Several experimental techniques have shown that the primary response of many materials comes from a heterogeneous distribution of independently relaxing nanoscale regions; but most Monte Carlo simulations have homogeneous correlations. Resolving this discrepancy may require including the energy needed to change the configurational entropy, which is often used in theoretical treatments of thermal fluctuations, but not in computer simulations. Here the local configurational entropy is shown to give a nonlinear correction to the Metropolis algorithm that restores conservation of energy, maintains maximum entropy, and yields heterogeneous correlations. The nonlinear correction also improves agreement between Monte Carlo simulations of the Ising model and measurements of specific heat and structural correlations from the Jahn–Teller distortion in LaMnO 3 .