Accelerated Random-Sweep Gibbs Sampling for Gaussian Graphical Models via Dual Normal Factor Graphs

What the math gives to ML

The paper gives a concrete acceleration strategy for random-sweep Gibbs sampling by moving a Gaussian graphical model into a Fourier-dual normal factor graph. The transferable assets are dual-domain sampling, exact spectral convergence formulas, and a covariance-complement identity derived from the Woodbury matrix identity. These tools are useful for graph-structured Gaussian latent variables in graph VAEs, Bayesian GNNs, and neural world models. The strongest experiments are to compare primal and dual Gibbs chains at equal computational cost and to use dual covariance statistics as a memory-efficient uncertainty representation.