Quantum computing · Experiment

Resolving and Suppressing Temporally Correlated Noise in Quantum Error Correction

Alec Eickbusch, Nick Noll, Michael Newman et al. (Google Quantum AI and Collaborators)

arXiv:2611.21578v1 ↗PDF ↗Submitted 16 November 2026

Question
Memory experiments, the standard benchmark of quantum error correction, mainly probe errors spread across many qubits at one time. Logical operations such as lattice surgery are also exposed to errors that repeat at one place over many cycles. How well does a superconducting processor suppress such timelike errors as surface-code patches grow, and what limits that suppression over long runs?
Result
On Google's 105-qubit Willow processor, surface-code stability experiments from \(3\times3\) to \(7\times7\) patches give a timelike block suppression factor \(\mu_T=1.47\pm0.25\), and together with memory experiments a spacetime factor \(\mu_{\mathrm{ST}}=1.42\pm0.03\) at an optimal duration of about \(d+1\) to \(d+2\) cycles. After roughly ten cycles the logical error falls only by a factor of about 1.1 per cycle; a leakage model with no free parameters and direct four-level readout trace this to data qubits heated into the third excited state \(\ket3\), which the routine leakage removal does not reset. A walking circuit that resets every qubit periodically removes this floor and reaches a logical error probability of at most \(3.2\times10^{-8}\) at 60 cycles. In offline analysis, stopping when the decoder is confident cuts spacetime error by 38% and mean duration by 34%.

The paper in detail

In a memory experiment a surface-code patch stores a logical qubit, and a logical error is a chain of faults stretching across the patch at roughly one time. The stability experiment is its dual. The patch has only one type of boundary, so the product of all stabilisers of that type is fixed to \(+1\); that product is the observable \(L_0\), and it can only be wrong if a chain of measurement faults stretches through time. With \(c\) rounds of stabiliser measurement, the experiment corrects up to \(\lfloor(c-1)/2\rfloor\) such faults, so its fault distance is the number of cycles. Lattice surgery and logical moves contain regions sensitive to both directions, which is why the authors treat the two experiments on equal footing.

Scaling stability patches from \(3\times3\) to \(7\times7\) on Willow shows two features. At a fixed number of cycles the logical error grows with patch width, roughly as \(w^2\), because a larger patch offers more places for a timelike chain. Against cycles, the error first falls steeply, as a Pauli noise model predicts, and then after 6 to 9 cycles bends onto a shallow slope of about 1.1 per cycle for every patch size. Combining memory and stability errors into a spacetime error, the authors find its minimum at \(c^*=4,5,7,7,9\) cycles for widths 3 to 7, longer than the usual rule of thumb \(c=d\), and an error budget in which readout and reset weigh 4.3 and 2.4 times more heavily for stability than for memory.

A simple model explains the knee: a defect that appears rarely and persists for several cycles randomises neighbouring stabilisers, and if it lasts about half the experiment it causes a logical fault. The authors identify the defect as leakage of a data qubit to its third excited state \(\ket3\). Leakage removal on Willow resets \(\ket2\) every cycle but has no mechanism for \(\ket3\), which is populated by passive heating and cleared only by relaxation or by transport to a measure qubit during a CZ gate. A device-level model with measured heating and transport rates and no free parameters reproduces the stability curves for all patch sizes out to 70 cycles, and the median detector firing rate (7.56% simulated against 7.26% measured). Reading the data qubits with four-level readout at the end of 14-cycle runs, 11 of 21 flagged failure events end with a qubit in \(\ket3\) and none in \(\ket2\).

Two remedies follow. A walking circuit alternates every qubit between data and measure roles, so every qubit passes through a multi-level reset, which clears \(\ket3\), every two cycles. It is worse than the standard circuit for short runs but removes the knee, reaching at most \(3.2\times10^{-8}\) at 60 cycles, and its remaining floor comes from a single ionising-radiation burst. The second remedy uses the freedom of a stability experiment to choose its own end: monitoring the decoder's complementary gap, a log-likelihood ratio between the two best explanations with different outcomes, and stopping once it is large lets the experiment end early after clean cycles and wait out bursts. Offline, this reduces timelike error more than a hundredfold at a mean hold of 12 cycles; on the \(7\times7\) patch, half the shots finish within 5 cycles and 99% within 18.

Transferable methods

  1. Stability experiments as timelike error metrology

    In a stability patch the observable is a product of stabilisers that should always be \(+1\), so only chains of faults stretched across time can flip it. The number of cycles plays the role of code distance, and the shape of the logical error against cycles separates short-lived Pauli noise from rare defects that persist.

    Under independent Pauli noise the stability error falls exponentially with cycles, \(P\approx a_1s_1^{-c}\). A defect that appears with probability \(p_{\mathrm{corr}}\) per qubit per cycle and randomises its neighbours causes a logical fault if it persists for at least \(\lfloor(c+1)/2\rfloor\) cycles, so its contribution scales as \(w^2p_{\mathrm{corr}}\Pr(\text{lifetime}\ge c/2)\). The sum \[P_L(c)=a_1s_1^{-c}+a_2s_2^{-c}\] then has a knee where the two terms cross, at \(c=\ln(a_1/a_2)/\ln(s_1/s_2)\). Editorial check: the fits in Table 1 put the crossings at 8 to 10 cycles.

    Editorial caution: if lifetimes were exponential with mean \(\tau\), the tail would give \(s_2=e^{1/(2\tau)}\), and the measured \(s_2\approx1.1\) would mean \(\tau\approx5\) cycles, shorter than the 10 to 15 cycles over which single-shot bursts are seen. The exit from \(\ket3\) combines relaxation with transport during CZ gates, so the simple exponential picture is only qualitative; the paper's quantitative agreement comes from the full multi-level model.

    The spacetime error averages the two experiments, \(P_{\mathrm{ST}}(c)=\tfrac12[P_M(c)+P_S(c)]\). Memory error grows about linearly with cycles while stability error falls exponentially. Editorial step: with \(P_M\approx c\,\varepsilon_M\) and \(P_S\approx a_1s_1^{-c}\), the minimum satisfies \(\varepsilon_M=a_1\ln s_1\,s_1^{-c^*}\), so \[c^*=\frac{\ln(a_1\ln s_1/\varepsilon_M)}{\ln s_1},\] which grows as the memory error per cycle shrinks with distance, consistent with \(c^*\) rising from 4 to 9.

    How to reuse it

    Any protocol whose failure needs a run of faults in time, such as repeated syndrome extraction, flag circuits or repeated readout, can use its duration as a dial: plotting failure against duration and looking for a change of slope reveals rare long-lived defects whose rate is far too small to see in one-shot benchmarks.

  2. Walking circuits and stopping on decoder confidence

    Leakage that the reset protocol cannot reach persists until something removes it. Walking the patch, so that every physical qubit alternates between data and measure roles, sends each qubit through a reset that clears all excited states every two cycles. And because a stability measurement can choose its own end time, a burst can be waited out instead of decoded through.

    The walking circuit alternates two cycle types that shuffle logical information between data and measure qubits, so every qubit is measured and reset with multi-level reset periodically, and a \(\ket3\) excursion lasts at most about two cycles instead of 10 to 15. The price is a larger circuit, about ten times worse in the Pauli regime at 10 cycles. At 60 cycles one failure occurred among \(n=3.2\times10^8\) windows. Editorial check: the likelihood-ratio interval with ratio 1000 solves \(np-\ln(np)-1=\ln1000\), giving \(np\approx10.2\) and \[p\le3.2\times10^{-8},\] at \(\sqrt{2\ln1000}\approx3.7\sigma\).

    For adaptive stopping, the decoder's complementary gap is the log-likelihood ratio between the most likely error set and the most likely one that predicts the other outcome. A gap near zero is a coin flip; the experiment continues until the gap exceeds a threshold that depends on the syndrome (21 nats in the example shot). This is a sequential test, trading mean duration against error: the mean stopping time falls by 34%, spacetime error by 38%, and timelike error by more than a factor of 100 at a mean hold of 12 cycles.

    How to reuse it

    When an error mechanism escapes the routine reset, schedule every physical element to pass periodically through the operation that does remove it. When a protocol's length is free, stopping on the decoder's own confidence, as in sequential hypothesis testing, can buy large error reductions at a modest average cost, provided the hardware can act on decoder output quickly.

Source: Sections II–IV (stability regimes, spacetime error, defect and leakage models), Sections V–VI (walking circuit, adaptive stopping). Based on the linked paper version. Derivations marked as editorial, and suggested uses in other problems, are editorial guidance rather than results of the paper.

Context & limitations

Why it matters. The stability experiment is the time-direction counterpart of the memory experiment, and this is its first scaling to a full device and hundreds of cycles. It exposes a specific error, leakage to \(\ket3\), that memory benchmarks did not reveal and that the authors' model predicts would overtake spatial errors near distance 11 to 13 if left alone. Because its fault distance is the number of cycles rather than the number of qubits, a \(3\times3\) patch can probe error floors near \(10^{-8}\), which makes it a cheap diagnostic before scaling up.

Limitations. The \(6\times6\) and \(7\times7\) patches ran in the \(Z\) basis only, and \(\mu_T\) carries a large uncertainty against a Pauli-model prediction of \(1.97\pm0.44\). The walking result rests on one logical failure among \(3.2\times10^8\) windows drawn from 999-cycle runs, so \(3.2\times10^{-8}\) is an upper bound at a likelihood ratio of 1000, and the walking circuit is about ten times worse at short durations. Adaptive stopping was analysed offline; in real time it needs decoder decisions within about a microsecond, and long tails in the hold time would complicate compilation. The crossover near distance 11 to 13 is a model extrapolation. The supplementary material, which specifies the leakage model and calibration, was not read. The paper contains no statement about AI use.

Description based on the full arXiv HTML of v1, read 2026-11-18: Sections I–VII of the main text with all figure captions and Table 1. Supplementary sections S1–S10 were not read. Editorial checks: one failure in \(3.2\times10^8\) windows gives a likelihood-ratio upper limit of \(3.198\times10^{-8}\) at ratio 1000, which corresponds to \(3.72\sigma\), matching the stated bound; the two-exponential fit of Table 1 for the \(3\times3\) patch predicts \(1.7\times10^{-6}\) at 60 cycles, 54 times the walking upper bound and 550 times its point estimate of \(3.1\times10^{-9}\); the crossings between the two fitted exponentials fall at 8 to 10 cycles across patch sizes; and the post-knee amplitudes \(a_2\) grow roughly as \(w^2\), with \(a_2/w^2\) between \(4.2\times10^{-5}\) and \(7.2\times10^{-5}\).