A Comparison of Quantum Gate Optimization Techniques

Abstract

Better quantum gates are likely key to enabling fault-tolerant, useful quantum computers. This thesis compares gate optimization techniques by simulating single-qubit and two-qubit gates for superconducting qubits. The primary focus is deep reinforcement learning. For single-qubit gates, the task is to optimize a π-pulse, while for two-qubit gates, the task is to optimize the Controlled-Z gate. The results indicate that using an ansatz for the gate’s pulse shape can enhance the performance of deep reinforcement learning, both for single-qubit and two-qubit gates, but only significantly for single-qubit gates. A simple square-pulse ansatz approximately halves the simulation time needed to reach the coherence limit for the single-qubit gate studied. The speed-up in simulation should translate to a speed-up in experiments as well. The thesis does not find evidence that the implemented deep reinforcement learning algorithm yields better quantum gates than a state-of-the-art black-box optimizer, despite the black-box optimizer being easier to implement experimentally. For a quantum gate defined by piece-wise constant controls, a low-pass filter seems to enhance the performance, at least if the filter is considered when optimizing. This indicates that piece-wise constant controls, for example, generated with deep reinforcement learning, are not hindered by the limited bandwidth of control electronics. Finally, the study highlights the importance of ZZ coupling to understanding Controlled-Z gates.