Optimal sampling strategies for tree-based
time series

R. Riedi
Rice University


In this paper we consider binary multiscale tree models of time
series. In this setting, the overall average of a time series is
represented by the tree root, the average over the first half of the
series by its left child node, and so on. The time series itself
is represented by the leaf nodes. We address the problem of
choosing a limited number of leaf nodes to provide the optimal
linear least-square estimator of the tree root. Such problems
arise in network traffic inference where the goal lies in
estimating the average traffic arrival rate based on a limited
number of traffic samples. The solution depends crucially on the
correlation structure in the time series.

Joint work with V. Ribeiro, R. Baraniuk