RUS  ENG
Full version
JOURNALS // Sibirskie Èlektronnye Matematicheskie Izvestiya [Siberian Electronic Mathematical Reports] // Archive

Sib. Èlektron. Mat. Izv., 2024 Volume 21, Issue 2, Pages A70–A81 (Mi semr1767)

Collection of papers, dedicated to 85-th birthday of academician Vladimir Gavrilovich Romanov (Editors: S.I. Kabanikhin, M.A. Shishlenin)

Regularized Cholesky decomposition method for finite bit width computing

Z. Zhang, V. Lyashev

Moscow Institute of Physics and Technology, 9 Institutskiy Lane Dolgoprudny City Moskovskaya oblast 141700, Moscow, Russia

Abstract: This research focuses on the Cholesky decomposition of symmetric positive definite matrices. While the Cholesky decomposition is known for its computational efficiency and numerical robustness, it may encounter decomposition failures when applied to ill-conditioned matrices with large condition numbers. To address these computational challenges, this paper proposes an improved probabilistic rounding error analysis method. This method can more accurately estimate the rounding errors and thereby guide the selection of the optimal diagonal loading value. The main contribution of this research is the determination of a diagonal loading value applicable to all positive definite matrices, ensuring the successful completion of Cholesky decomposition. In addition, taking into account the binary representation of numbers in computers, the diagonal loading value is converted to exponential form, allowing multiplication to be replaced by the floating-point bitwise operations. This approach is both practical and efficient, effectively solving the challenges posed by ill-conditioned matrices and limited computational precision.

Keywords: cholesky decomposition, diagonal loading, regularization, numerical robustness, low bit-width computations, probabilistic rounding error analysis.

UDC: 519.61

MSC: 65F22

Received November 21, 2024, published December 31, 2024

Language: English

DOI: 10.33048/semi.2024.21.A04



© Steklov Math. Inst. of RAS, 2025