Flow-Based Modeling Reconstructs Quantum States From Fewer Measurements

Researchers at the Massachusetts Institute of Technology, Stanford University, Tufts University, the University of California, Los Angeles, and Tsinghua University have developed QST-Flow, a new quantum state tomography framework that moves beyond traditional methods of modeling quantum information. QST-Flow represents data using “normalized, samplable neural densities” instead of a truncated density matrix, enabling more efficient processing of complex quantum states. The framework features two variants, QST-QFlow and QST-WFlow, which model the positive Husimi function and sign-changing Wigner functions, respectively, as trainable densities. This construction preserves quasiprobability normalization and enables exact density evaluation and direct sampling, along with importance-sampled learning from finite phase-space measurements without a fixed grid. Benchmarks using states like cat, binomial, and Gottesman-Kitaev-Preskill states demonstrate that QST-WFlow achieves improved reconstruction error compared with QST-CGAN, suggesting a quantifiable advancement in scalable, measurement-efficient phase-space tomography of nonclassical bosonic systems.

QST-Flow Framework for Continuous-Variable State Tomography

A new approach to quantum state tomography leverages machine learning to map complex quantum states with increased efficiency. This shift in modeling allows for more streamlined data processing crucial for understanding and manipulating quantum systems. QST-Flow distinguishes itself through two variants, QST-QFlow and QST-WFlow, each designed to handle different aspects of quantum behavior. QST-QFlow specifically models positive Husimi functions using a single normalizing flow, while QST-WFlow employs a “trainable difference of two normalized flows” to accurately represent sign-changing Wigner functions. This construction fundamentally preserves quasiprobability normalization, enabling exact density evaluation, direct sampling, and learning from limited phase-space measurements without the constraints of a fixed grid. Accurately capturing both positive and negative quasiprobabilities is critical for characterizing non-classical states, a long-standing challenge in the field.

Rigorous testing of QST-Flow’s capabilities involved benchmarking against quantum states, including cat, binomial, Gottesman-Kitaev-Preskill, number, and Fock states. This improvement extends beyond single-mode states, as the framework also handles more complex, multimode systems. The system exhibits robustness even with noisy Wigner data, a common issue in real-world experiments. Normalizing flows have recently become an important development in machine learning and are especially well suited to the positive part of the continuous-variable tomography problem. These flows map a simple density to a complex target density through an invertible transformation, offering a powerful tool for density estimation and simulation. The team focused on experimentally accessible phase-space representations, the Husimi function and the Wigner function, to fully determine a quantum state, recognizing that the Wigner function’s sign changes reveal nonclassicality.

Measurements of the Husimi function, directly measurable and non-negative, are a Gaussian-smoothed version of the Wigner function, retaining tomographic information. The Wigner function can be obtained from homodyne reconstruction or direct displaced-parity measurements. Finite-shot Wigner measurements are governed by binomial statistics.

Husimi and Wigner Functions in Phase-Space Reconstruction

Fully characterizing quantum states has long been constrained by the complexity of representing them, particularly for continuous-variable systems. Traditional methods, relying on truncated density matrices or discretized phase-space grids, struggle to scale with increasing numbers of modes and often sacrifice resolution of crucial non-Gaussian features. QST-Flow addresses these limitations by leveraging normalizing flows, a machine learning technique, to model experimentally accessible phase-space quasiprobability distributions with “normalized, samplable neural densities” rather than relying on conventional representations. This shift allows for more efficient processing and a potentially more accurate reconstruction of quantum states. QST-QFlow directly models the positive Husimi function using a single normalizing flow, trained through sample-based likelihood estimation.

Researchers are increasingly focused on streamlining quantum state tomography, and QST-Flow offers a potentially significant advance in efficiently characterizing continuous-variable quantum systems. This approach is particularly relevant as quantum systems grow in complexity, demanding more efficient methods for state reconstruction. The core innovation lies in how QST-Flow represents quasiprobability distributions. The resulting models provide direct sampling and exact density evaluation, circumventing the limitations of fixed grid methods often used in traditional tomography. Researchers validated the framework’s ability to accurately model the nuances of both the Husimi and Wigner functions, demonstrating its versatility, which is particularly crucial for applications where data acquisition is limited or subject to imperfections.

While conventional tomography relies on discretizing this space, either through Fock basis truncation or phase-space grids, these methods struggle with complex, non-Gaussian states and rapidly become computationally expensive as system size increases. This innovation allows for models to provide direct sampling and performs importance-sampled learning, a crucial advantage when dealing with limited data or measurement imperfections.