arXiv:2609.19534
Normalizing Flow-Based Bayesian Parameter Estimation for Noisy Quantum States
Hsien-Yi Hsieh, Yuan-Ting Liu, Juan Camilo Rodrıguez, Yang-Yi Lee, Po-Hang Wang, Ole Steuernagel, Chien-Ming Wu, Ray-Kuang Lee
Abstract
To extract complete information about experimental noisy quantum states, as well as correlations among different physical parameters, we develop an efficient machine learning-assisted framework with the normalizing flow based Bayesian parameter estimation (BPE). By embedding a physically interpretable parameter vector, ${s}$, into the quantum state ansatzes, $ρ({s})$, the corresponding parameter distribution $p({s}|D)$ is induced from BPE, based on the available experimental data $D$. The BPE allows us to perform parameter correlation analyses, providing powerful insights about how experimental parameters and related noise sources affect quantum features of the system. As an example, the flexibility of our framework is demonstrated using two different physical ansatzes on the optical cat state experiments, where a noisy single-photon state is added to impure squeezed state. With correlations among the Wigner function negativity, photon-addition fraction, noisy fraction, and pump powers, our framework gives interpretations for complicated experimental mechanisms in generating noisy non-Gaussian states. By leveraging normalizing flow-based neural networks and Bayesian uncertainty estimation, along with physical and experimental constraints incorporated naturally, the valuable information inferred in the posterior learning provides guidelines for experimentalists, as well as theorists, in identifying promising parameters to enhance desirable features.
AI Summary
The paper addresses inference of physically interpretable preparation and noise parameters from homodyne measurements, including their uncertainties and correlations, rather than reconstructing only a best-fit density matrix. It approximates the Bayesian posterior over parameters of a quantum-state ansatz using neural-spline normalizing flows trained by minimizing $D_{\mathrm{KL}}(q\Vert p)$. Analytic constraint mappings enforce bounded domains and normalized mixture weights. A 20% calibration subset sets empirical prior centers; the remaining 80% supplies inference data. The demonstration analyzes photon-added squeezed-state measurements across six OPO pump settings (15–40 mW) and thirteen stimulated-parametric-down-conversion pump settings (0.25–10 mW), with roughly 14,000–80,000 quadrature samples per setting. Two ansatzes combine squeezed, squeezed-thermal, and thermal components with a photon-added fraction $\eta$; the second adds branch-dependent displacements. Agreement is assessed through per-sample log-likelihood differences against iterative maximum-likelihood tomography and phase-resolved Jensen–Shannon divergence. For most settings, at least one ansatz has likelihood comparable to or above the tomography reconstruction. Increasing OPO power increases the inferred thermal fraction and weakens Wigner negativity, while $\eta$ stays comparatively stable. Increasing photon-addition pump power brings $\eta$ toward approximately 0.6, with nonmonotonic negativity. Posterior correlations between $\eta$ and $W(0,0)$ approach $-1$ in representative settings, whereas thermal-fraction correlations are weaker. Physical interpretations remain conditional on the ansatz: the paper notes that inappropriate models can bias them. Both ansatzes fit the 25 mW OPO dataset less well than tomography, suggesting omitted effects such as phase fluctuations or measurement imperfections. Each experimental setting is inferred independently, without sequential experimental updates.
Central claim may not hold: The abstract's claim of extracting complete information is unsupported because inference is conditional on restrictive physical ansatzes, empirical prior supports, and a variational posterior approximation. The claimed computational efficiency is not substantiated by runtimes, scaling results, or comparisons with alternative posterior-inference methods. The reported credible intervals lack reference-posterior comparisons or calibration tests establishing the accuracy of the learned posterior uncertainty. Representative-point likelihood comparisons do not establish Bayesian model support, and the evaluation split and positive likelihood differences relative to iMLE are insufficiently clarified.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 5 pages, 4 figures, with Supplementary Information