A Bayesian test for periodic signals in red noise

Many astrophysical sources, especially compact accreting sources, show strong, random brightness fluctuations with broad power spectra in addition to periodic or quasi-periodic oscillations (QPOs) that have narrower spectra. The random nature of the dominant source of variance greatly complicates the process of searching for possible weak periodic signals. We have addressed this problem using the tools of Bayesian statistics; in particular using Markov chain Monte Carlo techniques to approximate the posterior distribution of model parameters, and posterior predictive model checking to assess model fits and search for periodogram outliers that may represent periodic signals. The methods developed are applied to two example datasets, both long XMM-Newton observations of highly variable Seyfert 1 galaxies: RE J1034+396 and Mrk 766. In both cases a bend (or break) in the power spectrum is evident. In the case of RE J1034+396 the previously reported QPO is found but with somewhat weaker statistical significance than reported in previous analyses. The difference is due partly to the improved continuum modelling, better treatment of nuisance parameters, and partly to different data selection methods.

💡 Research Summary

The paper tackles the long‑standing problem of detecting weak periodic or quasi‑periodic oscillations (QPOs) in astrophysical time series that are dominated by red‑noise variability. Traditional Fourier‑based techniques assume a simple white‑noise background and therefore produce inflated false‑alarm rates when applied to data with a broad, power‑law‑like power spectrum. The authors adopt a fully Bayesian framework that simultaneously models the stochastic continuum and searches for outlying power‑spectral features that could indicate a periodic component.

The continuum is described by a “bent power‑law” (also called a broken or bending power spectrum) characterized by four parameters: a normalization A, a bend frequency f_bend where the slope changes, a low‑frequency slope α_low, and a high‑frequency slope α_high. Non‑informative priors are assigned to each parameter, and the likelihood is constructed from the logarithm of the periodogram assuming a χ² distribution with two degrees of freedom (the standard Whittle likelihood). Because the posterior distribution is analytically intractable, the authors employ Markov chain Monte Carlo (MCMC) sampling, combining Metropolis‑Hastings proposals with Gibbs updates to efficiently explore the parameter space. Convergence diagnostics (Gelman‑Rubin statistics, trace plots) confirm that the chains have mixed well, allowing reliable estimates of posterior means, credible intervals, and the full covariance structure among the parameters.

A key innovation is the use of posterior predictive checks. From the posterior sample the authors generate thousands of synthetic periodograms, each drawn from the fitted bent‑power‑law model. For every frequency bin they compute the 95 % predictive interval of the simulated powers. The observed periodogram is then compared to these intervals; any bin whose observed power lies outside the interval is flagged as an outlier. Because the predictive distribution already incorporates uncertainty in the continuum parameters, such outliers are statistically robust candidates for additional narrow‑band power, i.e., a QPO.

The methodology is applied to two long XMM‑Newton observations of highly variable Seyfert 1 galaxies: RE J1034+396 and Mrk 766. For RE J1034+396 the analysis reproduces the previously reported QPO near 2.7 mHz, but the Bayesian significance is lower (posterior predictive p‑value ≈ 0.02) than the frequentist values quoted in earlier work (≈ 10⁻³). The reduction in significance is attributed to (i) a more realistic bent‑power‑law description of the red‑noise continuum, (ii) proper propagation of nuisance‑parameter uncertainties through the posterior, and (iii) a different data‑selection strategy that avoids over‑optimistic trimming of high‑variability intervals. The posterior for the bend frequency is tightly constrained around 0.5 mHz, and the high‑frequency slope is steeper than earlier estimates, which together diminish the apparent excess power at the QPO frequency.

For Mrk 766 no statistically significant outliers are found; the observed periodogram lies comfortably within the 95 % predictive bands across the entire frequency range. The fitted continuum parameters (f_bend ≈ 0.8 mHz, α_low ≈ 0.9, α_high ≈ 2.2) fully account for the data, suggesting that any putative QPOs are either absent or below the detection threshold given the current exposure.

The authors also explore the impact of data selection. By comparing results obtained from the full observation with those derived from subsets selected for maximal variability, they demonstrate that aggressive trimming can artificially inflate the apparent significance of narrow‑band features. The Bayesian framework naturally penalizes such practices because the predictive checks remain calibrated to the full posterior uncertainty.

In the discussion, the paper emphasizes several advantages of the Bayesian approach: (1) simultaneous estimation of continuum and QPO parameters with full uncertainty quantification, (2) avoidance of ad‑hoc multiple‑trial corrections because the posterior predictive distribution already accounts for the number of frequency bins examined, and (3) a principled way to compare competing continuum models (e.g., simple power law vs. bent power law) via Bayes factors or posterior predictive p‑values. The authors propose extensions to include multiple QPO components, non‑stationary red‑noise models, and joint analyses of multi‑instrument datasets (e.g., combining XMM‑Newton and NuSTAR).

Overall, the study provides a rigorous statistical toolkit for the astrophysical community, enabling more reliable detection of weak periodic signals in the presence of complex red‑noise backgrounds and setting a new standard for time‑series analysis in high‑energy astrophysics.