Counterdiabaticity and the quantum approximate optimization algorithm

TL;DR

This paper links QAOA to counterdiabatic evolution, showing how to optimize QAOA angles for better approximation ratios and demonstrating improved performance over traditional adiabatic methods in solving combinatorial problems.

Contribution

The authors introduce a method to construct counterdiabatic QAOA angles that mimic adiabatic schedules, enhancing QAOA's performance and providing a new perspective on its connection to adiabatic quantum algorithms.

Findings

QAOA angles can be optimized using counterdiabatic principles.

Approximation ratio improves as 1 - C(p) ~ 1/p^μ.

QAOA outperforms finite-time adiabatic evolution in examples.

Abstract

The quantum approximate optimization algorithm (QAOA) is a near-term hybrid algorithm intended to solve combinatorial optimization problems, such as MaxCut. QAOA can be made to mimic an adiabatic schedule, and in the $p\to\infty$ limit the final state is an exact maximal eigenstate in accordance with the adiabatic theorem. In this work, the connection between QAOA and adiabaticity is made explicit by inspecting the regime of $p$ large but finite. By connecting QAOA to counterdiabatic (CD) evolution, we construct CD-QAOA angles which mimic a counterdiabatic schedule by matching Trotter "error" terms to approximate adiabatic gauge potentials which suppress diabatic excitations arising from finite ramp speed. In our construction, these "error" terms are helpful, not detrimental, to QAOA. Using this matching to link QAOA with quantum adiabatic algorithms (QAA), we show that the approximation ratio converges to one at least as $1-C(p)\sim 1/p^{\mu}$. We show that transfer of parameters between graphs, and interpolating angles for $p+1$ given $p$ are both natural byproducts of CD-QAOA matching. Optimization of CD-QAOA angles is equivalent to optimizing a continuous adiabatic schedule. Finally, we show that, using a property of variational adiabatic gauge potentials, QAOA is at least counterdiabatic, not just adiabatic, and has better performance than finite time adiabatic evolution. We demonstrate the method on three examples: a 2 level system, an Ising chain, and the MaxCut problem.