Why quantum error recovery needs more than entanglement

Quantum information can be encoded by letting a collection of qubits interact, distributing the information across a larger system. Recovering it after an error is a separate challenge. In a preprint submitted on November 10, 2026, Zheng-Hang Sun and colleagues investigate a deliberately restricted recovery procedure and connect its performance to how closely the encoding operations imitate random quantum transformations.

Their central distinction is between producing an apparently well-mixed state and producing an encoding operation with the right statistical structure. In the models studied, the first does not necessarily deliver the second.

Recovery with a filter

The researchers examine encoding-error-decoding protocols. Some qubits carry the logical information; additional helper qubits, called ancillas, begin in a known state. An encoding operation acts on both groups. An error then intervenes, after which the protocol applies the exact inverse of the encoder.

Finally, the ancillas are measured. Only trials in which they return to their initial state are retained. This filtering step is called postselection.

Recovery is measured by fidelity: how closely the retained logical state matches its input. Crucially, this is a conditional measure. High fidelity among accepted trials does not mean that most trials are accepted.

The errors studied are coherent rotations: each physical qubit undergoes the same rotation about its quantum Z axis. This is a specific error model, rather than a survey of every disturbance a device might experience.

What counts as random enough?

The benchmark is the Haar ensemble, the mathematical distribution that treats all unitary transformations equally. Sampling from it supplies an idealized standard for random encoders.

An ensemble need not reproduce every feature of Haar randomness to match certain averages. A unitary 2-design reproduces its second-moment statistics—the averages involving two copies of an operation and two copies of its complex conjugate.

The authors derive an inequality connecting departures from Haar-like recovery to the second unitary frame potential, a mathematical measure of those statistics. A recovery discrepancy therefore witnesses a departure from 2-design behavior. The reverse conclusion is not guaranteed: matching this particular recovery measurement does not by itself certify a 2-design.

There is also an averaging distinction. The inequality concerns an analytically tractable fidelity obtained by averaging the recovery numerator and acceptance probability separately, then taking their ratio. That need not equal the average of each encoder's conditional fidelity. Numerical comparisons support close agreement for the undoped matchgate examples tested, but the quantities remain mathematically distinct.

Adding gates changes the outcome

The first examples use matchgate circuits, which describe fermionic Gaussian transformations: a structured family of operations related to noninteracting fermions.

Without additional gates, the authors analytically show that the separately averaged fidelity tends to zero as system size grows at half code rate, for any fixed error angle strictly between zero and pi. A small system can appear protected at weak errors, yet that protective window shrinks with increasing size.

They then insert non-Gaussian gates between independently randomized matchgate layers. These additions break the Gaussian restriction and bring the ensemble toward a 2-design within its symmetry constraints. Recovery correspondingly approaches the Haar-like result.

For this architecture, their analysis identifies a gate-count scale proportional to the number of qubits multiplied by its logarithm for maintaining a fixed averaged recovery discrepancy as the system grows. This is not a universal prescription for quantum encoders; it depends on the circuit family and accuracy measure considered.

Entanglement is not the whole story

The researchers also generate encoders through evolution under Hamiltonians, the operators governing a system's dynamics.

A clean, non-integrable Ising model rapidly develops substantial entanglement and thermalizes in the example studied. Nevertheless, its encoding evolution does not establish a stable error-protecting regime. By contrast, Hamiltonians drawn from a Gaussian random-matrix ensemble reproduce a Haar-like recovery transition at the evolution time examined.

An experimentally motivated all-to-all spin model provides another numerical example. Increasing randomness in its couplings brings recovery closer to the Haar benchmark, alongside a reduction in the frame-potential excess. These are calculations, not a new hardware demonstration.

The result is a diagnostic connection, not a replacement for fault-tolerant quantum error correction. The protocol assumes inverse decoding, filters measurement outcomes, and can have exponentially small acceptance probabilities. Within those limits, recovery measurements offer evidence about encoding randomness that entanglement alone can miss.