Robust shallow shadows

We present a robust shadow estimation protocol for wide classes of shallow measurement circuits that mitigates noise as long as the effective measurement map is locally unitarily invariant. This is in practice an excellent approximation, encompassing for instance the case of ideal single-qubit Clifford gates composing the first circuit layer of an otherwise arbitrary circuit architecture and even non-Markovian, gate-dependent noise in the rest of the circuit. We argue that for approximately local noise the measurement channel has an efficient matrix-product (tensor-train) representation, and show how to estimate this directly from experimental data using tensor-network tools, eliminating the need for analytical or numeric calculations. We illustrate the relevance of our method with both numerics and proof-of-principle experiments on an IBM Q device. Numerically, we show that, while unmitigated shallow shadows with noisy circuits become more biased as the depth increases, robust ones become more accurate for relevant parameter regimes. Experimentally, we observe major bias reductions in two simple fidelity estimation tasks using 5-qubit circuits with up to 2 layers of entangling gates using the mitigated variant, of close to an order of magnitude for 10⁴ measurement shots, e. g. Under the practical constraints of current and near-term noisy quantum devices, our method maximally realizes the potential of shadow estimation with global rotations.