Reimer Kuehn (KCL)
Spectra of empirical auto-covariance matrices derived from time series
Abstract:
We compute spectral densities of empirical auto-covariance matrices of
second order stationary stochastic processes in the limit where the
number of time-lags included is large. Matrices of this type constitute
one way of randomizing Toeplitz matrices. While not Toeplitz themselves,
their averages are, and fluctuations about these averages decrease with
increasing sample size M. We look at a limit in which both the matrix
dimension N and the sample size M used to define empirical averages
diverge, with their ratio alpha=N/M kept fixed. One of our main results
is a remarkable scaling relation which relates spectral densities for
processes with correlations via a simple integral to the corresponding
spectral densities for processes without correlations (i.e. for
sequences of i.i.d variables), and we derive a closed form approximation
for the latter. Our results are well corroborated by simulations using
auto-regressive processes.