Bosonic Gaussian states describe laser light, thermal radiation, squeezed light and the modes of many continuous-variable quantum devices. They sit on the boundary between quantum and classical physics: a Gaussian state is fixed by a mean vector and a covariance matrix, exactly like a classical Gaussian random vector, yet some Gaussian states are entangled or squeezed. Chen, Anna Mele, Anna Mele and Preskill (arXiv:2609.26705) turn the question "when does a Gaussian state behave classically?" into a statistical one: how many copies of an unknown -mode Gaussian state does a learner need, if it may only measure the copies one at a time?
A classical -variate Gaussian is learned to total-variation error from samples. The paper shows that Gaussian states close to the vacuum cost copies under single-copy measurements, while states with enough thermal noise cost only the classical , and it determines the cost exactly in between.
Timeline. Bittel, Mele, Eisert and Mele (BMEM25, 2025) learn any -mode Gaussian state of energy at most with adaptive single-copy Gaussian measurements from copies. Chen et al. (CMF+26, 2026) show that the scaling is unavoidable for Gaussian (more generally, Wigner-positive) measurements, even for passive states. With entangled measurements the sample complexity of Gaussian state learning is (CFG+26, 2026). The present paper (2026) proves the single-copy lower bounds and the crossover theorem formalized here.
An -mode bosonic system has the Fock space with basis , . A state is a density operator: positive, trace class, trace one. The quadratures are and , collected in grouped order . A Gaussian state is a state with characteristic function . Here is the mean and the covariance; in this convention the vacuum has . A state is passive if it is Gaussian and block diagonal in the total photon number.
The promise with parameter is : every quadrature fluctuates at least above the vacuum level. Write for the class of all Gaussian states, with arbitrary mean, satisfying it.
The trace distance is . A learner succeeds on if its output has with probability at least .
An adaptive single-copy scheme on copies measures copy with a POVM (any measurement, Gaussian or not, with discrete or continuous outcomes) chosen as a function of the outcomes of copies , keeps no quantum memory, and outputs by a possibly randomized classical computation. Heterodyne detection is one such measurement: it projects onto coherent states and, on , returns a sample of .
The paper presents two main contributions: the cold-state lower bounds—refTheorem 4 for few-copy measurements and refTheorem 6 for non-adaptive measurements with known spectrum—and the warm-state crossover refTheorem 3. The platform permits one goal, so Theorem 3 is the goal; Theorems 4 and 6 are milestones of equal standing.
With , there are universal and such that for all , and :
Equivalently for , for and for . The constants are not fixed numerically; only the shape of the answer is claimed.
The upper bound: the heterodyne law Eq. (207); Gaussian concentration Eq. (210); Theorem 7 (a trace-distance bound with an -independent prefactor); Lemma 18; Theorem 8 ( copies); the bound of [BMEM25, Theorem 10]; Corollary 19. The lower bound: Lemma 2 (fugacity parametrization of passive states); Lemma 8; Propositions 1, 2 and 5; Lemma 12; Theorem 5 ( for ); Theorem 9 ( even for collective measurements).
Section 4. refTheorem 4 shows that possibly adaptive protocols using at most -copy measurements require copies, even for passive states with .
Section 6. refTheorem 6 gives an lower bound for non-adaptive single-copy learning of passive states with , even when their symplectic eigenvalues are known.
The milestones of Sections 4 and 6 were added on 2026-11-16 so that the mission covers every theorem of the paper.
Eq. (207), Eq. (210) and the bound are imported by the paper from [BMEM25] and not proved there; they are milestones so that the goal is unconditional. Theorem 1 combines the case of Theorem 4 with Theorem 6's non-adaptive, known-spectrum bound, while Theorem 2 is the introduction statement of Theorem 4.
The theorem locates the quantum-to-classical crossover for a learning task: a vacuum gap makes Gaussian states as easy to learn as classical Gaussians with simple heterodyne detection, while states within of the vacuum force a factor more copies on every non-entangled strategy. Entangled measurements reach in all regimes (CFG+26), so the cold-state gap is a separation between single-copy and collective learning. Theorem 9 shows the warm-state rate is optimal even for collective measurements.
None of these results is machine-checked. A formalization adds the platform's first sample-complexity theorems for continuous-variable states and a reusable layer: POVMs with measurable outcome spaces on Fock space, adaptive schemes and their transcript laws, collective schemes, the heterodyne law, the Fock-basis formula for , and Lebesgue measure and Schatten-norm balls on traceless Hermitian matrices (the Schatten norm itself is reused from the platform).
The upper bound needs a trace-distance estimate whose constant depends only on the vacuum gap. The usual two-sided perturbation bounds for Gaussian states lose a factor and give the wrong rate; Theorem 7 needs the one-sided order and relative entropy.
The lower bound must hold against adaptive, non-Gaussian measurements with continuous outcomes on an infinite-dimensional space. Reducing to finite-dimensional tomography only gives the cold-state rate, and a packing (Fano) argument only gives . The intermediate regime needs a posterior anti-concentration argument in which the state is not a linear function of its parameter and the likelihood ratios must be bounded in every photon-number sector.
States are coefficient arrays on Occupation (Fin n) = , with the published density, trace-norm and Borel structures of Dependency50.P33. Gaussianity is the published Support50.P14.GaussianAt, bridged to the paper's grouped quadratures and vacuum (P14 uses interleaved order and ). POVMs carry their Born law as a measure on an arbitrary measurable outcome space. Single-copy schemes are sequences of Markov kernels, and the estimator is a Markov kernel into states. In every statement the protocol is fixed before the state and constants are universal.
A trivializing formalization is ruled out: the lower bound quantifies over all outcome spaces, so heterodyne is among the schemes; the heterodyne law is defined from the Husimi density, and Eq. (207) pins its normalization; the class contains all means, not only passive states.
Welcome contributions: proofs of the imported milestones (Gaussian concentration, the heterodyne law), the fugacity representation, and general facts about Gaussian states in this Fock-space model.
namespace BosonicLearn
/-- **Theorem 3 (Tight transition of learnability).**
Chen, Anna Mele, Anna Mele, Preskill, *When are bosonic Gaussian states classical to learn?*, arXiv:2609.26705, Theorem 3, Eqs. (6)–(7), pp. 6–7: "There exist universal constants ε_0 > 0 and n_0 ∈ ℕ such that the following
holds for all n ≥ n_0 and 0 < ε < ε_0. Under the promise that an n-mode Gaussian state ρ(µ, Σ) satisfies
Σ ≥ (1/2 + ν)I for some ν > 0 that can depend on n, the necessary and sufficient number of copies to learn ρ to ε trace
distance with probability at least 2/3 using single-copy measurements is given by N = Θ((n²/ε²) min{n, 1 + ν^{−1}}).
… This bound can be achieved by heterodyne measurements for all values of ν > 0."
With `f(n, ε, ν) = (n²/ε²)·min{n, 1 + 1/ν}` there are universal `ε₀ > 0`, `n₀`, `c > 0`, `C > 0` such that for all
`n ≥ n₀`, `ε ∈ (0, ε₀)` and `ν > 0`:
1. (sufficiency, heterodyne) some non-adaptive protocol of `N ≤ C·f` independent heterodyne measurements and a
measurable estimator learns every Gaussian state with `Σ ⪰ (½ + ν)I` to trace distance `ε` with probability `≥ 2/3`;
2. (necessity) every adaptive single-copy scheme on `N` copies, with any measurable outcome space and any randomized
estimator, that learns every such state to trace distance `ε` with probability `≥ 2/3` has `N ≥ c·f`.
**Formalization Note.** `Θ` is read with universal constants placed, like `ε₀, n₀`, before `n, ε, ν`. The protocol is
chosen before the state. Eq. (7) restates (6) and is not a separate claim. -/
theorem theorem_3 :
∃ ε₀ : ℝ, 0 < ε₀ ∧ ∃ n₀ : ℕ, ∃ c : ℝ, 0 < c ∧ ∃ C : ℝ, 0 < C ∧
∀ n : ℕ, n₀ ≤ n → ∀ ε : ℝ, 0 < ε → ε < ε₀ → ∀ ν : ℝ, 0 < ν →
(∃ N : ℕ, (N : ℝ) ≤ C * ((n : ℝ) ^ 2 / ε ^ 2 * min (n : ℝ) (1 + 1 / ν)) ∧
∃ est : (Fin N → EuclideanSpace ℝ (Fin n ⊕ Fin n)) → State (Occupation (Fin n)),
Measurable est ∧
∀ ρ ∈ promiseClass n ν, ENNReal.ofReal (2 / 3) ≤ heterodyneSuccess N est ρ ε) ∧
(∀ (N : ℕ) (O : Type) [MeasurableSpace O] (S : AdaptiveScheme n N O)
(est : Kernel (Fin N → O) (State (Occupation (Fin n)))), IsMarkovKernel est →
(∀ ρ ∈ promiseClass n ν, ENNReal.ofReal (2 / 3) ≤ successProb S est ρ.coefficients ε) →
c * ((n : ℝ) ^ 2 / ε ^ 2 * min (n : ℝ) (1 + 1 / ν)) ≤ N) := by sorry
end BosonicLearn
Write . There are universal constants , and such that for all , all and all (which may depend on ), for the class of -mode Gaussian states with and arbitrary mean:
Equivalently, for , for , and for : a quantum-to-classical crossover in learnability between cold states (near the vacuum) and warm states (classical sample complexity).
Differences from the source:
No open leaves. Every sub-goal is proved or awaiting decomposition.