Repeated readout gives you many noisy answers to one question. You can clean them up afterwards with a filter, or you can intervene inside the circuit with quantum error correction. The two approaches are often discussed as if they were interchangeable, and they are not. The comparison comes down to a handful of procedural choices that a reader can check.

The Two Ways to Fix a Noisy Qubit

Repeated readout is the simplest noise-reduction trick available. Measure the same qubit many times, keep the majority answer, and the random flips mostly cancel. The method costs time and repetition, and it lives entirely in software downstream of the hardware.

Quantum error correction works differently. It spreads one logical qubit across several physical ones, measures stabiliser operators through ancilla qubits, and applies corrections mid-circuit. The measurement record is not the final product. The correction is.

The distinction matters because the two methods fail under different conditions. A filter can only reweight outcomes it has already seen. A correction scheme can prevent an error from propagating into the next gate. When the error mechanism is correlated across qubits or across time, that difference becomes the whole story.

A related piece on this site about superconducting qubit coherence loss makes a similar point from the materials side: where the noise enters the stack determines what any downstream scheme can remove.

What the Experiment Actually Measured

The quantity of interest is the logical error rate, not the physical fidelity of any single qubit. Physical fidelity tells you how good one gate or one measurement is. Logical error rate tells you how often the encoded information is wrong after the full cycle of syndrome extraction, decoding, and correction.

Repeated readout is a convenient test bed because it produces a clean baseline. Run the same hardware, the same noise environment, and the same number of measurement shots, then compare a filter pipeline against a correction pipeline. Any difference in the reported logical error rate is attributable to the pipeline rather than to the device.

Error suppression is usually reported relative to that baseline. A factor of two is a real effect. A factor of a hundred is a different regime. The number only means something when the baseline, the shot count, and the cycle definition travel with it.

This site has argued in a piece on calibration choices in noise budgets that the accounting frame shapes the conclusion as much as the measurement does. The same caution applies here.

Why Filtering Fails at Scale

Majority voting assumes errors are independent. If each measurement flips with some fixed probability and the flips are uncorrelated, then taking the majority of an odd number of shots drives the error down exponentially in the number of shots. That assumption is doing most of the work.

Real devices violate it. A drift in the readout resonator, a cosmic-ray event, or a slow charge rearrangement can flip several consecutive measurements in the same direction. Correlated noise defeats a filter because the majority is wrong together, and no amount of post-hoc voting recovers the lost information.

Post-processing also cannot repair damage done mid-circuit. If an error corrupts a qubit before the next entangling gate, the gate operates on corrupted input. A filter that sees only the final measurement record has no access to that intermediate state.

The threshold theorem sets the real bar. Below a certain physical error rate, a correction scheme can suppress the logical error rate to arbitrarily low levels by scaling the code. Above it, adding more physical qubits makes things worse. Filtering has no analogous threshold, which is why it cannot substitute for correction in the long run.

The Procedural Choices That Matter

Syndrome extraction cadence shapes everything. Measure the stabilisers too rarely and errors accumulate between rounds. Measure too often and the measurement circuitry itself becomes a dominant noise source. The cadence has to be tuned to the device's coherence time, and that tuning is a judgement call, not a constant.

Decoder latency is the second lever. A decoder that runs in real time can feed corrections back within the coherence window. A slower decoder that processes the record offline produces a better estimate of what happened, but the correction arrives too late to help the next round. There is a genuine trade between correction quality and correction speed.

Ancilla reset fidelity is the hidden variable in most published comparisons. If the ancilla is not properly reset between rounds, syndrome outcomes leak information from previous cycles, and the decoder sees a correlated record it was not designed to handle. Many reported logical error rates are sensitive to this parameter in ways the papers do not always foreground.

Calibration drift between runs is the fourth. A pipeline compared on Monday and a pipeline compared on Friday may not be compared at all, because the device has moved. Interleaving the two pipelines within a single calibration window is the only clean way to control for it.

Where the Comparison Gets Decided

Published comparisons rarely isolate a single procedural choice. A recent study on a superconducting transmon device, for example, reported that a filter pipeline achieved a logical error rate roughly twice that of a correction pipeline under independent noise, but the gap narrowed to near parity when the noise model included correlated errors. The exact numbers depend on the device and the cycle definition, but the pattern is common: the filter's advantage evaporates once the noise is correlated.

That study also highlighted the role of decoder latency. The correction pipeline used an offline decoder that took several milliseconds per syndrome round, while the filter pipeline had no such constraint. When the authors re-ran the comparison with a real-time decoder, the correction pipeline's logical error rate improved by a factor of about 1.5, enough to widen the gap. The lesson is that latency is not a fixed property of the approach; it is a design choice that can be optimized.

Another factor is the number of physical qubits per logical qubit. A code with a larger distance can suppress errors more effectively, but it also requires more ancillas and more complex decoding. The overhead trade-off is not often discussed in filter-versus-correction comparisons, yet it determines the resource cost of achieving a given logical error rate.

What This Does Not Show

A logical error rate improvement is not a demonstration of useful computation. It shows that the encoding and correction loop preserves information better than a filter does on this device. It says nothing about whether the logical qubit can run a long algorithm, because algorithmic performance depends on gate fidelities that the readout comparison does not probe.

Logical qubit counts remain far below what fault-tolerant computation requires. Published demonstrations operate with a handful of logical qubits, and the overhead per logical qubit is large. The gap to thousands is not a matter of engineering polish.

The threshold margin is also narrow. Devices sit close enough to threshold that a modest change in noise characteristics can flip the sign of the result. A scheme that wins today can lose after a firmware update, and that fragility is a property of the regime rather than of any particular team.

Finally, the result is hardware-specific. A conclusion drawn on superconducting transmons does not automatically transfer to trapped ions or neutral atoms, where the noise channels and the measurement timescales differ substantially.

How to Read the Next Claim

Ask for the logical error rate per cycle, not per experiment. The per-cycle figure is the one that scales, and it is the one that can be compared across devices with different cycle times.

Check whether the noise model includes correlated errors. A comparison run against an independent-flip model flatters the filter and understates the correction scheme's advantage.

Demand decoder latency numbers alongside the logical error rate. A result that reports only the offline decoding performance is measuring a different system from the one that would run in a real machine.

Compare against the same baseline, on the same device, within the same calibration window. If the two pipelines were characterised on different days, the comparison is not controlled.

Look for independent replication before treating any suppression factor as a property of the approach. A related piece on behavioural nudges versus auto-enrolment shows how often a headline effect shrinks once someone else runs the same comparison.