Small Field models with ACTPol and BICEP3 data -- Likelihood analysis

TL;DR

This paper uses Bayesian analysis with neural networks to study small field inflation models with recent CMB data, finding likely model coefficients and a tension between datasets.

Contribution

Introduces a neural network-based Bayesian approach to analyze small field inflation models using combined Planck, ACTPol, and BICEP3 data.

Findings

Most likely polynomial coefficients with tensor-to-scalar ratio r ≲ 0.03

Models are tuned with Δ ≳ 1/60

Identifies tension between ACTPol and Planck datasets

Abstract

We perform a Bayesian analysis for small field models of inflation, using the most recent datasets produced by Planck`18, ACTPol, and BICEP3. We employ Artificial Neural Networks (ANN) to perform analyses with model coefficients, instead of their proxy slow-roll parameters. The ANN connects the models with their projected scalar index $n_s$ and index running $\alpha$, in lieu of the less accurate Lyth-Riotto expressions. We recover the most likely coefficients for a sixth degree polynomial inflationary potential, which yields a tensor-to-scalar ratio $r\lesssim 0.03$. We do so for the case of joint Planck and ACTPol datasets, and for each dataset alone. The BICEP3 data is included in all three analyses. We show that these models are likely, with coefficients that are tuned to about $\Delta\gtrsim 1/60$. Curiously, we also find a significant tension between ACTPol and Planck datasets, which we try to account for.