Automated period detection from variable stars time series database

The exact period determination of a multi-periodic variable star based on its luminosity time series data is believed a task requiring skill and experience. Thus the majority of available time series analysis techniques require human intervention to some extent. The present work is dedicated to establish an automated method of period (or frequency) determination from the time series database of variable stars. Relying on the SigSpec method (Reegen 2007), the technique established here employs a statistically unbiased treatment of frequency-domain noise and avoids spurious (i. e. noise induced) and alias peaks to the highest possible extent. Several add-on’s were incorporated to tailor SigSpec to our requirements. We present tests on 386 stars taken from ASAS2 project database. From the output file produced by SigSpec, the frequency with maximum spectral significance is chosen as the genuine frequency. Out of 386 variable stars available in the ASAS2 database, our results contain 243 periods recovered exactly and also 88 half periods, 42 different periods etc. SigSpec has the potential to be effectively used for fully automated period detection from variable stars’ time series database. The exact detection of periods helps us to identify the type of variability and classify the variable stars, which provides a crucial information on the physical processes effective in stellar atmospheres.

💡 Research Summary

The paper presents an automated pipeline for detecting periodicities in variable‑star light curves by leveraging the SigSpec algorithm (Reegen 2007). Traditional period‑search techniques such as discrete Fourier transforms, Lomb‑Scargle periodograms, and Phase Dispersion Minimisation often require human supervision to set thresholds, resolve aliases, or interpret ambiguous peaks, which limits their scalability to the massive time‑series databases produced by modern surveys (ASAS, MACHO, OGLE, LSST, Gaia, etc.).

SigSpec distinguishes itself by analytically deriving the probability density function of white noise in the frequency domain, assigning each frequency a “spectral significance” (sig) and a cumulative significance (csig). These quantities provide a statistically unbiased measure of how unlikely a peak is to arise from noise, independent of the amplitude spectrum’s shape. The authors adopt the default SigSpec code but modify four key parameters in the runtime configuration file (SigSpec.ini):

1. lfreq – lower frequency limit set to 1/T, where T is the total time span of the observations, ensuring the search does not extend below the fundamental resolution.
2. nycoef – the Nyquist coefficient, set to 0.9, to accommodate the irregular sampling typical of ground‑based surveys; each interval Δt_k receives its own Nyquist frequency (2 Δt_k)⁻¹, and the global upper limit f_u is defined accordingly.
3. iterratio – the number of iterations is fixed at five rather than allowing the algorithm to stop when sig drops below 5, which balances computational load with the need to extract several dominant frequencies.
4. siglimit – lowered from the default value of 5 to 2 to enable detection of weaker signals in noisy data sets.

Two additional “add‑ons” are implemented to improve the robustness of the automated search. First, after each run the five frequencies with the highest cumulative sig are examined for integer‑multiple relationships; if a harmonic shows a larger sig than the fundamental, the algorithm promotes the fundamental to the top of the list. This corrects the common situation where a non‑sinusoidal light curve yields a stronger harmonic peak. Second, a filter excludes frequencies in the narrow band 0.99–1.01 d⁻¹, which correspond to daily aliases inherent to single‑site ground‑based observations.

The method is tested on the ASAS‑2 database, which contains light curves for 386 variable stars. After discarding one corrupted file and one with only two measurements, the remaining 384 series are processed automatically on a 2.8 GHz Linux workstation; the entire batch completes in 3 h 48 min. The results are categorized as follows:

• Exact matches: 243 stars (≈64 %) where the SigSpec‑derived period coincides with the published ASAS period.
• Integer or half‑integer multiples: 95 stars (≈25 %). In 88 cases the detected frequency is twice the published value, in 4 cases four times, and in 3 cases half the published frequency. These discrepancies are interpreted as manifestations of non‑sinusoidal waveforms where the fundamental frequency is accompanied by strong harmonics.
• Different periods: 42 stars (≈11 %) where no simple multiple relationship can be established; these may represent mis‑classifications, data quality issues, or genuinely new periodicities.
• Newly identified variables: 4 stars not previously listed in the ASAS catalog.

The authors argue that SigSpec’s statistically grounded significance measures allow a fully automated selection of the “true” period without subjective visual inspection, a major advantage over tools such as PERIOD04, PERANSO, or MUFRAN that still rely on human judgment for final validation. The processing speed and the ability to handle irregular sampling make the approach suitable for upcoming petabyte‑scale surveys like LSST and Gaia.

In conclusion, the study demonstrates that an automated pipeline built around SigSpec can reliably recover known periods, correctly identify harmonic relationships, and uncover previously unreported variability. Future work will extend the method to multi‑periodic stars, improve handling of sparse or highly noisy data, and integrate the pipeline into larger survey data‑reduction frameworks.