Two-Sample Testing can be as Hard as Structure Learning in Ising Models: Minimax Lower Bounds

Consider the following structural two-sample testing problem: given two sets of sample drawn from Ising models, determine whether the underlying network structure has changed. In [1], we showed that for Ising models over p variables with network structures that have degree bounded by d, under mild conditions on the model parameters, the sample complexity of this problem is very close to that of determining either of the network structures. Therefore, the naive scheme of learning and then comparing the structures of both sets of samples is near data-optimal. However, the minimax lower bounds in [1] relied on Ising models that differed in only one edge, which leads to the natural follow-up question: are large changes significantly easier to detect? We extend the previously developed framework to consider this problem, and show that, in a certain parameter regime, large changes do not provide any significant improvement in the number of necessary samples for reliable two-sample testing.