Geodesics of Quantum Feature Maps on the space of Quantum Operators

Abstract: Selecting a quantum feature is an essential step in quantum machine learning. There have been many proposed encoding schemes and proposed techniques to test the efficacy of a scheme. From the perspective of information retention, this paper considers the smooth Riemannian geometry structure of a point cloud and how an encoding scheme deforms this geometry once mapped to the space of quantum operators, $\SU(2^N)$. However, a Riemannian manifold structure of the codomain of a quantum feature map has yet to be formalized. Using a ground-up approach, this manuscript mathematically establishes a Riemannian geometry for a general class of Hamiltonian quantum feature maps that are induced from a Euclidean embedded manifold. For this ground-up approach, we first derive a closed form of a vector space and a respective metric, and prove there is a 1-1 correspondence from geodesics on the embedded manifold to the codomain of the encoding scheme. We then rigorously derive closed form equations to calculate curvature. The paper ends with an example with a subset of the Poincar\'e half-plane and two well-used feature maps.