Determine the cause of low efficiency on off-segment noise injections

Investigate whether a selection effect associated with non-Gaussian noise in off-segment LIGO data causes out-of-dataset errors that reduce BilbyFlow reweighting efficiency for synthetic waveform injections.

Background

BilbyFlow performs substantially worse on waveform injections into off-segment LIGO noise than on injections into Gaussian noise, even after restricting the source-parameter space to regions occupied by detected events. The authors observe that the corresponding efficiency does not notably improve after selection cuts, unlike the Gaussian-noise case.

The paper attributes this discrepancy tentatively to a possible secondary selection effect: off-segment data may contain non-Gaussian noise that is not represented in the training distribution, producing out-of-dataset errors in the neural posterior estimator. Because this explanation is explicitly unconfirmed, identifying the cause remains unresolved.

References

Although not confirmed, we suspect that this is due to another selection effect where the off-segment data used is more likely to have non-Gaussian noise leading to out-of-dataset errors in the framework.

— $\texttt{BilbyFlow}$: user-friendly neural posterior estimation for gravitational-wave astronomy  (2609.00766 - Pinchbeck et al., 1 Sep 2026) in Section 'Restricted parameter space efficiencies'