50 years of spin glass theory

David Sherrington, Scott Kirkpatrick · 2025

Read 50 years of spin glass theory by David Sherrington, Scott Kirkpatrick (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-24432.txt.

Introduction

David Sherrington and Scott Kirkpatrick look back on fifty years of spin glass theory, in a 2025 review. In 1975 two papers launched the field: one by Edwards and Anderson, and one by Sherrington himself with Kirkpatrick. Both tried to understand experiments hinting at a new phase of matter, a magnet whose atoms freeze into disorder rather than pattern. The review revisits the key ideas of those papers and how they were put to work.

Spin glass thinking later fed optimization, computing, and even neural network theory. The authors select the later developments they judge most consequential. Half a century on, the solvable model of a hopelessly disordered magnet remains one of physics' most exported ideas.

Catalog record

CREATORDavid Sherrington, Scott Kirkpatrick
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE10 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-24432

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

50 years of spin glass theory David Sherrington, Scott Kirkpatrick Abstract In 1975, two papers were published that together sparked major new directions, conceptual, mathematical and practically applicable, in several previously disparate fields of science. In this short review, we expose key aspects of their thinking, implementations and implications, along with a selection of further crucial and consequential developments. These papers were “Theory of spin glasses” by S.F.Edwards and P.W.Anderson (EA) S.F.Edwards and Anderson 197 and “Solvable Model of a Spin-Glass” by D.Sherrington and S.Kirkpatrick (SK) D.Sherrington and S.Kirkpatrick 1975 , both concerned with trying to understand recent experiments that suggested a new phase of matter. The experiments date from around 1970. They were on substitutional alloys of magnetic transition and non-magnetic noble metals and indicated that at low but finite concentration of the magnetic ions and low temperature the magnetic moments (spins) are individually ‘frozen’ in their orientation over nuclear spin resonance timescales, but in a quasi-random, non-periodic, fashion. It was named spin glass by analogy with positional amorphousness in normal glasses. Initially, transition to this behaviour as temperature was reduced seemed gradual, but in 1972 Cannella and Mydosh V.Cannella and J.A.Mydosh 1972 observed that it sharpened as the external field was reduced, suggesting a cooperative phase transition to a hitherto-unknown novel phase. However, there was no potentially satisfying theoretical explanation until EA took an interest. They used a statistical physics model approach but with many inspirationally-new concepts and mathematical tools. Among them were a novel but irregular mathematical procedure for evaluating quenched averages of observables, involving a combination of replication and limit-taking, {\langle{\ln Z}\rangle}=\lim_{n\rightarrow 0}{n^{-1}\{{Z^{n}}-1}\};Z^{n}={\Pi_{\alpha=1,..n}}Z_{\alpha} where the \alpha label replicas, a new type of order parameter, q^{\alpha\beta}=N^{-1}\sum_{i}{\langle{S_{i}^{\alpha}S_{i}^{\beta}}\rangle};{\alpha\neq\beta} , as well as replacing (experimental) random site disorder by ’easier-to-deal-with’ (theoretical) random bond disorder. With some further assumptions, approximations and ansätze they were able to demonstrate a phase transition to a new ‘amorphous’ cooperative spin ordering, EA’s paper immediately excited one of us (DS) by both its conceptually new ideas and its new methods of analysis, but he also found it challenging in evaluating the correctness of the assumptions and approximations. To test these further, he devised a related model that he expected should be solvable exactly, with an infinite-ranged and range-independent interaction distribution. Also, for calculational simplicity while maintaining conceptuality, he considered discrete Ising spins. Using the same two ansätze as EA, he derived Ising analogues of their equations and interested the other of us (SK) in further joint studies, both analytic and by computer simulation. These yielded a phase diagram in qualitative accord with experiment, but also led to discoveries that had major further consequence; (i) a prediction of a negative entropy at zero temperature, a result that is fundamentally forbidden for discrete variables and signalled a serious procedural error, (ii) evidence of a complex ‘energy landscape’ with many hills and valleys on many levels, hindering equilibration. More detailed computer simulations, reported in a subsequent longer paper (KS) Kirkpatrick and Sherrington 1978 , indicated a lower ground state energy than the calculation. There was a problem. EA and SK had both made a ‘natural’ ansatz for the order parameter q^{\alpha\beta} that it should have the same value for any non-identical pair of replicas {\alpha}\neq{\beta} , now known as ‘replica symmetry’ (RS). However, two years later, J.R.L. de Almeida and D.J Thouless de Almeida and Thouless 1978 showed that this choice was unstable against fluctuations in replica space. Two further years later, Giorgio Parisi Parisi 1979 devised a remarkable ‘replica symmetry breaking’ (RSB) ansatz that solved the negative entropy problem and also gave a lower ground state energy result close to the KS simulation. It involved radically new mathematical conceptualization and formulation, in which the order parameter is now a function q(x);0\leq x\leq 1 . In 1983 he provided its physical explanation Parisi 1983 and opened a new insightful physical conceptualization. Interest and activity exploded, leading to many further new discoveries Mézard et al. 1984 , such as ultrametricity, non-selfaveraging, unusual dynamics, breakdown of the usual fluctuation-dissipation theorem, extensions to new theoretical models and further new concepts and perspectives, e.g. to simple glasses and multiple meta-equilibria, leading eventually to Parisi’s Nobel Prize of 2021. While EA chose to study the symmetric zero-mean random exchange distribution, SK included provision for a non-zero mean to emulate the experimental observations of transitions from ferromagnetism to spin glass with lowering of the concentration of magnetic constituent. SK-AT-Parisi-Toulouse G.Toulouse 1980 : \overline{J_{ij}}=J_{0}/N,\overline{(J_{ij}^{2}-{\overline{J_{ij}}^{2}})}={{J_{0}}^{2}}/N Later, in 1982, John Hopfield Hopfield 1982 proposed and demonstrated by computer simulation a model system for distributed attractor memory and associative retrieval, in which the stored memories can be interpreted as an extension of SK to include many gauge-transformed ferromagnets, with corresponding attractor basins. Through an extension of SK analysis, Amit, Gutfreund and Sompolinsky D.Amit et al. 1987 demonstrated its resultant combination of ‘basins of attraction’ of stored memories, up to its critical retrieval capacity, their competitively constrained ’basin extents’ and passage to ’spin glass complexity’ beyond that capacity. Hopfield’s model did not include explicitly the ‘exchange-distribution’ variance term of EA, J in SK. Rather, for each retrieved memory, the non-retrieved memory terms in the Hamiltonian act as potentially-disruptive quenched ‘noise’. Both biological brains and artificially intelligent neural networks require mechanisms for learning as well as recalling and generalizing. Within an extension of the SK conceptualisation this can be translated to gradually modifying the interactions/synapses in a learning stage, to be later recalled or generalized in faster neural dynamics, either in a recurrent version (as one might model the brain) or a directed layered one as in most AI/’deep learning’ applications. EA and SK are examples of optimization in the presence of stochasticity, temperature noise. At the time of our collaboration, Scott K, at IBM Research, was involved in understanding problems of designing computer systems and their components. He recognised that extensions of the computer-simulation methods developed for spin glasses would be a natural and powerful generalization of the highly specialised and more limited approaches then used in the placement of components and the routing of signals between them. The generalization to other optimization problems led to the concept and application of ‘simulated annealing’ S.Kirkpatrick et al. 1983 . EA’s theoretical replication concept has also led to informative computer-simulational analogues, in which replica systems with identical quenched disorder and the same temperature level of noise are allowed to evolve independently and correlations between them studied to indicate spin glass-like transitions, critical properties and physical measures Martın-Mayor et al. 2023 . Another consequence within computer science has been the recognition of interest in ‘typical’ as well as ‘worst-case’ studies. As well as both EA and SK stimulating much further research in theoretical physics Charbonneau et al. 2023 , the SK model also provided new exciting challenges for mathematical physicists Bolthausen and Bovier 2007 Guerra 2003 and probabilists M.Talagrand 2011 Panchenko 2013 and led to several new developments in both subjects. Its rigorous solution by Michel Talagrand played a significant role in his award of the 2024 Abel Prize. So, the SK model is indeed ‘solvable’ but the solution has proven to be very subtle, as well as consequential. Prior to SK, mean field/infinite-range theory was mostly considered to be fairly trivial .

Read the complete text