Proposition 2 — for every fixed transcript
ProvedBosonicLearn.proposition_2bosonic-gaussian-statesp2o-gran-per-chapterp2o-plan-paperp2o-strengthenedp2o-v1quantum-informationquantum-learning
Assume , let with , and fix . For a transcript of outcomes, the -th POVM element is a rank-one with in the -photon sector, and
There are absolute and such that for , and every such transcript
Combined with Proposition 1 it shows that the posterior does not concentrate unless .
Differences from the source:
- generalization: the rank-one POVM elements |ϕ_t⟩ of a transcript of an adaptive scheme → arbitrary nonzero vectors ϕ_t in the k_t-photon sectors, any N; reason: every transcript has this form after the reduction of §5.4 (pp. 34–35); the proof treats each copy's vector separately
- reading: Z ∼ Unif(B_{op,b}(0)) with b a small constant → every b ∈ (0, 0.1]; reason: the proof uses only ‖Φ_{X₀}(Z)‖_op ≤ 4 (Lemma 8, b ≤ 0.1)
- encoding: log E_Z L ≥ −Cξ² Σ(…) → E_Z L ≥ exp(−Cξ² Σ(…)), the expectation a lower Lebesgue integral; reason: equivalent; avoids log 0
- encoding: ξ "sufficiently small", ξ < 1/32, 1/n ≤ ν ≤ 1 (§5.1) → ∃ ξ₀ ∈ (0, 1/32), ∀ ξ ∈ (0, ξ₀], n ≥ 1, ν ∈ [1/n, 1]; reason: the standing assumptions of §5.1
Preamble
import Mathlib import Definitions.Def_BosonicLearn_PhiMap import Definitions.Def_BosonicLearn_Ggood import Definitions.Def_BosonicLearn_herm0Ball import Definitions.Def_BosonicLearn_fugacityCoeff import Definitions.Def_BosonicLearn_fugacityFamily open QAlgorithms.Papers.Independent.Claimed50.P47 (Occupation) open MeasureTheory ProbabilityTheory open scoped ENNReal
Formal statement
namespace BosonicLearn
/-- **Proposition 2 (Likelihood ratio bound for a fixed transcript).**
Chen, Anna Mele, Anna Mele, Preskill, *When are bosonic Gaussian states classical to learn?*, arXiv:2609.26705, Proposition 2, Eq. (90), p. 38: "Fix X_0 ∈ G_good. For any transcript of measurement outcomes x := (z, k)
where k_t denotes the measured photon number of the t-th copy, log E_Z L_{Φ_{X_0}(Z)}(x) ≥ −Cξ² Σ_{t=1}^N (nν² + k_t²/n),
where Z ∼ Unif(B_{op,b}(0)) and C is some positive constant."
Setting (§5.1, §5.4, pp. 30, 34–35): `1/n ≤ ν ≤ 1`, `R_X = r(I + ξX)`, the POVM elements are rank one and
block diagonal, `|ϕ_t⟩ ∈ H_{k_t}^{(n)}`, and `L_X(x) = ∏_t ⟨ϕ_t|ρ_{R_X}|ϕ_t⟩ / ⟨ϕ_t|ρ_{R_{X_0}}|ϕ_t⟩` (Eq. (63)).
There are absolute `C > 0` and `ξ₀ ∈ (0, 1/32)` such that for all `n ≥ 1`, `ν ∈ [1/n, 1]`, `ξ ∈ (0, ξ₀]`, `X₀ ∈ G_good`,
`b ∈ (0, 0.1]`, all `N`, photon numbers `k_1, …, k_N` and nonzero vectors `ϕ_t` in the `k_t`-photon sectors,
$$\mathbb E_{Z∼\mathrm{Unif}(B_{op,b}(0))} \prod_{t} \frac{⟨ϕ_t|ρ_{R_{Φ_{X_0}(Z)}}|ϕ_t⟩}{⟨ϕ_t|ρ_{R_{X_0}}|ϕ_t⟩} \ge \exp\Big(-Cξ^2 \sum_{t=1}^N (nν^2 + k_t^2/n)\Big).$$
**Formalization Note.** The log form is exponentiated (no `log 0` junk), and the expectation is a lower Lebesgue
integral. Stated for arbitrary sector vectors `ϕ_t`, which include the rank-one POVM elements of every transcript of
every adaptive scheme (see the Differences list). -/
theorem proposition_2 :
∃ C : ℝ, 0 < C ∧ ∃ ξ₀ : ℝ, 0 < ξ₀ ∧ ξ₀ < 1 / 32 ∧
∀ n : ℕ, 1 ≤ n → ∀ ν : ℝ, 1 / (n : ℝ) ≤ ν → ν ≤ 1 → ∀ ξ : ℝ, 0 < ξ → ξ ≤ ξ₀ →
∀ X₀ ∈ Ggood n, ∀ b : ℝ, 0 < b → b ≤ 1 / 10 →
∀ (N : ℕ) (k : Fin N → ℕ) (ϕ : Fin N → Occupation (Fin n) → ℂ),
(∀ t u, ∑ i, u i ≠ k t → ϕ t u = 0) → (∀ t, ϕ t ≠ 0) →
ENNReal.ofReal (Real.exp (-(C * ξ ^ 2 * ∑ t, ((n : ℝ) * ν ^ 2 + (k t : ℝ) ^ 2 / n)))) ≤
∫⁻ Z, ENNReal.ofReal (∏ t,
(∑' u, ∑' v, star (ϕ t u) *
fugacityCoeff (fugacityFamily ν ξ (PhiMap X₀ ξ Z)) u v * ϕ t v).re /
(∑' u, ∑' v, star (ϕ t u) * fugacityCoeff (fugacityFamily ν ξ X₀) u v * ϕ t v).re)
∂(cond (Measure.addHaar : Measure (Herm0 n)) (herm0Ball ⊤ 0 b)) := by sorry
end BosonicLearn
Source
Chen, Anna Mele, Anna Mele, Preskill, When are bosonic Gaussian states classical to learn?, arXiv:2609.26705, Proposition 2, Eq. (90), p. 38
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.