arXiv:2609.25208
Adaptive thresholding for scalable measurement-based qubit reset
Qian Cao, Unnati Akhouri, Samuel Vizvary, Nissim Ofek, Wei Dai, Kater W. Murch
Abstract
Fast, high-fidelity qubit initialization is a key primitive for scalable quantum information processing, but conventional reset protocols either require long relaxation times or discard information contained in continuous measurement records. We demonstrate an adaptive measurement-based reset protocol for superconducting qubits that uses Bayesian inference to update the qubit-state estimate after each readout and dynamically adjust the feedback threshold. Unlike fixed-threshold or repeat-until-success (RUS) protocols, the method utilizes the full analog measurement history, enabling the reset decision to become progressively more conservative as confidence in ground-state preparation increases. Implemented with real-time FPGA-based feedback on a dispersively readout transmon qubit, the protocol achieves a ground-state initialization fidelity of $99.44\pm0.04\%$ after a small number of reset rounds. We further compare adaptive reset with RUS strategies in multi-qubit experiments and show that adaptive thresholding provides deterministic reset duration while maintaining a five-qubit simultaneous-reset fidelity of $98.78 \pm 0.19\%$. These results establish Bayesian adaptive thresholding as a practical and scalable route to fast qubit reset in superconducting quantum processors.
AI Summary
Measurement-based qubit reset usually converts analog readout into binary decisions, discarding confidence information; repeat-until-success (RUS) schemes also have stochastic completion times. This paper studies whether retaining measurement history can improve reset while fixing its duration in advance. The protocol maintains a Bayesian ground-state probability using calibrated Gaussian readout likelihoods. After each measurement, it conditionally applies an $X_\pi$ pulse when the posterior favors excitation, updates the probability to account for that pulse, and recalculates the next decision threshold. Increasing confidence in ground-state preparation makes subsequent feedback more conservative. The procedure is implemented in real time on an FPGA controller for superconducting transmons. Single-qubit initialization reaches $99.44\pm0.04\%$ fidelity, saturating after approximately six rounds despite single-shot assignment fidelity near $92\%$. Residual excitation is characterized using both Gaussian-mixture fits and an independent excited-to-second-excited-state Rabi method. QND Monte Carlo simulations reproduce measurement and threshold distributions but predict continued improvement beyond the experimental floor. Five-qubit simultaneous-reset fidelity reaches $96.27\pm0.08\%$ at higher readout power and $98.78\pm0.19\%$ at lower power. Multi-qubit comparisons use $7\,\mu\mathrm{s}$ rounds; fixed schedules of six or eleven rounds remove completion-time variability. An independent-qubit model gives RUS completion time growing approximately as $\ln N$. The stated limitations include measurement-induced transitions, higher transmon-level population, and relaxation and thermal excitation during finite-duration measurements. Lower readout power reduces measurement-induced transitions but requires more rounds. The demonstrated inference model assumes two-state Gaussian likelihoods and QND measurements; multilevel inference is left as an extension.
Error found: The demonstrated qubit-number-independent duration holds at fixed round count, not at fixed simultaneous-reset fidelity. For unequal Gaussian readout variances, the implemented single-threshold feedback rule does not generally reproduce the full Bayesian decision regions. The stated pi_max bounds the existence of decision thresholds rather than posterior confidence, and freezing a finite threshold beyond it can cause flips even when the Bayesian posterior favors the ground state. Section III.B reverses the posterior inequality for applying a reset pulse: with pi_i=P(g), the flip condition is pi_i<0.5, as correctly stated in the appendix.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 16 pages, 9 figures