1: \begin{abstract}

2: %Binary experiments are performed on $d$ different groups of people, and a statistician aims to estimate $\bold p:=(p_i)_{i\in [d]}$, the chances of positive outcomes within each group. The notion of almost exchangeability, introduced by de Finetti, describes experiments for which the division in groups is almost irrelevant. If nothing else is assumed about the model, a class of natural priors is indexed by networks $C$, measuring the relative discrepancies among categories, and a large parameter $A$, measuring the distance from complete exchangeability. We analyze the rate of convergence of a Gibbs sampler used to sample from these measures. We show that, for every fixed network $C$, $A^2$ steps are necessary and sufficient to mix in a suitable Wasserstein distance. The upper and lower bounds are explicit and depend on the network $C$ only through few spectral parameters.

3: %\end{abstract}

4: