Evidence for Super-Polynomial Advantage of QAOA over Unstructured Search

We compare the performance of several variations of the Quantum Alternating Operator Ansatz (QAOA) on constrained optimization problems. Specifically, we study the Clique, Ring, and Grover mixers as well as the traditional objective value and recently introduced threshold-based phase separators. These are studied through numerical simulation on k-Densest Subgraph, Maximum k-Vertex Cover, and Maximum Bisection problems of size up to n=18 on Erd\"os-Renyi graphs. We show that only one of these QAOA variations, the Clique mixer with objective value phase separator, outperforms Grover-style unstructured search, with a potentially super-polynomial advantage.