EstimatingtheL2norm.-People.csail.mit.edu.PDFVIP

L2 Norm Estimation MIT Piotr Indyk Lecture 2 L2 Norm Estimation Vector x : 1 2 ………………………m • A stream is a sequence of updates (i,a) x =x +a i i • Want to estimate ||x||2 up to 1±ε • Last week, we have seen how to do that for ||x|| : 0 – Space: (1/ε + log m)O(1) – Technique: • Linear sketches Sum (x)=∑ x for “random” sets S S i∈S i • (Somewhat messy) estimator • Today: two methods for estimating ||x|| +applications 2 – Alon-Matias-Szegedy - Really cute and simple – Johnson-Lindenstrauss - Need in future lectures • First: two digressions Lecture 2 Digression 1 • Our algorithm computes a linear sketch of the vector x: – Linear sketches Sum (x)=∑ x for “random” S i∈S i sets S • log(m)/ε values of T=1,1+ε, …, m • k sets S such that Pr[i ∈S ]=1/T j j – Can represent as a product of Ax, for a (log(m)/ε * k) x m 0-1 matrix A Lecture 2 Digression 2 • Our setup: – World: provides a stream, defining x – We: choose a random A – The method works with “high probability” • Comments: – Do not need to assume that a “source” generates x – Useful for composing algorithms, i.e., when x is itself an output of another algo