Limitations of tensor network approaches for optimization and sampling: A comparison to quantum and classical Ising machines

TL;DR

This paper compares tensor network algorithms to quantum and classical Ising machines for optimization and sampling, revealing limitations in solution quality and diversity, especially for large-scale problems, due to approximate contraction failures.

Contribution

It introduces a heuristic tensor network approach with branch-and-bound for Ising systems and benchmarks it against quantum annealers and bifurcation algorithms, highlighting its strengths and weaknesses.

Findings

Tensor network approach yields solutions within 0.1-1% of best but is slower.

For large problems, tensor networks are less accurate than Ising machines.

All methods can sample diverse low-energy states, but tensor networks find fewer such states.

Abstract

Optimization problems pose challenges across various fields. In recent years, quantum annealers have emerged as a promising platform for tackling such challenges. To provide a new perspective, we develop a heuristic tensor network (TN) based algorithm to reveal the low-energy spectrum of Ising spin-glass systems with interaction graphs relevant to present-day quantum annealers. Our deterministic approach combines a branch-and-bound search strategy with an approximate calculation of marginals via TN contractions. Its application to quasi-two-dimensional lattices with large unit cells of up to 24 spins, realized in current quantum annealing processors, requires a dedicated approach that utilizes sparse structures in the TN representation and GPU hardware acceleration. We benchmark our approach on random problems defined on Pegasus and Zephyr graphs with up to a few thousand spins, comparing it against the D-Wave Advantage quantum annealer and Simulated Bifurcation algorithm. Apart from the quality of the best solutions, we compare the diversity of low-energy states sampled by all the solvers. For the biggest considered i.i.d. problems with over 5000 spins, the state-of-the-art TN approach leads to solutions that are $0.1\%$ to $1\%$ worse than the best solutions obtained by Ising machines while being two orders of magnitude slower. We attribute those results to approximate contraction failures. For embedded tile planting instances, our approach gets to approximately $0.1\%$ from the planted ground state, a factor of $3$ better than the Ising solvers. While all three methods can output diverse low-energy solutions, e.g., differing by at least a quarter of spins with energy error below $1\%$, our deterministic branch-and-bound approach finds sets of a few such states at most. On the other hand, both Ising machines prove capable of sampling sets of thousands of such solutions.