Observation-Dependent Posterior Inflation for the Ensemble Kalman Filter

Ensemble-based Kalman filter (EBKF) algorithms are known to produce posterior ensembles whose variance is incorrect for a variety of reasons (e.g., nonlinearity and sampling error). It is shown here that the presence of sampling error implies that the true posterior error variance is a function of the latest observation, as opposed to the standard EBKF, whose posterior variance is independent of observations. In addition, it is shown that the traditional ensemble validation tool known as the "binned spread-skill'' diagram does not correctly identify this issue in the ensemble generation step of the EBKF, leading to an overly optimistic impression of the relationship between posterior variance and squared error. An updated ensemble validation tool is described that reveals the incorrect relationship between mean squared error (MSE) and ensemble variance, and gives an unbiased evaluation of the posterior variances from EBKF algorithms. Last, a new inflation method is derived that accounts for sampling error and correctly yields posterior error variances that depend on the latest observation. The new method has very little computational overhead, does not require access to the observations, and is simple to use in any serial or global EBKF.