Reverse annealing initializes a programmable quantum annealer in a chosen classical state, reintroduces quantum fluctuations down to a pause point, and returns to readout; repeated cycles form a configurable dissipative dynamics whose schedule knobs are routinely treated as a temperature. On mixed-frustration 12-qubit Ising instances on two D-Wave generations (Advantage2 and Advantage_system6.4), we measure schedule-driven redistribution of the subsystem readout and calibrate each schedule's effective scales, using exhaustive enumeration of the $2^{12}$ configurations, a locked analysis plan with a prior-knowledge disclosure, and regenerated classical reference families. First, schedule shape redistributes basin occupation: 11 of 20 random instances respond to a single-pause-to-two-pause change (15 of 20 for long-hold; block permutation, FDR-controlled), with shifts up to 38 percentage points and changes of dominant-configuration identity. Second, the single-qubit probe temperature does not transfer to the many-body problem: moment-matched inverse temperatures span 0.63-9.1 against a probe value of 7.22, and pause depth and duration each shift the scale on most instances (19/20 and 15/20). Third, no observation requires more than the tested classical families: with update budgets extended below one sweep per cycle, all 33 (device, instance, schedule) cells are reproduced within the tested family (minimum intersection-union $p=0.76$), and schedule-aware spin-vector Monte Carlo brings the late-window statistics within semiclassical reach. A pre-registered linear predictor fails on held-out instances, and an earlier pilot's two-pause enhancement claim is contradicted (Fisher $p=0.0013$). Reverse-anneal schedules are instance-specific basin-occupation controls whose effective scales must be calibrated per instance and schedule; nothing here licenses a quantum-mechanistic reading.
Comments
v3: substantive correction; v2's central claims are withdrawn (matched-temperature PT/SQA bracket, 3/18 PT-unmatched count, native-graph replication) after an internal reanalysis found calibration, units, estimator, and gauge-frame errors. All analyses regenerated under a locked prespecified plan; raw QPU records unchanged. Full details in the Note on version 3, page 1. 32 pages, 7 figures
Reverse annealing initializes a programmable quantum annealer in a chosen classical state, reintroduces quantum fluctuations down to a pause point, and returns to readout; repeated cycles form a configurable dissipative dynamics whose schedule knobs are routinely treated as a temperature. On mixed-frustration 12-qubit Ising instances on two D-Wave generations (Advantage2 and Advantage_system6.4), we measure schedule-driven redistribution of the subsystem readout and calibrate each schedule's effective scales, using exhaustive enumeration of the 212 configurations, a locked analysis plan with a prior-knowledge disclosure, and regenerated classical reference families. First, schedule shape redistributes basin occupation: 11 of 20 random instances respond to a single-pause-to-two-pause change (15 of 20 for long-hold; block permutation, FDR-controlled), with shifts up to 38 percentage points and changes of dominant-configuration identity. Second, the single-qubit probe temperature does not transfer to the many-body problem: moment-matched inverse temperatures span 0.63-9.1 against a probe value of 7.22, and pause depth and duration each shift the scale on most instances (19/20 and 15/20). Third, no observation requires more than the tested classical families: with update budgets extended below one sweep per cycle, all 33 (device, instance, schedule) cells are reproduced within the tested family (minimum intersection-union p=0.76), and schedule-aware spin-vector Monte Carlo brings the late-window statistics within semiclassical reach. A pre-registered linear predictor fails on held-out instances, and an earlier pilot's two-pause enhancement claim is contradicted (Fisher p=0.0013). Reverse-anneal schedules are instance-specific basin-occupation controls whose effective scales must be calibrated per instance and schedule; nothing here licenses a quantum-mechanistic reading.
Reverse annealing is widely used as a heuristic sampler on programmable quantum annealers, but it is often unclear whether the readout has erased its initialization memory or whether the resulting samples represent any calibrated target distribution. We introduce a subsystem-level validation protocol that pairs a memory order parameter M with the total-variation distance DTV between the measured subsystem readout and a fixed, independently calibrated conditional-Boltzmann reference, and apply it on two D-Wave QPU generations. The protocol delivers three things: a diagnostic that separates relaxed, memory-retaining and "wrong-basin" readouts; the isolation of relaxed-but-non-thermal trapping, in which a readout passes the initial-state-independence test while concentrating on the wrong basin; and an engineering control showing that SDK auto-scale, embedding choice and environment preparation shift apparent thresholds, so calibration metadata must be logged to interpret them. Relaxed ferromagnetic readouts are near-deterministic, so the small distances seen there are a consistency check rather than a thermometric measurement; the diagnostic earns its value in the opposite regime. Across 494 conditions on one processor, 113 pass the memory criterion and 110 of those agree with the reference at DTV<0.05, while a wrong-basin instance, reproduced at two Hamiltonian scales, passes the memory test yet reaches DTV>0.94; a 60-condition replication confirms that the diagnostic transfers to a second processor. We provide the protocol as a transferable, reproducible validation workflow for annealer-as-sampler applications.
Quadratic unconstrained binary optimization (QUBO) is the standard interface to quantum annealers, yet a single constrained task admits many QUBO encodings whose penalty choices reshape the energy landscape experienced by hardware. We study a Job Shop Scheduling instance using a two-parameter family of encodings controlled by penalty weights psum (one-hot/sum constraints) and ppair (precedence constraints). Sweeping (psum,ppair), we observe sharp transitions in feasibility and solver success across classical annealing-inspired heuristics and on a D-Wave Advantage processor. Going beyond solution probability, we treat the annealer as an open thermodynamic system and perform cyclic reverse-annealing experiments initialized from thermal samples, measuring the stochastic processor energy change. From the first two moments of this energy change we infer lower bounds on entropy production, work, and exchanged heat via thermodynamic uncertainty relations, and corroborate the observed trends with adiabatic master equation simulations. We find that the same encoding transitions that govern computational hardness also reorganize dissipation: weak penalties generate low-energy infeasible manifolds, while overly strong penalties suppress the effective problem energy scale and increase irreversibility, reducing the thermodynamic efficiency. Our results establish QUBO penalties as thermodynamic control knobs and motivate thermodynamics-aware encoding strategies for noisy intermediate-scale quantum annealers.
Annealing schedule control provides opportunities to better understand the manner and mechanisms by which putative quantum annealers operate. By appropriately modifying the annealing schedule to include a pause (keeping the Hamiltonian fixed) for a period of time, we show that it is possible to more directly probe the dissipative dynamics of the system at intermediate points along the anneal and examine thermal relaxation rates, for example, by observing the repopulation of the ground state after the minimum spectral gap. We provide a detailed comparison of experiments from a D-Wave device, simulations of the quantum adiabatic master equation, and a classical analogue of quantum annealing, spin-vector Monte Carlo, and we observe qualitative agreement, showing that the characteristic increase in success probability when pausing is not a uniquely quantum phenomena. We find that the relaxation in our system is dominated by a single timescale, which allows us to give a simple condition for when we can expect pausing to improve the time to solution, the relevant metric for classical optimization. Finally, we also explore in simulation the role of temperature whilst pausing as a means to better distinguish quantum and classical models of quantum annealers.
Programmable quantum annealers are used for optimization, probabilistic sampling, and simulation, but their performance is commonly reported without the energy exchanged during computation. Here we characterize the D-Wave quantum annealer as a closed thermodynamic cycle. From initial and final Ising energies and an effective temperature fitted to the output distribution, we obtain lower bounds on entropy production, environment energy exchange, work, and power. By varying the prepared distribution and the reverse annealing turning point, we map heater-, accelerator-, refrigerator-, and engine-compatible regimes in one dimensional chains and higher connectivity instances, and apply the same analysis to Advantage and Advantage2 hardware. For an encoded optimization problem, the measured processor energy change states whether final candidates improve or worsen the programmed objective on average. For sampling, the fitted temperature provides an operational measure of how strongly probability is concentrated among low energy configurations. The thermodynamic mode therefore adds information absent from solution quality or runtime alone: it distinguishes driven refinement, net heating, and heat pumping while quantifying their energetic consequences. This framework connects quantum optimization, probabilistic computing, statistical physics simulation, hardware diagnostics, and energy-aware assessment without assuming that a thermodynamic label alone determines computational performance.
Quantum annealing targets low-energy solutions of Ising/QUBO problems, but reliable assessment requires more than best-energy comparisons. This dissertation develops a benchmarking framework for D-Wave quantum annealers that combines strong classical baselines, sampling and diversity metrics, and thermodynamic cost.
Its first contribution, SpinGlassPEPS.jl, is a topology-aware tensor-network heuristic for optimization and sampling on Pegasus/Zephyr-like graphs. It maps Ising instances to local Potts clusters, represents the partition function with PEPS, and performs branch-and-bound search in probability space. Benchmarks show that it is a physically interpretable reference solver, though approximate contractions limit its competitiveness on the largest instances.
The second contribution treats quantum annealers as effective thermal machines, relating success probability and solution quality to dissipation, entropy production, and effective temperature. Carefully placed pauses can improve performance while reducing thermodynamic cost, although longitudinal fields may become harmful in paused schedules.
The thesis also introduces reinforcement-learning post-processing to improve returned samples and exact small-system simulations to probe annealing dynamics. Overall, it argues for quantum-annealing benchmarks that jointly measure algorithmic performance and physical expenditure.
Three recent experimental studies of D-Wave annealers each report an unexplained effective-temperature phenomenon: fitted inverse temperatures whose dependence on annealing time is left empirical, with sampling quality degrading at a critical frustration point [Pelofske, PRR 8, 033280 (2026)]; a coupling-independent temperature offset t-bar in [0.2,0.4] that violates the assumed T_eff ∝ 1/J scaling [Grattan et al., arXiv:2512.03162]; and hardware inverse temperatures a "consistent" factor of 5–7 above closed-system theory, attributed to uncontrollable hardware effects [Kim et al., PRE 113, 035302 (2026)]. We show that a single account, built from the diabatic effective-temperature relation of Gyhm et al. evaluated on the vendor's published fast-anneal schedules and one thermal bath, organizes all three. First, on real schedules the coherent inverse temperature is nearly linear in annealing time — not ∝ sqrt(t_a) as for linearly vanishing drivers — because the transverse field decays exponentially and dies by s ≈ 0.6; exact statevector simulation confirms the first-order integral to 0.4–4.2% in its validity window. Second, a units identity, beta_cyclic(t_a) = beta_correct(t_a/2pi), resolves the factor of 5–7: the published theoretical reference curves in this lane are quantitatively consistent with schedule data evaluated in cyclic rather than angular frequency units. The foundational paper's own schedule evaluations carry the rescale: its earlier single evaluation, beta ≈ 0.54 at 5 ns, equals our correct-convention value at t_a/2pi (0.539), and all 24 of its current tabulated beta and c_2 values that we can recompute from vendor schedule tables are reproduced at t_a/2pi to within 0.11%; a free-rescale fit to Kim et al.'s published simulation curve gives r = 6.40, with 2pi = 6.283 inside the fit's shallow region; what remains after the correction is consistent with open-system cooling. Third, the coupling-independent offset is a residual-excitation floor read through a domain-wall thermometer, t-bar ≈ 2/ln(1/rho): the implied kink densities (9e-5 to 4e-3) overlap published kink measurements, the excitation-floor form beats the constant-offset (noise) model on AIC for all four reported machines, and the logarithm compresses two and a half decades of rho into the narrow observed band, explaining its apparent universality. In the coherent nanosecond regime the closed-system law is an upper envelope on hardware inverse temperature; the shortfall is a candidate in-situ decoherence diagnostic, machine-resolved, and at the one matched cell where both machines are coherent the measured cross-machine ratio (2.63) agrees with the parameter-free prediction at the few-percent level. All hypotheses, tolerances, and kill criteria were registered before data extraction; failed and undecidable gates are reported alongside the successes. No quantum-advantage claim is made. AI use and validation: This work was produced in close collaboration with AI research assistants, directed by the author, who set the research direction, the standards of evidence, and all decisions about what is claimed and published. Anthropic's Claude (Fable 5) performed the analysis, simulation, extraction, and drafting; xAI's Grok (via Cursor) provided independent pre-publication review; and Anthropic's Claude Opus 5.5 performed a post-publication adversarial audit of the manuscript against its pinned evidence, with findings and dispositions logged in the project ledger and incorporated in subsequent versions. The registered-verdict protocol is designed so that no result rests on trusting a model's output: hypotheses, tolerances, and kill criteria were committed to the repository before any data was extracted; failed and undecidable gates are reported in the paper; every numeric claim traces to pinned, machine-readable extraction files and analysis code included in this deposit; figure data was extracted from the source PDFs' vector drawing operations rather than read by a model; and verbatim quotes were verified against raw page dumps.