https://arxiv.org/abs/2008.00466 (or view on SciRate)
A promising approach to achieve computational supremacy over the classical von Neumann architecture explores classical and quantum hardware as Ising machines. The minimisation of the Ising Hamiltonian is known to be NP-hard problem for certain interaction matrix classes, yet not all problem instances are equivalently hard to optimise. We propose to identify computationally simple instances with an `optimisation simplicity criterion'. Such optimisation simplicity can be found for a wide range of models from spin glasses to k-regular maximum cut problems. Many optical, photonic, and electronic systems are neuromorphic architectures that can naturally operate to optimise problems satisfying this criterion and, therefore, such problems are often chosen to illustrate the computational advantages of new Ising machines. We further probe an intermediate complexity for sparse and dense models by analysing circulant coupling matrices, that can be `rewired' to introduce greater complexity. A compelling approach for distinguishing easy and hard instances within the same NP-hard class of problems can be a starting point in developing a standardised procedure for the performance evaluation of emerging physical simulators and physics-inspired algorithms.
This looks at the Ising model as a universal model for spin physics. MaxCut is a natural problem (mapping to the Ising model without overhead). They look at \(k\)-regular unweighted MaxCut.
How do you know which instances of a problem class are hard or easy?
Mobius ladder is the easiest type of problem, then cyclic and complete graphs, all the way up to arbitrary 3-regular MaxCut.
They do a numerical study to investigate the “hardness” of intermediate graphs: randomly generate instances, sometimes rewire edges, solve with Gurobi. I personally don’t find this approach that inspiring, but Gurobi is a good proxy for “practically hard”.
For circulant graphs, the hardness (to solve with Gurobi) corresponds with the instance’s distance to the phase transition value. They observe this as a “large optimality gap”.
They propose a “simplicity criterion”: The ground state minimizer is in the same direction as the eigenvector of the coupling matrix (i.e. the maximal eigenvector is also the correct solution). This makes the locally optimal solution the optimal solution. Although this works, I think this criterion is extremely restrictive and hard to intuit when it’s true.