1: \begin{abstract}

2: We study nonparametric methods for the setting where multiple distinct networks are observed on the same set of nodes.

3: Such samples may arise in the form of replicated networks drawn from a common distribution, or in the form of

4: heterogeneous networks,

5: with the network generating process varying from one network to another,

6: e.g.~dynamic and cross-sectional networks.

7: Nonparametric methods for undirected networks have focused on estimation of the graphon model. 

8: While the graphon model accounts for nodal heterogeneity, it does not account for network heterogeneity,

9: a feature specific to applications where multiple networks 

10: are observed.

11: To address this setting of multiple networks,

12: we propose a \textit{multi-graphon} model which allows node-level as well as network-level heterogeneity.  

13: We show how information from multiple networks can be leveraged to enable estimation of the multi-graphon via standard nonparametric regression techniques, e.g.~kernel regression, 

14: orthogonal series estimation. We study theoretical properties of the proposed estimator establishing recovery of the latent nodal positions up to negligible error, 

15: and convergence of the multi-graphon estimator to the normal distribution.

16: Finite sample performance are investigated in a simulation study and application to

17:  two real-world networks---a dynamic contact network of ants and a collection of structural brain networks from different subjects---illustrate the utility of our approach. 

18: \end{abstract}

19: