pentapy

Welcome to pentapy

Purpose

pentapy is a toolbox to deal with pentadiagonal matrices in Python.

Pentadiagonal linear equation systems arise in many areas of science and engineering: e.g. when solving differential equations, in interpolation problems, or in numerical schemes like finite difference.

Installation

The package can be installed via pip. On Windows you can install WinPython to get Python and pip running.

pip install pentapy

There are pre-built wheels for Linux, MacOS and Windows for most Python versions.

To get the scipy solvers running, you have to install scipy or you can use the following extra argument:

pip install pentapy[all]

Instead of "all" you can also typ "scipy" or "umfpack" to get one of these specific packages.

Citation

If you use pentapy in your publication, please cite it:

Müller, (2019). pentapy: A Python toolbox for pentadiagonal linear systems. Journal of Open Source Software, 4(42), 1759, https://doi.org/10.21105/joss.01759

To cite a certain release, have a look at the Zenodo site: https://doi.org/10.5281/zenodo.2587158

References

The solver is based on the algorithms PTRANS-I and PTRANS-II presented by Askar et al. 2015.

Documentation and Examples

You can find the documentation under https://pentapy.readthedocs.org.

Solving a pentadiagonal linear equation system

This is an example of how to solve a LES with a pentadiagonal matrix.

import numpy as np
import pentapy as pp

size = 1000
# create a flattened pentadiagonal matrix
M_flat = (np.random.random((5, size)) - 0.5) * 1e-5
V = np.random.random(size) * 1e5
# solve the LES with M_flat as row-wise flattened matrix
X = pp.solve(M_flat, V, is_flat=True)

# create the corresponding matrix for checking
M = pp.create_full(M_flat, col_wise=False)
# calculate the error
print(np.max(np.abs(np.dot(M, X) - V)))

This should give something like:

4.257890395820141e-08

Performance

In the following a couple of solvers for pentadiagonal systems are compared:

• Solver 1: Standard linear algebra solver of Numpy np.linalg.solve
• Solver 2: scipy.sparse.linalg.spsolve
• Solver 3: Scipy banded solver scipy.linalg.solve_banded
• Solver 4: pentapy.solve with solver=1
• Solver 5: pentapy.solve with solver=2

The implementations of pentapy are almost one order of magnitude faster than the scipy algorithms for banded or sparse matrices.

The performance plot was created with perfplot. Have a look at the script: examples/03_perform_simple.py.

Requirements:

• NumPy >= 1.14.5

Optional

• SciPy
• scikit-umfpack

Contact

You can contact us via [e-mail].

License

MIT © 2019 - 2023