Detecting regime transitions in time series using dynamic mode decomposition (Q781788)

The authors apply the method of dynamic mode decomposition (DMD) and the Koopman operator technique to provide a computationally cheap and easily implementable method to detect transient dynamics and regime changes in time series. For chaotic dynamical systems evolving on a lower dimensional attractor, the dynamics with a finite number of Koopman modes is described; the number of modes needed is proportional to the attractor dimension. The authors focus only on reconstructing the observations and not on forecasting. They define the reconstruction error \[\mathcal{E}_{\text{DMD}}(t_{k}, r)=\frac{1}{m}\sum_{l=0}^{m-1}||x(t_{k+l})-\Phi \; \text{exp}(\Omega l \Delta t) b_{k}-\text{c.c.}||,\] with \(b_{k}=\Phi^{\dagger}x_{k}\) and c.c. being the complex conjugate. The reconstruction error of a dynamic mode decomposition is used to monitor the inability of a time series to resolve the fast relaxation towards the attractor as well as the effective dimension of the dynamics. The authors elucidate their method by detecting transient dynamics in the Kuramoto-Sivashinsky equation. They consider artificially generated time series obtained from a numerical simulation of the Kuramoto-Sivashinsky equation \(u_{t}+u u_{x}+\alpha u_{xx} + u_{xxxx}=0\). For fixed system length \(L\) the Kuramoto-Sivashinsky equation becomes chaotic upon increasing the driving \(\alpha\). The authors compute the reconstruction error \(\mathcal{E}_{\text{DMD}}(t, r)\) by means of a graphical analysis and show that during the transient dynamics the reconstruction error is large but drops off significantly when the dynamics has settled down on the limit cycle near \(t=16,000\). Afterwards a more rigorous discussion is provided. They also apply their method to study rigorously the regime change of the Northern Hemisphere atmospheric circulation dynamics and the North Atlantic Oscillation (NAO) in reanalysis data; their diagnostics detects the transition from a predominantly negative NAO to a predominantly positive NAO around 1970. Further, the regime change of the Southern Hemisphere atmospheric circulation dynamics is analyzed rigorously using data. In the conclusion, the authors mention that their method is cost-effective and is able to detect regime changes and transitory dynamics in the above both cases.