Atlas: Accelerating Quantum FTQC Compilation with Optimized Algebraic Rotation Synthesis
Abstract
As fault-tolerant quantum computing emerges, Clifford+T gate basis becomes a popular choice where we need to decompose single-qubit rotation gates into gates in this gate basis. Quantum rotation synthesis may lead to remarkable overhead in quantum circuit compilation. Considering variational quantum eigensolver (VQE) applications as an example, a conventional UCCSD ansatz in quantum chemistry can generate $O(n^4)$ rotations for $n$ qubits. Each optimization iteration updates the rotation angles, making rotation synthesis a recurring cost that can delay the VQE training loop. We identify two sources of delay in rotation synthesis: redundant computation during synthesis and uneven synthesis times that leave workers waiting for slow requests. Atlas addresses these costs with three optimizations: search pruning skips invalid synthesis candidates through statistical inference; adaptive arithmetic avoid unnecessary high-overhead big integer/float arithmetic when decomposing the target rotation gate; batch scheduling fills the gap between the synthesis time of varied rotations. Across various benchmarks, synthesis method, and compilation machines, Atlas can consistently accelerate the rotation synthesis process by up to 12.65x.