Universality classes for extreme-value statistics

Jean‐Philippe Bouchaud, Marc MézardJournal of Physics A Mathematical and General 1997 · DOI · arXiv

Abstract

The low temperature physics of disordered systems is governed by the statistics of extremely low energy states. It is thus rather important to discuss the possible universality classes for extreme value statistics. We compare the usual probabilistic classification to the results of the replica approach. We show in detail that one class of independent variables corresponds exactly to the so-called one step replica symmetry breaking solution in the replica language. This universality class holds if the correlations are sufficiently weak. We discuss the relation between the statistics of extremes and the problem of Burgers turbulence in decay.

Introduction

In the study of disordered systems, such as spin glasses or polymers in random media, the physics at very low temperatures is dominated by a small number of "extreme" states—the lowest possible energy configurations. To analyze these systems, physicists frequently employ the replica method, a mathematical technique that involves studying $n$ copies of a system and taking the limit as $n \to 0$. Within this framework, a specific scheme known as Replica Symmetry Breaking (RSB) is essential for obtaining physically correct results.

Simultaneously, the field of Extreme Value Statistics (EVS) in mathematics classifies the behavior of the maximum or minimum of a set of random variables into universality classes: Gumbel, Fréchet, and Weibull. This paper bridges these two disciplines by formalizing the link between RSB schemes and EVS universality classes, specifically examining how correlations between variables affect this classification.

Theoretical Framework

The authors utilize several key theoretical tools to establish this connection:

  • The Random Energy Model (REM): Derrida’s REM serves as the primary model where energy levels are treated as independent random variables.
  • Replica Method (1RSB): The "one-step" RSB ansatz is applied to calculate the free energy and participation ratios, which measure the concentration of the probability measure on a few states.
  • Scaling Analysis of EVS: The Gumbel distribution is derived for independent variables with tails that decay faster than any power law.
  • Cole-Hopf Transformation: This transformation maps the Burgers' equation for decaying turbulence onto a partition function problem identical to a d-dimensional REM.

The 1RSB-Gumbel Connection

A primary insight of this work is the proof that the 1RSB solution is the physical manifestation of the Gumbel universality class. When a system's energy states are independent and possess fast-decaying tails, it necessarily follows 1RSB logic at low temperatures. The "glass transition" in these models occurs when the probability measure condenses onto a finite number of extreme energy levels.

The function $\phi(e)$ characterizing the energy distribution above and below the critical temperature $T_c$.

As illustrated in Figure 1, the function $\phi(e)$, which characterizes the energy distribution, undergoes a qualitative change at the critical temperature $T_c$. This transition marks the point where the system's behavior becomes dominated by extreme values, a hallmark of the 1RSB phase.

Key Results and Applications

The paper provides several concrete results that solidify the 1RSB-Gumbel link:

  • Participation Ratios: The authors demonstrate that the moments $Y_k = \sum w_i^k$ (where $w_i$ are the weights of states) calculated via EVS match the 1RSB results exactly: $Y_k = \frac{\Gamma(k-\mu)}{\Gamma(k)\Gamma(1-\mu)}$.
  • Burgers' Correlation: They successfully reproduce the universal scaling function for velocity correlations in decaying turbulence, proving that the "shocks" in the fluid correspond to the "cusps" in the extreme energy landscape.
  • Robustness to Correlation: For Gaussian variables with correlations decaying as $r^{-(2-d-\eta)}$, the 1RSB/Gumbel class holds as long as $\eta > 2-d$. Below this threshold, the system likely shifts to continuous RSB.

Conclusion and Future Directions

While the link between 1RSB and the Gumbel class is well-defined, other EVS classes like Fréchet (power-law tails) do not fit easily into the standard 1RSB framework because they cause moments of the partition function to diverge. Additionally, the EVS equivalent of "continuous RSB" remains an open question, likely relating to variables with hierarchical, long-range correlations. Whether the ultrametric structure of states holds in finite-dimensional physical space remains a significant area for future research.

Summary updated

Abstract

The low temperature physics of disordered systems is governed by the statistics of extremely low energy states. It is thus rather important to discuss the possible universality classes for extreme value statistics. We compare the usual probabilistic classification to the results of the replica approach. We show in detail that one class of independent variables corresponds exactly to the so-called one step replica symmetry breaking solution in the replica language. This universality class holds if the correlations are sufficiently weak. We discuss the relation between the statistics of extremes and the problem of Burgers turbulence in decay.
BibTeX
@article{bouchaud1997universality,
  author = {Jean‐Philippe Bouchaud and Marc Mézard},
  title = {{Universality classes for extreme-value statistics}},
  journal = {Journal of Physics A Mathematical and General},
  year = {1997},
  doi = {10.1088/0305-4470/30/23/004},
}
Figure