Testing properties of graphs and functions

We define an analytic version of the graph property testing problem, which can be formulated as studying an unknown 2-variable symmetric function through sampling from its domain and studying the random graph obtained when using the function values as edge probabilities. We give a characterization of properties testable this way, and extend a number of results about ``large graphs'' to this setting. These results can be applied to the original graph-theoretic property testing.