Year: 2025
Type: article
Abstract: Singular Value Decomposition (SVD) is a very important matrix factorization technique in linear algebra which generalizes the eigenvalue decomposition to both non square and non symmetric matrices. This report explains the theoretical foundation of SVD by numerical examples and the comparison of SVD with eigenvalue decomposition on the basis of versatility. Theoretical derivations, including proofs of SVD existence and uniqueness, are presented with practical implementation using Python. Experiments include image compression via truncated SVD and dimensional reduction on the Iris dataset. The results indicate that only top k singular values are enough to retain the essential data features and reduce storage requirements. For example, image compression with 50 singular values results in a MSE of 0.05 and visually clear images, as well as 4D data can be reduced into 2D without losing discriminative patterns. Findings confirm that SVD is computationally stable and efficient, and provides robust solution for rank approximation, noise reduction, and feature extraction. The study demonstrates that SVD's capability to break down data into intelligible components make it to remain relevant in big data analytics, scientific computation and artificial intelligence, with foreseeable future improvements also likely to increase its applications.
Source: Theoretical and Natural Science
Author: Xinyi Zhao
Institution: New York University
Language: English
Topic: Matrix Theory and Algorithms
Subfield: Computational Theory and Mathematics
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4412495016