ℓp Subspace Embedding in Input Sparsity Time

Supratim Shit

Abstract: We study the distribution of matrices that can be used to preserve ℓp subspace embedding in input sparsity time, for integer p ∈ [2, ∞). We use the notion of power of two choice (Mitzenmacher, 2001) to design a distribution such matrices. For p = 2 case, we empirically compare our algorithm’s performance with an existing method such as CountSketch (Clarkson and Woodruff, 2017).