arXiv:2609.40279
Classical simulation of coherent crosstalk in surface codes
Andrew S. Darmawan, Yelyzaveta Kolesnyk, Robert Koenig
Abstract
We give a polynomial-time classical algorithm which samples from the syndrome distribution of a surface code state corrupted by coherent nearest-neighbor $ZZ$ crosstalk. Our algorithm complements the known efficient simulation algorithms for coherent single-qubit errors in the surface code. When both single-qubit coherent noise and coherent crosstalk are present, we find a complexity-theoretic obstruction to efficient simulation: unless the polynomial hierarchy collapses, there is no efficient classical algorithm for sampling from the syndrome distribution, even up to a constant multiplicative error. This is obtained by connecting the syndrome distribution under coherent noise to the output distribution of certain IQP circuits associated with a non-planar graph.
AI Summary
The paper asks when coherent nearest-neighbor $ZZ$ crosstalk in surface codes can be simulated classically without discarding interference. It considers code-capacity noise: an encoded state undergoes commuting coherent rotations, followed by ideal stabilizer measurement and Pauli recovery. For odd-distance rotated surface codes, the syndrome distribution under arbitrary edge-dependent $ZZ$ angles factorizes into distributions for two auxiliary unrotated surface codes of distance $(d+1)/2$ with single-qubit coherent $Z$ noise. Existing fermionic-linear-optics samplers therefore give exact syndrome sampling in $O(d^6)$ arithmetic operations per sample; the implementation takes about $5$ ms at $d=37$ ($1,369$ data qubits). The conditional logical action is determined solely by correction-weight parity. An even-weight recovery corrects this noise exactly at every coupling strength, whereas conventional MWPM can introduce logical errors. Simulations at $d=13,21,29,37$ find substantially larger MWPM residual errors for coherent crosstalk than for independently Pauli-twirled interactions. Finite-size fits extrapolate threshold angles $\theta_c/\pi=0.0342(4)$ and $0.0493(5)$, respectively, but poor goodness of fit prevents controlled asymptotic threshold estimates. Adding independently programmable single-qubit $Z$ rotations maps syndrome sampling to two-body IQP circuits on a king lattice. A postselected-universality construction establishes that polynomial-time sampling with pointwise relative error $\epsilon<1$ would collapse the polynomial hierarchy to its third level, already for angles in multiples of $\pi/8$. This hardness result does not cover constant-total-variation approximation or spatially uniform weak noise. Measurement errors and additional parity-breaking noise remain outside the efficient simulation result.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.