Robust directed tree approximations for networks of stochastic processes

We develop low-complexity algorithms to robustly identify the best directed tree approximation for a network of stochastic processes in the finite-sample regime. Directed information is used to quantify influence between stochastic processes and identify the best directed tree approximation in terms of Kullback-Leibler (KL) divergence. We provide finite-sample complexity bounds for confidence intervals of directed information estimates. We use these confidence intervals to develop a minimax framework to identify the best directed tree that is robust to point estimation errors. We provide algorithms for this minimax calculation and describe the relationships between exactness and complexity.