The Non-Gaussian Weak-Lensing Likelihood: A Multivariate Copula Construction and Impact on Cosmological Constraints

This paper presents a multivariate copula framework for constructing non-Gaussian weak-lensing likelihoods that better match simulated data than Gaussian approximations, finding that while this approach induces significant shifts in $S_8$ constraints for stage-III surveys, the effect becomes negligible for stage-IV surveys, suggesting standard Gaussian likelihoods remain sufficient for future large-scale analyses.

The Gist

Imagine you are a detective trying to solve a cosmic mystery: What is the universe made of, and how is it shaped?

To do this, astronomers use a tool called Weak Gravitational Lensing. Think of the universe as a giant, invisible trampoline. Massive objects (like galaxies and dark matter) create dips in this trampoline. When light from distant galaxies travels across the universe, it follows the curves of this trampoline, getting slightly stretched or twisted. By measuring these tiny distortions, astronomers can map out the invisible mass of the universe.

However, there's a catch. The universe isn't perfectly smooth; it's a bit "lumpy" and chaotic, especially on the largest scales. This creates a statistical problem for the detectives.

The Problem: The "Perfect Bell Curve" Assumption

For decades, when scientists analyzed this data, they used a mathematical shortcut. They assumed the data followed a Gaussian distribution (also known as a "Bell Curve").

• The Analogy: Imagine you are measuring the height of people in a city. Most people are average height, with fewer very short or very tall people. The data forms a perfect, symmetrical bell shape. This is easy to work with.
• The Reality: In the universe, especially on huge scales, the data is not a perfect bell curve. It's skewed, lopsided, and has "fat tails" (meaning extreme events happen more often than the bell curve predicts).

Using a Bell Curve to analyze lopsided data is like trying to fit a square peg into a round hole. It works okay for small, simple measurements, but as our telescopes get better (like the upcoming Stage-IV surveys), this "square peg" approach starts to give slightly wrong answers about the universe's secrets.

The Solution: The "Copula" Construction

The authors of this paper, Veronika Oehl and Tilman Tröster, have built a new, more flexible tool called a Copula Likelihood.

Here is how it works, using a creative analogy:

1. The Individual Players (Marginals):
First, they looked at each piece of data individually. They realized that while the group of data points is weird, each individual data point actually follows a very specific, known rule.

• Analogy: Imagine a choir. If you listen to just one singer, they might sing a perfect, complex note. But if you listen to the whole choir, the way they harmonize is messy and unpredictable.

2. The Conductor (The Copula):
The "Copula" is the mathematical conductor that figures out how these individual singers (data points) relate to each other.

• Instead of forcing the whole choir to sing a simple, symmetrical song (the Gaussian Bell Curve), the Copula takes the exact, complex notes each singer knows how to sing and figures out the exact rules of how they harmonize together.

3. The Result:
By combining the "exact notes" with the "exact harmony rules," they created a new map of the universe that is much more accurate than the old Bell Curve map.

What Did They Find?

The team tested this new method against computer simulations and real-world survey setups (like the Kilo-Degree Survey and the future LSST).

• Small Surveys (1,000 square degrees): When looking at smaller patches of the sky, the old method (Bell Curve) gave answers that were off by about one standard deviation. In the world of cosmology, that's a big deal! It's like a GPS telling you you're in the next town over when you're actually in the next street.
• Huge Surveys (10,000 square degrees): When looking at the massive sky maps of the future (Stage-IV surveys), the difference between the old method and the new Copula method became negligible.
  • Why? Because when you look at a huge area, the "Law of Large Numbers" kicks in. The chaos of the small lumps averages out, and the data starts to look more like a nice, smooth Bell Curve again.

The Takeaway

"Do we need this complicated new math?"

• For current, smaller surveys: Yes! It helps correct small errors that could lead us to the wrong conclusion about the nature of dark energy or dark matter.
• For the massive future surveys: Probably not strictly necessary for the final result, but it's good to have as a safety check. The authors suggest that while the big surveys will likely be fine with the old math, the shape of the survey map matters. If the map has weird holes or edges, the new Copula method is still the safest bet.

In summary: The authors built a smarter, more flexible mathematical tool to analyze the "lumpy" universe. They proved that while the old, simple tools work fine for the biggest pictures, we need this new, sophisticated tool to get the details right when we are looking at smaller, trickier slices of the cosmos.

Technical Summary

1. Problem Statement

Current and upcoming cosmological surveys (Stage-III and Stage-IV) aim for sub-percent-level precision in constraining cosmological parameters. A critical assumption in standard analyses is that the likelihood of two-point correlation functions (specifically cosmic shear, $\xi_+$) follows a Gaussian distribution. While the Central Limit Theorem often justifies this on small scales, the paper highlights that this assumption breaks down on large angular scales, where the sampling distributions of weak-lensing correlation functions are inherently non-Gaussian.

Existing methods to model this non-Gaussianity, such as Edgeworth expansions, have fundamental limitations: they can yield negative probability densities, are not guaranteed to be proper distributions, and lose accuracy away from the expansion point. Furthermore, while exact one-dimensional marginal likelihoods for masked Gaussian random fields have been derived (Oehl & Tröster 2025, "OT25"), constructing a full multivariate likelihood from these exact marginals has been computationally infeasible due to the complexity of the full characteristic function.

2. Methodology

The authors propose a framework to construct a multivariate non-Gaussian likelihood using Copula theory. This approach decouples the modeling of marginal distributions from the dependence structure between variables.

• Exact Marginals: The framework utilizes exact one-dimensional marginal likelihoods, $p(\xi^+_q)$, for weak-lensing correlation functions on arbitrarily masked skies. These are derived from the characteristic function of quadratic forms in pseudo-$a_{\ell m}$ coefficients (Gaussian random variables).
• Copula Construction:
  • The method maps the exact non-Gaussian marginals to a standard normal space using the inverse cumulative distribution function (CDF) of the marginals and the normal CDF (percent point function).
  • A Gaussian Copula is employed to model the dependence structure between data points. The covariance matrix used in the copula is the exact covariance of the correlation functions, $\Sigma_\xi$, derived from the pseudo-$a_{\ell m}$ covariance and the survey mask.
  • The final multivariate likelihood is constructed as:
    $$p(\boldsymbol{\xi}^+ | \boldsymbol{\theta}) = \rho_{\text{Gauss}}(\mathbf{z}) \prod_{q} p(\xi^+_q | \boldsymbol{\theta})$$
    where $\rho_{\text{Gauss}}$ is the Gaussian copula density and $\mathbf{z}$ represents the transformed data points.
• Computational Optimizations:
  • Eigenvalue Reduction: To compute marginals, the authors exploit the block-diagonal and sparse structure of the combination matrices ($M$) and the pseudo-$a_{\ell m}$ covariance ($\Sigma$). This allows the eigenvalue problem to be reduced to small $2 \times 2$ blocks for cross-correlations, making the calculation of exact marginals tractable.
  • Hybrid Approximation: For high multipoles ($\ell$), where the covariance matrix becomes too large, the authors apply a split: low-$\ell$ modes are treated with exact non-Gaussian marginals, while high-$\ell$ modes (where the distribution is effectively Gaussian due to the Central Limit Theorem) are approximated as Gaussian.
  • Scale-Dependent Marginals: For very small angular scales (sub-degree), the marginals are replaced with Gaussian distributions to further reduce computational cost without significant loss of accuracy.

3. Key Contributions

• First Application to Full Data Vectors: This is the first application of a copula likelihood with analytic non-Gaussian marginals to a full, multidimensional weak-lensing correlation-function data vector. Previous works were limited to power spectra or one-dimensional fields.
• Exact Marginal Integration: The framework successfully integrates the exact analytical marginals derived in OT25 into a multivariate context, avoiding the need to fit marginal distributions to simulations.
• Efficient Evaluation: By using the copula approach, the authors bypass the need to compute the full $N_d$-dimensional characteristic function (which requires retrieving eigenvalues $N^{N_d}$ times), enabling point-wise evaluation of the likelihood for Bayesian inference.

4. Results

The authors tested the framework using mock data vectors based on KiDS-1000 and Stage-IV (LSST/Euclid) survey geometries (1,000 deg² and 10,000 deg² masks).

• Validation against Simulations:
  • The copula likelihood was compared against $10^6$ simulated realizations of correlated Gaussian random fields.
  • The copula likelihood matched the simulated sampling distributions significantly better than a standard Gaussian likelihood, particularly in the tails and asymmetric regions of the distribution.
  • The Gaussian likelihood showed systematic deviations and a higher number of outliers compared to the simulation histograms.
• Impact on Cosmological Constraints ($S_8$ and $\Omega_m$):
  • 1,000 deg² Surveys (Stage-III): When analyzing large scales (excluding small scales < 30 arcmin), the use of the non-Gaussian copula likelihood resulted in a shift in the posterior mean of $S_8$ by approximately one standard deviation compared to the Gaussian likelihood. The shift was towards lower values of $S_8$. The shape of the posterior contours also changed significantly.
  • 10,000 deg² Surveys (Stage-IV): For larger survey areas, the shift in the posterior mean became negligible. The increased number of modes and the averaging effect of the larger survey area reduce the relative impact of the non-Gaussianity on the final constraints.
  • Scale Dependence: The impact of the likelihood choice is most pronounced when large angular scales (where non-Gaussianity is strongest) are included. Including small scales (which are Gaussian) dilutes the shift.
• Full Parameter Space Sampling: The framework was successfully applied to sample the full parameter space (varying $\Omega_m, S_8, h, n_s$, etc.) using MCMC (emcee). The results confirmed that the shifts observed in fixed-grid analyses persist in full sampling, and the shifts are not artifacts of the number of parameters varied.

5. Significance and Conclusion

• Validity of Gaussian Approximation: The study concludes that while Gaussian likelihoods are sufficient for Stage-IV surveys (large areas) where the non-Gaussian effects are averaged out, they may introduce significant biases (order of 1$\sigma$) for Stage-III surveys or analyses focusing specifically on large scales.
• Survey Geometry Sensitivity: The magnitude of the parameter shift depends heavily on the specific survey mask geometry and the structure of the data vector, not just the total survey area.
• Future Recommendations: The authors recommend that for final Stage-IV analyses (e.g., LSST, Euclid), a specific test using this framework should be conducted to ensure the mask geometry does not induce unexpected biases, even if the overall area is large.
• Generalizability: The copula framework is general and can be applied to other non-Gaussian summary statistics in cosmology, such as 21 cm maps, gravitational wave backgrounds, or combined galaxy clustering and weak lensing, provided the one-dimensional marginals are known.

In summary, Oehl and Tröster provide a robust, computationally feasible method to incorporate exact non-Gaussianity into weak-lensing likelihoods, demonstrating that while the impact on large-scale Stage-IV constraints may be minimal, the method is crucial for understanding biases in current and smaller-scale surveys.