Title :
On the impossibility of dimension reduction in ℓ1
Author :
Brinkman, Bo ; Charikar, Moses
Author_Institution :
Princeton Univ., NJ, USA
Abstract :
The Johnson-Lindenstrauss Lemma shows that any n points in Euclidean space (with distances measured by the ℓ2 norm) may be mapped down to O((log n)/ε2) dimensions such that no pairwise distance is distorted by more than a (1+ ε) factor. Determining whether such dimension reduction is possible in ℓ1 has been an intriguing open question. We show strong lower bounds for general dimension reduction in ℓ1. We give an explicit family of n points in ℓ1 such that any embedding with distortion δ requires nΩ(1δ2)/ dimensions. This proves that there is no analog of the Johnson-Lindenstrauss Lemma for ℓ1; in fact embedding with any constant distortion requires nΩ(1) dimensions. Further, embedding the points into ℓ1 with 1 + ε distortion requires n12 -O(εlog(1/ε))/ dimensions. Our proof establishes this lower bound for shortest path metrics of series-parallel graphs. We make extensive use of linear programming and duality in devising our bounds. We expect that the tools and techniques we develop will be useful for future investigations of embeddings into ℓ1.
Keywords :
computational complexity; duality (mathematics); graph theory; linear programming; Euclidean space; Johnson-Lindenstrauss Lemma; dimension reduction impossibility; duality; linear programming; lower bound; pairwise distance; point embedding; series-parallel graph; Algorithm design and analysis; Approximation algorithms; Clustering algorithms; Computer science; Embedded computing; Engineering profession; Extraterrestrial measurements; Linear programming; US Department of Energy; Upper bound;
Conference_Titel :
Foundations of Computer Science, 2003. Proceedings. 44th Annual IEEE Symposium on
Print_ISBN :
0-7695-2040-5
DOI :
10.1109/SFCS.2003.1238224