Bootstrapping Classical Shadows for Neural Quantum State Tomography

Wirawat Kokaew, Bohdan Kulchytskyy, Shunji Matsuura, Pooya RonagharXiv 2024 · DOI · arXiv

Abstract

We investigate the advantages of using autoregressive neural quantum states as ansatze for classical shadow tomography to improve its predictive power. We introduce a novel estimator for optimizing the cross-entropy loss function using classical shadows, and a new importance sampling strategy for estimating the loss gradient during training using stabilizer samples collected from classical shadows. We show that this loss function can be used to achieve stable reconstruction of GHZ states using a transformer-based neural network trained on classical shadow measurements. This loss function also enables the training of neural quantum states representing purifications of mixed states. Our results show that the intrinsic capability of autoregressive models in representing physically well-defined density matrices allows us to overcome the weakness of Pauli-based classical shadow tomography in predicting both high-weight observables and nonlinear observables such as the purity of pure and mixed states.

Introduction

Quantum State Tomography (QST) is a fundamental task in quantum information science, aimed at reconstructing the full state of a quantum system from measurements. However, as the number of qubits grows, the complexity of QST increases exponentially. Traditional methods like Classical Shadows offer an efficient way to predict many properties from a few measurements but often yield unphysical states and struggle with "high-weight" observables (complex correlations involving many qubits). Conversely, Neural Quantum States (NQS), particularly those based on autoregressive architectures like Transformers, provide a compact and physically valid representation. However, training NQS on shadow data has previously been unstable. This paper presents a robust framework that bridges these two approaches, leveraging the strengths of both.

Methodology

The researchers propose two key innovations to stabilize and enhance NQS training:

  • Shadow-Based Cross-Entropy (SCE) Loss: Standard cross-entropy training often treats all measurement outcomes equally. By incorporating weights derived from classical shadows, the SCE loss injects a stronger physical signal into the training process, facilitating faster convergence and higher accuracy.
  • Stabilizer-Based Importance Sampling: To compute gradients for the neural network, the authors propose sampling from the stabilizer states (the measurements) rather than the model itself. This shift from a reinforcement learning-style approach to a supervised-style approach significantly reduces the variance of the training gradients.
  • Purification Ansatz: For mixed (noisy) states, the density matrix is represented as a purified state in an enlarged Hilbert space. This ensures that the model always corresponds to a physically valid, positive semi-definite density matrix.

Results and Discussion

The proposed method was benchmarked on Greenberger–Horne–Zeilinger (GHZ) states, which are highly entangled and notoriously difficult to reconstruct. For an 8-qubit GHZ state, where previous NQS methods often failed to converge (stuck at 99% error), the new SCE loss achieved an infidelity as low as 5%.

Performance of the NQS reconstruction. (a) KL divergence as a function of system size. (b) KL divergence as a function of the number of shadows, demonstrating the convergence of the shadow-based estimator compared to empirical baselines.

As shown in the figures above, the shadow-based NQS reconstruction ($p^{sh}$) consistently outperforms raw empirical estimates ($p^{emp}$), especially as the number of shadows increases. The NQS effectively acts as a "physicality filter," ignoring the unphysical noise inherent in raw classical shadow data. This allows for the accurate prediction of high-weight Pauli observables and purity, even when using simple Pauli measurements that would typically be insufficient for such complex correlations.

Conclusion

By bootstrapping classical shadows with neural quantum states, this work provides a stable and accurate path for large-scale quantum state tomography. The method's ability to predict high-weight observables and maintain physicality makes it a promising tool for characterizing near-term quantum devices. Future work will focus on validating these techniques on physical quantum hardware and exploring further scaling with advanced transformer architectures.

Summary updated

Abstract

We investigate the advantages of using autoregressive neural quantum states as ansatze for classical shadow tomography to improve its predictive power. We introduce a novel estimator for optimizing the cross-entropy loss function using classical shadows, and a new importance sampling strategy for estimating the loss gradient during training using stabilizer samples collected from classical shadows. We show that this loss function can be used to achieve stable reconstruction of GHZ states using a transformer-based neural network trained on classical shadow measurements. This loss function also enables the training of neural quantum states representing purifications of mixed states. Our results show that the intrinsic capability of autoregressive models in representing physically well-defined density matrices allows us to overcome the weakness of Pauli-based classical shadow tomography in predicting both high-weight observables and nonlinear observables such as the purity of pure and mixed states.
BibTeX
@misc{kokaew2024bootstrapping,
  author = {Wirawat Kokaew and Bohdan Kulchytskyy and Shunji Matsuura and Pooya Ronagh},
  title = {{Bootstrapping Classical Shadows for Neural Quantum State Tomography}},
  year = {2024},
  eprint = {2405.06864},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2405.06864},
}
Figure