Closed-loop control for two-qubit gates with trapped ions

Technical Summary: Closed-Loop Control for Two-Qubit Gates with Trapped Ions

Problem Statement
State-of-the-art (SotA) implementations of two-qubit gates (2QG) in trapped-ion systems currently rely on open-loop control. These methods precompute control sequences based on simplified, closed-system models that assume ideal conditions. Consequently, they lack the reliability to meet fault-tolerant thresholds when faced with real-world disturbances such as motional heating, frequency drift, and decoherence. Once system parameters drift beyond robustness targets, open-loop schemes require computationally expensive re-optimization. Furthermore, these approaches are intrinsically blind to noise sources interacting with the system during gate execution.

Methodology
The authors propose a Closed-Loop Position Monitoring (CLPM) scheme that integrates real-time feedback with reinforcement learning (RL) to correct disturbances "on-the-fly."

1. Physical Architecture: The scheme introduces a spectator ion into the ion chain used for quantum logic processing. This ion couples to the computational ions via collective motional modes.
2. Continuous Monitoring: The spectator ion's position is continuously monitored by driving dipole transitions and detecting resultant fluorescence. A back mirror creates a standing wave, allowing the ion's position to be resolved via interference with its mirror image. The signal is processed by a time-correlated single-photon counting (TCSPC) system and an FPGA.
3. Mathematical Framework: The system dynamics are modeled using a stochastic quantum master equation (QME) derived from Quantum Trajectory Theory (QTT). The model incorporates:
   • Amplitude-modulated multimode-motional coupling (AM-MMC) gates.
   • Motional drift and heating.
   • Photon scattering recoil and spontaneous decay.
   • Light shifts (Autler-Townes and Bloch-Siegert).
   • The specific effects of the spectator ion's monitoring (spin-spin coupling, motion-mediated entanglement, and mirror-mode coupling).
4. Control Strategy: A Reinforcement Learning (RL) agent, specifically utilizing Proximal Policy Optimization (PPO), acts as the controller.
   • Input: The agent receives real-time estimates of the position ($x_k$) and momentum ($p_k$) of the motional modes, derived from the spectator ion's fluorescence.
   • Action: The agent adjusts the amplitudes of the bichromatic Raman lasers driving the gate ions.
   • Reward/Cost Function: The agent is trained to minimize gate infidelity ($I$). Two training approaches are explored: one based on episodic rewards (infidelity at the end of the gate) and another based on a cost function tracking the accumulation of the geometric phase toward the target $\pi/4$.

Key Contributions

• Feasibility of Spectator Monitoring: The paper demonstrates that incorporating a spectator ion is technically feasible for linear Paul traps. The authors show that the deleterious effects of position monitoring (spin coupling, motion-mediated errors, and mirror-mode coupling) are negligible compared to thermal effects, even in the absence of the spectator ion.
• Error Reduction: The inclusion of the spectator ion is expected to reduce the Bell-state preparation infidelity by an order of magnitude. The errors introduced by the monitoring process itself are calculated to be less than 10% of the total error budget.
• Real-Time Learning: The work synthesizes real-time quantum measurement with RL to devise robust control policies. Unlike global optimization algorithms that must be re-run when parameters drift, the RL agent learns to compensate for motional drift, heating, and light shifts dynamically.
• Validation Roadmap: The authors provide a three-stage roadmap for experimental validation: (1) verifying sympathetic cooling and signal strength, (2) reproducing predicted perturbative errors by manipulating laser detuning, and (3) validating the trained agent against engineered errors.

Results

• Signal Latency and Rate: The signal latency (photon travel time plus processing) is significantly lower than the typical gate duration (~100 $\mu$s), preserving the Markovian nature of the control loop. The photon detection rate is pushed to $1 \times 10^5 \text{ s}^{-1}$, providing sufficient information for trajectory reconstruction.
• Performance Comparison:
  • Episodic Reward Approach: The RL agent achieved a 33% improvement in infidelity over SotA methods, with a minimum infidelity of $46 \times 10^{-4}$.
  • Geometric Phase Approach: This method yielded the best results, achieving a 70% enhancement over SotA solutions, with a minimum infidelity of $14 \times 10^{-4}$ and an average of $85 \times 10^{-4}$.
• Pulse Characteristics: The RL-generated pulse sequences differed from standard SotA shapes. They utilized lower amplitudes (one-third lower than maximum allowed) and exhibited "Bell-like" shapes, which better satisfied the phase-space closure condition and reduced power-dependent disturbances.
• Trajectory Closure: The RL agents successfully navigated the non-convex solution space, ensuring that motional trajectories closed effectively (starting and ending points overlapped), a critical condition for high-fidelity gates.

Significance and Claims
The authors claim that their CLPM-scheme offers a pathway to significantly higher two-qubit gate fidelity with reduced calibration overhead. By shifting from static, open-loop models to a dynamic, data-driven closed-loop system, the approach mitigates the major obstacles of motional drift and decoherence in scalable ion-trap quantum computing. The paper asserts that the control method involves "learning" and correcting disturbances in real-time, making the system robust against parameter variations without the need for computationally expensive re-optimization. The work is presented as both theoretically sound and experimentally plausible, leveraging recent advancements in ion-position monitoring to achieve high-fidelity gates.