Johnson-Lindenstrauss lemma: embed to O(log n) dims with 1+ε distortion for k-means

FrnkNlsn · x · 2026-11-22

FrnkNlsn shares a teaching card on the Johnson-Lindenstrauss lemma: any n-point set in R^d can be embedded into k=O((1/ε²)·log n) dimensions with distortion bounded by 1±ε. A random projection matrix with i.i.d. standard normal entries (via Box-Muller) suffices, with sparse extensions. Practical upshot: high-dimensional embeddings can be reduced to O(log n) dims before running k-means, slashing compute with negligible structural loss.

Original post →

More from Research

Research channel →