Analysis of a Classical Matrix Preconditioning Algorithm

TL;DR

This paper provides the first theoretical convergence bounds for a classical matrix balancing algorithm used in numerical linear algebra, showing it converges in polynomial time with tight bounds.

Contribution

It proves a polynomial time convergence bound for a classical matrix balancing algorithm and characterizes matrices with order-independent limits, filling a longstanding gap in understanding.

Findings

Convergence bound of O(n^3 log(nρ/ε)) for the algorithm.

The bound is tight up to a log n factor.

Characterization of matrices with order-independent balancing limits.

Abstract

We study a classical iterative algorithm for balancing matrices in the $L_\infty$ norm via a scaling transformation. This algorithm, which goes back to Osborne and Parlett \& Reinsch in the 1960s, is implemented as a standard preconditioner in many numerical linear algebra packages. Surprisingly, despite its widespread use over several decades, no bounds were known on its rate of convergence. In this paper we prove that, for any irreducible $n\times n$ (real or complex) input matrix~$A$, a natural variant of the algorithm converges in $O(n^3\log(n\rho/\varepsilon))$ elementary balancing operations, where $\rho$ measures the initial imbalance of~$A$ and $\varepsilon$ is the target imbalance of the output matrix. (The imbalance of~$A$ is $\max_i |\log(a_i^{\text{out}}/a_i^{\text{in}})|$, where $a_i^{\text{out}},a_i^{\text{in}}$ are the maximum entries in magnitude in the $i$th row and column respectively.) This bound is tight up to the $\log n$ factor. A balancing operation scales the $i$th row and column so that their maximum entries are equal, and requires $O(m/n)$ arithmetic operations on average, where $m$ is the number of non-zero elements in~$A$. Thus the running time of the iterative algorithm is $\tilde{O}(n^2m)$. This is the first time bound of any kind on any variant of the Osborne-Parlett-Reinsch algorithm. We also prove a conjecture of Chen that characterizes those matrices for which the limit of the balancing process is independent of the order in which balancing operations are performed.