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

TL;DR

Jordan-RoPE uses complex Jordan blocks for non-semisimple relative positional encoding, outperforming RoPE on WikiText-103.

cs.LG 🔴 Advanced 2026-05-06 5 views
Yaobo Zhang
relative positional encoding complex Jordan blocks Transformer NLP machine learning

Key Findings

Methodology

The paper introduces a non-semisimple relative positional encoding method called Jordan-RoPE. It places a complex rotary eigenvalue and a nilpotent response in the same defective Jordan block, generating oscillatory-polynomial features like d e^{-γd}cos(ωd). This construction achieves a distance-modulated phase basis rather than merely adding a separate distance channel to RoPE.

Key Results

  • On a small WikiText-103 byte language model, the scaled-exact variant of Jordan-RoPE outperforms RoPE and direct-sum baselines within the Jordan family.
  • In a synthetic language-model task, the Jordan basis is useful when the target contains distance-modulated phase interactions.
  • On MusicNet, the scaled-exact order-three variant shows a natural positive case.

Significance

Jordan-RoPE offers a novel way of relative positional encoding by coupling complex rotary eigenvalues with nilpotent responses within the same Jordan block. This method excels in handling distance-modulated phase structures, making it potentially valuable for long-context modeling.

Technical Contribution

The technical contribution lies in proposing a non-semisimple relative positional encoding method that generates oscillatory-polynomial features. This method differs from existing RoPE and ALiBi methods, providing new theoretical guarantees and engineering possibilities.

Novelty

Jordan-RoPE is the first to couple complex rotary eigenvalues and nilpotent responses within the same Jordan block, generating a distance-modulated phase basis. This innovation fundamentally differs from existing RoPE and ALiBi methods.

Limitations

  • In long contexts, the polynomial growth of the nilpotent channels can be numerically or statistically awkward.
  • Scaled-exact variants require controlling long-lag growth.

Future Work

Future research could explore applying Jordan-RoPE on larger datasets and investigate its performance in other NLP tasks.

AI Executive Summary

Relative positional encoding plays a crucial role in Transformer models, yet existing methods like RoPE and ALiBi have limitations in long-context modeling. Jordan-RoPE addresses this by coupling complex rotary eigenvalues with nilpotent responses within the same Jordan block, generating oscillatory-polynomial features for a distance-modulated phase basis. Experimental results show that Jordan-RoPE outperforms RoPE and direct-sum baselines on a small WikiText-103 byte language model, especially in handling distance-modulated phase structures. It also performs well in synthetic language-model tasks and MusicNet. However, Jordan-RoPE may face polynomial growth issues in long contexts, and future research could optimize its numerical stability and scalability.

Deep Analysis

Background

Relative positional encoding in Transformers determines which previous representations should influence the current token. RoPE turns absolute rotations into a relative phase, while ALiBi adds a distance penalty directly to the logits. Although these methods differ operationally, they all choose the analytic form of the relative-position kernel.

Core Problem

Existing relative positional encoding methods perform poorly in long-context modeling, especially in handling distance-modulated phase structures. Achieving a distance-modulated phase basis without adding an independent distance channel is a key challenge.

Innovation

Jordan-RoPE couples complex rotary eigenvalues with nilpotent responses within the same Jordan block, generating oscillatory-polynomial features. This method provides a novel way of relative positional encoding and excels in handling distance-modulated phase structures.

Methodology

  • �� Place complex rotary eigenvalues and nilpotent responses in the same defective Jordan block
  • �� Generate oscillatory-polynomial features like d e^{-γd}cos(ωd)
  • �� Achieve a distance-modulated phase basis rather than adding a separate distance channel

Experiments

Experiments were conducted on a small WikiText-103 byte language model, comparing Jordan-RoPE with RoPE and direct-sum baselines. Additional tests were performed on synthetic language-model tasks and MusicNet.

Results

Results show that the scaled-exact variant of Jordan-RoPE outperforms RoPE and direct-sum baselines within the Jordan family, especially in handling distance-modulated phase structures.

Applications

Jordan-RoPE is potentially valuable for long-context modeling, especially in NLP tasks requiring distance-modulated phase structures.

Limitations & Outlook

While Jordan-RoPE excels in handling distance-modulated phase structures, it may face polynomial growth issues in long contexts. Future research could optimize its numerical stability and scalability.

Plain Language Accessible to non-experts

Imagine you're at a concert where each band member plays at different times. Traditional methods give each member a fixed schedule, while Jordan-RoPE allows them to adjust their playtime based on the audience's reaction. This flexibility makes the music more interactive and dynamic.

ELI14 Explained like you're 14

Hey, buddy! Imagine you're playing a game, and Jordan-RoPE is like a super-smart assistant that adjusts strategies in real-time based on your performance, making it easier for you to win! Isn't that cool?

Glossary

Relative Positional Encoding

Used in Transformers to determine which previous representations should influence the current token.

In this paper, it generates oscillatory-polynomial features.

Complex Jordan Block

A matrix form where complex eigenvalues and nilpotent responses are combined.

Used to generate a distance-modulated phase basis.

Oscillatory Polynomial Features

Features like d e^{-γd}cos(ωd) used to achieve a distance-modulated phase basis.

Generated by Jordan-RoPE.

Nilpotent Response

Combined with complex eigenvalues in a Jordan block to generate oscillatory-polynomial features.

Coupled with complex rotary eigenvalues in the same Jordan block.

Scaled-exact Variant

A Jordan-RoPE variant that controls long-lag growth.

Outperforms RoPE and direct-sum baselines in experiments.

Open Questions Unanswered questions from this research

  • 1 How can Jordan-RoPE be applied to larger datasets?
  • 2 What is the performance of Jordan-RoPE in other NLP tasks?

Applications

Immediate Applications

Long-context Modeling

Jordan-RoPE excels in handling long contexts, suitable for tasks requiring distance-modulated phase structures.

Long-term Vision

Natural Language Processing

Jordan-RoPE has potential applications across various NLP fields, especially in tasks requiring complex phase structures.

Abstract

Relative positional encodings determine which functions of query-key lag can enter the primitive attention logit. RoPE supplies a rotary phase, while ALiBi supplies an additive distance bias. Motivated by group-theoretic views of linear translation-invariant positional encodings, we study a non-semisimple case in which a complex rotary eigenvalue and a nilpotent response live in the same defective Jordan block. The resulting relative operator generates oscillatory-polynomial features such as $e^{-γd}\cos(ωd)$, $e^{-γd}\sin(ωd)$, $d e^{-γd}\cos(ωd)$, and $d e^{-γd}\sin(ωd)$, for causal lag $d=i-j\geq 0$. Thus the construction realizes a distance-modulated phase basis $d e^{iωd}$, rather than merely adding a separate distance channel to RoPE. We formulate Exact Jordan-RoPE as a non-semisimple one-parameter representation, give its real block form, and specify the contragredient query action required by non-orthogonal positional maps. We also distinguish this exact representation from stabilized variants whose bounded shear improves numerical behavior but breaks the exact group law. Kernel-level diagnostics and a Jordan-friendly synthetic language-model task show that the coupled Jordan basis is useful when the target contains distance-modulated phase interactions. On a small WikiText-103 byte language model, a scaled-exact variant improves over RoPE and direct-sum baselines within the Jordan family, while RoPE+ALiBi remains strongest overall. The evidence is structural rather than a broad performance claim.

cs.LG cs.CL