Lean module · Foundations
BanditRLProof.HeavyTailSourceConfidence
Arbitrary-log-confidence sample-index truncation. The target is the constant four confidence radius of BCL 2013 Lemma 1; algorithm regret remains separate.
Module map
Imports
BanditRLProof.HeavyTailUnshiftedMGF, BanditRLProof.HeavyTailConfidence, BanditRLProof.HeavyTailTuning
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.sourceTruncationThreshold
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.sourceTruncationThresholdReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
noncomputable def sourceTruncationThreshold (ε u L : ℝ) (s : ℕ) : ℝ
theorem
BanditRLProof.HeavyTail.sourceThreshold_pos
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.sourceThreshold_posReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem sourceThreshold_pos (ε u L : ℝ) (hu : 0 < u) (hL : 0 < L) (s : ℕ) : 0 < sourceTruncationThreshold ε u L s
theorem
BanditRLProof.HeavyTail.sourceThreshold_le_terminal
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.sourceThreshold_le_terminalReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem sourceThreshold_le_terminal (ε u L : ℝ) (hε : 0 ≤ ε) (hu : 0 ≤ u) (hL : 0 < L) (n s : ℕ) (hs : s < n) : sourceTruncationThreshold ε u L s ≤ (u*n/L)^(1/(1+ε))
theorem
BanditRLProof.HeavyTail.sourceThreshold_bias_average
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.sourceThreshold_bias_averageReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem sourceThreshold_bias_average (ε u L : ℝ) (hε : 0 ≤ ε) (hu : 0 < u) (hL : 0 < L) (n : ℕ) (hn : 0 < n) : (∑ s ∈ Finset.range n, u / (sourceTruncationThreshold ε u L s)^ε) / n ≤ (1+ε)*u^(1/(1+ε))*(L/n)^(ε/(1+ε))
theorem
BanditRLProof.HeavyTail.sourceThreshold_variance_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.sourceThreshold_variance_sumReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem sourceThreshold_variance_sum (ε u L : ℝ) (hε0 : 0 ≤ ε) (hε : ε ≤ 1) (hu : 0 < u) (hL : 0 < L) (n : ℕ) : (∑ s ∈ Finset.range n, u*(sourceTruncationThreshold ε u L s)^(1-ε)) ≤ n*u*((u*n/L)^(1/(1+ε)))^(1-ε)
theorem
BanditRLProof.HeavyTail.source_centered_sum_upper_tail_sharp
Compiled
One-sided centered-sum bound at the full raw-variable tilt 1/B.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.source_centered_sum_upper_tail_sharpReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem source_centered_sum_upper_tail_sharp {Ω : Type*} [MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ] (X : ℕ → Ω → ℝ) (ε u L : ℝ) (n : ℕ) (hn : 0 < n) (hXm : ∀ i, Measurable (X i)) (hi : iIndepFun X μ) (hε0 : 0 ≤ ε) (hε : ε ≤ 1) (hu : 0 < u) (hL : 0 < L) (hm : ∀ i, Integrable (fun ω => |X i ω|^(1+ε)) μ) (hraw : ∀ i, (∫ ω, |X i ω|^(1+ε) ∂μ) ≤ u) : μ.real {ω | 2*(u*n/L)^(1/(1+ε))*L ≤ ∑ i ∈ Finset.range n, (truncate (sourceTruncationThreshold ε u L i) (X i ω) - ∫ ω, truncate (sourceTruncationThreshold ε u L i) (X i ω) ∂μ)} ≤ Real.exp (-(5/4 : ℝ)*L)
theorem
BanditRLProof.HeavyTail.source_centered_sum_upper_tail
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.source_centered_sum_upper_tailReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem source_centered_sum_upper_tail {Ω : Type*} [MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ] (X : ℕ → Ω → ℝ) (ε u L : ℝ) (n : ℕ) (hn : 0 < n) (hXm : ∀ i, Measurable (X i)) (hi : iIndepFun X μ) (hε0 : 0 ≤ ε) (hε : ε ≤ 1) (hu : 0 < u) (hL : 0 < L) (hm : ∀ i, Integrable (fun ω => |X i ω|^(1+ε)) μ) (hraw : ∀ i, (∫ ω, |X i ω|^(1+ε) ∂μ) ≤ u) : μ.real {ω | 2*(u*n/L)^(1/(1+ε))*L ≤ ∑ i ∈ Finset.range n, (truncate (sourceTruncationThreshold ε u L i) (X i ω) - ∫ ω, truncate (sourceTruncationThreshold ε u L i) (X i ω) ∂μ)} ≤ Real.exp (-L)
theorem
BanditRLProof.HeavyTail.source_truncated_mean_upper_tail_log_sharp
Compiled
Constant-four upper deviation for arbitrary positive log confidence.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.source_truncated_mean_upper_tail_log_sharpReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem source_truncated_mean_upper_tail_log_sharp {Ω : Type*} [MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ] (X : ℕ → Ω → ℝ) (ε u L mean : ℝ) (n : ℕ) (hn : 0 < n) (hXm : ∀ i, Measurable (X i)) (hi : iIndepFun X μ) (hε0 : 0 ≤ ε) (hε : ε ≤ 1) (hu : 0 < u) (hL : 0 < L) (hX : ∀ i, Integrable (X i) μ) (hmean : ∀ i, (∫ ω, X i ω ∂μ) = mean) (hm : ∀ i, Integrable (fun ω => |X i ω|^(1+ε)) μ) (hraw : ∀ i, (∫ ω, |X i ω|^(1+ε) ∂μ) ≤ u) : μ.real {ω | 4*u^(1/(1+ε))*(L/n)^(ε/(1+ε)) ≤ (∑ i ∈ Finset.range n, truncate (sourceTruncationThreshold ε u L i) (X i ω))/n - mean} ≤ Real.exp (-(5/4 : ℝ)*L)
theorem
BanditRLProof.HeavyTail.source_truncated_mean_upper_tail_log
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.source_truncated_mean_upper_tail_logReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem source_truncated_mean_upper_tail_log {Ω : Type*} [MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ] (X : ℕ → Ω → ℝ) (ε u L mean : ℝ) (n : ℕ) (hn : 0 < n) (hXm : ∀ i, Measurable (X i)) (hi : iIndepFun X μ) (hε0 : 0 ≤ ε) (hε : ε ≤ 1) (hu : 0 < u) (hL : 0 < L) (hX : ∀ i, Integrable (X i) μ) (hmean : ∀ i, (∫ ω, X i ω ∂μ) = mean) (hm : ∀ i, Integrable (fun ω => |X i ω|^(1+ε)) μ) (hraw : ∀ i, (∫ ω, |X i ω|^(1+ε) ∂μ) ≤ u) : μ.real {ω | 4*u^(1/(1+ε))*(L/n)^(ε/(1+ε)) ≤ (∑ i ∈ Finset.range n, truncate (sourceTruncationThreshold ε u L i) (X i ω))/n - mean} ≤ Real.exp (-L)
theorem
BanditRLProof.HeavyTail.source_truncated_mean_upper_tail
Compiled
BCL 2013 Lemma 1 upper deviation, retaining its radius constant four. The non-strict bad event proved here is stronger than a strict upper-tail event.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Indexed settings: Heavy-tailed bandits
Canonical node identity
declaration:BanditRLProof.HeavyTail.source_truncated_mean_upper_tailReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem source_truncated_mean_upper_tail {Ω : Type*} [MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ] (X : ℕ → Ω → ℝ) (ε u δ mean : ℝ) (n : ℕ) (hn : 0 < n) (hXm : ∀ i, Measurable (X i)) (hi : iIndepFun X μ) (hε0 : 0 < ε) (hε : ε ≤ 1) (hu : 0 < u) (hδ : 0 < δ) (hδ1 : δ < 1) (hX : ∀ i, Integrable (X i) μ) (hmean : ∀ i, (∫ ω, X i ω ∂μ) = mean) (hm : ∀ i, Integrable (fun ω => |X i ω|^(1+ε)) μ) (hraw : ∀ i, (∫ ω, |X i ω|^(1+ε) ∂μ) ≤ u) : μ.real {ω | 4*u^(1/(1+ε))*(Real.log (1/δ)/n)^(ε/(1+ε)) ≤ (∑ i ∈ Finset.range n, truncate (sourceTruncationThreshold ε u (Real.log (1/δ)) i) (X i ω))/n - mean} ≤ δ
theorem
BanditRLProof.HeavyTail.truncate_neg
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.truncate_negReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem truncate_neg (B x : ℝ) : truncate B (-x) = -truncate B x
theorem
BanditRLProof.HeavyTail.source_truncated_mean_lower_tail
Compiled
Reflection supplies the other one-sided source confidence statement.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Indexed settings: Heavy-tailed bandits
Canonical node identity
declaration:BanditRLProof.HeavyTail.source_truncated_mean_lower_tailReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem source_truncated_mean_lower_tail {Ω : Type*} [MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ] (X : ℕ → Ω → ℝ) (ε u δ mean : ℝ) (n : ℕ) (hn : 0 < n) (hXm : ∀ i, Measurable (X i)) (hi : iIndepFun X μ) (hε0 : 0 < ε) (hε : ε ≤ 1) (hu : 0 < u) (hδ : 0 < δ) (hδ1 : δ < 1) (hX : ∀ i, Integrable (X i) μ) (hmean : ∀ i, (∫ ω, X i ω ∂μ) = mean) (hm : ∀ i, Integrable (fun ω => |X i ω|^(1+ε)) μ) (hraw : ∀ i, (∫ ω, |X i ω|^(1+ε) ∂μ) ≤ u) : μ.real {ω | 4*u^(1/(1+ε))*(Real.log (1/δ)/n)^(ε/(1+ε)) ≤ mean - (∑ i ∈ Finset.range n, truncate (sourceTruncationThreshold ε u (Real.log (1/δ)) i) (X i ω))/n} ≤ δ
theorem
BanditRLProof.HeavyTail.source_truncated_mean_lower_tail_log_sharp
Compiled
Sharper lower log-confidence tail, obtained by reflection.
Used in these reading views: Bandit Book
1. Finite bandits, traces, and regret
Canonical node identity
declaration:BanditRLProof.HeavyTail.source_truncated_mean_lower_tail_log_sharpReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem source_truncated_mean_lower_tail_log_sharp {Ω : Type*} [MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ] (X : ℕ → Ω → ℝ) (ε u L mean : ℝ) (n : ℕ) (hn : 0 < n) (hXm : ∀ i, Measurable (X i)) (hi : iIndepFun X μ) (hε0 : 0 ≤ ε) (hε : ε ≤ 1) (hu : 0 < u) (hL : 0 < L) (hX : ∀ i, Integrable (X i) μ) (hmean : ∀ i, (∫ ω, X i ω ∂μ) = mean) (hm : ∀ i, Integrable (fun ω => |X i ω|^(1+ε)) μ) (hraw : ∀ i, (∫ ω, |X i ω|^(1+ε) ∂μ) ≤ u) : μ.real {ω | 4*u^(1/(1+ε))*(L/n)^(ε/(1+ε)) ≤ mean - (∑ i ∈ Finset.range n, truncate (sourceTruncationThreshold ε u (L) i) (X i ω))/n} ≤ Real.exp (-(5/4 : ℝ)*L)