One-sided relative covariance bound under a Loewner gap: for ,
ProvedBosonicLearn.cov_perturbation_gapTheorem 7 in the form used in the proof of Theorem 8, proved without relative entropy.
Let and be -mode Gaussian states with the same mean, for some , and . With and ,
Comparison with BosonicLearn.theorem_7: the hypothesis on the symplectic eigenvalues () is replaced by the Loewner inequality , which implies it with and is the promise of Theorems 3 and 8; the mean is arbitrary but common; and the constant is replaced by . The scaling is the same. (The proof gives when .)
Proof idea. Both states are Gaussian mixtures of displaced thermal states of covariance . The extra noise is split as , : the first part is absorbed by a Cameron–Martin weight, the second displaces . The trace norm is bounded by duality (BosonicLearn.traceSize_le_of_pairing) and a Cauchy–Schwarz inequality weighted by ; the kernel identity BosonicLearn.thermal_displaced_kernel turns the resulting quantity into (BosonicLearn.gaussian_bilinear_exp_bound).
Formalization Note is (S⁻¹ * (S' - S) * S⁻¹ * (S' - S)).trace; hypotheses (S - (1/2 + ν) • 1).PosSemidef, (S' - S).PosSemidef.
import Mathlib import Definitions.Def_BosonicLearn_traceDist import Definitions.Def_BosonicLearn_IsGaussianState open QAlgorithms.Papers.Independent.Dependency50.P33 (State traceSize) open QAlgorithms.Papers.Independent.Claimed50.P47 (Occupation) open MeasureTheory open scoped ENNReal Matrix
namespace BosonicLearn
/-- **One-sided relative covariance bound under a Loewner gap: `½‖ρ(µ, Σ′) − ρ(µ, Σ)‖₁ ≤ √2 K_ν Δ_rel`.**
Chen, Anna Mele, Anna Mele, Preskill, *When are bosonic Gaussian states classical to learn?*, arXiv:2609.26705, Theorem 7, Eqs. (184)–(185), p. 55: "Let Σ̄ be an n-mode covariance matrix whose symplectic eigenvalues are all at
least ν₋ > 1/2, and let Σ′ be another covariance matrix satisfying Σ′ ⪰ Σ̄. … Then
½‖ρ(0, Σ′) − ρ(0, Σ̄)‖₁ ≤ (1/(2√(2γ₋))) Δ_rel", in the form used in the proof of Theorem 8 (p. 61), where
`Σ̄ ⪰ (½ + ν)I`.
Let `ρ(µ, Σ)` and `ρ(µ, Σ′)` be Gaussian states with the same mean, `Σ ⪰ (½ + ν)I` for some `ν > 0`, and `Σ′ ⪰ Σ`.
With `Δ_rel = ‖Σ^{−1/2}(Σ′ − Σ)Σ^{−1/2}‖_F` and `K_ν = (½ + ν)/√(ν(1 + ν))`,
`½‖ρ(µ, Σ′) − ρ(µ, Σ)‖₁ ≤ √2 · K_ν · Δ_rel`.
**Formalization Note.** This is Theorem 7 with the hypothesis on the symplectic eigenvalues replaced by the Loewner
inequality `Σ ⪰ (½ + ν)I` (which implies it, and is the promise of Theorems 3 and 8), a common mean `µ`, and the
constant `1/(2√(2γ₋)) = ¼√(2 + 1/ν)` replaced by `√2 K_ν ≤ 4 · ¼√(2 + 1/ν)`; the scaling `√(1 + 1/ν)` is the same.
`Δ_rel² = Tr(Σ⁻¹(Σ′ − Σ)Σ⁻¹(Σ′ − Σ))`. It is proved without relative entropy: both states are Gaussian mixtures of
displaced thermal states of covariance `(½ + ν)I`, and the trace norm is bounded by a `χ²`-divergence relative to
that thermal state (`thermal_displaced_kernel`, `gaussian_bilinear_exp_bound`, `traceSize_le_of_pairing`). -/
theorem cov_perturbation_gap {n : ℕ} (ρ ρ' : State (Occupation (Fin n)))
(μ : Fin n ⊕ Fin n → ℝ) (S S' : Matrix (Fin n ⊕ Fin n) (Fin n ⊕ Fin n) ℝ) (ν : ℝ) (hν : 0 < ν)
(hS : (S - (1 / 2 + ν) • (1 : Matrix (Fin n ⊕ Fin n) (Fin n ⊕ Fin n) ℝ)).PosSemidef)
(hord : (S' - S).PosSemidef)
(h : IsGaussianState ρ μ S) (h' : IsGaussianState ρ' μ S') :
traceDist ρ'.coefficients ρ.coefficients ≤
ENNReal.ofReal (Real.sqrt 2 * ((1 / 2 + ν) / Real.sqrt (ν * (1 + ν))) *
Real.sqrt (S⁻¹ * (S' - S) * S⁻¹ * (S' - S)).trace) := by sorry
end BosonicLearn