Agnostic Product Mixed State Tomography via Robust Statistics

We consider the problem of agnostic tomography with \emph{mixed state} ansatz, and specifically, the natural ansatz class of product mixed states. In more detail, given $N$ copies of an $n$-qubit state $ρ$ which is $ε$-close to a product mixed state $π$, the goal is to output a nearly-optimal product mixed state approximation to $ρ$. While there has been a flurry of recent work on agnostic tomography, prior work could only handle pure state ansatz, such as product states or stabilizer states. Here we give an algorithm for agnostic tomography of product mixed states which finds a product state which is $O(ε\log 1 / ε)$ close to $ρ$ which uses polynomially many copies of $ρ$, and which runs in polynomial time. Moreover, our algorithm only uses single-qubit, single-copy measurements. To our knowledge, this is the first efficient algorithm that achieves any non-trivial agnostic tomography guarantee for any class of mixed state ansatz. Our algorithm proceeds in two main conceptual steps, which we believe are of independent interest. First, we demonstrate a novel, black-box efficient reduction from agnostic tomography of product mixed states to the classical task of \emph{robustly learning binary product distributions} -- a textbook problem in robust statistics. We then demonstrate a nearly-optimal efficient algorithm for the classical task of robustly learning a binary product, answering an open problem in the literature. Our approach hinges on developing a new optimal certificate of closeness for binary product distributions that can be leveraged algorithmically via a carefully defined convex relaxation. Finally, we complement our upper bounds with a lower bound demonstrating that adaptivity is information-theoretically necessary for our agnostic tomography task, so long as the algorithm only uses single-qubit two-outcome projective measurements.