Abstract Andrey Lokhov

Title:  Predicting the behavior of quantum annealers with statistical learning

Abstract: Quantum annealers have a potential to provide a breakthrough in hard optimization and machine learning problems. Emerging physical implementations of quantum annealers are extremely sophisticated from the engineering point of view, and prediction of their performance remains a challenging problem. Here, we uncover the probabilistic relation between input and output of quantum annealers using the novel rigorous statistical learning tools. Extensive analysis of the output data allows us to check whether it satisfies the desired features assumed in the initial design of the device, to learn the machine's global response function, and to detect the echo of the chip architecture. In particular, our tests on D-Wave 2X and 2000Q quantum annealers revealed the presence of multi-body interactions and spurious next-nearest neighbor couplings between qubits as compared to the hardware-implemented topology of the chip. These results show how state-of-the-art statistical learning algorithms can quantify the performance of physical quantum annealers, suggest a path towards mitigating the effects of persistent biases inevitably present in every analog device, and guide the design of new hardware architectures.