Glossary · Term

rank-1 approximation

← all terms

Definition

Plain language

Approximating something complicated by the single most important pattern it contains.

As stated in the literature

A low-rank approximation that represents a matrix or operator as the outer product of two vectors, capturing its dominant mode.

Also called: rank-1 factorization, rank-one

Why it matters: It's the simplest meaningful compression of a matrix and a building block for interpretability and efficient model edits — capturing the main signal while throwing away the rest.

For example, you might summarize a thousand customers' purchasing patterns with one shared trend vector plus a single number per customer indicating how strongly they follow it.

Heard on the show

“The authors approximate that machinery with what's called a rank-1 factorization, and I'm going to skip the math because the result is the only part that matters.”
Episode 038 — How LLMs Get Persuaded: One Attention Head, A Tetrahedron, And A Single Dial

Mentioned in 1 episode

  1. 038
    How LLMs Get Persuaded: One Attention Head, A Tetrahedron, And A Single Dial