Abstract:The Kaczmarz and Gauss-Seidel methods aim to solve a linear $m \times n$ system $\boldsymbol{X} \boldsymbol{\beta} = \boldsymbol{y}$ by iteratively refining the solution estimate; the former uses random rows of $\boldsymbol{X}$ {to update $\boldsymbol{\beta}$ given the corresponding equations} and the latter uses random columns of $\boldsymbol{X}$ {to update corresponding coordinates in $\boldsymbol{\beta}$}. Interest in these methods was recently revitalized by a proof of Strohmer and Vershynin showing linear convergence in expectation for a \textit{randomized} Kaczmarz method variant (RK), and a similar result for the randomized Gauss-Seidel algorithm (RGS) was later proved by Lewis and Leventhal. Recent work unified the analysis of these algorithms for the overcomplete and undercomplete systems, showing convergence to the ordinary least squares (OLS) solution and the minimum Euclidean norm solution respectively. This paper considers the natural follow-up to the OLS problem, ridge regression, which solves $(\boldsymbol{X}^* \boldsymbol{X} + \lambda \boldsymbol{I}) \boldsymbol{\beta} = \boldsymbol{X}^* \boldsymbol{y}$. We present particular variants of RK and RGS for solving this system and derive their convergence rates. We compare these to a recent proposal by Ivanov and Zhdanov to solve this system, that can be interpreted as randomly sampling both rows and columns, which we argue is often suboptimal. Instead, we claim that one should always use RGS (columns) when $m > n$ and RK (rows) when $m < n$. This difference in behavior is simply related to the minimum eigenvalue of two related positive semidefinite matrices, $\boldsymbol{X}^* \boldsymbol{X} + \lambda \boldsymbol{I}_n$ and $\boldsymbol{X} \boldsymbol{X}^* + \lambda \boldsymbol{I}_m$ when $m > n$ or $m < n$.

Convergence properties of the randomized extended Gauss-Seidel and Kaczmarz methods

Refined upper bounds for the convergence of the randomized extended Kaczmarz and Gauss-Seidel algorithms

On Multi-Step Randomized Extended Kaczmarz Method for Solving Large Sparse Inconsistent Linear Systems

On Convergence Rates of Kaczmarz-Type Methods with Different Selection Rules of Working Rows

Iterative methods for solving factorized linear systems

Acceleration of Randomized Kaczmarz Method via the Johnson-Lindenstrauss Lemma

Randomized Methods for Linear Constraints: Convergence Rates and Conditioning

Convergence Behavior of Gauss-Newton's Method and Extensions of the Smale Point Estimate Theory.

Randomized Kaczmarz iteration methods: Algorithmic extensions and convergence theory

Randomized block Kaczmarz method with projection for solving least squares

Rows vs. Columns: Randomized Kaczmarz or Gauss-Seidel for Ridge Regression

On the convergence analysis of the greedy randomized Kaczmarz method

Randomized Kaczmarz method with adaptive stepsizes for inconsistent linear systems

Randomized Iterative Methods for Linear Systems

Preasymptotic convergence of randomized Kaczmarz method

Randomized Kaczmarz Algorithm for Inconsistent Linear Systems: An Exact MSE Analysis

Convergence Rates for Greedy Kaczmarz Algorithms, and Faster Randomized Kaczmarz Rules Using the Orthogonality Graph

On block Gaussian sketching for the Kaczmarz method

Nonlinear Kaczmarz algorithms and their convergence

Randomized Kaczmarz Converges Along Small Singular Vectors

Randomized Extended Kaczmarz for Solving Least-Squares