1: \begin{abstract}

2:  Many statistical estimators are defined as the fixed point of a

3:   data-dependent operator, with estimators based on minimizing a

4:   cost function being an important special case.  The limiting

5:   performance of such estimators depends on the properties of the

6:   population-level operator in the idealized limit of infinitely many

7:   samples.  We develop a general framework that yields bounds on

8:   statistical accuracy based on the interplay between the

9:   deterministic convergence rate of the algorithm at the population

10:   level, and its degree of (in)stability when applied to an empirical

11:   object based on $n$ samples.  Using this framework, we analyze both

12:   stable forms of gradient descent and some higher-order and unstable

13:   algorithms, including Newton's method and its cubic-regularized

14:   variant, as well as the EM algorithm. We provide applications of our

15:   general results to several concrete classes of models, including

16:   Gaussian mixture estimation, non-linear regression models, and informative

17:   non-response models.  We exhibit cases in which an unstable

18:   algorithm can achieve the same statistical accuracy as a stable

19:   algorithm in exponentially fewer steps---namely, with the number of

20:   iterations being reduced from polynomial to logarithmic in sample

21:   size $n$.

22: \end{abstract}

23: