Quantum Annealing vs QAOA for QUBO OptimizationAnalog Annealers Against Gate-Based Variational Circuits
Compare adiabatic quantum annealing with gate-based QAOA on the same QUBO and Ising optimization problems — hardware, constraint handling, scaling, and two real portfolio benchmarks.
Two Ways to Solve the Same QUBO
Quantum annealing and the Quantum Approximate Optimization Algorithm (QAOA) both target combinatorial optimization problems written as a quadratic unconstrained binary optimization (QUBO), or equivalently an Ising-model energy. They disagree on almost everything else. Annealing is an analog process: dedicated hardware evolves a quantum system from an easy initial Hamiltonian toward a problem Hamiltonian that encodes the QUBO cost, and if the schedule is slow enough and noise is manageable, the system settles into a low-energy state that corresponds to a good solution. QAOA is digital and hybrid: a gate-based circuit alternates a problem-specific cost unitary with a mixer unitary for p layers, parameterized by angles γ and β, while a classical optimizer runs outside the circuit and repeatedly updates those angles to reduce the expected cost.
The platform runs both. The quantum annealing pipeline solves a QUBO with the best available sampler — a D-Wave LeapHybrid solver or QPU when available, falling back to a classical simulated-annealing sampler (dimod) on a simulator otherwise. The QAOA pipeline runs the circuit on a simulator or small NISQ device, and can swap in a constraint-preserving mixer, such as a quantum walk mixer, in place of a penalty term.
- • Which hardware model and constraint-handling approach fits a given QUBO or Ising problem?
- • How do the two scale, in qubits, parameters, and noise sensitivity, as problem size grows?
- • What do the two approaches actually return on a real portfolio-optimization problem?
Why compare them
- Both target the same underlying objects — a QUBO cost function or its Ising-model equivalent — so the same problem can, in principle, be handed to either solver.
- Annealing runs on purpose-built analog hardware today; QAOA runs on gate-based hardware or a simulator, so the two sit on genuinely different hardware roadmaps.
- Running both against a shared classical baseline (tabu search, greedy, MILP) makes it possible to see where each one currently helps and where it does not.
Limits to keep explicit
- Neither has a proven, general quantum speedup; reported advantages are benchmark-specific and depend on error suppression, embedding, and problem structure.
- The two in-house results used here come from different problem instances, not a shared one — treat the numbers as evidence about each method, not a head-to-head race.
- Simulator runs avoid hardware noise but do not reproduce chain breaks, gate error, or qubit connectivity limits that a physical device would add.
Quantum Annealing vs QAOA at a Glance
Seven dimensions to check before picking a solver for a QUBO or Ising problem.
| Dimension | Quantum annealing | QAOA |
|---|---|---|
| Hardware model | Analog: continuous adiabatic evolution on purpose-built annealers (e.g. D-Wave), or a classical simulated-annealing sampler standing in for one | Digital: discrete gates on a universal gate-model quantum computer or a statevector/shot-based simulator |
| Constraint handling | Penalty terms folded into the QUBO objective (e.g. an exact-count term for a K-of-N constraint) | Penalty terms, or a constraint-preserving mixer — a quantum walk mixer restricts the search to the feasible subspace directly |
| Parameters to tune | Anneal time, schedule shape, chain strength, penalty coefficient | Number of layers p, angle pairs (γ, β) per layer, mixer choice, classical optimizer (COBYLA, SPSA, Nelder–Mead) |
| Scaling with problem size | Embedding overhead grows with graph density; on some benchmark families, time-to-solution scales with a smaller power-law exponent than a strong classical heuristic | Circuit depth grows with p and connectivity; shot budget and classical-optimizer iterations both grow with problem size |
| Noise behaviour | Chain breaks and embedding-induced errors on physical hardware; suppressed with techniques such as quantum annealing correction | Gate and readout noise accumulate with depth; deeper circuits (higher p) are more exposed, and parameters can concentrate or plateau |
| Typical problem sizes today | Tens of logical qubits in research PoCs; hybrid solvers on production annealers are built to reach far larger variable counts | Small graphs and instances — roughly 8–16 nodes for MaxCut, single-digit-to-low-double-digit qubit counts for constrained portfolio selection |
| Where each shines | Large, sparse combinatorial problems that map cleanly onto a QUBO with soft, penalty-encoded constraints | Problems with a hard constraint that must never be violated, where a problem-specific mixer exists, at sizes gate hardware can currently support |
A Worked Comparison: Two Portfolio-Optimization Benchmarks
The clearest evidence on this platform comes from quantum portfolio optimization, which runs both algorithms — but on two different formulations of portfolio selection, not the same instance. That distinction matters: these numbers show what each algorithm does on the constraint shape it is suited to, not which one wins a fair fight on identical inputs.
Quantum annealing: cardinality-constrained selection
The task is to choose exactly K of N assets to balance expected return against risk — a K-of-N cardinality constraint folded into the QUBO as a penalty term alongside the risk and return terms. The universe is ten large-cap tickers (AAPL, MSFT, GOOGL, AMZN, META, NVDA, JPM, XOM, V, PG) with daily prices from 2018 to 2024, with the constraint set to K = 6 of N = 10. The resulting QUBO runs on 10 logical qubits. This particular run used dimod's classical simulated-annealing sampler on a simulator — the fallback tier below D-Wave's LeapHybrid solver and QPU access in the platform's sampler order — and completed in about two minutes.
Sharpe
0.123
Annual return
40.4%
Annual volatility
18.25%
Qubits
10
Out of sample, the annealing-driven portfolio placed 2nd of 4 strategies, with a Sharpe ratio comparable to the top classical baseline. It beat Equal-Weight-All's 36.9% return and 14.6% volatility, and sat close to Top-K-Mean's 49.6% return, which came at a materially higher 22.3% volatility.
QAOA with a quantum walk mixer: knapsack selection
The task here is different in kind: choose the subset of assets that maximizes value under a hard budget or risk-capacity cap — a knapsack problem, not an exact count. On a 5-asset instance, constrained QAOA runs on 11 qubits on a simulator, with a Quantum Walk Mixer that keeps the search inside the feasible subspace instead of penalizing infeasible selections after the fact. The same instance is benchmarked against MILP and heuristic baselines, following the same setup detailed in QAOA vs. MILP.
Instance
5 assets
Qubits
11
Execution
Simulator
The run returns a feasible constrained-QAOA solution with an objective value comparable to the MILP and heuristic baselines, with per-constraint feasibility reporting and seed-controlled reproducibility — a different kind of result from the annealing case's Sharpe ratio, because the two are answering different questions about the same asset class.
When to Choose Quantum Annealing vs QAOA
The honest answer starts with your constraint, not with which algorithm sounds more advanced.
Choose quantum annealing when
- Your constraint is an exact count or another structure that folds cleanly into a QUBO penalty term.
- The problem is large and sparse, and you want a path to hybrid solvers that scale toward far larger variable counts than a gate-based circuit reaches today.
- You want a single tunable QUBO that can run on a classical simulated-annealing sampler today and move to hybrid or QPU hardware later without reformulating.
Choose QAOA when
- A constraint must never be violated, and a problem-specific mixer (a quantum walk mixer, for instance) can keep the search inside the feasible subspace.
- Your instance is small to medium — sizes gate-based simulators or NISQ devices can currently support at a reasonable circuit depth.
- You need a specific gate-model hardware target — QAOA is a circuit, and it does not run on an analog annealer such as a D-Wave system.
Who Should Use This Comparison
Use it when a QUBO or Ising problem needs a solver decision, not just a demo of either algorithm alone.
10 qubits
Annealing cardinality case
11 qubits
QAOA knapsack case
0.123
Annealing portfolio Sharpe ratio
Millions
Variables targeted by hybrid annealing solvers
Optimization and operations research teams
For teams with a QUBO- or Ising-shaped problem who need to know which hardware model and constraint-handling approach actually fits it.
Portfolio and risk teams
For teams choosing between a cardinality selection and a hard budget or risk-capacity cap, and wanting real numbers behind each formulation.
Quantum R&D leads
For teams deciding whether to pursue annealing hardware access, gate-model experimentation, or both, before committing engineering time.
How It Works on the Platform
One QUBO formulation, two solvers, one reproducible comparison.
Formulate
Write the problem as a QUBO/Ising cost function, with hard constraints as penalty terms or, for QAOA, a candidate mixer
Pick a solver
Route the same formulation to quantum annealing, to QAOA, or run both side by side against a classical baseline
Configure
Set anneal schedule and chain strength, or QAOA layers p, angles (γ, β), mixer, and classical optimizer
Execute
Run on a simulator today, with hybrid or QPU backends available for annealing as hardware access allows
Export
Compare against classical baselines and download seed-controlled code, metrics, and a reproducible report
Frequently Asked Questions
Is quantum annealing the same as QAOA?
No. Quantum annealing is an analog process: dedicated hardware continuously evolves toward a low-energy state of a problem Hamiltonian. QAOA is a digital, gate-based circuit that alternates a fixed number of cost and mixer layers, tuned by a classical optimizer running outside the circuit. Both can target the same QUBO or Ising cost function, but the physical mechanism and the hardware they need are different.
Which is better for portfolio optimization?
It depends on the constraint, not on which algorithm is faster. The in-house quantum-annealing case solves an exact K-of-N cardinality selection (choose exactly 6 of 10 assets), while the in-house QAOA case solves a knapsack-style hard budget cap on a 5-asset instance using a quantum walk mixer. Different constraint shapes, different formulations — see the full breakdown in quantum portfolio optimization.
Can QAOA run on a D-Wave?
No. D-Wave builds purpose-built quantum annealers, which perform analog adiabatic evolution — they do not implement arbitrary quantum gates. QAOA needs a universal gate-based quantum computer or a gate-model simulator. D-Wave's own hybrid solvers address the same QUBO problems directly through annealing and classical hybrid methods, not by running a QAOA circuit.
How many variables can quantum annealing handle today?
It depends on the sampler. The in-house cardinality-selection benchmark ran on 10 logical qubits for a 10-asset universe on a simulated-annealing sampler. On the hardware side, D-Wave's production Advantage2 annealer and its hybrid solvers are built to reach far larger variable counts for enterprise-style workflows, though native QPU problem size is still limited by qubit connectivity and embedding overhead.
Does either algorithm offer a proven quantum speedup?
Not a general one. On specific structured benchmark families, quantum annealing with error suppression has shown a smaller scaling exponent than a strong classical heuristic in time-to-solution — a possible, not proven, scaling advantage. QAOA's performance is heuristic and instance-dependent, and neither approach guarantees beating a good classical solver on an arbitrary problem.
Do I need real quantum hardware to see these results?
No. Every result referenced on this page ran on a simulator: a classical simulated-annealing sampler standing in for the annealer, and a circuit simulator for QAOA. The platform can route the annealing pipeline to a D-Wave hybrid solver or QPU when available, and it supports QAOA as well — both algorithms are available on the platform today.
Related pages
Algorithms, applied use cases and benchmarks connected to this page
Algorithms
Quantum Annealing for QUBO and Ising Problems
Map combinatorial optimization problems to QUBO and Ising form and benchmark quantum annealing with reproducible runs, baselines, and hardware notes.
Algorithms
QAOA for QUBO and Ising Optimization
Explore the Quantum Approximate Optimization Algorithm for QUBO and Ising problems with reproducible benchmarks, parameter schedules, and hardware notes.
Finance
Quantum Portfolio Optimization: QAOA, HHL and Annealing
Compare quantum portfolio optimization: QAOA with quantum walk mixers, HHL mean-variance allocation, and quantum annealing for cardinality-constrained QUBOs.
Comparisons
QAOA vs MILP for Portfolio Knapsack Optimization
Compare constrained QAOA with MILP on a 5-asset portfolio knapsack PoC using an 11-qubit simulator run, feasibility checks, and objective-quality audits.
Run Your QUBO Through Both Solvers
Compare quantum annealing and QAOA on your own optimization problem and review feasibility, objective value, and scaling against classical baselines.
Try your first case for free