The structure and realism of sensitivity perturbations and their interpretation as ‘Key Analysis Errors’



Adjoint sensitivity-based perturbations (e.g. ‘Key Analysis Errors’) minimizing the two-day forecast error have been calculated for December 2001 and January 2002 using three different initial-time norms. The goal has been to investigate if the perturbations can justifiably be interpreted as analysis error. We perform a systematic comparison against available observations, at initial time and as the perturbations evolve during the first 24 hours of forecasts. Ten-day forecasts have been run to verify that the medium-range forecasts from the perturbed analyses are better than the control forecasts. It is shown that the structure of the perturbations depends strongly on the initial-time norm, in experiments based on the energy norm, the background-error covariance norm, and the Hessian norm. For all three norms it is found that forecasts starting from the perturbed analyses are further away from the observations than forecasts from control analyses during the first approximately 12 hours of forecasts, a signal that is increased when the number of iterations is increased. From 12 hours onwards the perturbed forecasts are closer to observations than the control forecasts. We conclude that the sensitivity perturbations cannot justifiably be interpreted as analysis error as far as their detailed structure is concerned. This result has implications in applications that rely on sensitivity-based or singular vector-based approaches for detailed characterization of analysis error, e.g. some observation-targeting applications and the reduced-rank Kalman filter. Copyright © 2005 Royal Meteorological Society.