Standard-siren cosmology stops restarting from the prior: ASPIRE reuses old posteriors for a roughly 29-fold speed-up in a dark-siren case

A standard-siren cosmological analysis no longer has to start from the prior; posterior samples left by an earlier run can be transformed directly into the posterior under a new hypothesis. Previously every population model, galaxy catalogue or weighting scheme required its own full inference run, and information from the earlier run was discarded when the catalogue was updated. Across GWTC-5.0 dark-siren cases, ASPIRE posteriors agree with those from the nested sampler dynesty, with all 23 parameters falling below the chosen Jensen-Shannon divergence threshold.

The method first fits a normalising flow to the old posterior as a proposal that can be sampled and evaluated, then pushes it toward the new target along a chain of intermediate distributions. A normalising flow is a neural network that learns an invertible mapping between a simple base distribution and a target distribution; sequential Monte Carlo anneals by inverse temperature, resampling with importance weights and applying MCMC moves at each step. The annealing schedule is set by a target effective sample size, and by the final temperature step no proposal information remains in the posterior; in the case adding GLADE+ catalogue information to a spectral-siren result, sampling took 4.6 hours against 134.4 hours for dynesty.

The demonstrated scope is a set of dark-siren and spectral-siren cases using GWTC-5.0 data, the GLADE+ galaxy catalogue and the FullPop-4.0 mass model, and the reported speed-ups assume the proposal distribution already exists, excluding the cost of producing it.

Updating a GWTC-4.0 spectral-siren result to GWTC-5.0 changed the correlations inside the posterior, so four parameters tied to the Hubble constant had to be drawn from the prior; when broadening a tight fixed-cosmology posterior, reweighting and resampling cannot create samples where the proposal has none.