Machine-Learning-Enhanced Entanglement Detection Under Noisy Quantum Measurements

TL;DR

The paper tackles entanglement detection under noisy quantum measurements by learning a robust optimal entanglement witness (ROEW) via a distributionally robust SVM trained on Pauli-measurement features. By formulating a distributionally robust optimization framework with moment-based ambiguity sets, it handles unknown measurement noise while maintaining data efficiency. The authors validate the approach on Werner and Bell-diagonal state ensembles, showing high accuracy with limited labeled data and resilience to noise exceeding 10%. ROC analyses and confusion matrices indicate reliable discrimination between entangled and separable states, with near-perfect AUC in tested regimes. This work provides a practical tool for noise-robust quantum characterization, reducing measurement overhead and supporting scalable entanglement-based technologies.

Abstract

Quantum measurements are inherently noisy, hindering reliable entanglement detection and limiting the scalability of quantum technologies. While error mitigation and correction strategies exist, they often impose prohibitive resource overheads. Here, we introduce a machine-learning-based approach to achieve noise-resilient entanglement classification even with imperfect measurements. Using support vector machines (SVMs) trained on features extracted from Pauli measurements, we develop a robust optimal entanglement witness (ROEW) that remains effective under unknown measurement noise. By optimizing SVM parameters against worst-case errors, our protocol significantly outperforms conventional methods in classification accuracy. Numerical experiments demonstrate that ROEW achieves high-fidelity entanglement detection with minimal measurements, even when measurement errors exceed 10\%. This work bridges machine learning and quantum information science, offering a practical tool for noise-robust quantum characterization and advancing the feasibility of entanglement-based technologies in real-world settings.

Figures (9)

• Figure 1: Visualization of a two-dimensional dataset with two classes separated by a hyperplane. Blue circles represent separable states, while red circles denote entangled states. The decision boundary $\mathbf{v}^T\mathbf{x} + b = 0$ (solid line) and margins $\mathbf{v}^T\mathbf{x} + b = \pm 1$ (dashed lines) margin width $2/\|\mathbf{v}\|$ . Support vectors are highlighted in green.
• Figure 2: This 3D plot illustrates a robust chance-constrained SVM classifier applied to two-qubit Werner density matrices for distinguishing entangled from separable quantum states. The axes $c_1$, $c_2$ and $c_3$ represent features derived from the density matrix. The purple region contains separable states (shown as light purple dots), while the red squares represent entangled states. A dark blue plane labelled "EW" depicts the entanglement witness, which acts as the decision boundary learned by the robust SVM classifier. The black cross marked $\rho^{\textbf{x}}$ is a test quantum state, whose position relative to the separating plane indicates its classification. This visualization demonstrates the ability of the robust SVM to accurately separate quantum states under uncertainty.
• Figure 3: Distribution of F1 scores showing model performance across quantum state classifications in different noisy conditions rate .
• Figure 4: Confusion matrix for $\alpha = 0.9$.
• Figure 5: TP,TN, FP, FN counts by percentage of data.