1: \begin{abstract}

2:   We consider a group of agents connected by a social network who

3:   participate in {\em majority dynamics}: each agent starts with an

4:   opinion in $\{-1,+1\}$ and repeatedly updates it to match the

5:   opinion of the majority of its neighbors.

6: 

7:   We assume that one of $\{-1,+1\}$ is the ``correct'' opinion $S$,

8:   and consider a setting in which the initial opinions are independent

9:   conditioned on $S$, and biased towards it. They hence contain enough

10:   information to reconstruct $S$ with high probability. We ask whether

11:   it is still possible to reconstruct $S$ from the agents' opinions

12:   after many rounds of updates.

13: 

14:   While this is not the case in general, we show that indeed, for a

15:   large family of bounded degree graphs, information on $S$ is

16:   retained by the process of majority dynamics.

17: 

18:   Our proof technique yields novel combinatorial results on majority

19:   dynamics on both finite and infinite graphs, with applications to

20:   zero temperature Ising models.

21: \end{abstract}

22: