Researchers re-examined the famous Hubble constant result from the neutron-star merger GW170817 and found that the assumed prior on distance can substantially change the inferred value. Using a GPU-accelerated pipeline, they reran the full parameter inference under different distance priors and compared those direct results with the common practice of changing the prior after the fact by reweighting existing samples. The direct and reweighted results did not agree: changing the prior during sampling moved substantial probability into a high-H0 tail that reweighting largely missed.

The team built a fast analysis that combines a heterodyned (relative-binning) likelihood and a GPU-native nested sampling algorithm. Heterodyning reduces the cost of evaluating long gravitational-wave signals. Nested sampling is a method to draw samples and estimate the model evidence; they used 5,000 live points and report completing the full analysis in about 13 minutes on a single A100 graphics card. They applied this to the IMRPhenomXAS_NRTidalv3 waveform model for GW170817 and reran inference under four different distance priors.

When they imposed a uniform-in-distance prior during sampling instead of the original volumetric prior (which favors larger volumes at larger distances), the probability of H0 exceeding 120 km/s/Mpc rose from 0.017 to 0.159. The weighted median H0 shifted from 77.6 to 87.6 km/s/Mpc, even though the peak (the binned maximum a posteriori) stayed at 70.5 km/s/Mpc. By contrast, post-hoc reweighting of the baseline samples toward the uniform-in-distance prior recovered only P(H0>120)=0.041, roughly 17% of the increase seen when the prior was used during sampling.

The authors traced the discrepancy to a bimodality in the distance and binary inclination parameters. One branch corresponds to a nearer, more inclined binary that maps to higher H0. The volumetric prior used in the original analysis assigned very little prior weight to that branch, so the baseline run sampled it sparsely. Reweighting cannot reliably amplify an under-sampled region. The reweighted posterior also showed a much lower effective sample size, which independently signals that the reweighting failed to cover the target posterior properly.