Abstract
We present a gradient-based method to construct high-fidelity, two-qubit quantum gates in a system consisting of two transmon qubits coupled via a tunable coupler. In particular, we focus on single flux quantum (SFQ) pulses as a promising and scalable alternative to traditional control schemes that use microwave electronics. We develop a continuous embedding scheme to optimize these discrete pulses, taking advantage of auto-differentiation of our model. This approach allows us to achieve fSim-type gates with average gate fidelities on the order of 99.99% and CZ and CNOT gates with fidelities above 99.9%. Furthermore, we provide an alternative semi-analytical construction of these gates via an exact decomposition using a pair of fSim gates which leads to the reduction in memory required to store the associated pulse sequences.
Introduction
The scaling of superconducting quantum processors is currently hindered by the overhead of room-temperature microwave control electronics. As qubit counts increase, the density of coaxial cabling and the resulting heat load on the cryogenic environment become unsustainable. Single Flux Quantum (SFQ) technology offers a compelling solution by providing digital, ultra-fast voltage pulses generated at the cryogenic stage. However, the discrete nature of SFQ pulses—which are either present or absent—poses a significant challenge for traditional gradient-based optimization techniques typically used for analog pulse shaping.
Methodology
The researchers investigated a system of two transmon qubits connected by a tunable coupler, as shown in the schematic below. This architecture allows qubits to idle at the same frequency, simplifying single-qubit operations while reducing unwanted interactions.
To overcome the non-differentiability of discrete pulses, the authors employed a "continuous embedding" technique. During the optimization process, pulse amplitudes were allowed to vary continuously between 0 and 1. By incorporating penalty terms and logarithmic smoothing into the cost function, the optimizer was guided toward binary (0 or 1) solutions by the end of the training. This allowed the use of efficient gradient-descent algorithms (like L-BFGS-B) via the JAX library.
Two primary strategies were explored:
- Direct Optimization: Optimizing the entire pulse sequence for a target gate (e.g., CNOT) from scratch. This yields faster gates (~70-80 ns) but requires significant memory to store the complex bitstrings.
- Analytical Decomposition: Breaking down complex gates into native fSim gates and single-qubit rotations. This modular approach significantly reduces the memory footprint required for hardware implementation, requiring less than 200 bits of storage.
Results and Discussion
The optimization framework successfully produced high-fidelity gates across various configurations. The system achieved average gate fidelities of 99.99% for fSim gates and exceeded 99.9% for CZ and CNOT gates. Single-qubit rotations also performed exceptionally well, with fidelities ranging between 99.9% and 99.99%.
A key finding was the impact of the SFQ clock frequency on gate performance. Higher clock frequencies provide finer temporal resolution for the pulse sequences, leading to lower infidelities.
The study also highlighted a trade-off between gate speed and memory efficiency. While direct optimization produces the fastest gates, the analytical decomposition approach is highly suitable for current hardware with limited on-chip memory due to its significantly smaller bitstring requirements.
Conclusion
This work demonstrates that SFQ-based control can achieve fidelities competitive with traditional microwave schemes while offering a path toward scalable, cryogenically-integrated quantum control. Future efforts will focus on experimental validation on physical hardware and addressing the challenges of crosstalk in multi-qubit arrays. Finding the optimal balance between high clock frequencies and digital logic complexity remains a key task for hardware implementation.
Summary updated