Abstract
The measurement of the absolute neutrino mass scale from cosmological large-scale clustering data is one of the key science goals of the Euclid mission. Such a measurement relies on precise modelling of the impact of neutrinos on structure formation, which can be studied with $N$-body simulations. Here we present the results from a major code comparison effort to establish the maturity and reliability of numerical methods for treating massive neutrinos. The comparison includes eleven full $N$-body implementations (not all of them independent), two $N$-body schemes with approximate time integration, and four additional codes that directly predict or emulate the matter power spectrum. Using a common set of initial data we quantify the relative agreement on the nonlinear power spectrum of cold dark matter and baryons and, for the $N$-body codes, also the relative agreement on the bispectrum, halo mass function, and halo bias. We find that the different numerical implementations produce fully consistent results. We can therefore be confident that we can model the impact of massive neutrinos at the sub-percent level in the most common summary statistics. We also provide a code validation pipeline for future reference.
Introduction
Neutrinos are the "ghost particles" of the Standard Model. While we know they possess mass, the absolute mass scale remains one of the most significant unknowns in modern physics. In the early universe, neutrinos move at near-light speeds (free-streaming), which prevents matter from clumping together on small scales. This leaves a unique "suppression" signature in the large-scale structure of the universe.
The upcoming Euclid satellite mission intends to measure this signature to finally pin down the absolute neutrino mass scale. However, this requires incredibly precise theoretical models. If different simulation codes provide divergent answers for the same physics, the resulting cosmological constraints cannot be trusted. Neutrinos are notoriously difficult to simulate because they are fast and relativistic, requiring different numerical treatments than the slow-moving "cold" dark matter that dominates the universe.
Methodology
This study performed a "blind" code comparison, providing a standardized set of initial conditions to various research groups. The methods tested fall into three main categories:
- Particle-based methods: These treat neutrinos as a separate set of $N$-body particles. While physically intuitive, they can suffer from "shot noise" unless a massive number of particles is used to represent the smooth neutrino background.
- Mesh-based methods: These treat neutrinos as a fluid or a linear density field on a grid. This eliminates shot noise but must be carefully monitored to ensure it captures non-linear clustering if neutrinos become sufficiently "clumpy."
- Emulators and Surrogates: Fast, machine-learning-based tools (like
EuclidEmulator2orBACCOemulator) trained on previous simulations to predict results in milliseconds.
The study evaluated these across several summary statistics, including the power spectrum (matter clumping at different scales), the bispectrum (three-point correlations), and halo statistics (properties of dark matter clumps where galaxies form).
Datasets and Initial Conditions
The comparison utilized a newly generated suite of controlled simulations known as the Euclid Suite. This included 10 primary simulations with varying total neutrino masses ($\sum m_\nu$ from 0.0 to 0.6 eV). All codes started from the same "fixed-amplitude" initial conditions at redshift $z=127$, generated using the REPS and N-GenIC codes to ensure a fair baseline.
Key Results
Power Spectrum Consistency
The study confirms that all modern codes capture the characteristic "spoon-like" dip in the matter power spectrum caused by massive neutrinos with high consistency. All codes agree on the suppression of the power spectrum to within 1% for scales relevant to Euclid ($k < 1 \, h/\text{Mpc}$). Furthermore, the emulators tested are now accurate enough for real data analysis, matching full $N$-body simulations within 1-2%.
The Validity of Linear Treatment
One of the most important findings is that treating neutrinos using linear perturbation theory on a mesh is remarkably accurate. For the neutrino masses allowed by current experiments, the complex non-linear clumping of neutrinos themselves is a secondary effect that does not significantly bias the main dark matter results at the required precision levels.
Higher-Order Statistics and Halo Properties
In the bispectrum analysis, the results split into two camps: particle-based methods (noisy at large scales) and mesh-based methods (smooth). This highlights that for higher-order statistics, mesh-based neutrino treatments may be superior. Regarding dark matter halos:
- Halo Bias: The ratio of massive-to-massless neutrino bias showed excellent agreement, staying within 1% across almost all scales.
- Halo Mass Function: The codes agreed within ~3% on the number of dark matter halos formed, though some approximate codes (like COLA) tended to underestimate the number of small halos due to resolution limits.
Conclusion
The paper concludes that the "theory error" in current simulations is now smaller than the expected observational error from Euclid, clearing a major hurdle for the mission. While baryonic physics (gas dynamics) and redshift-space distortions remain areas for future integration, the core numerical pipeline for modelling massive neutrinos is now validated at the sub-percent level. This provides a rock-solid foundation for upcoming cosmological surveys to finally weigh the universe's most elusive particles.
Summary updated