Lean module · Foundations
BanditRLProof.Algorithms.HeavyTailSourceCounterexample
Finite counterexample to the literal BCL13 printed regret coefficient for the unchanged radius-four policy. Deterministic arms0,-1, raw second moment<=1, horizon2^50. The corrected upper bound remains a separate theorem.
Module map
Imports
BanditRLProof.Algorithms.HeavyTailSourceRegret, BanditRLProof.PullCountDecomposition
Imported by
Declarations
Open an item to read its exact compact statement and source link. Detailed teaching notes are linked when registered.
def
BanditRLProof.HeavyTail.SourceCounterexample.dropped
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.droppedReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
noncomputable def dropped (L : ℝ) (n : ℕ) : ℕ
theorem
BanditRLProof.HeavyTail.SourceCounterexample.truncate_neg_one
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.truncate_neg_oneReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem truncate_neg_one (L : ℝ) (hL : 0 < L) (s : ℕ) : truncate (sourceTruncationThreshold 1 1 L s) (-1) = if (s : ℝ)+1 < L then 0 else -1
theorem
BanditRLProof.HeavyTail.SourceCounterexample.dropped_filter
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.dropped_filterReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem dropped_filter (L : ℝ) (n : ℕ) : (Finset.range n).filter (fun s : ℕ => (s : ℝ)+1 < L) = Finset.range (dropped L n)
theorem
BanditRLProof.HeavyTail.SourceCounterexample.sum_truncate_neg_one
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.sum_truncate_neg_oneReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem sum_truncate_neg_one (L : ℝ) (hL : 0 < L) (n : ℕ) : (∑ s ∈ Finset.range n, truncate (sourceTruncationThreshold 1 1 L s) (-1)) = -(n : ℝ) + dropped L n
theorem
BanditRLProof.HeavyTail.SourceCounterexample.dropped_ge
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.dropped_geReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem dropped_ge (L : ℝ) (hL : 0 < L) (n : ℕ) (hn : L ≤ n) : L-1 ≤ (dropped L n : ℝ)
theorem
BanditRLProof.HeavyTail.SourceCounterexample.suboptimal_index_gt
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.suboptimal_index_gtReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem suboptimal_index_gt (H L n d : ℝ) (hH : 30 < H) (hL : 2*H-2 ≤ L) (hn : 0 < n) (hnM : n ≤ 32*H+5) (hd : L-1 ≤ d) : (1/100 : ℝ) < -1+d/n+4*Real.sqrt (L/n)
theorem
BanditRLProof.HeavyTail.SourceCounterexample.optimal_index_lt
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.optimal_index_ltReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem optimal_index_lt (H L m : ℝ) (hH : 0 < H) (hL0 : 0 ≤ L) (hL : L ≤ 2*H) (hm : 320000*H < m) : 4*Real.sqrt (L/m) < (1/100 : ℝ)
def
BanditRLProof.HeavyTail.SourceCounterexample.deterministicStream
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.deterministicStreamReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
noncomputable def deterministicStream : UCB.ArmRewardStream 2
def
BanditRLProof.HeavyTail.SourceCounterexample.trace
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.traceReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
noncomputable def trace : ActionTrace (Fin 2)
theorem
BanditRLProof.HeavyTail.SourceCounterexample.mean_zero
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.mean_zeroReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem mean_zero (t : ℕ) : SourcePolicy.robustMean (by decide) 1 1 deterministicStream 0 t = 0
theorem
BanditRLProof.HeavyTail.SourceCounterexample.mean_one
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.mean_oneReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem mean_one (t : ℕ) (ht : 2 ≤ t) : SourcePolicy.robustMean (by decide) 1 1 deterministicStream 1 t = -1 + dropped (sourceConfidenceLog t) (pullCount trace 1 t) / pullCount trace 1 t
theorem
BanditRLProof.HeavyTail.SourceCounterexample.radius_sqrt
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.radius_sqrtReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem radius_sqrt (t n : ℕ) : sourceConfidenceRadius 1 1 t n = 4 * Real.sqrt (sourceConfidenceLog t/n)
theorem
BanditRLProof.HeavyTail.SourceCounterexample.source_index_sub_gt
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.source_index_sub_gtReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem source_index_sub_gt (H : ℝ) (t : ℕ) (ht : 2 ≤ t) (hH : 30 < H) (hL : 2*H-2 ≤ sourceConfidenceLog t) (hnM : (pullCount trace 1 t : ℝ) ≤ 32*H+5) : (1/100 : ℝ) < SourcePolicy.robustMean (by decide) 1 1 deterministicStream 1 t + sourceConfidenceRadius 1 1 t (pullCount trace 1 t)
theorem
BanditRLProof.HeavyTail.SourceCounterexample.source_index_best_lt
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.source_index_best_ltReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem source_index_best_lt (H : ℝ) (t : ℕ) (ht : 2 ≤ t) (hH : 0 < H) (hL : sourceConfidenceLog t ≤ 2*H) (hm : 320000*H < (pullCount trace 0 t : ℝ)) : SourcePolicy.robustMean (by decide) 1 1 deterministicStream 0 t + sourceConfidenceRadius 1 1 t (pullCount trace 0 t) < (1/100 : ℝ)
theorem
BanditRLProof.HeavyTail.SourceCounterexample.forced_suboptimal
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.forced_suboptimalReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem forced_suboptimal (H : ℝ) (t : ℕ) (ht : 2 ≤ t) (hH : 30 < H) (hLlo : 2*H-2 ≤ sourceConfidenceLog t) (hLhi : sourceConfidenceLog t ≤ 2*H) (hnM : (pullCount trace 1 t : ℝ) ≤ 32*H+5) (hm : 320000*H < (pullCount trace 0 t : ℝ)) : trace t = 1
def
BanditRLProof.HeavyTail.SourceCounterexample.horizon
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.horizonReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
def horizon : ℕ
def
BanditRLProof.HeavyTail.SourceCounterexample.horizonHeight
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.horizonHeightReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
noncomputable def horizonHeight : ℝ
theorem
BanditRLProof.HeavyTail.SourceCounterexample.log_two_bounds
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.log_two_boundsReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem log_two_bounds : (3/5 : ℝ) < Real.log 2 ∧ Real.log 2 < 1
theorem
BanditRLProof.HeavyTail.SourceCounterexample.height_bounds
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.height_boundsReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem height_bounds : (30 : ℝ) < horizonHeight ∧ horizonHeight < 50
theorem
BanditRLProof.HeavyTail.SourceCounterexample.late_log_bounds
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.late_log_boundsReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem late_log_bounds (t : ℕ) (htlo : horizon/2 ≤ t) (hthi : t < horizon) : 2*horizonHeight-2 ≤ sourceConfidenceLog t ∧ sourceConfidenceLog t ≤ 2*horizonHeight
theorem
BanditRLProof.HeavyTail.SourceCounterexample.late_best_count
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.late_best_countReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem late_best_count (t : ℕ) (ht : horizon/2 ≤ t) (hthi : t ≤ horizon) (hcount : (pullCount trace 1 horizon : ℝ) ≤ 32*horizonHeight+5) : 320000*horizonHeight < (pullCount trace 0 t : ℝ)
theorem
BanditRLProof.HeavyTail.SourceCounterexample.finite_count_obstruction
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.finite_count_obstructionReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem finite_count_obstruction : 32*horizonHeight+5 < (pullCount trace 1 horizon : ℝ)
def
BanditRLProof.HeavyTail.SourceCounterexample.kernel
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.kernelReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
noncomputable def kernel : Kernel (Fin 2) ℝ
theorem
BanditRLProof.HeavyTail.SourceCounterexample.kernel_apply
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.kernel_applyReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
@[simp] theorem kernel_apply (a : Fin 2) : kernel a = Measure.dirac (if a = 0 then (0 : ℝ) else -1)
theorem
BanditRLProof.HeavyTail.SourceCounterexample.kernel_raw_moment
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Indexed settings: Heavy-tailed bandits
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.kernel_raw_momentReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem kernel_raw_moment (a : Fin 2) : Integrable (fun x : ℝ => |x|^(1+(1 : ℝ))) (kernel a) ∧ (∫ x : ℝ, |x|^(1+(1 : ℝ)) ∂kernel a) ≤ 1
theorem
BanditRLProof.HeavyTail.SourceCounterexample.kernel_mean
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.kernel_meanReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
@[simp] theorem kernel_mean (a : Fin 2) : realKernelMean kernel a = if a = 0 then 0 else -1
theorem
BanditRLProof.HeavyTail.SourceCounterexample.ae_deterministic
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.ae_deterministicReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem ae_deterministic : ∀ᵐ stream ∂UCB.armStreamMeasure kernel, stream = deterministicStream
theorem
BanditRLProof.HeavyTail.SourceCounterexample.kernel_best_mean
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.kernel_best_meanReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem kernel_best_mean : (⨆ a : Fin 2, realKernelMean kernel a) = 0
theorem
BanditRLProof.HeavyTail.SourceCounterexample.kernel_gap
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.kernel_gapReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem kernel_gap (a : Fin 2) : realMeanGap (realKernelMean kernel) a = if a = 0 then 0 else 1
theorem
BanditRLProof.HeavyTail.SourceCounterexample.regret_eq_count
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.regret_eq_countReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem regret_eq_count (action : ActionTrace (Fin 2)) (T : ℕ) : realMeanRegret (realKernelMean kernel) action T = (pullCount action 1 T : ℝ)
theorem
BanditRLProof.HeavyTail.SourceCounterexample.expected_regret_eq_count
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.expected_regret_eq_countReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem expected_regret_eq_count : (∫ stream, realMeanRegret (realKernelMean kernel) (SourcePolicy.robustAction (by decide) 1 1 stream) horizon ∂UCB.armStreamMeasure kernel) = (pullCount trace 1 horizon : ℝ)
theorem
BanditRLProof.HeavyTail.SourceCounterexample.printed_coefficient_counterexample
Compiled
The literal printed coefficient fails for the actual source-parameter policy under valid raw-second-moment stationary two-arm laws at a finite horizon.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.printed_coefficient_counterexampleReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem printed_coefficient_counterexample : 32 * Real.log (horizon : ℝ) + 5 < ∫ stream, realMeanRegret (realKernelMean kernel) (SourcePolicy.robustAction (by decide) 1 1 stream) horizon ∂UCB.armStreamMeasure kernel
theorem
BanditRLProof.HeavyTail.SourceCounterexample.printed_gap_sum
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.printed_gap_sumReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem printed_gap_sum : (∑ a ∈ Finset.univ.filter (fun a : Fin 2 => 0 < realMeanGap (realKernelMean kernel) a), (8 * (4 / realMeanGap (realKernelMean kernel) a) ^ (1 / (1 : ℝ)) * Real.log (horizon : ℝ) + 5 * realMeanGap (realKernelMean kernel) a)) = 32 * Real.log (horizon : ℝ) + 5
theorem
BanditRLProof.HeavyTail.SourceCounterexample.literal_printed_bound_false
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Indexed settings: Heavy-tailed bandits
Canonical node identity
declaration:BanditRLProof.HeavyTail.SourceCounterexample.literal_printed_bound_falseReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem literal_printed_bound_false : ¬ (∫ stream, realMeanRegret (realKernelMean kernel) (SourcePolicy.robustAction (by decide) 1 1 stream) horizon ∂UCB.armStreamMeasure kernel) ≤ ∑ a ∈ Finset.univ.filter (fun a : Fin 2 => 0 < realMeanGap (realKernelMean kernel) a), (8 * (4 / realMeanGap (realKernelMean kernel) a) ^ (1 / (1 : ℝ)) * Real.log (horizon : ℝ) + 5 * realMeanGap (realKernelMean kernel) a)