Bred
vectors are constructed as finite amplitude, finite time approximations of Lyapunov vectors. The
fact that we use them
locally, rather than globally, greatly reduces the number of independent bred vectors
needed (Kalnay et al, 2002).

Since the bred vectors represent the
instabilities of the underlying
flow, and these are finite time, finite space, it is reasonable to assume that there may be
instabilities that the
bred vectors “miss” if they are confined to a too small subspace.

Therefore, “sprinkling” bred vectors
with small random perturbations
not only mimics observational errors but keeps
them “young” and gives them a better chance to represent new instabilities that may
appear in the analysis errors.