Moving Beyond the DRW Model: Learning Quasar Variability with Generative Models
Organizations: Institute of Science and Technology Austria (ISTA), Am Campus 1, Klosterneuburg 3400, Austria · Department of Astronomy, Columbia University New York, NY 10027, USA · Department of Physics, Columbia University New York, NY 10027, USA
Abstract
The optical continuum variability of quasars is often described by a damped random walk (DRW) model, with characteristic damping timescales of hundreds of days and amplitudes of 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.
Figures & tables
| Term | Description |
|---|---|
| AR(1) | Auto-regressive process of order 1; a normalized version of the "Damped Random Walk (DRW)" model for quasar variability. See Eq. 4 |
| AR(2) | Auto-regressive process of order 2; a normalized version of the "Damped Harmonic Oscillator (DHO)" model for quasar variability. See Eq. 1 |
| Experiment | Description | Result |
|---|---|---|
| 1 | Can the generative model detect a small higher-order deviation from the DRW model (i.e. an AR(1) process)? | Yes, even for the two groups of light curves in which the average DRW PSD is identical (Fig. 2 ), the model detects and reproduces small deviations in the distribution of Fourier amplitudes (Fig. 3 : uniform ablation vs Fig. 4 : pure DRW) |
| 2 | Can the generative model learn and reproduce AR(2) light curves without being informed of the parametric form of the AR(2) model? | Yes, the model reproduces light curves whose fitted coefficients match the distribution of coefficients fitted to the data (Fig. 5 ). |
| 3 | Can the generative model infer the parameters of an individual light curve, and generate new light curves with the same parameters? | Yes, the coefficients of the generated light curves follow those of the conditioning light curves, with an accuracy comparable to fitting the correct model and simulating new light curves, and a somewhat larger spread (Fig. 6 ). |
| 4 | How do the generative model-generated light curves perform as a null hypothesis compared to other null hypotheses in a periodic signal detection exercise? | In the case of DRW, the generative model performs comparably to a null hypothesis in which the DRW parameters are fitted to each light curve (Fig. 7 , 8 ). In the case of a higher-order deviation to DRW, i.e. an AR(2) process, the generative model approaches the "oracle" scenario, in which the parameters are exactly known, outperforms fitting an AR(2) model to each light curve, and significantly outperforms a null hypothesis incorrectly assuming DRW (see Fig. 9 , 10 ). |
| 5 | If the training data is "contaminated" by periodic signals, does it affect detection performance of periodic signals, and can this be mitigated? | When the training data contains periodic signals (the smallest fraction tested), the detection performance decreases substantially (Fig. 11 ). Projecting the conditioning vector onto its leading principal components restores most of the detection performance, without retraining (Fig. 13 ). |
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
| Variant | Coverage |
|---|---|
| Both additional losses | 46.0–50.5% |
| InfoNCE only | 66.5–72.0% |
| Consistency only | 83.5–85.5% |
| Flow matching only | 82.0–85.5% |
| AR(1) posterior predictive | 88.0% |
| Variant | Bias | RMSE |
|---|---|---|
| Both auxiliary objectives | ||
| InfoNCE only | ||
| Consistency only | ||
| Flow matching alone | ||
| Direct reference |
| Comparison | |
|---|---|
| Training run 1 vs reference data | |
| Training run 2 vs reference data | |
| Training run 3 vs reference data | |
| Independent reference samples |
| Integration steps | Mean absolute error |
|---|---|
| 30 | 0.049 |
| 200 | 0.027 |
| 500 | 0.025 |
| Mean bias | Mean SD | Mean IQR | |
|---|---|---|---|
| Reference | 0.020 | 0.071 | 0.096 |
| Generative model | 0.022 | 0.083 | 0.114 |
| Reference | Generative model | |||||
|---|---|---|---|---|---|---|
| Mean | SD | 5–95% | Mean | SD | 5–95% | |
| 0.2 | 0.186 | 0.087 | 0.029–0.330 | 0.201 | 0.091 | 0.059–0.361 |
| 0.3 | 0.288 | 0.085 | 0.145–0.419 | 0.293 | 0.102 | 0.118–0.444 |
| 0.4 | 0.384 | 0.082 | 0.257–0.512 | 0.389 | 0.095 | 0.236–0.537 |
| 0.5 | 0.482 | 0.077 | 0.362–0.604 | 0.479 | 0.098 | 0.310–0.638 |
| 0.6 | 0.577 | 0.075 | 0.445–0.698 | 0.576 | 0.088 | 0.427–0.710 |
| Method | AUC [95% CI] | FPR at | FPR at |
|---|---|---|---|
| Fitted AR(1) | 0.9042 [0.8852, 0.9217] | 3.2% (32) [2.20, 4.49] | 0.2% (2) [0.02, 0.72] |
| Population AR(1), without fitting | 0.8291 [0.8019, 0.8546] | 5.3% (53) [3.99, 6.88] | 1.5% (15) [0.84, 2.46] |
| Oracle AR(1) | 0.9646 [0.9538, 0.9742] | 4.4% (44) [3.21, 5.86] | 0.5% (5) [0.16, 1.16] |
| Generative model (flow matching alone) | 0.9273 [0.9113, 0.9425] | 6.3% (63) [4.87, 7.99] | 1.0% (10) [0.48, 1.83] |
| Method | AUC | FPR (%), | FPR (%), |
|---|---|---|---|
| Fitted AR(1), misspecified | 0.8505 | 41.60 | 23.70 |
| Fitted AR(2) | 0.9180 | 3.00 | 0.00 |
| Oracle AR(2) | 0.9855 | 5.50 | 0.70 |
| Generative model, run 1 | 0.9616 | 3.50 | 0.10 |
| Generative model, run 2 | 0.9662 | 3.40 | 0.10 |
| Generative model, run 3 | 0.9665 | 6.20 | 0.80 |
| Method | AUC [95% CI] | FPR (%), [95% CI] | FPR (%), [95% CI] |
|---|---|---|---|
| AR(1) Baseline | 0.9699 [0.9568, 0.9819] | 4.00 [1.74, 7.73] | 0.00 [0.00, 1.83] |
| Nullasar (0%) | 0.9904 [0.9815, 0.9965] | 10.50 [6.62, 15.60] | 1.00 [0.12, 3.57] |
| Nullasar (0.5%) | 0.5735 [0.5173, 0.6309] | 9.50 [5.82, 14.44] | 1.50 [0.31, 4.32] |
| Nullasar (1%) | 0.5557 [0.4952, 0.6114] | 9.50 [5.82, 14.44] | 1.50 [0.31, 4.32] |
| Nullasar (5%) | 0.5387 [0.4807, 0.5967] | 7.50 [4.26, 12.07] | 1.00 [0.12, 3.57] |
| Nullasar (10%) | 0.5218 [0.4638, 0.5778] | 9.00 [5.42, 13.85] | 1.50 [0.31, 4.32] |
| Contamination | Mean AUC | SD across runs |
|---|---|---|
| 0% | 0.9802 | 0.0026 |
| 0.5% | 0.5771 | 0.0605 |
| 1% | 0.5659 | 0.0870 |
| 5% | 0.5419 | 0.0519 |
| 10% | 0.5309 | 0.0867 |
| 20% | 0.5472 | 0.0095 |
| Method | AUC [95% CI] | FPR (%), [95% CI] | FPR (%), [95% CI] |
|---|---|---|---|
| Nullasar (0%) / run 1 | 0.9813 [0.9681, 0.9925] | 9.00 [5.42, 13.85] | 2.00 [0.55, 5.04] |
| Nullasar (0%) / run 2 | 0.9821 [0.9678, 0.9928] | 10.50 [6.62, 15.60] | 2.00 [0.55, 5.04] |
| Nullasar (0%) / run 3 | 0.9773 [0.9616, 0.9903] | 14.00 [9.51, 19.59] | 3.00 [1.11, 6.42] |
| Nullasar (0.5%) / run 1 | 0.6071 [0.5522, 0.6612] | 8.00 [4.64, 12.67] | 2.00 [0.55, 5.04] |
| Nullasar (0.5%) / run 2 | 0.5074 [0.4488, 0.5641] | 13.00 [8.67, 18.47] | 2.50 [0.82, 5.74] |
| Nullasar (0.5%) / run 3 | 0.6168 [0.5644, 0.6694] | 13.00 [8.67, 18.47] | 2.50 [0.82, 5.74] |