Gaussian samples: and with probability
ProvedBosonicLearn.gaussian_sample_concentrationMean and covariance of Gaussian samples concentrate in relative (whitened) form, with failure probability .
Let be a positive definite real matrix, , and . Put
Let be i.i.d. , and . Then with probability at least , simultaneously
Relation to Eq. (210). This is the event of Eq. (210) (BosonicLearn.eq_210) with the deviation parameter left free. The three terms of the failure probability come from the norm of the whitened sample mean and from the smallest and largest singular values of the whitened data matrix. Taking gives probability at least ; taking , the choice of Eq. (209), this argument gives . For the proof of Theorem 8 the difference only changes the universal constant.
Role. It replaces BosonicLearn.eq_210 in the proof of Theorem 8 (heterodyne tomography), where is the covariance of the heterodyne outcomes.
Formalization Note Same event set as eq_210: samples are Measure.pi fun _ : Fin N => multivariateGaussian (WithLp.toLp 2 μ) C on EuclideanSpace ℝ (Fin n ⊕ Fin n), Loewner order is Matrix.PosSemidef of the difference, and the first inequality is the squared form of . The bound is 1 - ENNReal.ofReal (3 * exp (-t^2/2)). The proof imports the proved theorems GaussianMatrix.extreme_singular_values_deviation, GaussianMatrix.gaussian_concentration and GaussianMatrix.frobenius_second_moment of the mission series on Gaussian random matrices.
import Mathlib open MeasureTheory ProbabilityTheory open scoped ENNReal Matrix
namespace BosonicLearn
/-- **Mean and covariance concentration for `N` Gaussian samples, with failure probability `3e^{−t²/2}`.**
Chen, Anna Mele, Anna Mele, Preskill, *When are bosonic Gaussian states classical to learn?*, arXiv:2609.26705, §7.2, proof of Theorem 8, Eqs. (208)–(210), p. 61: "Set χ := √(2n) + √(2 log(2/δ)), ζ := 2χ/√N + 2χ²/N. The
Gaussian concentration bound of [BMEM25, Lemma 9, Eqs. (91), (94), and (95)] implies that, with probability at
least 1 − δ, ‖C^{−1/2}(µ̂ − µ)‖_2 ≤ χ/√N, (1 − ζ)C ⪯ Ĉ ⪯ (1 + ζ)C."
Let `C` be a positive definite `2n × 2n` matrix, `µ ∈ ℝ²ⁿ`, `N ≥ 1`, `t ≥ 0`, `χ = √(2n) + t`,
`ζ = 2χ/√N + 2χ²/N`, and `Y₁, …, Y_N` i.i.d. `N(µ, C)` with empirical mean `µ̂` and centered empirical
covariance `Ĉ`. Then with probability at least `1 − 3e^{−t²/2}` both
`(µ̂ − µ)ᵀ C⁻¹ (µ̂ − µ) ≤ χ²/N` and `(1 − ζ)C ⪯ Ĉ ⪯ (1 + ζ)C` hold.
**Formalization Note.** Same event as `eq_210`, with the deviation parameter `t` free and the failure
probability `3e^{−t²/2}`: one term for the norm of the whitened sample mean and two for the extreme singular
values of the whitened data matrix. Taking `t = √(2 log(3/δ))` gives probability at least `1 − δ`; with
`t = √(2 log(2/δ))` as in Eq. (209) this argument gives `1 − 3δ/2`. -/
theorem gaussian_sample_concentration {n : ℕ} (N : ℕ) (hN : 1 ≤ N) (t : ℝ) (ht : 0 ≤ t)
(μ : Fin n ⊕ Fin n → ℝ) (C : Matrix (Fin n ⊕ Fin n) (Fin n ⊕ Fin n) ℝ) (hC : C.PosDef) :
let χ : ℝ := Real.sqrt (2 * n) + t
let ζ : ℝ := 2 * χ / Real.sqrt N + 2 * χ ^ 2 / N
1 - ENNReal.ofReal (3 * Real.exp (-t ^ 2 / 2)) ≤
(Measure.pi fun _ : Fin N => multivariateGaussian (WithLp.toLp 2 μ) C)
{Y | let μhat : Fin n ⊕ Fin n → ℝ := fun a => (1 / (N : ℝ)) * ∑ j, WithLp.ofLp (Y j) a
let Chat : Matrix (Fin n ⊕ Fin n) (Fin n ⊕ Fin n) ℝ := Matrix.of fun a b =>
(1 / (N : ℝ)) * ∑ j, (WithLp.ofLp (Y j) a - μhat a) * (WithLp.ofLp (Y j) b - μhat b)
(μhat - μ) ⬝ᵥ (C⁻¹ *ᵥ (μhat - μ)) ≤ χ ^ 2 / N ∧
(Chat - (1 - ζ) • C).PosSemidef ∧ ((1 + ζ) • C - Chat).PosSemidef} := by sorry
end BosonicLearn