Ensemble Kalman Filter

Abstract The ensemble Kalman filter (EnKF) is a computational technique that leads to efficient approximate inference in state‐space models. Traditionally, this has been applied to data assimilation problems in which high‐dimensional spatial fields are observed sequentially through time. In essence, the EnKF approximates the traditional Kalman filter by using a sample (i.e., ensemble) of draws to represent the state process distribution. As new data become available, these ensemble members are updated by shifting toward the data. This shifting procedure is one of the key differences between the EnKF and reweighting‐based methods (e.g., particle filters), and allows the EnKF to avoid the degeneracy problem that is endemic to those methods, even for very high state dimensions. Although the EnKF implicitly assumes linear Gaussian state‐space models, it allows for nonlinear state propagation and is fairly robust to deviations from linearity and Gaussianity.