r → A i) : (boundary μ i).real {x | windowKey i x=(v,u) ∧ labelList i r x=w} = (boundary μ i).real {x | windowKey i x=(v,u)}*(sampledLaw i r u hr v).mass w := by classical let : IsProbabilityMeasure (boundary μ i) := boundary_isProbability μ hT i rw [←Alignment.sum_fibers (boundary μ i) {x | windowKey i x=(v,u) ∧ labelList i r x=w} (measurable_slot (i+1))] have hs (h : Fin (N (S (i+1)))) : (boundary μ i).real {x | x∈{x | windowKey i x=(v,u) ∧ labelList i r x=w} ∧ slot (i+1) x=h} = if sampledWindow i r u hr v h=w then μ.real {x | label (i+2) x=v}/((N (S i):ℝ)*N (S (i+1))) else 0 := by by_cases hw : sampledWindow i r u hr v h=w · rw [ite_eq_left hw] apply (measureReal_congr ?_).trans (boundary_slots_mass μ hT i v u h) filter_upwards [boundary_phase_zero μ i] with x hx apply propext constructor · rintro ⟨⟨hk,_⟩,hh⟩ exact ⟨congrArg Prod.fst hk,congrArg Prod.snd hk,hh⟩ · rintro ⟨hv,hu,hh⟩ have hk : windowKey i x=(v,u) := Prod.ext hv hu exact ⟨⟨hk,(labelList_key i r u hr v h x hx hk hh).trans hw⟩,hh⟩ · rw [ite_eq_right hw] apply (measureReal_congr (t := (∅ : Set Extension)) ?_).trans measureReal_empty filter_upwards [boundary_phase_zero μ i] with x hx apply propext change ((windowKey i x=(v,u) ∧ labelList i r x=w) ∧ slot (i+1) x=h) ↔ False constructor · rintro ⟨⟨hk,hlist⟩,hh⟩ exact hw ((labelList_key i r u hr v h x hx hk hh).symm.trans hlist) · exact False.elim simp only [hs,boundary_key_mass μ hT i v u] simp only [sampledLaw,push,uniform,Fintype.card_fin] rw [mul_sum] apply sum_congr rfl intro h _ split_ifs <;> simp [div_eq_mul_inv,mul_assoc,mul_left_comm,mul_comm] /-- The conditional law is the actual finite-parent sampled window law on all positive-probability fibres. No hypothesis about parent frequencies. -/ lemma conditional_window_eq (μ : Measure Extension) [IsProbabilityMeasure μ] (hT : MeasurePreserving T μ μ) (i r : ℕ) (u : Fin (N (S i))) (hr : u.val+r ≤ N (S i)) (v : A (i+2)) (hp : (boundary μ i).real {x | windowKey i x=(v,u)}≠0) : let : IsProbabilityMeasure (boundary μ i) := boundary_isProbability μ hT i conditionalObserved (boundary μ i) (labelList i r) (measurable_labelList i r) (windowKey i) (measurable_windowKey i) (v,u)=sampledLaw i r u hr v := by let : IsProbabilityMeasure (boundary μ i) := boundary_isProbability μ hT i apply mass_ext funext w have hh := conditionalObserved_mass (boundary μ i) (labelList i r) (measurable_labelList i r) (windowKey i) (measurable_windowKey i) (v,u) w rw [boundary_window_joint μ hT i r u hr v w] at hh exact (mul_left_cancel₀ hp) hh end SmoothObstruction.WordHierarchy namespace SmoothObstruction.WordHierarchy open MeasureTheory Set Finset WordParameters WordStep FiniteLaw lemma sampledLaw_prefix (i : ℕ) (a : A (i+2)) (t : Offsets (S i)) {r : ℕ} (hr : r ≤ D (S i)) : sampledLaw i r ((finCongr (q_mul_H (S i))) t.val) (by have _ht := t.property; change t.val.val+r ≤ N (S i); rw [←q_mul_H]; omega) a = push (parentWindow i a t) (takePrefix hr) := by classical rw [parentWindow,push_comp] have he : (fun j : Fin (groupLength (i+1)) => takePrefix hr (window (S i) t ((stage i).word ((stage (i+1)).word a j)))) = sampledWindow i r ((finCongr (q_mul_H (S i))) t.val) (by have _ht := t.property; change t.val.val+r ≤ N (S i); rw [←q_mul_H]; omega) a ∘ (finCongr (q_mul_H (S (i+1)))) := by funext j k simp only [Function.comp_apply,sampledWindow,children,Equiv.symm_apply_apply,takePrefix,window] congr 2 change push uniform _ = push uniform _ dsimp only [Function.comp_def] rw [he] exact (uniform_reindex (finCongr (q_mul_H (S (i+1)))) _).symm noncomputable def windowLower (i r : ℕ) : ℝ := (r:ℝ)*((1-2/10000)*Real.log (Fintype.card (A i))-(2*K (S i):ℕ)*Real.log 2)-2*Real.log 2 lemma sampled_entropy (i : ℕ) (a : A (i+2)) (u : Fin (N (S i))) (hu : u.val+D (S i) ≤ N (S i)) {r : ℕ} (hr : r ≤ D (S i)) : windowLower i r ≤ entropy (sampledLaw i r u (by omega) a) := by let t : Offsets (S i) := ⟨(finCongr (q_mul_H (S i))).symm u,by change u.val+D (S i) ≤ q (S i)*H (S i) rwa [q_mul_H]⟩ have hh := parent_prefix_entropy i a t hr rw [←sampledLaw_prefix] at hh simpa only [t,windowLower,Equiv.apply_symm_apply] using hh /-- The entropy of the genuine stationary aligned orbit window is bounded using the exactly uniform internal-slot law, not a selected parent law. -/ theorem boundary_window_entropy (μ : Measure Extension) [IsProbabilityMeasure μ] (hT : MeasurePreserving T μ μ) (i r : ℕ) (hr : r ≤ D (S i)) : let : IsProbabilityMeasure (boundary μ i) := boundary_isProbability μ hT i (∑ b : A (i+2) × Fin (N (S i)), if b.2.val+D (S i) ≤ N (S i) then (observed (boundary μ i) (windowKey i) (measurable_windowKey i)).mass b else 0)*windowLower i r ≤ entropy (observed (boundary μ i) (labelList i r) (measurable_labelList i r)) := by classical let : IsProbabilityMeasure (boundary μ i) := boundary_isProbability μ hT i dsimp only rw [observed_mixture (boundary μ i) (labelList i r) (measurable_labelList i r) (windowKey i) (measurable_windowKey i)] apply le_trans _ (entropy_mixture _ _) rw [sum_mul] apply sum_le_sum rintro ⟨v,u⟩ _ by_cases hu : u.val+D (S i) ≤ N (S i) · rw [ite_eq_left hu] by_cases hp : (boundary μ i).real {x | windowKey i x=(v,u)}=0 · change (boundary μ i).real {x | windowKey i x=(v,u)}*_ ≤ _ simp only [hp,zero_mul,observed,le_refl] · rw [conditional_window_eq μ hT i r u (by omega) v hp] exact mul_le_mul_of_nonneg_left (sampled_entropy i v u hu hr) measureReal_nonneg · rw [ite_eq_right hu,zero_mul] exact mul_nonneg measureReal_nonneg (entropy_nonneg _) end SmoothObstruction.WordHierarchy namespace SmoothObstruction.WordHierarchy open MeasureTheory Set Finset WordParameters WordStep FiniteLaw lemma D_pos (s : ℕ) : 0<D s := by unfold D; positivity lemma good_offset_count (s : ℕ) : #{u : Fin (N s) | u.val+D s ≤ N s}=N s-D s+1 := by have hD := D_le_N s have hp := D_pos s have he : (univ.filter (fun u : Fin (N s) => u.val+D s ≤ N s))= (univ.filter (fun u : Fin (N s) => u.val < N s-D s+1)) := by apply Finset.filter_congr intro u hu omega rw [he,Fin.card_filter_val_lt,min_eq_right (by omega)] lemma good_key_mass (μ : Measure Extension) [IsProbabilityMeasure μ] (hT : MeasurePreserving T μ μ) (i : ℕ) : let : IsProbabilityMeasure (boundary μ i) := boundary_isProbability μ hT i (∑ b : A (i+2) × Fin (N (S i)), if b.2.val+D (S i) ≤ N (S i) then (observed (boundary μ i) (windowKey i) (measurable_windowKey i)).mass b else 0)= (N (S i)-D (S i)+1:ℕ)/(N (S i):ℝ) := by classical let : IsProbabilityMeasure (boundary μ i) := boundary_isProbability μ hT i dsimp only rw [Fintype.sum_prod_type] simp only [observed,boundary_key_mass μ hT] have hs (v : A (i+2)) : (∑ u : Fin (N (S i)),if u.val+D (S i) ≤ N (S i) then μ.real {x | label (i+2) x=v}/N (S i) else 0)= (N (S i)-D (S i)+1:ℕ)*(μ.real {x | label (i+2) x=v}/N (S i)) := by rw [←sum_filter] simp only [sum_const,nsmul_eq_mul,good_offset_count] simp only [hs] rw [←mul_sum,←sum_div] have hm := (observed μ (label (i+2)) (measurable_label (i+2))).total change (∑ v,μ.real {x | label (i+2) x=v})=1 at hm rw [hm,mul_one_div] /-- The manuscript's explicit stationary entropy lower bound, with the exact loss from inadmissible offsets and an arbitrary invariant probability. -/ theorem stationary_window_entropy (μ : Measure Extension) [IsProbabilityMeasure μ] (hT : MeasurePreserving T μ μ) (i r : ℕ) (hr : r ≤ D (S i)) : let : IsProbabilityMeasure (boundary μ i) := boundary_isProbability μ hT i ((N (S i)-D (S i)+1:ℕ)/(N (S i):ℝ))*windowLower i r ≤ entropy (observed (boundary μ i) (labelList i r) (measurable_labelList i r)) := by have hh := boundary_window_entropy μ hT i r hr let : IsProbabilityMeasure (boundary μ i) := boundary_isProbability μ hT i dsimp only at hh ⊢ rw [good_key_mass μ hT i] at hh exact hh end SmoothObstruction.WordHierarchy namespace SmoothObstruction.WordHierarchy open MeasureTheory Set Finset WordParameters WordAlphabet WordStep FiniteLaw lemma log_card (i : ℕ) : Real.log (Fintype.card (A i))=ell (S i)*Real.log 2 := by change Real.log (Fintype.card (Alphabet (S i)))=_ rw [cardAlphabetReal] dsimp only [ell,alphabetSize,Real.logb] rw [div_mul_cancel₀ _ (ne_of_gt (Real.log_pos (by norm_num : (1:ℝ)<2)))] lemma K_large {i : ℕ} (hi : 1 ≤ i) : 4000 ≤ K (S i) := by have h := K_double_exponential initial_threshold i have hpow : 2 ≤ (2:ℕ)^i := by calc _ = (2:ℕ)^1 := by norm_num _ ≤ _ := Nat.pow_le_pow_right (by decide) hi have hh : (2:ℕ)^16 ≤ 2^(8*2^i) := Nat.pow_le_pow_right (by decide) (by omega) exact (by norm_num : 4000 ≤ (2:ℕ)^16).trans (hh.trans h) lemma bad_offset_small (i : ℕ) : (D (S i):ℝ)/(N (S i):ℝ) ≤ 1/2000 := by have hk := (step_conditions initial_threshold i).1 change 256 ≤ K (S i) at hk have he : K (S i)/8+11 ≤ K (S i) := by omega have hp : 2048*D (S i) ≤ N (S i) := by calc _ = 2^(K (S i)/8+11) := by rw [pow_add]; norm_num [D]; ring _ ≤ _ := Nat.pow_le_pow_right (by decide) he have hr : (2000:ℝ)*(D (S i):ℝ) ≤ N (S i) := by exact_mod_cast (show 2000*D (S i) ≤ N (S i) by omega) have hN : (0:ℝ)<N (S i) := by exact_mod_cast N_pos (S i) apply (div_le_iff₀ hN).mpr linarith lemma good_offset_large (i : ℕ) : (1999/2000:ℝ) ≤ (N (S i)-D (S i)+1:ℕ)/(N (S i):ℝ) := by have hD := D_le_N (S i) have hN : (0:ℝ)<N (S i) := by exact_mod_cast N_pos (S i) have hb := bad_offset_small i rw [Nat.cast_add,Nat.cast_sub hD,Nat.cast_one] apply (le_div_iff₀ hN).mpr have hb' := (div_le_iff₀ hN).mp hb nlinarith lemma windowLower_large {i r : ℕ} (hi : 1 ≤ i) (hr : 1 ≤ r) : (2497/2500:ℝ)*r*Real.log (Fintype.card (A i)) ≤ windowLower i r := by have hk : (4000:ℝ)≤K (S i) := by exact_mod_cast K_large hi have he := (K_bounds (S i)).1 have hK : 2*(K (S i):ℝ) ≤ ell (S i)/2000 := by nlinarith have h1 : (2:ℝ) ≤ ell (S i)/2000 := by nlinarith have hlog := Real.log_pos (by norm_num : (1:ℝ)<2) have hr' : (1:ℝ)≤r := by exact_mod_cast hr have hepos : 0 ≤ ell (S i) := ell_nonneg _ have hklog := mul_le_mul_of_nonneg_right hK hlog.le have h1log := mul_le_mul_of_nonneg_right h1 hlog.le have hklogr := mul_le_mul_of_nonneg_left hklog (show (0:ℝ)≤r by positivity) have hcost : 2*Real.log 2 ≤ (r:ℝ)*(ell (S i)/2000*Real.log 2) := by exact h1log.trans (le_mul_of_one_le_left (by positivity) hr') rw [windowLower,log_card] push_cast nlinarith /-- Uniform stationary window entropy, with the manuscript's decimal constant. It is valid already at level one beyond the fixed initial level, independently of the stationary probability and independently of the window length. Entropy here uses natural logarithms, so both sides have the same conversion factor. -/ theorem stationary_window_entropy_996 (μ : Measure Extension) [IsProbabilityMeasure μ] (hT : MeasurePreserving T μ μ) {i r : ℕ} (hi : 1 ≤ i) (hr : 1 ≤ r) (hrD : r ≤ D (S i)) : let : IsProbabilityMeasure (boundary μ i) := boundary_isProbability μ hT i (996/1000:ℝ)*r*Real.log (Fintype.card (A i)) ≤ entropy (observed (boundary μ i) (labelList i r) (measurable_labelList i r)) := by let : IsProbabilityMeasure (boundary μ i) := boundary_isProbability μ hT i dsimp only have hh := stationary_window_entropy μ hT i r hrD dsimp only at hh have hl := windowLower_large hi hr have hg := good_offset_large i have hlog : 0 ≤ Real.log (Fintype.card (A i)) := by rw [log_card] exact mul_nonneg (ell_nonneg _) (Real.log_pos (by norm_num : (1:ℝ)<2)).le have hpos : 0 ≤ (2497/2500:ℝ)*r*Real.log (Fintype.card (A i)) := by positivity calc _ ≤ (1999/2000:ℝ)*((2497/2500:ℝ)*r*Real.log (Fintype.card (A i))) := by nlinarith _ ≤ ((N (S i)-D (S i)+1:ℕ)/(N (S i):ℝ))*windowLower i r := mul_le_mul hg hl hpos (by positivity) _ ≤ _ := hh end SmoothObstruction.WordHierarchy namespace SmoothObstruction.WordHierarchy open WordParameters WordAlphabet WordStep WordConcatenation MeasureTheory Set Finset lemma graph_block_room (i k : ℕ) (hk : k ≤ K (S i)/8) : 2*2^k ≤ H (S i) := by have hK := (step_conditions initial_threshold i).1 change 256 ≤ K (S i) at hK have he : k+1 ≤ K (S i)/2 := by omega calc _ = 2^(k+1) := by rw [pow_succ]; omega _ ≤ _ := Nat.pow_le_pow_right (by decide) he lemma group_site_lt (i : ℕ) (u : Fin (N (S i))) {r : ℕ} (hr : r<H (S i)) : H (S i)*(u.val/H (S i))+r < N (S i) := by have hd : u.val/H (S i)<q (S i) := by apply (Nat.div_lt_iff_lt_mul (H_pos (S i))).mpr rw [q_mul_H] exact u.isLt have hh : (u.val/H (S i)+1)*H (S i) ≤ q (S i)*H (S i) := Nat.mul_le_mul_right _ hd rw [q_mul_H] at hh nlinarith lemma children_graph (i k : ℕ) (hk : k ≤ K (S i)/8) (a : A (i+1)) (u : Fin (N (S i))) (j : ℕ) (hr : u.val%H (S i)+j+2^k < H (S i)) : let h₁ : u.val+j < N (S i) := by have hh := group_site_lt i u (show u.val%H (S i)+j<H (S i) from lt_of_le_of_lt (Nat.le_add_right _ _) hr) have he := Nat.div_add_mod u.val (H (S i)) omega let h₂ : u.val+j+2^k < N (S i) := by have hh := group_site_lt i u hr have he := Nat.div_add_mod u.val (H (S i)) omega form (S i) (children i a ⟨u.val+j,h₁⟩).val (children i a ⟨u.val+j+2^k,h₂⟩).val=0 := by dsimp only have hK := (step_conditions initial_threshold i).1 change 256 ≤ K (S i) at hK have hkk : k<K (S i) := by omega let b : Fin (q (S i)) := ⟨u.val/H (S i),by apply (Nat.div_lt_iff_lt_mul (H_pos (S i))).mpr rw [q_mul_H] exact u.isLt⟩ have hh := (stage i).relations ((Fintype.equivFin (A (i+1))) a) b (u.val%H (S i)+j) k hkk hk hr have he₁ : (finCongr (q_mul_H (S i))).symm (⟨u.val+j,by have hh := group_site_lt i u (show u.val%H (S i)+j<H (S i) from lt_of_le_of_lt (Nat.le_add_right _ _) hr) have he := Nat.div_add_mod u.val (H (S i)); omega⟩ : Fin (N (S i))) = finProdFinEquiv (b,⟨u.val%H (S i)+j,lt_of_le_of_lt (Nat.le_add_right _ _) hr⟩) := by apply Fin.ext simp only [finProdFinEquiv_apply_val] change u.val+j = u.val%H (S i)+j+H (S i)*(u.val/H (S i)) have he := Nat.mod_add_div u.val (H (S i)) omega have he₂ : (finCongr (q_mul_H (S i))).symm (⟨u.val+j+2^k,by have hh := group_site_lt i u hr have he := Nat.div_add_mod u.val (H (S i)); omega⟩ : Fin (N (S i))) = finProdFinEquiv (b,⟨u.val%H (S i)+j+2^k,hr⟩) := by apply Fin.ext simp only [finProdFinEquiv_apply_val] change u.val+j+2^k = u.val%H (S i)+j+2^k+H (S i)*(u.val/H (S i)) have he := Nat.mod_add_div u.val (H (S i)) omega simpa only [children,Family.word,he₁,he₂] using hh /-- The genuine two-window graph event on a boundary-started orbit. -/ def graphWindows (i k : ℕ) (x : Extension) : Prop := ∀ j : Fin (2^k-1), form (S i) (label i ((T^[j.val*length 40000 i]) x)).val (label i ((T^[(j.val+2^k)*length 40000 i]) x)).val=0 lemma graphWindows_of_good (i k : ℕ) (hk : k ≤ K (S i)/8) (x : Extension) (hx : phase i x=0) (hgood : (slot i x).val%H (S i)+(2*2^k-1) ≤ H (S i)) : graphWindows i k x := by intro j have hD : 0<(2:ℕ)^k := by positivity have hj := j.isLt have hr : (slot i x).val%H (S i)+j.val+2^k < H (S i) := by omega have hh := children_graph i k hk (label (i+1) x) (slot i x) j.val hr dsimp only at hh have he := Nat.div_add_mod (slot i x).val (H (S i)) have h₁ : (slot i x).val+j.val<N (S i) := by have hh := group_site_lt i (slot i x) (show (slot i x).val%H (S i)+j.val<H (S i) by omega) omega have h₂ : (slot i x).val+(j.val+2^k)<N (S i) := by have hh := group_site_lt i (slot i x) hr omega rw [label_advance_child i j.val x hx h₁,label_advance_child i (j.val+2^k) x hx h₂] simpa only [Nat.add_assoc] using hh end SmoothObstruction.WordHierarchy namespace SmoothObstruction.WordHierarchy open WordParameters WordAlphabet WordStep WordConcatenation MeasureTheory Set Finset FiniteLaw lemma boundary_slot_mass (μ : Measure Extension) [IsProbabilityMeasure μ] (hT : MeasurePreserving T μ μ) (i : ℕ) (u : Fin (N (S i))) : (boundary μ i).real {x | slot i x=u}=1/(N (S i):ℝ) := by let : IsProbabilityMeasure (boundary μ i) := boundary_isProbability μ hT i rw [←Alignment.sum_fibers (boundary μ i) {x | slot i x=u} (measurable_label (i+2))] have he (v : A (i+2)) : {x | x∈{x | slot i x=u} ∧ label (i+2) x=v}= {x | windowKey i x=(v,u)} := by ext x simp only [Set.mem_ofPred_eq,windowKey,Prod.mk.injEq,and_comm] simp only [he,boundary_key_mass μ hT] rw [←sum_div] have hm := (observed μ (label (i+2)) (measurable_label (i+2))).total change (∑ v,μ.real {x | label (i+2) x=v})=1 at hm rw [hm] lemma residue_sum (i c : ℕ) (hc : c ≤ H (S i)) : (∑ u : Fin (N (S i)), if u.val%H (S i)<c then (1:ℝ)/N (S i) else 0)= (c:ℝ)/H (S i) := by classical let e : Fin (q (S i)) × Fin (H (S i)) ≃ Fin (N (S i)) := finProdFinEquiv.trans (finCongr (q_mul_H (S i))) have he (v : Fin (q (S i)) × Fin (H (S i))) : (e v).val%H (S i)=v.2.val := by change (v.2.val+H (S i)*v.1.val)%H (S i)=v.2.val simp only [Nat.add_mod,Nat.mul_mod,Nat.mod_self,zero_mul,Nat.zero_mod,Nat.add_zero, Nat.mod_eq_of_lt v.2.isLt] rw [←e.sum_comp] simp only [he,Fintype.sum_prod_type] have hs : (∑ r : Fin (H (S i)),if r.val<c then (1:ℝ)/N (S i) else 0)= (c:ℝ)/N (S i) := by rw [←sum_filter,sum_const,nsmul_eq_mul,Fin.card_filter_val_lt,min_eq_right hc,mul_one_div] simp only [hs,sum_const,Finset.card_univ,Fintype.card_fin,nsmul_eq_mul] rw [←q_mul_H (S i),Nat.cast_mul] have hq : (q (S i):ℝ)≠0 := by exact_mod_cast (q_pos (S i)).ne' field_simp lemma boundary_residue_mass (μ : Measure Extension) [IsProbabilityMeasure μ] (hT : MeasurePreserving T μ μ) (i c : ℕ) (hc : c ≤ H (S i)) : (boundary μ i).real {x | (slot i x).val%H (S i)<c}=(c:ℝ)/H (S i) := by classical let : IsProbabilityMeasure (boundary μ i) := boundary_isProbability μ hT i rw [←Alignment.sum_fibers (boundary μ i) {x | (slot i x).val%H (S i)<c} (measurable_slot i)] have hs (u : Fin (N (S i))) : (boundary μ i).real {x | x∈{x | (slot i x).val%H (S i)<c} ∧ slot i x=u}= if u.val%H (S i)<c then (1:ℝ)/N (S i) else 0 := by split_ifs with hu · have he : {x | x∈{x | (slot i x).val%H (S i)<c} ∧ slot i x=u}={x | slot i x=u} := by ext x simp only [Set.mem_ofPred_eq] constructor · exact And.right · intro hx; exact ⟨by simpa only [hx] using hu,hx⟩ rw [he,boundary_slot_mass μ hT] · have he : {x | x∈{x | (slot i x).val%H (S i)<c} ∧ slot i x=u}=∅ := by ext x simp only [Set.mem_ofPred_eq,Set.mem_empty_iff_false,iff_false] rintro ⟨hr,hx⟩ exact hu (by simpa only [hx] using hr) rw [he,measureReal_empty] simp only [hs] exact residue_sum i c hc /-- The two lists of the actual system pass all graph tests simultaneously except for the precise group-boundary fraction (2D-2)/H. -/ theorem boundary_graph_probability (μ : Measure Extension) [IsProbabilityMeasure μ] (hT : MeasurePreserving T μ μ) (i k : ℕ) (hk : k ≤ K (S i)/8) : 1-((2*2^k-2:ℕ):ℝ)/H (S i) ≤ (boundary μ i).real {x | graphWindows i k x} := by let : IsProbabilityMeasure (boundary μ i) := boundary_isProbability μ hT i have hD : 0<(2:ℕ)^k := by positivity have hH := graph_block_room i k hk let c := H (S i)-(2*2^k-1)+1 have hc : c ≤ H (S i) := by dsimp only [c]; omega have he : {x : Extension | (slot i x).val%H (S i)<c}= {x | (slot i x).val%H (S i)+(2*2^k-1)≤H (S i)} := by ext x change (slot i x).val%H (S i)<c ↔ (slot i x).val%H (S i)+(2*2^k-1)≤H (S i) dsimp only [c] omega have hm := boundary_residue_mass μ hT i c hc rw [he] at hm have hle : (boundary μ i).real {x | (slot i x).val%H (S i)+(2*2^k-1)≤H (S i)} ≤ (boundary μ i).real {x | graphWindows i k x} := by apply ENNReal.toReal_mono (measure_ne_top _ _) apply measure_mono_ae filter_upwards [boundary_phase_zero μ i] with x hx exact fun hg => graphWindows_of_good i k hk x hx hg rw [hm] at hle have hc' : c=H (S i)-(2*2^k-2) := by dsimp only [c]; omega have hsub : 2*2^k-2 ≤ H (S i) := by omega rw [hc',Nat.cast_sub hsub] at hle have hHp : (H (S i):ℝ)≠0 := by exact_mod_cast (H_pos (S i)).ne' convert hle using 1 field_simp end SmoothObstruction.WordHierarchy end section open Finset namespace SmoothObstruction.FiniteLaw open MeasureTheory Set Finset variable {X α : Type*} [MeasurableSpace X] [Fintype α] [MeasurableSpace α] [MeasurableSingletonClass α] /-- Every finite observable's real expectation agrees with integration in its original probability space. No conditional normalization occurs. -/ theorem observed_integral (μ : Measure X) [IsProbabilityMeasure μ] (f : X → α) (hf : Measurable f) (g : α → ℝ) : (∑ a,(observed μ f hf).mass a*g a)=∫ x,g (f x) ∂μ := by calc _ = ∫ a,g a ∂(μ.map f) := by rw [integral_fintype (Integrable.of_finite (μ:=μ.map f))] apply sum_congr rfl intro a _ simp only [observed,Measure.real,Measure.map_apply hf (measurableSet_singleton a), smul_eq_mul] rfl _ = _ := integral_map hf.aemeasurable (measurable_of_countable g).aestronglyMeasurable /-- A lower bound on a success event is a lower bound in the UNCONDITIONED law. In particular .95 mass at statistic .96 gives expectation .912. -/ theorem integral_lower_of_event (μ : Measure X) [IsProbabilityMeasure μ] (Z : X → ℝ) (hZi : Integrable Z μ) (hZn : ∀ x,0≤Z x) (G : Set X) (hG : MeasurableSet G) (a b : ℝ) (ha : 0≤a) (hGm : b≤μ.real G) (hZG : ∀ x∈G,a≤Z x) : a*b≤∫ x,Z x ∂μ := by calc a*b ≤ μ.real G*a := by nlinarith _ = ∫ x,G.indicator (fun _ => a) x ∂μ := by rw [integral_indicator_const a hG,smul_eq_mul] _ ≤ _ := integral_mono_ae ((integrable_const a).indicator hG) hZi (Filter.Eventually.of_forall (fun x => by by_cases hx : x∈G · simpa only [indicator_of_mem hx] using hZG x hx · simpa only [indicator_of_notMem hx] using hZn x)) end SmoothObstruction.FiniteLaw end end OAI
import OAI.Dynamics.SmoothObstruction.WordHierarchy
namespace OAI
section
namespace SmoothObstruction.WordFlatten
open Finset FiniteLaw WordLaw SequentialLaw
variable {A : Type*} [Fintype A] [DecidableEq A] {q H : ℕ}
omit [Fintype A] in
lemma test_flatten (m : Fin (q*H) → Option A) (w : Fin q → Fin H → A) :
test m (flattenEquiv w) = ∏ b, test (fun r => m (finProdFinEquiv (b,r))) (w b) := by
unfold test
calc
_ = ∏ br : Fin q × Fin H, mark (m (finProdFinEquiv br)) (w br.1 br.2) := by
simpa only [flattenEquiv, Equiv.coe_fn_mk,Equiv.symm_apply_apply] using
(finProdFinEquiv.prod_comp (fun j => mark (m j) (flattenEquiv w j))).symm
_ = _ := Fintype.prod_prod_type _
omit [Fintype A] [DecidableEq A] in
lemma count_grouped (m : Fin (q*H) → Option A) :
(∑ b : Fin q, (univ.filter (fun r : Fin H => (m (finProdFinEquiv (b,r))).isSome)).card) =
(univ.filter (fun r => (m r).isSome)).card := by
simp only [card_eq_sum_ones, sum_filter]
calc
_ = ∑ br : Fin q × Fin H, if (m (finProdFinEquiv br)).isSome then 1 else 0 :=
(Fintype.sum_prod_type (fun br : Fin q × Fin H =>
if (m (finProdFinEquiv br)).isSome then (1:ℕ) else 0)).symm
_ = _ := finProdFinEquiv.sum_comp (fun r => if (m r).isSome then 1 else 0)
variable {V : Type*} [AddCommGroup V] [Module (ZMod 2) V] [FiniteDimensional (ZMod 2) V]
[Fintype V] [DecidableEq V] (B : LinearMap.BilinForm (ZMod 2) V) {K : ℕ}
omit [FiniteDimensional (ZMod 2) V] in
/-- W3 itself: the actual single conditioned law controls every arbitrary
partial list, with a single factor two, including lists crossing groups. -/
theorem flat_sublist (hd : K+1 ≤ Module.finrank (ZMod 2) V) (hK : 1 ≤ K)
(E : Finset (Fin q → Fin H → {v : V // v ≠ 0}))
(hE : 1/2 ≤ prob (rawLaw B hd H q) E)
(m : Fin (q*H) → Option {v : V // v ≠ 0}) :
expect (flatLaw B hd E hE) (test m) ≤
2 * ((2 : ℝ)^(2*K) / Fintype.card {v : V // v ≠ 0}) ^
(univ.filter (fun r => (m r).isSome)).card := by
unfold flatLaw
rw [expect_transport]
have ht : ((test m) ∘ flattenEquiv) =
(fun w => ∏ b, test (fun r => m (finProdFinEquiv (b,r))) (w b)) := by
funext w
exact test_flatten m w
rw [ht]
have hh := conditioned_sublist B hd hK E hE (fun b r => m (finProdFinEquiv (b,r)))
simpa only [count_grouped] using hh
end SmoothObstruction.WordFlatten
namespace SmoothObstruction.FiniteLaw
open Finset FiniteEntropy
variable {α β γ : Type*} [Fintype α] [Fintype β] [Fintype γ]
noncomputable def uniform [Nonempty α] : Law α where
mass := fun _ => (Fintype.card α : ℝ)⁻¹
nonneg := fun _ => by positivity
total := by simp [Fintype.card_ne_zero]
noncomputable def mixture (P : Law α) (Q : α → Law β) : Law β where
mass b := ∑ a,P.mass a*(Q a).mass b
nonneg b := sum_nonneg (fun a _ => mul_nonneg (P.nonneg a) ((Q a).nonneg b))
total := by
rw [sum_comm]
simp only [←mul_sum,(Q _).total,mul_one,P.total]
noncomputable def TV (P Q : Law α) : ℝ := (1/2:ℝ)*∑ a,|P.mass a-Q.mass a|
noncomputable def entropy (P : Law α) : ℝ := natEntropy P.mass
lemma TV_nonneg (P Q : Law α) : 0 ≤ TV P Q := by unfold TV; positivity
lemma TV_symm (P Q : Law α) : TV P Q=TV Q P := by simp only [TV,abs_sub_comm]
lemma TV_triangle (P Q R : Law α) : TV P R ≤ TV P Q+TV Q R := by
unfold TV
rw [←mul_add,←sum_add_distrib]
exact mul_le_mul_of_nonneg_left (sum_le_sum (fun a _ => abs_sub_le (P.mass a) (Q.mass a) (R.mass a))) (by norm_num)
lemma push_sub [DecidableEq β] (P Q : Law α) (f : α → β) (b : β) :
(push P f).mass b-(push Q f).mass b = ∑ a,if f a=b then P.mass a-Q.mass a else 0 := by
classical
simp only [push,←sum_sub_distrib]
apply sum_congr rfl
intro a _
split_ifs <;> ring
lemma TV_push [DecidableEq β] (P Q : Law α) (f : α → β) : TV (push P f) (push Q f) ≤ TV P Q := by
classical
unfold TV
apply mul_le_mul_of_nonneg_left _ (by norm_num)
simp_rw [push_sub]
calc
_ ≤ ∑ b,∑ a,|if f a=b then P.mass a-Q.mass a else 0| := sum_le_sum (fun b _ => abs_sum_le_sum_abs _ _)
_ = ∑ a,|P.mass a-Q.mass a| := by
rw [sum_comm]
simp only [apply_ite abs,abs_zero,Finset.sum_ite_eq,Finset.mem_univ,ite_true]