extreme-value-theory-17c588e0·1 events·first seen Aliases: Extreme Value Theory
A new arXiv technical report proposes a 'Certify-then-Rectify' framework that augments HNSW approximate nearest neighbor search with formal accuracy guarantees. The approach uses a distribution-free statistical certifier to assess search quality at runtime, escalating to exact retrieval only when needed, and reinterprets the HNSW graph as a geometric spanner to bound worst-case distances via Extreme Value Theory. Evaluations on benchmark datasets show the method preserves HNSW's average-case speed while guaranteeing worst-case correctness, outperforming comparable approaches.