Learning by Confusion Locates a Parametrization-Dependent Fisher Information Maximum

Abstract

Learning by confusion (LBC) is a machine-learning method for locating phase transitions without assuming a known order parameter. Its characteristic accuracy curve contains a central maximum that is interpreted as the transition point, but what determines the location and finite-size scaling of this maximum has remained unclear. We derive an exact balance condition for the Bayes-optimal confusion peak and show that, for equilibrium systems with a single swept parameter, the peak closely follows the maximum of the corresponding Fisher information. This immediately gives its finite-size scaling and reveals an unexpected dependence on how the data are sampled: the same configurations arranged on grids uniform in $T$, $\ln T$, or $β=1/T$ can yield different finite-size pseudocritical trajectories. Tests on the Ising, Blume-Capel, Potts, and Baxter-Wu models confirm the prediction across continuous, tricritical, and first-order transitions. Our results identify the statistical quantity measured by LBC and show that finite-size machine-learning estimates can depend on the parametrization of the training data as well as on the underlying physics.