Reproduce “Adapting to noise distribution shifts in flow-based gravitational-wave inference” on OpenResearch

About this paper

Jonas Wildberger, Maximilian Dax, Stephen R. Green, Jonathan Gair, Michael Pürrer, Jakob H. Macke, Alessandra Buonanno, Bernhard Schölkopf

DINGO estimates source properties from gravitational-wave detector data, conditioning on the noise power spectral density (PSD). The authors generate synthetic PSDs by shifting and broadening a latent distribution fit to O2 noise, using one early O3 observation. Across real events, mean JSD was 1.4 × 10⁻³ nat for synthetic training and 2.7 × 10⁻³ nat for early-noise training. This O3 evidence applies when future PSDs fall within training support.

Key insights

• Designing an interpretable and partially factorized latent space for PSDs allows for targeted, physically informed interventions to adapt training distributions to new detector states with minimal new data.
• Deliberately broadening the synthetic PSD distribution in the latent space effectively ensures that the training distribution's support covers future, unseen noise characteristics, making inference robust to distribution shifts.
• Leveraging prior domain knowledge about PSD structure (e.g., distinct broad-band and spectral line components) within the latent variable model is crucial for achieving effective generalization and data-efficient adaptation.

Summary by alphaXiv.