Benchmarking a boson sampler with Hamming nets

Analyzing the properties of complex quantum systems is crucial for further development of quantum devices, yet this task is typically challenging and demanding with respect to the required amount of measurements. Special attention to this problem appears within the context of characterizing outcomes of noisy intermediate-scale quantum devices, which produce quantum states with specific properties so that it is expected to be hard to simulate such states using classical resources. In this work, we address the problem of characterization of a boson sampling device, which uses the interference of input photons to produce samples of nontrivial probability distributions that at certain condition are hard to obtain classically. For realistic experimental conditions the problem is to probe multiphoton interference with a limited number of the measurement outcomes without collisions and repetitions. By constructing networks on the measurement outcomes, we demonstrate the possibility to discriminate between regimes of indistinguishable and distinguishable bosons by quantifying the structures of the corresponding networks. Based on this, we propose a machine-learning-based protocol to benchmark a boson sampler with unknown scattering matrix. Notably, the protocol works in the most challenging regimes of having a very limited number of bitstrings without collisions and repetitions. As we expect, our framework can be directly applied for characterizing boson sampling devices that are currently available in experiments.