Fast and Stable Gradient Approximation for Bilinear Forms of Hermitian Matrix Functions

Navjot Singh, Kipton Barros, Xiaoye Sherry Li · 2026

Read Fast and Stable Gradient Approximation for Bilinear Forms of Hermitian Matrix Functions by Navjot Singh, Kipton Barros, Xiaoye Sherry Li (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-12801.txt.

Description

Singh, Barros and Li propose a fast, stable gradient approximation for bilinear forms of Hermitian matrix functions. Such quantities show up across scientific computing and probabilistic machine learning. With large matrices, Lanczos approximates them well, but differentiating them is hard. Existing methods backpropagate through the Lanczos recurrence, needing reorthogonalization for stability, or apply Arnoldi to a doubled block matrix. Their forward-only method reuses the Lanczos pass with minimal overhead, and its error scales with the Lanczos residual norm.

Catalog record

CREATORNavjot Singh, Kipton Barros, Xiaoye Sherry Li
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE48940
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.12801

Browse the full library · About the Tower of Babel