A small but powerful interface to generate theoretical likelihoods from ``UVPSpec`` objects. status

Full Documentation#

Read the docs.


  • Ingests data output from hera_pspec: power spectra, covariance matrices and window functions.

  • Agnostic to theory code (i.e. run 21cmFAST or ARES or any other model)

  • Outputs a log-likelihood to be used in parameter inference, but is sampler agnostic.


Clone/download the repo and pip install ., or pip install git+git://

If developing:

git clone
cd pspec_likelihood
pip install -e .[dev]
pre-commit install


Import like this:

from pspec_likelihood import DataModelInterface, Gaussian

To construct a likelihood, you first need to construct the DataModelInterface, for which you will specify the data, its covariance, a window function, and a model both for the theory and the systematics. This class contains all the methods required to compute the model/systematics and transform it consistently to data-space.

Secondly, you need to construct a PSpecLikelihood, via one of its concrete sub-classes. Examples of such subclasses are Gaussian and GaussianLinearSystematics. The reason these are their own class, instead of being part of the DataModelInterface, is for the sake of modularity and extensibility. This allows different actual likelihoods to be computed given the data, and new likelihoods to be implemented with ease. The basic requirement of a PSpecLikelihood subclass is that it must implement the loglike(theory_params, sys_params) method, which goes and computes the actual log-likelihood given a set of parameters. It has access to the DataModelInterface object through its model attribute. So, eg.:

likelihood = Gaussian(
    model = DataModelInterface(...)

likelihood.loglike(theory_params, sys_params)


From v0.1.0, pspec_likelihood will be using strict semantic versioning, such that increases in the major version have potential API breaking changes, minor versions introduce new features, and patch versions fix bugs and other non-breaking internal changes.

If your package depends on pspec_likelihood, set the dependent version like this:



Indices and tables#