Jingpeng Zhang of Sun Yat-sen University and colleagues have developed a new framework for recovering quantum states in continuous-variable (CV) systems, overcoming a key limitation of current machine-learning-based quantum error mitigation (QEM) methods. Existing QEM techniques typically demand training data covering the complete evolution of a quantum state, a process that is experimentally challenging.
The researchers introduce an extrapolative QEM approach using a time-conditioned Swin Transformer, which learns a correction map by embedding evolution time and capturing nonlocal phase-space correlations. Their results demonstrate accurate state recovery even in the long-time regime, establishing a practical route to mitigating noise without exhaustive training data.
Continuous-Variable Systems and Quantum State Degradation
Numerical simulations reveal that the new extrapolative quantum error mitigation (QEM) framework accurately recovers quantum states even when noise accumulates over extended periods, a regime where conventional methods falter. This capability addresses a critical limitation of existing machine-learning-based QEM techniques, which typically demand comprehensive training data spanning the entire duration of a quantum process. Current machine-learning approaches to QEM for CV systems require substantial resources for quantum state tomography, the process of reconstructing the Wigner function that represents the quantum state in phase space.
The team’s framework circumvents this challenge by learning a correction map that accounts for the continuous accumulation of noise and captures nonlocal phase-space correlations, a combination previously difficult to achieve. This is accomplished through a time-conditioned Swin Transformer, a neural network architecture that explicitly embeds the evolution time using adaptive layer normalization.
The adaptive layer normalization allows the model to extrapolate beyond the training horizon, effectively predicting the state’s evolution even for timeframes not included in the training data. The framework operates by applying a fiducial sequence, a combination of forward and backward evolution under the system’s Hamiltonian, to the quantum state. While these operations ideally cancel each other out in the absence of noise, environmental factors introduce decoherence, creating a noisy state.
The Swin Transformer then learns to map this noisy state back to the original, target state, effectively reversing the effects of the noise. The performance of this new approach was contrasted with a convolutional neural network (CNN) U-Net architecture, a common tool in image processing adapted for QEM. The simulations considered both Markovian and non-Markovian noise models, representing different types of environmental disturbances.
The framework may be extended to multimode continuous-variable systems, potentially increasing its versatility and applicability to more complex quantum information processing tasks. They also suggest incorporating physics-informed constraints into the learning process, which could further enhance the model’s accuracy and efficiency. The ability to accurately reconstruct quantum states without exhaustive training data represents an advancement for CV quantum systems.
This reduction in data requirements could substantially lower the experimental overhead associated with QEM, making it more accessible and practical for a wider range of applications, including quantum sensing, optical communication, and quantum computing. The framework’s success in handling both Markovian and non-Markovian noise further solidifies its potential as a robust and versatile tool for mitigating errors in CV quantum systems.
Time-Conditioned Swin Transformer for Error Mitigation
Researchers at Sun Yat-sen University and Tsinghua University have developed a time-conditioned Swin Transformer that mitigates errors in continuous-variable (CV) quantum systems, achieving high fidelity even when extrapolating beyond the initial training period. This new framework explicitly incorporates the progression of time into its architecture, using adaptive layer normalization to model the continuous accumulation of noise. By embedding the evolution time, the Swin Transformer learns a “correction map” capable of accounting for the subtle distortions that accumulate over longer timescales, while simultaneously capturing complex relationships within the quantum state’s phase space.
The design allows the model to function effectively without access to ideal, noiseless quantum states, a significant advantage in practical experimental settings where perfect control is often unattainable. Numerical simulations under both Markovian and non-Markovian noise demonstrate accurate state recovery in the long-time regime, where existing approaches deteriorate.
The researchers found that the Swin Transformer’s success stems from its ability to dynamically adapt its normalization scale through adaptive layer normalization, a technique that allows the model to follow continuous trajectories and extract fine features from the noisy data. This is particularly advantageous in environments where memory effects amplify subtle distortions over time. The input channels varied slightly, using five channels for Markovian dynamics and two (Wigner function plus delta tau) for non-Markovian dynamics.
Nonlocal Phase-Space Correlations Enable Long-Time Recovery
Existing methods typically falter when asked to extrapolate beyond the timeframes used to build their correction maps, a limitation stemming from their inability to capture the continuous accumulation of noise and subtle, long-range correlations within phase space. This new framework, based on a time-conditioned Swin Transformer, addresses this challenge by explicitly modeling noise dynamics as a function of evolution time and incorporating nonlocal feature extraction.
The architecture’s success hinges on its ability to discern faint structural correlations within degraded Wigner distributions, where environmental dissipation has diminished fine quantum features. Unlike convolutional neural networks, which struggle to track the memory-driven modulations of phase-space distributions over extended periods, the Swin Transformer uses a self-attention mechanism to sensitively extract weak signals and adapt reconstruction based on the evolution time, denoted as τ.
Numerical simulations reveal a significant performance difference; while conventional CNN-based methods exhibit a decline in reconstruction accuracy when extrapolating beyond the training horizon, the Swin Transformer maintains a cosine similarity of approximately 0.92 with the target state, preserving characteristic non-monotonic behavior. This preservation of structural correlations is a consequence of the model’s capacity to capture the underlying law governing noise accumulation. The framework processes Wigner functions through an encoder-decoder pipeline with skip connections, enabling multi-scale feature extraction while preserving fine phase-space structure.
This hierarchical approach allows the model to identify and reconstruct subtle nonlocal patterns that emerge as dissipation progresses, patterns that would otherwise be lost to conventional methods. The performance metric, cosine similarity, quantifies the structural fidelity of the reconstructed quantum state, with values approaching unity indicating successful recovery of the target state.
The design of this system extends beyond simply improving reconstruction accuracy. It also offers a pathway to reduce the experimental overhead associated with quantum error mitigation. By accurately reconstructing quantum states well beyond the temporal range covered by training data, this framework substantially lowers those requirements. The team demonstrated the framework’s efficacy across both Markovian and non-Markovian noise models, showing its versatility and applicability to a broad range of quantum systems.
The test horizon extends to t ∈, encompassing both the interpolation regime and the extrapolation regime (], allowing for a rigorous assessment of the model’s ability to generalize beyond observed data. This capability is particularly important for complex quantum information processing tasks, where long-time evolution and subtle distortions can significantly impact the fidelity of results. The architecture’s success is not solely dependent on the Swin Transformer itself, but on the synergistic combination of nonlocal feature extraction and continuous time conditioning.
Extrapolative QEM Reduces Exhaustive Training Data Needs
Numerical simulations reveal a framework capable of accurately recovering quantum states even when the duration of the simulation extends to twice the length of the initial training data, a feat previously unattainable with existing quantum error mitigation techniques. 0 arbitrary units and subsequently testing its performance up to 2.0 units. This testing regime encompassed both interpolation, where the model predicts within the range of training data, and extrapolation, where it predicts beyond it, demonstrating the framework’s ability to generalize to previously unseen temporal regimes.
The framework employs an iterative inference strategy for non-Markovian noise, using delta-tau steps to refine the error correction process. This configuration simulates realistic conditions encountered in CV quantum systems, where photon loss and dephasing are primary sources of degradation. The team’s approach differs from previous methods by explicitly modeling the continuous accumulation of noise, allowing for more accurate extrapolation beyond the training horizon. By learning from data within a limited training horizon, the framework minimizes the need for exhaustive data collection across the entire temporal range of interest.
This is particularly valuable for CV applications, where the collection of extensive training data can be experimentally costly and time-consuming. The framework’s ability to accurately recover quantum states beyond the training horizon represents a step toward realizing the full potential of CV quantum systems for a wide range of applications.
👉 More information
🗞 Extrapolative quantum error mitigation in continuous-variable systems beyond the training horizon
✍️ Jingpeng Zhang, Shengyong Li, Jie Han, Qianchuan Zhao, Jing Zhang and Ze-Liang Xiang
🧠 DOI: http://link.aps.org/doi/10.1103/z48d-38tn




See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.
