We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.DS

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Data Structures and Algorithms

Title: Faster Kernel Matrix Algebra via Density Estimation

Abstract: We study fast algorithms for computing fundamental properties of a positive semidefinite kernel matrix $K \in \mathbb{R}^{n \times n}$ corresponding to $n$ points $x_1,\ldots,x_n \in \mathbb{R}^d$. In particular, we consider estimating the sum of kernel matrix entries, along with its top eigenvalue and eigenvector.
We show that the sum of matrix entries can be estimated to $1+\epsilon$ relative error in time $sublinear$ in $n$ and linear in $d$ for many popular kernels, including the Gaussian, exponential, and rational quadratic kernels. For these kernels, we also show that the top eigenvalue (and an approximate eigenvector) can be approximated to $1+\epsilon$ relative error in time $subquadratic$ in $n$ and linear in $d$.
Our algorithms represent significant advances in the best known runtimes for these problems. They leverage the positive definiteness of the kernel matrix, along with a recent line of work on efficient kernel density estimation.
Subjects: Data Structures and Algorithms (cs.DS); Machine Learning (cs.LG); Numerical Analysis (math.NA)
Cite as: arXiv:2102.08341 [cs.DS]
  (or arXiv:2102.08341v2 [cs.DS] for this version)

Submission history

From: Cameron Musco [view email]
[v1] Tue, 16 Feb 2021 18:25:47 GMT (152kb,D)
[v2] Thu, 17 Jun 2021 18:18:57 GMT (176kb,D)

Link back to: arXiv, form interface, contact.