A complementary note to ‘A lag‐1 smoother approach to system‐error estimation’: the intrinsic limitations of residual diagnostics

Abstract

Recently, this author studied an approach to the estimation of system error based on combining observation residuals derived from a sequential filter and fixed lag‐1 smoother. While extending the methodology to a variational formulation, experimenting with simple models and making sure consistency was found between the sequential and variational formulations, the limitations of the residual‐based approach came clearly to the surface. This note uses the sequential assimilation application to simple nonlinear dynamics to highlight the issue. Only when some of the underlying error statistics are assumed known is it possible to estimate the unknown component. In general, when considerable uncertainties exist in the underlying statistics as a whole, attempts to obtain separate estimates of the various error covariances are bound to lead to misrepresentation of errors. The conclusions are particularly relevant to present‐day attempts to estimate observation‐error correlations from observation residual statistics. A brief illustration of the issue is also provided by comparing estimates of error correlations derived from a quasi‐operational assimilation system and a corresponding Observing System Simulation Experiments framework.