PJ-RoPE: A Fourier-Jet-Affine Position Space for Relative Attention

In three linesPJ-RoPE unifies RoPE's Fourier phase, Jordan-RoPE's finite jets, and ALiBi's affine recency into a single learnable relative-position space. The framework separates scalar PJ-bias kernels from exact PJ-rotary feature transforms, with adaptive sector diagnostics and LC/rapidity coordinates to stabilize high-order jets.

Summary generated by Claude — human-verified