Statistical properties of Fourier-based time-lag estimates

TL;DR

This paper analyzes the statistical properties of Fourier-based time-lag estimates in X-ray astronomy, providing guidelines to minimize bias and accurately assess errors in AGN studies.

Contribution

It offers a comprehensive investigation combining analytical and numerical methods to understand biases and errors in Fourier-based time-lag estimates for AGN data.

Findings

Sampling and finite duration bias the estimates downward.

Frequency smoothing can introduce low-frequency bias.

Low signal-to-noise ratios reduce coherence and bias estimates.

Abstract

The study of X-ray time-lag spectra in active galactic nuclei (AGN) is currently an active research area, since it has the potential to illuminate the physics and geometry of the innermost region (i.e. close to the putative super-massive black hole) in these objects. To obtain reliable information from these studies, the statistical properties of time-lags estimated from data must be known as accurately as possible. Aims: We investigated the statistical properties of Fourier-based time-lag estimates (i.e. based on the cross-periodogram), using evenly sampled time series with no missing points. Our aim is to provide practical `guidelines' on estimating time-lags that are minimally biased (i.e. whose mean is close to their intrinsic value) and have known errors.} Methods: Our investigation is based on both analytical work and extensive numerical simulations. The latter consisted of generating artificial time series with various signal-to-noise ratios and sampling patterns/durations similar to those offered by AGN observations with present and past X-ray satellites. We also considered a range of different model time-lag spectra commonly assumed in X-ray analyses of compact accreting systems. Results: Discrete sampling, binning and finite light curve duration cause the mean of the time-lag estimates to have a smaller magnitude than their intrinsic values. Smoothing (i.e. binning over consecutive frequencies) of the cross-periodogram can add extra bias at low frequencies. The use of light curves with low signal-to-noise ratio reduces the intrinsic coherence, and can introduce a bias to the sample coherence, time-lag estimates, and their predicted error.