Conclusions

The application of the RGCV criterion as a stopping rule for non-linear iterative algorithms appears to have both analytic and experimental justification. The comparison to the -square criterion shows that the RGCV criterion more closely approximates the minimum mean-square error stopping point, and does so without any prior knowledge of the noise present in the system.

Unfortunately, there is one major drawback to the RGCV criterion, namely computation. The normal R-L iteration involves two convolutions. The computation of uses previously computed convolutions and adds an additional three convolutions. Thus, by using the RGCV criterion, one increases the computation time by approximately a factor of 2.5. The advantage of the RGCV that may counter this disadvantage is the fact that the RGCV stopping rule is autonomous.