1: \begin{abstract}

2: The numerical properties of algorithms for finding the intersection of sets depend

3: to some extent on the regularity of the sets, but even more importantly 

4: on the regularity of the intersection. The alternating projection 

5: algorithm of von Neumann has been shown to converge locally at a linear rate

6: dependent on the regularity modulus of the intersection.  In many applications, 

7: however, the sets in question come from inexact measurements that are 

8: matched to idealized models.  It is unlikely that any such problems in 

9: applications will enjoy metrically regular intersection, let alone set intersection.

10: We explore a regularization strategy that generates an intersection with 

11: the desired regularity properties.  The regularization, however, 

12: can lead to a significant increase in computational complexity.  In a further refinement, 

13: we investigate and prove linear convergence of an approximate alternating projection

14: algorithm.  The analysis provides a regularization strategy that fits naturally with many 

15: ill-posed inverse problems, and a mathematically  

16: sound stopping criterion for extrapolated, approximate algorithms.  The theory is  demonstrated on 

17: the phase retrieval problem with experimental data.  The conventional early termination applied 

18: in practice to unregularized, consistent problems in diffraction imaging can be justified fully in the 

19: framework of this analysis providing, for the first time, proof of convergence of alternating approximate 

20: projections for finite dimensional, consistent phase retrieval problems.  

21: \end{abstract}

22: