We show that every symmetric normed space admits an efficient nearest neighbor search data structure with doubly-logarithmic approximation. Specifically, for every , , and every -dimensional symmetric norm , there exists a data structure for -approximate nearest neighbor search over for -point datasets achieving query time and space. The main technical ingredient of the algorithm is a low-distortion embedding of a symmetric norm into a low-dimensional iterated product of top- norms. We also show that our techniques cannot be extended to general norms.
View on arXiv