Year: 2026
Type: preprint
Abstract: Singular value thresholding (SVT) is a core subroutine in low-rank matrix recovery and quantum machine learning, but suffers from high computational complexity. The quantum singular value transformation (QSVT) algorithm improves performance, yet its resource usage still depends polynomially on the full rank \(r\) or dimension \(\sqrt{pq}\), leading to prohibitive costs for large-scale matrices. Based on the mathematical theory of low-rank matrix approximation, we propose a variational quantum algorithm framework, termed Variational Quantum Low-Rank SVT (VQLR-SVT). This framework utilizes a low-complexity variational quantum circuit to extract the low-dimensional submanifold spanned by the effective rank \(r_\epsilon\) of the matrix, and directly learns the threshold function from the phase estimation register, thereby compressing high-dimensional control information into a single flag qubit and implicitly realizing a low-dimensional approximation. Theoretical analysis shows that, under the mild assumption of variational training convergence, the resource complexity of QSVT operations can be reduced from \(\text{poly}(r)\) to \(\text{poly}(r_\epsilon)\), where \(r_\epsilon \ll r\). We rigorously prove the accuracy guarantee and complexity upper bound of the pre-compression algorithm using matrix perturbation theory. Subsequently, through five numerical simulation experiments, we validate the efficiency advantage in terms of resource consumption versus effective rank, the theoretical upper bound of output fidelity, circuit expressibility testing, convergence of variational quantum circuit training trajectories, and cross-validation with classical algorithms along with resource comparison. Experimental results demonstrate that in typical scenarios with rapidly decaying singular values, VQLR-SVT can reduce quantum gate calls by more than \(50\%\) while maintaining fidelity above 0.97, and the quantum resources exhibit scale-independence with respect to matrix dimensions.
Source: Research Square
Authors: Yiwei Chen, Xianhua Song, Zhixin Zhang
Institution: Harbin University of Science and Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236556911
