ResearchPod Summary
Continuous-variable (CV) quantum state tomography is traditionally limited by the need to represent states in a truncated Fock basis or on a fixed phase-space grid. As the number of modes increases, these methods suffer from an exponential growth in computational cost and difficulty in resolving fine non-Gaussian structures, such as interference fringes in Wigner functions. The authors seek a more scalable, measurement-efficient approach that can represent these states directly in continuous phase space.
QST-Flow utilizes normalizing flows—invertible neural networks that map simple probability distributions to complex target densities—to represent quantum states. The framework consists of two primary variants:
Both variants are trained using importance-sampled losses, allowing the model to focus on relevant regions of phase space without requiring a predefined grid.
The authors demonstrate that QST-Flow accurately reconstructs a variety of non-Gaussian states, including cat, binomial, and GKP states. In benchmarks against the state-of-the-art QST-CGAN, QST-WFlow achieved lower reconstruction error on noisy Wigner data. Furthermore, the framework successfully extends to multimode systems, where it avoids the memory and computational bottlenecks associated with tensor-product grid representations. By treating tomography as a generative modeling problem, the authors provide a flexible, scalable tool for characterizing nonclassical bosonic systems.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.