Identifiability of flow distributions from link measurements with applications to computer networks

Abstract

We study the problem of identifiability of distributions of flows on a graph from aggregate measurements collected on its edges. This is a canonical example of a statistical inverse problem motivated by recent developments in computer networks. In this paper (i) we introduce a number of models for multi-modal data that capture their spatio-temporal correlation, (ii) provide sufficient conditions for the identifiability of nth order cumulants and also for a special class of heavy tailed distributions. Further, we investigate conditions on network routing for the flows that prove sufficient for identifiability of their distributions (up to mean). Finally, we extend our results to directed acyclic graphs and discuss some open problems.