Moving Beyond the DRW Model: Learning Quasar Variability with Generative Models

Nadav Bojan Sellam, Kevin Park, Alex Bronstein, Zoltán HaimanarXiv 2026 · arXiv

TL;DR

Standard statistical models for quasar variability, like the Damped Random Walk (DRW), are often too simple, leading to biased results when searching for rare signals like binary black holes. This paper introduces Nullasar, a generative model trained via flow matching that learns complex, stochastic quasar variability directly from data, providing a robust, data-driven "null hypothesis" for astrophysical discovery.

What Problem Does This Paper Address?

Quasars—supermassive black holes actively devouring matter—flicker in brightness stochastically. Astronomers use statistical models of this "noise" to search for periodic signals that might indicate a pair of orbiting black holes (SMBHBs). The industry standard is the Damped Random Walk (DRW) model.

However, the DRW is often a poor fit for real-world data. If your "null hypothesis" (your model of background noise) is wrong, you run into two major problems:

  1. False Positives: Mistaking a complex random fluctuation for a genuine periodic signal.
  2. False Negatives: A model that is too flexible might "swallow" a real signal, treating it as part of the noise.

This paper addresses the need for a noise model that doesn't rely on rigid parametric assumptions but instead learns the true statistical nature of quasar variability from large datasets.

Methods

The authors present Nullasar, a generative model based on Continuous Normalizing Flows (CNF) and trained using flow matching.

  • Architecture: It uses a 1D U-Net to parameterize a velocity field that transforms simple Gaussian noise into realistic quasar light curves.
  • Domain: Uniquely, the model operates in the frequency domain, acting directly on Fourier coefficients. This allows it to explicitly represent global spectral structures and long-range correlations.
  • Conditioning: An encoder network maps an observed light curve to a latent vector $z$. Nullasar then generates new realizations that are statistically consistent with that specific observation.
  • Robustness: To handle "contaminated" training data (e.g., datasets that accidentally include the periodic signals you're trying to find), the authors propose projecting the latent conditioning vector onto its leading Principal Components (PCA).

Datasets

The study relies on large-scale synthetic datasets where the "ground truth" is known:

  • AR(1) / DRW: The standard baseline model.
  • AR(2): A more complex model capable of producing quasi-periodic oscillations, used to test model mis-specification.
  • Ablation Datasets: Custom datasets designed with identical power spectra but different higher-order statistics (e.g., Rayleigh vs. Uniform Fourier amplitude distributions) to see if the model can detect subtle differences.
  • Training Size: Typically 100,000 light curves per experiment.

Key Insights and Ideas

  • Beyond the Power Spectrum: Standard methods often only look at the Power Spectral Density (PSD). Nullasar successfully learns higher-order statistics, such as the specific shape of the distribution of Fourier amplitudes, which is invisible to PSD-only methods.
  • Population-Informed Regularization: When fitting a model to a single, noisy light curve, parametric fits can go wild. Because Nullasar is trained on a whole population, it "knows" which noise parameters are likely, providing a natural regularizing effect that prevents it from over-fitting to signals.
  • The Contamination Trap: A purely data-driven model is too good at its job. If 0.5% of your training data contains periodic signals, the model learns that periodicity is "normal" noise, destroying its ability to detect those signals later.
  • The PCA Fix: The authors found that the "noise" of quasars occupies the primary directions of the model's latent space, while rare signals (like binaries) live in the "tails." By keeping only the top 6 PCA components of the latent vector, they could filter out the influence of periodic signals in the training set without retraining the model.

Key Results

  • Statistical Fidelity: Nullasar perfectly reproduced the non-standard "Uniform-amplitude" distribution in ablation tests, whereas standard DRW models failed.
  • Superior Detection: In the challenging AR(2) scenario, Nullasar achieved an AUC of 0.96 for signal detection, significantly outperforming standard parametric fitting (AUC 0.92) and crushing the mis-specified AR(1) model (AUC 0.85).
This figure illustrates how using the wrong null hypothesis (AR(1) instead of AR(2)) leads to frequency-dependent errors in p-values, causing both false positives and negatives.
  • Calibration: The $p$-values generated by Nullasar are well-calibrated, meaning the probability of a false alarm matches the theoretical expectation.

Limitations and Future Work

  • Uniform Sampling: The current model assumes data points are collected at perfectly regular intervals. Real telescopes deal with gaps due to weather and seasonal cycles.
  • Fixed Length: Nullasar is currently limited to a fixed sequence length ($T=256$). Extending this to variable lengths and longer timescales is a priority.
  • Multi-band Data: Future versions could incorporate color information (multi-filter data from surveys like LSST) to further distinguish black hole binaries from stochastic noise.

Nullasar is publicly available on GitHub.

Summary updated

Abstract

The optical continuum variability of quasars is often described by a damped random walk (DRW) model, with characteristic damping timescales $τ_{\rm DRW}$ of hundreds of days and amplitudes $σ_{\rm DRW}$ of $\mathcal{O}(0.1)$ magnitudes. This model has been widely used as a null hypothesis to assess the significance of periodic signals in quasar light curves in the search for supermassive black hole binaries, the significance of extreme quasar variability like tidal disruption events (TDEs) and changing-look (CL) quasars, or to model time lags in reverberation mapping. However, the DRW model is known to be incomplete, and an incorrect null hypothesis can lead to biased significance estimates. We investigate the impact of such null-hypothesis mis-specification on the detection of periodic signals in synthetic quasar light curves and introduce the generative model Nullasar, capable of learning quasar variability directly from time-series data. We demonstrate that Nullasar, trained with flow matching, can learn delicate statistical properties of stochastic light-curve variability without explicitly assuming any noise model. We further demonstrate that Nullasar can learn correlated distributions of noise parameters and generate new light-curve realizations from a range of deviations from the Gaussian DRW noise model, approximating parametric descriptions of stochastic processes. When used as a null hypothesis for detecting periodic signals in synthetic light curves, Nullasar performs comparably to fitting the correct DRW model to each light curve, and outperforms parametric fitting when the variability follows a more flexible AR(2) model. Finally, we identify practical limitations, including sensitivity to contamination of the training set by non-null signals, and show that projecting the conditioning vector of the model onto its leading principal components largely mitigates it.
BibTeX
@misc{sellam2026moving,
  author = {Nadav Bojan Sellam and Kevin Park and Alex Bronstein and Zoltán Haiman},
  title = {{Moving Beyond the DRW Model: Learning Quasar Variability with Generative Models}},
  year = {2026},
  eprint = {2611.15040},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.15040},
}
Figure