Abstract
Current benchmarks for mid-circuit measurements (MCMs) are limited in scalability or the types of error they can quantify, necessitating new techniques for quantifying their performance. Here, we introduce a theory for learning Pauli noise in MCMs and use it to create MCM cycle benchmarking, a scalable method for benchmarking MCMs. MCM cycle benchmarking extracts detailed information about the rates of errors in randomly compiled layers of MCMs and Clifford gates, and we demonstrate how its results can be used to quantify correlated errors during MCMs on current quantum hardware. Our method can be integrated into existing Pauli noise learning techniques to scalably characterize and benchmark wide classes of circuits containing MCMs.
Introduction
Mid-circuit measurements (MCMs) are critical components of modern quantum computing architectures, serving as the "eyes" necessary for quantum error correction (QEC) and adaptive algorithms. Despite their importance, MCMs are inherently noisy. Measuring a subset of qubits mid-computation can disturb neighboring qubits through crosstalk or introduce complex, non-stochastic errors that are difficult to track and mitigate.
Existing characterization tools like Quantum Process Tomography (QPT) provide comprehensive noise profiles but fail to scale beyond a few qubits. Conversely, standard Randomized Benchmarking (RB) scales well but typically yields only a single fidelity metric, obscuring the specific Pauli error channels at play. To bridge this gap, we present MCM Cycle Benchmarking (MCM-CB), a scalable protocol designed to learn the full Pauli noise profile of MCMs.
Methodology: MCM Cycle Benchmarking
The core of the MCM-CB protocol is the modeling of MCM noise as a "Uniform Stochastic Instrument" (USI). In this framework, errors are treated as random Pauli bit-flips occurring either immediately before or after the measurement process. To ensure that real-world hardware noise—which can be coherent and complex—conforms to this stochastic model, we employ randomized compiling. This technique involves inserting random Pauli gates into the circuit to "twirl" the noise into a simpler, manageable Pauli form.
As shown in Figure 2, the protocol runs subexperiments where a specific circuit layer is repeated $d$ times. By observing the decay of the expectation values as $d$ increases, we extract decay constants. These constants are then processed using a Walsh-Hadamard Transform to calculate the exact probabilities of specific Pauli errors, such as bit-flips on a single qubit or correlated multi-qubit errors.
Experimental Results and Correlated Noise
We validated MCM-CB using both Stim simulations and experiments on IBM Quantum processors (ibm_osaka and ibm_torino). In simulations, the method accurately estimated true instrument fidelity across diverse noise regimes, including sparse and depolarizing models with crosstalk.
On physical hardware, specifically the 4-qubit test on ibm_osaka, a striking observation was made: the "idling" qubits (those not being measured) often suffered more noise than the qubit undergoing measurement. Furthermore, the total infidelity of the idling subsystem was significantly lower than the sum of individual qubit infidelities. This discrepancy is a "smoking gun" for correlated errors.
Figure 3 illustrates the power of MCM-CB in identifying these specific error channels. On the IBM hardware, we successfully pinpointed dominant ZZ-type correlations between adjacent idling qubits. These are likely caused by residual couplings in the transmon hardware that are exacerbated during the measurement pulse of a neighbor. Identifying these specific patterns is crucial for tailoring error-correcting codes to the actual noise of the device.
Discussion and Scalability
A primary advantage of MCM-CB is its scalability. The number of samples required to characterize a layer is independent of the total number of qubits ($n$), making it suitable for "utility-scale" quantum processors. While the current implementation assumes surrounding gates are Clifford, and the resulting fidelity is technically a lower bound, the method provides unprecedented detail compared to traditional benchmarking.
Future work involves integrating MCM-CB into "Pattern Transfer Graphs" (PTGs). This integration would allow for the simultaneous characterization of entire syndrome extraction cycles—the fundamental building blocks of fault-tolerant quantum computing.
Conclusion
MCM Cycle Benchmarking provides a robust, scalable framework for quantifying the noise associated with mid-circuit measurements. By revealing the detailed Pauli error rates and identifying damaging correlated noise, this method offers a vital tool for hardware developers and QEC researchers aiming to build reliable, large-scale quantum computers.
Summary updated