Abstract
Online data assimilation techniques such as ensemble Kalman filters and particle filters lose accuracy dramatically when presented with an unlikely observation. Such an observation may be caused by an unusually large measurement error or reflect a rare fluctuation in the dynamics of the system. Over a long enough span of time it becomes likely that one or several of these events will occur. Often they are signatures of the most interesting features of the underlying system and their prediction becomes the primary focus of the data assimilation procedure. The Kuroshio or Black Current that runs along the eastern coast of Japan is an example of such a system. It undergoes infrequent but dramatic changes of state between a small meander during which the current remains close to the coast of Japan, and a large meander during which it bulges away from the coast. Because of the important role that the Kuroshio plays in distributing heat and salinity in the surrounding region, prediction of these transitions is of acute interest. Here the authors focus on a regime in which both the stochastic forcing on the system and the observational noise are small. In this setting large deviation theory can be used to understand why standard filtering methods fail and guide the design of the more effective data assimilation techniques. Motivated by this analysis the authors propose several data assimilation strategies capable of efficiently handling rare events such as the transitions of the Kuroshio. These techniques are tested on a model of the Kuroshio and are shown to perform much better than standard filtering methods.
