Abstract:
|
In the multivariate Normal case, a graph constrains the covariance matrix of (large-dimensional) data by imposing a pattern of zeroes in its inverse, and its decomposability ensures comparison between two nested models can be localized. Heteroscedascity can then be thought of as resulting from a slowly varying sequence of underlying graphs. For $B-1$ changepoints, the state space consists of: (i) inter-arrival times $N = (N_1, \cdots, N_B); $ (ii) a Markov chain of decomposable graphs $G = (G_1, \cdots, G_B), $ with one-step transition probabilities $P(h,g) $ given by $1/d(h), $ where $d(h)$ is the number of decomposable graphs $g$ such that at most one change (dropping or adding one edge) separates $g$ from $h;$ (iii) and an unconstrained covariance matrix $\Sigma^*, $ which, for each $b,$ is the completion of the constrained covariance matrix $\Sigma_b$ relative to $G_b.$ A carefully designed proposal strategy takes advantage of the form of the transition matrix $\{P(h,g)\}, $ and allows for a relatively expeditious exploration of the state space via a Metropolis-Hastings chain, facilitating posterior estimation of $\Sigma^*$ and prediction of $G.$
|