Understanding Conditional Expectation, Variance, and Precision Matrices

  • #1
MAXIM LI
6
2
Homework Statement
$$E(u_{i,j}\vert u_{-i,j}) = \frac{1}{a}(u_{i-1,j}+u_{i+1,j}+u_{i,j-1}+u_{i,j+1})$$
Relevant Equations
$$E(u_{i,j}\vert u_{-i,j}) = \frac{1}{a}(u_{i-1,j}+u_{i+1,j}+u_{i,j-1}+u_{i,j+1})$$
My question relates to subsection 2.2.1 of [this article][1]. This subsection recalls the work of Lindgren, Rue, and Lindström (2011) on Gaussian Markov Random Fields (GMRFs). The subsection starts with a two-dimensional regular lattice where the 4 first-order neighbours of $u_{i,j}$ are identified. The article defines the full conditional distribution through the expectation $$E(u_{i,j}\vert u_{-i,j}) = \frac{1}{a}(u_{i-1,j}+u_{i+1,j}+u_{i,j-1}+u_{i,j+1})$$ and variance $$Var(u_{i,j}\vert u_{-i,j}) = \frac{1}{a}.$$
This is then redefined in terms of the precision matrix, where the upper right quadrant is
\begin{array}
-1 \\
a & -1
\end{array}

Extending to second-order neighbours (i.e. the neighbours of first-order neighbours), the precision matrix becomes (again, just the upper right quadrant)
\begin{array}
-1 \\
-2a & 2 \\
4+a^2 & -2a & 1.
\end{array}

I am new to this topic and am trying to understand where the expressions for the conditional expectation and variance came from, and how the precision matrices were derived. I'd appreciate a fulsome explanation and derivation for both the first-order and second-order case. I tried looking in the book 'Gaussian Markov Random Fields
Theory and Applications' and this looks very similar to a conditional autoregression (CAR) model, defined in Chapter 1. However, here the full conditionals are written down as

$$
x_i \vert \mathbf{x}_{-i} \sim N\left(\sum_{j\neq i}\beta_{ij}x_{j},\kappa_i^{-1} \right)
$$

and the elements of the corresponding precision matrix are stated to be ##Q_{ii} = \kappa_i## and ##Q_{ij} = -\kappa_{i}\beta_{ij}## for ##i\neq j##This seems to be more general, which leaves me wondering how the conditional mean and variance at the start of this post were derived (along with the precision matrices). Where did a come from and why did we scale by this amount? Any help addressing this is much appreciated.

Note that ##\mathbf{x}_{-i}## means the vector of random variables excluding ##x_i##.

[1]: https://becarioprecario.bitbucket.io/spde-gitbook/ch-intro.html#sec:spde
 

FAQ: Understanding Conditional Expectation, Variance, and Precision Matrices

What is conditional expectation?

Conditional expectation is the expected value of a random variable given that another random variable or set of variables is known. It provides a way to update our predictions or expectations based on new information. Mathematically, if \(X\) and \(Y\) are random variables, the conditional expectation of \(X\) given \(Y\) is denoted as \(E(X|Y)\).

How is conditional variance defined?

Conditional variance measures the variability of a random variable given that another variable is known. It is defined as the variance of the random variable around its conditional expectation. Formally, if \(X\) and \(Y\) are random variables, the conditional variance of \(X\) given \(Y\) is \(Var(X|Y) = E[(X - E(X|Y))^2|Y]\).

What is a precision matrix?

A precision matrix, also known as the inverse covariance matrix, is the inverse of the covariance matrix of a multivariate random variable. It provides information about the partial correlations between variables, indicating the strength and direction of the linear relationship between pairs of variables, controlling for the effect of other variables.

How do you compute the conditional expectation in a multivariate normal distribution?

In a multivariate normal distribution, the conditional expectation can be computed using the mean and covariance matrix of the distribution. If \(X\) is partitioned into \(X_1\) and \(X_2\), the conditional expectation of \(X_1\) given \(X_2 = x_2\) is \(E(X_1|X_2 = x_2) = \mu_1 + \Sigma_{12}\Sigma_{22}^{-1}(x_2 - \mu_2)\), where \(\mu\) is the mean vector and \(\Sigma\) is the covariance matrix partitioned conformably.

Why is the precision matrix important in statistical modeling?

The precision matrix is important because it provides insights into the conditional independence structure of the variables. In graphical models, for instance, the precision matrix helps identify which variables are conditionally independent given others, simplifying the model and making it more interpretable. Additionally, it is useful in regularization techniques like Lasso and Ridge regression to improve model stability and performance.

Back
Top