Definition
Plain language
A small sphere drawn around a cluster of points so a search can skip the whole cluster at once when even its best-case member can't qualify.
As stated in the literature
In Louver-style attention sparsification, a tight enclosing sphere (center plus radius) over a group of key vectors; a single dot-product-plus-radius test upper-bounds any contained key's attention score, letting the index prune entire clusters that provably cannot exceed the threshold.
Also called: bounding balls, bounding sphere
Why it matters: It makes large-scale search dramatically faster by ruling out entire clusters with a single test, without missing any item that could have qualified.
For example, instead of checking a thousand stored items one by one, a search draws a sphere around a group and skips the whole group when even its closest possible member falls short.
Heard on the show
“The first idea is bounding balls.”Episode 036 — Sparse Attention Was the Wrong Frame. Treat It as Geometry Instead.