# The First Glimpse into Pseudolikelihood

##### Posted on Feb 12, 2019

This post caught a glimpse of the pseudolikelihood.

As Wikipedia said, a pseudolikelihood is an approximation to the joint probability distribution of a collection of random variables.

Given a set of random variables, $X=X_1,\ldots,X_n$ and a set $E$ of dependencies between these random variables, where $\{X_i,X_j\}\not\in E$ implies $X_i$ is conditionally independent of $X_j$ given $X_i$’s neighbors, the pseudolikelihood of $X=x=(x_1,x_2,\ldots,x_n)$ is

In a word, the pseudolikelihood is the product of the variables, with each probability conditional on the rest of the data.

Murphy (2012) provides a motivation $\eqref{eq:pseudoll}$ base on MRFs.

Consider an MRF in log-linear form:

where $c$ indexes the cliques.

The objective for maximum likelihood is

The derivative for the weights of a particular clique, $c$, is given by

Let the gradient be zero, we will get moment matching:

One alternative to MLE is to maximize the pseudo likelihood, defined as follows:

which is also called the composite likelihood.

The PL approach can be illustrated by the following figure for a 2d grid. This objective is generally fast to compute since each full conditional $p(y_{id}\mid \y_{i,-d},\btheta)$ only requires summing over the states of a single node, $y_{id}$, in order to compute the local normalizing constant.

By the way, there are many methods to compute the MLE/MAP, as show in the following table.

Published in categories Note