Temperature-dependent performance of the erasure machine

Interaction strengths in spin systems can be estimated from the maximum likelihood estimate (MLE) method based on information of spin configurations. The application of the standard MLE method has been known to face serious problems for large data sets since it demands huge computation cost. We test the performance of the erasure machine, which has been recently proposed to overcome such drawbacks of MLE, for the Sherrington–Kirkpatrick model and two-dimensional Ising model. In this paper, we focus on how the performance of the erasure machine varies with the temperature and find that it exhibits the best performance not at criticality, but at somewhat higher temperature. We also attempt to explain such temperature-dependent performance of the erasure machine.