Generalized Svetlichny Inequalities as a Cross-Platform Benchmark for Quantum Processors

AI summary

Overview

Research area: Quantum computing hardware characterization and benchmarking, grounded in quantum foundations (multipartite Bell nonlocality, generalized Svetlichny inequalities).

Technical level: Intermediate. Readers should be comfortable with qubits, GHZ states, Bell-type inequalities, and the idea of readout-error mitigation; the paper is otherwise self-contained and largely empirical.

Scope (one sentence): The paper defines a single hardware-independent benchmark based on violations of generalized Svetlichny inequalities and applies it unchanged to three real quantum processors of two different architectures, from N = 2 to N = 10 qubits, with and without readout-error mitigation.

What This Paper Is About

Most benchmarks of quantum processors measure local or quasi-local quantities—single- and two-qubit gate fidelities, randomized benchmarking, quantum volume—which say little about whether a device can generate and preserve large-scale multipartite quantum correlations as a whole. The authors ask whether a violation of the generalized Svetlichny inequality, an exponentially growing and highly fragile multipartite quantity, can serve as a practical, device-independent benchmark that is still sensitive to the imperfections a real processor accumulates during execution. They answer by running the identical logical circuit—a GHZ-state preparation with fixed theoretical measurement angles—on two trapped-ion processors (AQT Marmot and AQT IBEX Q1) and one superconducting processor (IQM Euro-Q-Exa System 1, EQE1) from 2 up to 10 qubits.

Key Contributions

1. A cross-platform Svetlichny benchmarking protocol. The authors formulate and validate a protocol whose logical specification is independent of the target hardware but which remains sensitive to compilation and execution imperfections, and they apply it unchanged—no per-device tuning of measurement settings—to three processors spanning two architectures.

2. A scalable, full-correlator evaluation instead of sampling. Rather than adapting the resource state to device connectivity or estimating the Bell operator from a randomly sampled subset of measurement contexts, they reconstruct every individual correlator of the generalized Svetlichny inequality on real hardware, reaching N = 10 qubits, which requires 2^10 = 1,024 correlators.

3. A normalization that makes subsystem sizes comparable. They define the normalized violation V_N = S_N / 2^(N−1), which removes the exponential growth of the classical bound (|S_N^±| ≤ 2^(N−1)) and places all devices and all N on a common scale, with V_N > 1 marking a genuine violation and the ideal GHZ prediction at V_N = sqrt(2).

4. A separation of measurement noise from state degradation. Local readout mitigation (tractable up to N = 10) and correlated readout mitigation (restricted to N ≤ 6 by its exponential calibration cost) are applied to the same data, allowing the authors to quantify how much of the observed loss of violation is attributable to measurement imperfections rather than to destruction of the correlations themselves.

Main Findings

• Trapped-ion devices hold the violation to ten qubits. AQT Marmot and AQT IBEX Q1 sustain statistically significant Svetlichny violations across the entire explored range up to N = 10 qubits, remaining in the nonclassical region throughout.

• The superconducting device loses the violation beyond six qubits under raw conditions. IQM EQE1 shows performance comparable to the trapped-ion devices at N = 2–3, crosses the hybrid-local-realistic bound at N = 6, and degrades well below the classical bound for N ≥ 7 in unmitigated execution.

• Mitigation recovers the superconducting device up to eight qubits. Local readout mitigation extends IQM EQE1's accessible nonclassical region to approximately N = 8 qubits.

• The mitigation gain is strongly asymmetric across platforms. For both AQT processors, relative improvement from local and correlated mitigation stays below approximately 9% at every subsystem size; for IQM EQE1 the relative improvement grows rapidly with subsystem size, reaching approximately 27% at N = 10.

• Measurement noise is not the dominant limit on trapped ions. Because mitigation yields only modest corrections on Marmot and IBEX Q1, the authors attribute their residual degradation primarily to intrinsic decoherence and gate imperfections affecting the multipartite state itself.

• Two devices with the same quantum volume differ measurably. AQT IBEX Q1 consistently attains values closer to the ideal sqrt(2) than Marmot across most of the explored range, although their confidence intervals converge toward N = 9–10; quantum volume is identical for the two devices (128) despite this gap.

• Marmot tracks its simulator only at small sizes. Marmot's values follow the AQT simulator predictions closely only at N = 2–3; beyond that a persistent gap in V_N emerges and stays roughly constant in magnitude up to N = 10.

• Deeper transpiled circuits did not mean worse performance. Table 1 shows the AQT and IQM transpiled circuits need an identical number of single- and two-qubit gates at every subsystem size (logical circuit depth rising from 5 at N = 2 to 13 at N = 10; IQM depths 4 to 20; AQT depths 5 to 29), yet the AQT implementation is consistently deeper because each pair of chained RXX gates is separated by a physical RZ correction the compiler cannot absorb, whereas the IQM compiler merges the phase correction via virtual-Z tracking.

• Noise-model simulations anticipated the hardware result. Among seven simulated backends, only FakeAQT and FakeAlgiers remain in the nonclassical regime up to N = 10, while FakeBrisbane, FakeMontrealV2, FakeCusco, FakeSingaporeV2 and FakeTorontoV2 cross into the HLR regime before reaching N = 10, with all simulated points obtained at 2,000 shots and 95% confidence intervals from 20 repetitions per data point.

• A stated limitation on what the test certifies. Because all qubits sit in one package and measurements cannot be spacelike separated, the protocol does not provide a loophole-free certification of genuine multipartite nonlocality; following earlier work, the authors describe it as simulating nonlocality, and note the correlations can alternatively be read as a contextuality phenomenon.

Methodology in Plain English

The team prepares GHZ states—maximally entangled states of N qubits that give the largest possible Svetlichny violation—using one standard circuit that is written once and then automatically translated (transpiled) into each platform's native gate set. They measure along fixed angles taken from the original Svetlichny formalism rather than tuning angles per device, and they evaluate every term of the inequality, which means 2^N distinct correlators, reconstructed one by one on the hardware.

Each device and each subsystem size N from 2 to 10 gets an effective budget of 2,000 shots: Marmot caps jobs at 200 shots, so ten such jobs were aggregated, while IBEX Q1 and EQE1 accepted a single 2,000-shot job. This whole procedure is repeated five times per device and size, and the reported values are means with 95% confidence intervals from a Student's t-distribution with four degrees of freedom. To separate measurement error from genuine loss of correlation, the raw counts are then corrected with two readout-error mitigation schemes: a local one that assumes independent single-qubit assignment errors and stays tractable to N = 10, and a correlated one that reconstructs the full 2^N × 2^N assignment matrix but becomes prohibitively expensive beyond N = 6. Before touching hardware, the same circuits are simulated under a calibrated AQT noise model and six Qiskit FakeProvider superconducting noise models to establish a reference range.

Why This Matters

Impact on research. The work turns a foundational multipartite-nonlocality test into an operational benchmark, and shows that system-level collective performance is not captured by local metrics. It gives the community a protocol whose logical definition is hardware-independent yet whose hardware instantiation is sensitive enough to distinguish two devices from the same vendor with identical quantum volume (128). It also gives a concrete, quantitative account of how much apparent loss of nonlocal signal on a superconducting device is measurement noise rather than destroyed correlation.

Real-world applications (where such benchmarking matters):

• Comparing and selecting quantum hardware for procurement, where headline gate fidelities may not reflect the ability to run large entangled workloads.
• Validating cloud-accessible quantum processors from different providers on a common, reproducible scale.
• Co-design of compilers and native gate sets, since the analysis isolates the effect of transpilation choices such as virtual-Z tracking and unavoidable RZ corrections between chained RXX gates.
• Deciding when error mitigation is worth its cost, particularly the trade-off between local mitigation, which remains tractable at all sizes, and correlated mitigation, whose calibration grows exponentially.

Industry relevance. The results are immediately actionable for hardware providers and supercomputing hosts: they show that gate count and transpiled depth considered in isolation do not predict collective performance, and that readout quality—long a secondary concern next to gate fidelity—can dominate the large-N signal on superconducting hardware. The 10-qubit demonstrations on both trapped-ion platforms also show the benchmark is not merely a small-N diagnostic.

Future Directions

• Scalable measurement strategies. Because evaluating the Svetlichny operator requires 2^N measurement configurations and correlated mitigation requires a 2^N × 2^N assignment matrix, further scaling will need sampling strategies or reduced-measurement formulations, and the authors flag structured approximations of the assignment matrix and model-free mitigation approaches as candidate directions.
• Other resource states and circuit families. Graph states and linear cluster states are named as a natural extension, since their Bell inequalities show improved noise resistance at large N; comparing them against GHZ protocols on the same hardware would clarify how entanglement topology and connectivity shape robustness. The authors also note their earlier work on Clifford, Clifford+T, and Haar-random circuits, to which the same on-hardware evaluation applies.
• Interpolating mitigation schemes. Developing scalable mitigation that falls between local and correlated approaches, retaining accuracy without exponential calibration overhead, is identified as an important direction for future multipartite benchmarking.
• Clarifying the contextuality reading. The relation between observed loss of violation and loss of the underlying nonclassical resource is left open: the authors stress that absence of a violation does not imply the total disappearance of multipartite correlations, which may remain physically present while hidden by readout noise.

Target Audience

Quantum computing experimentalists and benchmarking specialists, especially those comparing hardware across platforms or designing characterization suites for NISQ devices; compiler and control-stack engineers interested in how transpilation and readout choices propagate into large-entanglement performance; quantum foundations researchers who want to see Bell-type multipartite inequalities used as an engineering diagnostic; and research infrastructure staff at supercomputing centers evaluating or procuring quantum systems. It is accessible to graduate students with basic familiarity with entanglement and Bell inequalities, and assumes no specialist knowledge of trapped-ion or superconducting hardware.

Authors’ abstract

A cross-platform benchmark should be device-independent yet sensitive to imperfections accumulated during execution. Building on prior work, we formulate and validate a protocol through demonstrations based on generalized Svetlichny inequality violations (a simulation, rather than a loophole-free certification, of multipartite nonlocality) for larger subsystem sizes. We apply it unchanged to three processors, the trapped-ion Marmot and IBEX Q1 (both manufactured by AQT) and the superconducting Euro-Q-Exa System 1 (IQM), using identical GHZ-state circuits with fixed theoretical measurement settings from N=2 to N=10 qubits. Both trapped-ion processors sustain violations up to ten qubits, while the superconducting processor loses them beyond six qubits under unmitigated (raw) conditions; readout-error mitigation recovers violations up to eight qubits, with a substantial fraction of the degradation attributable to measurement imperfections. These processors provide a heterogeneous testbed for validating the protocol's cross-platform applicability and discriminating power, demonstrating its practicality as a common framework for assessing collective quantum performance.