[cv,wx] = best_dir_cov(x,a,sx,wd)

x    D x P     :
a    P x W     : Same as in best_dir, but sx is compulsory.
sx   P x 1     :

wd (W+D) x 1   : ML estimate of [w;d]

cv (W+D)x(W+D) : Covariance of the ML estimator at [w;d]

wx (W+D)x(P*D) : derivatives of ML estimate wrt to observations