Skip to content
Quantum.world

The Signal

Google preprint links Willow’s lingering errors to third-state leakage

A Google Quantum AI preprint traces a slowdown in Willow’s error suppression to qubits heating into the third excited state. It reports that a walking stability circuit lowered the measured logical-error floor to no more than 3.20 × 10⁻⁸ at 60 cycles.

Google preprint links Willow’s lingering errors to third-state leakage

Persistent errors on Willow

A Google Quantum AI preprint reports that errors persisting across correction cycles limit error suppression on its Willow processor, and describes a circuit intended to suppress them. The study uses a 105-qubit superconducting processor to run a surface-code stability experiment, a test the researchers use to probe errors that build up over time.

Source arxiv.org

The team scaled the test from patch sizes of 3 × 3 to 7 × 7 and measured a timelike block suppression factor of 1.47 ± 0.25. The preprint also reports that extending the test across more cycles produced a slowdown in error suppression, rather than the single-exponential behavior expected for uncorrelated local errors.

Heating identified as the source

The researchers attribute the slowdown to data qubits heating into the third excited state, written as |3⟩. A qubit in that state can disrupt nearby stabilizer measurements over multiple cycles; the paper says these correlated detection patterns lasted about 10–15 cycles.

The preprint says Willow’s data-qubit leakage-removal operation is calibrated to remove leakage into the |2⟩ state, not |3⟩. It reports that a qubit in |3⟩ can remain there for roughly ten cycles, until it leaves through natural relaxation or leakage transport during controlled-Z gates, which can move it to a measure qubit for reset.

As a direct check, the team ran a 14-cycle stability experiment on a 3 × 3 patch and used a readout that distinguishes the |0⟩, |1⟩, |2⟩ and |3⟩ states. Among 21 flagged events, 52% ended with the flagged data qubit in |3⟩, while none ended in |2⟩. The authors note that an outlier reading may indicate still higher states, an issue they say remains under investigation.

A circuit that moves the encoded information

To address the error, the researchers implemented a walking stability circuit that alternates between two cycle types. The sequence shifts logical information between data and measure qubits, allowing all qubits to be read and periodically reset; with a multi-level reset that clears higher states, the procedure can also remove |3⟩ leakage.

In the reported experiment, the walking circuit reached a logical error probability floor of no more than 3.20 × 10⁻⁸ at 60 cycles. The preprint describes that result as two orders of magnitude below the standard circuit construction at the same cycle count, and says ionizing-radiation error bursts limited performance.

Adaptive stopping remains offline

The study also examined stopping error-correction cycles based on the decoder’s confidence in its result. In an offline analysis, this approach reduced spacetime logical error by 38% and time overhead by 34% compared with the optimal fixed-cycle strategy.

That analysis was not a live, adaptive-control demonstration. The authors say real-time stopping would require fast decoder decisions and feedback on the order of 1 microsecond; they also note that variable-duration operations could make algorithm compilation challenging.

Sources

  1. Resolving and Suppressing Temporally Correlated Noise in Quantum Error Correction
  2. arXiv (full text)

Related stories

Explore

More articles