Sparsity Lower Bounds for Dimensionality Reducing Maps
Jelani Nelson
Member, School of Mathematics
January 22, 2013
Abstract:
We give near-tight lower bounds for the sparsity required in several dimensionality reducing linear maps. In particular, we show:
(1) The sparsity achieved by [Kane-Nelson, SODA 2012] in the sparse Johnson-Lindenstrauss lemma is optimal up to a log(1/eps) factor.
(2) RIP_2 matrices preserving k-space vectors in R^n with the optimal number of rows must be dense as long as k less than n / polylog(n).
(3) Any oblivious subspace embedding with 1 non-zero entry per column and preserving d-dimensional subspaces in R^n must have Omega(d^2) rows, matching an upper bound of [Nelson-Nguyen, 2012] for constant distortion.
Joint work with Huy Lê Nguyen (Princeton).

Computer Science and Discrete Mathematics (CSDM)

свернуть