Abstract: About sixty years ago, the anomalous magnetic response of certain magnetic alloys drew the attention of theoretical physicists. It soon became clear that understanding these systems, now called spin glasses, would give rise to a new branch of statistical physics. As physical materials, spin glasses were found to be as useless as they were exotic. They have nevertheless been recognized as paradigmatic examples of complex systems with applications to problems as diverse as neural networks, amorphous solids, biological molecules, social and economic interactions, information theory and constraint satisfaction problems. This paper provides a comprehensive review and synthesis of the Random Energy Model (REM) and its hierarchical successor, the Generalized Random Energy Model (GREM), which serve as foundational models for understanding these complex phenomena.
Introduction
In the late 1970s, the Sherrington-Kirkpatrick (SK) model of spin glasses introduced a massive theoretical challenge: the "replica trick" used to solve it led to non-physical results unless one assumed a bizarre structure called Replica Symmetry Breaking (RSB). RSB implies that a system doesn't just settle into one state, but a complex, fluctuating landscape of many possible states.
The authors address the need for a simplified framework to test these radical ideas. While the SK model is notoriously difficult, the REM (introduced by Derrida in 1980) provides a limit where the energies of different configurations are completely uncorrelated. This paper re-examines these models to explain why they work, how they relate to more complex $p$-spin models, and how they describe the "sluggish" behavior of glasses.
Methodology
The paper is divided into two primary methodological approaches:
1. Equilibrium Analysis
The authors use the microcanonical ensemble to derive the phase diagram of the REM. They employ Parisi’s RSB scheme to show how the "replica trick" can be applied to these models. They also utilize Poisson Point Processes (specifically Ruelle cascades) to describe the statistics of the lowest energy levels, which dominate the system at low temperatures. The analysis is based on:
- The REM: $2^N$ configurations with independent, identically distributed (i.e.d.) Gaussian energies.
- The GREM: A hierarchical version of the REM where energies are correlated through a tree-like structure.
- The $p$-spin SK Model: A more complex model that reduces to the REM in the limit where $p \to \infty$.
2. Dynamical Analysis
To study how these systems evolve, the authors use Markov jump processes (specifically Random Hopping and Metropolis dynamics). They introduce the concept of a "clock process"—a mathematical tool that tracks the cumulative time the system spends trapped in various energy valleys. By scaling these processes, they map the system's evolution onto stable subordinators (a type of Lévy process).
Key Insights and Results
- The $p \to \infty$ Bridge: The REM is the exact limit of the $p$-spin SK model. This justifies using the REM to understand the "one-step RSB" transition found in more realistic glasses.
- Freezing and Overlaps: In the low-temperature phase, the system becomes "frozen." The probability of two random configurations being identical (the "overlap") doesn't settle to a single value but fluctuates even in the thermodynamic limit—a hallmark of RSB.
- The Arcsine Law of Aging: One of the most striking insights is that "aging"—the phenomenon where a glass relaxes more slowly the longer it has been left alone—is universal. The authors show that for both the REM and the $p$-spin model, the correlation between the system's state at two different times follows the generalized arcsine law.
- Phase Transition: The REM undergoes a sharp phase transition at a critical temperature $T_c$. Below this, the entropy per spin is zero, and the system is dominated by a handful of low-energy states.
- Metropolis vs. Hopping: The paper proves that even though Metropolis dynamics are much harder to analyze than simple random hopping, they yield the same universal aging behavior in the REM.
Discussion and Future Directions
While the extensive part of the free energy is well-understood, "sub-leading" corrections remain subtle and differ between models. Furthermore, while the REM's Metropolis dynamics are solved here, doing the same for the $p$-spin SK model remains an open challenge due to the more complex "valley" structure of its energy landscape. The paper suggests that a more advanced replica approach, allowing for fluctuating block sizes, might be needed to capture finite-size effects and temperature-dependent overlaps more accurately.
Summary updated