Throw out the model covariance matrix

I won't say the covariance of the model can always be ignored, but since it must be a prior given, I would like to find some way around it. Thus I suggest that every data space be augmented till it has the dimensionality and completeness required to determine a solution. It this cannot be done fully, it should still be done to the extent feasible. For example we might omit evanescent waves, but not waves blocked by an aperture. Reinaldo and I are doing just that in tomography.

The covariance matrix of the residual in data space (missing and observed) seems a reasonable thing to estimate--unlike the covariance matrix of the model. I think this matrix should not be thought of as a covariance matrix of the solution, but as an interpolation function for plotting the solution.