Lean module · Frontier
BanditRLProof.Algorithms.StochasticGradientBanditExponentialAudit
This module formalizes the source constant C_eta and Equation (8) from the two-arm proof of Baudry--Johnson--Vary--Pike-Burke--Rebeschini (NeurIPS 2025). For an almost-everywhere measurable reward supported on [-1, 1], it derives the exact second-order moment-generating-function inequality used by the paper. It also proves the source comparison C_eta <= exp (2 * eta) for nonnegative eta.
Module map
Imports
BanditRLProof.Algorithms.StochasticGradientBanditTwoArmRate
Imported by
BanditRLProof, BanditRLProof.Algorithms.StochasticGradientBanditConditionalExponentialAudit
Declarations
Open an item to read its exact compact statement and source link. Detailed teaching notes are linked when registered.
theorem
BanditRLProof.StochasticGradientBandit.two_mul_abs_pow_div_factorial_add_two_le
Compiled
A factorial comparison used to dominate the shifted exponential series.
Used in these reading views: Bandit Book
10. Automation, resources, and open routes
Canonical node identity
declaration:BanditRLProof.StochasticGradientBandit.two_mul_abs_pow_div_factorial_add_two_leReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem two_mul_abs_pow_div_factorial_add_two_le (x : Real) (n : Nat) : 2 * |x| ^ n / ((n + 2).factorial : Real) <= |x| ^ n / (n.factorial : Real)
def
BanditRLProof.StochasticGradientBandit.sourceC
Compiled
The source constant `C_eta = 2 * sum_{n >= 0} (2 * eta)^n / (n + 2)!`.
Used in these reading views: Bandit Book
10. Automation, resources, and open routes
Canonical node identity
declaration:BanditRLProof.StochasticGradientBandit.sourceCReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
noncomputable def sourceC (eta : Real) : Real
theorem
BanditRLProof.StochasticGradientBandit.sourceC_terms_summable
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
10. Automation, resources, and open routes
Canonical node identity
declaration:BanditRLProof.StochasticGradientBandit.sourceC_terms_summableReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem sourceC_terms_summable (eta : Real) : Summable (fun n : Nat => (2 * eta) ^ n / ((n + 2).factorial : Real))
theorem
BanditRLProof.StochasticGradientBandit.sourceC_nonneg
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
10. Automation, resources, and open routes
Canonical node identity
declaration:BanditRLProof.StochasticGradientBandit.sourceC_nonnegReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem sourceC_nonneg (eta : Real) (heta : 0 <= eta) : 0 <= sourceC eta
theorem
BanditRLProof.StochasticGradientBandit.sourceC_mono
Compiled
Monotonicity needed to replace the time-varying source constants in the Theorem-1 recurrences by a common `C_eta`.
Used in these reading views: Bandit Book
10. Automation, resources, and open routes
Canonical node identity
declaration:BanditRLProof.StochasticGradientBandit.sourceC_monoReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem sourceC_mono {eta eta' : Real} (heta : 0 <= eta) (hle : eta <= eta') : sourceC eta <= sourceC eta'
theorem
BanditRLProof.StochasticGradientBandit.sourceC_le_exp_two_mul
Compiled
The source comparison following Theorem 1: `C_eta <= exp (2 * eta)`.
Used in these reading views: Bandit Book
10. Automation, resources, and open routes
Canonical node identity
declaration:BanditRLProof.StochasticGradientBandit.sourceC_le_exp_two_mulReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem sourceC_le_exp_two_mul (eta : Real) (heta : 0 <= eta) : sourceC eta <= Real.exp (2 * eta)
def
BanditRLProof.StochasticGradientBandit.expTailTwo
Compiled
The exponential-series tail beginning at degree two.
Used in these reading views: Bandit Book
10. Automation, resources, and open routes
Canonical node identity
declaration:BanditRLProof.StochasticGradientBandit.expTailTwoReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
noncomputable def expTailTwo (x : Real) : Real
theorem
BanditRLProof.StochasticGradientBandit.expTailTwo_terms_summable
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
10. Automation, resources, and open routes
Canonical node identity
declaration:BanditRLProof.StochasticGradientBandit.expTailTwo_terms_summableReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem expTailTwo_terms_summable (x : Real) : Summable (fun n : Nat => x ^ (n + 2) / ((n + 2).factorial : Real))
theorem
BanditRLProof.StochasticGradientBandit.exp_eq_one_add_self_add_expTailTwo
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
10. Automation, resources, and open routes
Canonical node identity
declaration:BanditRLProof.StochasticGradientBandit.exp_eq_one_add_self_add_expTailTwoReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem exp_eq_one_add_self_add_expTailTwo (x : Real) : Real.exp x = 1 + x + expTailTwo x
theorem
BanditRLProof.StochasticGradientBandit.expTailTwo_le_of_abs_le
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
10. Automation, resources, and open routes
Canonical node identity
declaration:BanditRLProof.StochasticGradientBandit.expTailTwo_le_of_abs_leReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem expTailTwo_le_of_abs_le {x y : Real} (hxy : |x| <= y) : expTailTwo x <= expTailTwo y
theorem
BanditRLProof.StochasticGradientBandit.sq_div_two_mul_sourceC_abs_div_two
Compiled
No declaration docstring is present; use the chapter context and exact statement below.
Used in these reading views: Bandit Book
10. Automation, resources, and open routes
Canonical node identity
declaration:BanditRLProof.StochasticGradientBandit.sq_div_two_mul_sourceC_abs_div_twoReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem sq_div_two_mul_sourceC_abs_div_two (q : Real) : q ^ 2 / 2 * sourceC (|q| / 2) = expTailTwo |q|
theorem
BanditRLProof.StochasticGradientBandit.exp_mul_le_sourceEqEight
Compiled
Pointwise form of source Equation (8).
Used in these reading views: Bandit Book
10. Automation, resources, and open routes
Canonical node identity
declaration:BanditRLProof.StochasticGradientBandit.exp_mul_le_sourceEqEightReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem exp_mul_le_sourceEqEight (q reward : Real) (hreward : |reward| <= 1) : Real.exp (q * reward) <= 1 + q * reward + q ^ 2 / 2 * sourceC (|q| / 2)
theorem
BanditRLProof.StochasticGradientBandit.integral_exp_mul_le_sourceEqEight
Compiled
Expectation form of source Equation (8), with integrability explicit.
Used in these reading views: Bandit Book
10. Automation, resources, and open routes
Canonical node identity
declaration:BanditRLProof.StochasticGradientBandit.integral_exp_mul_le_sourceEqEightReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem integral_exp_mul_le_sourceEqEight {Omega : Type*} [MeasurableSpace Omega] (mu : MeasureTheory.Measure Omega) [MeasureTheory.IsProbabilityMeasure mu] (q : Real) (reward : Omega -> Real) (hrewardIntegrable : MeasureTheory.Integrable reward mu) (hexpIntegrable : MeasureTheory.Integrable (fun omega => Real.exp (q * reward omega)) mu) (hreward : ∀ᵐ omega ∂mu, |reward omega| <= 1) : (∫ omega, Real.exp (q * reward omega) ∂mu) <= 1 + q * (∫ omega, reward omega ∂mu) + q ^ 2 / 2 * sourceC (|q| / 2)
theorem
BanditRLProof.StochasticGradientBandit.integral_exp_mul_le_sourceEqEight_of_ae_abs_le_one
Compiled
Equation (8) from measurability and almost-sure support in `[-1, 1]`.
Used in these reading views: Bandit Book
10. Automation, resources, and open routes
Canonical node identity
declaration:BanditRLProof.StochasticGradientBandit.integral_exp_mul_le_sourceEqEight_of_ae_abs_le_oneReading membership is not a proof dependency. Exact assumptions remain in the Lean statement.
theorem integral_exp_mul_le_sourceEqEight_of_ae_abs_le_one {Omega : Type*} [MeasurableSpace Omega] (mu : MeasureTheory.Measure Omega) [MeasureTheory.IsProbabilityMeasure mu] (q : Real) (reward : Omega -> Real) (hrewardMeasurable : MeasureTheory.AEStronglyMeasurable reward mu) (hreward : ∀ᵐ omega ∂mu, |reward omega| <= 1) : (∫ omega, Real.exp (q * reward omega) ∂mu) <= 1 + q * (∫ omega, reward omega ∂mu) + q ^ 2 / 2 * sourceC (|q| / 2)