Rotary embeddings encode relative position exactly
ProofAdvanced⏱ 30 minDeep Learningroperelative positionrotationextrapolation
Problem
Prove that for the block-diagonal rotation of RoPE, and contrast with additive sinusoidal encodings, where relative position is only linearly recoverable.
Write a complete argument. Every inequality you use should be one you could state precisely.
Solution
Opening this before you have a real attempt on paper is the fastest way to feel like you understand something you cannot reproduce.
Related problems
All problems| Solved | Problem | Field | Type | Difficulty | Time |
|---|---|---|---|---|---|
| When does attention dominate the MLP? | Deep Learning | Numerical | 30 min | ||
| Attention is permutation equivariant | Deep Learning | Proof | 30 min | ||
| Implement multi-head attention from scratch run | Deep Learning | Coding | 3 hours | ||
| The softmax Jacobian as a variance, and saturation | Deep Learning | Proof | 30 min | ||
| Why attention divides by the square root of d_k | Deep Learning | Proof | 30 min |