Quantum error correction assumes errors are random and independent. Real errors are not. Coherent errors — systematic rotations from miscalibrated gates, correlated drifts from shared control lines — accumulate constructively rather than averaging out. The standard response: correct errors actively and frequently to prevent accumulation.
Beale and colleagues (arXiv:2602.21076) show that doing less works better. Virtual Pauli frame updates — passive bookkeeping that tracks errors without physically correcting them — outperform active correction for coherent errors. The lazy strategy matches the performance of idealized independent noise models, even when the physical noise is highly correlated.
The mechanism: active correction applies physical operations that interact with the coherent error structure, potentially amplifying correlations across rounds. Passive correction simply relabels the Pauli frame in classical software, applying no physical operation at all. The absence of physical intervention prevents the correction from coupling to the error's coherent structure. The correction works by not touching the system.
For single-qubit Hamiltonian noise beyond distance 3, perturbation theory shows that virtual corrections decohere the error process — they randomize the Pauli frame, which breaks the temporal correlations that coherent errors rely on to accumulate. Within a single syndrome extraction round, correlations may build up. But across rounds, the frame randomization from passive correction disrupts them.
The result inverts the standard intuition about error correction aggressiveness. For independent errors, active correction is fine — there are no correlations to amplify. For coherent errors, active correction risks feeding the correlation structure. Passive correction starves it. The optimal response to correlated noise is not more intervention but less.
The general observation: when the problem has internal structure (correlations, coherence, systematic bias), aggressive correction can couple to that structure and make things worse. Doing nothing — or doing the minimum — avoids creating feedback between the correction and the error. Sometimes the best response to a structured problem is an unstructured one.