Abstract
We present a mean field model for spin glasses with a natural notion of distance built in, namely, the Edwards-Anderson model on the diluted $D$-dimensional unit hypercube in the limit of large $D$. We show that finite $D$ effects are strongly dependent on the connectivity, being much smaller for a fixed coordination number. We solve the nontrivial problem of generating these lattices. Afterward, we numerically study the nonequilibrium dynamics of the mean field spin glass. Our three main findings are the following: (i) the dynamics is ruled by an infinite number of time sectors, (ii) the aging dynamics consists of the growth of coherent domains with a nonvanishing surface-volume ratio, and (iii) the propagator in Fourier space follows the $p^4$ law. We study as well the finite $D$ effects in the nonequilibrium dynamics, finding that a naive finite size scaling ansatz works surprisingly well.
Introduction
In the study of disordered systems, spin glasses represent a fundamental challenge due to the conflict between two primary theoretical frameworks: the Droplet picture and Replica Symmetry Breaking (RSB). While the Droplet picture provides a simple spatial intuition, RSB offers a sophisticated Mean Field (MF) description. However, traditional MF models, such as the Sherrington-Kirkpatrick model, lack a sense of physical distance, making it difficult to analyze spatial structures like coherent domains.
Fernández et al. address this gap by introducing the Edwards-Anderson model on a high-dimensional hypercube. By placing spins on a $D$-dimensional hypercube and controlling their connectivity, they create a system that behaves like a Mean Field model in the large-$D$ limit while retaining a natural metric for distance.
Methodology: The Hypercube Lattice
The researchers utilized a diluted $D$-dimensional unit hypercube. A critical methodological discovery was that standard Poissonian dilution (randomly removing links) introduced significant finite-size effects that obscured the underlying physics. To mitigate this, the authors developed a novel algorithm to generate lattices with fixed connectivity ($z=6$). This approach ensures the graph remains isotropic and locally "tree-like," significantly accelerating convergence to the infinite-dimensional limit.
Distance was defined using the "postman metric" (the minimum number of edges between vertices), which is more appropriate for hypercube geometries than Euclidean distance. Simulations were performed using Multispin Coding, allowing for the parallel processing of 64 independent systems on a single CPU word, enabling the study of systems up to $D=22$ (approximately 4.2 million spins).
Key Insights and Results
The Surface-Volume Ratio and RSB
A central finding of the study is the behavior of the surface-to-volume ratio of growing domains. In the Droplet theory, domains are expected to be compact, leading to a vanishing surface-volume ratio as they grow. Conversely, the hypercube model shows that this ratio remains constant, suggesting that domains are fractal-like or "spongy." This provides strong numerical evidence in favor of the RSB picture in a spatial context.
Aging and Domain Growth
The nonequilibrium dynamics are characterized by the growth of a coherence length, $\xi(t_w)$, which scales roughly as $\log t_w$. The spatial correlation functions collapse under specific scaling laws, confirming that aging is governed by an infinite hierarchy of time sectors rather than a single relaxation time.
The $p^4$ Propagator
The study also provides a rare numerical confirmation of the $1/p^4$ propagator law in Fourier space. This specific mathematical prediction of RSB theory describes how spatial correlations decay, and its verification on the hypercube validates the model's ability to capture complex spatial MF phenomena.
Conclusion
By successfully implementing the Edwards-Anderson model on a high-dimensional hypercube, the authors have created a powerful "playground" for testing glassy phenomena. Their results not only support the RSB framework but also demonstrate that spatial dynamics—such as domain growth and fractal surfaces—can be rigorously studied within a Mean Field architecture. Future work may extend this framework to optimization problems and other complex disordered systems.
Summary updated