ERROR TRANSPORTATION IN SVD-BASED TENSOR-TRAIN APPROXIMATIONS

Read full text Summarise with VIRAT
Authors: Yassine Zniyed, Rémy Boyer
Cited by: 0
References: 33

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Tensor Train (TT) decompositions provide an efficient framework for representing and approximating high-dimensional tensors while mitigating the curse of dimensionality. Among the most widely used construction methods are SVD-based algorithms, in particular, the standard TT-SVD and its hierarchical generalization, TT-HSVD, which allows for flexible dimension trees and parallel computation. Despite their practical success, a unified algorithm-level understanding of the approximation error of SVD-like schemes for TT, especially for TT-HSVD on general dimension trees, has been lacking. In this work, we develop a comprehensive error analysis for SVD-like algorithms for TT approximation algorithms by explicitly tracking algorithmic matrices and the transport operators that propagate local truncations to the ambient tensor. We derive exact global error decompositions, establish a priori error bounds, and study quasi-optimality guarantees with respect to the best TT approximation of prescribed ranks. Our analysis shows that non-isometric transport operators on general trees induce data-dependent conditioning factors, while in the standard chain case these factors vanish, yielding sharp Pythagorean error identities. Finally, we introduce a transport-normalized variant of the TT-HSVD algorithm that achieves the best possible worstcase quasi-optimality bound within our framework, and we validate the theory through numerical experiments.

Fields of Study & Subjects

Tensor decomposition and applications Mathematical Approximation and Integration Model Reduction and Neural Networks

Publication details

Journal / source
HAL
Publication date
16 March 2026
Language
EN
Document type
Preprint