How Theory-Informed Priors Affect DESI Evidence for Evolving Dark Energy

Abstract

Recent measurements of baryon acoustic oscillations (BAO) from the Dark Energy Spectroscopic Instrument (DESI) have been interpreted to suggest that dark energy may be evolving. In this work, we examine how prior choices affect such conclusions. Specifically, we study the biases introduced by the customary use of uniform priors on the Chevallier-Polarski-Linder (CPL) parameters, $w_0$ and $w_a$, when assessing evidence for evolving dark energy. To do so, we construct theory-informed priors on $(w_0, w_a)$ using a normalizing flow (NF), trained on two representative quintessence models, which learns the distribution of these parameters conditional on the underlying $Λ$CDM parameters. In the combined $\textit{Planck}$ CMB + DESI BAO analysis we find that the apparent tension with a cosmological constant in the CPL framework can be reduced from $\sim 3.1σ$ to $\sim 1.3σ$ once theory-informed priors are applied, rendering the result effectively consistent with $Λ$CDM. For completeness, we also analyze combinations that include Type Ia supernova data, showing similar shifts toward the $Λ$CDM limit. Taken together, the observed sensitivity to prior choices in these analyses arises because uniform priors - often mischaracterized as "uninformative" - can actually bias inferences toward unphysical parameter regions. Consequently, our results underscore the importance of adopting physically motivated priors to ensure robust cosmological inferences, especially when evaluating new hypotheses with only marginal statistical support. Lastly, our NF-based framework achieves these results by post-processing existing MCMC chains, requiring $\approx 1$ hour of additional CPU compute time on top of the base analysis - a dramatic speedup over direct model sampling that highlights the scalability of this approach for testing diverse theoretical models.