Bounding Large-Scale Bell Inequalities

Bell inequalities are an important tool for studying non-locality, however quickly become computationally intractable as the system size grows. We consider a novel method for finding an upper bound for the quantum violation of such inequalities by combining the NPA hierarchy, the method of alternating projections, and the memory-efficient optimisation algorithm L-BFGS. Whilst our method may not give the tightest upper bound possible, it often does so several orders of magnitude faster than state-of-the-art solvers, with minimal memory usage, thus allowing solutions to problems that would otherwise be intractable. We benchmark using the well-studied I3322 inequality as well as a more general large-scale randomized inequality RXX22. For randomized inequalities with 130 inputs either side (a first-level moment matrix of size 261x261), our method is 100x faster than both MOSEK and SCS whilst giving a bound only 2% above the optimum.