1: \begin{abstract}

2: \noindent 

3: Consider a low temperature stochastic Ising model in the phase

4: coexistence regime with Markov semigroup $P_t$. A fundamental and still

5: largely open problem is the understanding of the long time behavior of

6: $\d_\h P_t$ when the initial configuration $\h$ is sampled from a highly

7: disordered state $\nu$ (e.g. a product Bernoulli measure or a high

8: temperature Gibbs measure).  Exploiting recent progresses in the

9: analysis of the mixing time of Monte Carlo Markov chains for discrete

10: spin models on a regular $b$-ary tree $\Tree^b$, we tackle the above

11: problem for the Ising and hard core gas (independent sets) models on

12: $\Tree^b$. If $\nu$ is a biased product Bernoulli law then, under

13: various assumptions on the bias and on the thermodynamic parameters, we

14: prove $\nu$-almost sure weak convergence of $\d_\h P_t$ to an extremal Gibbs

15: measure (pure phase) and show that the limit is approached at

16: least as fast as a stretched exponential of the time $t$. In the

17: context of randomized algorithms and if one considers the Glauber

18: dynamics on a large, finite tree, our results prove fast local

19: relaxation to equilibrium on time scales much smaller than the true

20: mixing time, provided that the starting point of the chain is not taken

21: as the worst one but it is rather sampled from a suitable distribution.

22: \end{abstract}

23: