Low-variance observable estimation with informationally-complete measurements and tensor networks

Abstract: We propose a method for providing unbiased estimators of multiple observables with low statistical error by utilizing informationally (over)complete measurements and tensor networks. The technique consists of an observable-specific classical optimization of the measurement data based on tensor networks leading to low-variance estimations. Compared to other observable estimation protocols based on classical shadows and measurement frames, our approach offers several advantages: (i) it can be optimized to provide lower statistical error, resulting in a reduced measurement budget to achieve a specified estimation precision; (ii) it scales to a large number of qubits due to the tensor network structure; (iii) it can be applied to any measurement protocol with measurement operators that have an efficient representation in terms of tensor networks. We benchmark the method through various numerical examples, including spin and chemical systems in both infinite and finite statistics scenarios, and show how optimal estimation can be found even when we use tensor networks with low bond dimensions.