Numerical Linear Algebra with Applications Applied Mathematics

Total Least Squares With the Tensor Einstein Product: Algorithms and Regularization

Yimin Yang, Yimin Wei
Wiley · 2026-11-19 · DOI 10.1002/nla.70129

Abstract

ABSTRACT This paper extends the classical matrix TLS framework to tensor formulations under the Einstein product, allowing multidimensional data to be handled when both the mapping operator and the observation tensor contain errors. For the single right‐hand‐side case, we extend the Rayleigh quotient iteration preconditioned conjugate gradient (RQI‐PCG) method to tensors, yielding an efficient solver, and incorporate Golub–Kahan bidiagonalization (GKB) and Tikhonov regularization (TR) for ill‐posed problems. Since GKB‐TLS and TR‐TLS are restricted to single right‐hand‐side problems, we further introduce a randomized SVD approach for general multi‐right‐hand‐side tensor TLS in the low‐rank setting, which includes the single right‐hand‐side case as a special case. All methods exploit the tensor operator structure and evaluate the Einstein product directly without forming the large unfolded matrix, leading to significant computational savings, especially for multi‐right‐hand‐side problems where joint processing is more efficient than solving each right‐hand side independently. Numerical experiments on synthetic ill‐posed problems and image processing demonstrate the effectiveness and efficiency of the proposed tensor solvers.

Read at publisher →