Overview
This research focuses on the quality of approximate balanced truncation methods, specifically by establishing global error bounds for reduced systems derived using approximate Gramians. Model reduction techniques are crucial for managing the computational complexity associated with numerical simulations of large-scale dynamic systems, particularly those encountered in the study of complex physical phenomena. Among these techniques, balanced truncation-based methods are recognized for their theoretical soundness, primarily due to the existence of a comparatively tight global error bound on the approximation error between the original and the reduced system.
Research Context
Model reduction is a fundamental approach for simplifying the numerical simulation of large-scale dynamic systems. For linear time-invariant dynamic systems, two predominant types of model reduction methods are employed: Krylov subspace-based methods and balanced truncation-based methods. The latter class, balanced truncation, is distinguished by its theoretical underpinnings, which include a global error bound for the approximation error. This error bound quantifies the discrepancy between the full-order system and its reduced counterpart.
A critical condition for the establishment of this existing error bound is the availability of exact controllability and observability Gramians. In practical computational scenarios, particularly for large-scale systems, exact Gramians are not numerically obtainable. Instead, these Gramians must be calculated numerically, and for systems of substantial size, the viable approach involves computing low-rank approximations of these Gramians. Subsequently, an approximate balanced truncation procedure is performed using these approximations.
A significant gap in the existing theoretical framework is that the established global error bound, which relies on exact Gramians, does not rigorously apply to reduced systems generated via approximate Gramians. This paper's central objective is to address this discrepancy.
Approach
The research undertakes the task of establishing global error bounds that are applicable to reduced systems resulting from approximate balanced truncation. The methodology is designed to bridge the theoretical gap that exists when Gramians are approximated rather than exactly known. This involves developing new theoretical formulations that account for the inaccuracies introduced by using low-rank approximations of the controllability and observability Gramians.
Findings
The primary finding of this research is the establishment of global error bounds specifically tailored for reduced systems derived through approximate balanced truncation. These new bounds address the limitation of previous theoretical frameworks, which were predicated on the availability of exact Gramians. By providing these new error bounds, the research ensures that the quality of approximation can be rigorously assessed even when low-rank approximations of Gramians are utilized in the model reduction process.
Specifically, the study indicates that:
- The existing global error bound for balanced truncation is dependent on the availability of exact controllability and observability Gramians.
- For large-scale systems, exact Gramians are not numerically available; instead, low-rank approximations are computed.
- The existing global error bound is not applicable to reduced systems obtained via approximate Gramians.
- New global error bounds have been established for reduced systems generated through approximate balanced truncation.
Why This Matters
The development of new global error bounds for approximate balanced truncation is significant because it extends the theoretical rigor of model reduction techniques to computationally feasible scenarios. It allows for a more accurate assessment of approximation quality in practical applications where exact Gramians are unobtainable and low-rank approximations are necessary for large-scale dynamic systems.