1: \begin{abstract}

2: We propose two new methods based on Kaczmarz's method that produce a regularized solution to noisy tomography problems. These methods exhibit semi-convergence when applied to inverse problems, and the aim is therefore to stop near the semi-convergence point.  Our approach is based on an error gauge that is constructed by pairing Kaczmarz's method with its reverse-ordered method; we stop the iterations when this error gauge is minimum. Our first proposed method stops when the error gauge is at a minimum, the second uses the error gauge to determine step sizes.

3: Our numerical experiments demonstrate that our two methods are superior to the standard Kaczmarz method equipped with state-of-the-art statistical stopping rules.

4: Even compared to Kaczmarz's method equipped with an oracle that provides the exact error -- and is thereby able to stop at the best possible iterate -- our methods perform better in almost 90\% of our test cases.

5: In terms of computational cost, our methods are a little cheaper than standard Kaczmarz equipped with a statistical stopping rule.

6: 

7: 

8: 

9: % \mh{We propose a new stopping rule for Kaczmarz's method based on general principles in scientific computing \mhq{(or numerical analysis?)}.

10: % We define an error gauge based on using Kaczmarz's method and its reversed-order version in tandem.

11: % The error gauge is demonstrated to give a highly \mhq{(omit highly?)} accurate stopping rule.

12: % We also use the error gauge to compute step sizes, which results in a monotonically convergent method.

13: % Using several numerical examples, we demonstrate that both methods perform better than an oracle that provides the exact error.

14: % In terms of cost, our methods are a little cheaper than state-of-the-art stopping rules such as GCV and UPRE.}

15: \end{abstract}

16: