Abstract:In this work we develop a low-rank algorithm for the computation of low-rank approximations to the large-scale Lyapunov operator $\varphi$-functions. Such computations constitute the important ingredient to the implementation of matrix-valued exponential integrators for large-scale stiff matrix differential equations where the (approximate) solutions are of low rank. We evaluate the approximate solutions of $LDL^T$-type based on a scaling and recursive procedure. The key parameters of the method are determined using a quasi-backward error bound combined with the computational cost. Numerical results demonstrate that our method can be used as a basis of matrix-valued exponential integrators for solving large-scale differential Lyapunov equations
and differential Riccati equations.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: when solving large - scale matrix differential equations (MDEs), how to efficiently calculate the action of the Lyapunov operator ϕ - function. Specifically, the paper focuses on how to approximately solve these ϕ - functions by low - rank algorithms when using matrix - valued exponential integrators. This is especially important when dealing with large - scale stiff matrix differential equations with low - rank characteristics, such as differential Lyapunov equations and differential Riccati equations.
### Specific Problem Description
1. **Background and Challenges**:
- Large - scale matrix differential equations (MDEs) play an important role in many fields, such as optimal control, model order reduction, semi - discretization of two - dimensional parabolic partial differential equations, etc.
- In particular, differential Lyapunov equations (DLEs) and differential Riccati equations (DREs) usually have non - homogeneous terms and initial values of low - rank.
- Exponential integrators are widely used to solve stiff problems because of their exact treatment of the linear part and explicit treatment of the nonlinear part, but the core computational problem is to solve the action of the operator ϕ - function on certain matrices.
2. **Limitations of Existing Methods**:
- Many time - integration methods (such as BDF, Rosenbrock methods, splitting methods, and Krylov - projection - based methods) can take advantage of the low - rank structure, but there are still deficiencies in efficiency and accuracy when numerically solving large - scale DLEs and DREs.
- Traditional numerical methods (such as standard numerical quadrature formulas) may require a large amount of computing time to obtain high - precision results, and low - precision approximations may lead to numerical instability.
3. **New Method Proposed in the Paper**:
- A low - rank algorithm based on LDLT decomposition is proposed for efficiently calculating the Lyapunov operator ϕ - function.
- This method combines scaling and recursive processes and is implemented by matrix - vector multiplication, thereby reducing storage and computational complexity.
- By selecting optimal parameters (such as the polynomial order m and the scaling parameter s), the accuracy and efficiency of the calculation are ensured.
### Goal
The goal of the paper is to develop a novel and efficient low - rank algorithm that can quickly and accurately calculate the Lyapunov operator ϕ - function, thereby providing a solid foundation for exponential integrators of large - scale matrix differential equations. This method not only improves computational efficiency but also maintains high numerical stability and is suitable for solving large - scale stiff problems.