Comparing thresholds of XZZX and rotated surface codes under spatio-temporal correlated noise

Abstract

While independent and identically distributed noise is typically assumed in quantum computing research, complex error models such as correlated and spatio-temporal errors remain underexplored. In this paper, the performance of $\mathrm{XZZX}$ and rotated surface codes is evaluated under these complex noise models. Standard graph-based decoders struggle with the hyperedges generated by correlated errors on bias-tailored codes. To address this, we employ the Minimum Weight Parity Factor (MWPF) decoder, which natively processes syndrome hypergraphs without artificial edge decomposition. Our simulations show that under strong Z-biased circuit-level correlated noise ($η=100$), the threshold of the rotated code degrades to $\approx 0.36\%$, while the $\mathrm{XZZX}$ code drops to $\approx 0.70\%$. Furthermore, a distinct non-monotonic threshold dependence is observed for the $\mathrm{XZZX}$ code across different bias ratios. While the $\mathrm{XZZX}$ code effectively absorbs high-density Z-type hyperedges at extreme biases, it exhibits a threshold minimum at moderate biases ($η\approx 10$) because transverse X and Y errors actively bridge the diagonal strings, breaking the code's one-dimensional decoupling symmetry. These findings demonstrate that utilizing hypergraph-aware decoders is critical, as the performance of bias-tailored codes is highly sensitive to the geometric structure of spatio-temporal correlated noise.