Quantum Error Correction, Noise & Decoherence

Quantum noise includes stochastic and coherent gate errors, preparation and readout error, relaxation, dephasing, leakage, crosstalk, drift, and correlated faults. T1 characterizes energy relaxation; T2 characterizes loss of phase coherence and is constrained by relaxation as well as pure dephasing. A calibrated channel model is an approximation tied to a device, time window, pulse schedule, and experiment—not a universal description of hardware.

Quantum error correction encodes logical information redundantly without copying an unknown state. Stabilizer measurements reveal an error syndrome while preserving logical amplitudes; a decoder infers a correction or updates a Pauli frame. A distance-d code corrects up to floor((d-1)/2) arbitrary errors under its code model. Repeated syndrome extraction is itself noisy, so fault-tolerant gadgets must prevent one fault from spreading into too many errors.

Error suppression and error mitigation are not error correction. Suppression reduces physical noise exposure; mitigation estimates less-biased observables using extra samples or model assumptions but does not create a protected logical state. Fault tolerance combines codes, repeated syndrome extraction, decoding, logical operations, state preparation, measurement, and resource factories. The threshold theorem applies only when a specified noise model and complete implementation keep effective faults below a scheme- and model-dependent threshold; being below a quoted hardware number is not sufficient by itself.

Production evidence must retain physical and logical error definitions, code and distance, circuit and schedule, syndrome rounds, decoder and weights, calibration window, leakage handling, post-selection, shot accounting, confidence intervals, and resource overhead. The production invariant is logical-error accountability: improvement is claimed only when end-to-end logical failure at matched workload and success criteria is lower than the relevant unencoded baseline, with all discarded trials, decoder latency, correlated noise, and scaling assumptions visible.

The overhead of error correction is the number that decides architecture, and it is worth carrying exactly. A rotated surface code of distance d uses d squared data qubits and d squared minus one measurement qubits, so 2d squared minus 1 physical qubits carry one logical qubit: 49 at distance 5, 241 at distance 11, and 1,249 at distance 25. Below threshold, logical error falls roughly as the ratio of physical to threshold error raised to the power (d+1)/2, which in practice is summarized by a suppression factor per two units of distance. Measured results are the honest anchor here: Google's 2024 surface-code experiment reported a logical error per cycle of about 0.143 percent at distance 7 and a suppression factor near 2.14 for each increase of distance by two, which is the first demonstration of the exponential trend running in the right direction rather than a claim that an algorithmic machine exists. Extrapolating that factor tells you the price of a target: reaching a logical error rate of 10 to the minus 10 from 10 to the minus 3 requires roughly seven doublings, hence about distance 21 to 25, hence over a thousand physical qubits for each logical one before any magic-state factory is counted — and the non-Clifford gates that factories supply typically dominate the qubit budget of a real algorithm.

Two things routinely turn a promising code experiment into an unusable one, and neither is visible in a single-round fidelity number. The first is decoder latency. Syndrome extraction on superconducting hardware runs on the order of a microsecond per round, and the decoder has to consume rounds at least as fast as they are produced; a decoder that takes 2 microseconds per round falls a microsecond further behind every round, so the backlog grows without bound and the Pauli frame needed to interpret a logical measurement is never ready. This is why decoder throughput, not just decoder accuracy, belongs in the reported result, and why real-time decoding is an engineering programme rather than a post-processing step. The second is leakage: a transmon excited out of the computational subspace into the second excited state is not a Pauli error, so a decoder built on a Pauli noise model mis-corrects it, and the leaked population persists across rounds and spreads through two-qubit gates, producing exactly the time-correlated faults the threshold theorem's independence assumptions exclude. Leakage-reduction units that return the population to the computational subspace, and a decoder informed by leakage detection, are what keep the measured suppression factor from stalling as distance grows. The related discipline is to keep suppression, mitigation and correction in separate columns of any report. Dynamical decoupling and zero-noise extrapolation improve an estimated observable while consuming shots — mitigation overhead grows exponentially in circuit volume — and neither produces a protected logical qubit that could carry a state through an arbitrarily long computation.

Break-even claims need a stated baseline or they mean nothing, and the baseline is easy to choose flatteringly. A logical qubit outperforming the average physical qubit on the same chip is a weaker result than one outperforming the best physical qubit, which is weaker again than one outperforming the best physical qubit at a matched workload and matched wall-clock duration — and only the last supports the sentence teams want to write. Two adjacent distinctions matter as much. A pseudothreshold is the physical error rate at which a specific code and decoder at a specific distance stops hurting, measured on one device with one noise model; the threshold in the threshold theorem is an asymptotic property of a scheme under an assumed noise model, and the two are routinely confused in press summaries. Idle memory performance is also not gate performance: a code that preserves a logical state for a hundred rounds says nothing about logical operations, which require lattice surgery or transversal gates plus distillation, and where most of the eventual overhead lives. State the code, distance, rounds, decoder and its latency, the physical error rates measured in the same calibration window, the post-selection rate if any, and the baseline being beaten — and the claim can be checked instead of believed.