Quantum information needs active protection
Classical error correction can copy bits and compare redundant values. An unknown quantum state cannot be copied, and direct measurement generally disturbs it. Quantum error-correcting codes instead distribute logical information across many physical qubits and repeatedly measure parity relationships called syndromes. The syndromes reveal evidence about errors without directly reading the encoded state. A classical decoder infers a likely correction or updates the interpretation of later measurements.
This process adds components and operations, each capable of failing. Error correction helps only when physical operations are sufficiently good that the protection gained exceeds the errors introduced by the larger circuit. The boundary is the threshold. Below it, increasing code distance should suppress logical error; above it, more qubits can make the result worse. Demonstrating this scaling behaviour is more informative than showing one encoded state that survives one selected experiment.
Below threshold is a scaling result
Google's Willow experiment compared surface-code memories of increasing distance. The reported logical-error suppression factor was 2.14 ± 0.02 for each distance increase of two. A distance-7 memory using 101 qubits reached 0.143% ± 0.003% logical error per correction cycle. Its lifetime exceeded that of the best constituent physical qubit by a factor of 2.4 ± 0.3. These are measured results with uncertainty, and together they show encoded performance improving with added protection.
They do not mean arbitrary algorithms can now run fault-tolerantly. The experiment primarily protected a memory, while computation requires logical gates, routing, state preparation, measurement and repeated correction across many logical qubits. Total failure accumulates over operations. A system intended to perform billions of logical operations needs a logical error per operation far below the per-cycle values demonstrated today, with resource overhead that remains buildable.
Decoding is part of the computer
Syndrome measurements arrive continuously and must be interpreted quickly enough that the control system can keep pace. Willow's real-time distance-5 decoder reported an average latency of 63 microseconds while the error-correction cycle was 1.1 microseconds. This is possible through a streaming pipeline, but future systems must manage throughput, tail latency and backlog across many codes. A decoder with excellent offline accuracy but insufficient real-time throughput cannot support a large fault-tolerant machine.
Machine-learning decoders such as AlphaQubit can improve error identification, but they introduce their own training, generalization, hardware and verification questions. Decoders must remain effective as calibration drifts, noise correlations change and code distance grows. Reports should include average and tail latency, compute resources, training data, accuracy under distribution shift and the effect on logical error. Classical infrastructure is not ancillary; it closes the feedback loop that makes the quantum system work.
Code efficiency can change the hardware equation
The surface code is attractive because it tolerates local errors on a two-dimensional layout, but it can require many physical qubits per logical qubit. A 2024 Nature paper on quantum low-density parity-check codes presented an end-to-end theoretical protocol with a 0.7% threshold in its circuit noise model. Under an assumed 0.1% physical error, the authors estimated that 12 logical qubits could be preserved for nearly one million syndrome cycles using 288 physical qubits, compared with nearly 3,000 for a surface code.
That is a modelled comparison, not a hardware demonstration. LDPC approaches can demand more complex connectivity, syndrome circuits and decoders. Architecture decisions must include wiring, fabrication, calibration, leakage, loss and control, not code rate alone. The most efficient abstract code may be difficult to realize on a particular platform. QEC is therefore a co-design problem spanning device physics, code theory, layout, cryogenics or optics, classical processing and software.
Logical-qubit counts need operational context
A neutral-atom experiment reported programmable operations with up to 48 logical qubits, including a hypercube circuit on 48 logical qubits encoded in 128 physical atoms. A Microsoft and Quantinuum preprint reported four logical qubits formed from 30 trapped-ion qubits, with entangled logical-state error rates ranging from 4.7 to 800 times lower than physical comparisons depending on code and post-selection. These results explore different regimes and cannot be ranked by logical-qubit count alone.
Post-selection is particularly important. Discarding runs with detected faults can produce high-fidelity retained results, useful for experiments and early algorithms, but lowers yield and differs from active correction that keeps computation running. A report should state the acceptance fraction, whether errors were detected or corrected, how many syndrome rounds occurred, which logical gates were used and whether the comparison to physical qubits involved equivalent circuits.
A scorecard for credible QEC progress
At the physical layer, report distributions of gate, measurement, reset, leakage and loss errors, plus stability over time. At the code layer, report distance, number of data and ancilla qubits, logical error per cycle, suppression factor and uncertainty. At the decoder layer, report accuracy, compute cost, throughput and tail latency. At the logical-operation layer, report gate set, fidelity, depth, yield and total success probability. At the system layer, report wiring, cooling, calibration and control overhead.
The strongest milestone is reproducible improvement along several axes, not a record isolated from its cost. Below-threshold memory is foundational; universal fault-tolerant computation remains further away. Research programmes should publish measured values separately from simulation and targets, test correlated and rare errors, and show how an architecture scales beyond one protected state. Quantum error correction is the metric that matters because it converts better components into a path toward longer computation, but only when the entire feedback system closes.
Scope and limitations
The experiments use different hardware, codes, noise models and post-selection policies, so direct ranking is inappropriate. The LDPC overhead comparison is theoretical and assumption-dependent. Logical memories and small encoded circuits do not establish application-scale fault tolerance. Rare correlated errors, manufacturing yield and control-system scaling remain active research areas.
References
Source review: 20 August 2026. Quantitative values retain their original definitions, periods, and boundaries.
- 01Quantum error correction below the surface code threshold
Nature · 2025
www.nature.com ↗ - 02Learning high-accuracy error decoding for quantum processors
Nature · 2024
www.nature.com ↗ - 03High-threshold and low-overhead fault-tolerant quantum memory
Nature · 2024
www.nature.com ↗ - 04Logical quantum processor based on reconfigurable atom arrays
Nature · 2023
www.nature.com ↗ - 05Demonstration of logical qubits and repeated error correction with better-than-physical error rates
arXiv · 2024
arxiv.org ↗ - 06Theory: Turning noisy intermediate-scale quantum information processing into practical quantum computing
National Institute of Standards and Technology · 2025
www.nist.gov ↗