We conjecture that this is because the
bred vectors

collapse as they “age” into a subspace
which is too small: the
stochastic perturbations due to nonlinearity and convection, etc., are not enough to
avoid this collapse. As a result,
some directions of growing errors are missed.
In order to avoid this problem, we
added random perturbations
to the bred vectors at the beginning of the 12-hr integrations. This reduced by 40%
the errors with the global
bred vectors (Corazza et al 2002).

With the Ott et al (2002) local bred
vector Kalman Filtering approach
the improvements are even more dramatic.