Paper detail

Optimization and learning of quantum programs

A programmable quantum processor is a fundamental model of quantum computation. In this model, any quantum channel can be approximated by applying a fixed universal quantum operation onto an input state and a quantum `program' state, whose role is to condition the operation performed by the processor. It is known that perfect channel simulation is only possible in the limit of infinitely large program states, so that finding the best program state represents an open problem in the presence of realistic finite-dimensional resources. Here we prove that the search for the optimal quantum program is a convex optimization problem. This can be solved either exactly, by minimizing a diamond distance cost function via semi-definite programming, or approximately, by minimizing other cost functions via gradient-based machine learning methods. We apply this general result to a number of different designs for the programmable quantum processor, from the shallow protocol of quantum teleportation, to deeper schemes relying on port-based teleportation and parametric quantum circuits. We benchmark the various designs by investigating their optimal performance in simulating arbitrary unitaries, Pauli and amplitude damping channels.