Low-depth gradient measurements can improve convergence in variational hybrid quantum-classical algorithms

A broad class of hybrid quantum-classical algorithms known as "variational algorithms" have been proposed in the context of quantum simulation, machine learning, and combinatorial optimization as a means of potentially achieving a quantum speedup on a near-term quantum device for a problem of practical interest. Such algorithms use the quantum device only to prepare parameterized quantum states and make simple measurements. A classical controller uses the measurement results to perform an optimization of a classical function induced by a quantum observable which defines the problem. While most prior works have considered optimization strategies based on estimating the objective function and doing a derivative-free or finite-difference-based optimization, some recent proposals involve directly measuring observables corresponding to the gradient of the objective function. The measurement procedure needed requires coherence time barely longer than that needed to prepare a trial state. We prove that strategies based on such gradient measurements can admit substantially faster rates of convergence to the optimum in some contexts. We first introduce a natural black-box setting for variational algorithms which we prove our results with respect to. We define a simple class of problems for which a variational algorithm based on low-depth gradient measurements and stochastic gradient descent converges to the optimum substantially faster than any possible strategy based on estimating the objective function itself, and show that stochastic gradient descent is essentially optimal for this problem. Importing known results from the stochastic optimization literature, we also derive rigorous upper bounds on the cost of variational optimization in a convex region when using gradient measurements in conjunction with certain stochastic gradient descent or stochastic mirror descent algorithms.