Since we use the Morss and Emanuel (2001) simulation system, we
know the “truth” and
analysis/forecast errors.
We compare the three methods showing
a)the total analysis error for a simulated year over the whole
domain, and
b)the background error superimposed with the analysis
increments.
The 3D-Var parameters have been carefully tuned to get the best
results.
Note that, because it is based on a statistically optimized
background error
covariance, 3D-Var cannot capture the “errors of the day”.
The Local Bred Vector Kalman Filtering (Ott et al, 2002) do
capture the errors of the day. However, the bred vectors may “inbreed” to
much and
miss some growing directions.
In order to avoid this, we add random perturbations (as if they
were observational
errors) to the bred vectors.