Jordan-RoPE: Non-Semisimple Relative Positional Encoding via Complex Jordan Blocks
Not on the map. Why?
Abstract (arXiv)
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-γ dcos(ω d), e-γ dsin(ω d), d e-γ dcos(ω d), and d e-γ dsin(ω d), for causal lag d=i-j≥ 0. Thus the construction realizes a distance-modulated phase basis d eiω d, 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.
Key facts. Cited by 1 paper in the map (data through 14 Oct 2026). Builds on 26 of its 31 references that are in the map. First posted on arXiv 5 May 2026.
Builds on
26 of 31 references.
References · 26
- 2017Attention Is All You Needcited by 60,361
- 2021RoFormer: Enhanced Transformer with Rotary Position Embeddingcited by 4,703
- 2016Pointer Sentinel Mixture Modelscited by 3,096
- 2021Efficiently Modeling Long Sequences with Structured State Spacescited by 2,465
- 2019Transformer-XL: Attentive Language Models Beyond a Fixed-Length Contextcited by 1,850
- 2018Self-Attention with Relative Position Representationscited by 1,120
- 2021Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolationcited by 1,103
- 2023Extending Context Window of Large Language Models via Positional Interpolationcited by 740
- 2020HiPPO: Recurrent Memory with Optimal Polynomial Projectionscited by 605
- 2023YaRN: Efficient Context Window Extension of Large Language Modelscited by 601
- 2023Hyena Hierarchy: Towards Larger Convolutional Language Modelscited by 393
- 2022Diagonal State Spaces are as Effective as Structured State Spacescited by 350
14 more
- 2023The Impact of Positional Encoding on Length Generalization in Transformerscited by 327
- 2024LongRoPE: Extending LLM Context Window Beyond 2 Million Tokenscited by 313
- 2018Enabling Factorized Piano Music Modeling and Generation with the MAESTRO Datasetcited by 217
- 2018Music Transformercited by 211
- 2016FMA: A Dataset For Music Analysiscited by 207
- 2022A Length-Extrapolatable Transformercited by 127
- 2022How to Train Your HiPPO: State Space Models with Generalized Orthogonal Basis Projectionscited by 122
- 2022KERPLE: Kernelized Relative Positional Embedding for Length Extrapolationcited by 86
- 2016Learning Features of Music from Scratchcited by 58
- 2023Universality of Linear Recurrences Followed by Non-linear Projections: Finite-Width Guarantees and Benefits of Complex Eigenvaluescited by 36
- 2024LieRE: Lie Rotational Positional Encodingscited by 18
- 2023Algebraic Positional Encodingscited by 10
- 2024Provable Benefits of Complex Parameterizations for Structured State Space Modelscited by 9
- 2025Group Representational Position Encodingcited by 5
Built on by
Citing papers per year: 2026: 1.
1 paper.