Solutions differ between neal and QPU
I'm still rather new to QC. I have found a way to implement an optimization problem (basically a matrix multiplication) into the QUBO formulation. The solution I want to obtain is a string of binary numbers in the (0,1) representation that are later interpreted as N integers. It follows the treatment described in [O'Malley and Vesselinov]. The QPU is a chimera graph (16x16x4).
If I run the optimization with neal, I always find the exact solution to Ax=y. However, if I use the QPU, it never finds the minimum even if I execute up to 10,000 cycles. By comparing the energy found by the QPU to the exact solution, they are numerically close but not the same, and bit-wise the two solutions do not really match. Interestingly, if I force the QPU to start from the state found by neal, and run a reverse annealing, the system remains in the same minimum.
I'm not sure what is causing this issue. I wonder if more experienced users can point me to similar examples where this behaviour was observed before, or can give me any useful suggestion to track down the origin of the problem.