Quantum annealing is a computational method that uses quantum fluctuations to search for low-energy configurations of a system. An optimization problem is expressed as an energy function, with better solutions corresponding to lower energies. The system begins under conditions that favor a readily prepared quantum state, and its governing interactions are gradually changed toward those representing the problem. Measurements at the end produce candidate solutions. Unlike a gate-based quantum computer, a quantum annealer generally performs this evolution through continuously controlled physical interactions rather than a sequence of discrete logical operations.
The method connects quantum mechanics, statistical physics, and mathematical optimization. Its central motivation is that quantum transitions, including tunneling, may explore some difficult energy landscapes differently from classical thermal search. Quantum annealing is studied both as an optimization technique and as a way to investigate interacting quantum systems. Its performance depends on the problem, the evolution schedule, environmental effects, and the physical implementation; the presence of quantum behavior alone does not establish a computational advantage.
Origins and development
Quantum annealing grew out of efforts to find low-energy states in systems with many competing interactions. In a frustrated magnetic system, satisfying one interaction may prevent another from being satisfied, producing numerous local energy minima. Related mathematical structures arise in scheduling, graph partitioning, and other discrete optimization problems. This connection made methods from statistical physics especially influential in the development of optimization algorithms.
A major classical predecessor was simulated annealing, introduced into optimization in influential 1983 work by Scott Kirkpatrick, C. Daniel Gelatt, and Mario Vecchi. It adapts the idea of cooling a material toward a low-energy configuration. At high effective temperature, a search can accept moves that increase its objective function; as the temperature decreases, such moves become less frequent. This offers a mechanism for escaping local minima that would trap a purely greedy procedure. Quantum annealing substitutes, or supplements, thermal fluctuations with quantum fluctuations controlled by a parameter such as a transverse magnetic field.
Several early contributions established different aspects of the idea. In 1994, A. B. Finnila and collaborators described quantum annealing as a method for minimizing multidimensional functions. In 1998, Tadashi Kadowaki and Hidetoshi Nishimori analyzed annealing in a transverse-field Ising model, a formulation that became central to later research. Their work compared reductions of quantum fluctuations with classical cooling in disordered spin systems. Experimental work by J. Brooke and colleagues, published in 1999, explored quantum annealing in the disordered magnetic material , linking the computational proposal to laboratory studies of slow magnetic relaxation.
A closely related development was adiabatic quantum computation. Edward Farhi and collaborators formulated this approach in 2000 and presented a study of a satisfiability problem in 2001. It emphasizes evolving sufficiently slowly that a system follows a low-energy instantaneous eigenstate. Quantum annealing subsequently became a broader term, encompassing finite-time, finite-temperature, and deliberately nonadiabatic protocols as well as approximately adiabatic ones.
Commercial development centered on D-Wave Systems, whose 2011 D-Wave One system was marketed with a 128-qubit processor. Later generations increased physical qubit counts and connectivity, including the Advantage generation introduced in 2020 with more than 5,000 physical qubits. These machines made programmable annealing available for experiments at scales beyond small laboratory demonstrations. Their development also sharpened distinctions among observing quantum effects, improving a particular benchmark, and establishing a general computational speedup—three claims that require different kinds of evidence.
Energy functions and annealing Hamiltonians
The standard mathematical description uses a time-dependent Hamiltonian, the operator that specifies the system’s energy and generates its quantum evolution. A common form is
where is the annealing time. The driver Hamiltonian creates quantum fluctuations, while the problem Hamiltonian encodes the objective. Typically, decreases and increases during the anneal. Their precise forms, endpoints, and accessible control ranges depend on the device or algorithm.
For an Ising-type optimization problem, the problem Hamiltonian is commonly written
Here is a Pauli operator acting on qubit , and its measured eigenvalues are . The coefficients specify local biases, while specify pairwise interactions. In this sign convention, a positive favors opposite values of and ; a negative value favors equal values. Reading all qubits yields a classical configuration whose objective value can be calculated directly.
The usual driver is a transverse field,
Because does not commute with , this term induces transitions between classical configurations. Its ground state is a product of equal superpositions of the two computational-basis states. Although this initial state is straightforward, interactions during the anneal can generate quantum entanglement and strongly correlated states. As the driver weakens, the final measurement increasingly reflects the classical energy landscape defined by .
In an ideal isolated system, evolution obeys the Schrödinger equation,
This equation describes probability amplitudes, not a classical trajectory moving among configurations. A landscape picture is useful for intuition, but its valleys and barriers represent an objective over discrete configurations rather than necessarily a literal spatial potential.
The overall scale of the Hamiltonian is physically important. Multiplying all objective coefficients by a positive constant leaves the mathematical minimizer unchanged, yet changes energy differences relative to temperature, control noise, and available evolution time. Consequently, two mathematically equivalent objective functions can behave differently on hardware. Annealing schedules, coefficient normalization, and the driver’s strength are therefore part of the computational method rather than merely implementation details.
Adiabatic evolution, tunneling, and thermal dynamics
The quantum adiabatic theorem provides the idealized rationale for slow annealing. Under suitable conditions, a system initially in a nondegenerate ground state remains close to the instantaneous ground state if its Hamiltonian changes sufficiently slowly. A central quantity is the spectral gap,
between the ground state and first excited state. Small gaps generally make adiabatic evolution more demanding. Bounds on the required time also depend on derivatives of the Hamiltonian, transition matrix elements, boundary conditions, and the desired accuracy; there is no universal runtime formula based on the minimum gap alone.
An avoided crossing illustrates the difficulty. Two low-energy states can approach one another as the annealing parameter changes without becoming exactly degenerate. If the system crosses this region too quickly, it may acquire excited-state population. In some problem families, the minimum gap becomes exponentially small as the number of variables grows. A ground-state-following strategy may then require exponentially long evolution, despite using quantum mechanics.
Quantum tunneling motivates a different, complementary picture. Thermal activation over a barrier is suppressed approximately by , where is the barrier height. Tunneling depends on a barrier’s width and shape as well as its height. This suggests an advantage for some tall, narrow barriers that are difficult for thermal search. In a spin problem, however, “width” concerns changes among configurations, often involving several spins. An advantage for one barrier geometry does not imply an advantage for a full optimization instance containing many competing bottlenecks.
Physical annealers also interact with their surroundings. Decoherence, energy relaxation, and thermal excitation make them open quantum systems rather than perfectly isolated realizations of the Schrödinger equation. When a relevant gap is comparable to or smaller than , thermal transitions can strongly affect the populations. Relaxation can sometimes return population to lower-energy states, while thermal excitation can remove it. Short coherence times therefore do not by themselves rule out quantum-assisted transitions, but they do rule out treating the entire computation as an ideal coherent evolution without further evidence.
A further effect is dynamical freeze-out: transitions among relevant states become too slow compared with the remaining schedule. The eventual output may then reflect populations established earlier in the anneal rather than thermal equilibrium under the final Hamiltonian. Pauses, changes in annealing rate, and alternative paths can modify these populations. Some useful protocols deliberately allow nonadiabatic transitions, so the best practical schedule need not be the slowest one.
Encoding optimization problems
Many annealing workflows begin with a quadratic unconstrained binary optimization problem, or QUBO:
The substitution converts this objective into an Ising energy, with . The conversion changes coefficients and introduces a constant offset but preserves the correspondence between minimizing assignments. QUBO and Ising formulations are therefore two common interfaces to the same class of binary quadratic problems.
Constraints are often represented through penalties. If a set of binary variables must satisfy
a nonnegative penalty
can be added to the objective. Because , this expression remains quadratic. It is zero exactly when one variable is selected and positive otherwise. Similar constructions encode resource limits, assignments, and consistency conditions, sometimes using auxiliary binary variables to represent integer slack.
The choice of penalty strength is consequential. If is too small, violating a constraint may improve the penalized objective. If it is unnecessarily large, the full coefficient range may have to be compressed to fit hardware limits, reducing the physical energy differences between feasible solutions. Correct modeling therefore involves more than producing an algebraically valid expression: it requires balancing constraint enforcement against precision, noise, and finite temperature.
Graph problems offer direct examples. In weighted maximum cut, vertices are assigned spins, and an edge contributes to the cut when its endpoint spins differ. For nonnegative edge weights , the cut value is
Maximizing is equivalent to minimizing a corresponding Ising objective. Scheduling problems often require less direct encodings, such as one binary variable for every possible job–time assignment. These encodings may introduce large numbers of variables and interactions before physical implementation is considered.
Higher-order objectives can sometimes be reduced to quadratic form by introducing auxiliary variables and consistency penalties. This preserves selected properties of the optimization problem only when the reduction is constructed correctly and penalties are sufficient. The extra variables increase both logical size and embedding cost. Many relevant problems exhibit NP-hardness, but representing them on an annealer does not remove that difficulty: the output remains a candidate whose quality and feasibility must be assessed.
Computational power and classical relationships
Quantum annealing is not identical to universal quantum computation. General adiabatic quantum computation can simulate the standard circuit model with polynomial overhead under suitable constructions. That equivalence applies to sufficiently expressive Hamiltonians and controlled evolutions; it does not imply that a commercial annealer with a restricted interaction set can efficiently execute every quantum circuit. Available couplings, control precision, geometry, state preparation, and measurement all matter.
This distinction is particularly important for transverse-field Ising annealers. Their Hamiltonians are commonly stoquastic in the computational basis: off-diagonal matrix elements are real and nonpositive. This structure avoids the usual sign problem in certain path-integral Monte Carlo formulations. It enables useful classical simulation approaches, but does not guarantee efficient simulation. A sampling process can still mix very slowly, and reproducing real-time quantum dynamics can be much harder than representing equilibrium statistical weights.
One classical comparator is simulated quantum annealing, which commonly uses quantum Monte Carlo methods to sample an equilibrium distribution associated with an annealing Hamiltonian. It can capture some consequences of quantum fluctuations without running on quantum hardware. Its algorithmic dynamics, however, are not the physical dynamics of a quantum annealer. Similar performance on a benchmark may indicate a useful classical explanation or competitor; it does not establish identical microscopic mechanisms.
Other comparisons involve conventional simulated annealing, parallel tempering, branch-and-bound, integer programming, and specialized graph algorithms. Their strengths depend on the problem structure. An annealer may perform well against a simple local search while remaining slower than a tailored classical solver. Computational complexity therefore requires attention to the precise task, the instance family, and how resources scale—not merely whether an individual calculation uses quantum effects.
Annealing also differs from the circuit algorithms associated with prominent quantum speedups. Shor’s algorithm exploits structured arithmetic and Fourier transforms, while Grover’s algorithm provides a quadratic query advantage for unstructured search. Neither is automatically implemented by slowly reducing a transverse field over an arbitrary objective function. Constructing a low-energy encoding of a problem is not equivalent to constructing an efficient quantum algorithm for it.
For optimization, the relevant outcomes may include exact solutions, approximate solutions, or samples below a target energy. These tasks have different difficulty and verification requirements. Calculating a returned assignment’s energy is often easy; proving that no better assignment exists may be much harder. No general theorem establishes that standard quantum annealing efficiently solves all NP-hard optimization problems.
Physical hardware and connectivity
The best-known programmable annealers use superconducting circuits operating at temperatures of order tens of millikelvin. A common implementation employs flux qubits, in which circulating-current states provide the two computational states. External controls adjust local energy biases, interactions, and the tunneling amplitude between states. These devices belong to the broader family of superconducting qubits, although their design and operating regime differ from those of many gate-based superconducting processors.
Programmable couplers implement pairwise interactions. Their physical arrangement defines the hardware connectivity graph: two variables can interact directly only where an appropriate coupler exists. D-Wave’s Chimera architecture arranged qubits into linked cells with bipartite connectivity, while the later Pegasus architecture increased connectivity. A large physical-qubit count does not imply that an equally large arbitrary logical graph can be represented directly.
Mapping a logical graph onto a sparse hardware graph often requires minor embedding. A logical variable is represented by a connected set of physical qubits, commonly called a chain, with strong ferromagnetic interactions intended to keep them aligned. Couplings between chains implement the logical interactions. This makes sparse hardware more flexible, but it consumes qubits and introduces additional choices into the computation.
Chain strength creates a practical trade-off. Weak chains can break, producing inconsistent physical representations of a logical variable. Excessively strong chains can force the programmed problem coefficients to be reduced relative to the device’s available range. Decoding then becomes part of the solver: a broken chain may be resolved by majority vote or by an energy-aware classical procedure. Embedding, chain strength, and decoding can all change measured performance.
Analog imperfections affect the implemented Hamiltonian. Bias offsets, coupler errors, cross-talk, drift, and limited coefficient precision can perturb a programmed objective, particularly when relevant energy differences are small. Calibration reduces such effects but does not eliminate them. Spin-reversal transformations—changing selected spin signs and adjusting coefficients consistently—allow logically equivalent instances to be programmed in different physical forms. Aggregating these runs can reduce sensitivity to some systematic biases.
Other experimental platforms, including trapped ions and neutral atoms, support related analog evolutions and optimization encodings. Their native interactions and controls differ from superconducting annealers, so results must be interpreted in the context of the implemented model. Across platforms, scaling useful capability involves more than adding physical qubits: connectivity, noise, control resolution, and reproducibility determine which logical problems remain accessible.
Programming workflows and annealing protocols
A typical workflow begins with a mathematical model, not with hardware submission. The objective and constraints are translated into a binary quadratic formulation, simplified, scaled, and checked against small examples. Unsupported interactions are then embedded into the processor graph. A software interface sends coefficients and schedule settings to the device and receives multiple measured configurations, usually accompanied by energies and occurrence counts.
Repeated sampling is fundamental because an anneal is stochastic. Identical programmed instances can return different assignments, including excited states and invalid solutions to a penalized model. Samples must therefore be decoded, evaluated against the original constraints, and ranked using the original objective. Classical postprocessing may repair violations or improve local choices. Such processing is legitimate in a practical solver, but its contribution should be distinguished from the contribution of the quantum hardware.
Forward annealing begins with a strong driver and evolves toward the problem Hamiltonian. Reverse annealing instead starts from a selected classical configuration, increases quantum fluctuations partway, and then reduces them again. This permits exploration around an incumbent solution. It is not simply a forward anneal played backward: initialization and the full schedule determine the resulting search behavior. Reverse annealing is consequently useful to study as part of iterative classical–quantum workflows.
Pauses and quenches provide additional control. A pause holds the system at an intermediate schedule point, potentially allowing relaxation or redistribution among states. A rapid quench moves toward a measurement-compatible endpoint, although it can itself alter populations. Where supported, individual anneal offsets change the timing experienced by different qubits. These controls can help investigate bottlenecks, but they introduce parameters that must be tuned without giving one benchmark method an unfair advantage.
Large applications frequently use decomposition. A classical controller selects a subproblem small enough for the annealer, fixes or summarizes the remaining variables, calls the device, and combines the returned solution with a larger incumbent. Other hybrid methods coordinate an annealer with population-based search, tabu search, or mathematical programming. The resulting solver can be useful even when its logical problem is much larger than the hardware.
Attributing performance in such systems requires careful experiments. A useful test replaces the quantum component with a strong classical subsolver while retaining the rest of the workflow. This helps separate the value of decomposition, preprocessing, and software engineering from any value supplied by quantum evolution. Processor anneal time, programming time, repeated-readout costs, network delay, and classical processing should likewise be reported separately.
Benchmarking and quantum speedup
A meaningful performance claim must specify a task. Possible tasks include finding a ground state, reaching a target objective value, producing diverse low-energy configurations, or sampling from a specified distribution. An annealer can perform well on one of these and poorly on another. For example, repeatedly returning one good configuration may be useful for optimization but unsuitable for a sampling application that requires accurate relative probabilities.
For a fixed target, a common metric is time to solution. If independent runs succeed with probability , the number needed to reach a desired cumulative success probability is
Multiplying by per-run time gives a basic estimate. Real deployments may require additional terms for programming, readout, decoding, and communication. Correlations between samples, device drift, and uncertainty in the estimated success probability can complicate this calculation.
The annealing time itself should be optimized. Longer anneals may increase success probability while reducing the number of trials possible within a fixed budget. Conversely, an apparent scaling advantage can result from comparing a fixed short anneal against a poorly tuned classical method. Fair studies tune each solver using comparable procedures and separate tuning instances from evaluation instances. Random instance generation, precision requirements, and the definition of success must also be explicit.
Scaling is more informative than a single timing ratio. A claim of quantum speedup depends on how resources change as problem size or difficulty increases relative to an appropriate classical baseline. It also depends on the scope of the comparison: outperforming one selected classical algorithm is weaker than outperforming the best available classical methods. Finite-size experiments can reveal important behavior without establishing an asymptotic result.
Instance structure is particularly influential. Problems designed to contain narrow barriers or particular tunneling events can test physical mechanisms. Industrial problems, by contrast, may contain regularities exploited by presolve, decomposition, or specialized optimization. Strong classical baselines may include parallel tempering, cluster updates, tensor-network methods, or exact solvers, depending on the task. Classical GPU acceleration and parallel computing also affect practical comparisons.
Evidence of entanglement, tunneling, or agreement with quantum models establishes aspects of device behavior, not automatically computational superiority. Conversely, the absence of a speedup on one benchmark does not show that quantum effects were absent. These are separate experimental questions requiring separate measurements and controls.
Applications in optimization, sampling, and physics
Optimization is the most visible application area. Research prototypes have formulated traffic routing, production scheduling, graph partitioning, portfolio selection, facility location, and resource allocation as annealing problems. Their suitability depends heavily on encoding overhead. A task with a compact binary quadratic model may fit naturally, whereas detailed scheduling constraints or large numerical ranges can require many auxiliary variables and penalties.
Practical evaluation must preserve the original task. A routing model, for example, may approximate travel times, discretize available choices, or omit operational constraints to obtain an implementable QUBO. Solving that QUBO accurately does not necessarily solve the original routing problem to the same standard. Comparisons should therefore assess feasibility and objective quality in the original application, including any repair procedures. Demonstrating an encoding is an important first step, but not evidence of an end-to-end advantage.
Sampling provides a different use. Instead of seeking only the lowest energy, an application may need many configurations with controlled statistical properties. An equilibrium model would assign probabilities of the form
where is the partition function and . Actual annealer samples need not follow this distribution for the programmed final energy: non-equilibrium evolution, freeze-out, residual quantum effects, and control errors can all matter. An effective-temperature fit is therefore a model to test, not an automatic characterization.
This issue is important in machine learning. Annealers have been investigated for training energy-based models such as Boltzmann machines, whose learning procedures can require samples from difficult probability distributions. They have also been explored for feature selection and other combinatorial subproblems. These applications face competition from classical sampling algorithms and optimized digital hardware. Useful performance depends on sampling quality and total training cost, not merely the speed of generating device readouts.
Programmable annealers also function as experimental platforms for many-body physics. Researchers can study disordered magnets, frustrated lattices, spin-glass behavior, and non-equilibrium dynamics under controlled changes of interactions. Here the goal may be to reproduce or explore a physical process rather than to optimize an industrial objective. Relevant observables include correlations, defect densities, response functions, and scaling near phase transitions.
Such experiments connect quantum annealing with quantum simulation. A processor that is not advantageous for general optimization may still provide a useful realization of a particular quantum model. Its scientific value then depends on faithful control, accessible observables, and validation against theory or trusted calculations in regimes where comparison is possible.
Limits and active research questions
The central unresolved issue is where quantum dynamics supplies a reproducible computational benefit after realistic costs are included. This is narrower than asking whether an annealer is quantum and broader than asking whether it solves a selected benchmark quickly. A useful characterization must identify the relevant instance structure, the resources required, and the classical methods against which any benefit persists.
One research direction changes the evolution path rather than merely increasing its duration. Additional driver interactions or intermediate “catalyst” terms can alter avoided crossings and phase transitions. Nonstoquastic interactions, which do not retain the standard nonpositive off-diagonal form, have been investigated as ways to modify difficult dynamics. Such modifications can improve particular models but are harder to implement and do not provide a general guarantee of faster optimization.
Another direction addresses noise and protection. Physical redundancy, consistency checks, and energy penalties can reduce selected error mechanisms. These ideas differ from the full quantum error correction used in fault-tolerant circuit architectures. An embedded chain is also not automatically a protected logical qubit in that sense: its primary role is to represent one variable across a connectivity-limited device. Every protection scheme must be assessed against the additional qubits, couplings, and control it consumes.
Analog precision presents a separate scaling challenge. As logical problems grow, physically implementing the intended objective can become harder even when enough qubits are available. Small coefficient perturbations may reorder nearly degenerate configurations. Robust formulations seek to preserve useful solutions under such perturbations, while calibration and statistical techniques attempt to reduce their impact. The interaction between mathematical conditioning and hardware error remains important for practical problem design.
Verification is similarly task-dependent. A returned configuration can be checked for constraint violations and its energy computed, yet global optimality usually needs an independent bound or certificate. For sampling, verification involves distributional properties rather than only low energies. For many-body dynamics beyond classically tractable sizes, researchers rely on combinations of small-system agreement, limiting cases, symmetries, independent observables, and cross-platform comparisons.
Research also continues on schedule learning, reverse annealing, improved connectivity, better open-system models, and hybrid algorithms. Each addresses a different possible bottleneck: difficult spectral structure, environmental transitions, embedding overhead, or inefficient orchestration. Progress is consequently measured through specific capabilities and controlled comparisons rather than physical qubit count alone.