Two Gaussian measures , have a common part of mass
ProvedBosonicLearn.gaussian_shift_overlapbosonic-gaussian-statesquantum-informationquantum-learning
A total-variation bound for two Gaussian measures with the same covariance.
Let be a positive semidefinite real matrix, and . There is a measure on with
Equivalently, the total variation distance of the two Gaussians is at most (the exact value is ). For invertible and a shift , .
Role. Classical counterpart of Lemma 18 of the paper: it controls the part of a mean shift that can be absorbed in the mixing measure of a classical Gaussian state (BosonicLearn.lemma_18_classical).
Formalization Note The shift is taken in the range of , so no inverse is needed and singular is allowed. The measures are Mathlib's multivariateGaussian on EuclideanSpace ℝ ι; the statement is purely classical probability.
Preamble
import Mathlib open MeasureTheory ProbabilityTheory open scoped ENNReal Matrix
Formal statement
namespace BosonicLearn
/-- **Overlap of two Gaussian measures with the same covariance.**
Chen, Anna Mele, Anna Mele, Preskill, *When are bosonic Gaussian states classical to learn?*, arXiv:2609.26705, Lemma 18 (first-moment perturbation), Eq. (204), p. 60 (classical counterpart: two Gaussians with equal covariance and
shifted means are close in total variation).
Let `T` be a positive semidefinite real matrix, `μ` a vector and `w` a vector. The Gaussian measures `N(μ, T)` and
`N(μ + Tw, T)` have a common part of mass at least `1 − √(wᵀTw)/√(2π)`: there is a measure `κ` with `κ ≤ N(μ, T)`,
`κ ≤ N(μ + Tw, T)` and `κ(ℝᵈ) ≥ 1 − √(wᵀTw)/√(2π)`. Equivalently their total variation distance is at most
`√(wᵀTw)/√(2π)`.
**Formalization Note.** The shift is taken in the range of `T` (`Tw`), so that no inverse of `T` is needed;
for invertible `T` and shift `δ`, `wᵀTw = δᵀT⁻¹δ`. The measures are Mathlib's `multivariateGaussian`. -/
theorem gaussian_shift_overlap {ι : Type} [Fintype ι] [DecidableEq ι] (μ : EuclideanSpace ℝ ι)
(T : Matrix ι ι ℝ) (hT : T.PosSemidef) (w : ι → ℝ) :
∃ κ : Measure (EuclideanSpace ℝ ι),
κ ≤ multivariateGaussian μ T ∧
κ ≤ multivariateGaussian (μ + WithLp.toLp 2 (T *ᵥ w)) T ∧
1 - Real.sqrt (w ⬝ᵥ (T *ᵥ w)) / Real.sqrt (2 * Real.pi) ≤ (κ Set.univ).toReal := by sorry
end BosonicLearn
Source
Chen, Anna Mele, Anna Mele, Preskill, When are bosonic Gaussian states classical to learn?, arXiv:2609.26705, Lemma 18 (first-moment perturbation), Eq. (204), p. 60