# Factor graph

Jump to: navigation, search
Not to be confused with Graph factorization.

In probability theory and its applications, a factor graph is a particular type of graphical model, with applications in Bayesian inference, that enables efficient computation of marginal distributions through the sum-product algorithm. One of the important success stories of factor graphs and the sum-product algorithm is the decoding of capacity-approaching error-correcting codes, such as LDPC and turbo codes.

A factor graph is an example of a hypergraph, in that an arrow (i.e., a factor node) can connect more than one (normal) node.

When there are no free variables, the factor graph of a function f is equivalent to the constraint graph of f, which is an instance to a constraint satisfaction problem.

## Definition

A factor graph is a bipartite graph representing the factorization of a function. Given a factorization of a function $g(X_1,X_2,\dots,X_n)$,

$g(X_1,X_2,\dots,X_n) = \prod_{j=1}^m f_j(S_j),$

where $S_j \subseteq \{X_1,X_2,\dots,X_n\}$, the corresponding factor graph $G=(X,F,E)$ consists of variable vertices $X=\{X_1,X_2,\dots,X_n\}$, factor vertices $F=\{f_1,f_2,\dots,f_m\}$, and edges $E$. The edges depend on the factorization as follows: there is an undirected edge between factor vertex $f_j$ and variable vertex $X_k$ when $X_k \in S_j$. The function is tacitly assumed to be real-valued: $g(X_1,X_2,\dots,X_n) \in \Bbb{R}$.

Factor graphs can be combined with message passing algorithms to efficiently compute certain characteristics of the function $g(X_1,X_2,\dots,X_n)$, such as the marginal distributions.

## Examples

An example factor graph

Consider a function that factorizes as follows:

$g(X_1,X_2,X_3) = f_1(X_1)f_2(X_1,X_2)f_3(X_1,X_2)f_4(X_2,X_3)$,

with a corresponding factor graph shown on the right. Observe that the factor graph has a cycle. If we merge $f_2(X_1,X_2)f_3(X_1,X_2)$ into a single factor, the resulting factor graph will be a tree. This is an important distinction, as message passing algorithms are usually exact for trees, but only approximate for graphs with cycles.

## Message passing on factor graphs

A popular message passing algorithm on factor graphs is the sum-product algorithm, which efficiently computes all the marginals of the individual variables of the function. In particular, the marginal of variable $X_k$ is defined as

$g_k(X_k) = \sum_{X_{\bar{k}}} g(X_1,X_2,\dots,X_n)$

where the notation $X_{\bar{k}}$ means that the summation goes over all the variables, except $X_k$. The messages of the sum-product algorithm are conceptually computed in the vertices and passed along the edges. A message from or to a variable vertex is always a function of that particular variable. For instance, when a variable is binary, the messages over the edges incident to the corresponding vertex can be represented as vectors of length 2: the first entry is the message evaluated in 0, the second entry is the message evaluated in 1. When a variable belongs to the field of real numbers, messages can be arbitrary functions, and special care needs to be taken in their representation.

In practice, the sum-product algorithm is used for statistical inference, whereby $g(X_1,X_2,\dots,X_n)$ is a joint distribution or a joint likelihood function, and the factorization depends on the conditional independencies among the variables.

The Hammersley–Clifford theorem shows that other probabilistic models such as Markov networks and Bayesian networks can be represented as factor graphs; the latter representation is frequently used when performing inference over such networks using belief propagation. On the other hand, Bayesian networks are more naturally suited for generative models, as they can directly represent the causalities of the model.