The Convergence of Analyses Produced by Overlapping Data Assimilation Streams

When conducting a reanalysis for a long period, it is common to do so using multiple, parallel computational schemes. A period of overlap is included to foster continuity between successive streams. At the GMAO, the overlap period is one year. By the end of that period, the time -mean analysis for corresponding overlapped months are almost identical, as desired. The variances of differences at corresponding analysis times within each month, however, do not converge to zero. Instead their monthly variances converge to values that are a significant fraction of the estimated variances of analysis errors. This occurs although the overlapping streams use the same observations, assimilation model, and assimilation algorithm, differing only in the background information applied at the beginning of each overlapping period.