openAI-math-dataset / lean /OAI /Computability /MaxCut /Games /CanonicalAddress.lean
roseking's picture
Add lean source and binary files (part 61)
ea76aac verified
Raw History Blame Contribute Delete
20.3 kB
import Mathlib
import OAI.Computability.MaxCut.Games.TranslationTarget
import OAI.Computability.MaxCut.Encoding.TableKeysCompleteness
namespace OAI
/-! Direct canonical-body addresses. The fixed list of canonical words is
interpreted in big-endian order, matching forward Horner steps `B*acc+digit`.
Unused addresses represent isolated vertices. These are semantic encoding and
polynomial-size results, not an assumed machine implementation. -/
namespace MaxCutGames.Reduction.CanonicalAddress
open CanonicalEncoding
/-- Most-significant-word first, with leading zero words retained in the
fixed-length domain. Mathlib's `ofDigits` uses the opposite list order. -/
def radix (B : Nat) (words : List Nat) : Nat := Nat.ofDigits B words.reverse
theorem radix_eq_foldl (B : Nat) (words : List Nat) :
radix B words = words.foldl (fun acc digit => B * acc + digit) 0 := by
unfold radix
rw [Nat.ofDigits_eq_foldr, List.foldr_reverse]
simp only [Nat.cast_id, Nat.add_comm]
@[simp] theorem radix_nil (B : Nat) : radix B [] = 0 := rfl
/-- This is precisely the arithmetic operation of `MachineRadixStep`. -/
theorem radix_append_word (B : Nat) (words : List Nat) (digit : Nat) :
radix B (words ++ [digit]) = B * radix B words + digit := by
simp only [radix_eq_foldl, List.foldl_append, List.foldl_cons, List.foldl_nil]
theorem radix_lt_pow_length {B : Nat} (hB : 1 < B) (words : List Nat)
(bounded : ∀ w ∈ words, w < B) : radix B words < B ^ words.length := by
unfold radix
simpa only [List.length_reverse] using
Nat.ofDigits_lt_base_pow_length hB
(fun w hw => bounded w (List.mem_reverse.mp hw))
/-- Fixed width makes leading zero words unambiguous. -/
theorem radix_injective_of_length_eq {B : Nat} (hB : 1 < B)
{xs ys : List Nat} (sameLength : xs.length = ys.length)
(xsBounded : ∀ w ∈ xs, w < B) (ysBounded : ∀ w ∈ ys, w < B)
(sameAddress : radix B xs = radix B ys) : xs = ys := by
have reversed : xs.reverse = ys.reverse :=
Nat.ofDigits_inj_of_len_eq hB (by simpa only [List.length_reverse] using sameLength)
(fun w hw => xsBounded w (List.mem_reverse.mp hw))
(fun w hw => ysBounded w (List.mem_reverse.mp hw)) sameAddress
have h := congrArg List.reverse reversed
simpa only [List.reverse_reverse] using h
/-- One larger than the proved bound on every canonical word. -/
def base (n m s d : Nat) : Nat := wordBound n m s d + 1
@[simp] theorem base_eq (n m s d : Nat) :
base n m s d = n + m + 2^s + 2^d + 4 := by
simp [base, wordBound, Nat.add_assoc]
theorem one_lt_base (n m s d : Nat) : 1 < base n m s d := by
have h : 3 ≤ wordBound n m s d := by
unfold wordBound
exact Nat.le_add_left 3 _
unfold base
omega
theorem bodyWords_lt_base {n m k s d : Nat} (body : Body n m k s d)
(w : Nat) (hw : w ∈ bodyWords body) : w < base n m s d :=
Nat.lt_succ_of_le (bodyWords_bounded body w hw)
def capacity (n m k s d : Nat) : Nat := (base n m s d) ^ (1 + 9*k)
theorem capacity_pos (n m k s d : Nat) : 0 < capacity n m k s d :=
pow_pos (Nat.zero_lt_of_lt (one_lt_base n m s d)) _
def bodyAddress {n m k s d : Nat} (body : Body n m k s d) :
Fin (capacity n m k s d) :=
⟨radix (base n m s d) (bodyWords body), by
simpa only [capacity, bodyWords_length] using
radix_lt_pow_length (one_lt_base n m s d) (bodyWords body) (bodyWords_lt_base body)⟩
@[simp] theorem bodyAddress_val {n m k s d : Nat} (body : Body n m k s d) :
(bodyAddress body).val = radix (base n m s d) (bodyWords body) := rfl
theorem bodyAddress_injective {n m k s d : Nat} :
Function.Injective (bodyAddress (n := n) (m := m) (k := k) (s := s) (d := d)) := by
intro x y h
apply bodyWords_injective
exact radix_injective_of_length_eq (one_lt_base n m s d)
(by simp only [bodyWords_length]) (bodyWords_lt_base x) (bodyWords_lt_base y)
(congrArg Fin.val h)
@[simp] theorem bodyAddress_eq_iff {n m k s d : Nat} (x y : Body n m k s d) :
bodyAddress x = bodyAddress y ↔ x = y := bodyAddress_injective.eq_iff
/-- For fixed k,s,d, the entire direct-address space is a polynomial in n+m. -/
noncomputable def capacityPolynomial (k s d : Nat) : Polynomial Nat :=
(Polynomial.X + Polynomial.C (2^s + 2^d + 4)) ^ (1 + 9*k)
theorem capacityPolynomial_eval (n m k s d : Nat) :
(capacityPolynomial k s d).eval (n + m) = capacity n m k s d := by
simp [capacityPolynomial, capacity, base_eq, Nat.add_assoc]
end MaxCutGames.Reduction.CanonicalAddress
/-! Finite noise tables for the executable outer reduction. Positions, including
positions containing equal vectors, remain distinct sample outcomes. -/
namespace MaxCutGames.Integration.NoiseTables
open BinaryLinear MaxCutGames.Reduction.ActualSource
variable {s d : Nat}
/-- Reindexing changes the names of outcomes, retaining their uniform law. -/
def reindex (g : SplitGadget s d) {I : Type} [Fintype I] [Nonempty I]
(e : I ≃ g.NoiseIndex) : SplitGadget s d where
f := g.f
equivariant := g.equivariant
NoiseIndex := I
noiseFintype := inferInstance
noiseNonempty := inferInstance
noise i := g.noise (e i)
theorem reindex_stabilityError (g : SplitGadget s d)
{I : Type} [Fintype I] [Nonempty I] (e : I ≃ g.NoiseIndex) :
(reindex g e).stabilityError = g.stabilityError := by
apply Fintype.expect_equiv ((Equiv.refl (Ambient s d)).prodCongr e)
intro x
rfl
theorem reindex_kernelProbability (g : SplitGadget s d)
{I : Type} [Fintype I] [Nonempty I] (e : I ≃ g.NoiseIndex)
{P : Type} [AddCommGroup P] [Module F2 P]
(S : Ambient s d →ₗ[F2] P) :
SplitGadget.kernelProbability (reindex g e) S =
SplitGadget.kernelProbability g S := by
classical
apply Fintype.expect_equiv e
intro i
rfl
/-- Only a nonempty finite list of binary vectors is required as static noise
data. The correcting function is a separate witness used in the proof. -/
structure Table (s d : Nat) where
vectors : List (Ambient s d)
nonempty : vectors ≠ []
namespace Table
def gadget (T : Table s d) (f : Ambient s d → Alphabet s)
(hf : ∀ x c, f (x + (c, 0)) = f x + c) : SplitGadget s d where
f := f
equivariant := hf
NoiseIndex := Fin T.vectors.length
noiseFintype := inferInstance
noiseNonempty := ⟨⟨0, List.length_pos_iff.mpr T.nonempty⟩⟩
noise := T.vectors.get
def ofEnumeration (g : SplitGadget s d) (en : NoiseEnumeration g) : Table s d where
vectors := en.indices.map g.noise
nonempty := by simpa using en.nonempty
@[simp] theorem ofEnumeration_length (g : SplitGadget s d) (en : NoiseEnumeration g) :
(ofEnumeration g en).vectors.length = en.indices.length := by
simp [ofEnumeration]
/-- The inverse equivalence is used only to prove the probability identity;
the forward noise lookup is ordinary finite-list indexing. -/
noncomputable def positionEquiv (g : SplitGadget s d) (en : NoiseEnumeration g) :
Fin (ofEnumeration g en).vectors.length ≃ g.NoiseIndex := by
classical
exact (finCongr (ofEnumeration_length g en)).trans
(en.nodup.getEquivOfForallMemList en.indices en.complete)
theorem noise_positionEquiv (g : SplitGadget s d) (en : NoiseEnumeration g)
(i : Fin (ofEnumeration g en).vectors.length) :
((ofEnumeration g en).gadget g.f g.equivariant).noise i =
g.noise (positionEquiv g en i) := by
change (en.indices.map g.noise)[i.val] = g.noise en.indices[i.val]
exact List.getElem_map _
theorem stabilityError_ofEnumeration (g : SplitGadget s d) (en : NoiseEnumeration g) :
((ofEnumeration g en).gadget g.f g.equivariant).stabilityError =
g.stabilityError := by
apply Fintype.expect_equiv
((Equiv.refl (Ambient s d)).prodCongr (positionEquiv g en))
intro x
change (if g.f (x.1 + _) = g.f x.1 then (0 : ℚ) else 1) = _
rw [noise_positionEquiv]
rfl
theorem kernelProbability_ofEnumeration (g : SplitGadget s d)
(en : NoiseEnumeration g) {P : Type} [AddCommGroup P] [Module F2 P]
(S : Ambient s d →ₗ[F2] P) :
SplitGadget.kernelProbability ((ofEnumeration g en).gadget g.f g.equivariant) S =
SplitGadget.kernelProbability g S := by
classical
apply Fintype.expect_equiv (positionEquiv g en)
intro i
rw [noise_positionEquiv]
/-- Exact table conversion retains both mathematical gadget guarantees. -/
theorem guarantees_ofEnumeration (g : SplitGadget s d) (en : NoiseEnumeration g)
{ζ ν : ℚ} {r : Nat}
(hstable : g.stabilityError ≤ ζ)
(hdisp : ∀ (P : Type) [AddCommGroup P] [Module F2 P]
(S : Ambient s d →ₗ[F2] P),
r ≤ Module.finrank F2 (S.comp (alphabetEmbedding s d)).range →
SplitGadget.kernelProbability g S ≤ ν) :
((ofEnumeration g en).gadget g.f g.equivariant).stabilityError ≤ ζ ∧
∀ (P : Type) [AddCommGroup P] [Module F2 P]
(S : Ambient s d →ₗ[F2] P),
r ≤ Module.finrank F2 (S.comp (alphabetEmbedding s d)).range →
SplitGadget.kernelProbability
((ofEnumeration g en).gadget g.f g.equivariant) S ≤ ν := by
constructor
· simpa only [stabilityError_ofEnumeration] using hstable
· intro P _ _ S hS
rw [kernelProbability_ofEnumeration]
exact hdisp P S hS
end Table
end MaxCutGames.Integration.NoiseTables
namespace MaxCutGames.Integration.TableReduction
open BinaryLinear NoiseTables MaxCutGames.Reduction.ActualSource
open MaxCutGames.Reduction.ActualGame MaxCutGames.Foundations.Target
variable {s d : Nat}
def tableEnumeration (T : Table s d) (f : Ambient s d → Alphabet s)
(hf : ∀ x c, f (x + (c, 0)) = f x + c) : NoiseEnumeration (T.gadget f hf) where
indices := List.finRange T.vectors.length
nodup := List.nodup_finRange _
complete i := List.mem_finRange i
/-- The projection is used only to fill an erased witness field. It is never
used when generating constraints or evaluating their permutation tables. -/
def tableSkeleton (T : Table s d) : SplitGadget s d :=
T.gadget Prod.fst (fun _ _ => rfl)
/-- A total executable map with finite vector data and no nonlinear-map input. -/
def output (S : Source) (k : Nat) (T : Table s d) : Instance (2 ^ s) :=
outputInstanceWithEnumeration S k (tableSkeleton T)
(tableEnumeration T Prod.fst (fun _ _ => rfl))
/-- Every correcting-map witness generates exactly the same numbered instance,
including the ordered constraint list and all permutation entries. -/
theorem output_eq_with_witness (S : Source) (k : Nat) (T : Table s d)
(f : Ambient s d → Alphabet s) (hf : ∀ x c, f (x + (c, 0)) = f x + c) :
output S k T =
outputInstanceWithEnumeration S k (T.gadget f hf) (tableEnumeration T f hf) := by
rfl
theorem completeAt_output_iff (error : RationalError) (S : Source) (k : Nat)
(T : Table s d) (f : Ambient s d → Alphabet s)
(hf : ∀ x c, f (x + (c, 0)) = f x + c) :
CompleteAt error (output S k T) ↔
CompleteAt error (outputInstance S k (T.gadget f hf)) := by
rw [output_eq_with_witness S k T f hf]
exact completeAt_outputInstanceWithEnumeration_iff _ _ _ _ _
theorem soundAt_output_iff (error : RationalError) (S : Source) (k : Nat)
(T : Table s d) (f : Ambient s d → Alphabet s)
(hf : ∀ x c, f (x + (c, 0)) = f x + c) :
SoundAt error (output S k T) ↔
SoundAt error (outputInstance S k (T.gadget f hf)) := by
rw [output_eq_with_witness S k T f hf]
exact soundAt_outputInstanceWithEnumeration_iff _ _ _ _ _
theorem edge_ofEnumeration (S : Source) (k : Nat) (g : SplitGadget s d)
(en : NoiseEnumeration g)
(ω : Outcome S k ((Table.ofEnumeration g en).gadget g.f g.equivariant)) :
edge S k ((Table.ofEnumeration g en).gadget g.f g.equivariant) ω =
edge S k g (ω.1, Table.positionEquiv g en ω.2.1, ω.2.2) := by
simp only [edge, leftQuery, rightQuery]
erw [Table.noise_positionEquiv g en ω.2.1]
theorem acceptance_ofEnumeration (S : Source) (k : Nat) (g : SplitGadget s d)
(en : NoiseEnumeration g) (label : Fin (vertexCount S k s d) → Fin (2 ^ s)) :
acceptanceProbability S k ((Table.ofEnumeration g en).gadget g.f g.equivariant) label =
acceptanceProbability S k g label := by
apply Fintype.expect_equiv ((Equiv.refl (Query S k s d)).prodCongr
((Table.positionEquiv g en).prodCongr (Equiv.refl (Dual k))))
intro ω
exact congrArg
(fun c : Constraint (vertexCount S k s d) (2 ^ s) =>
if c.satisfied label then (1 : ℚ) else 0)
(edge_ofEnumeration S k g en ω)
theorem rate_ofEnumeration (S : Source) (k : Nat) (g : SplitGadget s d)
(en : NoiseEnumeration g) (label : Fin (vertexCount S k s d) → Fin (2 ^ s)) :
GapSemantics.satisfactionRate
(outputInstance S k ((Table.ofEnumeration g en).gadget g.f g.equivariant)) label =
GapSemantics.satisfactionRate (outputInstance S k g) label := by
exact (acceptanceProbability_eq_count S k _ label).symm.trans
((acceptance_ofEnumeration S k g en label).trans
(acceptanceProbability_eq_count S k g label))
/-- Normalization to a concrete table preserves the original completeness
threshold exactly; there is no additional approximation loss. -/
theorem completeAt_ofEnumeration_iff (error : RationalError)
(S : Source) (k : Nat) (g : SplitGadget s d) (en : NoiseEnumeration g) :
CompleteAt error (output S k (Table.ofEnumeration g en)) ↔
CompleteAt error (outputInstance S k g) := by
rw [completeAt_output_iff error S k _ g.f g.equivariant]
rw [GapSemantics.completeAt_iff, GapSemantics.completeAt_iff]
constructor
· rintro ⟨label, h⟩
exact ⟨label, h.trans_eq (rate_ofEnumeration S k g en label)⟩
· rintro ⟨label, h⟩
exact ⟨label, h.trans_eq (rate_ofEnumeration S k g en label).symm⟩
theorem soundAt_ofEnumeration_iff (error : RationalError)
(S : Source) (k : Nat) (g : SplitGadget s d) (en : NoiseEnumeration g) :
SoundAt error (output S k (Table.ofEnumeration g en)) ↔
SoundAt error (outputInstance S k g) := by
rw [soundAt_output_iff error S k _ g.f g.equivariant]
rw [GapSemantics.soundAt_iff, GapSemantics.soundAt_iff]
constructor
· intro h label
exact (rate_ofEnumeration S k g en label).symm.trans_le (h label)
· intro h label
exact (rate_ofEnumeration S k g en label).trans_le (h label)
end MaxCutGames.Integration.TableReduction
/-! Injective vertex addresses may leave unused output vertices. The actual
instance is obtained by mapping endpoints in the original constraint list;
constraint order, multiplicity, and permutation tables are retained exactly. -/
namespace MaxCutGames.Integration.VertexEmbedding
open MaxCutGames.Foundations.Target
open GapSemantics TranslationTarget
def embedConstraint {n m q : Nat} (f : Fin n → Fin m) (c : Constraint n q) :
Constraint m q where
source := f c.source
target := f c.target
permutation := c.permutation
@[simp] theorem embedConstraint_satisfied {n m q : Nat} (f : Fin n → Fin m)
(c : Constraint n q) (labeling : Fin m → Fin q) :
(embedConstraint f c).satisfied labeling = c.satisfied (labeling ∘ f) := rfl
@[simp] theorem embedConstraint_permutation {n m q : Nat} (f : Fin n → Fin m)
(c : Constraint n q) : (embedConstraint f c).permutation = c.permutation := rfl
theorem count_map {n m q : Nat} (f : Fin n → Fin m)
(constraints : List (Constraint n q)) (labeling : Fin m → Fin q) :
countSatisfied labeling (constraints.map (embedConstraint f)) =
countSatisfied (labeling ∘ f) constraints := by
induction constraints with
| nil => rfl
| cons c cs ih =>
by_cases hc : c.satisfied (labeling ∘ f) = true <;> simp [countSatisfied, ih, hc]
def embedInstance {q m : Nat} (g : Instance q) (f : Fin g.vertices → Fin m)
(_hf : Function.Injective f) : Instance q where
vertices := m
constraints := g.constraints.map (embedConstraint f)
nonempty h := g.nonempty (List.map_eq_nil_iff.mp h)
@[simp] theorem embedInstance_vertices {q m : Nat} (g : Instance q)
(f : Fin g.vertices → Fin m) (hf : Function.Injective f) :
(embedInstance g f hf).vertices = m := rfl
theorem embedInstance_constraints {q m : Nat} (g : Instance q)
(f : Fin g.vertices → Fin m) (hf : Function.Injective f) :
(embedInstance g f hf).constraints = g.constraints.map (embedConstraint f) := rfl
@[simp] theorem embedInstance_length {q m : Nat} (g : Instance q)
(f : Fin g.vertices → Fin m) (hf : Function.Injective f) :
(embedInstance g f hf).constraints.length = g.constraints.length := List.length_map _
theorem count_embed {q m : Nat} (g : Instance q)
(f : Fin g.vertices → Fin m) (hf : Function.Injective f)
(labeling : Fin m → Fin q) :
countSatisfied labeling (embedInstance g f hf).constraints =
countSatisfied (labeling ∘ f) g.constraints := count_map f g.constraints labeling
theorem satisfactionRate_embed {q m : Nat} (g : Instance q)
(f : Fin g.vertices → Fin m) (hf : Function.Injective f)
(labeling : Fin m → Fin q) :
satisfactionRate (embedInstance g f hf) labeling =
satisfactionRate g (labeling ∘ f) := by
unfold satisfactionRate
simp only [embedInstance, count_map, List.length_map]
/-- The extension is only used as a semantic witness, not by the instance constructor. -/
noncomputable def extendLabeling {n m q : Nat} (f : Fin n → Fin m)
(labeling : Fin n → Fin q) (defaultLabel : Fin q) (v : Fin m) : Fin q := by
classical
exact if h : ∃ u, f u = v then labeling (Classical.choose h) else defaultLabel
@[simp] theorem extendLabeling_apply {n m q : Nat} (f : Fin n → Fin m)
(hf : Function.Injective f) (labeling : Fin n → Fin q)
(defaultLabel : Fin q) (v : Fin n) :
extendLabeling f labeling defaultLabel (f v) = labeling v := by
classical
have hx : ∃ u, f u = f v := ⟨v, rfl⟩
rw [extendLabeling, dite_eq_left hx]
exact congrArg labeling (hf (Classical.choose_spec hx))
theorem extendLabeling_comp {n m q : Nat} (f : Fin n → Fin m)
(hf : Function.Injective f) (labeling : Fin n → Fin q) (defaultLabel : Fin q) :
extendLabeling f labeling defaultLabel ∘ f = labeling := by
funext v
exact extendLabeling_apply f hf labeling defaultLabel v
theorem soundAt_embed {q m : Nat} (error : RationalError) (g : Instance q)
(f : Fin g.vertices → Fin m) (hf : Function.Injective f)
(hg : SoundAt error g) : SoundAt error (embedInstance g f hf) := by
intro labeling
change (Fin m → Fin q) at labeling
simp only [embedInstance, count_map, List.length_map]
exact hg (labeling ∘ f)
theorem completeAt_embed {q m : Nat} (error : RationalError) (g : Instance q)
(f : Fin g.vertices → Fin m) (hf : Function.Injective f) (hq : 0 < q)
(hg : CompleteAt error g) : CompleteAt error (embedInstance g f hf) := by
obtain ⟨labeling, hl⟩ := hg
refine ⟨extendLabeling f labeling ⟨0, hq⟩, ?_⟩
simp only [embedInstance, count_map, List.length_map, extendLabeling_comp f hf]
exact hl
theorem completeAt_embed_iff {q m : Nat} (error : RationalError) (g : Instance q)
(f : Fin g.vertices → Fin m) (hf : Function.Injective f) (hq : 0 < q) :
CompleteAt error (embedInstance g f hf) ↔ CompleteAt error g := by
constructor
· rintro ⟨labeling, hl⟩
change (Fin m → Fin q) at labeling
refine ⟨labeling ∘ f, ?_⟩
simpa only [embedInstance, count_map, List.length_map] using hl
· exact completeAt_embed error g f hf hq
theorem soundAt_embed_iff {q m : Nat} (error : RationalError) (g : Instance q)
(f : Fin g.vertices → Fin m) (hf : Function.Injective f) (hq : 0 < q) :
SoundAt error (embedInstance g f hf) ↔ SoundAt error g := by
constructor
· intro hg labeling
have hl := hg (extendLabeling f labeling ⟨0, hq⟩)
simpa only [embedInstance, count_map, List.length_map, extendLabeling_comp f hf] using hl
· exact soundAt_embed error g f hf
theorem isTranslationInstance_embed_iff {q s m : Nat}
(coordinates : Fin q ≃ BinaryLinear.Vector s) (g : Instance q)
(f : Fin g.vertices → Fin m) (hf : Function.Injective f) :
IsTranslationInstance coordinates (embedInstance g f hf) ↔
IsTranslationInstance coordinates g := by
constructor
· intro hg c hc
exact hg (embedConstraint f c) (List.mem_map.mpr ⟨c, hc, rfl⟩)
· intro hg c hc
obtain ⟨d, hd, rfl⟩ := List.mem_map.mp hc
exact hg d hd
end MaxCutGames.Integration.VertexEmbedding
end OAI