Using domain walls to benchmark fast quantum anneals

AI summary

Overview

Research area: Quantum physics — experimental benchmarking of quantum annealers using frustrated one-dimensional Ising spin chains.

Technical level: Intermediate. The core ideas (domain walls, transverse-field Ising Hamiltonians, Boltzmann distributions) are explained from scratch, but familiarity with quantum annealing terminology helps.

Scope (1 sentence): The paper reports domain-wall-position statistics from short frustrated Ising chains run on a D-Wave Advantage 4.1 processor using its fast anneal protocol, fits those statistics with coherent-evolution and thermal models, and proposes the "tail" of the domain-wall distribution as a platform-independent benchmark of coherence time and noise.

What This Paper Is About

A ferromagnetic spin chain whose two ends are pinned to opposite polarizations must contain at least one domain wall, and where that wall sits depends sensitively on whether the chain is behaving quantum-mechanically (a delocalized particle-in-a-box, giving a sine-squared profile) or classically and thermally (a uniform or U-shaped profile). Prior work on older D-Wave hardware saw only the U-shape produced by averaging over random hardware fields and thermal fluctuations, never the quantum sine-squared shape. This paper uses D-Wave's newer fast anneal feature, with anneal times far shorter than the coherence time, to look for the coherent quantum distribution and to use its shape — specifically how far the distribution extends toward the chain ends — as a diagnostic of hardware quality.

Key Contributions

1. Reproduction of prior benchmarking results on newer hardware. The characteristic thermal-plus-disorder U-shaped domain-wall distribution reported by Chancellor et al. (npj Quantum Inf. 8, 73, 2022) is recovered on a D-Wave Advantage 4.1 QPU (Pegasus topology), and the standard and fast anneal protocols are shown to agree at a 20 μs anneal time.

2. Observation of the coherent sine-squared distribution. Using the fast anneal protocol at 0.01 μs, the experiment produces a fair approximation of the sine-squared particle-in-a-box profile that the original work anticipated but could not access, because anneal times were previously limited to 0.5 μs minimum versus 0.005 μs with fast anneal.

3. Quantitative extraction of hardware noise and temperature parameters. A random-field strength of ζ = 0.101 ± 0.007 GHz is fitted from the 0.005 μs data, and the operating temperature is estimated as 17.53 ± 2.34 mK by combining standard- and fast-anneal fits, consistent within one standard deviation with the reported 15.4 mK.

4. Proposal of a distribution-tail metric. The extent of the domain-wall distribution tail is put forward as a platform-independent metric that simultaneously signals coherence time and noise levels in any analogue device able to instantiate an Ising chain.

Main Findings

• Quantum shape appears at intermediate fast times: At 0.01 μs the domain-wall distribution is a fair approximation of the sine-squared particle-in-a-box profile, the coherent behaviour expected for an ideal chain but not observed in earlier work.

• Thermal U-shape returns at longer times: At 0.03 μs and 20 μs the profiles show the emergence of the U-shape associated with thermalization, matching the earlier Chancellor et al. result.

• Shortest anneal gives a distinct M-shape: At 0.005 μs the profile is slightly skewed and M-shaped — U-shaped in the middle with turned-down ends — indicating the domain wall has not had time to localize across the whole chain. This is supported by a high rate of multi-domain-wall outcomes: 2,927.1 ± 11.2 out of 10^4 samples, averaged over 468 chains per run.

• Post-selection yields single-domain-wall counts: Averaged over 468 chains per run out of 10^4 runs, single-domain-wall counts are 7072.9 ± 11.2 (0.005 μs), 9119.4 ± 17.4 (0.01 μs), 9960.9 ± 26.7 (0.03 μs) and 9762.9 ± 21.3 (20 μs).

• Disorder strength: Fitting closed-system simulations to the 0.005 μs data gives ζ = 0.101 ± 0.007 GHz for the residual random longitudinal fields.

• Temperature estimates agree across protocols: Dimensionless fit values c_fit = βζ are 0.278 ± 0.055 (standard, L = 10) and 0.276 ± 0.042 (fast, L = 12), giving 17.46 ± 3.67 mK and 17.59 ± 2.92 mK respectively, with an average of 0.277 ± 0.034 and 17.53 ± 2.34 mK.

• Coherence dominates at 0.005 μs but not at 0.01 μs: Ideal simulations (no random fields, no thermal effects) lie within roughly one standard deviation of the 0.005 μs data but far from the 0.01 μs data. Adding random fields alone matches 0.01 μs "pretty well, but not perfect." A linear mixture with a thermal distribution fits well with mixing weight b ≈ 0 (95% percentile confidence interval [0.000, 0.056]) at 0.005 μs and b = 0.105 ± 0.025 at 0.01 μs.

• Protocol equivalence: At 20 μs, standard and fast anneal protocols produce the same U-shaped profile, confirming the different fast-anneal mechanisms are equivalent in that regime.

• Reported crossover timescale: The paper cites prior work placing the crossover to thermalized behaviour between 20 ns and 30 ns in the D-Wave hardware used, and its own Monte Carlo simulations suggest the adiabatic limit is approached around 20 ns.

• Note on the excerpt: The provided paper content ends mid-Section IV.3.3, so the detailed simulations characterizing the proposed tail metric in Section V are not available here.

Methodology in Plain English

The team programmed short chains of coupled spins onto a D-Wave Advantage 4.1 chip, pinning the two end spins in opposite directions so that exactly one domain wall must exist. They set the pinning field to twice the coupling strength (h = 2J) so the ends mostly stay fixed, and then ran the chains many times, recording where the domain wall ended up.

Two protocols were compared. The standard protocol ramps the control fields smoothly with time, with a minimum anneal time of 0.5 μs. The fast protocol instead holds the external flux constant — set through user-supplied flux biases — and permits anneal times down to 0.005 μs. Chain lengths were L = 10 (i.e. N + 1 = 10, matching the earlier study) for standard anneals and L = 12 for fast anneals. Anneal times were 0.005, 0.01, 0.03 and 20 μs. Each problem instance was run 10^4 times in four batches of 2000 to 3000, and a Hamiltonian-path heuristic was used for a one-to-one minor embedding so that 562 chains of length 10 or 468 chains of length 12 could run simultaneously. Gauge averaging was applied to remove gauge-dependent systematic errors, and results were post-selected for single-domain-wall outcomes.

On the theory side, the authors modelled the hardware's residual random longitudinal fields as uncorrelated Gaussian variables of zero mean and variance ζ², much weaker than the coupling, and constant over the dynamics. They computed expected distributions two ways: exact time evolution of the full time-dependent Hamiltonian from a uniform superposition (for the coherent regime), and a Boltzmann distribution averaged over many disorder realizations (for the thermalized regime). Fitting used Monte Carlo ensembles — 5×10^5 samples over a range of ζ values with polynomial interpolation across 16 Chebyshev nodes between [0, 1/π] GHz, and 10^7 samples for β with nodes between [0, 1] — with Kullback–Leibler divergence minimized as a stand-in for maximum likelihood, and errors estimated by resampling 10^5 times. Numerical error was dominated by ensemble standard error, roughly 0.1% for the ζ fit and 0.01% for the β fit. Simulations ran on laptops and desktop workstations.

Why This Matters

Impact on research. The work supplies a concrete, reproducible protocol for separating coherent quantum behaviour from thermal and disorder effects on an annealer, and cross-checks a parameter (temperature) against the vendor's reported value. Because the chain problem is simple enough to simulate exactly, it gives a rare meeting point between device measurement and first-principles modelling, and it turns a qualitative question ("is this thing quantum?") into a measurable shape statistic. The tail metric is deliberately platform-independent: any device that can instantiate an Ising chain — superconducting flux qubits, neutral atoms, or a gate-based machine running QAOA — can in principle be scored the same way, which matters for comparing special-purpose and general-purpose hardware in the NISQ and early fault-tolerant eras.

Real-world applications (the paper frames these as the industrially relevant problems that motivate quantum annealing, and does not itself report deployed applications):

• Combinatorial optimization problems mapped onto Ising Hamiltonians, which the introduction identifies as the target class for annealing hardware.
• MAX-2-SAT, cited as a specific problem where diabatic quantum annealing has been shown to outperform adiabatic QA.
• Quantum simulation of non-equilibrium dynamics, phase transitions and magnetic frustration, which the introduction notes state-of-the-art flux-qubit and neutral-atom devices can now access.
• Optimization formulations that encode discrete variables in domain-wall positions, following Chancellor's earlier work on ferromagnetic chains.

Industry relevance. Better benchmarks directly inform procurement and roadmapping decisions for annealer users, since metrics like the distribution tail can distinguish "the hardware is noisy" from "the schedule was too fast" — a distinction that raw success-rate numbers obscure. The fast-anneal regime examined here is also the regime in which diabatic annealing, a candidate route to quantum advantage, actually operates.

Future Directions

• Formalizing and testing the tail metric. Section V is described as providing further simulations to characterize the tail's properties; validating it across multiple devices, chain lengths and topologies is the natural next step.
• Applying the benchmark to other platforms. The authors explicitly note that the coherent and adiabatic regimes should both be accessible on more coherent hardware such as neutral atoms, and that implementing spin-chain benchmarks in a QAOA setting would enable cross-platform comparisons between special- and general-purpose hardware.
• Scaling beyond short chains. The current fits rely on short chains so that Monte Carlo simulation is tractable on desktops; extending the analysis to larger chains and to systems with more than one domain wall is an open modelling challenge.
• Disentangling coherence, noise and schedule. The fitted thermal mixing weight grows from b ≈ 0 at 0.005 μs to b = 0.105 ± 0.025 at 0.01 μs, leaving open how best to locate the coherence-time/thermalization crossover precisely and how it depends on device calibration.

Target Audience

Quantum annealing experimentalists and hardware teams who need practical, low-overhead device benchmarks; theorists working on open-system and diabatic dynamics in programmable spin systems; benchmarking and standards groups such as the UK National Physical Laboratory effort referenced in the paper; and researchers comparing special-purpose annealers with gate-model approaches such as QAOA. Readers wanting an accessible worked example of how to connect a simple spin-chain experiment to a quantitative noise-and-temperature model will benefit most.

Authors’ abstract

We consider the distribution of domain walls arising in ground states of frustrated 1-dimensional Ising spin chains with anti-parallel boundary spins. For ideal chains, a sine-squared particle-in-a-box distribution of domain walls is expected in the quantum setting, and a uniform distribution expected classically. However, for chains with random imperfections, prior work [Chancellor et al., npj Quantum Inf. 8, 73, 2022] found that the distribution in a physical implementation corresponded to neither. Instead, the domain wall distribution forms a U-shaped profile arising from a classical model with random fields, where the combination of averaging over these fields and thermal averaging produces the characteristic U-shape. With D-Wave's newer fast anneal feature, we access anneal times shorter than the coherence time and observe the expected coherent quantum behavior in the form of a sine-squared distribution. For the shortest available anneal time, the dynamics have not yet reached the sine-squared shape and instead display a different U-shape in the middle of the chains. This motivates using the extent of the domain wall distribution tail in this regime as a platform independent metric that provides a signature of both coherence time and noise levels.