1: \begin{abstract}

2: 

3: We present global 

4: convergence rates for a line-search method which is based on random

5: first-order models and directions whose 

6: quality is ensured only with certain probability. We show that in terms of the order of the accuracy, the evaluation 

7: complexity of such a method is the same as  its counterparts that use

8: deterministic accurate models; the use of 

9: probabilistic models only increases the complexity by a constant, which depends on the probability of the models being good. 

10: We particularize and improve  these results in the convex and strongly convex case. 

11: 

12: We also analyze a probabilistic cubic regularization 

13: variant that allows approximate probabilistic second-order models and show improved complexity bounds 

14: compared to probabilistic first-order methods; again, as a function of the accuracy, the probabilistic cubic regularization

15: bounds are of the same (optimal) order as for the deterministic case.

16: 

17: %Line-search methods are an important class of algorithms for

18: %unconstrained optimization which rely on  approximately computing a local descent direction

19: %and computing a step along this direction in such a way that a sufficient decrease is achieved. 

20: %To ensure that sufficient decrease is possible the direction has to satisfy certain requirements.

21: 

22: %Often, in practical applications,  ensuring that the direction properties are always what is needed is

23: %prohibitively expensive or impossible. This may be due to the fact that derivative information about the objective function is not available

24: %or because full gradient (and Hessian)  are too expensive to compute, or a model of the true objective function is too expensive to optimize accurately.

25: 

26: %In this paper we present the analysis of a line-search method which is based on random models and direction whose quality is ensured only with certain probability.

27: \end{abstract}

28: