First-hand information for everyone
A long‑running stability test on a 105‑qubit superconducting processor revealed that the logical error probability does not fall with a single exponential, as expected for uncorrelated errors. Instead, after about ten cycles a knee appears and the decay rate slows. This sum‑of‑exponentials behavior is consistent with temporally correlated data‑qubit leakage into the third excited state –|3”. The leakage model, which incorporates higher‑state transport during syndrome extraction, reproduces the measured logical error probability across all patch sizes with no free fitting parameters.
| Patch Size | a1 | s1 | a2(×10−3) | s2 |
|---|---|---|---|---|
| 3×3 | 0.29±0.07 |
| 2.46±0.12 |
| 0.38±0.05 |
| 1.094±0.004 |
| 4×4 | 0.40±0.08 | 2.17±0.08 | 0.69±0.08 | 1.090±0.003 |
|---|
| 5×5 | 0.62±0.10 | 2.15±0.07 | 1.67±0.16 | 1.094±0.003 |
|---|
| 6×6∗ | 0.77±0.11 | 2.13±0.06 | 2.58±0.23 | 1.109±0.003 |
|---|
| 7×7∗ | 0.93±0.10 | 1.98±0.04 | 2.83±0.21 | 1.096±0.002 |
|---|
To combat this leakage‑induced slowdown, the team implemented a walking stability circuit that repeatedly dismantles the –|3” population through multi‑level reset on all qubits. Experimentally the walking circuit achieved a logical‑failure floor of <3.2×10−8 at 60 cycles—two orders of magnitude lower than the non‑walking configuration. The improvement is captured by the paper’s measured suppression factors: μ_T=1.47±0.25 for timelike errors, μ_S=1.41±0.02 for spacelike errors, and μ_ST=1.42±0.03 for the combined spacetime case.

The trade‑off is an overhead that raises the logical‑error probability by roughly an order of magnitude in the Pauli‑error regime, but the walking circuit eliminates the so‑called –|3” knee and reaches the low error floor within a realistic time frame.
These findings highlight a common, hard‑to‑detect source of logical errors in surface‑code memories and demonstrate a practical circuit tweak that can significantly extend memory lifetime without tuning device parameters.