[Paper Review] Dark Energy Survey Year 3 results: likelihood-free, simulation-based $w$CDM inference with neural compression of weak-lensing map statistics

N Jeffrey, L Whiteway+89 more|arXiv (Cornell University)|Mar 4, 2024

Gaussian Processes and Bayesian Inference4 citations

TL;DR

This paper presents a likelihood-free, simulation-based inference framework for $w$CDM cosmology using Dark Energy Survey Year 3 weak-lensing maps, combining neural compression of map statistics (power spectra, peak counts, and direct CNN-based map compression) with forward-modelled mock data to achieve precise parameter constraints. The method improves dark energy parameter precision by over a factor of two compared to power spectrum-only inference, yielding $w < -0.80$ at 68% credible interval with $\Omega_{\rm m} = 0.283^{+0.020}_{-0.027}$ and $S_8 = 0.804^{+0.025}_{-0.017}$.

ABSTRACT

We present simulation-based cosmological $w$CDM inference using Dark Energy Survey Year 3 weak-lensing maps, via neural data compression of weak-lensing map summary statistics: power spectra, peak counts, and direct map-level compression/inference with convolutional neural networks (CNN). Using simulation-based inference, also known as likelihood-free or implicit inference, we use forward-modelled mock data to estimate posterior probability distributions of unknown parameters. This approach allows all statistical assumptions and uncertainties to be propagated through the forward-modelled mock data; these include sky masks, non-Gaussian shape noise, shape measurement bias, source galaxy clustering, photometric redshift uncertainty, intrinsic galaxy alignments, non-Gaussian density fields, neutrinos, and non-linear summary statistics. We include a series of tests to validate our inference results. This paper also describes the Gower Street simulation suite: 791 full-sky PKDGRAV dark matter simulations, with cosmological model parameters sampled with a mixed active-learning strategy, from which we construct over 3000 mock DES lensing data sets. For $w$CDM inference, for which we allow $-1

Motivation & Objective

• To improve cosmological parameter inference in the $w$CDM model using weak gravitational lensing data from DES Year 3.
• To develop a simulation-based inference framework that fully propagates complex statistical uncertainties—such as shape noise, photometric redshift errors, and non-Gaussianity—through mock data.
• To leverage deep learning for neural compression of high-dimensional lensing map statistics (power spectra, peak counts, and direct map representations) to enhance inference efficiency and accuracy.
• To validate the inference pipeline using a large, publicly released simulation suite (Gower Street) with 791 full-sky cosmological simulations and over 10,000 mock surveys.
• To demonstrate the superiority of map-level inference via convolutional neural networks over traditional two-point statistics in constraining dark energy and matter density.

Proposed method

• The method employs simulation-based inference (likelihood-free inference), using forward-modelled mock data to estimate posterior distributions without requiring explicit likelihood functions.
• Neural compression is applied to three types of lensing summary statistics: power spectra, peak counts, and raw lensing maps via convolutional neural networks (CNNs) for end-to-end map-level compression.
• A mixed active-learning strategy samples cosmological parameters (including $\Omega_{\rm m}, w, \sigma_8, h, \Omega_{\rm b}h^2, n_s, m_\nu$) across 791 full-sky pkdgrav3 simulations to construct the Gower Street simulation suite.
• The mock data include realistic observational effects: sky masks, non-Gaussian shape noise, shape measurement bias, source clustering, photometric redshift uncertainty, intrinsic alignments, and non-linear density fields.
• Neural density estimation is used to learn the likelihood from simulated data, enabling posterior validation and uncertainty quantification.
• Posterior distributions are validated through multiple consistency checks, including comparison to Planck and standard DES two-point function results.

Experimental results

Research questions

• RQ1 Can map-level inference using deep learning improve cosmological constraints in $w$CDM beyond traditional two-point statistics?
• RQ2 How does neural compression of lensing map statistics (power spectra, peak counts, and direct maps) affect inference precision and computational efficiency?
• RQ3 To what extent can simulation-based inference with realistic forward modeling reduce systematic biases in weak-lensing cosmology?
• RQ4 What is the impact of including non-Gaussian and non-linear summary statistics on dark energy parameter constraints?
• RQ5 How does the combination of power spectra and CNN-compressed maps compare to power spectra alone in constraining $\Omega_{\rm m}$, $S_8$, and $w$?

Key findings

• The combination of power spectra and CNN-based map-level inference yields the most constraining results, improving dark energy parameter precision ($\Omega_{\rm DE}, w$) by more than a factor of two compared to power spectrum-only inference.
• The most precise constraints are $\Omega_{\rm m} = 0.283^{\pm 0.020}_{-0.027}$, $S_8 = 0.804^{+0.025}_{-0.017}$, and $w < -0.80$ at 68% credible interval.
• The method successfully propagates complex statistical uncertainties—such as shape noise, photometric redshift errors, and non-Gaussianity—through the full inference pipeline via forward-modelled mocks.
• The Gower Street simulation suite, comprising 791 full-sky cosmological simulations and over 10,000 non-independent mock surveys, enables robust training and validation of the inference framework.
• Validation tests confirm consistency with external constraints from Planck CMB and standard DES two-point function likelihoods, supporting the reliability of the posterior estimates.
• Map-level inference via CNNs captures non-Gaussian features in lensing maps that are missed by power spectra, leading to enhanced sensitivity to cosmological parameters.