Lifting: Quantum State Discrimination Framework for Mid-Circuit Measurements

AI summary

Overview

Research area: Quantum computing hardware — superconducting qubit readout, mid-circuit measurement (MCM), and quantum state discrimination (QSD).

Technical level: Advanced. The paper combines classical statistical learning theory (Bayes risk, Bhattacharyya coefficients), superconducting readout physics (dispersive readout, resonator photon effects), and hardware experiments on IBM processors. Readers will need comfort with probability densities and quantum circuit notation.

Scope (one sentence): The paper introduces "Lifting," a readout framework that couples a measured qubit to auxiliary qubits, decodes their joint I–Q signals with a neural network, and calibrates that decoder in the same circuit context as the workload, with a supporting theory of why extra readout channels can lower the Bayes risk floor.

What This Paper Is About

Mid-circuit measurements are essential for error correction and adaptive circuits, but on superconducting hardware they are noisy. The classical step that turns an analog readout signal into a 0 or 1 — quantum state discrimination — is bounded by the Bayes risk of whatever signal set it is given, so better neural networks alone cannot push error lower. The authors' goal is to change what the decoder sees (joint signals from repetition-encoded auxiliary qubits) and when it is trained (calibration captured at the workload's actual measurement points), to escape both the Bayes-risk floor of single-qubit readout and the mismatch caused by state deterioration during circuit execution.

Key Contributions

1. Characterization of within-circuit state deterioration. The authors show, on a five-cycle qubit-reuse stabilizer workload (QR-Stabilizer) on ibm_fez, that readout I–Q distributions drift further from SPAM (state-preparation-and-measurement) references as successive measurements accumulate, and that discrimination accuracy declines accordingly. They frame this as calibration context mismatch, distinct from long-term parameter drift.

2. A theory for decoding joint I–Q readout signals. Theorem IV.1 proves that, for a fixed joint distribution, retaining an additional readout channel cannot increase the optimal error R*_{2d}, and gives sufficient conditions under which an overlap-based upper bound contracts: R*_{2d} ≤ ½ ρ_d ≤ ½ ρ_1 ∏_{k=2}^{d} γ_{max,k}, which becomes geometric, ½ ρ_1 q^{d-1}, when all factors are bounded by a common q < 1. Proposition IV.2 shows a sufficiently expressive feedforward network can approach R*_{2d} within any ε > 0.

3. A practical encoding choice. The paper identifies distance-two encoding (one auxiliary qubit, d = 2) as the practical operating point, balancing the discrimination gain from an extra readout channel against the gate and readout errors introduced by implementing a larger encoding.

4. Workload-level hardware validation. Experiments across three IBM processors (ibm_fez, ibm_kingston, ibm_marrakesh) evaluate the framework on repeated stabilizer measurement, parity-check (QEC-5, five rounds), and teleportation workloads.

Main Findings

• Joint readout halves the estimated risk at d = 2. On ibm_fez, using 10,000 shots per preparation label, the equal-prior kernel density estimation (KDE) risk estimate falls from 14.7% for the target channel alone to 5.9% for the joint target-plus-auxiliary readout.

• Neural joint decoding beats hard repetition voting at distance two. Relative error reductions over distance-two hard repetition decoding are 80.4% on ibm_fez, 82.5% on ibm_kingston, and 79.8% on ibm_marrakesh. Curves for the two methods coincide at d = 1.

• Hard voting scales poorly and irregularly. The binary repetition comparator loses noticeably at d = 2 and shows no consistent gain from d = 3 to d = 4; at even distances ties are possible and their resolution changes the outcome.

• Lifting generally exceeds a temporal (repeated-measurement) comparator. Both neural schemes receive 2d scalar inputs; they start coincident, and are near-tied on ibm_fez at the largest distance. The authors note this compares two implemented schemes with different circuit and acquisition conditions, not a single isolated noise mechanism.

• in situ calibration helps, but only partially. Comparing SPAM-trained and cycle-matched single-channel decoders on QR-Stabilizer target shots, the improvement is concentrated in early measurements, while later-cycle errors remain high — motivating the addition of extra readout information rather than calibration changes alone.

• The empirical cost model fits the hardware trend. A phenomenological product model with fitted parameters κ = 0.45 and α = 9.03 tracks measured assignment accuracy versus encoding distance on ibm_fez.

• Workload-specific accuracy and success results are not available in the supplied text. The content is truncated in Section V.3 (per-cycle QR-Stabilizer results), so the reported accuracy for QR-Stabilizer, QEC-5, and teleportation workloads, the specific confidence-filter thresholds and retained-shot coverages, and the exact calibration shot counts are not reported here.

Methodology in Plain English

The approach has three moving parts.

Encoding. Just before a target qubit is measured, the protocol borrows d − 1 nearby qubits, initializes them to |0⟩, and applies CNOT gates from the target to each. This copies the target's computational-basis value across several qubits, producing the state a|0⟩^⊗d + b|1⟩^⊗d. All d qubits are then read out simultaneously with multiplexed dispersive readout, yielding 2d numbers per shot (an I and Q value for each qubit).

Decoding. Rather than thresholding each qubit and taking a majority vote, the framework concatenates all 2d values into one feature vector and feeds it to a small neural network — a multilayer perceptron with 2d input nodes, ReLU hidden layers (one hidden layer of 32 neurons in the MCM experiments), and a sigmoid output. It is trained with binary cross-entropy. The intuition is that disagreement patterns and analog confidence levels across the auxiliary channels carry information that hard 0/1 votes discard.

Calibration in context. Instead of training on isolated state-preparation-and-measurement circuits, the calibration circuits reproduce the workload's circuit depth, spectator qubit activity, and measurement timing, and place the encoded readout at the same measurement locations. This produces labeled reference data whose I–Q distributions better match what target shots actually look like mid-circuit.

An optional confidence filter rejects shots whose analog observations sit too close to the decision boundary — using a Gaussian likelihood score for the QR-Stabilizer workload and a neural posterior score for QEC-5 and teleportation — trading the number of retained shots for higher accuracy among those retained.

Why This Matters

Impact on research. The paper reframes readout improvement as an information problem rather than a model capacity problem. Because Bayes risk is a hard floor for any classifier on a fixed observation space, the result gives a principled argument for spending auxiliary qubits and gates to expand that space, plus explicit conditions under which the benefit compounds geometrically with distance. The distinction between calibration mismatch and irreducible distributional overlap is also a useful diagnostic separation for the readout community.

Real-world applications:

• Quantum error correction — repeated syndrome extraction depends on many mid-circuit measurements per logical cycle; reducing per-measurement assignment error directly lowers logical error rates.
• Dynamic / adaptive circuits — outcome-dependent gate selection and correction-frame updates propagate readout errors into subsequent operations, so more reliable discrimination limits error amplification.
• Qubit-reuse architectures — resetting and reassigning measured qubits reduces footprint but exposes exactly the state-deterioration effect this paper characterizes.
• Real-time control hardware — the compact MLP input structure is compatible with the FPGA-based, low-latency inference line of work the paper cites.

Industry relevance. The experiments run entirely on IBM superconducting hardware, and the framework is designed as a drop-in change only at measurement locations, leaving the rest of the circuit untouched. The finding that distance two is the sweet spot is directly actionable for hardware teams deciding how many ancillas to spend per readout.

Future Directions

• Closing the later-cycle gap. In in situ calibration alone, later-cycle errors stayed high. The truncated portion of the paper presumably tests whether joint decoding closes that gap, but the mechanism of deterioration — relaxation, measurement backaction, gate errors, residual resonator photons, or ancilla errors — is explicitly not isolated by the present data, and the authors say so.
• Separating the encoding benefits from implementation costs. Theorem IV.1 compares coordinates of a fixed joint distribution, whereas physically increasing distance also changes gates and measurement conditions. The authors note that separately executed distance sweeps need not share the required marginals, so a cleaner experimental design would be needed to validate the conditional-overlap assumptions directly.
• Scaling beyond distance two. Larger distances give smaller or device-dependent gains; whether better encodings, better ancilla initialization, or cavity-depletion techniques can shift the optimum to higher d is open.
• Confidence filtering as a system-level knob. The trade between retained-shot coverage and accepted-outcome accuracy is a policy question for applications — for instance, whether a QEC decoder can tolerate erased shots more cheaply than misassigned ones. The paper leaves the operating-point analysis to future work, and the specific thresholds and coverage numbers are not in the supplied text.

Target Audience

Quantum hardware and architecture researchers working on superconducting readout and mid-circuit measurement; quantum error correction engineers who care about per-measurement fidelity budgets; machine-learning researchers interested in learned decoders for physical-layer signals; and graduate students with a background in quantum information and classical statistical learning who want a worked example of connecting a Bayes-risk argument to a hardware experiment. The early sections are accessible to readers who know dispersive readout; the theorem and its proof sketch require comfort with Bhattacharyya coefficients and conditional densities.

Authors’ abstract

Quantum computing promises new capabilities beyond classical computation. Mid-circuit measurements (MCMs) enable error correction and outcome-dependent control, but their reliability is limited by noise on superconducting hardware. These operations depend on accurate quantum state discrimination (QSD), which assigns measured signals to state labels. Neural network-based (NN-based) QSD improves this assignment, yet its error remains bounded below by the Bayes risk of the available readout distributions. A complementary challenge is state deterioration as successive measurements accumulate and circuit execution advances, making static calibration less representative of the workload. We present Lifting, a QSD framework that addresses both challenges through repetition encoding, neural decoding, and in situ calibration. Encoding supplies joint in-phase and quadrature (I-Q) readout signals from the target and auxiliary qubits, while contextual calibration trains the decoder for the conditions encountered during execution. Our theoretical analysis shows that retaining additional channels from a fixed joint distribution cannot increase Bayes risk and establishes sufficient conditions for contraction of an overlap-based upper bound. Experiments across three IBM processors evaluate the framework on repeated stabilizer measurement, parity-check, and teleportation workloads.