1: \begin{abstract}

2: We study how the learning rate affects the performance of

3: a relaxed randomized Kaczmarz algorithm for solving $A x \approx b + \varepsilon$,

4: where $A x =b$ is a consistent linear system and $\varepsilon$ has independent mean zero random entries. 

5: We derive a scheduled learning rate which optimizes a bound on the expected error that is sharp in certain cases;

6: in contrast to the exponential convergence of the standard randomized Kaczmarz algorithm, our optimized bound 

7: involves the reciprocal of the Lambert-$W$ function of an exponential.

8: \end{abstract}

9: