Fast Random Projections Using Lean Walsh Transforms
Abstract
We present a kappa x d random projection matrix that is applicable to vectors x belongs to R(exp d) in O(d) operations if d greater than or equal to k(exp 2 + delta-prime) . Here, kappa is the minimal Johnson Lindenstrauss dimension and delta is arbitrarily small. The projection succeeds, with probability 1-1/n, in preserving vector lengths, up to distortion epsilon, for all vectors such that || x ||(sub infinity) less than or equal to || x ||(sub 2)k(exp -1/2)d(exp -delta) (for arbitrary small delta). Sampling based approaches are either not applicable in linear time or require a bound on || x ||(sub infinity) that is strongly dependant on d. Our method overcomes these shortcomings by rapidly applying dense tensor power matrices to incoming vectors.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 2007
- Accession Number
- ADA478655
Entities
People
- Amit Singer
- Edo Liberty
- Nir Ailon
Organizations
- Yale University