Spectrum estimation with uniformly and non-uniformly sampled data: some challenges and strategies

Radhendushka Srivastava

Abstract

The power spectral density of a continuous time wide sense stationary stochastic process is the Fourier transform of its autocovariance function. A common nonparametric estimator of this function is a smoothed version of the periodogram (the Fourier transform of the sampled autocovariance function), which is based on uniformly spaced samples of the continuous time process. It is well known that this estimator is inconsistent when the underlying power spectral density does not have support restrcited to an suitable interval. For this reason, alternative estimators based on irregularly spaced samples have been proposed. The most common of these estimators is based on sampling according to a Poisson process. Even though Poisson and other stochastic sampling schemes involve somewhat high-tech theory, these are generally more difficult to implement than regular sampling. In this talk, we shall show that one does not necessarily have to use stochastic sampling for consistent estimation of the power spectral density. For this purpose, we would use `shrinking asymptotics', where the rate of uniform sampling is allowed to go to infinity at a suitable rate as the sample size goes to infinity. We shall also look into the issues of optimal rate, comparison with stochastic sampling based estimators, extension to multivariate time series and limiting distributions. A possible criticism of `shrinking asymptotics' is that the inter-sample distance cannot go to zero in real life sampling systems. For the problem of estimation under the constraint of a minimum separation between successive samples, we investigate the possibility of consistent estimation through stochastic sampling, and come across some surprising results. These findings are expected to lead to a more complete understanding of the worth of stochastic sampling.