Linear-time classical approximate optimization of cubic-lattice classical spin glasses (2501.17267v5)

Abstract: Computing low-energy configurations (i.e., approximate optimization) of classical spin glasses is of relevance to both condensed matter and combinatorial optimization. Recent theoretical work opens the possibility to make its time complexity with quantum annealing generically polynomial, and D-Wave experiments can now achieve approximate optimization of cubic-lattice classical Ising spin glasses with $\sim$$10^4$ spins. It is therefore timely to ask which short-range classical spin glasses are good candidates for demonstrating quantum advantage in the time complexity of heuristic approximate optimization. One intuition is to consider models with very rugged energy landscapes in configuration space. However, here we provide evidence that short-range classical spin glasses may be approximately optimized in linear time and space with a very simple deterministic tensor-network heuristic regardless of ruggedness. On the cubic lattice with up to 50$\times$50$\times$50 spins, we obtain energy errors of $\lesssim$3% for the $\pm J$ model used in recent D-Wave experiments, and $\lesssim$5% for much more rugged planted-solution instances. For cubic-lattice-Ising reductions of unweighted Max-Cut on random 3-regular graphs with up to 300 vertices, we find energy errors of $<$1% and approximation ratios of about 72-88%. These results inform the search for quantum advantage and suggest an efficient classical method for generating warm starts for other spin-glass optimization algorithms. Our algorithm is amenable to massive parallelization and may also allow for low-power, accelerated implementations with photonic matrix-multiplication hardware.