Abstract:We study ill-conditioned positive definite matrices that are disturbed by the sum of $m$ rank-one matrices of a specific form. We provide estimates for the eigenvalues and eigenvectors. When the condition number of the initial matrix tends to infinity, we bound the values of the coordinates of the eigenvectors of the perturbed matrix. Equivalently, in the coordinate system where the initial matrix is diagonal, we bound the rate of convergence of coordinates that tend to zero.
What problem does this paper attempt to address?
### What problems does this paper attempt to solve?
This paper mainly studies the **asymptotic estimation problem of eigenvalues and eigenvectors of a positive definite matrix in ill - conditioned state after being perturbed by a rank - one matrix in a specific form**. Specifically, the author focuses on the change of eigenvector coordinates of the perturbed matrix \(A(m)\) when the condition number of the initial matrix \(B\) tends to infinity.
#### Main problem description
1. **Research object**:
- Given a \(d\times d\) symmetric matrix \(B\) whose eigenvalues and eigenvectors are known.
- Consider the perturbed matrix \(A(m)=B + \sqrt{B}\sum_{i = 1}^m [v(i)][v(i)]^T\sqrt{B}\), where \(v(i)\in\mathbb{R}^d\) is a set of vectors.
2. **Objective**:
- Provide estimates of eigenvalues and eigenvectors of the perturbed matrix \(A(m)\).
- When the condition number of the initial matrix \(B\) tends to infinity, give the convergence rate of eigenvector coordinates.
3. **Specific problems**:
- The paper pays special attention to how the coordinates of the perturbed eigenvector \(e(m)_i\) change when \(B\) is highly ill - conditioned (i.e., the condition number is large).
- By introducing the relative perturbation bound, the author hopes to prove that the convergence rate of these eigenvector coordinates is related to the eigenvalues of the initial matrix \(B\), rather than directly depending on the norm of the perturbation.
#### Key formulas
- **Convergence rate of eigenvector coordinates**:
\[
\left|\left[e(m)_i\right]_j\right|\leq C_m\sqrt{\frac{\min\{\lambda_i,\lambda_j\}}{\max\{\lambda_i,\lambda_j\}}}
\]
where \(C_m\) is a constant depending on \(m\) and \(d\).
- **Bounds of eigenvalues**:
\[
\lambda_i\leq\nu_i\leq\lambda_i\left(1 + md\cdot\max_{k = 1,\dots,m}\|v(k)\|_\infty^2\right)
\]
#### Application background
- Such problems have important applications in fields such as optimization algorithms and Markov chain Monte Carlo methods. For example, in the CMA - ES (Covariance Matrix Adaptation Evolution Strategy) algorithm, the update of the covariance matrix involves a similar perturbation problem, so understanding the change of eigenvectors is crucial for analyzing the stability and convergence of the algorithm.
In short, this paper aims to provide a powerful tool for theoretical analysis in related fields by accurately estimating the eigenvectors of ill - conditioned matrices.