1: \begin{abstract}

2: \begin{singlespace}

3: \noindent The Kaczmarz's alternating projection method has been widely

4: used for solving a consistent (mostly over-determined) linear system

5: of equations $A\bm{x}=\bm{b}$. Because of its simple iterative nature

6: with light computation, this method was successfully applied in computerized

7: tomography. Since tomography generates a matrix $A$ with highly coherent

8: rows, randomized Kaczmarz algorithm is expected to provide faster

9: convergence as it picks a row for each iteration at random, based

10: on a certain probability distribution. It was recently shown that

11: picking a row at random, proportional with its norm, makes the iteration

12: converge exponentially in expectation with a decay constant that depends

13: on the scaled condition number of $A$ and not the number of equations.

14: Since Kaczmarz's method is a subspace projection method, the convergence

15: rate for simple Kaczmarz algorithm was developed in terms of subspace

16: angles. This paper provides analyses of simple and randomized Kaczmarz

17: algorithms and explain the link between them. It also propose new

18: versions of randomization that may speed up convergence.\end{singlespace}

19: 

20: \end{abstract}

21: