arXiv:2610.31885
Optimal Control of a Fluxonium Circuit with the Q-PRONTO Numerical Method
John Barta, Brian D. Isakov, Pablo Aramburu Sanchez, Jieqiu Shao, Tamás Kalmár, Trevyn F. Q. Larson, Marco M. Nicotra, András Gyenis
Abstract
Optimizing the control of large quantum systems is a resource-intensive task, where the convergence rate of the underlying iterative schemes is crucial. The recently developed Quantum Projection Operator-Based Newton Method for Trajectory Optimization (Q-PRONTO) offers a unique quadratic convergence for controlling quantum states, while most model-based solvers guarantee only an approximately linear convergence rate. As a proof of concept, here, we design and benchmark single-qubit gates optimized by Q-PRONTO based on the circuit Hamiltonian of a fluxonium superconducting qubit. To ensure that we can apply a long string of arbitrary pulses to the circuit, we modify the firmware associated with the Quantum Instrumentation Control Kit (QICK) by altering its memory allocation. We characterize the performance of the optimal solutions using standard randomized benchmarking and achieve a total error per gate of $6.6\times10^{-4}$.
AI Summary
The paper tests whether Q-PRONTO, an existing Newton-based quantum optimal-control method with local quadratic convergence, can generate experimentally usable gates for a superconducting fluxonium. It addresses experimental implementation rather than demonstrating a computational speedup over other optimizers. A spectroscopy-derived circuit model is truncated to three levels, and Q-PRONTO designs an $X_{\pi/2}$ pulse with penalties on drive amplitude, rapid waveform variation, and occupation outside the computational subspace. The fluxonium operates at its half-flux sweet spot with transition frequency $225\,\mathrm{MHz}$, $T_1\approx263\,\mu\mathrm{s}$, and echo pure-dephasing time $T_\phi^E\approx155\,\mu\mathrm{s}$. Modified QICK firmware reallocates memory to store complete arbitrary-waveform sequences: approximately 2 MB supports $152.4\,\mu\mathrm{s}$ at 6.88 GSamples/s. Clifford sequences combine optimized $X$ pulses with waiting-time $Z$ rotations. Randomized benchmarking using 200 random sequences per length and 512 repetitions per sequence gives decay parameter $p=0.99867\pm0.00002$ for $178\,\mathrm{ns}$ pulses, corresponding to average error per Clifford $(1-p)/2=6.6\times10^{-4}$. A white-noise coherence estimate for a single $178\,\mathrm{ns}$ pulse is $6.1\times10^{-4}$; this proximity is interpreted as evidence of coherence-limited performance. The demonstrated waveform largely resembles an enveloped resonant sinusoid. Hardware and design limitations prevented testing shorter pulses, where optimized waveforms differ more substantially. The current implementation does not explicitly optimize robustness to experimental uncertainty, and penalty weights are chosen empirically. Benefits for larger systems, beyond-rotating-wave control, two-qubit gates, and closed-loop operation remain future directions.
Error found: The paper reports an RB error of 6.65×10⁻⁴ per compiled Clifford but subsequently attributes it to the optimized pulse without providing a conversion to error per Xπ/2 pulse. The coherence-limited-performance conclusion compares error per compiled Clifford with a coherence estimate for one 178 ns pulse, without accounting for the Clifford compilation's average duration.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.