A comparison of Bayesian and frequentist confidence intervals in the presence of a late Universe degeneracy

Abstract

Hubble tension is a problem in one-dimensional (1D) posteriors, since local $H_0$ determinations are only sensitive to a single parameter. Projected 1D posteriors for $\Lambda$CDM cosmological parameters become more non-Gaussian with increasing effective redshift when the model is fitted to redshift-binned data in the late Universe. We explain mathematically why this non-Gaussianity arises and show using observational Hubble data (OHD) that Markov Chain Monte Carlo (MCMC) marginalisation leads to 1D posteriors that fail to track the $\chi^2$ minimum at $68\%$ confidence level in high redshift bins. To gain a second perspective, we resort to profile likelihoods as a complementary technique. Doing so, we observe that $z \gtrsim 1$ cosmic chronometer (CC) data currently prefers a non-evolving (constant) Hubble parameter over a Planck-$\Lambda$CDM cosmology at $\sim 2 \sigma$. Within the Hubble tension debate, it is imperative that subsamples of data sets with differing redshifts yield similar $H_0$ values. In addition, we confirm that MCMC degeneracies observed in 2D posteriors are not due to curves of constant $\chi^2$. Finally, on the assumption that the Planck-$\Lambda$CDM cosmological model is correct, using profile likelihoods we confirm a $>2 \sigma$ discrepancy with Planck-$\Lambda$CDM in a combination of CC and baryon acoustic oscillations (BAO) data beyond $ z \sim 1.5$. This confirms a discrepancy reported earlier with fresh methodology.