Documentation

LeanPool.Superdiffusion.Algsuperdiff.Probability.GammaSigmaFiniteTriangle

GammaSigmaFiniteTriangle #

Ported from scottnarmstrong/Superdiffusion at 7f20476385d4af418f110303085ad0740c9eba46. Module imports, support namespace and local option directives adapted for Lean Pool.

Measurability-free finite triangle inequality for O_{Γ_σ} #

ABK26, Lemma l.Gamma.sigma.triangle (e.Gamma.sigma.triangle), in the finite case.

What this adds over CoarseGraining #

CoarseGraining already proves a finite generalized triangle inequality, Homogenization.IndependentSums.isBigO_finset_sum_of_isBigO_gammaSigma, but it carries two hypotheses that are painful downstream:

The requirement is unnecessary. The finite triangle has a direct union-bound proof: the tail event of the sum is contained in a finite union of the summands' tail events, and MeasureTheory.measureReal_biUnion_finset_le holds for completely arbitrary sets. This module gives that proof. No measurability hypothesis appears anywhere below, and scales are only assumed nonnegative.

The constant #

For an N-term sum the scale produced here is (1 + log N)^{1/σ} · ∑ a_i. This is both sharper and more explicit than CoarseGraining's gammaTriangleConst σ (which is σ-dependent and enormous). The mechanism: splitting the budget as a_i · (λ t) with λ = (1 + log N)^{1/σ} makes each of the N summand tails at most exp(-(1 + log N) t^σ), and N · exp(-(1 + log N) t^σ) ≤ exp(-t^σ) exactly because t^σ ≥ 1.

Main results #

References #

theorem Algsuperdiff.Probability.isBigO_gammaSigma_finset_sum_of_nonneg_scales {Ω : Type u_1} {ι : Type u_2} [MeasurableSpace Ω] {μ : MeasureTheory.Measure Ω} [MeasureTheory.IsFiniteMeasure μ] (s : Finset ι) {X : ι → Ω → ℝ} {a : ι → ℝ} {σ : ℝ} (hσ : 0 < σ) (hs : s.Nonempty) (ha : ∀ i ∈ s, 0 ≤ a i) (hX : ∀ i ∈ s, HCPolySupport.IndependentSums.IsBigO μ (HCPolySupport.IndependentSums.gammaSigma σ) (X i) (a i)) :
HCPolySupport.IndependentSums.IsBigO μ (HCPolySupport.IndependentSums.gammaSigma σ) (fun (ω : Ω) => ∑ i ∈ s, X i ω) ((1 + Real.log ↑s.card) ^ σ⁻¹ * ∑ i ∈ s, a i)

Measurability-free finite generalized triangle inequality for the stretched-exponential class Γ_σ (l.Gamma.sigma.triangle, finite case).

If X i = O_{Γ_σ}(a i) for each i in a nonempty finite index set s, with nonnegative scales a i, then

∑_{i ∈ s} X i = O_{Γ_σ}( (1 + log #s)^{1/σ} · ∑_{i ∈ s} a i ).

Unlike CoarseGraining's isBigO_finset_sum_of_isBigO_gammaSigma, no measurability of the summands is required — the proof is a pure union bound, and measureReal_biUnion_finset_le needs no measurable sets.

Two-term measurability-free Γ_σ triangle inequality, constant (1 + log 2)^{1/σ}.

Three-term measurability-free Γ_σ triangle inequality, constant (1 + Real.log 3) ^ σ⁻¹.