Abstract
Due to the complexity of the space of quantum many-body states, the computation of expectation values by statistical sampling is, in general, a hard task. Neural network representations of such quantum states, which can be physically implemented by neuromorphic hardware, could enable efficient sampling. A scheme is proposed that leverages this capability to speed up sampling from so-called neural quantum states encoded by a restricted Boltzmann machine. Due to the complex network parameters, a direct hardware implementation is not feasible. We overcome this problem by considering a phase-reweighting scheme for sampling expectation values of observables. Applying our method to a set of paradigmatic entangled quantum states we find that, in general, the phase-reweighted sampling is subject to a form of sign problem, which renders the sampling computationally costly. The use of neuromorphic chips could allow reducing computation times and thereby extend the range of tractable system sizes.
- Received 1 August 2019
DOI:https://doi.org/10.1103/PhysRevB.100.195120
©2019 American Physical Society