Quantum Error Correction Explained: Surface Codes and Magic States

Quantum error correction protects a logical qubit by spreading its state across several physical qubits and repeatedly measuring checks that reveal errors without revealing the encoded information. It works only when the checks, control operations and decoding are reliable enough to outweigh the errors they introduce.

The latest change is specific. Google’s hardware demonstration of magic state cultivation, first posted in December 2025, reached Science on 15 October 2026. It improves the preparation of a resource needed for computation. To understand why that matters, separate three tasks: protecting a state, operating on it and supplying the harder operations an algorithm needs.

Why a quantum backup is different

A classical repetition code can protect a bit by storing three copies and taking a majority vote. An arbitrary unknown qubit cannot be protected by making three independent perfect copies. Quantum encoding instead distributes its information into a joint state.

The right-hand side is one encoded state, not three copies of the original superposition. In this simple code, measuring all three qubits in the computational basis would reveal the logical bit and erase the relative phase information. Measuring an appropriate parity check can instead leave the encoded amplitudes intact.

This distinction, developed in early work including Shor’s quantum-memory code, is more precise than saying that “looking at a qubit always destroys it.” A measurement’s effect depends on what is measured and the state being measured. Error correction deliberately chooses measurements that do not distinguish the logical alternatives.

A worked example: find a bit flip without reading the qubit

Use the three-qubit encoding above and suppose at most one bit flip occurs. Measure two checks: Z₁Z₂ and Z₂Z₃. Each returns +1 for matching computational-basis values and −1 for different values. Both components of the undamaged encoded state give the same pair of answers.

If the middle qubit flips, the state becomes α|010⟩ + β|101⟩. Both checks now return −1. That identifies the middle qubit under the single-bit-flip assumption, while revealing neither α nor β. Applying another bit flip to that qubit restores the original encoded state.

Single-bit-flip syndrome table

On a narrow screen, swipe the table sideways. The first column stays fixed.

The table is a deliberately limited teaching example. Two bit flips can imitate a different single flip, and this code does not protect against phase errors. Full quantum codes must protect both kinds of information. Steane’s work helped establish constructions that do so.

What a syndrome tells you

The collection of check outcomes is an error syndrome. An auxiliary, or ancilla, qubit can interact with data qubits and then be measured to extract one check. Across repeated rounds, changes in these outcomes help distinguish data errors from faulty measurements.

A stabilizer-code description makes this systematic. Correctable errors take the encoded state into distinguishable syndrome sectors, without exposing the protected amplitudes. A general single-qubit error can be expanded in the operators I, X, Y and Z. A code that corrects the relevant basis errors also corrects their linear combinations.

This does not mean every noisy operation simply becomes one of four identifiable outcomes. Different errors can share a syndrome, and some act as an undetected logical error. The decoder chooses a recovery consistent with the observations and a noise model; its inference can fail.

How surface codes turn more qubits into better protection

A surface code arranges data qubits and local checks on a two-dimensional layout. Complementary checks diagnose bit and phase errors. Local connectivity is attractive because each measurement involves a small group of nearby qubits.

The code distance d is the minimum weight of a nontrivial logical Pauli operator. Under ideal syndrome extraction, a distance-d code corrects up to ⌊(d − 1)/2⌋ arbitrary physical-qubit errors. Real fault-tolerant operation must also account for errors during gates and measurements.

A common rotated surface-code patch uses d² data qubits and d² − 1 check qubits: about 2d² in total before routing or other workspace. That is an illustrative layout, not a universal conversion between physical and logical qubits.

Below the relevant error threshold, increasing distance can reduce the logical error rate. Above it, extra operations can create more trouble than the larger code removes. The threshold depends on the code, circuit, noise and decoder. It is not one accuracy percentage that applies to every machine.

What Willow’s below-threshold memory actually showed

Google’s surface-code memory paper appeared online on 9 December 2024, before its 2025 Nature issue. Increasing the code distance by two reduced logical error by a factor of 2.14 ± 0.02 in the reported experiment.

The distance-7 implementation used 101 qubits and reached a logical error probability of 0.143% ± 0.003% per correction cycle. Its logical memory lifetime exceeded that of the best constituent physical qubit by a factor of 2.4 ± 0.3. Those are memory benchmarks, not error rates for a complete algorithm.

The same work demonstrated real-time decoding at distance 5, with an average latency of 63 microseconds and a correction-cycle time of 1.1 microseconds. Latency and throughput are different: processing can be pipelined, so every answer need not arrive within one cycle. The system must still avoid an ever-growing backlog and deliver decisions when subsequent operations require them.

Why a reliable memory still needs magic states

A memory preserves information. A computer must transform it. Many fault-tolerant constructions handle Clifford operations efficiently, but Clifford circuits with stabilizer inputs and Pauli measurements can be simulated efficiently on a classical computer. Universal quantum computing needs an additional resource.

A common choice is a non-Clifford T gate, implemented by consuming a specially prepared magic state. The state can be prepared separately, checked, and then used in a gate-teleportation procedure. It supplies a transformation beyond the stabilizer operations used for much of the error-correction machinery.

Magic state distillation combines noisy resource states to produce cleaner ones, with a cost in qubits, operations and rejected outcomes. The alternative cultivation proposal by Gidney, Shutty and Jones improves one encoded state through carefully chosen checks and code growth. Its theoretical resource estimates depend on the assumed noise and acceptance procedure.

The 2026 result: cleaner states, with most attempts discarded

Rosenfeld and colleagues’ hardware study reports a 40-fold reduction in characterised magic-state error, with fidelity 0.9999(1) and 8% of attempts retained. The parenthetical uncertainty is 0.0001. The fidelity bound comes from postselected characterisation, including filtering during preparation and measurement. It is not a full-computation success probability.

The team also transferred a state from a distance-3 colour code into a distance-5 surface code. The author manuscript does not report the same fidelity after that transfer. Submitted on 15 December 2025, it predates the October 2026 journal publication; the publication is not a new measurement.

As a bookkeeping illustration, 8% retention corresponds to 12.5 attempts per retained experimental record on average. Because selection includes characterisation, this does not establish a production rate for usable, unmeasured states. A working factory must deliver such states, with preparation time, routing and rejected attempts included in its cost. A 40-fold reduction in state error is not a 40-fold computational speedup.

Why practical cultivation costs remain under study

A separate preprint by Mehta and colleagues, submitted on 5 October 2026, examines the difference between benchmarking a resource state and delivering one that remains available for computation. Measuring the final state can help assess performance, but destroys the state being assessed.

The authors study postselection using only nondestructive information available during the circuit. In some simulated regimes, they recover earlier efficiency estimates; in others, matching the target logical error requires greater resources. The result is a modelling analysis, not a remeasurement overturning Google’s hardware data.

Our reading is that both studies can be informative at once. A hardware experiment can establish that cultivation improves state quality, while practical resource estimates still need revision. Until an integrated machine produces and consumes these states at the necessary rate, the complete computational cost remains an engineering question.

Can other codes reduce the qubit overhead?

Quantum low-density parity-check codes offer another route. In a 2024 theoretical and numerical study, Bravyi and colleagues described a memory protecting 12 logical qubits using 288 physical qubits, including check qubits, under specified circuit-noise assumptions. This was a proposed protocol with simulated performance, not a hardware demonstration of those specifications.

The advantage comes with a connectivity requirement that differs from a simple planar nearest-neighbour surface-code layout. A higher encoding rate therefore cannot be compared fairly with a surface-code patch without also matching target error, runtime, available connections and operations.

This is why a headline ratio of physical to logical qubits is incomplete. A useful comparison asks what information survives, for how long, under which operations and at what measured error rate. The best code for one hardware architecture need not be the best for another.

What still separates these results from useful computation

Quantum error correction is established theory with experimental demonstrations of protected memories and increasingly capable logical operations. It must continually counter noise and decoherence. Leakage, correlated faults, inaccurate checks and limited classical decoding can all defeat assumptions that work in a simplified model.

The unresolved task is integration: reliable logical gates, adequate resource-state supply, fast decisions and an error budget that survives an entire useful algorithm. Our assessment is that below-threshold memory and magic state cultivation address different parts of that task. Neither alone establishes practical quantum advantage.

The cultivation hardware work is now peer-reviewed; the October practical-cost analysis remains a preprint. Google’s announcement is company-authored, and the paper comes from Google Quantum AI and collaborators. The distinction between measured performance and projected system cost should remain visible as both develop.