Jordan-RoPE: Non-Semisimple Relative Positional Encoding via Complex Jordan Blocks

Yaobo Zhang · 2026

Read Jordan-RoPE: Non-Semisimple Relative Positional Encoding via Complex Jordan Blocks by Yaobo Zhang (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-2605-04217.txt.

Description

Yaobo Zhang studies how neural networks tell a model where each word sits relative to others, a step that shapes what attention can learn. Standard methods add either a rotating phase or a simple distance penalty, each giving the model a limited set of patterns.

Zhang combines a rotating part and a fading part inside one mathematical block, producing patterns that wave and decay together, with distance itself scaling the wave. He gives the exact form, a stable variant that trades exactness for safer numbers, and tests at the kernel level plus on a synthetic task designed for this structure.

Catalog record

CREATORYaobo Zhang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE60 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.04217

Browse the full library · About the Tower of Babel