Informations pratiques sur cette ressource

Description de la ressource pédagogique

Description (résumé)

Bordenave and Caputo (2014) defined a notion of entropy for probability distributions on rooted graphs with finite expected degree at the root. When such a probability distribution rho has finite BC entropy Sigma(rho), the growth in the number of vertices n of the number of graphs on n vertices whose associated rooted graph distribution is close to rho is as d/2 n log n + Sigma(rho) n + o(n), where d is expected degree of the root under rho. We develop the parallel result for probability distributions on marked rooted graphs. Our graphs have vertex marks drawn from a finite set and directed edge marks, one towards each vertex, drawn from a finite set. The talk will focus on presenting an overview of the technical details of this extension We are motivated by the interpretation of a discrete time stochastic process taking values in a finite set Theta as the local weak limit of long strings of symbols from Theta. We argue that probability distributions on marked rooted graphs are the natural analogs of stochastic process models for *graphical data*, by which we mean data indexed by the vertices and edges of a sparse graph rather than by linearly ordered time. Our extension of the BC entropy can then be argued to be the natural extension, in the world of graphical data, of the Shannon entropy rate in the world of time series. We illustrate this viewpoint by proving a lossless data compression theorem analogous to the basic lossless data compression theorem for time series. Joint work with Payam Delgosha.

"Domaine(s)" et indice(s) Dewey

• Probabilités, Statistiques mathématiques, Mathématiques appliquées (519)

Thème(s)