Documentation

LeanPool.ConcentrationInequalities.FreedmanBernstein

Freedman / Bernstein inequality for martingales (adapted sums of sub-gamma increments) #

The variance-based (Freedman) martingale tail: a finite adapted sum of conditionally sub-gamma increments satisfies P(∑ Yᵢ ≥ ε) ≤ exp(−ε²/(2(∑Vᵢ + cε))). Mathlib has Azuma (sub-Gaussian martingales) but not the sub-gamma / Freedman version. Built on the sub-gamma MGF and Bernstein tail of Contrib.BennettBernstein, with kernel-level tower additivity of the variance factors.

Sorry-free and axiom-clean [propext, Classical.choice, Quot.sound].

structure Contrib.Freedman.Kernel.HasSubgammaMGF {Ω : Type u_1} {Ω' : Type u_2} { : MeasurableSpace Ω} {mΩ' : MeasurableSpace Ω'} (X : Ω) (V c : NNReal) (κ : ProbabilityTheory.Kernel Ω' Ω) (ν : MeasureTheory.Measure Ω' := by volume_tac) :

Kernel-valued sub-gamma MGF bound. Integrability is required for every real parameter (rather than only the effective domain of the estimate), as in Mathlib's kernel-valued sub-Gaussian definition; this is what makes the tower argument composable.

Instances For
    theorem Contrib.Freedman.Kernel.HasSubgammaMGF.ae_integrable_exp_mul {Ω : Type u_1} {Ω' : Type u_2} { : MeasurableSpace Ω} {mΩ' : MeasurableSpace Ω'} {ν : MeasureTheory.Measure Ω'} {κ : ProbabilityTheory.Kernel Ω' Ω} {X : Ω} {V c : NNReal} (h : HasSubgammaMGF X V c κ ν) (t : ) :
    ∀ᵐ (ω' : Ω') ν, MeasureTheory.Integrable (fun (y : Ω) => Real.exp (t * X y)) (κ ω')
    theorem Contrib.Freedman.Kernel.HasSubgammaMGF.ae_forall_integrable_exp_mul {Ω : Type u_1} {Ω' : Type u_2} { : MeasurableSpace Ω} {mΩ' : MeasurableSpace Ω'} {ν : MeasureTheory.Measure Ω'} {κ : ProbabilityTheory.Kernel Ω' Ω} {X : Ω} {V c : NNReal} (h : HasSubgammaMGF X V c κ ν) :
    ∀ᵐ (ω' : Ω') ν, ∀ (t : ), MeasureTheory.Integrable (fun (ω : Ω) => Real.exp (t * X ω)) (κ ω')
    theorem Contrib.Freedman.Kernel.HasSubgammaMGF.memLp_exp_mul {Ω : Type u_1} {Ω' : Type u_2} { : MeasurableSpace Ω} {mΩ' : MeasurableSpace Ω'} {ν : MeasureTheory.Measure Ω'} {κ : ProbabilityTheory.Kernel Ω' Ω} {X : Ω} {V c : NNReal} (h : HasSubgammaMGF X V c κ ν) (t : ) (p : NNReal) :
    MeasureTheory.MemLp (fun (ω : Ω) => Real.exp (t * X ω)) (↑p) (ν.bind κ)
    @[simp]
    theorem Contrib.Freedman.Kernel.HasSubgammaMGF.zero_measure {Ω : Type u_1} {Ω' : Type u_2} { : MeasurableSpace Ω} {mΩ' : MeasurableSpace Ω'} {κ : ProbabilityTheory.Kernel Ω' Ω} {X : Ω} {V c : NNReal} :
    HasSubgammaMGF X V c κ 0
    theorem Contrib.Freedman.Kernel.HasSubgammaMGF.congr {Ω : Type u_1} {Ω' : Type u_2} { : MeasurableSpace Ω} {mΩ' : MeasurableSpace Ω'} {ν : MeasureTheory.Measure Ω'} {κ : ProbabilityTheory.Kernel Ω' Ω} {X : Ω} {V c : NNReal} {Y : Ω} (h : HasSubgammaMGF X V c κ ν) (h' : X =ᵐ[ν.bind κ] Y) :
    HasSubgammaMGF Y V c κ ν
    theorem Contrib.Freedman.Kernel.HasSubgammaMGF.prodMkLeft_compProd {Ω : Type u_1} {Ω' : Type u_2} { : MeasurableSpace Ω} {mΩ' : MeasurableSpace Ω'} {ν : MeasureTheory.Measure Ω'} {κ : ProbabilityTheory.Kernel Ω' Ω} {c : NNReal} {Ω'' : Type u_3} {mΩ'' : MeasurableSpace Ω''} {Y : Ω''} {W : NNReal} {η : ProbabilityTheory.Kernel Ω Ω''} (h : HasSubgammaMGF Y W c η (ν.bind κ)) :
    theorem Contrib.Freedman.Kernel.HasSubgammaMGF.integrable_exp_add_compProd {Ω : Type u_1} {Ω' : Type u_2} { : MeasurableSpace Ω} {mΩ' : MeasurableSpace Ω'} {ν : MeasureTheory.Measure Ω'} {κ : ProbabilityTheory.Kernel Ω' Ω} {X : Ω} {V c : NNReal} {Ω'' : Type u_3} {mΩ'' : MeasurableSpace Ω''} {Y : Ω''} {W : NNReal} [MeasureTheory.SFinite ν] {η : ProbabilityTheory.Kernel (Ω' × Ω) Ω''} [ProbabilityTheory.IsZeroOrMarkovKernel η] (hX : HasSubgammaMGF X V c κ ν) (hY : HasSubgammaMGF Y W c η (ν.compProd κ)) (t : ) :
    MeasureTheory.Integrable (fun (ω : Ω × Ω'') => Real.exp (t * (X ω.1 + Y ω.2))) (ν.bind (κ.compProd η))
    theorem Contrib.Freedman.Kernel.HasSubgammaMGF.add_compProd {Ω : Type u_1} {Ω' : Type u_2} { : MeasurableSpace Ω} {mΩ' : MeasurableSpace Ω'} {ν : MeasureTheory.Measure Ω'} {κ : ProbabilityTheory.Kernel Ω' Ω} {X : Ω} {V c : NNReal} {Ω'' : Type u_3} {mΩ'' : MeasurableSpace Ω''} {Y : Ω''} {W : NNReal} [MeasureTheory.SFinite ν] {η : ProbabilityTheory.Kernel (Ω' × Ω) Ω''} [ProbabilityTheory.IsZeroOrMarkovKernel η] (hX : HasSubgammaMGF X V c κ ν) (hY : HasSubgammaMGF Y W c η (ν.compProd κ)) :
    HasSubgammaMGF (fun (p : Ω × Ω'') => X p.1 + Y p.2) (V + W) c (κ.compProd η) ν

    Tower-property additivity of variance factors for two consecutive kernel increments.

    theorem Contrib.Freedman.Kernel.HasSubgammaMGF.add_of_indep {Ω : Type u_1} {Ω' : Type u_2} { : MeasurableSpace Ω} {mΩ' : MeasurableSpace Ω'} {ν : MeasureTheory.Measure Ω'} {κ : ProbabilityTheory.Kernel Ω' Ω} {X : Ω} {V c : NNReal} {Ω'' : Type u_3} {mΩ'' : MeasurableSpace Ω''} {Y : Ω''} {W : NNReal} [MeasureTheory.SFinite ν] {η : ProbabilityTheory.Kernel Ω Ω''} [ProbabilityTheory.IsZeroOrMarkovKernel η] (hX : HasSubgammaMGF X V c κ ν) (hY : HasSubgammaMGF Y W c η (ν.bind κ)) :
    HasSubgammaMGF (fun (p : Ω × Ω'') => X p.1 + Y p.2) (V + W) c (κ.compProd (ProbabilityTheory.Kernel.prodMkLeft Ω' η)) ν

    Additivity in the usual sequential-kernel formulation.

    def Contrib.Freedman.HasCondSubgammaMGF {Ω : Type u_1} (m : MeasurableSpace Ω) { : MeasurableSpace Ω} (hm : m ) [StandardBorelSpace Ω] (X : Ω) (V c : NNReal) (μ : MeasureTheory.Measure Ω := by volume_tac) [MeasureTheory.IsFiniteMeasure μ] :

    Conditional sub-gamma MGF, expressed using the conditional-expectation kernel.

    Equations
    Instances For
      theorem Contrib.Freedman.HasCondSubgammaMGF.integrable_exp_mul {Ω : Type u_1} {m : MeasurableSpace Ω} {X : Ω} {V c : NNReal} {hm : m } [StandardBorelSpace Ω] {μ : MeasureTheory.Measure Ω} [MeasureTheory.IsFiniteMeasure μ] (h : HasCondSubgammaMGF m hm X V c μ) (t : ) :
      MeasureTheory.Integrable (fun (ω : Ω) => Real.exp (t * X ω)) μ
      theorem Contrib.Freedman.HasCondSubgammaMGF.ae_condExp_le {Ω : Type u_1} {m : MeasurableSpace Ω} {X : Ω} {V c : NNReal} {hm : m } [StandardBorelSpace Ω] {μ : MeasureTheory.Measure Ω} [MeasureTheory.IsFiniteMeasure μ] (h : HasCondSubgammaMGF m hm X V c μ) {t : } (ht : 0 t) (htc : c * t < 1) :
      ∀ᵐ (ω : Ω) μ, μ[fun (x : Ω) => Real.exp (t * X x) | m] ω Real.exp (V * t ^ 2 / (2 * (1 - c * t)))
      structure Contrib.Freedman.MeasureSubgammaMGF {Ω : Type u_1} { : MeasurableSpace Ω} (X : Ω) (V c : NNReal) (μ : MeasureTheory.Measure Ω := by volume_tac) :

      Measure-valued sub-gamma MGF bound, including exponential integrability.

      Instances For
        theorem Contrib.Freedman.MeasureSubgammaMGF.of_map {Ω : Type u_1} {Ω' : Type u_2} { : MeasurableSpace Ω} {mΩ' : MeasurableSpace Ω'} {X : Ω} {V c : NNReal} {μ : MeasureTheory.Measure Ω'} {Y : Ω'Ω} (hY : AEMeasurable Y μ) (h : MeasureSubgammaMGF X V c (MeasureTheory.Measure.map Y μ)) :
        MeasureSubgammaMGF (X Y) V c μ
        theorem Contrib.Freedman.MeasureSubgammaMGF.trim {Ω : Type u_1} {m : MeasurableSpace Ω} {μ : MeasureTheory.Measure Ω} {X : Ω} {V c : NNReal} (hm : m ) (hXm : Measurable X) (hX : MeasureSubgammaMGF X V c μ) :
        MeasureSubgammaMGF X V c (μ.trim hm)
        theorem Contrib.Freedman.MeasureSubgammaMGF.add_of_hasCondSubgammaMGF {Ω : Type u_1} {m : MeasurableSpace Ω} {μ : MeasureTheory.Measure Ω} {X : Ω} [StandardBorelSpace Ω] [MeasureTheory.IsFiniteMeasure μ] {Y : Ω} {V W c : NNReal} (hm : m ) (hX : MeasureSubgammaMGF X V c (μ.trim hm)) (hY : HasCondSubgammaMGF m hm Y W c μ) :
        MeasureSubgammaMGF (X + Y) (V + W) c μ
        theorem Contrib.Freedman.MeasureSubgammaMGF.sum_of_hasCondSubgammaMGF {Ω : Type u_1} { : MeasurableSpace Ω} {μ : MeasureTheory.Measure Ω} [StandardBorelSpace Ω] {Y : Ω} {V : NNReal} {c : NNReal} { : MeasureTheory.Filtration } [MeasureTheory.IsZeroOrProbabilityMeasure μ] (h_adapted : MeasureTheory.StronglyAdapted Y) (h0 : MeasureSubgammaMGF (Y 0) (V 0) c μ) (n : ) (h_subG : i < n - 1, HasCondSubgammaMGF ( i) (Y (i + 1)) (V (i + 1)) c μ) :
        MeasureSubgammaMGF (fun (ω : Ω) => iFinset.range n, Y i ω) (∑ iFinset.range n, V i) c μ
        theorem Contrib.Freedman.subgamma_bernstein_tail {Ω : Type u_1} { : MeasurableSpace Ω} {μ : MeasureTheory.Measure Ω} [StandardBorelSpace Ω] {Y : Ω} {V : NNReal} {c : NNReal} { : MeasureTheory.Filtration } [MeasureTheory.IsZeroOrProbabilityMeasure μ] (h_adapted : MeasureTheory.StronglyAdapted Y) (h0 : MeasureSubgammaMGF (Y 0) (V 0) c μ) (n : ) (h_subG : i < n - 1, HasCondSubgammaMGF ( i) (Y (i + 1)) (V (i + 1)) c μ) {ε : } (hV : 0 < iFinset.range n, V i) (hc : 0 < c) ( : 0 ε) :
        μ.real {ω : Ω | ε iFinset.range n, Y i ω} Real.exp (-ε ^ 2 / (2 * ((∑ iFinset.range n, V i) + c * ε)))

        Freedman/Bernstein upper tail for a finite adapted sum of conditionally sub-gamma increments.