Now we face the problem of approximating the Cq coefficients. Here using a Taylor expansion does not help much, so we decided to freeze the coefficients (make them constant)
and look at the predicted rates for the worst possible scenario, i.e., when the coefficients are very anisotropic. We also note that, as remarked before, the present analysis is valid when the algorithm has almost reached convergence. At this stage, the change in the values of the coefficients is almost null (they remain almost constant), and the edges in the image have reached their sharpest form; i.e., the coefficients are very anisotropic in these regions. These two observations back up our decision to use constant coefficients in the LFA analysis.