Comparing classical and quantum QUBO solvers: optimality, trajectory diagnostics, and limits of solver selection

Comparing solvers for quadratic unconstrained binary optimization (QUBO) requires distinguishing solution quality from optimality certification. We compare classical, tensor-network, and fixed-schedule quantum approaches on 180 benchmark instances from ten families with 14-48 variables, supplemented by harder synthetic instances with known optimal solutions, extending to 128 variables. On the main benchmark, a general-purpose branch-and-cut solver matches all reference energies and certifies 176 instances within a 600,s cap. When the known optima are moved to randomly chosen bitstrings, the search comparison changes: at n = 48 the branch-and-cut solver finds only four of ten optima and certifies none, whereas parallel tempering reaches every optimum in every tested run at a 60,s budget. Structural and trajectory diagnostics explain aspects of tensor-network performance. A treewidth bound ranks trajectory bond-dimension demand at Spearman $ρ= 0.78$ in sample and $0.73$ out of sample, but provides only a coarse runtime estimate. Controlled experiments identify excited-state trapping during warm-started density-matrix renormalization group (DMRG) sweeps and demonstrate mitigation by restarting the variational search. These diagnostics do not yield a validated solver-selection rule. The tested linear-ramp quantum approximate optimization algorithm (LR-QAOA) runs offer no solution-quality improvement over the classical comparators. In the separate Wishart scaling study, on the harder setting, none of the classical methods tested reaches the optimum at the sampled sizes from n = 72 onward within the tested budgets. While this does not provide evidence for quantum advantage, it identifies a domain for further investigation.