Lomb-Scargle Period Analysis

• Lomb-Scargle period analysis is a method that generalizes Fourier techniques to detect periodic signals from unevenly sampled data.
• Bayesian extensions like BGLST incorporate linear trends and heteroscedastic noise, enabling robust estimation even in the presence of offsets.
• Efficient computation is achieved through analytical marginalization of nuisance parameters, supporting rigorous model comparison and multi-harmonic extensions.

The Lomb-Scargle periodogram is a foundational statistical method for period analysis in unevenly sampled time series, particularly within astronomical research. It generalizes the classical Fourier power spectrum to accommodate arbitrary sampling and heteroscedastic errors, allowing for rigorous detection and characterization of periodic signals, trends, and offsets. Extensions such as the Generalised Lomb-Scargle (GLS), Bayesian Generalised Lomb-Scargle (BGLS), and the Bayesian Generalised Lomb-Scargle Periodogram with Trend (BGLST) further enhance robustness to long-term drifts and non-uniform noise, making these techniques optimal for recovering cycle periods that may span the dataset duration (Olspert et al., 2017).

1. Mathematical Formalism and Probabilistic Foundation

BGLST extends the classical Lomb-Scargle methodology by embedding the period-finding task in a linear probabilistic regression framework. The generative model is:

$$y(t_i)\;=\;A\cos\bigl(2\pi f\,t_i\bigr)\;+\;B\sin\bigl(2\pi f\,t_i\bigr) \;+\;\alpha\,t_i\;+\;\beta \;+\;\epsilon_i$$

where $f$ is frequency, $A, B$ are amplitudes, $\alpha$ is the linear slope, $\beta$ the intercept, and $\epsilon_i \sim \mathcal{N}(0, \sigma_i^2)$ is Gaussian noise (possibly heteroscedastic). Zero-mean Gaussian priors are placed on the coefficients, while the frequency parameter is assigned a uniform prior over the scan range. The marginal posterior for $f$, the BGLST spectrum, is obtained by analytically integrating over nuisance parameters:

$$p(f|D) \propto \frac{1}{\sqrt{\det\Lambda(f)}} \exp\bigl[-\tfrac12 E_\mathrm{min}(f)\bigr]$$

where $\Lambda(f)$ is the posterior precision matrix of $(A, B, \alpha, \beta)$, and $E_\mathrm{min}(f)$ is the penalized sum-of-squares minimized with respect to those parameters (Olspert et al., 2017).

2. Computational Implementation and Algorithmic Steps

For each trial frequency, the BGLST algorithm proceeds as follows:

• Precompute weighted sums on the observations: $\sum w_i\cos^2(2\pi f t_i)$, $\sum w_i\sin^2(2\pi f t_i)$, and other mixed moments, with $w_i=1/\sigma_i^2$; these include cross-terms with time, necessary for capturing linear trends.
• Construct and invert the small ($4\times4$) posterior precision matrix $\Lambda(f)$, compute its determinant, and analytically eliminate linear coefficients to calculate $E_\mathrm{min}(f)$.
• Evaluate and store the (log-)posterior for each frequency.
• Identify frequencies with significant peaks in $p(f|D)$. Closed-form posteriors for amplitudes, slopes, offsets at the maximum likelihood frequency can be recovered.
• Extend to multi-harmonic models if required, albeit with increased complexity and parameters.

This pipeline preserves the computational scaling of GLS (O($N N_f$)), requiring only O($N$) operations per frequency evaluation but solving a constant-size linear algebraic system at each step (Olspert et al., 2017).

3. Treatment of Trends, Offsets, and Heteroscedastic Noise

Whereas classical LS forces zero mean and ignores long-term drifts, BGLST:

• Incorporates an explicit time-dependent trend ($\alpha t_i$) directly into the regression; this is essential when periods being searched for are comparable to the dataset span.
• Handles unknown or dynamic DC offsets by allowing a floating intercept ($\beta$).
• Enables seamless modeling of both constant and heteroscedastic noise by setting $w_i = 1/\sigma_i^2$. If individual error variances are unavailable, BGLST supports empirical estimates via sliding windows for sufficiently dense data.
• Avoids erroneous “detrending” steps that can destroy or bias the periodic signal capture, especially problematic near the trend/cycle ambiguity boundary (Olspert et al., 2017).

4. Comparative Performance: Synthetic and Benchmark Data

Extensive validation using synthetic datasets and real-star examples demonstrates (Olspert et al., 2017):

• In the absence of true trend, BGLST and GLS yield indistinguishable results under uniform or moderately gapped sampling (at large sample size).
• For light curves with unknown DC offsets, classical LS is severely biased, while GLS and BGLST remain unbiased.
• In tests with strong trends: both “detrend+GLS” and GLS alone degrade rapidly as trend strength increases, while BGLST maintains accurate period estimates; pre-detrending can even obliterate detectable periodic signals.
• For heteroscedastic noise ratios $\max(\sigma_i)/\min(\sigma_i) \gtrsim 2$, explicit modeling of noise variance by BGLST substantially improves period recovery compared to unweighted LS/GLS; benefits are observed for ratios above 1.5, even when variances are only empirically estimated.

5. Practical Guidelines, Diagnostics, and Model Selection

BGLST provides a seamless Bayesian upgrade to classical periodograms:

• Gaussian priors on coefficients regularize against pathological fits (e.g. sharp slopes at low frequency).
• Uniform frequency prior ensures direct comparability to classical spectrum; posterior normalization is explicit, allowing relative probability and Bayes factor discrimination between candidate periods.
• Trend vs. cycle ambiguity: if cycle period exceeds data span, BGLST trades model power consistently between trend slope and sinusoidal amplitude, avoiding arbitrary cutoff effects of third-party detrending.
• Model comparison against null (trend-only) fit is feasible via Bayesian Information Criteria or Bayes factors.
• For multi-component signals, iterative “cleaning” or pre-whitening can be used, but full Bayesian multi-harmonic fits are feasible in principle.

Synthetic and real data confirm that in cases with unknown offsets, trends, or time-varying errors (particularly at low cycle counts), BGLST outperforms classical methods, especially compared to ad hoc “detrend then periodogram” workflows (Olspert et al., 2017).

6. Extensions and Limitations

BGLST is extensible to:

• Multi-harmonic (sum-of-sinusoid) models, but incurs increased complexity.
• Arbitrary error distributions if required (non-Gaussian, correlated), with adapted marginalization strategies.
• Time-varying activity cycles, provided sufficient data density for local variance estimation.

Limitations include reliance on linearity in all coefficients except frequency, and increased complexity for highly multi-component signals or if extensive non-linearities are present (Olspert et al., 2017).

7. Context and Impact on Astronomical Time Series Analysis

BGLST has established itself as the optimal procedure for period estimation when unknown long-term trends and noise heterogeneity are possible, particularly in the domain of stellar magnetic cycle studies, where the sought periods approach dataset spans and where reliable separation of cosmic and instrumental trends is fundamentally ambiguous. The method's rigorous treatment of trends and probabilistic model comparison is critical for robust astrophysical inference, especially in low-SNR regimes and uneven seasonal sampling (Olspert et al., 2017).

In summary, the Bayesian Generalised Lomb-Scargle Periodogram with Trend is the state-of-the-art technique for extracting periodicities from time series contaminated by trends, offsets, and non-uniform noise, routinely outperforming traditional detrending or naive periodogram approaches across diverse astronomical contexts.