Optimizing Optical Searches for Supermassive Black Hole Binaries in AGN Light Curves: Fourier versus Bayesian Periodicity Detection

Technical Summary: Optimizing Optical Searches for Supermassive Black Hole Binaries in AGN Light Curves

Problem Statement
The detection of Supermassive Black Hole Binaries (SMBHBs) via electromagnetic (EM) observations relies on identifying periodic brightness variations in Active Galactic Nuclei (AGN) light curves. These variations are hypothesized to arise from relativistic Doppler boosting (RDB) and interactions between circumbinary and mini-disks. However, intrinsic AGN variability, modeled as a Damped Random Walk (DRW), creates a stochastic "red noise" background that mimics periodicity. This challenge is exacerbated by sparse sampling and short observational baselines in current and upcoming surveys (Catalina Real-Time Transient Survey [CRTS], Zwicky Transient Facility [ZTF], and the upcoming Legacy Survey of Space and Time [LSST]). The primary difficulty lies in distinguishing genuine binary signals from DRW-only fluctuations while minimizing False Positive Rates (FPR) without sacrificing True Positive Rates (TPR).

Methodology
The authors simulate a population of AGN light curves representing CRTS, ZTF, and LSST observational properties. The simulations combine a DRW process (to model stochastic variability) with an injected sinusoidal signal (to model SMBHB orbital motion). The study evaluates three distinct periodicity detection methods:

1. Nested Bayesian Sampler (NBS): Utilizes the Ultranest package to perform Bayesian model selection between a pure DRW model and a combined DRW+Sinusoid model. Model preference is determined using the Bayesian Information Criterion ($\Delta$BIC).
2. Generalized Lomb-Scargle Periodogram (GLSP): Implemented via PyAstronomy, this Fourier-based method detects periodicity in unevenly sampled data. Model selection relies on False Alarm Probability (FAP) calculations, including Gaussian ($FAP_{Gauss}$), global bootstrap ($FAP_{global}$), and local bootstrap ($FAP_{local}$) variations.
3. Weighted Wavelet Z-transform (WWZ): Implemented via the wwz package, this method uses localized wavelets to detect evolving periodic signals. Model selection utilizes various FAP definitions ($FAP_{local-local}$, $FAP_{local-avg}$, $FAP_{avg-local}$) based on bootstrap simulations against a DRW background.

To address the computational burden of LSST (projected $\sim10^7$ light curves), the authors propose a "database procedure" for rapid triage. Instead of generating unique bootstrap distributions for every candidate, they pre-compute periodogram distributions for two "representative" LSST survey windows and use these to estimate FAPs for all candidates.

Key Results

• Method Performance: The NBS ($\Delta$BIC) is identified as the most robust classifier, producing the most "complete" samples (high Genuine True Positive Rate, TPR(P)) with minimal false positives. Among Fourier methods, the GLSP with local bootstrap FAP ($FAP_{local}$) offers the best balance of completeness and low FPR, while $FAP_{Gauss}$ is computationally efficient but less accurate for red noise. The WWZ performs well for mid-to-long periods but struggles with short-period signals due to sparse sampling artifacts.
• Parameter Dependence: Detection success for all methods is positively correlated with injected sinusoidal amplitude ($A_{in}$) and negatively correlated with DRW variability amplitude ($\sigma_{in}$). All methods show reduced sensitivity as the injected period approaches the observational baseline ($T_{obs}$), though the NBS can detect signals with fewer than two full cycles.
• Joint Model Selection: Combining NBS and GLSP model selection significantly reduces FPRs (by more than an order of magnitude) with only a marginal impact on TPR(P). Adding WWZ selection further reduces FPR but at the cost of a significantly lower TPR(P).
• LSST Triage: The database-driven $FAP_{Dlocal}$ classifier enables rapid triage of LSST data. Using strict thresholds, this method achieves a TPR of $\sim40\%$ and an FPR of $\sim0.5\%$ with minimal computational cost (processing $\sim10^7$ light curves in $\sim10-30$ hours).
• False Positives: DRW-only light curves that pass all selection criteria tend to exhibit high-amplitude stochastic variations that mimic periodicity. These "convincing" false positives often output low periods and amplitudes, suggesting that combining NBS and GLSP effectively filters these cases.

Significance and Claims
The paper claims to provide a robust pipeline for the EM side of a joint EM+Gravitational Wave (GW) SMBHB detection effort. Its primary contributions are:

1. Validation of a Tiered Search Strategy: The authors recommend a workflow where large datasets are first triaged using efficient Fourier methods (GLSP) to filter the bulk of non-periodic sources, followed by rigorous Bayesian analysis (NBS) on the remaining candidates.
2. Scalability for LSST: The proposed database procedure for FAP calculation offers a computationally feasible path to analyze the massive volume of LSST data, enabling the identification of rare SMBHB candidates within a realistic timeframe.
3. Multi-messenger Context: The study positions these EM detection methods as complementary to Pulsar Timing Array (PTA) GW searches. By identifying EM candidates with high confidence, the pipeline facilitates targeted GW follow-up and joint parameter estimation, potentially constraining SMBHB chirp masses and source parameters more accurately than GW or EM methods alone.

The authors conclude that while no single method is perfect, the combination of NBS and GLSP model selection offers the optimal trade-off for identifying genuine SMBHB candidates in the era of large-scale photometric surveys.