Singular vectors hold the promise to be able to identify those regions of the atmosphere where more accurate initial conditions are needed the most to reduce the risk of large forecast errors at a given range of, say up to 2 days. In these so-called sensitive regions supplementary observations are likely to be more useful for the forecast than observations outside of the sensitive regions.

Whether the singular vectors actually point to the most sensitive regions depends on the usefulness of the involved approximations. First, the tangent-linear model needs to capture the essence of the dynamics of short-range forecast errors. Second, the initial time norm in the singular vector computation must be a good approximation of the true analysis error covariances. However, our knowlegde of the statistics of the analysis error is very limited. The presentation will address the sensitivity of the sensitive regions diagnosed from singular vectors to the choice of the initial time norm used in their computation. Singular vectors computed with the total energy metric and singular vectors computed with the more elaborate Hessian metric will be compared. The role of observations in the latter metric for the structure and geographical location of the singular vectors will be discussed.