The Extreme Landscape

For independent, identically distributed random variables, extreme value theory is complete. The maximum of n samples, properly rescaled, converges to one of three universal distributions: Gumbel, Fréchet, or Weibull. The classification depends only on the tail behavior of the underlying distribution. This is classical, clean, and insufficient for physics.

Physical systems have correlations. The maximum of n correlated random variables does not, in general, converge to any of the three classical distributions. The rescaling, the limiting distribution, and the universality class all change when correlations are present.

The paradigmatic examples: time series generated by random walks, eigenvalues of random matrices, and interfaces in the KPZ universality class. For the maximum of a Brownian motion, the rescaled limit involves the Airy distribution. For the largest eigenvalue of a random matrix, the Tracy-Widom distribution. For KPZ fluctuations, the Tracy-Widom distribution appears again but through a different mechanism — the connection is deep and not yet fully understood.

The applications reach across statistical physics. The Random Energy Model — Derrida’s model of a disordered system — maps extreme value statistics to thermodynamic phase transitions: the ground state energy IS the minimum of correlated random variables, and the phase transition occurs when the extreme value distribution changes character. Stochastic search problems — the time for a random searcher to find a target — are controlled by the minimum of first-passage times, another extreme value problem.

Each physical system introduces its own correlation structure, and each correlation structure produces its own extreme value distribution. The taxonomy of universality classes in extreme value theory mirrors the taxonomy of universality classes in statistical mechanics.