On Structured Variational Approximations

The problem of approximating a probability distribution occurs frequently in many areas of applied mathematics, including statistics, communication theory, machine learning, and the theoretical analysis of complex systems such as neural networks. Saul and Jordan (1996) have recently proposed a powerful method for efficiently approximating probability distributions known as structured variational approximations. In structured variational approximations, exact algorithms for probability computation on tractable substructures are combined with variational methods to handle the interactions between the substructures which make the system as a whole intractable. In this note, I present a mathematical result which can simplify the derivation of structured variational approximations in the exponential family of distributions. 1 Introduction Belief networks provide a well-understood graphical framework for expressing the interactions between random variables. Such networks have proven useful ..