orpheus
Efficient estimators for higher-order correlation functions

orpheus is a high-performance Python package for the estimation of second-, third-, and fourth-order correlation functions of scalar and polar fields such as weak lensing shear. To make these calculations computationally tractable, orpheus makes use of a multipole decomposition of the N>2 correlation functions and combines it with hierarchical spatial algorithms, with the computationally intensive operations implemented in parallelized C kernels.

This framework makes the estimation of higher-order statistics feasible for ongoing and forthcoming stage-IV cosmological surveys containing hundreds of millions of objects. As a ballpark estimate, orpheus can accurately determine how the $10^{18}$ triangles formed by a catalogue of one million objects are distributed across configuration-space bins within a few CPU minutes, with the computational complexity scaling approximately linear with the numer of objects.

Quickstart

orpheus is installable from PyPI, so a simple pip install orpheus-npcf should get you the latest pre-compiled version.

The computation of any higher-order correlation function follows the same patten; below we give an example to compute third-order shear statistics. For a fully worked example see the introductory tutorial notebook

import orpheus 

yourshapecat = load_your_catalog()
binning = define_your_binning()

# Define catalog and correlator
cat = orpheus.SpinTracerCatalog(yourshapecat)
ggg = orpheus.GGGCorrelation(binning)

# Process the catalog to obtain 3pcf in multipole basis
ggg.process(cat)

# Transform to real-space basis
ggg.multipoles2npcf()

# Optionally compute integrated statistics
apradii = ...
map3 = ggg.computeMap3(apradii)

Featured statistics

Currently, orpheus contains estimators for the following correlation functions.

In addition, orpheus also implements direct estimators for aperture statistics of arbitrary order for pure scalar and polar correlators.

All correlators are featured in the tutorial notebooks. In there you can also find worked examples on how to measure the statistics on realistic catalogs on the celestial sphere, how to handle tomography and how to customize the level of accuracy of the estimators.

Performance

We showcase the performance of orpheus using the example third-order shear correlations. In the panels below we compare the scaling properties of orpheus' different implementation of the multipole-based estimator and compare it to traditional estimators operating purely in real space:

The different orpheus implementations show substantially better scaling than their real-space counterparts. In particular:

• More configurations: Both tree-based estimators show a much more shallow scaling when increasing the larges search distance; the DoubleTree in particular nearly plateaus.
• Larger datesets: The DoubleTree estimator maintains near-linear scaling with survey depth, making it well suited to increasingly large catalogues.
• Parallel performance: the implementation achieves near-ideal strong scaling up to 32 threads, with diminishing returns at higher thread counts. This is partly due to the non-scaling spatial organisation of the catalog which becomes less significant for larger datasets.

The figure can be reproduced with python benchmarks/scaling.py run followed by python benchmarks/scaling.py plot. The same script can also be used to perform a similar scaling test on your machine.

Numerical validation

The estimators implemented in orpheus are validated against reference results based on a field with analytically tractable correlation functions and aperture statistics. These tests are used to verify the implementations and to quantify the accuracy of the available approximation schemes.

Below we show outcome of the test suite; for a full reference including the derivation of the expressions and for a motivation of the test setup we refer to the following notes.

Installation

orpheus is on PyPI:

pip install orpheus-npcf

Building from source instead compiles a parallelised C/C++ extension at install time. Besides the python requirements you therefore need a C compiler with OpenMP support (GCC, or Clang together with libomp). No external C libraries are needed: the small HEALPix subset used by the curved-sky estimators is shipped with orpheus, in orpheus/src/healpix/.

In case you want to compile the package from source you can do this by first cloning the github repository

git clone git@github.com:lporth93/orpheus.git

or

git clone https://github.com/lporth93/orpheus.git

Then navigate to the cloned directory and install:

cd orpheus
pip install .

For more detailed installation instuctions, troubleshooting and optional flags please consult the corresponding section in the documentation.

Citations

orpheus implements and extends methods developed in several papers. Please cite the publications relevant to the estimators used in your work:

• Three-point functionality: Porth et al. 2024
• Four-point functionality: Porth et al. 2025 and Silvestre-Rosello et al. 2025
• Direct estimator functionality: Porth & Smith 2022
• Two-point functionality: Please provide a reference to the official GitHub repository in a footnote
• Fully spherical estimators: Please also cite the original HEALPix paper: Gorski+2005

In each of the papers, you can find the main equations implemented in orpheus.