OAI/Probability/TraceReconstruction/Propagation.lean
64 declarations, 0 contain sorry, 0 declared as axioms
Propagation
Moment restriction, scale margins, sharp budget closure, clipping and error propagation, and dense/warmup overlap.
- theorem
OAI.TraceReconstruction.FiniteLaw.evenMomentSq_mono_orderline 25 - def
OAI.TraceReconstruction.OracleTracePair.restrictline 48 - def
OAI.TraceReconstruction.ConcreteLevelPair.oracleline 56Source documentation
Null-mass outcomes are assigned zero score. This proves, rather than assumes, the pointwise positivity needed by the finite simulation API.
def ConcreteLevelPair.oracle {q N Δ d ω F : ℝ} {hq : q∈Set.Icc 0 1} {i k : ℕ} (C : ConcreteLevelPair q hq N Δ d ω F i k) : OracleTracePair q C.left C.right k C.error (N^(-d)) := by …Mentions: OAI.TraceReconstruction.ConcreteLevelPair, OAI.TraceReconstruction.OracleTracePair
- theorem
OAI.TraceReconstruction.sourceScaleProduct_step_leline 101 - theorem
OAI.TraceReconstruction.sourceHalfMoment_stepline 117 - def
OAI.TraceReconstruction.sourceGapFloorline 138 - theorem
OAI.TraceReconstruction.sourceGapFloor_posline 140 - theorem
OAI.TraceReconstruction.sourceGapFloor_le_oneline 142 - theorem
OAI.TraceReconstruction.sourceEpsilon_gap_smallline 148 - theorem
OAI.TraceReconstruction.sourceHalfMoment_strongline 162 - theorem
OAI.TraceReconstruction.source_clipping_power_deltaline 170 - def
OAI.TraceReconstruction.sourceWorkSizeline 212 - theorem
OAI.TraceReconstruction.sourceWorkSize_posline 215 - theorem
OAI.TraceReconstruction.sourceCancellationOrder_lowerline 222 - theorem
OAI.TraceReconstruction.source_smoothing_absorbline 226 - theorem
OAI.TraceReconstruction.sparse_power_absorbline 250Source documentation
The actual sparse count controls both extra forced roles and clipping.
theorem sparse_power_absorb {N ε gap W L x : ℝ} (hN : 1≤ N) (he : 0≤ε) (hg : 2*ε≤ gap) (hW : 0≤ W) (hL : 1≤ L) (hx : x≤ W*L*N^(-gap)) (hfac : 8*W*L^2≤ N^ε) : x≤ N^(-gap/2)/8 ∧ x≤1/2 ∧ x*L≤1 := by … - theorem
OAI.TraceReconstruction.source_weight_squareline 276 - theorem
OAI.TraceReconstruction.source_high_tail_marginline 289 - theorem
OAI.TraceReconstruction.rpow_coefficient_absorbline 321 - theorem
OAI.TraceReconstruction.sharp_energy_absorbline 331Source documentation
Closing the three explicit sharp classes, with the source's fixed finite degree cutoff and no hidden dependence on N in their constants.
theorem sharp_energy_absorb {N ε d M Fdec gap Cr Cf E : ℝ} {R : ℕ} (hN : 1≤ N) (he : ε≤1) (hM : d+(R:ℝ)+1≤ M) (hd : 0≤ d) (_hR : 0≤(R:ℝ)) (hbad : 384*Cr≤ N^ε) (hforce : 192*Cf≤ N^ε) (htail : 1536≤ N^ε) (hfdec : d+ε≤ Fdec) (hgap : d+ε≤ gap*(R:ℝ)) (hE : E ≤ Cr*N^R*((N^(-20*M))^2+(N^(-20*M))^2)+Cf*N^(-Fdec)+8*N^(-gap*(R:ℝ))) : 64*E≤ N^(-d) := by … - theorem
OAI.TraceReconstruction.sourceDimension_boundsline 364 - theorem
OAI.TraceReconstruction.totalScoreExponent_monoline 369 - theorem
OAI.TraceReconstruction.sourceTargetScale_boundsline 375 - theorem
OAI.TraceReconstruction.ConcreteLevelPair.error_leline 382 - theorem
OAI.TraceReconstruction.source_warmup_score_marginline 389 - theorem
OAI.TraceReconstruction.source_main_score_marginline 404 - theorem
OAI.TraceReconstruction.sourceSparseCount_ge_oneline 438 - theorem
OAI.TraceReconstruction.source_block_energy_closeline 446Source documentation
Scalar closure of the exact sparse/forced block budget. All assumptions here are explicit elementary coefficient bounds, not oracle conclusions.
theorem source_block_energy_close (T n : ℕ→ℕ) (s b R : ℕ) (ω N ε d dnew K W M gap : ℝ) (hN : 1 ≤ N) (he : 0 ≤ ε) (he1 : ε ≤ 1) (hK : 0 ≤ K) (hW : 0 ≤ W) (hd : 0 ≤ dnew) (hM : dnew+(R:ℝ)+1 ≤ M) (hg : 2*ε ≤ gap) (hf : dnew+ε ≤ forcedDecay (b+1) ε d) (hR : dnew+ε ≤ gap/2*(R:ℝ)) (hfac : 8*W*((2^s:ℕ):ℝ)^2 ≤ N^ε) (hbad : 384*sourceBadRoleFactor b R ≤ N^ε) (hforce : 192*(sourceForcedRoleFactor b R*W^(2^(b+1))*sourceSkeletonPrefactor s b K) ≤ N^ε) (htail : 1536 ≤ N^ε) (hleaf : (SourceLeafCount T (b+1):ℝ) ≤ N) (hcount : 1 ≤ sourceSparseCount T s b) (k : ℕ) (L : ℝ) (hw : ((6*(2*k:ℕ):ℝ)*L)^2 ≤ W*N^(-scaleA (b+1) 0 ε*d)) …Mentions: OAI.TraceReconstruction.forcedDecay, OAI.TraceReconstruction.sourceBadRoleFactor, OAI.TraceReconstruction.sourceForcedRoleFactor, OAI.TraceReconstruction.sourceSkeletonPrefactor, OAI.TraceReconstruction.SourceLeafCount, OAI.TraceReconstruction.sourceSparseCount, OAI.TraceReconstruction.scaleA
- theorem
OAI.TraceReconstruction.source_level_energyline 496Source documentation
Instantiation at the manuscript's literal block dimension, integer schedules, source cutoff, moment envelope, and full next score exponent.
theorem source_level_energy (B D i ell h k : ℕ) (C N F ω K : ℝ) (hB : 2 ≤ B) (hi : i<D) (hN : 1 ≤ N) (hC : 1 ≤ C) (hF : 1 ≤ F) (hω : 0<ω) (H : SourceArithmeticBounds C N F (sourceEpsilon B D) (sourceCancellationOrder B D) (sourceDimension B i) ell h) (hk : (k:ℝ) ≤ sourceMomentEnvelope B D) (hK : 200*(C^(2*sourceDimension B i)*F)/ω ≤ K) (hfac : 8*sourceWeightEnvelope B D*((2^(sourceCancellationOrder B D):ℕ):ℝ)^2 ≤ N^(sourceEpsilon B D)) (hbad : 384*sourceBadRoleFactor (sourceDimension B i-1) (sourceDegreeCutoff B) ≤ N^(sourceEpsilon B D)) (hforce : 192*(sourceForcedRoleFactor (sourceDimension B i-1) (sourceDegreeCutoff B)* (sourceWeightEnvelope B D)^(2^(sourceDimension B i))* sourceSkeletonPrefactor (sourceCancellationOrder B D) (sourceDimension B i-1) K) ≤ N^(sourceEpsilon B D)) (htail : 1536 ≤ N^(sourceEpsilon B D)) …Mentions: OAI.TraceReconstruction.SourceArithmeticBounds, OAI.TraceReconstruction.sourceEpsilon, OAI.TraceReconstruction.sourceCancellationOrder, OAI.TraceReconstruction.sourceDimension, OAI.TraceReconstruction.sourceMomentEnvelope, OAI.TraceReconstruction.sourceWeightEnvelope, OAI.TraceReconstruction.sourceBadRoleFactor, OAI.TraceReconstruction.sourceDegreeCutoff, OAI.TraceReconstruction.sourceForcedRoleFactor, OAI.TraceReconstruction.sourceSkeletonPrefactor
- structure
OAI.TraceReconstruction.SourceNumericMarginsline 608 - theorem
OAI.TraceReconstruction.sourceWeightEnvelope_nonnegline 621 - theorem
OAI.TraceReconstruction.sourceForcedRoleFactor_nonnegline 625 - theorem
OAI.TraceReconstruction.source_forced_prefactor_expUpperline 629 - theorem
OAI.TraceReconstruction.source_overlap_prefactor_expUpperline 644 - theorem
OAI.TraceReconstruction.source_numeric_margins_eventuallyline 657 - theorem
OAI.TraceReconstruction.source_count_envelopeline 710 - theorem
OAI.TraceReconstruction.source_cap_envelopeline 719 - theorem
OAI.TraceReconstruction.source_local_envelopeline 728 - theorem
OAI.TraceReconstruction.SourceArithmeticBounds.count_boundline 740 - theorem
OAI.TraceReconstruction.sourceErrorUnit_smallline 755 - def
OAI.TraceReconstruction.sourceTailDenominatorline 780 - def
OAI.TraceReconstruction.sourceTailExponentline 783 - theorem
OAI.TraceReconstruction.sourceTailDenominator_posline 786 - theorem
OAI.TraceReconstruction.sourceTailDenominator_expUpperline 791 - theorem
OAI.TraceReconstruction.sourceWorkSize_polyline 798 - theorem
OAI.TraceReconstruction.sourceTailExponent_expUpperline 802 - theorem
OAI.TraceReconstruction.source_tails_eventuallyline 811 - theorem
OAI.TraceReconstruction.source_exponential_error_budgetline 830Source documentation
Turning the common uniform exponential bound into the two exact scheduled-error terms, including the division by the clipping threshold.
theorem source_exponential_error_budget {B D b s ell k : ℕ} {C N F ω f n : ℝ} (H : SourceArithmeticBounds C N F (sourceEpsilon B D) s b ell k) (hB : 2 ≤ B) (hb : b ≤ B) (hC : 0<C) (hF : 0<F) (hω : 0<ω) (hn : 1 ≤ n) (hN : 1 ≤ N) (hNn : N ≤ n) (hf512 : 512 ≤ f) (hfmeta : 9600*(C^(2*b)*F)/ω^4 ≤ f) (htail : 2*(B:ℝ)*Real.exp (-(N^(sourceEpsilon B D)/f)) ≤ n^(-sourceTailExponent B D)) : sourceEffectiveHiddenTail (sourceBlockWindows N b (sourceEpsilon B D)) (sourceArithmeticTrials C N b (sourceEpsilon B D) ell k) b ω ≤ n^(-200*(totalScoreExponent B D+B+1)) ∧ sourceFitTail (sourceBlockWindows N b (sourceEpsilon B D)) (sourceArithmeticTrials C N b (sourceEpsilon B D) ell k) b / N^(-20*sourceWorkSize B D) ≤ n^(-200*(totalScoreExponent B D+B+1)) ∧ sourceFitTail (sourceBlockWindows N b (sourceEpsilon B D)) …Mentions: OAI.TraceReconstruction.SourceArithmeticBounds, OAI.TraceReconstruction.sourceEpsilon, OAI.TraceReconstruction.sourceTailExponent, OAI.TraceReconstruction.sourceEffectiveHiddenTail, OAI.TraceReconstruction.sourceBlockWindows, OAI.TraceReconstruction.sourceArithmeticTrials, OAI.TraceReconstruction.totalScoreExponent, OAI.TraceReconstruction.sourceFitTail, OAI.TraceReconstruction.sourceWorkSize
- theorem
OAI.TraceReconstruction.rpow_constant_absorbline 899 - theorem
OAI.TraceReconstruction.overlap_sqrt_budgetline 909 - theorem
OAI.TraceReconstruction.source_dense_variance_budgetline 916Source documentation
Generic dense-template estimate, used for the first warmup step and every step of a post-warmup block.
theorem source_dense_variance_budget {N ε e a K L V h : ℝ} {j : ℕ} (hN : 1 ≤ N) (_hK : 0 ≤ K) (hL : 0 ≤ L) (hh : 0 ≤ h) (hcount : h ≤ (K+1)*N^a) (hV : V ≤ N^(-e)) (hc : 4096*(K+1)*L^j ≤ N^ε) (he : ε+a ≤ e) : h*(L^j*V) ≤ 1/4096 := by … - theorem
OAI.TraceReconstruction.source_warmup_variance_budgetline 934Source documentation
Exact scalar one-window refinement. Its quadratic sparse term is what makes the second warmup overlap possible at old score exponent 11/5.
theorem source_warmup_variance_budget {N ε e M K W L w T h E : ℝ} (hN : 1 ≤ N) (hK : 0 ≤ K) (hW : 0 ≤ W) (hL : 0 ≤ L) (hT : 0 ≤ T) (hh : 0 ≤ h) (hcount : h ≤ (K+1)*N) (hwindow : T ≤ N) (hw : w^2 ≤ W*N^(-e)) (henergy : E ≤ 64*(72*T*w^2*(N^(-20*M))^2+8*(w^2*L)^2)) (hbad : 8192*(4608*(K+1)*L*W) ≤ N^ε) (hquad : 8192*(512*(K+1)*W^2*L^3) ≤ N^ε) (hebad : ε+2-e-40*M ≤ 0) (hequad : ε+1-2*e ≤ 0) : h*(L*E) ≤ 1/4096 := by … - theorem
OAI.TraceReconstruction.sourceBlockLength_factorline 981 - theorem
OAI.TraceReconstruction.source_delta_power_budgetline 994 - theorem
OAI.TraceReconstruction.source_repeated_error_budgetline 1009Source documentation
Uniform repeated old simulation error; this is the error term in every dense nested-overlap estimate.
theorem source_repeated_error_budget {n N ε K E t h : ℝ} {B D : ℕ} (hB : 2 ≤ B) (hn : 1 ≤ n) (hN : 1 ≤ N) (hNn : N ≤ n) (he : 0 ≤ ε) (he1 : ε ≤ 1) (hh : 0 ≤ h) (_hK : 0 ≤ K) (hcount : h ≤ (K+1)*N) (hcost : t*E ≤ ((D:ℝ)+1)*n^(-100*(totalScoreExponent B D+B+1))*N) (hmargin : 64*(K+1)*((D:ℝ)+1) ≤ N^ε) : 2*h*(t*E) ≤ 1/32 := by … - theorem
OAI.TraceReconstruction.source_old_error_costline 1035 - theorem
OAI.TraceReconstruction.normalized_block_errorline 1056Source documentation
The normalized error invariant propagates through actual leaf products, with one fresh delta paid by clipping and fitting.
theorem normalized_block_error {E Δ clip fit : ℝ} {ell newell leaves i : ℕ} (hell : 1 ≤ newell) (hE : E ≤ ((i:ℝ)+1)*Δ*ell) (hΔ : 0 ≤ Δ) (hleaf : leaves*ell ≤ newell) (hnew : clip+fit ≤ Δ) : (clip+(leaves:ℝ)*E+fit)/(newell:ℝ) ≤ ((i:ℝ)+2)*Δ := by … - theorem
OAI.TraceReconstruction.source_level_clippingline 1075Source documentation
All three genuine clipping losses, and fitting, share the same global error unit. The reference power is the actual half-moment at this scale.
theorem source_level_clipping (B D j s b : ℕ) (T t : ℕ→ℕ) (n ω a gap L : ℝ) (hB : 2 ≤ B) (hn : 4 ≤ n) (hg : sourceGapFloor B ≤ gap) (hmeta : sourceEffectiveHiddenTail T t (b+1) ω ≤ n^(-200*(totalScoreExponent B D+B+1))) (hfit : sourceFitTail T t (b+1)/a ≤ n^(-200*(totalScoreExponent B D+B+1))) (hfit' : sourceFitTail T t (b+1) ≤ n^(-200*(totalScoreExponent B D+B+1))) (hhalf : ((6*(2*sourceHalfMoment B D n j:ℕ):ℝ)*L)^2*sourceSparseCount T s b ≤ 1/2) (hscale : ((6*(2*sourceHalfMoment B D n j:ℕ):ℝ)*L)^2*sourceSparseCount T s b ≤ (sourceTargetScale B D n j)^(-gap/2)/8) : sourceClippingCost T t s b (sourceHalfMoment B D n j) ω a L ≤ 1/2 ∧ sourceClippingCost T t s b (sourceHalfMoment B D n j) ω a L+sourceFitTail T t (b+1) ≤ n^(-100*(totalScoreExponent B D+B+1)) := by …Mentions: OAI.TraceReconstruction.sourceGapFloor, OAI.TraceReconstruction.sourceEffectiveHiddenTail, OAI.TraceReconstruction.totalScoreExponent, OAI.TraceReconstruction.sourceFitTail, OAI.TraceReconstruction.sourceHalfMoment, OAI.TraceReconstruction.sourceSparseCount, OAI.TraceReconstruction.sourceTargetScale, OAI.TraceReconstruction.sourceClippingCost
- theorem
OAI.TraceReconstruction.source_one_window_clippingline 1096Source documentation
At one window the root metadata can still fail; only the fitting loss vanishes. We charge the root metadata rather than discard it.
theorem source_one_window_clipping (B D j s : ℕ) (T t : ℕ) (n ω a gap L : ℝ) (hB : 2 ≤ B) (hn : 4 ≤ n) (hg : sourceGapFloor B ≤ gap) (hmeta : sourceEffectiveHiddenTail (fun _=>T) (fun _=>t) 1 ω ≤ n^(-200*(totalScoreExponent B D+B+1))) (hhalf : ((6*(2*sourceHalfMoment B D n j:ℕ):ℝ)*L)^2*(2:ℝ)^s ≤ 1/2) (hscale : ((6*(2*sourceHalfMoment B D n j:ℕ):ℝ)*L)^2*(2:ℝ)^s ≤ (sourceTargetScale B D n j)^(-gap/2)/8) : sourceClippingCost (fun _=>T) (fun _=>t) s 0 (sourceHalfMoment B D n j) ω a L ≤ 1/2 ∧ sourceClippingCost (fun _=>T) (fun _=>t) s 0 (sourceHalfMoment B D n j) ω a L ≤ n^(-100*(totalScoreExponent B D+B+1)) := by …Mentions: OAI.TraceReconstruction.sourceGapFloor, OAI.TraceReconstruction.sourceEffectiveHiddenTail, OAI.TraceReconstruction.totalScoreExponent, OAI.TraceReconstruction.sourceHalfMoment, OAI.TraceReconstruction.sourceTargetScale, OAI.TraceReconstruction.sourceClippingCost
- theorem
OAI.TraceReconstruction.source_overlap_prefactorsline 1118 - theorem
OAI.TraceReconstruction.source_literal_dense_overlapline 1149Source documentation
Complete dense-overlap numerical criterion for the literal schedules, including the error carried by the old exact simulator.
theorem source_literal_dense_overlap {C N F ε E Δ V K W L n e : ℝ} {s b ell k j i D B : ℕ} (H : SourceArithmeticBounds C N F ε s b ell k) (hB : 2 ≤ B) (hb : b ≤ B) (hC : 1 ≤ C) (hF : 0 ≤ F) (hn : 1 ≤ n) (hN : 1 ≤ N) (hNn : N ≤ n) (he : 0 ≤ ε) (he1 : ε ≤ 1) (hK : 0 ≤ K) (hW : 0 ≤ W) (hL : 1 ≤ L) (hLdef : L=(2:ℝ)^s) (hG : C^(2*b)*F ≤ K) (hi : i ≤ D) (hj : j ≤ b) (ha : 0 ≤ scaleA b j ε) (ha1 : scaleA b j ε ≤ 1) (hE : E ≤ ((i:ℝ)+1)*Δ*ell) (hΔ : Δ=n^(-100*(totalScoreExponent B D+B+1))) (hV : V ≤ N^(-e)) (hgap : ε+scaleA b j ε ≤ e) (hmargin : 64*(K+1)*((D:ℝ)+1) ≤ N^ε) (hoverlap : 1099511627776*(K+1)*(W+1)^2*L^(B+2) ≤ N^ε) : Real.sqrt (((sourceArithmeticCount C N b ε ell k j+1:ℕ):ℝ)* …Mentions: OAI.TraceReconstruction.SourceArithmeticBounds, OAI.TraceReconstruction.scaleA, OAI.TraceReconstruction.totalScoreExponent, OAI.TraceReconstruction.sourceArithmeticCount
- theorem
OAI.TraceReconstruction.SourceArithmeticBounds.length_positiveline 1187 - theorem
OAI.TraceReconstruction.SourceArithmeticBounds.copies_le_targetline 1195 - theorem
OAI.TraceReconstruction.source_warmup_error_budgetline 1209Source documentation
Both contributions to the error in the second warmup overlap are paid by the common delta margin. The old term uses T₀ ≤ P₀, not an independent full-block overlap hypothesis.
theorem source_warmup_error_budget {C N F ε E Δ clip K n : ℝ} {s ell k i D B : ℕ} (H : SourceArithmeticBounds C N F ε s 2 ell k) (hB : 2 ≤ B) (hC : 1 ≤ C) (hF : 0 ≤ F) (hn : 1 ≤ n) (hN : 1 ≤ N) (hNn : N ≤ n) (he : 0 ≤ ε) (he1 : ε ≤ 1) (hK : 0 ≤ K) (hG : C^4*F ≤ K) (hi : i ≤ D) (hE : E ≤ ((i:ℝ)+1)*Δ*ell) (hE0 : 0 ≤ E) (hΔ : Δ=n^(-100*(totalScoreExponent B D+B+1))) (hclip : clip ≤ Δ) (hmargin : 64*(K+1)*((D:ℝ)+1) ≤ N^ε) : 2*((sourceArithmeticCount C N 2 ε ell k 2+1:ℕ):ℝ)* ((sourceArithmeticP C N 2 ε ell k 1:ℝ)* (clip+(sourceBlockWindows N 2 ε 0:ℝ)*E)) ≤ 1/8 := by …Mentions: OAI.TraceReconstruction.SourceArithmeticBounds, OAI.TraceReconstruction.totalScoreExponent, OAI.TraceReconstruction.sourceArithmeticCount, OAI.TraceReconstruction.sourceArithmeticP, OAI.TraceReconstruction.sourceBlockWindows