redistpy 2.0.0

A novel reinterpretation method for high-energy physics results.

Overview

This is the implementation of a reweighing method for the reinterpretation of binned analyses in high energy physics. The method is based on calculating the change in the distributions of observables, given changes in the decay channel's kinematic distributions.

Check out the paper here.

Installation

You can install redist from pypi

pip install redistpy

Manual installation

You can install manually with

# Clone this repository
git clone https://github.com/lorenzennio/redist.git

# Install
pip install -e redist

Dependencies

This implementation is based on the pyhf software for statistical inference. The examples use the EOS software to calculate theoretical predictions.

Gradients with JAX (optional)

redist runs on any pyhf backend. On the JAX backend the likelihood becomes differentiable, so it can be used with jax.grad, jax.jit and jax.vmap:

pip install redistpy[jax]

import pyhf, jax, jax.numpy as jnp
from redist import modifier

pyhf.set_backend("jax")          # set the backend *before* building the modifier

model = modifier.load("model.json", alt_dist, null_dist)
grad = jax.grad(lambda pars: model.logpdf(pars, data)[0])(pars)

The distributions must accept broadcast arrays and be written with operations JAX can trace: jax.numpy rather than scipy, and no Python if on parameter values. Adaptive quadrature cannot be traced, so the fallback described below does not apply here.

Theory codes that are not written in JAX — EOS, for example — therefore cannot be differentiated through. Using them still works on the NumPy backend.

Bayesian inference (optional)

If you want to perform Bayesian inference with redist (or pyhf) you'll need to install bayesian_pyhf.

You can do so with::

pip install git+https://github.com/malin-horstmann/bayesian_pyhf.git

For visualization of the posterior distribution, corner is very useful:

pip install corner

Bin integrals

The bin integrals use fixed-order Gauss-Legendre quadrature, which calls each distribution once with one broadcast array per kinematic dimension. A theory code that can only be evaluated a point at a time — EOS, for example — is detected when the modifier is built and falls back to adaptive quadrature, one to two orders of magnitude slower. cmod.quad reports which rule was picked, and quad="gauss" or quad="nquad" forces one.

Contact

If you come across a bug, have an issue or a question, please file an issue. For further inquiries, you can talk to us via Discord.

License

MIT