A fully Gaussian quantum Stein's lemma

A central yet mysterious problem in quantum information is the asymptotic distinguishability of quantum states under restricted measurements, and this paper investigates how Gaussian restrictions affect hypothesis testing for multimode bosonic states. The main finding establishes a single-letter formula for the Gaussian Stein exponent by separating its contribution into first and second moments, revealing that collective Gaussian measurements offer no asymptotic advantage when considering centred Gaussian states under specific conditions.

The Gist

The asymptotic error exponent is characterized by the regularised Gaussian-measured relative entropy, and this quantity admits a sharp structural decomposition where the dependence on the first moments is intrinsically single-letter and can be optimized independently of the covariance contribution.

Asymptotic Characterization of the Stein Exponent

The Gaussian Stein exponent, defined as

SteinG(rhoG∥sigmaG) = limgamma→0+ liminf n→∞ −1/n log betan,epsilonrho⊗n Gsigma⊗n G (1), is exactly characterized by the regularisation of the corresponding restricted measured relative entropy:

SteinG(rhoG∥sigmaG) = limn→∞ 1/n D Grho⊗n Gsigma⊗n G (3).

Decoupling of First and Second Moments

The paper establishes a sharp structural decomposition for the Gaussian Stein exponent, showing that the dependence on the first moments can be solved exactly:

SteinG(rhoG∥sigmaG) = SteinG(rho0 G∥sigma0 G) + (s − t)⊤W−1(s − t), where rho0 G and sigma0 G are states with zero first moments, and W is the covariance matrix (4). This demonstrates that the displacement information is intrinsically single-letter and can be optimized independently.

Conditions for Additivity of Covariance Contribution

The core difficulty lies in the additivity problem concerning the second-moment term, which is regularised. The paper identifies sufficient conditions under which this covariance contribution becomes additive over the number of copies:

Theorem 2 states that if the reference covariance matrix W has a flat symplectic spectrum (there exists a symplectic matrix S such that S W S−1 = w12m) and [V, e Ω V e Ω−1] = 0, then D G V⊗n G W⊗n G = n D G V G W G (8).

Conclusion on Collective Measurements

When the conditions of Theorem 2 are satisfied, the regularisation can be removed and the Gaussian Stein exponent simplifies to a single-letter optimisation over a single-copy Gaussian measurement, along with the explicit displacement contribution:

SteinG(rhoG∥sigmaG) = supgamma≥iΩ 1/2 log det(W + gamma)det(V + gamma) + 1/2 Tr(W + gamma−1)(V − W) (9). For centred states belonging to this class, collective Gaussian measurements provide no asymptotic advantage over independent single-copy measurements.

Operational Interpretation

The separation of moments has a direct operational interpretation: the contribution of the first moments can be isolated asymptotically by a passive Gaussian mixing of copies that concentrates the displacement difference into a single output copy, leaving covariance matrices unchanged. The remaining copies are then devoted to covariance discrimination, which in additive cases can be extracted using only single-copy Gaussian measurements.

Single-Mode Example

For two centred, single-mode Gaussian states with the same covariance matrices but different first moments (coherent states), the problem reduces to a single-copy second moment optimisation under Theorem 7. This leads to a closed form expression for SteinG(rhoG∥sigmaG) involving the displacement difference and inverse covariance matrix:

SteinG(rhoG∥sigmaG) = (s − t)⊤W−1(s − t), where W is the common covariance matrix (38).

Pure Multimode States

The conditions of Theorem 7 are satisfied for all pure multimode Gaussian states, as a symplectic transformation can map the reference covariance matrix to a form where [V, e Ω V e Ω−1] = 0, thereby ensuring additivity and the removal of regularisation. This confirms that collective Gaussian measurements are unnecessary for pure multimode pairs.

Final Characterization

Combining Theorem 1 and Theorem 7 yields the final single-letter characterisation for arbitrary Gaussian states:

SteinG(rhoG∥sigmaG) = D G(V∥W) + (s − t)⊤W−1(s − t), where D G(V∥W) is the single-letter optimisation over a single-copy measurement of the covariance matrices (91). This formulation holds for all pairs satisfying Theorem 7.

Based on the provided scientific paper, here are specific improvements for AI systems that leverage these theoretical results:

The paper establishes a single-letter formula for the Gaussian Stein exponent, which characterizes the asymptotic distinguishability rate of quantum states under restricted measurements (Gaussian measurements). This framework can be used to design more efficient and robust quantum-inspired AI systems.

Here are the specific improvements:

1. Improvements to Quantum State Discrimination and Classification:

2. Improved Robustness in Sensor Data Processing:

3. Enhanced Resource-Efficient Quantum Machine Learning Algorithms (QML):

1. Improvements to Quantum State Discrimination and Classification:

The paper's core result is the characterization of the Gaussian Stein exponent, which provides a sharp asymptotic bound on how well one can distinguish between two quantum states using only Gaussian measurements.

An improved AI system could perform:

• A real-time classification of complex quantum states (e.g., those arising from noisy experimental data) by minimizing the Gaussian Stein distance rather than relying on general, computationally expensive quantum state tomography.

• The system can distinguish between two closely related Gaussian distributions (like thermal vs. vacuum states) with high asymptotic precision, even when only limited Gaussian measurements are available. This is crucial for tasks like quantum sensing where experimental constraints limit the measurement apparatus to Gaussian observables (e.g., homodyne detection).

• It allows for the construction of optimal measurement strategies that maximize distinguishability, directly improving the performance of quantum classifiers in noisy environments by accounting for measurement limitations theoretically derived from Gaussian constraints.

2. Improved Robustness in Sensor Data Processing:

The paper demonstrates a structural decomposition where the overall error exponent separates into a single-letter term related to first moments (displacement) and a covariance-only term related to second moments (noise/correlations).

An improved AI system could perform:

• Noise characterization and compensation in quantum sensors. By isolating the first moment contribution, the system can precisely account for classical phase or amplitude shifts (displacement) in sensor readings, allowing it to focus its computational effort on discriminating between subtle differences in noise, squeezing, and correlations encoded in the covariance matrices.

• Adaptive measurement strategies for data streams where measurement constraints are Gaussian. The system can dynamically decide whether to prioritize optimizing for displacement information or noise discrimination based on the specific data being processed, leading to more efficient processing pipelines compared to using a single, fixed collective measurement strategy.

3. Enhanced Resource-Efficient Quantum Machine Learning Algorithms (QML):

The paper provides sufficient conditions (Theorem 7) under which the many-copy optimization problem simplifies into a single-copy optimization over Gaussian seeds, effectively removing the need for complex collective measurements across many copies.

An improved AI system could perform:

• Reduced complexity in large-scale QML tasks. For problems requiring discrimination of many copies (e.g., estimating parameters from noisy entangled states), the system can leverage Theorem 7 to simplify the required measurement hardware and strategy, reducing the complexity from a collective multi-mode measurement problem to a single, local Gaussian measurement per copy.

• Design of resource-efficient QML circuits. Since the optimal strategy for centered states is known (by decoupling moments), researchers can design quantum circuits that only need to be optimized for covariance matrices, leading to smaller, more practical quantum algorithms that achieve near-optimal performance under Gaussian constraints.