Fast and effortless computation of profile likelihoods using CONNECT

TL;DR

This paper introduces a fast, neural network-based method for computing profile likelihoods in cosmology, significantly reducing computational time and enabling detailed likelihood analyses previously hindered by high costs.

Contribution

The authors develop a novel approach combining CONNECT and gradient-based basin-hopping to efficiently compute profile likelihoods, achieving 1-2 orders of magnitude speed-up over traditional methods.

Findings

Achieved high-precision likelihood calculations with minimal bias.

Enabled generation of cosmological triangle plots using likelihood maximization.

Demonstrated method on multiple cosmological models with consistent results.

Abstract

The frequentist method of profile likelihoods has recently received renewed attention in the field of cosmology. This is because the results of inferences based on the latter may differ from those of Bayesian inferences, either because of prior choices or because of non-Gaussianity in the likelihood function. Consequently, both methods are required for a fully nuanced analysis. However, in the last decades, cosmological parameter estimation has largely been dominated by Bayesian statistics due to the numerical complexity of constructing profile likelihoods, arising mainly from the need for a large number of gradient-free optimisations of the likelihood function. In this paper, we show how to accommodate the computational requirements of profile likelihoods using the publicly available neural network framework CONNECT together with a novel modification of the gradient-based $basin$-$hopping$ optimisation algorithm. Apart from the reduced evaluation time of the likelihood due to the neural network, we also achieve an additional speed-up of 1$-$2 orders of magnitude compared to profile likelihoods computed with the gradient-free method of $simulated$ $annealing$, with excellent agreement between the two. This allows for the production of typical triangle plots normally associated with Bayesian marginalisation within cosmology (and previously unachievable using likelihood maximisation because of the prohibitive computational cost). We have tested the setup on three cosmological models: the $\Lambda$CDM model, an extension with varying neutrino mass, and finally a decaying cold dark matter model. Given the default precision settings in CONNECT, we achieve a high precision in $\chi^2$ with a difference to the results obtained by CLASS of $\Delta\chi^2\approx0.2$ (and, importantly, without any bias in inferred parameter values) $-$ easily good enough for profile likelihood analyses.