arXiv:2610.18482
Lifting: Quantum State Discrimination Framework for Mid-Circuit Measurements
Chi Zhang, Neel Vora, Yilun Xu, Gang Huang, Fan Chen, Lei Jiang
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.
AI Summary
The paper addresses mid-circuit readout discrimination on superconducting processors, focusing on overlapping I–Q distributions and calibration mismatch caused by preceding circuit operations. Its Lifting framework combines CNOT-based repetition encoding, joint analog neural decoding, cycle-matched calibration, and optional confidence filtering. Experiments use three IBM processors, with repeated-stabilizer, parity-check, and computational-basis teleportation workloads. For channels drawn from one fixed joint distribution, the analysis proves that retaining additional observations cannot increase optimal classification error. With equal priors and uniformly bounded conditional Bhattacharyya overlaps, it obtains $R^*_{2d}\leq\frac12\rho_1\prod_{k=2}^{d}\gamma_{\max,k}$, giving geometric decay of this upper bound if every factor is bounded by a common constant below one. These assumptions are not established experimentally. A distance-two acquisition gives KDE-based overlap-risk estimates of 14.7% using the target channel and 5.9% using joint readout. Distance-two neural decoding reduces assignment error by approximately 80–83% relative to hard repetition decoding across the three devices. On Fez, combined encoding and contextual calibration increases unfiltered five-cycle stabilizer success from 54.2% to 58.2%; confidence filtering raises retained-shot success to 61.5%. The paper distinguishes adding features from physically enlarging encoding circuits, which introduces gates, resets, and measurement costs. Some comparisons use different evaluation populations or transferred digital-error estimates. Filtering discards shots, and stabilizer-filter coverage is data dependent. Corrections are interpreted offline: the parity-check success metric reduces to final-syndrome classification, while teleportation tests input-bit classification rather than arbitrary-state fidelity. No real-time neural feedback or corrected-state fidelity is demonstrated.
Error found: The per-cycle QR-Stabilizer accuracy comparison uses target-workload shots for ordinary neural discrimination but held-out calibration shots for Lifting, so it cannot establish improved discrimination on the same workload population.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.