Rylan Schaeffer

Kernel Papers

Undirected Graphical Models

An undirected graphical model is a different way of representing probabilistic relationships on a graph. An undirected graph \(G=(V, E)\) also represents variables using vertices. Rather than starting with factorization, it is easier to start with separation.

Undirected Separation: For a given undirected graph \(G+(V, E)\), \(A, B \subset V\) are separated w.r.t. \(C \subset V\) if every path between any \(a \in A, b \in B\) passes through some \(c \in C\).

Global Markov Property: A distribution \(p\) satisfies this property if it satisfies all conditional independence statements \(x_A \perp x_B \lvert x_C\) for all disjoint sets \(A, B, C \in V\) such that \(A, B\) are separated w.r.t. \(C\).

Pairwise Markov Property: \(\forall (i, j) \not\in E, \, x_i \perp x_j \lvert x_{V \setminus \{i, j\}}\)

Factorization: Let \(cl^*(G)\) be the set of maximal cliques on \(G\). Then the Boltzmann (or Gibbs) distribution on \(G\) is defined as

\[p(x) = \frac{1}{Z} \exp(- \sum_{c \in cl^*(G)} \Psi (x_c))\]

where \(x_c\) are the nodes in the maximal clique and \(\Psi: X^{\lvertc\lvert} \rightarrow \mathbb{R}_{\geq 0}\) are non-negative functions called potentials. Cliques can be intuited from a physics perspective. If undirected edges describe interactions between objects, then a maximial clique is a group of particles that all interact. That interaction has some energy, and the sum of all the interactions has a total energy.

Claim: Factorization \(\Rightarrow\) Undirected Global Markov Property.

Hammersley-Clifford Theorem: If \(p(x) > 0\), then factorization w.r.t. undirected \(G\) \(\Leftrightarrow\) undirected global Markov property \(\Leftrightarrow\) undirected local Markov property.

Gaussian Undirected Graphs

As the name implies, all variables have Normal distributions. Properties:

  1. Potentials can have at most pairwise interactions between variables because \(J\) creates pairwise terms and \(h\) creates single terms, but there are no other higher interaction terms.

  2. \[x_i \perp x_j \Leftrightarrow \Sigma_{ij} = 0\]
  3. \(x_i \perp x_j | \text{everything else} \Leftrightarrow J_{ij} = 0\). This only holds when conditioning on everything else. Consequently, to test \(x_i \perp x_j | x_k\), first marginalize till only \(x_i, x_j, x_k\) remain.