Generalized Wedderburn Rank Reduction

Oskar Kędzierski
2024-06-06
Abstract:We generalize the Wedderburn rank reduction formula by replacing the inverse with the Moore--Penrose pseudoinverse. In particular, this allows one to remove the non--singularity of a certain matrix from assumptions. The results implies in a straightforward way Nystroem, CUR decompositions, meta-factorization, and a result of Ameli, Shadden. We investigate which properties of the matrix are inherited by the generalized Wedderburn reduction. Reductions leading to the best low-rank approximation are explicitly described in terms of singular vectors. We give a self--contained calculation of the range and the nullspace of the projection $A(BA)^+B$ and prove that any projection can be expressed in this way.
Numerical Analysis
What problem does this paper attempt to address?
The core problem that this paper attempts to solve is to generalize the classical Wedderburn rank - reduction formula by introducing the Moore - Penrose pseudoinverse, thereby relaxing the requirement for matrix non - singularity. Specifically, the author hopes: 1. **Remove the assumption of matrix non - singularity**: The traditional Wedderburn rank - reduction formula depends on the invertibility of the matrix. In this paper, by using the Moore - Penrose pseudoinverse, the formula can be applied to a wider range of matrix types. 2. **Derive a generalized rank - reduction formula**: By replacing the inverse matrix with the pseudoinverse matrix, the author derives a more general rank - reduction formula, which can handle the cases of non - square matrices and singular matrices. 3. **Explore the properties of generalized Wedderburn rank - reduction**: Study the property changes of matrices after generalized Wedderburn rank - reduction, especially which properties will be preserved or changed. 4. **Apply to other matrix decomposition methods**: Prove that this generalized formula can directly lead to other important matrix decomposition methods such as Nystrom decomposition and CUR decomposition, and give an explicit description of the best low - rank approximation. ### Main contributions - **Generalize the Wedderburn rank - reduction formula**: By introducing the Moore - Penrose pseudoinverse, the requirement for matrix non - singularity is removed. - **Prove the effectiveness of related matrix decompositions**: Such as Nystrom decomposition, CUR decomposition, etc. These decomposition methods can be directly derived from the generalized Wedderburn rank - reduction formula. - **Provide a new proof of the projection formula**: Re - prove the projection formula based on basic principles without relying on the Zlobec formula. - **Give the conditions for the optimal low - rank approximation**: Use the singular value decomposition (SVD) to describe the selection conditions of matrices \(X\) and \(Y\) that lead to the best low - rank approximation. ### Formula summary The generalized Wedderburn rank - reduction formula is as follows: \[ \text{rank}(A-(AX)(Y^{*}AX)^{+}(Y^{*}A))=\text{rank}(A)-k \] where \(A\in K^{m\times n}\), \(X\in K^{n\times p}\), \(Y\in K^{m\times q}\), and \(\text{rank}(Y^{*}AX) = k\). The key to this formula is to replace the inverse matrix \(M^{-1}\) in the traditional formula with the Moore - Penrose pseudoinverse \((Y^{*}AX)^{+}\), so that the formula can be applied to a wider range of matrix types. ### Practical applications This formula and its generalized form can be applied to various numerical linear algebra problems. Especially when dealing with large - scale data sets, it can effectively perform matrix dimension reduction and feature extraction, thereby improving computational efficiency and model performance. Through these improvements and generalizations, this paper provides new tools and methods for the fields of matrix analysis and numerical computation, which has important theoretical and practical significance.