Matrix Balancing in Lp Norms: A New Analysis of Osborne's Iteration

TL;DR

This paper analyzes the convergence of Osborne's matrix balancing algorithm in various Lp norms, especially L1, providing new bounds and resolving open problems in the field.

Contribution

It offers the first analysis of Osborne's iteration in the L1 norm, establishing convergence bounds and lower bounds, and compares different implementation strategies.

Findings

Convergence in L1 norm with greedy order within specific iteration bounds.

Round-robin implementation converges in a different iteration complexity.

A lower bound on the convergence rate of any implementation is established.

Abstract

We study an iterative matrix conditioning algorithm due to Osborne (1960). The goal of the algorithm is to convert a square matrix into a balanced matrix where every row and corresponding column have the same norm. The original algorithm was proposed for balancing rows and columns in the $L_2$ norm, and it works by iterating over balancing a row-column pair in fixed round-robin order. Variants of the algorithm for other norms have been heavily studied and are implemented as standard preconditioners in many numerical linear algebra packages. Recently, Schulman and Sinclair (2015), in a first result of its kind for any norm, analyzed the rate of convergence of a variant of Osborne's algorithm that uses the $L_{\infty}$ norm and a different order of choosing row-column pairs. In this paper we study matrix balancing in the $L_1$ norm and other $L_p$ norms. We show the following results for any matrix $A = (a_{ij})_{i,j=1}^n$, resolving in particular a main open problem mentioned by Schulman and Sinclair. 1) We analyze the iteration for the $L_1$ norm under a greedy order of balancing. We show that it converges to an $\epsilon$-balanced matrix in $K = O(\min\{\epsilon^{-2}\log w,\epsilon^{-1}n^{3/2}\log(w/\epsilon)\})$ iterations that cost a total of $O(m + Kn\log n)$ arithmetic operations over $O(n\log w)$-bit numbers. Here $m$ is the number of non-zero entries of $A$, and $w = \sum_{i,j} |a_{ij}|/a_{\min}$ with $a_{\min} = \min\{|a_{ij}|:\ a_{ij}\neq 0\}$. 2) We show that the original round-robin implementation converges to an $\epsilon$-balanced matrix in $O(\epsilon^{-2}n^2\log w)$ iterations totalling $O(\epsilon^{-2}mn\log w)$ arithmetic operations over $O(n\log w)$-bit numbers. 3) We demonstrate a lower bound of $\Omega(1/\sqrt{\epsilon})$ on the convergence rate of any implementation of the iteration.