Wavelet-based graph inference using multiple testing

Graph-based representation enables to outline efficiently interactions between sensors and as such has encountered a growing interest. For example in neurosciences, the graph of interactions between brain regions has shed lights on evolution of diseases. In this paper, we describe a whole procedure which estimates the graph from multivariate time series. First correlations using wavelet decomposition of the signals are estimated. Bonferroni (1935)'s procedure on multiple correlation testing is then used. We prove theoretically that the Family Wise Error Rate (FWER) is asymptotically controlled for any graph structures. We implement our approach on small-world graph structures, with signals possibly having long-memory properties. This structure is inspired by real data examples from resting-state functional magnetic resonance imaging. The control is confirmed graphically. Numerical simulations illustrate the behavior of the bias and the power of our proposed approach.