PixPoint

Correlation functions of HEALPix maps, fast.

Documentation · Getting started · Installation · Citing

PixPoint measures two-point, three-point, integrated three-point and one-point statistics of cosmological fields on HEALPix maps: cosmic shear, galaxy clustering, galaxy-galaxy lensing, convergence. You give it maps and weights, it gives you labelled results. The pixel pairs of a footprint are found once and reused for every map, so a simulation suite costs one pair search and one fast pass per map, on a CPU or, with the same code, on NVIDIA GPUs.

from pixpoint import Global2PCF

g2 = Global2PCF(nside=512, mask=mask, nbins=20, theta_min=2.5, theta_max=250)
g2.preprocess()                          # find the pixel pairs once (save_pairs / load_pairs to reuse)

res = g2.gg(shear_maps, weights)         # shear_maps (nz, 2, npix), weights (nz, npix)
res.theta                                # bin centres, arcmin
res.xi_p, res.xi_m                       # (n_combinations, nbins)
res.combinations                         # which tomographic bins each row is

g2.kk(delta_maps, w_cov).xi              # clustering w(theta)
g2.kg(delta_maps, shear_maps, w_lens, w_source).xi   # galaxy-galaxy lensing gamma_t

What it measures

Method names follow TreeCorr's field codes: g is a shear field, k a scalar field such as a density contrast or convergence map. Every estimator is tomographic, and every result carries the labels of its rows.

Why PixPoint

• Exact pixel-pair estimators. The two-point functions are ratios of weighted sums over every pair of unmasked pixels, and agree with TreeCorr on the pixel centres to rounding.
• Find pairs once, measure many maps. A full-footprint tomographic shear 2PCF of DES Y3 at nside 512 (6.7 × 10⁸ pairs) takes 33 ms per map on an A100 after a one-off pair search; the same code runs on a laptop CPU.
• Everything in one vocabulary. Shear, clustering, galaxy-galaxy lensing and 3×2pt from one object; 3PCF multipoles for any mix of spin-0 and spin-2 fields; the integrated 3PCF with its aperture filters; aperture-mass PDFs and moments.
• Built for suites. Streaming HDF5 writers with provenance and resume, loaders that hide the upload behind the measurement, multi-GPU splitting, and an optional treecode that cuts the pair count at large separations.
• Checked. Every GPU kernel has a NumPy twin and a brute-force reference test at machine precision; mypy --strict clean; ships py.typed.

Installation

Python 3.12 or newer:

pip install pixpoint             # CPU
pip install "pixpoint[gpu]"      # + CuPy for NVIDIA GPUs (CUDA 12 build)

The gpu extra installs cupy-cuda12x, which runs on any driver that supports CUDA 12 or newer; for a CUDA 13 build install pixpoint and cupy-cuda13x instead. PixPoint picks the GPU automatically. Details, including the development install, in the installation guide.

Learn more

• Concepts: maps versus catalogues, weights and masks, what the estimators compute.
• Two-point functions, three-point functions, integrated 3PCF, one-point statistics.
• Running on GPUs and files and streaming.
• Conventions (shear signs, bin labels, the 3PCF projection) and the formal estimator definitions.

PixPoint was called CosmoFuse before version 8.0.0; import CosmoFuse is now import pixpoint, nothing else in the API changed, and old pair files are still read.

Citing

Until PixPoint has a software paper of its own, please cite the paper it was developed for, Gebauer et al. 2026, SBi3PCF: simulation-based inference with the integrated 3PCF, JCAP 06 (2026) 036 (arXiv:2510.13805), and name the PixPoint version you used. Please also cite the papers behind the estimators you use. The BibTeX entry and those papers are on the Citing page.

Contributing

Issues and pull requests are welcome; see CONTRIBUTING.md. PixPoint is released under the MIT licence.