Classifying Quasar Types Without a Spectrum

The paper demonstrates that Slepian Wavelet Variance (SWV) can be used to classify Type 1 and Type 2 quasars using only irregularly sampled photometric light curves, without the need for spectroscopy. The authors analyze 516 Type 1 and 238 Type 2 quasars from the MILLIQUAS catalogue, using Zwicky Transient Facility (ZTF) light curves. They find that Type 1 quasars typically exhibit a parabola-like variance curve with a minimum near 10 days and rising variance at longer timescales, while Type 2 quasars show an almost monotonic decline in variance with increasing timescale. Using agglomerative hierarchical clustering with complete linkage on the variance curves, they achieve a recovery rate of 99% for Type 1 and 87% for Type 2 quasars. The 34 misclassified quasars show variability behavior opposite to their spectral classification, and inspection of their spectra reveals a variety of types including BAL quasars, broad-line quasars, Seyferts, blazars, and X-ray candidates, suggesting these could be changing-look quasars or objects whose spectral type has evolved since their spectra were taken. The authors also interpret the variance curves physically: for Type 1 quasars, short-term variability (2 0 to 2 squared days) may trace X-ray reprocessing from the corona, while long-term variability (peaking near 2 7 days) may reflect thermal fluctuations in the outer accretion disk, with the minimum around 2 3-2 4 days marking a transition between these regimes. For Type 2 quasars, the weak long-term variability is attributed to stable emitters like host galaxy starlight and the narrow-line region, while short-term variability may leak through the clumpy dust torus. The authors note that SWV offers advantages over structure functions and the Damped Random Walk model because it is model-independent and provides a direct view of variability across multiple timescales without assuming a specific form for the data. They also verify that the classification is not affected by redshift selection effects or survey baseline limitations, and they discuss potential degeneracies with other variable sources (stars, normal galaxies), suggesting that additional color or emission-line diagnostics would be needed for broader classification tasks.

Improvements to AI Systems:

1. Model-Independent Time-Series Feature Extraction for Irregular Data

• Implement SWV as a preprocessing layer in AI pipelines for astronomical time-series classification, replacing or augmenting structure functions and DRW fits. This removes assumptions about stochastic processes and handles unevenly sampled data natively.

• The improved system can directly compute multi-timescale variance features from raw light curves without interpolation or gap-filling, reducing bias and computational overhead.

2. Hierarchical Clustering with Interpretable Variance-Curve Signatures

• Use agglomerative clustering (complete linkage) on SWV curves as an unsupervised pretraining step for downstream classifiers. The resulting dendrogram can serve as a prior for semi-supervised learning, especially when labeled spectra are scarce.

• The improved system can automatically group quasars by variability morphology, flagging outliers (e.g., the 34 misclassified objects) for targeted follow-up—acting as an anomaly detector for changing-look AGN or spectral evolution.

3. Physical-Regime-Aware Classification via Timescale Decomposition

• Train a neural network or gradient-boosted model on SWV curve segments (e.g., short-term 2 0–2 squared days, transition 2 3–2 4 days, long-term 2 7 days) as separate input channels. This forces the model to learn regime-specific variability physics rather than global curve shape.

• The improved system can distinguish Type 1 vs. Type 2 quasars with >99% and >87% recovery, respectively, and can also output a “regime importance” map—explaining whether classification was driven by reprocessing, disk fluctuations, or torus leakage.

4. Redshift- and Baseline-Robust Generalization

• Augment training data with synthetic SWV curves generated under varying redshift, cadence, and survey duration (e.g., ZTF vs. LSST). Use domain adaptation to make the classifier invariant to these observational biases.

• The improved system can be deployed on new surveys without retraining, maintaining high accuracy even when light-curve baselines are shorter or sampling is sparser.

5. Multi-Wavelength and Multi-Epoch Spectral Evolution Prediction

• Integrate SWV features with color or emission-line diagnostics (as suggested by the authors) into a multimodal transformer. The SWV branch handles temporal variability; the spectral branch handles instantaneous classification.

• The improved system can predict whether a quasar’s spectral type has changed since its last spectroscopic observation, flagging candidates for spectroscopic re-observation (e.g., BAL, blazar, or changing-look transitions).

6. Uncertainty-Aware Clustering for Contamination Detection

• Replace hard clustering with a Bayesian hierarchical model (e.g., Dirichlet process mixture) over SWV curves. This yields posterior probabilities for each object belonging to Type 1, Type 2, or an unknown class.

• The improved system can quantify classification confidence, automatically isolating the 34 misclassified quasars as high-uncertainty objects, and prioritize them for human review or additional photometric bands.

7. Transfer Learning to Other Variable Sources

• Pretrain a feature extractor on the quasar SWV curves, then fine-tune on other variable astrophysical sources (e.g., stars, normal galaxies) with small labeled sets. The SWV representation is physically motivated and transferable across object types.

• The improved system can classify a mixed sky survey into quasars, stars, and galaxies with minimal new labels, using the same variance-curve features, while flagging degeneracies for follow-up.

Distinguishing between Type 1 and Type 2 quasars is important because it helps us understand accretion regimes, black hole mass scaling, disk instabilities and feedback processes in active galaxies. Although spectroscopy provides robust classification, it does not scale well with the millions of quasars observed in modern surveys, as it requires substantial time and resources to acquire a good spectrum. On the photometry side, quasar light curves are always irregularly sampled and affected by the specifics of photometric surveys, making them difficult to analyze. In this work, we show that we can use irregularly sampled light curves from the Zwicky Transient Facility to classify quasar types without a spectrum, using Slepian Wavelet Variance. This technique allows us to decompose the variance of light curves into multiple timescales. We use agglomerative hierarchical clustering to classify 516 Type 1 and 238 Type 2 quasars from the MILLIQUAS catalogue, solely based on their wavelet variance curves. We obtain a recovery rate of 99% for Type 1 and 87% for Type 2 quasars, and the few misclassified quasars show the opposite variability behaviour to their spectral type. In contrast to structure functions and the Damped Random Walk model, Slepian Wavelet Variance offers a complementary, model-independent view of variability across short and long timescales.