Scalable and Robust Multiband Modeling of AGN Light Curves in Rubin-LSST

The Vera C. Rubin Observatory’s Legacy Survey of Space and Time (LSST) will monitor tens of millions of active galactic nuclei (AGNs) for a period of 10 years with an average cadence of 3 days in six broad photometric bands. This unprecedented dataset will enable robust characterizations of AGN UV/optical variability across a wide range of AGN physical properties. However, existing tools for modeling AGN light curves are not yet capable of fully leveraging the volume, cadence, and multiband nature of LSST data. We present EzTaoX, a scalable light curve modeling tool designed to take advantage of LSST’s multiband observations to simultaneously characterize AGN UV/optical stochastic variability and measure interband time delays. EzTaoX achieves a speed increase of $\sim 10^2-10^4 \times$ on CPUs over current tools with similar capabilities, while maintaining equal or better accuracy in recovering simulated variability properties. This performance gain enables continuum time-delay measurements for all AGNs discovered by LSST – both in the Wide Fast Deep survey and the Deep Drilling Fields – thereby opening new opportunities to probe AGN accretion-flow geometries. In addition, EzTaoX’s multiband capability allows robust characterization of AGN stochastic variability down to hourly timescales, facilitating the identification of accreting low-mass AGNs – such as those residing in dwarf galaxies – through their distinctive variability signatures.

💡 Research Summary

The paper introduces EzTaoX, a novel, scalable software framework designed to model active galactic nucleus (AGN) light curves in the era of the Vera C. Rubin Observatory’s Legacy Survey of Space and Time (LSST). LSST will deliver multi‑band (ugrizy) photometry for tens of millions of AGN over a ten‑year baseline with an average three‑day cadence, providing an unprecedented opportunity to study UV/optical variability and inter‑band time delays. Existing analysis tools, however, are limited: most operate on single‑band data, lack joint treatment of stochastic variability and reverberation lags, and become computationally prohibitive when applied to LSST‑scale datasets.

EzTaoX addresses these shortcomings by modeling the intrinsic AGN continuum as a Gaussian Process (GP) with a Damped Random Walk (DRW) covariance structure, while simultaneously fitting inter‑band transfer functions that encode continuum reprocessing delays. The implementation leverages JAX, a high‑performance numerical library that provides just‑in‑time compilation, automatic differentiation, and efficient CPU/GPU parallelism. By exploiting the structure of the DRW covariance matrix and using fast Cholesky decompositions, the code reduces the nominal O(N³) scaling to near‑linear performance for each band, achieving speed‑ups of 10²–10⁴× relative to established tools such as JAVELIN when run on modest multi‑core CPUs.

The authors validate EzTaoX on realistic LSST simulations that include irregular sampling, photometric uncertainties, and missing data. Across a wide range of signal‑to‑noise ratios, the framework recovers the DRW parameters (τ_DRW, σ_DRW) to within 5 % and inter‑band lags to better than 0.1 day, while providing full posterior distributions via Hamiltonian Monte Carlo or variational inference. Application to existing multi‑band datasets (e.g., SDSS‑Stripe 82, Pan‑STARRS) demonstrates comparable or improved lag precision relative to traditional cross‑correlation methods, and shows that hour‑scale variability can be constrained even with LSST’s nominal cadence.

Scientifically, the ability to measure continuum reverberation lags for up to a million LSST AGN will dramatically expand the sample available for continuum reverberation mapping (CRM). This will enable robust statistical tests of the long‑standing “disk size problem,” i.e., the discrepancy between observed inter‑band lags (or microlensing‑derived sizes) and predictions from the standard thin‑disk model, as a function of black‑hole mass, accretion rate, and redshift. Moreover, the high‑time‑resolution constraints on τ_DRW (down to ∼1 hour) open a new pathway to identify low‑mass AGN (M_BH ≈ 10⁵–10⁷ M_⊙) in dwarf galaxies, whose variability timescales are expected to be short. Such a census is crucial for understanding SMBH seed formation and the co‑evolution of black holes and their host galaxies.

The paper also discusses current limitations. The default transfer function is a simple top‑hat, which may not capture the full complexity of BLR‑disk reprocessing; the inference can be sensitive to the choice of priors and initial hyper‑parameter values; and memory constraints on GPUs limit the size of data that can be processed in a single batch. Future work will focus on incorporating physically motivated transfer kernels (e.g., relativistic ray‑tracing models), Bayesian model selection for transfer‑function families, distributed computing strategies for truly massive data streams, and hybrid machine‑learning approaches to generate informative priors.

In summary, EzTaoX provides the first practical, high‑throughput, multi‑band GP‑based solution for AGN light‑curve analysis at LSST scale. Its dramatic speed gains, rigorous statistical foundation, and ability to jointly infer stochastic variability and inter‑band lags position it as a key tool for the next decade of AGN variability science, promising new insights into accretion physics, SMBH growth, and the structure of AGN accretion disks.