Analysis of stochastic Lanczos quadrature for spectrum approximation

05/13/2021
βˆ™
by   Tyler Chen, et al.
βˆ™
0
βˆ™

The cumulative empirical spectral measure (CESM) Ξ¦[𝐀] : ℝ→ [0,1] of a nΓ— n symmetric matrix 𝐀 is defined as the fraction of eigenvalues of 𝐀 less than a given threshold, i.e., Ξ¦[𝐀](x) := βˆ‘_i=1^n1/nx1D7D9[ Ξ»_i[𝐀]≀ x]. Spectral sums tr(f[𝐀]) can be computed as the Riemann–Stieltjes integral of f against Ξ¦[𝐀], so the task of estimating CESM arises frequently in a number of applications, including machine learning. We present an error analysis for stochastic Lanczos quadrature (SLQ). We show that SLQ obtains an approximation to the CESM within a Wasserstein distance of t | Ξ»_max[𝐀] - Ξ»_min[𝐀] | with probability at least 1-Ξ·, by applying the Lanczos algorithm for ⌈ 12 t^-1 + 1/2βŒ‰ iterations to ⌈ 4 ( n+2 )^-1t^-2ln(2nΞ·^-1) βŒ‰ vectors sampled independently and uniformly from the unit sphere. We additionally provide (matrix-dependent) a posteriori error bounds for the Wasserstein and Kolmogorov–Smirnov distances between the output of this algorithm and the true CESM. The quality of our bounds is demonstrated using numerical experiments.

READ FULL TEXT

page 9

page 17

page 18

research
βˆ™ 11/28/2022

A posteriori error bounds for the block-Lanczos method for matrix function approximation

We extend the error bounds from [SIMAX, Vol. 43, Iss. 2, pp. 787-811 (20...
research
βˆ™ 06/17/2021

Error bounds for Lanczos-based matrix function approximation

We analyze the Lanczos method for matrix function approximation (Lanczos...
research
βˆ™ 12/07/2020

The Spectral-Domain 𝒲_2 Wasserstein Distance for Elliptical Processes and the Spectral-Domain Gelbrich Bound

In this short note, we introduce the spectral-domain 𝒲_2 Wasserstein dis...
research
βˆ™ 12/09/2019

Error control for statistical solutions

Statistical solutions have recently been introduced as a an alternative ...
research
βˆ™ 06/15/2021

Non-asymptotic convergence bounds for Wasserstein approximation using point clouds

Several issues in machine learning and inverse problems require to gener...
research
βˆ™ 08/30/2012

An Improved Bound for the Nystrom Method for Large Eigengap

We develop an improved bound for the approximation error of the NystrΓΆm ...
research
βˆ™ 04/12/2019

The Lanczos Algorithm Under Few Iterations: Concentration and Location of the Ritz Values

We study the Lanczos algorithm where the initial vector is sampled unifo...

Please sign up or login with your details

Forgot password? Click here to reset