MatchCake

_{Logo by Tarik El-Khateeb / Xanadu}

Description

MatchCake is a Python package that provides a new PennyLane device for simulating a specific class of quantum circuits called Matchgate circuits or matchcircuits. These circuits are made with matchgates, a class of restricted quantum unitaries that are parity-preserving and operate on nearest-neighbor qubits. These constraints lead to matchgates being classically simulable in polynomial time.

Additionally, this package provides quantum kernels made with scikit-learn API allowing the use matchcircuits as kernels in quantum machine learning algorithms. One way to use these kernels could be in a Support Vector Machine (SVM).

Note that this package is built on PennyLane and PyTorch. This means that only the NumPy and PyTorch backends are compatible. Other backends provided by Autoray, such as JAX and TensorFlow, are not supported. We highly recommend using PyTorch as the backend when working with MatchCake.

Requirements

Every one of these Python versions is tested on Linux (x86-64 and aarch64) and on Windows on each pull request, in continuous integration. On Linux, a glibc based distribution is required. PyTorch publishes no musl wheels, so MatchCake cannot be installed on Alpine.

macOS is not supported. MatchCake contains no platform specific code and may well run there, but it is not tested. If you need macOS support, please open an issue and we will look into adding it to the continuous integration matrix.

Installation

MatchCake is published on PyPI as matchcake.

Set up a project or an environment

Install MatchCake into a virtual environment and not into the system interpreter, which recent Linux distributions refuse with error: externally-managed-environment. If you do not have one yet, create it:

uv init myproject && cd myproject                    # uv
poetry new myproject && cd myproject                 # poetry (then cap requires-python, see below)
python -m venv .venv && source .venv/bin/activate    # plain virtual environment

On Windows, activate the virtual environment with .venv\Scripts\activate instead.

Install MatchCake

A few things useful to know during the installation:

• uv add and poetry add are project commands: they record a dependency in the pyproject.toml of the project you are standing in. In a directory without one they stop with No pyproject.toml found in current directory or any parent directory. Note that uv add also searches parent directories, so running it inside an unrelated project modifies that project. Use uv pip install matchcake when you only want the package in an environment.
• poetry add matchcake additionally requires your project to put an upper bound on its Python range, for example requires-python = ">=3.11,<3.15". poetry new writes an open ended range such as >=3.12, and poetry then declines to resolve because torchpfaffian supports Python <3.15 only.
• The source row needs git available on your PATH. It installs the latest release from the main branch; see below for the development branch.

Last unstable version

To install the development branch, use the @dev reference:

pip install "git+https://github.com/MatchCake/MatchCake@dev"

The uv equivalent is uv add "matchcake @ git+https://github.com/MatchCake/MatchCake@dev".

PyTorch build and installation size

MatchCake depends on PyTorch, and on Linux the default PyTorch wheel on PyPI is the CUDA build. A plain pip install matchcake therefore downloads PyTorch together with around twenty NVIDIA packages and occupies roughly 6 GB, whether or not the machine has a GPU. This comes from how PyTorch is packaged, not from MatchCake.

For a CPU only environment, install PyTorch from the PyTorch CPU index first, then install MatchCake on top. The requirement is already satisfied at that point, so nothing pulls the CUDA build in, and the result is about 1.6 GB with no NVIDIA packages:

pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install matchcake

Swap the index for a specific CUDA version, for instance https://download.pytorch.org/whl/cu128 for CUDA 12.8 or https://download.pytorch.org/whl/cu130 for CUDA 13.0.

With uv, declare the index in your own pyproject.toml and add torch alongside matchcake:

[[tool.uv.index]]
name = "pytorch-cpu"
url = "https://download.pytorch.org/whl/cpu"
explicit = true

[tool.uv.sources]
torch = { index = "pytorch-cpu" }

uv add matchcake torch

Listing torch explicitly is important: [tool.uv.sources] only redirects dependencies that your own project declares, so a torch pulled in solely through matchcake still comes from PyPI.

MatchCake declares cpu, cu128 and cu130 extras, but they only select these indexes when MatchCake itself is built from a clone, through [tool.uv.sources] in its pyproject.toml. That mechanism belongs to uv and is not part of the metadata published to PyPI, so pip install "matchcake[cpu]" and uv add matchcake --extra cpu resolve exactly the same PyTorch as no extra at all. Use the index based recipe above. The extras are documented for contributors in CONTRIBUTING.md.

Quick Usage Preview

Quantum Circuit Simulation with MatchCake

import matchcake as mc
import pennylane as qml
import numpy as np
from pennylane.ops.qubit.observables import BasisStateProjector

# Create a Non-Interacting Fermionic Device
nif_device = mc.NonInteractingFermionicDevice(wires=4)
initial_state = np.zeros(len(nif_device.wires), dtype=int)

# Define a quantum circuit
def circuit(params, wires, initial_state=None):
    qml.BasisState(initial_state, wires=wires)
    for i, even_wire in enumerate(wires[:-1:2]):
        idx = list(wires).index(even_wire)
        curr_wires = [wires[idx], wires[idx + 1]]
        mc.operations.CompRxRx(params, wires=curr_wires)
        mc.operations.CompRyRy(params, wires=curr_wires)
        mc.operations.CompRzRz(params, wires=curr_wires)
    for i, odd_wire in enumerate(wires[1:-1:2]):
        idx = list(wires).index(odd_wire)
        mc.operations.fSWAP(wires=[wires[idx], wires[idx + 1]])
    projector: BasisStateProjector = qml.Projector(initial_state, wires=wires)
    return qml.expval(projector)

# Create a QNode
nif_qnode = qml.QNode(circuit, nif_device)
qml.draw_mpl(nif_qnode)(np.array([0.1, 0.2]), wires=nif_device.wires, initial_state=initial_state)

# Evaluate the QNode
expval = nif_qnode(np.random.random(2), wires=nif_device.wires, initial_state=initial_state)
print(f"Expectation value: {expval}")

Data Classification with MatchCake

from matchcake.ml.kernels import FermionicPQCKernel
from matchcake.ml.visualisation import ClassificationVisualizer
from sklearn import datasets
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import MinMaxScaler
from sklearn.pipeline import Pipeline
from sklearn.svm import SVC

# Load the iris dataset
X, y = datasets.load_iris(return_X_y=True)
x_train, x_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=0)

# Create and fit the model
pipeline = Pipeline([
    ('scaler', MinMaxScaler(feature_range=(0, 1))),
    ('kernel', FermionicPQCKernel(n_qubits=4, rotations="X,Z").freeze()),
    ('classifier', SVC(kernel='precomputed')),
])
pipeline.fit(x_train, y_train)

# Evaluate the model
test_accuracy = pipeline.score(x_test, y_test)
print(f"Test accuracy: {test_accuracy * 100:.2f}%")

# Visualize the classification
viz = ClassificationVisualizer(x=X, n_pts=1_000)
viz.plot_2d_decision_boundaries(model=pipeline, y=y, show=True)

Tutorials

• MatchCake Basics
• Compute Expectation Values with MatchCake
• Iris Classification with MatchCake
• Nystroem Kernel Approximation

For Developers

To contribute to the development of MatchCake, please refer to the contributing guidelines.

Notes

• This package is still in development and some features may not be available yet.
• The documentation is still in development and may not be complete yet.

About

This work was supported by the Ministère de l'Économie, de l'Innovation et de l'Énergie du Québec through its Research Chair in Quantum Computing, an NSERC Discovery grant, and the Canada First Research Excellence Fund.

Important Links

• Documentation at https://MatchCake.github.io/MatchCake/.
• Github at https://github.com/MatchCake/MatchCake/.

Found a bug or have a feature request?

• Click here to create a new issue.

License

Apache License 2.0

Citation

Repository:

@misc{matchcake_Gince2023,
  title={MatchCake},
  author={Jérémie Gince},
  year={2023},
  publisher={Université de Sherbrooke},
  url={https://github.com/MatchCake/MatchCake},
}

Fermionic Machine Learning Paper

Fermionic Machine Learning is a work presented at the 2024 IEEE International Conference on Quantum Computing and Engineering (QCE). The paper compares unconstrained quantum kernel methods with constraint-based kernels derived from matchgate (free-fermionic) circuits, and benchmarks their performance on supervised classification tasks. All free-fermionic kernels considered in this work were simulated using MatchCake.

IEEE Xplore paper:

@INPROCEEDINGS{10821385,
  author={Gince, Jérémie and Pagé, Jean-Michel and Armenta, Marco and Sarkar, Ayana and Kourtis, Stefanos},
  booktitle={2024 IEEE International Conference on Quantum Computing and Engineering (QCE)},
  title={Fermionic Machine Learning},
  year={2024},
  volume={01},
  number={},
  pages={1672-1678},
  keywords={Runtime;Quantum entanglement;Computational modeling;Benchmark testing;Rendering (computer graphics);Hardware;Kernel;Integrated circuit modeling;Quantum circuit;Standards;Quantum machine learning;quantum kernel methods;matchgate circuits;fermionic quantum computation;data classification},
  doi={10.1109/QCE60285.2024.00195}
}

ArXiv paper:

@misc{gince2024fermionic,
      title={Fermionic Machine Learning},
      author={Jérémie Gince and Jean-Michel Pagé and Marco Armenta and Ayana Sarkar and Stefanos Kourtis},
      year={2024},
      eprint={2404.19032},
      archivePrefix={arXiv},
      primaryClass={quant-ph}
}