Syndrome-Resolved Decoding of Qubit Readout Errors

AI summary

Overview

• Research area: Superconducting qubit readout and quantum measurement characterization (experimental quantum physics / quantum computing hardware).
• Technical level: Intermediate. The protocol is conceptually simple (repeated binary readout plus known gates plus pattern counting), but the surrounding discussion assumes familiarity with qubit readout fidelity, QNDness, and leakage.
• Scope (1 sentence): The paper introduces and experimentally validates the Readout Error Decoder (RED), a protocol that turns repeated thresholded single-shot readout records into a channel-resolved budget of assignment, bit-flip, and short- and long-lived leakage errors, demonstrated on two fluxonium qubits and one transmon.

What This Paper Is About

Readout errors in superconducting qubits come from several physically distinct processes: incorrect state assignment, bit flips inside the computational subspace, and leakage out of it. Standard metrics such as measurement fidelity and repeatability collapse all of these into single numbers, so devices with similar scores can fail for completely different reasons. The paper's goal is to recover the individual error channels from nothing more than repeated binary (0/1) readout records interleaved with known gates, without needing to resolve the leaked state directly in the IQ plane.

Key Contributions

1. The RED protocol: A decoder that interleaves repeated readouts with a fixed gate chosen from {I, X, X-τ-X}, computes a correlation record between the expected and measured bit strings, and groups runs of broken correlations into syndromes S₁, S₂, and S₃₊. These syndromes map onto bit-flip, assignment, and long-lived leakage errors respectively, while short-lived leakage is extracted as the excess S₁ under X-τ-X relative to the I baseline.
2. A complete decoder dictionary: Five measurement sequences (I prepared in |g⟩, I prepared in |e⟩, X, X-τ-X prepared in |g⟩, X-τ-X prepared in |e⟩) with branch-resolved (|g⟩ vs |e⟩) estimators for each channel, summarized in the paper's Tables 1 and 2.
3. Quantitative validation by controlled error injection: A π₀₁ pulse applied with programmed probability P_b produces a bit-flip rate that tracks P_b with near-unit slope while the assignment channel stays nearly flat; a π₁₂ pulse applied with probability P_f produces a leakage signal with a fitted slope of 0.887.
4. Cross-platform, cross-regime demonstration: RED is applied to Fluxonium 1, Fluxonium 2, and a transmon, covering both long-lived leakage (transmon) and short-lived leakage (fluxonium), with an X-τ-X delay sweep yielding a leakage lifetime τ_leak = 0.44 μs.

Main Findings

• Similar totals, different physics: Fluxonium 1 and Fluxonium 2 have similar total RED errors (1.79% and 1.88%) but different compositions. On Fluxonium 1, bit-flip and short-lived leakage contribute comparably; on Fluxonium 2, bit flips dominate. The transmon has a larger total (5.11%) and additionally shows long-lived leakage.
• Channel-resolved error budget at the optimal readout duration (Table 3): Fluxonium 1 at t_RO = 11 μs: assignment 0.11%, bit-flip |e⟩ 1.13% / |g⟩ 0.46%, short-lived leakage |e⟩ 1.54% / |g⟩ 0.23%, total 1.79%, versus measurement infidelity 0.28% and QNDness error 0.72%. Fluxonium 2 at 6.0 μs: assignment 0.10%, bit-flip |e⟩ 1.94% / |g⟩ 1.27%, short-lived leakage |e⟩ 0.03% / |g⟩ 0.31%, total 1.88%, versus measurement infidelity 0.53% and QNDness error 2.05%. Transmon at 6.5 μs: assignment 1.10%, bit-flip |e⟩ 2.90% / |g⟩ 1.51%, long-lived leakage |e⟩ 3.49% / |g⟩ 0.12%, total 5.11%, versus measurement infidelity 5.43% and QNDness error 4.41%.
• Assignment errors fall with readout time: On both Fluxonium 1 and the transmon, the S₂-derived assignment rate decreases rapidly with integration time as SNR improves, and follows a Gaussian-overlap model within the fitted range.
• Bit flips grow with readout time: Branch-resolved bit-flip rates grow approximately linearly over the fitted range, yielding T↑/↓ = 2990/1230 μs for fluxonium and 532/346 μs for the transmon, i.e. T_q = 872 μs and 210 μs. These are shorter than the independently measured T₁ = 2130 ± 360 μs and 274 ± 30 μs, which the authors attribute to the driven readout opening additional bit-flip channels beyond idle T₁.
• Two leakage regimes, one decoder: On the transmon, long-lived leakage appears as S₃₊ on the X sequence and grows with readout duration from |e⟩ while the |g⟩ contribution stays below a reference threshold set at 10% of the largest decoded error channel. On Fluxonium 1, S₃₊ stays at the noise floor, but the excess S₁ of X-τ-X over I reveals short-lived leakage from |e⟩ while the |g⟩ contribution stays near the noise floor.
• Leakage lifetime measurement: Sweeping the inter-pulse delay τ (fixed τ = 1.5 μs in the duration sweep) gives an exponential fit with τ_leak = 0.44 μs for the |e⟩ branch; the |g⟩ branch S₁ rate is almost independent of τ, indicating bit-flip dominance there.
• Branch asymmetry is device-dependent: Fluxonium 2 shows more nearly balanced |g⟩ and |e⟩ leakage contributions, so readout-induced leakage depends on qubit and readout parameters.
• Known limitation quantified: Simulations with known error rates show the dominant channel is recovered accurately over a broad parameter range, but the absolute accuracy of the smaller channel degrades when the two rates are strongly imbalanced; the qualitative hierarchy remains informative.

Methodology in Plain English

The qubit is measured many times in a row, and each measurement is reduced to a single bit (0 or 1) by a threshold. Between consecutive measurements the team applies a known operation: nothing (I), a bit flip (X), or a flip-wait-flip (X-τ-X), waiting at least 10/κ after each readout so the resonator has rung down before the next gate. Because the starting state and the gates are known, the expected bit string is known too. The decoder compares how the measured string changes from bit to bit with how it should have changed, marking each disagreement as a broken correlation.

The pattern of consecutive broken correlations identifies the culprit. A bit flip changes the qubit state, and later measurements remain consistent with the new state, so it produces a single isolated break (S₁). An assignment error corrupts exactly one reported bit, breaking the correlation on both sides of that readout (S₂). Long-lived leakage keeps the qubit out of the computational subspace for several cycles, producing three or more consecutive breaks (S₃₊). Short-lived leakage decays back before the next readout, leaving no persistent signature, so the team uses the flip-wait-flip sequence: if leakage made the first flip fail and the population returns during the wait, the second flip succeeds and the event shows up as an extra S₁ compared with the plain I sequence. Counting first-error events and normalizing by the number of valid readouts preceding the first error gives error probabilities, which the decoder dictionary converts into per-channel rates. Each sequence used 50 consecutive readouts repeated 500 times per preparation branch; the injection experiments used 100 consecutive readouts with at least 20 random masks and at least 500 shots per mask.

Why This Matters

Scalar benchmarks hide the mechanism behind a readout error, and the mechanism determines the fix. RED extracts the mechanism using hardware and data that experiments already produce, and it does so without resolving non-computational states in the IQ plane or requiring readout that is fast compared with the leakage lifetime. The paper shows directly that measurement infidelity and QNDness error at the same operating point can obscure leakage entirely, since a binary discriminator can map leaked population onto the same outcome.

• Quantum error correction: QND readout is a prerequisite for syndrome extraction, and knowing whether residual errors are assignment, bit-flip, or leakage determines whether the fix is better discrimination, faster integration, or leakage reduction.
• Real-time feedback and conditional control: Feedback protocols inherit readout-induced back-action; channel-resolved budgets tell hardware teams which back-action to suppress.
• Measurement-based quantum computation: Cluster-state protocols depend on destructive but clean single-qubit measurements, where leakage and bit flips have different consequences.
• Device benchmarking and calibration across platforms: The same decoding logic is proposed as adaptable to semiconductor spin qubits, trapped ions, and neutral-atom arrays, since it needs only repeated binary single-shot readout, known inter-readout operations, and errors with characteristic temporal patterns.

Industry relevance: Readout is one of the dominant overheads in superconducting processor operation, and channel-resolved error budgets give hardware and control teams a diagnostic that separates readout-duration trade-offs (assignment versus bit flip) from leakage, which is directly actionable for calibrating commercial and laboratory quantum processors.

Future Directions

• Faster readout: The measurements reported here do not use parametric amplification, so readout durations are typically several microseconds. Adding parametric amplification to reach the sub-microsecond regime would resolve how individual error channels develop during measurement.
• Microscopic origins: Combining channel-resolved dynamics with sweeps of readout power and frequency could constrain the physical sources of readout-induced errors, including intrinsic multi-photon transitions and processes involving two-level systems or spurious electromagnetic modes.
• Extension to other qubit modalities: Adapting RED to semiconductor spin qubits, trapped ions, and neutral-atom arrays, with syndrome definitions tailored to each platform's error processes.
• Syndrome ambiguity: Different error processes can produce the same syndrome and overlapping errors introduce higher-order contamination; improving how the decoder handles strongly imbalanced rates and correlated multi-error events remains open.

Target Audience

Experimentalists and theorists working on superconducting qubit readout, quantum error correction, and qubit benchmarking; hardware and control engineers who calibrate quantum processors; and researchers on other platforms (spins, ions, neutral atoms) interested in a readout-diagnostics protocol that requires only repeated binary single-shot measurements and known inter-readout gates. Readers need basic familiarity with qubit readout, fidelity, T₁, and leakage to follow the results, but the protocol itself is approachable for graduate students.

Authors’ abstract

Readout errors in superconducting qubits include incorrect state assignment, bit flips within the computational subspace, and leakage to non-computational states. These errors require different mitigation strategies, yet conventional measurement fidelity and repeatability metrics do not resolve their individual contributions. Previous repeated-readout studies have shown that temporal correlations reveal additional information about readout errors. Here we introduce the Readout Error Decoder (RED), which converts repeated thresholded binary records into a channel-resolved error budget. Controlled error injections validate the channel-specific response of the decoder. We demonstrate RED on both fluxonium and transmon qubits and across distinct leakage-lifetime regimes. RED identifies the error channels limiting readout performance without requiring dedicated discrimination of non-computational states, providing a basis for targeted readout optimization across different qubit modalities.