Generalized Mermin Inequalities for Benchmarking Large-Scale GHZ States

AI summary

Overview

Research area: Quantum information science — multipartite Bell tests, quantum nonlocality certification, and benchmarking of quantum processors.

Technical level: Advanced. The paper assumes familiarity with Bell inequalities, GHZ states, stabilizer formalism, and statistical confidence bounds, though the underlying logic can be grasped by a nonexpert.

Scope (one sentence): The paper introduces a family of finite-setting generalized Mermin Bell inequalities whose extra measurement settings tighten classical bounds, and uses them to certify nonlocality in GHZ states of up to 80 qubits on a superconducting processor.

What This Paper Is About

Multipartite Bell tests can benchmark a quantum processor using only measured correlations, but at large scale the correlations decay under noise and the number of terms in a standard Bell expression grows exponentially, making certification impractical. This paper asks whether the Bell functional itself — not the state preparation — can be redesigned to fight back, by adding measurement settings that lower the classical threshold while leaving the ideal quantum value untouched. The authors build that generalized inequality, derive its analytic classical bounds, and test it on GHZ states up to 80 qubits.

Key Contributions

1. A finite-setting generalized Mermin family. The authors define an operator indexed by the number of local measurement settings m and the system size n, built from m coplanar equatorial measurement directions. At m = 2 it reduces to the standard Mermin inequality up to normalization.

2. Analytic certification bounds organized for large-scale use. They derive closed-form local hidden-variable bounds for powers-of-two m (and even n, with the odd-n expression in the Supplement), grouping-model bounds that certify nonlocality depth, and adopt a biseparable quantum benchmark following Pál and Vértesi's multisetting extended-parity-game bound.

3. A sampling-based estimator with rigorous statistics. Instead of enumerating exponentially many correlators, they estimate the Bell operator by uniformly random sampling of settings, treating setting-level averages as bounded random variables and assigning p-values via the one-sided empirical Bernstein bound.

4. Experimental demonstration on 80-qubit GHZ states. Using the Zuchongzhi 3.1 superconducting processor, they show Bell ratios growing exponentially with system size, certify nonlocality depth 14, and show that raising m strengthens both the Bell ratio and depth certification — all from measured correlators with no readout correction, tomography, or model-based mitigation.

Main Findings

• The quantum value is invariant while the classical bound drops. For the powers-of-two setting choices considered, every signed term in the operator contributes +1 on an ideal GHZ state, so the normalized quantum value saturates at 1 for any m. The local bound, however, decays as α_m^n, with α_2 = 1/√2 ≈ 0.707 falling to α_8 ≈ 0.641.

• Improved noise-robustness scaling in principle. The ideal Bell-violation ratio scales as 1.414^n for the standard Mermin inequality versus 1.561^n for m = 8.

• Measured exponential scaling. The experimental Bell ratio for m = 8 scales as (1.5168 ± 0.0024)^n, substantially exceeding the standard two-setting scaling.

• Correlations are essentially unchanged; the threshold is what moves. At n = 80, the measured Bell-operator expectation values remain near 0.1 for both m = 2 and m = 8, and the settings-level empirical Bernstein test still gives small p_EB upper bounds for both. The gain comes from lower classical thresholds, not better state preparation.

• Fixed-size advantage. For the 80-qubit GHZ state, the experimental Bell ratio increases monotonically as m rises from 2 to 4, 8, 16, and 32, and once m > 2 the observed ratio exceeds the theoretical ceiling of the standard Mermin inequality.

• Deeper nonlocality certification. At n = 60 the generalized test certifies depth 12 for m = 8 versus depth 10 for m = 2; at n = 80 the advantage grows to depth 14 versus depth 10. The depth trend is non-monotonic, which the authors attribute to a staircase competition between discrete grouping-model thresholds and decoherence-driven correlator decay.

• Favorable biseparable comparison at moderate size. At n = 16, ⟨M_{2,16}⟩exp = 0.7446 exceeds 1/√2, and ⟨M{8,16}⟩_exp = 0.7128 exceeds 1/(8 sin(π/16)) ≈ 0.6407 — almost doubling the gap to the relevant biseparable threshold.

• Honest scope statement. The experiment is not loophole-free and is not presented as a device-independent Bell test; it is a correlation-only Bell benchmark.

Methodology in Plain English

The inequality. Take n qubits in a GHZ state. At each qubit, measure one of m directions lying in the equator of the Bloch sphere, chosen so that on the ideal state each allowed combination of outcomes carries a predictable sign. Averaging those signed correlators gives an operator whose ideal value is 1 no matter how many settings are used. What changes with m is how large the value can be if the qubits were governed by classical local instructions — and that ceiling falls as m rises, so a noisy state that barely beats the classical threshold at m = 2 can clear it easily at m = 8.

The statistics. The operator has m^(n−1) terms, far too many to measure one by one. The authors instead pick settings uniformly at random, measure M = 1500 shots per sampled setting, average within each setting to get a number in [−1, 1], and average those across N sampled settings. They deliberately count N — the number of settings — rather than the total shot count K = NM as the statistical sample size, a conservative choice that does not assume independence among shots within a setting. Randomizing the setting order throughout the run keeps slow hardware drift from being tied systematically to any particular term. Confidence in rejecting local, biseparable, or grouping-model explanations is then assigned with a one-sided empirical Bernstein bound.

The hardware. GHZ states are built by expanding entanglement outward from a central qubit along the chip's native connectivity, with the entangling graph chosen by an automated optimization script for shallow, hardware-aligned compilation. A phase-protection window — a Hadamard frame plus dynamical-decoupling pulses — suppresses phase noise during idle time before the equatorial analysis layer is applied. Across all comparisons of m, the preparation circuit, qubit subset, dynamical-decoupling sequence, and total sampling budget were held fixed; only the final measurement layer changed.

Why This Matters

Bell certification usually trades scale for rigor: large multipartite experiments lean on fidelity or coherence diagnostics and correlation-depth witnesses rather than direct Bell tests. This work shows that a well-chosen Bell functional can extract a sharper certificate from the same noisy state, changing only the classical threshold. That reframes the design question for near-term hardware — certification strength depends on both the state and the functional used to interrogate it.

Potential applications (motivated by the paper's framing):

• Processor benchmarking. A correlation-only benchmark for large quantum processors that requires no tomography and no noise model.
• Device-independent cryptography. The paper notes a connection, in more idealized architectures, to DI cryptography and self-testing — though this implementation is explicitly not loophole-free.
• Communication-complexity advantages. Also cited as a downstream motivation for multipartite nonlocality.
• Entanglement-structure certification. Nonlocality-depth bounds distinguish how many parties genuinely share nonclassical correlations, relevant wherever entanglement is a resource.

Industry relevance: Superconducting processors at the 100-qubit scale are now industrial and national-lab assets, and buyers and operators need benchmarks that cannot be gamed by calibration choices. A benchmark built purely from measured correlators and analytical bounds, with no readout correction or model-based mitigation entering the estimate, is attractive precisely because it makes no assumptions about the device's noise.

Future Directions

1. Closing the loopholes. The authors explicitly leave open the path to a loophole-free, genuinely device-independent version of this test on chip, noting compatibility with "future Bell architectures."

2. Extending beyond powers-of-two settings. The closed-form local bound is given for m = 2^r with even n, with the odd-n expression deferred to the Supplement; a fully general-m analytic treatment would widen the family's reach.

3. Porting the strategy to other platforms. The paper describes the certification strategy as "in principle portable across hardware platforms," inviting tests on trapped ions, neutral atoms, or photonics.

4. Understanding the depth staircase. The non-monotonic nonlocality-depth trend — temporary increases at intermediate n, plateaus or declines at larger n — is explained qualitatively as a competition between discrete grouping thresholds and correlator decay, but a predictive model would help target where depth certification is maximized.

Target Audience

Quantum information researchers working on nonlocality, Bell inequalities, and entanglement certification; experimentalists running large-scale GHZ or other multipartite state preparation on superconducting or other platforms; quantum-computing benchmarking and characterization teams who need device-level figures of merit that do not depend on a noise model; and graduate students with a background in quantum mechanics who want a worked example of how an abstract Bell functional becomes a practical, statistically rigorous benchmark.

Authors’ abstract

Multipartite Bell tests provide a correlation-only route to benchmarking quantum processors, but their application at large scales is hindered by the rapid decay of many-body correlators under noise and exponentially many terms in conventional Bell expressions. Here we address these scalability obstacles by introducing a finite-setting generalized Mermin family of state-tailored Bell inequalities with analytic certification bounds, in which the measurement-setting number $m$ provides an additional certification dimension complementary to the system size $n$. We show that, for the powers-of-two setting choices considered here, increasing $m$ leaves the ideal normalized multipartite quantum value unchanged while lowering the relevant classical bounds, thereby strengthening the Bell-violation ratios and yielding an improved noise-robustness scaling compared to the standard Mermin inequality. We test this construction experimentally on a programmable superconducting processor by preparing Greenberger-Horne-Zeilinger (GHZ) states of up to 80 qubits. Using randomized sampling for direct Bell-operator estimation, we observe Bell ratios that grow exponentially with system size, certify a nonlocality depth of 14, and show that increasing $m$ strengthens both the Bell ratio and depth certification. All results are obtained solely from measured correlators and analytical bounds, without readout correction, tomography, or model-based mitigation. Generalized Mermin inequalities therefore provide a sharper Bell benchmark for noisy large-scale GHZ states.