Abstract:We present improved approximation bounds for the Moore-Penrose inverses of banded matrices, where the bandedness is induced by an arbitrary metric space. We show that the pseudoinverse of a banded matrix can be approximated by another banded matrix, and the error of approximation is exponentially small in the bandwidth of the approximation. An intuitive corollary can be obtained: the off-diagonal blocks of the pseudoinverse decay exponentially with the distance between the node sets associated with row and column indices, on the given metric space. Our bounds are expressed in terms of the bound of singular values of the system. For saddle point systems, commonly encountered in optimization, we provide the bounds of singular values associated with standard regularity conditions (linear independence constraint qualifications and second-order sufficiency conditions). Our bounds improve previously reported ones (Demko,1984; Bickel, 2012; Shin, 2022). Remarkably, our bounds allow us to establish a perturbation bound for continuous-domain optimal control problems by analyzing the asymptotic limit of their finite difference discretization, which has been challenging with previously reported bounds.
What problem does this paper attempt to address?
The core problem that this paper attempts to solve is to improve the approximate bounds of the Moore - Penrose pseudoinverse of banded matrices and apply it to the continuous - time linear - quadratic control problem. Specifically, the authors propose new methods to prove that the pseudoinverse of a banded matrix can be approximated by another banded matrix with exponential - level precision, and the error of this approximation decays exponentially as the bandwidth increases.
### Main contributions of the paper include:
1. **Generalized banded structure**: Generalize the concept of banded matrices to arbitrary metric spaces, so that the main approximation results have more extensive applications.
2. **Improved approximate bounds**: Provide tighter approximate bounds for the Moore - Penrose pseudoinverse of banded matrices, improving the results in the existing literature. These new bounds are applicable not only to positive definite matrices, but also to indefinite matrices and rectangular matrices.
3. **Application to optimal control problems**: Utilize the above - improved bounds to analyze the perturbation bounds of the continuous - time linear - quadratic optimal control problem, especially to solve the problem of the error bounds diverging when the grid size tends to zero during the discretization process.
### Specific content analysis:
- **Definition and properties of banded matrices**: The paper defines banded matrices indexed by finite metric spaces and proves that these matrices maintain their banded characteristics under transposition, addition, and multiplication operations.
- **Main theorem**: For a banded matrix \(A\), its Moore - Penrose pseudoinverse \(A^{+}\) can be approximated by a banded matrix \(\tilde{A}\) with exponential - level precision, and the error bound depends on the singular value range of the original matrix and the upper bound of the bandwidth.
- **Polynomial approximation strategy**: By using techniques such as Chebyshev polynomials, construct an odd - polynomial approximation of \(1/x\), thereby deriving tighter error bounds.
- **Singular value bounds of saddle - point systems**: Derive the upper and lower bounds of the singular values for saddle - point systems common in optimization problems, which helps to further analyze the convergence and stability of the discretized system.
### Practical applications:
These theoretical results can be applied to multiple fields, such as accelerating the convergence speed of iterative solvers as pre - conditioners, analyzing the convergence rate of decomposition algorithms for large - scale optimization problems, and problems such as distributed control and stochastic predictive control.
In general, through in - depth research on the pseudoinverse of banded matrices, this paper provides a new tool to handle the discretization problem of infinite - dimensional linear systems, especially having important application value in the field of optimal control.