Spatio-temporal graphical modeling with innovations based on multi-scale diffusion kernel

A random field of interest is observed on an undirected spatial graph over time, thereby providing a time series of dependent random fields. We propose a general modeling procedure which has the potential to explicitly quantify intrinsic and extrinsic fluctuations of such dynamical system. We adopt a paradigm in which the intrinsic fluctuations correspond to a process of latent diffusion on the graph arising from stochastic interactions within the system, whereas the extrinsic fluctuations correspond to a temporal drift reflecting the effects of the environment on the system. We start with a spatio-temporal diffusion process which gives rise to the latent spatial process. This makes a bridge with the conventional Wold representation, for which the latent process represents the innovation process, and beyond that with the stochastic differential equation associated to the Fokker-Planck dynamic. The innovation process is modeled by a Gaussian distribution whose covariance matrix is defined by a multi-scale diffusion kernel. This model leads to a multi-scale representation of the spatio-temporal process. We propose a statistical procedure to estimate the multi-scale structure and the model parameters in the case of the vector autoregressive model with drift. Modeling and estimation tasks are illustrated on simulated and real biological data.