Optimal data compression for Lyman-$\alpha$ forest cosmology

TL;DR

This paper introduces a near-maximal score compression method for Lyman-alpha forest cosmology data, reducing data dimensionality while preserving information and enabling reliable goodness-of-fit tests, demonstrated on mock and real survey data.

Contribution

It develops a lossless, unbiased compression technique that maintains the full information content of Lyman-alpha correlation data for cosmological inference.

Findings

Compression is lossless and unbiased on realistic mocks and survey data.

Additional data beyond model parameters are necessary for meaningful goodness-of-fit tests.

Compressed covariance matrices are better conditioned and suitable for limited mock samples.

Abstract

The Lyman-$\alpha$ (Ly$\alpha$) three-dimensional correlation functions have been widely used to perform cosmological inference using the baryon acoustic oscillation (BAO) scale. While the traditional inference approach employs a data vector with several thousand data points, we apply near-maximal score compression down to tens of compressed data elements. We show that carefully constructed additional data beyond those linked to each inferred model parameter are required to preserve meaningful goodness-of-fit tests that guard against unknown systematics, and to avoid information loss due to non-linear parameter dependencies. We demonstrate, on suites of realistic mocks and DR16 data from the Extended Baryon Oscillation Spectroscopic Survey, that our compression approach is lossless and unbiased, yielding a posterior that is indistinguishable from that of the traditional analysis. As an early application, we investigate the impact of a covariance matrix estimated from a limited number of mocks, which is only well-conditioned in compressed space.