Part of the speed problem demonstrated by the recent thread on the ML, might be from the fact that the KF algorithm isn't switching to the steady state calculations. It checks to see that the F matrix is 1. to switch. In this case, it quickly converges to close to 1, but it looks like it's approaching it asymptotically. I wonder if we can relax this and maintain good precision results.
We can get a decent speedup in some cases for (F_mat - 1) < 1e-X, but the tolerance isn't general and the precision - speed trade-off isn't clear. Many tests fail when I lessen it. We could try also a tolerance for (F_mat_last - F_mat), but this takes more than a one line change, and I don't have time right now.
(F_mat - 1) < 1e-X
Closing. No longer an issue with recent Cython speedups.