Improving QAOA to find approximate QUBO solutions in O(1) shots

Abstract: Quantum Approximate Optimization Algorithm (QAOA) provides one of the most promising quantum frameworks for addressing discrete optimization problems with broad real-world applications, particularly quadratic unconstrained binary optimization (QUBO) problems. However, the widely used hybrid quantum--classical implementation faces significant challenges due to statistical noise and barren plateaus. A prominent approach to mitigate these issues is fixed-point QAOA (fpQAOA), where circuit parameters are trained on small problem instances and subsequently transferred to larger ones. In this work, we develop a modified fpQAOA scheme that combines (i) considering the probability of achieving a target approximation ratio (AR) rather than requiring the exact optimum, (ii) setting the number of layers equal to the problem size with the sine--cosine encoding of QAOA angles, and (iii) rescaling the problem coefficient matrices to unit Frobenius norm. We demonstrate that this combination leads to a decreasing median number of shots required to obtain approximate solutions as the problem size increases, with ARs being within a few percent from the optimum for the considered problem classes. Extrapolation of these results suggests an $O(1)$ shot complexity while retaining at most quadratic circuit depth, underscoring the potential of our approach to overcome key bottlenecks in QAOA implementations and scalability estimations. Remarkably, omitting even a single one of the modifications (i)--(iii) results in exponential growth of the number of shots required as the problem size increases.