Threshold separation for sums of geometric variables, uniform in the scale of the means (binomial thinning)
ProvedBosonicLearn.geometric_threshold_scaledA version of BosonicLearn.geometric_pgf_threshold whose constant does not depend on the bound of the means; it is what makes the photon-counting argument of Lemma 12 work for hot states (, Theorem 9).
There is an absolute constant such that for every : let () and let be the laws on of , sums of independent geometric variables with means , given through for . Put . If and , then there is a threshold with
Proof idea: binomial thinning with probability maps a sum of geometric variables with means to one with means (the generating function is evaluated at ), and a thinned threshold test is a mixture of threshold tests (Abel summation), so the case applies.
Formalization Note Laws are sequences ℕ → ℝ with non-negative entries determined by the generating-function identities; tails are ∑' k, if τ < k then p k else 0.
import Mathlib
namespace BosonicLearn
/-- **Threshold separation for sums of geometric variables, uniformly in the scale of the means.**
Chen, Anna Mele, Anna Mele, Preskill, *When are bosonic Gaussian states classical to learn?*, arXiv:2609.26705, §5.7, Lemmas 13–14 and Eqs. (137)–(142) (threshold separation for the photon count), used in §8 for hot states
(`ν ≥ 1`, proof of Theorem 9, p. 66).
There is an absolute constant `c > 0` such that for every `Λ ≥ 1`: let `0 ≤ λ⁻_i ≤ λ⁺_i ≤ Λ` and let `p⁺, p⁻` be the laws
on `ℕ` of sums of independent geometric variables with means `λ^±_i` (given through their generating functions
`∑ₖ p^±_k zᵏ = ∏ᵢ 1/(1 + λ^±_i(1 − z))`, `0 ≤ z ≤ 1`). Put `δ = ∑ᵢ(λ⁺_i − λ⁻_i)`. If `0 < δ` and
`δ² ≤ 1 + ∑ᵢ λ⁻_i(1 + λ⁻_i)`, then there is a threshold `τ` with
`Pr(K₊ > τ) − Pr(K₋ > τ) ≥ c (δ/Λ) / √(1 + ∑ᵢ (λ⁻_i/Λ)(1 + λ⁻_i/Λ))`.
**Formalization Note.** This is `geometric_pgf_threshold` (whose constant depends on the bound `Λ` of the means) made
uniform: binomial thinning with probability `1/Λ` maps a sum of geometric variables with means `λ_i` to one with means
`λ_i/Λ ≤ 1`, and a thinned threshold test is a mixture of threshold tests. -/
theorem geometric_threshold_scaled :
∃ c : ℝ, 0 < c ∧ ∀ Λ : ℝ, 1 ≤ Λ →
∀ (m : ℕ) (lp lm : Fin m → ℝ), (∀ i, 0 ≤ lm i) → (∀ i, lm i ≤ lp i) → (∀ i, lp i ≤ Λ) →
∀ (pp pm : ℕ → ℝ), (∀ k, 0 ≤ pp k) → (∀ k, 0 ≤ pm k) →
(∀ z : ℝ, 0 ≤ z → z ≤ 1 →
HasSum (fun k => pp k * z ^ k) (∏ i, (1 + lp i * (1 - z))⁻¹)) →
(∀ z : ℝ, 0 ≤ z → z ≤ 1 →
HasSum (fun k => pm k * z ^ k) (∏ i, (1 + lm i * (1 - z))⁻¹)) →
0 < ∑ i, (lp i - lm i) →
(∑ i, (lp i - lm i)) ^ 2 ≤ 1 + ∑ i, lm i * (1 + lm i) →
∃ τ : ℕ, c * ((∑ i, (lp i - lm i)) / Λ) /
Real.sqrt (1 + ∑ i, lm i / Λ * (1 + lm i / Λ)) ≤
(∑' k, if τ < k then pp k else 0) - ∑' k, if τ < k then pm k else 0 := by sorry
end BosonicLearn