Overview
This paper introduces an algebraic framework for subspace pruning, aimed at systematically improving the accuracy of Koopman operator approximations in finite-dimensional spaces. The proposed method focuses on refining the invariance error, which is quantified by principal angles between a candidate subspace and its image under the Koopman operator. The framework establishes a geometric equivalence between consistency-based methods and principal-vector pruning, leading to a hybrid strategy.
To address issues of numerical stability and scalability, the hybrid strategy balances between multiple and single principal vector pruning. The research also details an efficient numerical update scheme based on rank-one modifications designed to reduce computational complexity in tracking principal angles. The subspace derived from these pruning algorithms is then utilized to construct a lifted linear model for state prediction, which considers the trade-offs between enhanced invariance and minimized state reconstruction error.
Research Context
The accuracy of Koopman operator approximations, particularly over finite-dimensional spaces, is fundamentally dependent on their invariance properties. These properties can be rigorously quantified through the principal angles between a selected subspace and its transformation under the Koopman operator. Improving these approximations often involves methods to refine the invariance error. Existing approaches may encounter limitations related to numerical stability or scalability, prompting the need for more robust and efficient pruning strategies.
Approach
The research first establishes a unified algebraic framework for subspace pruning. Within this framework, a geometric equivalence between consistency-based methods and principal-vector pruning is identified. Building on this insight, a hybrid strategy is introduced, which incorporates both multiple and single principal vector pruning techniques. This hybrid approach is designed to enhance numerical stability and scalability.
The methodology includes the derivation of error bounds specific to the retention of both approximate and external eigenfunctions. This derivation highlights that the multi-vector approach can mitigate numerical drift, which is an inherent issue in sequential pruning processes. For scalability, an efficient numerical update scheme was developed. This scheme leverages rank-one modifications, substantially reducing the computational complexity associated with tracking principal angles.
Finally, the subspace resulting from the application of these pruning algorithms is used to construct a lifted linear model. This model is engineered for state prediction and aims to balance the trade-offs between achieving better invariance and minimizing the error in state reconstruction. Simulations were conducted to evaluate the effectiveness of the proposed approach.
Findings
- The proposed algebraic framework provides a unified method for subspace pruning to refine invariance error in Koopman operator approximations.
- A geometric equivalence exists between consistency-based methods and principal-vector pruning.
- A hybrid strategy, balancing multiple and single principal vector pruning, enhances numerical stability and scalability.
- Error bounds were derived for the retention of approximate and external eigenfunctions.
- The multi-vector approach mitigates numerical drift observed in sequential pruning.
- An efficient numerical update scheme based on rank-one modifications reduces the computational complexity of tracking principal angles by an order of magnitude.
- The subspace obtained from the pruning algorithms can be used to construct a lifted linear model for state prediction, balancing invariance improvement with state reconstruction error minimization.
- Simulations demonstrated the effectiveness of the approach.
Why This Matters
The development of methods that systematically refine the invariance error in Koopman operator approximations can lead to more accurate and reliable representations of dynamical systems. By improving numerical stability and scalability, this work addresses practical challenges in applying Koopman-based techniques to larger and more complex systems. The ability to construct a lifted linear model that effectively balances invariance and state reconstruction accuracy could facilitate more robust state prediction in various applications.