Core Concepts
The authors propose two new Kaczmarz methods, the Maximum Residual Block Kaczmarz (MRBK) method and the Maximum Residual Average Block Kaczmarz (MRABK) method, to efficiently solve large consistent linear systems. The MRBK method selects the block with the largest residual to eliminate at each iteration, while the MRABK method avoids the computation of pseudo-inverse by projecting the current iterate onto each row of the selected block and averaging them with different extrapolation steps.
Abstract
The paper introduces two new Kaczmarz methods for solving large consistent linear systems:
Maximum Residual Block Kaczmarz (MRBK) Method:
Partitions the rows of the coefficient matrix A into blocks {AV1, AV2, ..., AVt}.
In each iteration, selects the block Vik with the largest residual norm, i.e., ik = arg max1≤i≤t ∥bVi - AVixk∥2, and updates the solution by projecting onto the hyperplane corresponding to AVik.
Proves the convergence of the MRBK method and provides an upper bound on its convergence rate.
Maximum Residual Average Block Kaczmarz (MRABK) Method:
Also partitions the rows of A into blocks, but avoids the computation of pseudo-inverse by projecting the current iterate onto each row of the selected block and averaging them with different extrapolation steps.
Proves the convergence of the MRABK method and provides an upper bound on its convergence rate.
Shows that the MRABK method has a faster convergence rate than the MRBK method due to the lower computational cost.
The authors compare the proposed methods with other Kaczmarz variants, such as the Greedy Randomized Kaczmarz (GRK), Maximal Residual Kaczmarz (MRK), Randomized Block Kaczmarz (RBK), Greedy Block Kaczmarz (GBK), and Greedy Randomized Block Kaczmarz (GRBK) methods, through numerical experiments. The results demonstrate the superiority of the MRBK and MRABK methods in terms of both iteration steps and computational time.
Stats
The authors provide the following key figures and metrics in the paper:
The convergence factor of the MRBK method is ρMRBK = 1 - σ²min(A) / (β(t-1)).
The convergence factor of the MRABK method is ρMRABK = 1 - (2ω - ω²)σ²min(A) / (βt), where ω is the extrapolation step size.
The speed-up value of the MRBK method compared to the MRK method (SU1) ranges from 7.35 to 100.99.
The speed-up value of the MRBK method compared to the GRBK method (SU2) ranges from 1.19 to 1.77.
The speed-up value of the MRABK method compared to the MRBK method (SU3) ranges from 2.10 to 4.11.
Quotes
"The MRBK method accelerates the MRK method naturally by utilizing row block iterations instead of single row iterations."
"The MRABK method has the shortest computing time among all the above methods, and its speed-up value relative to the MRBK method (SU3) is at least 2.42 and up to 2.75."
"Compared to the MRK method, the MRBK method exhibits a minimum speed-up value of 1.04 and a maximum speed-up value of 28.04."