Abstract
We propose a regression algorithm that utilizes a learned dictionaryoptimized for sparse inference on D-Wave quantum annealer. In this regressionalgorithm, we concatenate the independent and dependent variables as ancombined vector, and encode the high-order correlations between them into adictionary optimized for sparse reconstruction. On a test dataset, thedependent variable is initialized to its average value and then a sparsereconstruction of the combined vector is obtained in which the dependentvariable is typically shifted closer to its true value, as in a standardinpainting or denoising task. Here, a quantum annealer, which can presumablyexploit a fully entangled initial state to better explore the complex energylandscape, is used to solve the highly non-convex sparse coding optimizationproblem. The regression algorithm is demonstrated for a lattice quantumchromodynamics simulation data using a D-Wave 2000Q quantum annealer and goodprediction performance is achieved. The regression test is performed using sixdifferent values for the number of fully connected logical qubits, between 20and 64, the latter being the maximum that can be embedded on the D-Wave 2000Q.The scaling results indicate that a larger number of qubits gives betterprediction accuracy, the best performance being comparable to the bestclassical regression algorithms reported so far.