Metrics for sparse graphs

Recently, Bollobás, Janson and Riordan introduced a very general family of random graph models, producing inhomogeneous random graphs with Θ(n) edges. Roughly speaking, there is one model for each kernel, i.e., each symmetric measurable function from [0, 1] to the non-negative reals, although the details are much more complicated, to ensure the exact inclusion of many of the recent models for large-scale real-world networks. A different connection between kernels and random graphs arises in the recent work of Borgs, Chayes, Lovász, Sós, Szegedy and Vesztergombi. They introduced several natural metrics on dense graphs (graphs with n vertices and Θ(n) edges), showed that these metrics are equivalent, and gave a description of the completion of the space of all graphs with respect to any of these metrics in terms of graphons, which are essentially bounded kernels. One of the most appealing aspects of this work is the message that sequences of inhomogeneous quasi-random graphs are in a sense completely general: any sequence of dense graphs contains such a subsequence. Alternatively, their results show that certain natural models of dense inhomogeneous random graphs (one for each graphon) cover the space of dense graphs: there is one model for each point of the completion, producing graphs that converge to this point. Our aim here is to briefly survey these results, and then to investigate to what extent they can be generalized to graphs with o(n) edges. Although many of the definitions extend in a simple way, the connections between the various metrics, and between the metrics and random graph models, turn out to be much more complicated than in the dense case. We shall prove many partial results, and state even more conjectures and open problems, whose resolution would greatly enhance the currently rather unsatisfactory theory of metrics on sparse graphs. This paper deals mainly with graphs with o(n) but ω(n) edges: a companion paper will discuss the (more problematic still) case of extremely sparse graphs, with