Mean field theory

What Is Mean Field Theory?

Mean field theory, abbreviated MFT and closely related to the self-consistent field methods used in quantum chemistry, is an approximation technique for many-body systems in which the interactions felt by any one component are replaced by an averaged, or mean, effect of all the others. The substitution turns an intractable coupled problem into a single-body problem whose solution must reproduce the average that was assumed, a requirement known as the self-consistency condition. Because the fluctuations that the averaging discards are precisely what makes many-body statistics hard, mean field theory is best understood as a controlled first approximation rather than an exact method.

The approach originates with Pierre Weiss's 1907 molecular field model of ferromagnetism and with the Curie-Weiss and Bragg-Williams treatments that followed. It is now standard across statistical physics, condensed matter theory, probability, and machine learning, wherever a large system of interacting variables must be summarized by a small number of order parameters.

Self-Consistency in the Ising Model

The Ising model gives the cleanest illustration. Each spin interacts with its lattice neighbors, but in the mean field treatment those neighbors are replaced by an effective field proportional to the average magnetization, giving a transcendental equation in which the magnetization appears on both sides. Solving it yields a nonzero magnetization below a critical temperature and zero above it, so the approximation predicts a continuous phase transition and a critical temperature scaling with the coordination number. Tutorial treatments comparing mean-field, effective-field, and exact results for Ising and Ising-like models show where the simple version succeeds and where effective-field corrections that retain some short-range correlations do better.

Accuracy, Dimension, and Fluctuations

Mean field theory becomes exact in the limit of infinite spatial dimension or infinite-range interactions, since each site then averages over so many neighbors that fluctuations vanish. In low dimensions it fails in specific, well-characterized ways: it predicts a phase transition for the one-dimensional Ising chain, where none exists, and it gives critical exponents such as a magnetization exponent of one half that disagree with exact two-dimensional results. The Ginzburg criterion quantifies when fluctuations dominate, and for the Ising universality class the upper critical dimension is four, above which mean field exponents are correct. Renormalization group methods were developed in part to supply what the approximation cannot. Sharpened error bounds, including information inequalities for the mean-field approximation, express the gap in terms of the Kullback-Leibler divergence between the true distribution and the best factorized one.

Variational Formulation and Machine Learning

Mean field theory can be derived variationally: among all product distributions, choose the one minimizing the Kullback-Leibler divergence to the target, which is equivalent to maximizing a lower bound on the log partition function or evidence. This framing carries the method directly into Bayesian statistics, where naive mean field variational inference underlies scalable approximations for topic models, Gaussian mixtures, and variational autoencoders, and coordinate ascent updates each factor while holding the others fixed. Work relating mean-field approximation to convex hierarchies connects the physics-derived lower bound to relaxations developed in theoretical computer science. Extensions include dynamical mean field theory for strongly correlated electrons, replica and cavity methods for spin glasses, and mean field games for large populations of interacting agents.

Applications

Mean field theory has applications in a wide range of disciplines, including:

  • Condensed matter physics, in magnetism, superconductivity, and correlated electron materials
  • Approximate Bayesian inference and probabilistic graphical models
  • Deep learning theory, including analyses of signal propagation and infinite-width networks
  • Epidemic and population modeling on networks
  • Economics and control, through mean field game formulations of large-agent systems
  • Polymer science and soft matter, via self-consistent field theory of block copolymers
Loading…