Classical Learning Surrogates Enhance Quantum Error Mitigation Efficiency

TLDR: Surrogate-Enabled Zero-Noise Extrapolation (S-ZNE) is a new quantum error mitigation technique that uses classical machine learning models to predict noisy quantum circuit outcomes. This approach drastically reduces the quantum measurement overhead typically associated with conventional Zero-Noise Extrapolation (ZNE), requiring only a constant number of measurements for an entire family of parameterized quantum circuits. S-ZNE achieves accuracy comparable to traditional ZNE, as validated by numerical experiments on up to 100-qubit systems for tasks like ground-state energy estimation and quantum metrology, offering significant resource savings.

Quantum computing holds immense promise for solving complex problems beyond the reach of classical computers. However, the practical application of these powerful machines is currently hampered by an inherent challenge: noise. Near-term quantum processors are susceptible to errors and decoherence, which degrade the fidelity of computations and make it difficult to obtain reliable results. While the ultimate goal is full-scale quantum error correction, which requires vast resources, Quantum Error Mitigation (QEM) techniques have emerged as crucial near-term solutions to improve computational accuracy with relatively low overhead.

Addressing the Measurement Burden in Quantum Error Mitigation

One of the most widely adopted QEM techniques is Zero-Noise Extrapolation (ZNE). ZNE works by systematically amplifying the noise levels in a quantum circuit, performing measurements at these amplified levels, and then extrapolating the results back to an ideal, noiseless outcome. While effective in principle, conventional ZNE faces a significant practical hurdle: a substantial measurement overhead. This issue becomes particularly problematic when applying ZNE to families of quantum circuits that are parameterized by classical inputs, such as those used in variational quantum algorithms or quantum simulations. In such scenarios, each distinct parameter value necessitates an independent and often costly execution of the full ZNE protocol, leading to an unsustainable demand for quantum resources.

Introducing Surrogate-Enabled ZNE (S-ZNE): A Scalable Solution

To address this critical bottleneck, researchers have developed an innovative framework called Surrogate-Enabled ZNE (S-ZNE). This groundbreaking approach leverages classical learning surrogates to perform the zero-noise extrapolation process almost entirely on the classical side, drastically reducing the need for repeated quantum measurements. A key advantage of S-ZNE is its superior scalability: unlike conventional ZNE, whose measurement cost scales linearly with the number of circuits, S-ZNE requires only a constant measurement overhead for an entire family of quantum circuits.

The core philosophy behind S-ZNE is to decouple the intensive data acquisition phase from the actual quantum execution. After an initial, one-time training phase, where a limited set of quantum measurements is used to train the classical surrogates, S-ZNE can mitigate errors for a wide range of parameterized circuits with different inputs purely through classical computation. This means that once the surrogates are trained, new inputs can be processed without any additional quantum measurement overhead, offering a fundamental scaling advantage.

The Three Pillars of S-ZNE

The S-ZNE framework operates through three distinct stages:

1. Data Collection

This initial step involves a one-time investment in quantum measurements. Noise scaling techniques, such as unitary folding, are applied to generate a comprehensive training dataset. At various artificially amplified noise levels, different classical inputs are fed into the quantum circuit, and the resulting expectation values are measured. This collected data forms the essential foundation for training the classical surrogates.

2. Surrogate Modeling

Once the training dataset is prepared, classical learning surrogates are trained. These surrogates are essentially machine learning models designed to accurately predict the expectation values of target observables under specific noise conditions. A significant benefit is that both the training and inference processes for these surrogates are entirely classical, making them highly efficient and fast.

3. Surrogate-Enabled Extrapolation

With the optimized classical learning surrogates in place, S-ZNE can then complete the ZNE estimation purely on the classical side. For any new input, the surrogates predict the corresponding expectation values at all relevant noise factors. These predictions are then extrapolated to the zero-noise limit, yielding an error-mitigated estimate with significantly reduced quantum overhead compared to traditional methods.

Theoretical analysis of S-ZNE indicates that, under many practical scenarios, it achieves an error scaling that is provably comparable to conventional ZNE. This is a crucial finding, as it suggests that the substantial resource savings do not come at the cost of mitigation quality or accuracy.

Validating S-ZNE in Practice

The effectiveness and efficiency of S-ZNE have been rigorously confirmed through extensive numerical experiments on quantum systems with up to 100 qubits. These experiments focused on two distinct and important applications:

Ground-State Energy Estimation: Using variational quantum algorithms (VQAs) for quantum many-body systems, specifically the one-dimensional transverse field Ising model and the Heisenberg model, S-ZNE demonstrated comparable accuracy to conventional ZNE. Critically, it achieved this while drastically reducing the sampling overhead. For example, in one optimization scenario, S-ZNE consumed 37.5 times fewer measurements over the entire optimization process.
Quantum Metrology: In tasks based on Ramsey interferometry with a 100-qubit GHZ probe state, S-ZNE successfully reconstructed the ideal phase estimation signals. It achieved an error mitigation performance statistically comparable to conventional ZNE, but with an impressive 80% reduction in quantum measurement cost during the inference stage.

The researchers also explored a hybrid S-ZNE approach, where classical learning surrogates are employed only at high noise levels, while direct quantum measurements are retained for low noise levels. This offers a flexible trade-off between accuracy and efficiency, further enhancing the practicality and adaptability of the framework for diverse quantum tasks.

Also Read:

A Promising Future for Quantum Error Mitigation

The introduction of S-ZNE represents a significant step forward in making quantum error mitigation more practical and scalable for near-term quantum processors. By effectively decoupling data acquisition from quantum execution and leveraging the power of classical learning surrogates, S-ZNE directly addresses the critical challenge of measurement overhead. This innovation paves the way for more robust, efficient, and ultimately, more useful quantum computations. The approach provides a versatile template that can potentially be extended to other quantum error mitigation protocols, opening new avenues for research and development in this rapidly evolving field. For more in-depth information, you can read the full research paper here.