Linear Generative Models for Spatial and Spatio-Temporal Data

36-467/36-667

18 October 2018

\[ \newcommand{\Expect}[1]{\mathbb{E}\left[ #1 \right]} \newcommand{\Var}[1]{\mathrm{Var}\left[ #1 \right]} \newcommand{\Cov}[1]{\mathrm{Cov}\left[ #1 \right]} \newcommand{\Neighbors}{N} \]

In our last episode

Today

Estimating AR(1)

\[ X(t) = a+b X(t-1) + \epsilon(t) \]

Estimating \(b\): Yule-Walker approach

Estimating \(b\): Ordinary least squares approach

Estimating \(b\): Maximum likelihood approach

Estimating AR(1)

In space, no one can decide which way to go

Conditional autoregressive (CAR) spatial models

The Gaussian CAR model

The Gaussian CAR model

Some points about CAR models

The “Gibbs sampler” trick

Simultaneous autoregressions (SAR):

Simultaneous autoregressions (SAR):

Spatio-temporal

Summary

Backup: Deriving the global distribution of a linear-Gaussian CAR

(after Guttorp (1995, 206–7))

\[\begin{eqnarray} \log{\frac{p(\mathbf{X}=\mathbf{x})}{p(\mathbf{X}=\mathbf{0})}} & = & \sum_{i=1}^{n}{\log{p(X_i=x_i|X_{-i}=x_{-i;i+1})} - \log{p(X_i=0|X_{-i}=x_{-i;i+1})}} ~ \text{(Brooks representation, see backup)}\\ & = & -\frac{1}{2\tau^2}\left(\sum_{i=1}^{n}{\left(x_i - \sum_{j=1}^{i-1}{b_{ij} x_j}\right)^2 - \left(0 - \sum_{j=1}^{i-1}{b_{ij} x_j}\right)^2}\right)\\ & = & -\frac{1}{2\tau^2}\left(\sum_{i=1}^{n}{x_i^2 - 2 x_i \sum_{j=1}^{i-1}{b_{ij} x_j}}\right)\\ & = & \frac{1}{2\tau^2}\left(\sum_{i=1}^{n}{x_i^2 - x_i (\mathbf{b} \mathbf{x})_{i}}\right) ~ \text{(symmetry of}~\mathbf{b})\\ & = & \frac{1}{2\tau^2}\mathbf{x}^T(\mathbf{I}-\mathbf{b})\mathbf{x}\\ \end{eqnarray}\]

Backup: Deriving the global distribution of a linear-Gaussian SAR

(after Guttorp (1995, 219–21, Exercise 4.1))

Backup: Converting an SAR to a CAR

Backup: The “Brooks representation” result

After Guttorp (1995, 7, Proposition 1.1)

\(\mathbf{X}\) is an \(n\)-dimensional random vector. Assume \(p(\mathbf{x}) > 0\) for all \(\mathbf{x}\). Then \[ \frac{p(\mathbf{x})}{p(\mathbf{y})} = \prod_{i=1}^{n}{\frac{p(x_i|x_{1:i-1}, y_{i+1:n})}{p(y_i|x_{1:i-1}, y_{i+1:n})}} \] Since \(\sum_{\mathbf{x}}{p(\mathbf{x})}=1\), this implies that the conditional distributions uniquely fix the whole distribution

Backup: Proof of the Brooks representation result

Again, after Guttorp (1995, 7)

\[\begin{eqnarray} p(\mathbf{x}) & = & p(x_n|x_{1:n-1}) p(x_{1:n-1})\\ & = & p(x_n|x_{1:n-1})p(x_{1:n-1})\frac{p(y_n|x_{1:n-1})}{p(y_n|x_{1:n-1})}\\ & = & \frac{p(x_n|x_{1:n-1})}{p(y_n|x_{1:n-1})}p(x_{1:n-1}, y_n)\\ & = & \frac{p(x_n|x_{1:n-1})}{p(y_n|x_{1:n-1})} p(x_{n-1}| x_{1:n-2}, y_n) p(x_{1:n-2}, y_n)\\ & = & \frac{p(x_n|x_{1:n-1})}{p(y_n|x_{1:n-1})} \frac{p(x_{n-1}| x_{1:n-2}, y_n)}{p(y_{n-1}| x_{1:n-2}, y_n)}p(x_{1:n-2}, y_{n-1:n})\\ & \vdots &\\ & = & \frac{p(x_n|x_{1:n-1})}{p(y_n|x_{1:n-1})} \frac{p(x_{n-1}| x_{1:n-2}, y_n)}{p(y_{n-1}| x_{1:n-2}, y_n)} \ldots \frac{p(x_1|y_{2:n})}{p(y_1|y_{2:n})} \end{eqnarray}\]

References

Guttorp, Peter. 1995. Stochastic Modeling of Scientific Data. London: Chapman; Hall.