OAI/Probability/TraceReconstruction/LayoutRates.lean

GitHub source file

75 declarations, 0 contain sorry, 0 declared as axioms

LayoutRates

Integer block schedules, rounding, length and trial margins, sparse gaps, and forced-anchor decay.

  1. def OAI.TraceReconstruction.sourceWindowSize line 23

    Source documentation

    The source integer window size, with zero-based nesting index.

    def sourceWindowSize (N : ℝ) (b : ℕ) (ε : ℝ) (j : ℕ) : ℕ :=
    …
  2. def OAI.TraceReconstruction.sourceLocalTrials line 27

    Source documentation

    At a nonroot node the source trial count is an even integer.

    def sourceLocalTrials (P h : ℕ) : ℕ := 2*(P/(100*(2*h+1)))
  3. def OAI.TraceReconstruction.sourceRootTrials line 30

    Source documentation

    The root schedule count: the literal floor after subtracting its window.

    def sourceRootTrials (N : ℝ) (S T h : ℕ) : ℕ :=
    …
  4. structure OAI.TraceReconstruction.SourceArithmeticStage line 35

    Source documentation

    Literal length and overlap-count state of one nested source block. copies and trials refer to the transition into this state.

    structure SourceArithmeticStage where
    …
  5. def OAI.TraceReconstruction.sourceArithmeticStages line 41
  6. def OAI.TraceReconstruction.sourceArithmeticP line 53
  7. def OAI.TraceReconstruction.sourceArithmeticTrials line 56
  8. def OAI.TraceReconstruction.sourceArithmeticLength line 59
  9. def OAI.TraceReconstruction.sourceArithmeticCount line 62
  10. theorem OAI.TraceReconstruction.sourceLocalTrials_even line 65
  11. theorem OAI.TraceReconstruction.sourceLocalTrials_budget line 68
  12. theorem OAI.TraceReconstruction.sourceLocalTrials_reserves line 74
  13. theorem OAI.TraceReconstruction.sourceWindowSize_even line 92
  14. theorem OAI.TraceReconstruction.sourceArithmeticLength_step line 97
  15. theorem OAI.TraceReconstruction.sourceArithmeticCount_step line 101
  16. theorem OAI.TraceReconstruction.sourceArithmetic_inner_trials line 105
  17. theorem OAI.TraceReconstruction.sourceArithmetic_root_copies line 110
  18. theorem OAI.TraceReconstruction.sourceArithmetic_count_lift line 117
  19. theorem OAI.TraceReconstruction.sourceWindowSize_bounds line 129

    Source documentation

    Rounding the source windows loses at most a factor two.

    theorem sourceWindowSize_bounds {N ε : ℝ} {b j : ℕ}
        (h : 8 ≤ N^(scaleKappa b (j+1) ε)) :
        N^(scaleKappa b (j+1) ε)/2 ≤ (sourceWindowSize N b ε j:ℝ) ∧
          (sourceWindowSize N b ε j:ℝ) ≤ N^(scaleKappa b (j+1) ε) := by
    …

    Mentions: OAI.TraceReconstruction.scaleKappa, OAI.TraceReconstruction.sourceWindowSize

  20. theorem OAI.TraceReconstruction.sourceLocalTrials_bounds line 141

    Source documentation

    The nonroot even trial count is between one and two times its unrounded quotient once the quotient is at least two.

    theorem sourceLocalTrials_bounds (P h : ℕ) (hlarge : (200:ℝ)*(2*h+1) ≤ P) :
        (P:ℝ)/(100*(2*h+1)) ≤ (sourceLocalTrials P h:ℝ) ∧
          (sourceLocalTrials P h:ℝ) ≤ (P:ℝ)/(50*(2*h+1)) := by
    …

    Mentions: OAI.TraceReconstruction.sourceLocalTrials

  21. theorem OAI.TraceReconstruction.source_floor_length_bounds line 159
  22. theorem OAI.TraceReconstruction.sourceRootTrials_floor line 173

    Source documentation

    The real floor and integer division versions of the root are identical whenever its remaining ambient count is nonnegative.

    theorem sourceRootTrials_floor {N : ℝ} {S T h : ℕ}
        (hT : T ≤ ⌊N/(S:ℝ)⌋₊) :
        sourceRootTrials N S T h=⌊((⌊N/(S:ℝ)⌋₊:ℝ)-T)/(2*h+1)⌋₊ := by
    …

    Mentions: OAI.TraceReconstruction.sourceRootTrials

  23. theorem OAI.TraceReconstruction.source_root_length_bounds line 180
  24. theorem OAI.TraceReconstruction.source_ceil_count_bounds line 211
  25. theorem OAI.TraceReconstruction.source_scale_target_step line 237
  26. theorem OAI.TraceReconstruction.sourceWindowSize_le_previous line 245
  27. theorem OAI.TraceReconstruction.source_root_margins line 261

    Source documentation

    The root's subtraction and rounding margins follow from the one subpower comparison N^epsilon >= 8F.

    theorem source_root_margins {N X u F : ℝ} {S h T : ℕ}
        (hS : 1 ≤ S) (hX : (S:ℝ) ≤ X) (hF : 1 ≤ F) (hu : 8*F ≤ u)
        (hh : (h:ℝ) ≤ F*S) (hT : (T:ℝ) ≤ X) (hrel : X^2*u=N) :
        T ≤ ⌊N/(S:ℝ)⌋₊ ∧ (2*(h:ℝ)+2)*S ≤ N/2 := by
    …
  28. theorem OAI.TraceReconstruction.source_arithmetic_inner_copies line 282
  29. theorem OAI.TraceReconstruction.source_arithmetic_root_copies line 287
  30. theorem OAI.TraceReconstruction.source_arithmetic_lengths line 303

    Source documentation

    All nested lengths and overlap counts for the literal floors/ceilings. Only the scalar subpower margin is needed for every rounding step.

    theorem source_arithmetic_lengths {C N F ε : ℝ} {b ell k : ℕ}
        (hC : 2 ≤ C) (hN : 1 ≤ N) (hF : 1 ≤ F) (he : 0 ≤ ε)
        (ha : ∀j ≤ b,0 ≤ scaleA b j ε)
        (hsmall : 8*C^(2*b)*F ≤ N^ε)
        (hell : N^(scaleA b 0 ε)/2 ≤ (ell:ℝ)) (hell' : (ell:ℝ) ≤ N^(scaleA b 0 ε))
        (hk : ell ≤ k) (hk' : (k:ℝ) ≤ F*ell) :
        ∀j ≤ b,
          N^(scaleA b j ε)/2 ≤ (sourceArithmeticLength C N b ε ell k j:ℝ) ∧
          (sourceArithmeticLength C N b ε ell k j:ℝ) ≤ N^(scaleA b j ε) ∧
          sourceArithmeticLength C N b ε ell k j ≤ sourceArithmeticCount C N b ε ell k j ∧
          (sourceArithmeticCount C N b ε ell k j:ℝ) ≤
            C^(2*j)*F*sourceArithmeticLength C N b ε ell k j := by
    …

    Mentions: OAI.TraceReconstruction.scaleA, OAI.TraceReconstruction.sourceArithmeticLength, OAI.TraceReconstruction.sourceArithmeticCount

  31. theorem OAI.TraceReconstruction.source_copies_bounds line 381
  32. theorem OAI.TraceReconstruction.source_inner_trials_scale line 391
  33. theorem OAI.TraceReconstruction.source_root_trials_scale line 420
  34. def OAI.TraceReconstruction.sourceDegreeCutoff line 452
  35. def OAI.TraceReconstruction.sourceCancellationOrder line 454
  36. def OAI.TraceReconstruction.sourceChannelLogConstant line 457
  37. theorem OAI.TraceReconstruction.sourceChannelLogConstant_pos line 459
  38. theorem OAI.TraceReconstruction.deletion_inverse_cube_ge_one line 463
  39. theorem OAI.TraceReconstruction.deletion_inverse_le_cube line 468
  40. theorem OAI.TraceReconstruction.sourceOverlapConstant_log_bound line 477
  41. theorem OAI.TraceReconstruction.sourceOverlapConstant_ge_two line 497
  42. theorem OAI.TraceReconstruction.source_count_factor_exp_bound line 506

    Source documentation

    The whole block-count loss, including every nesting and every earlier block, is bounded by a single exp(O_B((D+1)Lambda)).

    theorem source_count_factor_exp_bound {q : ℝ} (hq : q∈Set.Ioc 0 (1/2:ℝ))
        (B D b i : ℕ) (hb : b ≤ B) (hi : i ≤ D) :
        8*(sourceOverlapConstant q)^(2*b)*(sourceOverlapConstant q)^(2*B*i+2) ≤
          Real.exp ((8+sourceChannelLogConstant*(2*B+2))*((D:ℝ)+1)*q^(-3:ℤ)) := by
    …

    Mentions: OAI.TraceReconstruction.sourceOverlapConstant, OAI.TraceReconstruction.sourceChannelLogConstant

  43. theorem OAI.TraceReconstruction.source_uniform_count_margin line 539
  44. theorem OAI.TraceReconstruction.sourceEpsilon_power_small line 565

    Source documentation

    Tiny epsilon estimates are uniform in the varying block depth.

    theorem sourceEpsilon_power_small (B D m : ℕ) (hm : m ≤ 4*B+100) :
        (2:ℝ)^m*sourceEpsilon B D ≤ 1 := by
    …

    Mentions: OAI.TraceReconstruction.sourceEpsilon

  45. theorem OAI.TraceReconstruction.scaleA_mono_index line 574
  46. theorem OAI.TraceReconstruction.scaleA_le_geometric line 581
  47. theorem OAI.TraceReconstruction.sourceKappa_three_epsilon line 592
  48. theorem OAI.TraceReconstruction.sourceA_nonneg line 622
  49. def OAI.TraceReconstruction.sourceInteriorExponent line 632

    Source documentation

    The interior scale mass excludes the bottom window, as required by the sparse role count rather than the much larger total leaf count.

    def sourceInteriorExponent (b : ℕ) (ε : ℝ) : ℝ :=
    …
  50. theorem OAI.TraceReconstruction.two_power_sum_shift line 635
  51. theorem OAI.TraceReconstruction.sourceInteriorExponent_upper line 641
  52. theorem OAI.TraceReconstruction.scaleA_zero_simple_lower line 662
  53. theorem OAI.TraceReconstruction.source_sparse_gap line 670

    Source documentation

    Quantitative sparse gap, keeping its growth in d. Both the two-dimensional warmup and the post-warmup dimension satisfy this common estimate.

    theorem source_sparse_gap (B D b : ℕ) (d : ℝ) (hb1 : 1 ≤ b) (_hb : b ≤ B)
        (hd : 11/5 ≤ d) (hcases : b=2 ∨ (b=B ∧ (2:ℝ)^B+2 ≤ d)) :
        (d+1)/(128*(2:ℝ)^(2*B)) ≤
          scaleA b 0 (sourceEpsilon B D)*d-sourceInteriorExponent b (sourceEpsilon B D) := by
    …

    Mentions: OAI.TraceReconstruction.scaleA, OAI.TraceReconstruction.sourceEpsilon, OAI.TraceReconstruction.sourceInteriorExponent

  54. theorem OAI.TraceReconstruction.totalScoreExponent_step line 724
  55. theorem OAI.TraceReconstruction.totalScoreExponent_after_warmup line 730
  56. theorem OAI.TraceReconstruction.source_level_sparse_gap line 739
  57. def OAI.TraceReconstruction.sourceBlockWindows line 761

    Source documentation

    Levels outside the finite source block are irrelevant. Their harmless positive extension permits the global-index API of the finite tree lemmas.

    def sourceBlockWindows (N : ℝ) (b : ℕ) (ε : ℝ) (j : ℕ) : ℕ :=
    …
  58. structure OAI.TraceReconstruction.SourceArithmeticBounds line 764
  59. theorem OAI.TraceReconstruction.source_arithmetic_geometry line 780

    Source documentation

    Exact layout inequalities for the literal recursive integer construction. The hypotheses are just two scalar subpower margins and the exact scale identities, not any probabilistic producer or new overlap premise.

    theorem source_arithmetic_geometry {C N F ε : ℝ} {s b ell k : ℕ}
        (hC : 2 ≤ C) (hN : 1 ≤ N) (hF : 1 ≤ F) (he : 0 ≤ ε)
        (ha : ∀j ≤ b,0 ≤ scaleA b j ε) (hκ : ∀j<b,3*ε ≤ scaleKappa b (j+1) ε)
        (hsmall : 1200*C^(2*b)*F ≤ N^ε) (hmotif : 48*((2^s:ℕ):ℝ) ≤ N^ε)
        (hell : N^(scaleA b 0 ε)/2 ≤ (ell:ℝ)) (hell' : (ell:ℝ) ≤ N^(scaleA b 0 ε))
        (hk : ell ≤ k) (hk' : (k:ℝ) ≤ F*ell) :
        SourceArithmeticBounds C N F ε s b ell k := by
    …

    Mentions: OAI.TraceReconstruction.scaleA, OAI.TraceReconstruction.scaleKappa, OAI.TraceReconstruction.SourceArithmeticBounds

  60. theorem OAI.TraceReconstruction.SourceArithmeticBounds.layout line 855
  61. theorem OAI.TraceReconstruction.SourceArithmeticBounds.window_positive line 884
  62. def OAI.TraceReconstruction.binaryTreeExponent line 902

    Source documentation

    Exact exponent of the full binary forced skeleton, in the same bottom-up order as binaryTreeBudget.

    def binaryTreeExponent (e : ℕ → ℝ) : ℕ → ℝ
      | 0 => 0
      | b+1 => e b+2*binaryTreeExponent e b
  63. theorem OAI.TraceReconstruction.binaryTreeExponent_sum line 906
  64. theorem OAI.TraceReconstruction.binaryTreeBudget_mono line 922
  65. theorem OAI.TraceReconstruction.binaryTreeBudget_scaled line 937
  66. theorem OAI.TraceReconstruction.binaryTreeBudget_const line 955
  67. theorem OAI.TraceReconstruction.source_interior_product line 966
  68. theorem OAI.TraceReconstruction.sourceForcedBudget_degree_factor line 974

    Source documentation

    Pull all extra selected leaves into the sparse-degree small factor. Only the mandatory 2^(b+1) bottom leaves pay the old score exponent.

    theorem sourceForcedBudget_degree_factor (T n : ℕ → ℕ) (s b r : ℕ) (ω w : ℝ)
        (hr : 2^(b+1) ≤ r) :
        w^(2*r)*sourceForcedBudget T n s b r ω=
          (OccupiedParityTree.shapeRoleConstant r (b+1):ℝ)*
          (w^2*((2^s:ℕ):ℝ)*(sourceSparseCount T s b:ℝ))^(r-2^(b+1))*
          (w^2)^(2^(b+1))*
          ((2*((2^s:ℕ):ℝ)^3*hiddenBinomialCap (n b) ω)*
            binaryTreeBudget (fun j => 4*((2^s:ℕ):ℝ)^6*
              hiddenBinomialCap (sourceScheduledTrials T n (j+1)) ω*hiddenBinomialCap (n j) ω) b) := by
    …

    Mentions: OAI.TraceReconstruction.sourceForcedBudget, OAI.TraceReconstruction.OccupiedParityTree.shapeRoleConstant, OAI.TraceReconstruction.sourceSparseCount, OAI.TraceReconstruction.hiddenBinomialCap, OAI.TraceReconstruction.binaryTreeBudget, OAI.TraceReconstruction.sourceScheduledTrials

  69. theorem OAI.TraceReconstruction.sourceForcedBudget_degree_le line 993

    Source documentation

    The remaining extra-degree factor costs no scale power.

    theorem sourceForcedBudget_degree_le (T n : ℕ → ℕ) (s b r : ℕ) (ω w : ℝ)
        (hr : 2^(b+1) ≤ r)
        (hsmall : w^2*((2^s:ℕ):ℝ)*(sourceSparseCount T s b:ℝ) ≤ 1) :
        w^(2*r)*sourceForcedBudget T n s b r ω ≤
          (OccupiedParityTree.shapeRoleConstant r (b+1):ℝ)*(w^2)^(2^(b+1))*
          ((2*((2^s:ℕ):ℝ)^3*hiddenBinomialCap (n b) ω)*
            binaryTreeBudget (fun j => 4*((2^s:ℕ):ℝ)^6*
              hiddenBinomialCap (sourceScheduledTrials T n (j+1)) ω*hiddenBinomialCap (n j) ω) b) := by
    …

    Mentions: OAI.TraceReconstruction.sourceSparseCount, OAI.TraceReconstruction.sourceForcedBudget, OAI.TraceReconstruction.OccupiedParityTree.shapeRoleConstant, OAI.TraceReconstruction.hiddenBinomialCap, OAI.TraceReconstruction.binaryTreeBudget, OAI.TraceReconstruction.sourceScheduledTrials

  70. theorem OAI.TraceReconstruction.hiddenBinomialCap_rpow line 1021
  71. theorem OAI.TraceReconstruction.forcedDecay_binary line 1049

    Source documentation

    The exact exponent in the skeleton budget equals the manuscript's forcedDecay, not an arbitrary favorable surrogate.

    theorem forcedDecay_binary (b : ℕ) (ε d : ℝ) (hb : 1 ≤ b) :
        forcedDecay b ε d=ε/2+
          binaryTreeExponent (fun j=>scaleKappa b (j+2) ε/2+ε) (b-1)+
          (2:ℝ)^b*scaleA b 0 ε*d := by
    …

    Mentions: OAI.TraceReconstruction.forcedDecay, OAI.TraceReconstruction.binaryTreeExponent, OAI.TraceReconstruction.scaleKappa, OAI.TraceReconstruction.scaleA

  72. def OAI.TraceReconstruction.sourceSkeletonPrefactor line 1064
  73. theorem OAI.TraceReconstruction.sourceSkeletonPrefactor_nonneg line 1067
  74. theorem OAI.TraceReconstruction.source_forced_core_bound line 1073
  75. theorem OAI.TraceReconstruction.source_forced_degree_decay line 1128

    Source documentation

    The sharp forced degree has the full decay gain, after all actual source caps and the old-score factor have been multiplied.

    theorem source_forced_degree_decay (T n : ℕ → ℕ) (s b r : ℕ) (ω N ε d K W w : ℝ)
        (hN : 0 < N) (hK : 0 ≤ K) (hW : 0 ≤ W) (hr : 2^(b+1) ≤ r)
        (hsmall : w^2*((2^s:ℕ):ℝ)*(sourceSparseCount T s b:ℝ) ≤ 1)
        (hw : w^2 ≤ W*N^(-scaleA (b+1) 0 ε*d))
        (hn : ∀j ≤ b,hiddenBinomialCap (n j) ω ≤ K*N^(-ε/2))
        (hg : ∀j < b,hiddenBinomialCap (sourceScheduledTrials T n (j+1)) ω ≤
          K*N^(-(scaleKappa (b+1) (j+2) ε+ε)/2)) :
        w^(2*r)*sourceForcedBudget T n s b r ω ≤
          (OccupiedParityTree.shapeRoleConstant r (b+1):ℝ)*W^(2^(b+1))*
            sourceSkeletonPrefactor s b K*N^(-forcedDecay (b+1) ε d) := by
    …

    Mentions: OAI.TraceReconstruction.sourceSparseCount, OAI.TraceReconstruction.scaleA, OAI.TraceReconstruction.hiddenBinomialCap, OAI.TraceReconstruction.sourceScheduledTrials, OAI.TraceReconstruction.scaleKappa, OAI.TraceReconstruction.sourceForcedBudget, OAI.TraceReconstruction.OccupiedParityTree.shapeRoleConstant, OAI.TraceReconstruction.sourceSkeletonPrefactor, OAI.TraceReconstruction.forcedDecay