arXiv:2610.13589
Decoder- and Noise-Aware Optimization Framework for Clifford-Deformed Quantum Error-Correcting Codes
Julius Rye Bønnelykke, Benjamin Joecker, Peter Krogstrup, Vincent Philippe Michal, Mark Kamper Svendsen
Abstract
Practical fault-tolerant quantum computing requires qubit-efficient quantum error correction (QEC). Clifford deformations (CDs) modify the stabilizers of a QEC code to lower its logical error rate at no additional resource cost. However, the number of possible CDs grows exponentially with the number of qubits, making exhaustive search infeasible beyond small codes. In this work, we further show that under circuit level noise the optimal CD depends strongly on both the noise model and decoder. To make systematic CD optimization possible, we introduce Evolution-Based Best-arm Identification (EBBI), a shot-efficient search algorithm that treats the circuit, noise model and decoder as a black box. We benchmark EBBI's performance on the rotated surface code, under two strongly $Z$-biased circuit-level noise models with the PyMatching decoder. At distance 3, EBBI finds the optimal CD using up to $\sim 1100\times$ fewer shots than random search. At distance 5, when run at the target error rate, the median EBBI run converges to the CDs we believe to be optimal, while random search never converges. Under a $Z$-biased SI1000-like noise model, EBBI discovers a new CD, the B-XZZX code, which we generalized to larger distances. It differs from XZZX on only a fraction $\mathcal{O}(d^{-1})$ of the qubits at code distance $d$, yet its relative advantage grows from $\approx 5\%$ at $d = 3$ to $\approx 12\%$ at $d = 11$. We thus show that the B-XZZX pattern found by EBBI can be systematically scaled to larger code distances. Our work represents a first step towards scalable CD optimization for general QEC codes.
AI Summary
Clifford deformations tailor stabilizer codes to biased noise without changing qubit count or connectivity, but searching over local deformations is exponentially expensive. This paper addresses optimization with circuit-level noise and a specified practical decoder, showing that the preferred deformation can depend strongly on both. Evolution-Based Best-arm Identification (EBBI) combines evolutionary mutations with Bayesian allocation of simulation shots and pruning of candidates using beta-posterior credible intervals. It treats circuit simulation and decoding as a black box. Benchmarks use rotated surface codes at $d=3,5$, PyMatching, and two circuit-level noise models with bias $\eta=500$. Exhaustive evaluation of all $3^9$ reduced-set deformations establishes the distance-3 optima. Relative to fixed-shot uniform random search, EBBI reaches these optima using up to approximately $1100\times$ fewer shots. At distance 5 and the target physical error rate, median runs reach the conjectured optima; random search does not. Under an SI1000-like noise model, the search finds bordered XZZX (B-XZZX), differing from XZZX on $2(d-1)$ boundary qubits. Extending this pattern through $d=11$ yields a logical-error reduction growing from approximately $5\%$ to $12\%$. Exhaustive distance-3 comparisons also find deformation-performance correlations between PyMatching and BP-OSD as low as $0.30$. Direct searches are demonstrated only through distance 5, without optimality guarantees there. EBBI can become trapped or permanently prune the best candidate; increasing search noise reduces sampling costs but worsens convergence in some settings. Larger-distance advantages, their mechanism, and performance under spatially nonuniform device noise remain unresolved. The random-search baseline also has sampling-noise and early-termination limitations.
Error found: Algorithm EBBI adds a candidate's cumulative shot count to the global total after resampling, rather than adding only the newly executed shots, producing an incorrect shot-budget stopping rule. The main text specifies 1.25 S_i additional shots per sampling step, while Algorithm EBBI specifies S_i additional shots. The stated mutation procedure does not imply the claimed 80% single-qubit mutation rate: conditional on a nonzero initial change, the rates are approximately 87.5% and 86.8% for the two benchmark sizes.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.