Theorem 5.1 — Gaussian measurements reach infidelity from copies (corrected: max(1, log log κ) for log log‖V⁻¹‖)
OpenAgnGaussTomo.Supplement.theorem_5_1There are universal constants and such that for every number of modes , every , and there is an adaptive scheme that uses only Gaussian measurements, on
copies, with an estimator that always outputs a pure Gaussian state , such that the following holds. For every -mode state (a density operator on the Fock space, mixed allowed, of any energy), every pure Gaussian state with covariance , , and every with , with probability at least ,
Here is the infidelity, , the covariance is normalized so that the vacuum has , and a Gaussian measurement is a POVM with a mean-zero Gaussian seed state . The scheme measures the copies one at a time, each with a Gaussian measurement chosen from the earlier outcomes, and it is given only , , and .
This is the positive half of the paper's separation between Gaussian and non-Gaussian measurements: Gaussian measurements alone improve the warm start to the semi-agnostic guarantee in infidelity, while Theorem 5.6 shows they cannot reach efficiently.
Differences from the source:
- correction: log log ‖V^{-1}‖_op in the copy count → L(κ) = max(1, log log κ) for an upper bound κ ≥ ‖V^{-1}‖_op; reason: log log ‖V^{-1}‖_op ≤ 0 for ‖V^{-1}‖_op ≤ e (and −∞ at 1), so the printed count allows zero copies; but for one mode and ρ one of the coherent states |α⟩, |−α⟩ with |α| = 1 (V = I, opt = 0) a zero-copy algorithm outputs a fixed pure state, whose infidelity with one of them is at least 0.24, since d_tr(|α⟩, |−α⟩) = √(1 − e^{−4}) > 0.98 and d_tr = √d_IF for pure states, contradicting the guarantee for ε < 0.24; L agrees with log log once κ ≥ e^e. This is the correction of the published Theorems 3.4, 3.6 and 3.8
- reading: the algorithm's resources depend on ‖V^{-1}‖_op → the algorithm is given an upper bound κ ≥ 1 on ‖V^{-1}‖_op and must work for every state with ‖V^{-1}‖_op ≤ κ; reason: the algorithm runs log log ‖V^{-1}‖_op unsqueezing rounds (p. 15), so this quantity must be available to it; an algorithm that knows only an upper bound is at least as strong
- reading: Õ(n² ε^{-1} log log ‖V^{-1}‖_op log⁵(1/δ)) copies → N ≤ C·n²·ε^{-1}·L(κ)·log⁵(1/δ)·log(e + n + 1/ε + 1/δ + L(κ))^C; reason: Õ hides polylogarithmic factors of the parameters (κ enters only through L(κ)); one universal C; ε^{-1} multiplies every factor, as printed
- reading: O(opt) + ε → C·opt + ε with the same universal constant C; reason: O hides an absolute constant; one constant for the copy count and the error is equivalent to two
- reading: opt ≤ opt̄ → opt̄ is a universal constant in (0, 1), chosen with C before n and every other parameter; opt is any number with d_IF(ρ, ψ) ≤ opt ≤ opt̄; reason: Theorem 3.8 and Theorem 1.1 call opt̄ a universal constant in (0, 1); Theorem 5.1 uses the same symbol
- reading: given ε > 0 and δ ∈ (0, 1/2), there exists an algorithm → the scheme and estimator depend on n, ε, δ and κ only, not on ρ, ψ, V or opt; reason: opt does not enter the copy count; an algorithm that does not know it is at least as strong
- reading: uses only Gaussian measurements → every round's POVM is a Gaussian measurement (seed: a mean-zero Gaussian state) whose outcome also carries an independent uniform random number in [0, 1]; reason: the paper's algorithms are randomized (Proposition 2.9, Corollary 5.5), and classical randomness carries no quantum information; this is the published IsGaussianScheme of Theorems 3.4 and 3.6
- generalization: there exists an algorithm → an adaptive single-copy scheme of Gaussian measurements on a number N of copies fixed in advance, with an estimator that is a function of the transcript; reason: this is one kind of algorithm using only Gaussian measurements, so the existence claim specializes to the source's
- encoding: with probability at least 1 − δ → an inner probability: a measurable set of transcripts of probability ≥ 1 − δ on which the guarantee holds; reason: the guarantee need not define a measurable event; same meaning for measurable events
- encoding: covariance V in the order (q₁, …, qₙ, p₁, …, pₙ) → the interleaved order (q₁, p₁, …, qₙ, pₙ) of the published GaussianAt; reason: a relabelling of the same quadratures; ‖V^{-1}‖_op is unchanged
import Mathlib import Definitions.Def_OptGaussTomo_boseGauss import Definitions.Def_AgnGaussTomo_Succeeds import Definitions.Def_AgnGaussTomo_IsGaussianPOVM import Definitions.Def_AgnGaussTomo_infid import Definitions.Def_AgnGaussTomo_LL import Definitions.Def_AgnGaussTomo_opNorm open QAlgorithms.Papers.Independent.Dependency50.P33 (State) open QAlgorithms.Papers.Independent.Claimed50.P47 (Occupation) open QAlgorithms.Papers.Independent.Support50.P14 (GaussianAt InterleavedQuadrature OscillatorCovariance)
namespace AgnGaussTomo.Supplement
/-- **Theorem 5.1: Gaussian measurements refine to `O(opt) + ε` in infidelity, with the corrected
`max(1, log log κ)`.** Arulandu, Chen, Chen, Li, Ma, *Finding Gaussian Structure in Bosonic States*,
arXiv:2610.06810 (version of 3 Nov 2026), Theorem 5.1, printed p. 49 (PDF p. 52): "Let ρ be an n-mode
state where d_IF(ρ, |ψ⟩⟨ψ|) ≤ opt ≤ opt̄ for pure Gaussian state |ψ⟩ with covariance V. Then, given
ε > 0 and δ ∈ (0, 1/2), there exists an algorithm takes Õ(n² ε^{-1} log log ‖V^{-1}‖_op log⁵(1/δ))
copies of ρ and uses only Gaussian measurements to output a pure Gaussian state |ψ̂⟩ satisfying
d_IF(ρ, |ψ̂⟩⟨ψ̂|) ≤ O(opt) + ε with probability at least 1 − δ."
**Formalization Note.**
1. Correction (as in the published Theorems 3.4, 3.6, 3.8): `log log ‖V^{-1}‖_op` → `LL κ =
max(1, log log κ)`; the printed factor is `≤ 0` (or `−∞`) for `‖V^{-1}‖_op ≤ e`, which would allow
zero copies, and zero copies cannot learn the coherent states `|±α⟩`, `|α| = 1`.
2. The algorithm is an adaptive scheme on a number `N` of copies fixed in advance, every round a
Gaussian measurement (`IsGaussianScheme`, seed form of p. 10, with an independent uniform random
number for classical randomness), and a deterministic estimator of the transcript. It is given
`n, ε, δ` and an upper bound `κ ≥ ‖V^{-1}‖_op`, and never sees `ρ`, `ψ`, `V` or `opt`.
3. `Õ(X)` in the copy count is `C·X·log(e + n + 1/ε + 1/δ + LL κ)^C`: one product, `ε^{-1}` multiplies
every factor. `O(opt)` in the error is `C·opt` with the same universal `C`; `opt̄ ∈ (0, 1)` is universal.
Both are chosen before every other parameter.
4. Every output is a pure Gaussian state (`OptGaussTomo.bosePure`); `d_IF(ρ, |ψ⟩⟨ψ|) = 1 − ⟨ψ|ρ|ψ⟩`
(`infid`, pure second argument). `opt` is any number with `d_IF(ρ, ψ) ≤ opt ≤ opt̄`. -/
theorem theorem_5_1 : ∃ optBar : ℝ, 0 < optBar ∧ optBar < 1 ∧ ∃ C : ℝ, 0 < C ∧
∀ (n : ℕ) (ε δ κ : ℝ), 0 < ε → 0 < δ → δ < 1 / 2 → 1 ≤ κ →
∃ (N : ℕ) (S : BosonicLearn.AdaptiveScheme n N ((Fin n ⊕ Fin n → ℝ) × ℝ))
(est : (Fin N → (Fin n ⊕ Fin n → ℝ) × ℝ) → State (Occupation (Fin n))),
(N : ℝ) ≤ C * (n : ℝ) ^ 2 / ε * LL κ * Real.log (1 / δ) ^ 5 *
Real.log (Real.exp 1 + n + 1 / ε + 1 / δ + LL κ) ^ C ∧
IsGaussianScheme S ∧
(∀ x, est x ∈ OptGaussTomo.bosePure n) ∧
∀ (ρ ψ : State (Occupation (Fin n))) (m : InterleavedQuadrature (Fin n) → ℝ)
(V : OscillatorCovariance (Fin n)) (opt : ℝ),
ψ ∈ OptGaussTomo.bosePure n → GaussianAt ψ.coefficients m V → opNorm V⁻¹ ≤ κ →
infid ρ.coefficients ψ.coefficients ≤ opt → opt ≤ optBar →
Succeeds S ρ.coefficients
(fun x => infid ρ.coefficients (est x).coefficients ≤ C * opt + ε) δ := by sorry
end AgnGaussTomo.Supplement