DES Year 3 shear data get their first integrated three-point cosmology, and adding it to two-point statistics roughly doubles the figure of merit

The DES Year 3 weak-lensing shear catalogue can now constrain flat ΛCDM parameters with two-point correlation functions and the integrated three-point correlation function together, and the three-point statistic is detected in the data at a signal-to-noise ratio of 12.8. Previously this same DES Y3 shear dataset drew its information mainly from two-point statistics, which are sensitive only to the second moments of the density field and miss the higher-order information in the non-Gaussian late-time matter distribution. The joint data vector has a signal-to-noise ratio of 33.0, against 31.0 for the two-point correlation functions and 12.8 for the integrated three-point correlation function, with the covariance estimated from 16,000 fiducial footprints of the forward model.

The joint analysis yields posterior medians and 68% credible intervals for Ωm, S8 and w, and roughly doubles the figure of merit in the Ωm–S8–w parameter space over two-point statistics alone. With two-point correlation functions alone, Ωm and S8 are constrained along one degeneracy direction, while the three-point statistic constrains a different combination of the two, so the joint posterior sits where the two constraints intersect. On the single DES Y3 data realisation, adding the i3PCF improves Ωm by 22%, S8 by 26% and w by 49%, and raises the Ωm–S8–w figure of merit by 99%; this improvement lies between the 65th and 77th percentiles of the same quantity over 800 fiducial noise realisations and 1250 test cases.

The inference does not assume a Gaussian likelihood; it learns the likelihood of the compressed data vector from forward simulations with a normalising flow. Higher-order statistics generally have no analytic likelihood, and the widely used Gaussian likelihood approximation need not hold for them. The 1440-element joint data vector is first compressed linearly with MOPED to 17 coefficients, then a neural spline flow is trained on 60,000 CosmoGridV1 simulations whose forward model includes the survey mask, shape noise, intrinsic alignment, baryonic feedback, photometric redshift uncertainty and shear calibration bias.

The analysis was carried out blind, and the data were unblinded only after coverage tests, tests of the accuracy of the learned likelihoods and internal consistency tests had all passed. Earlier DES Y3 simulation-based inference analyses mostly blinded at the posterior level, whereas here every galaxy ellipticity in the catalogue was rescaled by an unknown factor. The TARP coverage of all eight inference pipelines lies within the binomial band, and all pipelines pass the coverage tests; the DES Y3 measurement sits at the 56th percentile of the Mahalanobis distance under the learned likelihood in the joint analysis, indistinguishable from a typical mock.

The result applies to the DES Y3 weak-lensing shear catalogue, using an effective area of 4143 square degrees and four tomographic source bins, with a lower scale cut about twice the pixel scale that excludes the small scales most affected by baryonic feedback. The intended readers are cosmology researchers working on higher-order weak-lensing statistics and simulation-based inference.

The forward model rests on the single CosmoGridV1 simulation suite, and recovery of the input cosmology could not be tested on simulations from an independent N-body code; the prior covers only flat ΛCDM at a fixed total neutrino mass, so there is no corresponding wCDM analysis; intrinsic alignments enter only through the NLA model without redshift evolution; and the MOPED compression vectors are fixed at one fiducial cosmology, which can lose information but does not bias the result.