pyBumpHunter 0.5.0

Python implementation of the BumpHunter algorithm used by HEP community.

pyBumpHunter

This is a python version of the BumpHunter algorithm, see arXiv:1101.0390, G. Choudalakis, designed to find localized excess (or deficit) of events in a 1D or 2D distribution.

The main BumpHunter function will scan a data distribution using variable-width window sizes and calculate the p-value of data with respect to a given background distribution in each window. The minimum p-value obtained from all windows is the local p-value. To cope with the "look-elsewhere effect" a global p-value is calculated by performing background-only pseudo-experiments.

The BumpHunter algorithm can also perform signal injection tests where more and more signal is injected in toy data until a given signal significance (global) is reached (signal injection not available in 2D yet).

Content

• pyBumpHunter : The pyBumpHunter package
• example : Folder containing a set of example scripts and notebooks
• example/results : Folder containing the outputs of example scripts
• test : Folder containing the testing scripts (based on pytest)
• data/data.root : Toy data used in the examples and tests
• data/gen_data.C : Code used to generate the toy data with ROOT

Dependencies

Requires Python >= 3.9.

BumpHunter depends on the following python libraries :

• numpy
• scipy
• matplotlib

pyBumpHunter wiki

Examples

The examples provided in example.py and test.ipynb require the uproot package in order to read the data from a ROOT software file.

The data provided in the example consists of three histograms: a steeply falling 'background' distribution in a [0,20] x-axis range, a 'signal' gaussian shape centered on a value of 5.5, and a 'data' distribution sampled from background and signal distributions, with a signal fraction of 0.15%. The data file is produced by running gen_data.C in ROOT.

In order to run the example script, simply type python3 example.py in a terminal.

You can also open the example notebook with jupyter or binder.

• Bump hunting:

• Tomography scan:

• Test statistics and global p-value:

See the wiki for a detailed overview of all the features offered by pyBumpHunter.

To do list

• Run BH on 2D histograms

Authors and contributors

Louis Vaslin (main developper), Julien Donini

Thanks to Samuel Calvet for his help in cross-checking and validating pyBumpHunter against the (internal) C++ version of BumpHunter developped by the ATLAS collaboration.