# PyMatching

PyMatching is a Python package for decoding quantum codes with the minimum-weight perfect matching (MWPM) decoder, and is designed to be fast and easy to use.

While a Python package such as NetworkX can also be used to implement MWPM, it is far too slow to be used for large fault-tolerance simulations, which often require matching graphs with many thousands of nodes. On the other hand, the widely used C++ BlossomV library is fast, but using it to decode quantum codes also requires path-finding algorithms, which must also be implemented in C++ for a fast implementation. Furthermore, attempting to solve the full matching problem even with BlossomV can become prohibitively expensive for matching graphs with more than a few thousand nodes, since the complexity is worse than quadratic in the number of nodes. BlossomV is also not open-source since it does not have a permissive license.

PyMatching is typically faster than a BlossomV/C++ implementation of the full matching problem, while being easy to use in conjunction with numpy, scipy and NetworkX using the Python bindings. The core algorithms and data structures are implemented in C++ for good performance (with the help of the open-source LEMON and Boost Graph libraries), using a local variant of the matching decoder given in the Appendix of https://arxiv.org/abs/2010.09626, which empirically has an average runtime roughly linear in the number of nodes and gives the same output as full matching in practice. Since PyMatching uses the open-source LEMON C++ library for the Blossom algorithm, which has similar performance to Kolmogorov’s BlossomV library, both PyMatching and its dependencies have permissive licenses. PyMatching can be applied to any quantum code for which defects come in pairs (or in isolation at a boundary), and it does not require knowledge of the specific geometry used.

Compared to a pure Python NetworkX implementation of MWPM, PyMatching can be orders of magnitude faster, as shown here for the toric code under an independent noise model at $$p=0.05$$:

For more information, please also see the PyMatching paper. To make a feature request or report a bug, please visit the PyMatching GitHub repository.