Distinguish near-degenerate QUBO objectives from incomplete Dirac-3 convergence

Compare the QUBO energies of Dirac-3-selected feature subsets with Gurobi’s certified optimum to determine whether the observed subset diversity arises from near-degenerate objective landscapes or incomplete convergence of the stochastic Dirac-3 solver.

Background

The paper reports that Dirac-3 often selects subsets with downstream classifier performance different from Gurobi’s certified QUBO optima, while cross-method and repeated-run Jaccard similarities can be close to the random-subset expectation. Two explanations are proposed: many subsets may have nearly equal QUBO energy, or the stochastic solver may not have converged because only a limited number of samples were used. Comparing each Dirac-3 solution’s objective value with Gurobi’s certified optimum would resolve this ambiguity and clarify whether the observed diversity reflects the optimization landscape or solver behavior.