Inferring couplings in networks across order-disorder phase transitions

In recent years scientific progress depends increasingly on the ability to infer statistical models from high-dimensional data, yet a full picture of how inference works remains elusive. Here we take a step towards a quantitative understanding of the intrinsic interplay between statistical models, inference methods and data structure. We consider a reconstruction of ferromagnetic Ising models on Erdős-Renyi random graphs by direct coupling analysis (DCA)--a highly successful, global statistical approach to modeling amino acid sequence data. But DCA's success is not universal. Indeed we find that a much simpler local statistical approach outperforms DCA when data are limited. By tuning the temperature of the Ising model, we show that DCA excels at low temperatures where local methods are more severely blighted by the emergence of macroscopic order. We find further that inference is maximally performant near the order-disorder phase transition. But instead of a manifestation of critical data distributions, our analysis reveals that this behavior results from macroscopic ordering at low temperatures and large thermal noise at high temperatures both of which are detrimental to inference. Our work offers an insight into the regime in which DCA operates so successfully and more broadly how inference interacts with the nature of data-generating distributions.