Reducing the Dimensions of AGN Lightcurve Manifolds

TL;DR

The paper tackles the problem of unifying diverse AGN variability behavior across multiple wavelengths and surveys. It proposes an unsupervised approach that learns a low-dimensional manifold of multi-band light-curve morphology using GP-based unification and UMAP with DTW distances, applied to two distinct samples. The results show coherent organization of variability space, with CLAGN transitions and TDEs occupying distinct regions in Sample A, and smooth correlations with independent spectroscopic/host properties in Sample B, all without explicit physical modeling. This demonstrates a practical, assumption-light method to integrate time-domain data, reveal connections to physical diagnostics, and guide spectroscopic follow-up, with broad applicability to future surveys like LSST.

Abstract

The Active Galactic Nuclei (AGN) glossary is vast and complex. Depending on selection method, observing wavelength, and brightness, AGNs are assigned distinct labels, yet the relationship between different selection methods and the diversity of time-domain behavior within and across classes remains difficult to characterize in a unified framework. Changing-look AGNs (CLAGNs), which transition between classifications over time, further complicate this picture. In this work, we learn a data-driven, low-dimensional representation of multi-wavelength photometric light curves of AGNs, in which the structure of the projected manifold correlates with AGN class and independent spectroscopic properties. Using the NASA Fornax Science Platform, we assemble light curves from ZTF, Pan-STARRS, Gaia, and WISE/NEOWISE for two samples: (1) a heterogeneous set of $\sim$2000 AGNs spanning $z \lesssim 1$, including SDSS quasars, variability-selected sources, and CLAGNs; and (2) a homogeneous sample of $\sim$65000 narrow-line AGNs at $z \approx 0.1$ with well-characterized optical emission-line measurements. Without using class labels during training, the learned manifolds organize variability-selected AGNs into coherent regions of the low-dimensional space, distinguish between turn-on and turn-off CLAGNs, and place tidal disruption events in distinct regions. Manifold coordinates correlate with key spectroscopic and host-galaxy properties, including stellar mass, [OIII] luminosity, and D$_n$(4000), demonstrating that heterogeneous multi-band variability can be combined in a purely data-driven manner to recover correlations with independent physical diagnostics, without requiring explicit physical modeling. These results show that manifold learning offers a practical, assumption-light approach for integrating time-domain surveys and prioritizing spectroscopic follow-up.

Figures (10)

• Figure 1: Light curves of two AGNs at $z=0.277$ and $z=0.79$, queried from Gaia, ZTF, Pan-STARRS, and WISE, shown in the left and right panels, respectively. The bottom panels are zoomed-in views of the ZTF bands.
• Figure 2: Unified time arrays for the same two objects shown in Figure \ref{['fig:sample_lightcurve']}. The ZTF $g$ and $r$ bands, along with the WISE W1 and W2 bands, are displayed. Gaussian Process Regression is shown with solid lines and shaded regions, and nearest-neighbor linear interpolation is depicted with dashed lines.
• Figure 3: The redshift distribution of the two fully distinct samples is presented here. The left panel corresponds to first one with approximately 2,000 AGNs; the inset pie chart shows the number of objects from each sub-sample. The right panel comrises of approximately 65,000 Type-2 AGNs from Kauffmann2003. The inset plot shows the sample on the BPT diagram, color-coded by the mean fractional variation of the lightcurves. All objects in this subsample sit in the composite and AGN parts of the BPT by construct.
• Figure 4: UMAP projections of the WISE W1 light curves for sample A, generated using different combinations of distance metrics (columns) and interpolation methods (rows). This sample includes AGNs compiled from archival variability catalogs, including changing-look AGNs (CLAGNs) and tidal disruption events (TDEs). Each point represents a source, color-coded by the mean fractional variation of its light curve. The color scale ranges from dark blue (low variability) to yellow (high variability).These projections allow us to assess how different preprocessing choices affect the learned manifold structure. Among the configurations, Gaussian Process (GP) interpolation combined with the Dynamic Time Warping (DTW) distance metric produces the most coherent structure, with a clear gradient in fractional variation. The presence of this gradient indicates that variability amplitude is a key factor shaping the manifold, suggesting that this configuration best preserves physically meaningful variability information.
• Figure 5: This figure presents UMAP projections of the WISE W1 light curves for the first AGN sample, using GP regression and the DTW distance metric (as in the bottom-right panel of Figure 4). Points are color-coded by mean brightness, redshift, and mean fractional variation (left to right). As expected, given that the manifold was constructed from normalized lightcurves, there is no clear structure associated with redshift or brightness. In contrast, a pronounced gradient is observed with mean fractional variation, indicating that variability amplitude plays a dominant role in shaping the learned manifold. This confirms that the embedding captures physically meaningful differences in AGN variability behavior.