← All Courses

From Mean Field Theory to Variational Autoregressive Networks

Background for reading Wu, Wang & Zhang, "Solving Statistical Mechanics Using Variational Autoregressive Networks" (arXiv:1809.10606). Builds from the variational principle through naïve mean field theory to the one-layer VAN architecture, with worked examples throughout.

Prerequisites: Multivariable calculus, basic probability (Bayes' rule, conditional distributions), some familiarity with statistical mechanics (partition function, Boltzmann distribution). No machine learning background assumed — the neural network connection is built from scratch.
Part I — Variational Free Energy & Mean Field Theory
01

Variational Free Energy & Naïve Mean Field for the Ising Model

The variational principle, KL divergence, the factorized ansatz, NMF self-consistency equations. Applications to the 1D Ising ring (spurious transition) and the 2D square lattice (comparison with Onsager's exact solution).

Sections 1–4 · 6 figures · Landau expansion · Critical exponents
Part II — From NMF to Autoregressive Networks
02

From NMF to Variational Autoregressive Networks

The autoregressive decomposition, Bernoulli conditionals with lower-triangular weights, the bias question (paper vs. notes conventions), why one layer captures pairwise correlations, worked N=3 example, and the VAN training loop.

Sections 5–10 · Bias discussion · Small-W expansion · REINFORCE gradient