Abstract
A major challenge in developing quantum computing technologies is to accomplish high precision tasks by utilizing multiplex optimization approaches, on both the physical system and algorithm levels. Loss functions assessing the overall performance of quantum circuits can provide the foundation for many optimization techniques. In this paper, we use the quadratic error loss and the final-state fidelity loss to characterize quantum circuits. We find that the distribution of computation error is approximately Gaussian, which in turn justifies the quadratic error loss. It is shown that these loss functions can be efficiently evaluated in a scalable way by sampling from Clifford-dominated circuits. We demonstrate the results by numerically simulating ten-qubit noisy quantum circuits with various error models as well as executing four-qubit circuits with up to ten layers of two-qubit gates on a superconducting quantum processor. Our results pave the way towards the optimization-based quantum device and algorithm design in the intermediate-scale quantum regime.
Introduction
In the current "Noisy Intermediate-Scale Quantum" (NISQ) era, the primary hurdle to useful quantum computing isn't just building more qubits—it's managing the noise they produce. To optimize a quantum algorithm or a hardware device, researchers need a way to measure exactly how much "loss" or error a specific circuit accumulates. However, traditional methods for characterizing noise either don't scale to many qubits or fail to capture the complex, time-dependent errors found in real hardware.
Standard techniques like Quantum Process Tomography (QPT) provide full information but scale exponentially, becoming impossible beyond a few qubits. Randomized Benchmarking (RB) scales well but only provides an average gate fidelity, often missing spatial and temporal correlations (like crosstalk) that occur in actual algorithm execution. This paper addresses this gap by presenting a scalable solution: using Clifford sampling to efficiently evaluate the error loss of large quantum circuits.
Methods: Clifford Sampling and the Frame Concept
The researchers focus on two primary metrics: quadratic error loss (the squared difference between noisy and ideal results) and final-state fidelity loss. The core innovation is the use of Clifford Sampling, which relies on dividing a circuit into "frame operations" (fixed entangling gates, initialization, and measurement) and "variable operations" (single-qubit gates).
Mathematically, the Clifford group is a "unitary 2-design." This means that if you average a quadratic function (like error loss) over the entire space of possible unitary gates (Haar measure), the result is identical to averaging over just the Clifford group. Because Clifford circuits can be efficiently simulated on classical computers (via the Gottesman-Knill theorem), the "ideal" error-free result needed for the loss function can be computed easily, even for very large circuits. To account for "gate-dependent" noise, the authors introduce a hybrid approach that mixes Clifford gates with a few general unitary gates.
Key Insights and Experimental Results
The study validated the method through numerical simulations of 10-qubit circuits and experiments on a 4-qubit superconducting Xmon processor. One of the most striking findings is that for large, random unitary circuits, the distribution of computation error is approximately Gaussian. This provides a strong physical justification for using "quadratic loss" as a standard metric for error.
Furthermore, the authors prove and experimentally verify that the loss measured by sampling general unitaries ($L_U$) is equivalent to the loss measured by sampling Clifford gates ($L_C$). This "Clifford equivalence" is the key to the method's scalability, as it allows complex circuit performance to be estimated using classically simulatable components.
The resource cost of the method scales polynomially with the number of qubits and circuit depth. On the 4-qubit processor, the authors used their Clifford loss function to optimize the Rabi frequency of a two-qubit gate, reducing the error loss by more than 10%, proving the method's utility for device calibration.
Conclusion
While challenges remain—such as extremely strong gate-dependent noise or the classical overhead of hybridization in high-precision scenarios—this method provides a powerful tool for NISQ-era optimization. The authors suggest this approach could be integrated directly into the training loops of variational quantum algorithms (like VQE or QAOA) to perform "learning-based" error mitigation in real-time, paving the way for more robust quantum computation in the intermediate-scale regime.
Summary updated