1: \begin{abstract}

2: This article establishes the performance of stochastic blockmodels in

3: addressing the co-clustering problem of partitioning a binary array

4: into subsets, assuming only that the data are generated by a

5: nonparametric process satisfying the condition of separate

6: exchangeability.\vspace*{1pt} We provide oracle inequalities with rate of

7: convergence $\mathcal{O}_P (n^{-1/4} )$ corresponding to profile

8: likelihood maximization and mean-square error minimization, and show

9: that the blockmodel can be interpreted in this setting as an optimal

10: piecewise-constant approximation to the generative nonparametric model.

11: We also show for large sample sizes that the detection of co-clusters

12: in such data indicates with high probability the existence of

13: co-clusters of equal size and asymptotically equivalent connectivity in

14: the underlying generative process.

15: \end{abstract}