Clustering by quantum annealing on three-level quantum elements qutrits

TL;DR

This paper explores how using three-level quantum elements called qutrits in quantum neural networks can improve clustering efficiency, reducing resource requirements and accelerating hierarchical data partitioning compared to traditional qubit-based methods.

Contribution

The paper demonstrates the advantages of qutrit-based quantum clustering over qubit-based approaches, including fewer resources and faster hierarchical clustering, through numerical simulations and Hamiltonian modeling.

Findings

Qutrits require fewer resources than qubits for clustering.

Clustering on qutrits is more efficient for three-cluster problems.

Hierarchical clustering is accelerated using qutrits.

Abstract

Clustering is grouping of data by the proximity of some properties. We report on the possibility of increasing the efficiency of clustering of points in a plane using artificial quantum neural networks after the replacement of the two-level neurons called qubits represented by the spins S = 1/2 by the three-level neurons called qutrits represented by the spins S = 1. The problem has been solved by the slow adiabatic change of the Hamiltonian in time. The methods for controlling a qutrit system using projection operators have been developed and the numerical simulation has been performed. The Hamiltonians for two well-known cluster-ing methods, one-hot encoding and k-means ++, have been built. The first method has been used to partition a set of six points into three or two clusters and the second method, to partition a set of nine points into three clusters and seven points into four clusters. The simulation has shown that the clustering problem can be ef-fectively solved on qutrits represented by the spins S = 1. The advantages of clustering on qutrits over that on qubits have been demonstrated. In particular, the number of qutrits required to represent data points is smaller than the number of qubits by a factor of log2 N / log3 N . Since, for qutrits, it is easier to partition the data points into three clusters rather than two ones, the approximate hierarchical procedure of data partition-ing into a larger number of clusters is accelerated. At the exact data partition into more than three clusters, it has been proposed to number the clusters by the numbers of states of the corresponding multi-spin subsys-tems, instead of using the numbers of individual spins. This reduces even more the number of qutrits (N log3 K instead of NK ) required to implement the algorithm.