Bispectrum-window convolution via Hankel transform

Kevin Pardede, Federico Rizzo, Matteo Biagetti, Emanuele Castorina, E. Sefusatti, Pierluigi Monaco

TL;DR

Researchers have developed a high-performance method to account for survey geometry—the "window function"—when analyzing the galaxy bispectrum. By leveraging a 2D-FFTLog algorithm and Hankel transforms, they reduced a complex high-dimensional convolution into a simple matrix multiplication that executes in under a second, enabling precise cosmological constraints from upcoming galaxy surveys.

What Problem Does This Paper Address?

In cosmology, we study the Large-Scale Structure (LSS) of the universe to understand dark energy and gravity. While the "power spectrum" (the 2-point correlation) is the workhorse of the field, the bispectrum (the 3-point correlation) contains vital information about non-linear gravity and how galaxies trace dark matter (galaxy bias).

However, we never observe a perfect, infinite universe. We see it through a "window"—the specific footprint of a telescope's survey and its selection depth. This window distorts the signal, "blurring" the theoretical bispectrum. To compare a theoretical model to real data, you must perform a "convolution" of the model with the window function.

For the bispectrum, this convolution involves a daunting 6D integral. Standard numerical integration is far too slow for a Likelihood analysis (like MCMC), where the model must be recalculated thousands of times. Previous attempts often relied on crude approximations that fail on large scales, exactly where some of the most interesting physics (like primordial non-Gaussianity) resides.

Methods

The authors propose a method to perform the exact convolution using a 2D-FFTLog algorithm.

  1. Hankel Transforms: The core idea is to transform the bispectrum into configuration space (real space), where the convolution becomes a simple multiplication with the window's own 3-point correlation function.
  2. 2D-FFTLog: FFTLog is a standard tool in cosmology for performing fast Hankel transforms on logarithmically spaced grids. The authors extend this to two dimensions to handle the two independent wavevectors ($k_1, k_2$) that define a triangle in the bispectrum.
  3. Mixing Matrix: The result of this process is a "mixing matrix." Once this matrix is pre-computed for a specific survey geometry, the convolution for any new theoretical model is reduced to a basic matrix-vector multiplication.
  4. Coordinate Systems: They carefully reconcile different coordinate definitions (like Scoccimaro’s vs. others) to ensure the azimuthal angles are handled correctly during the transform.

Datasets

The authors validated their method using several high-quality cosmological datasets:

  • Minerva Simulations: A set of 298 N-body simulations used to test if the method can recover the correct "bias" parameters (how galaxies cluster relative to dark matter) in a realistic setting.
  • Pinocchio Mocks: A massive set of 10,000 fast mock catalogs. This large volume allowed the authors to suppress statistical noise and isolate tiny systematic errors in their algorithm.
  • Spherical Toy Model: To test the math in a controlled environment, they used a perfectly spherical window function, which allows for an analytical (exact) solution to compare against.

Key Insights and Ideas

  • Speed vs. Accuracy: The "1D approximation" (a common shortcut that treats the bispectrum convolution like a power spectrum convolution) is surprisingly resilient but ultimately inaccurate, showing 6-7% errors on large scales. The 2D-FFTLog method brings this error down to the ~1% level.
  • The "Mixing Matrix" Paradigm: By shifting the heavy lifting to a pre-computation step (which takes a few hours), the actual analysis step becomes nearly instantaneous. This makes the bispectrum as "cheap" to evaluate as the power spectrum in a likelihood pipeline.
  • Squeezed Configurations: The method is particularly robust for "squeezed" triangles (where one side is much shorter than the others). These shapes are the most sensitive to the window function and are crucial for detecting physics from the early universe.

Key Results

  • Computational Efficiency: The matrix multiplication takes less than one second for a standard set of 360 triangle configurations.
  • Unbiased Constraints: In a full MCMC analysis of the Minerva simulations ($100\, h^{-3} \, {\rm Gpc}^3$ total volume), the method successfully recovered the underlying physical parameters (halo bias) without any systematic offset.
  • Precision: Comparison with 10,000 mocks showed that the 2D-FFTLog method matches the "ground truth" at the percent level, significantly outperforming previous approximations.

Limitations and Future Work

  • Complex Geometries: This paper primarily tests a spherical window. Real-world surveys have complex masks with holes and jagged edges. While the math holds, measuring the window's 3-point correlation function for complex masks is the next technical hurdle.
  • Redshift Space Multipoles: While the theoretical framework for redshift space (where galaxy positions are distorted by their motion) is presented, the numerical tests were performed in real space. Testing higher-order multipoles (quadrupole, etc.) is a necessary next step.
  • Wide-Angle Effects: The authors used the "plane-parallel" approximation. For very wide-area surveys, additional corrections might be needed to account for the fact that the "line-of-sight" changes across the sky.

Summary updated

Abstract

Abstract We present a method to perform the exact convolution of the model prediction for bispectrum multipoles in redshift space with the survey window function. We extend a widely applied method for the power spectrum convolution to the bispectrum, taking advantage of a 2D-FFTlog algorithm. As a preliminary test of its accuracy, we consider the toy model of a spherical window function in real space. This setup provides an analytical evaluation of the 3-point function of the window, and therefore it allows to isolate and quantify possible systematic errors of the method. We find that our implementation of the convolution in terms of a mixing matrix shows differences at the percent level in comparison to the measurements from a very large set of mock halo catalogs. It is also able to recover unbiased constraints on halo bias parameters in a likelihood analysis of a set of numerical simulations with a total volume of 100 h -3 Gpc 3 . For the level of accuracy required by these tests, the multiplication with the mixing matrix is performed in the time of one second or less.
BibTeX
@article{pardede2022bispectrum,
  author = {Kevin Pardede and Federico Rizzo and Matteo Biagetti and Emanuele Castorina and E. Sefusatti and Pierluigi Monaco},
  title = {{Bispectrum-window convolution via Hankel transform}},
  journal = {Journal of Cosmology and Astroparticle Physics},
  year = {2022},
  doi = {10.1088/1475-7516/2022/10/066},
}

Loading related content...

Figure