Quantum annealing uses a slowly varying transverse-field Ising Hamiltonian in analog hardware (chiefly D-Wave machines) to find low-energy solutions of binary optimisation problems, and after two decades it remains a usable, commercially accessible device whose advantage over good classical solvers is still narrow, contested and problem-specific.
Summary
Quantum annealing (QA) attacks combinatorial optimisation by physics rather than by circuits. A problem is first rewritten as a quadratic unconstrained binary optimisation (QUBO) problem, equivalently an Ising spin model, so that the answer is the ground state of an energy function. The hardware is initialised in the easy ground state of a strong transverse field and the Hamiltonian is then swept slowly towards the problem Ising Hamiltonian; if the sweep is slow relative to the minimum energy gap, the system should stay in the ground state and be read out as the solution. Quantum fluctuations, rather than the thermal fluctuations of classical simulated annealing, drive the escape from local minima. This is analog, not gate-based, computation: there is no error correction, and noise and small gaps set the limits.
The parameters that decide whether QA is useful are mostly unglamorous. First, formulation cost: constraints must be folded in as quadratic penalty terms, whose weights have to be tuned, and penalty-free reformulations are an active research line precisely because penalties waste qubits and distort the landscape. Second, embedding overhead: the QUBO graph must be mapped onto the fixed sparse hardware graph, which inflates qubit counts and degrades performance as instances grow. Third, the spectral gap and freeze-out dynamics, which determine whether the anneal is adiabatic at all and which are largely invisible to direct hardware measurement, motivating dedicated numerical toolboxes. Fourth, scheduling tricks: reverse annealing, pauses, quench-based mid-anneal readout, all of which have narrow optimal windows.
In practice almost every applied study in the current literature is hybrid: the instance is decomposed or clustered, subproblems go to a D-Wave QPU or to D-Wave’s hybrid solvers, and the results are compared against Gurobi, simulated annealing and metaheuristics. The term also increasingly denotes an algorithm rather than a machine: simulated quantum annealing and path-integral formulations run on classical CPUs, on molecular-dynamics frameworks, and on probabilistic p-bit hardware, and in at least one benchmark the classical path-integral variant beat the quantum hardware pipeline.
So the assessment question is not whether the machines run. They run, at hundreds to thousands of qubits, and outside groups can book time on them. The question is whether the physical annealer beats the best classical alternative on any problem anyone cares about, at a size anyone cares about.
Evidence base
Open questions
- Does the compressed-sensing relaxation-gap advantage survive at n=128 and beyond once embedding overhead is properly accounted for, or does it vanish exactly where it becomes commercially interesting?
- Is the H=14 collapse in the greenhouse benchmark a noise/gap limit intrinsic to analog annealing, or an artefact of the Leap Hybrid BQM workflow and time limits used?
- Can any quantum-annealing dynamics result be shown to resist belief-propagation tensor-network simulation in two and three dimensions, and at what qubit count?
- Do penalty-free encodings and reverse-annealing schedules compound into a scaling improvement, or only shift the constant factor at fixed problem size?
Assessment drafted 2026-08-31 from up to 18 KB sources using the technology-scorecard framework; scores are a draft read pending review.