Abstract:In this paper, two randomized block Kaczmarz methods to compute inner inverses of any rectangular matrix A are presented. These are iterative methods without matrix multiplications and their convergence is proved. The numerical results show that the proposed methods are more efficient than iterative methods involving matrix multiplications for the high-dimensional matrix.
What problem does this paper attempt to address?
This paper aims to solve the problem of calculating the inner inverse within a rectangular matrix. Specifically, the author proposes two new algorithms based on the randomized block Kaczmarz method for calculating the inner inverse of an arbitrary rectangular matrix \(A\). These algorithms are iterative methods that do not require matrix multiplication operations, and their convergence has been proven. Numerical experiments show that for high - dimensional matrices, the proposed methods are more efficient than iterative methods involving matrix multiplication.
### Background and Motivation
The inner inverse plays an important role in solving linear equations, least - squares problems, and describing the properties of linear transformations, and is widely used in engineering fields such as robotics, big - data analysis, network learning, and sensor fusion. Traditional methods for calculating the inner inverse include the Moore - Penrose generalized inverse, singular value decomposition (SVD), and block - matrix methods, etc. However, these methods usually require a large amount of computing time, especially when dealing with large - scale matrices.
### Research Problem
The main research problem in this paper is to develop efficient algorithms for calculating the inner inverse of an arbitrary rectangular matrix, especially for large - scale matrices. Existing methods usually involve matrix multiplication and have a high computational cost. Therefore, the author proposes two new randomized block Kaczmarz methods to reduce the computational complexity and improve the computational efficiency.
### Method Overview
1. **Projection Randomized Block Kaczmarz Method (MII - PRBK)**:
- In each iteration, this method selects a row vector \(A_{i,:}\) and updates the current estimate matrix \(X^{(k)}\) by projecting onto the corresponding hyperplane.
- The update formula is:
\[
X^{(k + 1)}=X^{(k)}+\frac{A_{i,:}^*(A_{i,:}-A_{i,:}X^{(k)}A)A^{\dagger}}{\|A_{i,:}\|_2^2}
\]
- Where \(i\) is selected with probability \(p_i = \frac{\|A_{i,:}\|_2^2}{\|A\|_F^2}\).
2. **Randomized Averaged Block Kaczmarz Method (MII - RABK)**:
- In each iteration, this method selects a row vector \(A_{i,:}\) and updates the current estimate matrix \(X^{(k)}\) by projecting onto a convex combination of multiple hyperplanes.
- The update formula is:
\[
X^{(k + 1)}=X^{(k)}+\alpha\frac{A_{i,:}^*(A_{i,:}-A_{i,:}X^{(k)}A)A^*}{\|A_{i,:}\|_2^2}
\]
- Where \(\alpha\) is a step - size parameter, satisfying \(0 < \alpha < \frac{2}{\|A\|_2^2}\), and \(i\) is selected with probability \(p_i=\frac{\|A_{i,:}\|_2^2}{\|A\|_F^2}\).
### Numerical Experiments
The author verifies the effectiveness of the proposed methods through numerical experiments. The experimental results show that the MII - PRBK method performs best in terms of running time, while the MII - RABK method performs better in terms of the number of iterations. In addition, these two methods have significant advantages when dealing with large - scale matrices, especially when the computational cost of matrix multiplication is high.
### Conclusion
The randomized block Kaczmarz methods proposed in this paper provide an efficient method for calculating the inner inverse of an arbitrary rectangular matrix. These methods are particularly suitable for large - scale matrices because they only need to use one row of the matrix in each iteration and do not require matrix multiplication. Future work will focus on further optimizing the algorithm performance, extending to the calculation of other generalized inverses, and designing a randomized block Kaczmarz algorithm with an adaptive step - size.