arXiv:2610.15975
Scalable Exact Non-Pauli Noise Simulation for Fault-Tolerant Quantum Circuits
Yuewen Hou, Gokul Subramanian Ravi
Abstract
Fault-tolerant quantum design relies on simulating noisy quantum error correction (QEC) circuits with hundreds to thousands of physical qubits and large circuit depths. High-throughput stabilizer simulators, such as Stim, are the de facto tools for this regime. However, these simulators typically support only Clifford gates and stochastic Pauli noise. As a result, more realistic device noise models-such as coherent control errors and thermal relaxation-are often approximated using Pauli Twirling. This replacement offers speed but can introduce inaccuracies. To address this, we propose QuasiStim for scalable exact QEC simulation of realistic noise, with QuasiFold as a compiler frontend and DeltaFrame as a sampling backend. QuasiFold lowers sampling overhead by merging equivalent non-Pauli sites in the quantum circuit. DeltaFrame samples the resulting stabilizer trajectories by reusing precomputed predicted references instead of resimulating the full tableau for each shot. QuasiFold can reduce the number of shots by over $100\times$ on bivariate-bicycle (BB) codes. Compared to a per-shot tableau simulator, DeltaFrame is $166\times$--$1331\times$ faster on rotated surface-code memory under coherent $R_Z$ noise as $d$ grows from 5 to 35. QuasiStim also makes the PTA bias measurable: the gap is non-negligible and varies with noise strength, decoder, and circuit schedule. QuasiStim will be open-sourced upon publication at https://github.com/isaachyw/QuasiStim
AI Summary
Fault-tolerant circuit studies commonly replace coherent errors and relaxation by stochastic Pauli noise to enable fast stabilizer sampling. This paper asks how to simulate the specified non-Pauli channels at QEC scale and quantify the bias introduced by this replacement. QuasiStim combines stabilizer-channel quasiprobability decomposition (QPD) with two components. QuasiFold fuses compatible rotations, removes unobservable rotations, and absorbs neighboring Pauli noise when this reduces negativity. DeltaFrame precomputes branch-dependent measurement-reference offsets, groups sites using pairwise interaction tests, and samples heterogeneous trajectories using batched Pauli frames. QPD is unbiased with exact trajectory execution, but its sampling overhead scales as $\Gamma^2$, where $\Gamma$ is the product of local decomposition one-norms. Benchmarks cover surface, color, and bivariate-bicycle memories, logical operations, and magic-state cultivation. QuasiFold reduces the nominal shot overhead by $1.16$–$38.3\times$ on surface codes and over $100\times$ on bicycle codes. DeltaFrame reports $166$–$1331\times$ lower per-trajectory cost than independent tableau simulation for coherent noise at surface-code distances $5$–$35$. These timing tests bypass reference verification and parts of preprocessing. Comparisons with Pauli-twirled simulations find logical-error-rate differences dependent on noise strength, decoder, and schedule, reaching approximately $6\times$ in a cultivation setting. Despite the title, DeltaFrame is not universally exact: missed higher-order or inter-block interactions produce measured reference-mismatch rates up to $2\times10^{-5}$; none were observed in tested relaxation cases. Negativity still prevents unrestricted high-distance coherent-noise studies within the available budget, motivating defect-restricted or segmentwise benchmarks. Other stated limitations include greedy compilation, exponential block enumeration, decoding cost, and no treatment of leakage or loss.
Central claim may not hold: QuasiStim is advertised as exact, but its DeltaFrame backend explicitly uses reference predictions without a general exactness guarantee and reports nonzero record and silent-error rates. The headline sampling-speed benchmarks disable reference-precomputation steps and verification, so they measure frame-execution throughput rather than verified exact simulation throughput. Large-distance per-trajectory benchmarks do not establish practical unrestricted coherent-noise LER estimation, which the paper reports as unaffordable even at distances 5 and 7. The claim that reference errors do not affect LER is unsupported because signed QPD weights can amplify mismatch-induced bias by Γ, and silent-error counts omit other decoder-relevant mismatches.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.