2016/02/22 by Kennedy, Christopher, Thijs Laarhoven, Rachel Ward · 7 citations
Computer Science · Engineering · Mathematics · #Advanced Image and Video Retrieval Techniques #Advanced Neural Network Applications #Algorithm #Combinatorics #Computation #Computer science #Discrete mathematics #Double hashing #Dynamic perfect hashing #Gaussian #Hash function #Hash table #K-independent hashing #Locality-sensitive hashing #Mathematics #Perfect hash function #Physics #Polytope #Randomness #Robotics and Sensor-Based Localization #Sensitivity (control systems) #Statistics #cs.CC #cs.CG #cs.CR #cs.DS
paper · pdf · open access · doi:10.4230/lipics.mfcs.2017.7
published in DROPS (Schloss Dagstuhl – Leibniz Center for Informatics) (Schloss Dagstuhl – Leibniz Center for Informatics) · 18 pages, 4 figures
openalex publication_date 2016/02/22 · openalex created_date 2016/06/24 · arxiv created 2017/02/19 · arxiv updated 2017/12/08 · openalex updated_date 2026/08/06
A celebrated technique for finding near neighbors for the angular distance involves using a set of random hyperplanes to partition the space into hash regions [Charikar, STOC 2002]. Experiments later showed that using a set of orthogonal hyperplanes, thereby partitioning the space into the Voronoi regions induced by a hypercube, leads to even better results [Terasawa and Tanaka, WADS 2007]. However, no theoretical explanation for this improvement was ever given, and it remained unclear how the resulting hypercube hash method scales in high dimensions. In this work, we provide explicit asymptotics for the collision probabilities when using hypercubes to partition the space. For instance, two near-orthogonal vectors are expected to collide with probability (1/pi)d in dimension d, compared to (1/2)d when using random hyperplanes. Vectors at angle pi/3 collide with probability (sqrt[3]/pi)d, compared to (2/3)d for random hyperplanes, and near-parallel vectors collide with similar asymptotic probabilities in both cases. For c-approximate nearest neighbor searching, this translates to a decrease in the exponent rho of locality-sensitive hashing (LSH) methods of a factor up to log2(pi) ~ 1.652 compared to hyperplane LSH. For c = 2, we obtain rho ~ 0.302 for hypercube LSH, improving upon the rho ~ 0.377 for hyperplane LSH. We further describe how to use hypercube LSH in practice, and we consider an example application in the area of lattice algorithms.