- The paper introduces Peaked Random Circuits as a system-level fidelity benchmark that uses peak identification and continuous fidelity error to expose interference loss from cumulative gate errors and decoherence.
- PRCs deliver fidelity estimates comparable to Quantum Volume while offering higher sensitivity to coherent noise and independently tunable width and depth for testing scalability, connectivity, and coherence limits.
- The framework was demonstrated on IQM superconducting and AQT trapped-ion platforms, but its value as a predictor of application performance remains unvalidated and may depend on noise models, optimization quality, and sampling costs.
Motivation and positioning
As quantum hardware diversifies across superconducting, trapped-ion, neutral-atom, and photonic platforms, architecture-agnostic system-level benchmarks become necessary for cross-platform comparison. This paper repurposes Peaked Random Circuits (PRCs), originally proposed by Aaronson and Zhang as a verifiable quantum-advantage construction (2605.25983), as a fidelity benchmark for NISQ devices. The central idea: a PRC produces a deterministic "peak" bitstring amid a uniform noise floor, so a device's ability to recover that peak directly quantifies how well it preserves interference structure under cumulative gate errors, decoherence, and connectivity constraints.
The authors argue this addresses gaps left by Quantum Volume (QV) (2605.25983), which compresses performance into a single scalar for square circuits and requires exponentially expensive classical simulation to verify heavy outputs. PRCs offer three claimed advantages: instantaneous verification (the target bitstring is known by construction), higher signal-to-noise ratio (probability concentrated on one outcome rather than spread over many), and architectural flexibility (width and depth are decoupled, unlike QV's square-circuit requirement). These claims are plausible but rest on the assumption that peak recovery is a faithful proxy for general computational reliability—a point the paper does not fully establish empirically against application-level workloads.
Circuit construction and benchmark design
The PRCs follow the mirror architecture of the foundational papers: d/2 layers of random two-qubit unitaries in a 1D brick-wall pattern form the random half R, followed by d/2 layers forming the peaking half P(θ), giving C=P(θ)R. The peaking-half parameters are optimized to maximize ∣⟨s∣RP(θ)∣0n⟩∣2, concentrating probability on target bitstring s from the all-zero input. The circuits were built as tensor networks using quimb, with a hybrid optimization schedule of 5,000 L-BFGS-B iterations followed by 10,000 Adam iterations, which the authors report as an empirical trade-off between peakedness and tractability.
A practical detail worth noting: skewed random-to-peaking layer ratios either reduced achievable peak probability within fixed optimization time or required additional runtime, so near-symmetric halves were retained. For odd depths, the random half is shortened by one layer. Alternative peak bitstrings are realized by inserting NOT gates into the final layer; a control experiment with 10-qubit circuits peaking at $00$, $01$, $10$, R0 (five repetitions each) showed no discernible metric variation, justifying use of the default all-zero peak. Two-qubit circuits are an edge case handled with sequential gate alignment since the brick-wall layout does not fit.
Difficulty scales with both register size and depth: each additional brick-wall layer introduces new entangling bonds across the lattice, stressing coherence time and gate-fidelity budgets. The benchmark matrix spans R1 qubits and R2 layers, executed on IQM's superconducting platform and AQT's trapped-ion platform.
Evaluation metrics
Two metrics are introduced. Peak identification is a binary event: whether the most frequent sampled bitstring equals the target R3. Fidelity error R4 quantifies the contrast between the observed peak probability and the ideal, providing a continuous measure of how hardware noise blurs the output distribution. Together they probe the device's capability boundary: identification fails once noise pushes a spurious bitstring above the true peak, while R5 degrades continuously before that threshold, giving earlier and finer-grained warning than the binary criterion alone.
Running the full circuit matrix on both platforms yields high-precision fidelity estimates comparable to QV while exhibiting greater sensitivity to interference effects—the paper's headline claim. Because PRCs concentrate signal in a single sharp peak, small coherent errors that would be averaged out in heavy-output statistics manifest visibly in R6. The width–depth decoupling additionally allows independent stress tests: depth probes coherence-limited performance, width probes scalability under connectivity constraints (notably relevant given the strictly nearest-neighbor brick-wall coupling pattern).
An implication of the sensitivity result is that PRC-based metrics can resolve degradation mechanisms—such as structured interference loss under realistic noise—that scalar QV values conflate or obscure. Conversely, this same sensitivity means the metric may penalize platforms differently depending on their native noise character (e.g., crosstalk-dominated superconducting devices versus slower but higher-fidelity trapped-ion gates), so cross-platform rankings should be interpreted with the noise model in mind.
Limitations and open questions
Several limitations are stated or implicit. First, the benchmark's validity as a proxy for computational reliability rests on the assumption that preserving a single engineered peak correlates with performance on useful algorithms; no application-level validation is performed. Second, the optimization of R7 is classical and its quality bounds the achievable ideal peakedness—circuits whose optimization did not converge to high peak probability would confound hardware attribution, and the paper reports the chosen iteration budget only as an empirical trade-off without a convergence guarantee. Third, experiments cover two platforms and one coupling topology; extension to all-to-all connectivity (e.g., trapped ions with arbitrary pair selection) or neutral atoms remains untested. Fourth, shot-count requirements for reliable identification at large R8 grow as the peak-to-floor ratio shrinks, though the paper does not quantify this scaling. Open questions include whether PRC-derived fidelity tracks algorithmic benchmarks such as QAOA or variational workloads, and whether the metric remains discriminating at register sizes where classical verification of the ideal distribution becomes infeasible.
Conclusion
This paper converts PRCs from a quantum-advantage proposal into a practical system-level fidelity benchmark, introducing peak-identification and fidelity-error metrics validated on commercial superconducting and trapped-ion hardware. The results indicate precision comparable to Quantum Volume with superior resolution of interference-sensitive noise effects, and the width–depth decoupling enables targeted stress testing unavailable in square-circuit benchmarks. The framework's main open dependency is demonstrated correlation with application-level performance, which future work must establish before PRCs can serve as a standalone reliability standard.