1: \begin{abstract}

2:   We introduce a new, elementary method for studying random

3:   differences in arithmetic progressions and convergence phenomena

4:   along random sequences of integers. We apply our method to obtain

5:   significant improvements on two results. The first improvement is

6:   the following: Let $\ell$ be a positive integer and let $\mathbbm

7:   u_1\ge \mathbbm u_2\ge \dots$ be a decreasing sequence of

8:   probabilities satisfying $\mathbbm u_n\cdot n^{1/(\ell +1)}\to

9:   \infty$. Let $R=R^\omega$ be the random sequence we get by selecting

10:   the natural number $n$ with probability $u_n$. If $A$ is a set of

11:   natural numbers with positive upper density, then $A$ contains an

12:   arithmetic progression $a, a+r,a+2r,\dots,a+\ell r$ of length $\ell

13:   +1$ with difference $r\in R^\omega$. The best earlier result (by

14:   M. Christ and us) was the condition $\mathbbm u_n\cdot

15:   n^{2^{-\ell+1}}\to \infty$ with a logarithmic speed.  The new bound

16:   is better when $\ell \ge 4$.

17: 

18:   The other improvement concerns almost everywhere convergence of

19:   double ergodic averages: we (randomly) construct a sequence

20:   $r_1<r_2<\dots$ of positive integers so that for any $\epsilon>0$ we

21:   have $r_n/n^{2-\epsilon}\to \infty$ and for any measure preserving

22:   transformation $T$ of a probability space the averages

23:   \begin{equation}

24:     \frac1N\sum_{n<N}T^nF_1(x)T^{r_n}F_2(x)

25:   \end{equation}

26:   converge for almost every $x$. Our earlier best result was the

27:   $r_n/n^{(1+1/14)-\epsilon}\to \infty$ growth rate on the sequence

28:   $(r_n)$.

29: \end{abstract}

30: