In
Corazza et al, 2002, we showed that the use of the standard 3D-Var error covariance matrix
augmented with global
bred vectors reduced by about 20% the analysis errors.

However, when we added random errors
(as if they were observation
errors) to the bred vector at the beginning of each integration, the analysis errors
were much smaller, more
than 40% smaller than the regular 3D-Var.

We conjecture that this is because, in
the standard breeding
approach (Toth and Kalnay, 1993, 1997), the bred vectors tend to collapse into a
subspace which is too small.