Detection of correlated components in multivariate Gaussian models

In this paper, the problem of detecting correlated components in a p-dimensional Gaussian vector is considered. In the setup considered, s unknown components are correlated with a known covariance structure. Hence, there are equation possible hypotheses for the unknown set of correlated components. Instead of taking a full-vector observation at each time index, in this paper we assume that the observer is capable of observing any subset of components in the vector. With this flexibility in taking observations, the observer is interested in finding the optimal sampling strategy to maximize the error exponent (per sample) of the multi-hypothesis testing problem. We show that, when the correlation of these s components is weak, it is optimal for the observer to take full-vector observations; when the correlation is strong, the strategy of taking full-vector observation is not optimal anymore, and the optimal sampling strategy increases the detection error exponent by 25% at least, compared with the full-vector observation strategy.